EVIDENCE FROM THE ENGLISH LONGITUDINAL STUDY OF AGEING. in the Faculty of Humanities

Transcription

1 IS WORKING BEYOND STATE PENSION AGE BENEFICIAL FOR HEALTH? EVIDENCE FROM THE ENGLISH LONGITUDINAL STUDY OF AGEING A thesis submitted to the University of Manchester for the degree of PhD in the Faculty of Humanities 2014 KATHERINE MATTHEWS SCHOOL OF SOCIAL SCIENCES CCSR

2 1 Contents List of tables... 7 Thesis abstract Declaration and Copyright Statement Chapter 1: Introduction Rationale for the research: ageing and state pension age in the UK Research questions The health effects of working beyond SPA The impact of work quality on health in later-life Overview of the thesis Chapter 2: Systematic Review of the Health Effects of Later-Life Employment Ageing, employment and health: theoretical perspectives Method and results of the initial search for literature An overview of the studies included in the review Key outcomes of the systematic review Depression Self-rated health Other health outcomes Cognitive functioning Physical functioning Mental health (excluding depression) The role of work quality Overview and conclusions Chapter 3: Meta-analysis of studies concerning the effects of later-life employment on health outcomes Calculating the pooled estimate in meta-analysis Measuring heterogeneity in meta-analysis Heterogeneity and the choice of final model: fixed-effects versus random-effects Choice of statistic to measure heterogeneity Q, τ 2 and I Rationale and data for present meta-analysis Inclusion criteria Meta-analysis and repeated measures design... 88

3 2 3.4 Definition of the variables analysed in the meta-analysis Depression Self-rated health Employment Retirement Study populations Meta-analysis of Cohen s d from different effect types Conversion of all effects to Cohen s d Results of the meta-analysis of combined effects: later-life employment on depression Results of the meta-analysis of combined effects: later-life employment on self-rated health Subgroup analysis Identification of potential sources of heterogeneity Crude data versus study-provided estimates Inclusion versus exclusion of early retirees Repeated measures versus cross-sectional design Single-gender versus mixed-gender groups Country-specific studies (UK and USA) Results of the subgroup analysis of later-life employment on depression Subgroup analysis of later-life employment and depression: heterogeneity Subgroup analysis of later-life employment and depression: pooled estimates Results of the subgroup analysis of later-life employment on self-rated health Subgroup analysis of later-life employment and self-rated health: heterogeneity Subgroup analysis of later-life employment and self-rated health: pooled estimates Conclusions, limitations and rationale for future research Conclusions of the meta-analysis Limitations and rationale for further research on the topic of later-life employment and health outcomes Chapter 4: Data and Research Methodology Key Methodological aims of the research Bias and causality in observational studies Randomisation: the Gold Standard Selection bias and confounding in observational studies

4 3 4.3 Data and methodology for the current research Overview of the data: The English Longitudinal Study of Ageing (ELSA) Selecting the sample The exposure and control variables Later-life employment and retirement Work quality The outcome variables Depression Self-rated health Cognitive function Justification for use of the outcome variables The covariates Statistical analysis Dealing with unobserved confounding in observational studies Summary of the chapter Chapter 5: Sample and Descriptive Analysis Comparing the selected sample to the ELSA population Associations between later-life employment, socio-demographic characteristics and health Associations between employment status and health Associations between employment status, gender and health Associations between employment status, wealth and health Associations between employment status, property ownership and health Associations between employment status, socio-economic classification and health Associations between employment status, qualification level and health Associations between employment status, other socially productive activities and health Associations between employment status, ADL score and health Work Characteristics of the selected sample Associations between self-employment and health Associations between employment type and health Associations between hours worked, employer and health Associations between work quality and retirement, socio-demographic characteristics and health

5 Associations between work quality and retirement and health Associations between work quality and retirement, gender and health Associations between work quality and retirement, socio-demographic characteristics and health Associations between work quality and retirement and other activities Associations between work quality and retirement, ADL score and health Overview of chapter Overview of the sample characteristics Overview of the relationships between health and employment Heterogeneity, selection bias and the forthcoming analyses Chapter 6: Health Trajectories and Later-Life Employment using Spline Models Data used in this chapter Results of the piecewise spline regression analysis Depression Self-rated health Cognitive function Conclusions and discussion Chapter 7: Propensity Score Matching The propensity score matching approach Rationale for use of propensity score matching Assumptions of the propensity score Matching on the propensity score Propensity score matching techniques Nearest neighbor matching Caliper matching Radius matching Kernel matching Considerations to be made regarding choice of algorithm Data analysis using propensity score matching: the health effects of later-life employment Selection of covariates for calculation of the propensity score Results of propensity score matching Assessing the quality of the propensity score matching The relationship between work quality and health in later-life

6 The sample Estimation of the propensity score Matching on the propensity score Poor quality versus good quality employment Poor quality work versus retirement Good quality work versus retirement Assessing the quality of propensity score matching Sensitivity analysis with Rosenbaum bounds Conclusions and evaluation of the method Overview of chapter Chapter 8: Conclusions and Discussion Overview of key findings and study limitations Review of previous findings Original research Applying the study findings to theories of working and retirement in later life Study limitations Policy implications and future research Policy implications of the research Future research Final summary and conclusion References Appendix 1: Meta-analysis formulae A. Calculation of study and pooled effects B. Statistics to Assess heterogeneity C. Transformation of effects to Cohen s d Appendix 2: Studies excluded from meta-analysis Appendix 3: Sensitivity analysis following meta-analysis: removal of the study by Dave et al (2006) A. Sensitivity analysis of later-life employment on depressive symptomatology B. Sensitivity analysis of later-life employment on self-rated health Appendix 4: Subgroup Meta-Analysis Appendix 5: Supporting sample descriptive analysis Appendix 6: Propensity score matching formulae A. Matching algorithms

7 6 6B Assessing covariate bias C. Sensitivity analysis using Rosenbaum bounds Appendix 7: Full results of covariate imbalance testing (all workers compared to retirees) Depression nearest neighbour matching Depression caliper matching Depression radius matching Depression kernel matching Self-rated health nearest neighbour matching Self-rated health caliper matching Self-rated health radius matching Self-rated health kernel matching Cognitive function nearest neighbour matching Cognitive function caliper matching Cognitive function radius matching Cognitive function kernel matching Appendix 8: Results of matching using additional algorithms (work quality and retirement) Appendix 9: Full results of covariate imbalance testing (analyses of work quality) Poor quality work, compared to good quality work, on depression Poor quality work, compared to retirement, on depression Good quality work, compared to retirement, on self-rated health Appendix 10: Results of logistic regression for the comparison of results of the sensitivity analysis using Rosenbaum bounds

8 7 List of tables Chapter 2 Table 2.1: Results of PubMed search for systematic literature review Table 2.2: Overview of the studies included in the systematic review Table 2.3: Findings from literature regarding depressive symptomatology Table 2.4: Findings from literature regarding self-rated health Chapter 3 Table 3.1: Studies included in meta-analysis and details of how inclusion criteria are met. Table 3.2: Meta-analysis of combined effects (d) later-life employment on depression final model. Table 3.3: Study estimates and calculated weight values Table 3.4: Meta-analysis of combined effects (d) later-life employment on suboptimal self-rated health. Table 3.5: Study estimates and calculated weight values Table 3.6: Key statistics from final meta-analysis and subgroup analyses later-life employment on depression Table 3.7: Key statistics from final meta-analysis and subgroup analyses later-life employment on suboptimal self-rated health. Chapter 4 Table 4.1: Wellbeing and employment activity at the wave prior to reaching retirement age Chapter 5 Table 5.1: Characteristics of the core ELSA sample by wave Table 5.2: Characteristics of the selected sample by wave

9 8 Table 5.3: Health at retirement age by employment activity at SPA Table 5.4: Employment activity by gender and wave Table 5.5: Employment activity by gender and partner s employment status at SPA Table 5.6: Health at SPA by gender and employment activity Table 5.7: Health at SPA by wealth and employment activity Table 5.8: Optimal wealth by employment activity at SPA Table 5.9: Property ownership by employment activity at SPA Table 5.10: Health at SPA by property ownership and employment activity Table 5.11: Employment activity at SPA by socio-economic classification at the previous wave Table 5.12: House ownership and wealth by employment status and NS- SEC category at SPA Table 5.13: Health at SPA by NS-SEC category and employment activity Table 5.14: Employment activity at SPA by qualification status Table 5.15: Health at SPA by qualification level and employment activity Table 5.16: Other activities by employment status at SPA Table 5.17: Health at SPA by other activities and employment status Table 5.18: ADL score by employment status at SPA Table 5.19: Health at SPA by ADL score and employment activity Table 5.20: Employment activity by self-employment status Table 5.21: Health at SPA by self-employment status Table 5.22: Employment activity at SPA by employment type Table 5.23: Health at SPA by employment type Table 5.24: Employment status by hours worked prior to SPA annd hours worked at SPA

10 9 Table 5.25: Mean health scores at SPA by hours worked Table 5.26: Hours worked prior to SPA by hours worked at SPA Table 5.27: Working for the same employer at the wave prior to reaching SPA Table 5.28: Mean health scores at SPA by employer at the previous wave Table 5.29: Employment activity at SPA by work quality Table 5.30: Health at SPA by work quality/retirement Table 5.31: Work quality/retirement by gender Table 5.32: Health at SPA by gender and work quality/retirement Table 5.33: Socio-demographic characteristics by work quality/retirement at SPA Table 5.34: Health at SPA by socio-demographic characteristics and work quality/retirement Table 5.35: Other activities by work quality/retirement at SPA Table 5.36: ADL score by work quality/retirement at SPA Table 5.37: Mean wealth at SPA by ADL score and work quality/retirement Chapter 6 Table 6.1: Sample size at each centred wave, according to employment status and type. Table 6.2: Results of piecewise regression on depression (CES-D score) Table 6.3: Results of piecewise regression on self-rated health Table 6.4: Results of piecewise regression on cognitive function Chapter 7 Table 7.1 Results of the probit model to estimate the propensity score for continuing later-life employment.

11 10 Table 7.2: Summary statistics of the propensity scores (all workers) Table 7.3: Results of propensity score matching on all work versus retirement Table 7.4: Covariate imbalance before and after radius matching (all workers) Table 7.5 Results of the probit model to estimate the propensity scores for analyses of work quality and retirement Table 7.6: Summary statistics of the propensity scores (work stratified by quality) Table 7.7: Results of propensity score matching on work quality and retirement Table 7.8: Covariate imbalance before and after radius matching (work stratified by quality) Table 7.9: Results of Rosenbaum-bounds sensitivity analysis Appendix 2 Table 2A: List of studies excluded from the meta-analysis and justification for exclusion Appendix 3 Table 3A.1: Meta-analysis of combined effects (d) later-life employment on depression sensitivity analysis without Dave et al. (2006) Table 3A.2: Study estimates and calcualted weight values without Dave et al. (2006) - depression Table 3A.3: Meta-analysis of combined effects (d) later-life employment on suboptimal self-rated health sensitivity analysis without Dave et al. (2006). Table 3A.4: Study estimates and calcualted weight values without Dave et al. (2006) self-rated health Appendix 4

12 11 Table 4A.1: Binary responses for study-level dummy variables analysis of depression Table 4A.2: Binary responses for study-level dummy variables analysis of self-rated health Appendix 5 Table 5A.1: Wealth and employment information by private pension status (S.E.) Appendix 7 Table 7A.1: Covariate imbalance testing for the analysis of depression using nearest neighbour matching (full results) Table 7A.2: Covariate imbalance testing for the analysis of depression using caliper matching (full results) Table 7A.3: Covariate imbalance testing for the analysis of depression using radius matching (full results) Table 7A.4: Covariate imbalance testing for the analysis of depression using kernel matching (full results) Table 7A.5: Covariate imbalance testing for the analysis of self-rated health using nearest neighbour matching (full results) Table 7A.6: Covariate imbalance testing for the analysis of self-rated health using caliper matching (full results) Table 7A.7: Covariate imbalance testing for the analysis of self-rated health using radius matching (full results) Table 7A.8: Covariate imbalance testing for the analysis of self-rated health using kernel matching (full results) Table 7A.9: Covariate imbalance testing for the analysis of cognitive function using nearest neighbour matching (full results) Table 7A.10: Covariate imbalance testing for the analysis of cognitive function using caliper matching (full results) Table 7A.11: Covariate imbalance testing for the analysis of cognitive function using radius matching (full results)

13 12 Table 7A.12: Covariate imbalance testing for the analysis of cognitive function using kernel matching (full results) Appendix 8 Table 8A: Results of propensity score matching using nearest neighbour, caliper and kernel matching algorithms unmatched results and ATT (tstatistics). Appendix 9 Table 9A.1: Covariate imbalance testing for the analysis of poor quality work, compared to good quality work, on depression (full results) Table 9A.2: Covariate imbalance testing for the analysis of poor quality work, compared to retirement, on depression (full results) Table 9A.3: Covariate imbalance testing for the analysis of good quality work, compared to retirement, on self-rated health (full results) Appendix 10 Table 10A: Results of the logistic regression model to determine effect sizes comparable to results of the sensitivity analysis using Rosenbaum bounds

14 13 The University of Manchester Katherine Matthews Doctor of Philosophy Is working beyond state pension age beneficial for health? Evidence from the English Longitudinal Study of Ageing 2014 Thesis abstract Objectives: Extending working lives is a major strategy in policy responses to ageing populations. This is currently being implemented by means of the increasing UK state pension age. However, the health effects of such changes are highly debatable. A systematic review conducted by this thesis revealed that previous research on the topic has provided a diverse set of findings. One of the reasons for the lack of agreement between previous studies is the high degree of heterogeneity in the study samples of older adults. This is statistically revealed by a meta-analysis conducted in this study. The research presented within this thesis examines whether extending working lives is beneficial for health, and focuses on the importance of accounting for quality of work when considering these effects. Methods: The study used respondents from waves 1 to 5 of the English Longitudinal Study of Ageing who worked until state pension age and then entered either later-life employment or retirement. Linear spline regressions examined trajectories of depression, self-rated health and cognitive function across the retirement age period, stratified by work quality and retirement. Propensity score matching was subsequently used to estimate unbiased treatment effects of extended working as opposed to retirement, and then of poor and good quality work individually in relation to retirement. Results: The spline models indicated entering retirement from work was associated with a significant change in patterns of depression and self-rated health, but continuation of work was not. Retiree trajectories consistently showed poorer outcomes than those of respondents who were working. The results of the propensity score matching found no significant differences in health on the basis of belonging to the group of overall workers compared to retirees. However when work was stratified on the basis of its quality, significant differences became apparent. Belonging to the group of poor quality workers was associated with significantly worse depression than belonging to both the good quality workers and retirees, and belonging to the group of good quality worker was associated with significantly better self-rated health than belonging to the group of retirees. Discussion: The heterogeneous sociodemographic and health characteristics of the older working population should be taken into account when examining impacts of employment on health. Failure to account for differences in quality of work may lead to the incorrect assumption that extended employment is beneficial to the health of all workers. If older people are going to be encouraged to work for longer periods of time, beneficial effects need to apply to all working groups. Employers need to ensure adjustments to individual working patterns and environments are made in order to suit the needs of an ageing workforce.

15 14 Declaration and Copyright Statement Statement of Authorship Katherine Matthews is the sole author of the work presented within this thesis. Contributions to the thesis in the form of theoretical and methodological guidance were provided by Professors Tarani Chandola and James Nazroo and Doctor Neil Pendleton. Declaration No portion of the work referred to in the thesis has been submitted in support of an application for another degree or qualification of this or any other university or other institute of learning. Copyright Statement i. The author of this thesis (including any appendices and/or schedules to this thesis) owns certain copyright or related rights in it (the Copyright ) and s/he has given The University of Manchester certain rights to use such Copyright, including for administrative purposes. ii. Copies of this thesis, either in full or in extracts and whether in hard or electronic copy, may be made only in accordance with the Copyright, Designs and Patents Act 1988 (as amended) and regulations issued under it or, where appropriate, in accordance with licensing agreements which the University has from time to time. This page must form part of any such copies made. iii. The ownership of certain Copyright, patents, designs, trademarks and other intellectual property (the Intellectual Property ) and any reproductions of copyright works in the thesis, for example graphs and tables ( Reproductions ), which may be described in this thesis, may not be owned by the author and may be owned by third parties. Such Intellectual Property and Reproductions cannot and must not be made available for use without the prior written permission of the owner(s) of the relevant Intellectual Property and/or Reproductions. iv. Further information on the conditions under which disclosure, publication and commercialisation of this thesis, the Copyright and any Intellectual Property and/or Reproductions described in it may take place is available in the University IP Policy (see in any relevant Thesis restriction declarations deposited in the University Library, The University Library s regulations (see and in The University s policy on Presentation of Theses.

16 15 Chapter 1: Introduction This first chapter of the thesis presents an introduction to the topic of interest to the research: the health effects of later-life employment, and the role of work quality within any relationships which might exist. Firstly, a rationale for the research is presented, with reference to the importance of understanding the subject. Secondly, the specific research questions to be investigated are outlined, and their relevant hypotheses presented. Thirdly, a brief overview of the thesis is provided.

17 Rationale for the research: ageing and state pension age in the UK Extending working lives is a major economic strategy in UK policy responses to the ageing population. A 2012 report by the ONS states the number of people of normal pension age and above who remain in work almost doubled to a total of 1.4 million between the years 1993 and The report subsequently states this increase is not simply due to rising numbers of older people, but that the percentage of older people 1 remaining in the workforce has also increased within this time period, from 7.6 per cent in 1993 to 12.0 per cent in State pension age is currently increasing for women, and by 2018 will increase for both men and women simultaneously. It is essential to the successful implementation of these policies that older people remain in the workforce for as long as possible. As a result, there is an increased importance in understanding the health consequences of continued workforce participation. Furthermore, an understanding of who is working later in life, as well as the reasons for doing so, is essential in ensuring these changes to policy reach their potential. One of the key factors influencing changes to state pension age (SPA) is the increasing life expectancy of people in industrialised countries, including the UK. If retirement ages do not change to reflect this, there will be increasing financial and welfare burdens on the economy as retirement funding is expected to last for longer periods of time. It is currently not clear whether these changes will have beneficial or detrimental effects on the health of the elderly population. The research presented within this thesis provides a means of assessing how extended working affects health outcomes, and discusses the implications for policies concerning employment in later-life. Despite recent changes to policy, a significant proportion of older people retire before reaching SPA (Weyman et al. 2012). Research has also demonstrated the preference of most older adults is to work for as long as is possible if adjustments can be made to the working environment to suit the individual needs of older employees (Vickerstaff et al. 2008). There are likely to be a variety of factors influencing an individual s decision to exit the workforce as retirement nears, and an 1 The term older people in this instance refers to people of state pension age and above.

18 17 understanding of these is essential if older populations are to be successfully encouraged to spend longer periods of time in employment. While evidence demonstrates early retirement to be most strongly driven by poor health, particularly among those retiring in their early fifties (Humphrey et al. 2003), a report by Phillipson & Smith (2005) discusses a range of push and pull factors which bear an influence on whether or not an individual decides to continue work. Manual work, high workplace stress and low reciprocity are all negative reasons as to why people leave prior to reaching SPA, but early exit is also influenced by positive factors such as the affordability to leave work while young enough to pursue favoured activities. Positive factors leading to an increased likelihood of continued employment include higher grade occupational statuses, opportunities for career training and development, the option for flexible working and participation in self-employment, while the main negative influences centre on financial insecurity and unaffordability of retirement. The effects of continued working are likely to be strongly impacted by the factors influencing it. As will be discussed throughout the work, continuation of employment among individuals who work in optimal conditions might be preferential and associated with beneficial health and wellbeing outcomes. This relates to the cost-effective idea of a compression of morbidity (Fries 2002), by which good health is maintained through the prolonged continuation of enjoyable work roles, and periods of morbidity thereafter are delayed and subsequently shorter. However, it is also important to remember that many individuals may be forced to continue work when their health is in decline, through lack of a financial viability of retirement. It is these individuals who are particularly important to consider when implementing policies to increase retirement age, as prolonged working here may be detrimental to aspects of both physical and mental health and wellbeing. Recognising older workers as a heterogeneous group is essential if changes to policy are to be cost-effective to the economy and beneficial to society.

19 Research questions The work presented within this thesis centres around two key research questions, the first of which is broad, and the second of which is broken down into two separate parts. Each chapter of the thesis places a focus on each of these research questions in turn, with the descriptive and main analyses providing answers to both. i. What are the health effects of working beyond SPA? ii. What impact does work quality have on the health of those who continue to work beyond SPA? a. How does poor quality employment affect health among those who continue to work past SPA compared to good quality employment? b. How does poor and good quality employment affect health compared to those who retire at SPA? The health effects of working beyond SPA Using longitudinal data, the first part of the research focuses on the broad question of how later-life employment impacts on health outcomes. Previously conducted studies using observational data are incomparable due to their heterogeneous populations and definitions of key concepts. This is discussed in greater detail in Chapters 2 and 3, which provide a systematic review and meta-analysis of recent literature on the subject. The thesis will then examine trajectories of health across the retirement age period in order to ascertain whether these differ on the basis of economic activity in later life. Finally, propensity score matching will be used to account for the differences in the characteristics of individuals who work beyond retirement age and those who do not and establish whether or not significant effects exist on the basis of this. The study hypothesises that large selection biases will be present when analysing workers as a homogeneous group, and stratification of work by some means of its type or quality will be necessary to determine the extent to which it might be beneficial to older populations The impact of work quality on health in later-life It cannot be assumed that working has the same effect on all individuals, and a key source of variation is likely to arise from the quality of the work an individual carries out. The growing number of people working later in life, and the projected

20 19 continuation of this increase, means it is important to look beyond the effects of employment on average, and consider how employment in older age can be managed in order to provide the greatest protective effect on health. The second research question examines the role of work quality in terms of an effort-reward imbalance scale, and investigates differences in health outcomes on the basis of whether individuals work in poor or good quality employment. The two work quality groups will then be compared to those who retire, in order to examine whether or not continued employment leads to better health among older populations, or whether retirement is the most beneficial option.

21 Overview of the thesis The work presented throughout the following thesis provides an extensive investigation into the effects of working beyond normal retirement age in consideration of the issues of selection bias which might confound results. Chapter 2 examines literature already published on the topic of later-life employment and different aspects of health and wellbeing, from which the overall conclusion is made that differences between study methods are too great to establish whether any overall conclusion can be drawn. Chapter 3 furthers the investigation of previous literature by using meta-analysis to establish pooled effects and to quantify the levels of between-study heterogeneity present within the research area. As might be expected, these levels of heterogeneity are too high to fairly compare the studies. A subsequent subgroup analysis, segregating studies by specific characteristics, such as those which provide effects for men and women separately, and those which include only individuals who retire post SPA, demonstrates that consideration of specific individual characteristics, such as retirement age and gender, can reduce betweenstudy heterogeneity. The following sections of the thesis focus on the analysis carried out to answer the research questions outlined in section 1.2. Chapter 4 provides a description of the ELSA data and the importance of the strict sample selection process used in light of the heterogeneous findings presented in chapters 2 and 3. A description of the forthcoming methodology to be used is also detailed here, alongside the rationale for using spline models to examine trajectories of health and then propensity score matching to deal with issues of selection bias. Chapters 5 and 6 provide some preliminary and descriptive sample analyses using the ELSA data. Chapter 5 demonstrates initial differences in health and sociodemographic characteristics which appear to exist on the basis of whether an individual continues to work or retires once SPA is reached. The results appear particularly interesting when employment is stratified on the basis of its quality. Chapter 6 uses piecewise linear spline regression to examine trajectories of wellbeing according to employment activity at normal retirement age. An examination of the impact on health outcomes of the variables used throughout the study is made, and differences in intercepts and slope gradients before and after

22 21 retirement age are discussed. Entering retirement is associated with a significant worsening of depression and self-rated health, yet the results for all groups continuing work remain non-significant. Chapter 7 presents the results from the propensity score matching approach used to account for selection bias. Calculation of the propensity score and its potential effectiveness is detailed, before the main results are presented. Here, once observable sources of bias are accounted for, no significant differences in health are present on the basis of whether an individual continues to work or retires in the first instance, but significant differences do exist on the basis of work quality. Poor work quality is associated with significantly higher levels of depression than both continuation of good quality work and retirement, and continuation of good quality work is associated with significantly better self-rated health than retirement. A sensitivity analysis, to examine how great an unobservable source of bias would need to be in order to significantly affect the findings, shows an acceptable level of robustness of results. A discussion of the thesis key results, alongside their limitations and implications, is provided in Chapter 8.

23 22 Chapter 2: Systematic Review of the Health Effects of Later-Life Employment This chapter provides an overview of theoretical perspectives concerning ageing, employment and health, and a systematic review of empirical evidence on the topic of the health effects of later-life employment. The process by which studies were selected for inclusion in the review is detailed, alongside tables of study descriptions and their relevance to the research. Several health outcomes are discussed, and a key focus is placed on the commonly examined outcomes which are of interest to the later sections of this study: depression, self-rated health and cognitive function. Findings will be applied to popular theoretical perspectives concerning ageing, employment and health. The chapter will close with a brief summary of key findings and conclusions.

24 Ageing, employment and health: theoretical perspectives The relationship between later-life working and health is likely to be dependent on a complex set of theoretical mechanisms, strongly influenced by an individual s personal and workplace characteristics. Some of these characteristics, which are likely to influence decisions on later-life working, were outlined in the previous chapter, and will be discussed in greater detail later in the work. Before examining empirical evidence concerning the effects of later-life employment, some of these mechanisms will be discussed in this section of the chapter. Two key perspectives concerning later-life employment and health, which are often intertwined with one another, are those concerning role continuity and activity theory. Continuity theory (Atchley 1989) exploits the idea that a continuation of socially meaningful roles across the life course and into older age, such as those provided by paid employment, leads to better health outcomes. Activity theory, similarly, maintains that remaining in productive and worthwhile social roles is an essential aspect of the healthy ageing process (George 1993). In both instances, a failure to remain engaged in socially worthwhile activities is detrimental to health and wellbeing. The ability to maintain better health through continued work roles ties in with ideas surrounding compression of morbidity, through which the delayed onset of poorer health which might accompany workforce exit leads to longer periods of healthy life and shorter periods of morbidity thereafter (Fries 2002). This idea is central in the implementation of cost-effective policies to extend working lives in the face of ageing populations. While continuity and activity theories may suggest the effect of continued employment in later-life to be beneficial, it is important to recognise this to be dependent on individual circumstances. Remaining in the workforce beyond pensionable age may, indeed, be beneficial to those who work in rewarding roles, with low levels of stress and good levels of reciprocity, but this is unlikely to be the case for those who are employed in poor quality or stressful occupations (Siegrist et al. 2008; Godin et al. 2005; Niedhammer et al. 2004, Pikhart et al. 2004; Stansfeld et al. 1999; Niedhammer et al. 1998). Similarly, continued employment may be beneficial to individuals who are physically and mentally capable of remaining in the workforce for longer periods of time, but the same may not be true for those with

25 24 health limitations who are forced to work through unaffordability of retirement (Ekerdt 1983). Extended periods of manual work might serve a protective effect on health decline when it is carried out by healthy individuals (Lemon et al. 1972), but conversely may worsen health decline when carried out by individuals already suffering poorer health (Ekerdt 1983). In instances of job-lock such as these, whereby the individual feels unable to exit the workforce despite it being the preferred option, adverse health effects are more likely to be observed (Benjamin et al. 2008), and retirement would be the healthier option. Ideas surrounding both continuity and activity theory can be extended into the retirement process and are not solely based on extended workforce participation. Roles similar to those carried out across the life-span through paid employment can often be replicated by means of voluntary work or engagement in socially rewarding leisure activities. In turn, this again can lead to a continued sense of social identity and high life satisfaction (Drentea 2002). The decision to retire is a complex process and the transition itself is a heterogeneous path, again often dependent on individual and workplace circumstances (Van Solinge & Henkens 2007; Phillipson & Smith 2005). While some may leave the workforce completely and focus on either socially productive activities such as volunteering, or leisure activities which may promote good wellbeing, others enter a period of bridge employment, the term given to more flexible work, such as that which is part-time or includes less demanding roles. This occurs after exit from one s main career but prior to leaving the workforce completely (Feldman 1994). While, from the perspectives of continuity and activity theories, periods of bridge employment may be beneficial to health in that it maintains engagement in socially productive roles (Bonsdorff et al. 2009; Zhan et al. 2009; Phillipson 2004) in more flexible working environments, participation in bridge employment itself is strongly related to favourable individual and workplace characteristics (Wang et al. 2009; Kim & Feldman 2000), and again, its beneficial nature is likely to be dependent on both the reasons for participating in it as well as worker ability. Another popular perspective on the mechanisms through which continued working affects health is that concerning disengagement theory (Hochschild 1975; Cumming & Henry 1961). Here, workforce exit is seen as a natural part of the decline in

26 25 activity and social life that accompanies ageing. The discontinuation of socially productive roles is replaced with the opportunity to take part in rewarding leisure activities, many of which may encourage better physical and cognitive functioning, as well as mental wellbeing (Bound & Waidmann 2007; Drentea 2002; Coyle 1994). The idea also ties in with the cultural-institutional approach to retirement, which maintains that workforce exit at the time seen as normal and expected by society leads to better health outcomes than those who retire either earlier or later than state pension age (Börsch-Supan & Jürges 2009). The theory also relates to ideas of role strain and role strain reduction theory, which state that continuation of physically or mentally stressful roles, such as those which might be associated with certain types of paid employment, can be detrimental to a range of health outcomes (Kim & Moen 2002; Ekerdt 1983), and retirement is a means of exiting such circumstances. However, it should be noted again than mechanisms through which the stress of employment is alleviated by retirement are heterogeneous and dependent on individual circumstances. For instance, retirement is only likely to be beneficial to health when it can be afforded and enjoyed, and activities can be carried out to replace those lost by leaving the workforce (Scherger et al. 2011; Drentea 2002; Herzog & House 1991). In cases where retirement is taken and unfavourable roles, such as providing care for ill family members, are assumed, health may be negatively affected (Evandrou & Glaser 2004; Dentinger & Clarkberg 2002). Additionally, enjoyment of retirement is likely to be dependent on whether the decision was voluntary (Van Solinge & Henkens 2007; Szinovacz & Davey 2005; Peretti & Wilson 1975). Studies have shown involuntary workforce exit at retirement age to be associated with poorer chances of replacing lost meaningful roles and sufficient levels of income, which in turn leads to poorer physical and mental health outcomes (Bender 2012; Calvo et al. 2009; Gallo et al. 2000). This effect may be particularly detrimental following involuntary retirement due to health reasons, which may be especially likely to negatively impact activity options thereafter (Hershey & Henkens 2007). The importance of work roles in older age should be considered in terms of a lifecourse perspective (Kim & Moen 2002; Moen 1996). Again, this may relate to the aforementioned importance of adhering to the cultural norms of retirement expected throughout the life-course. However, it should be understood that the role of work is

27 26 likely to differ for men and women across the lifespan, and therefore effects of its continuation or exit are likely to be gendered (Kim & Moen 2002; Quick & Moen 1998; Moen 1996). Men are historically likely to have a stronger attachment to their role as a worker than women, as they participate in the workforce for longer and more continuous periods of time (Davey & Szinovacz 2004; Kim & Moen 2002; Myers & Booth 1996). As a result, they may find exit from the worker role particularly stressful, especially if they are entering retirement before their spouse (Szinovacz & Davey 2004) or involuntarily (Van Solinge & Henkens 2007). Women are historically more likely to have shorter periods of work across the life-course, and therefore transitions into retirement may be easier (Kim & Moen 2002; Han & Moen 1999). Similarly, women are more likely to participate in poorer quality and lower grade employment, which again lead to an easier transition out of the workforce (Elder & Schmidt 2004; Bosma et al. 1998). These gendered relationships will be considered in greater detail further in the chapter. The life-course perspective also offers strength to a human capital approach of laterlife working, which states that the cumulative advantages of lifelong good quality working, preceded by educational attainment, may lead to beneficial financial effects of postponing retirement (Dannefer 2003; Kosloski et al. 2001). In turn, this participation in good quality and enjoyable work, and receipt of financial reward, may lead to better health outcomes (Pikhart et al. 2001). Again, however, this idea is dependent on the type of work carried out being sufficiently financially rewarding, and on the worker s ability to remain within it. The same approach must also highlight the cumulative disadvantages brought about by lifelong participation in poor quality and unrewarding work. This is likely to lead to social disadvantage, which is related to poor outcomes of health and wellbeing across the life-course (Chandola et al. 2007; Blane et al. 1999; Marmot et al. 1991), again worsened when the individual feels a sense of job-lock and inability to exit work at retirement age (Ekerdt 1983). The research presented herein places a focus on the importance of accounting for work quality when examining impacts of employment on the health of older workers. An effort-reward imbalance model (Siegrist 1996) is used to measure work quality, and works on the underlying theoretical basis that health is adversely affected when contracted job roles are not reciprocated by means of sufficient salary, esteem and

28 27 security. The model may be specifically relevant to the study of older workers in that continuation of poor quality roles might be particularly prevalent at parts of the life course in which few occupational alternatives are offered (Seigrist et al. 2004). Herein lies another example of job-lock, whereby older workers are not only forced into continued work through financial incentives, but also through lack of potential mobility. Emotional wellbeing is then impacted by these negative job characteristics, as successful self-regulation is, in part, controlled by receipt of adequate reward for social roles (Siegrist et al. 2002). Continued employment in suboptimal effortreward conditions leads to a greater strain on health (Siegrist et al. 2008; Godin et al. 2005; Niedhammer et al. 2004, Pikhart et al. 2004; Stansfeld et al. 1999; Niedhammer et al. 1998). Further details of the effort-reward imbalance scale are provided in section The work presented within this thesis follows the theoretical perspective that continuation of productive roles and activities, such as paid employment, are beneficial to health and wellbeing as long as they are well reciprocated and enjoyable. However, a focus will also be placed on the importance of considering the negative impacts of job lock and role strain when regarding those individuals who are working in poorer health and in poorly reciprocated conditions. The complex mechanisms through which later-life working impacts health will be referred to throughout the work in relation to the findings produced.

29 Method and results of the initial search for literature The first stage of the literature review was a search of the PubMed database for any publication which may have been relevant to the research question concerning health effects of later-life working. Table 2.1 presents the initial results of the search. The original number of studies identified, alongside the number deemed of use to the work, is presented by each search term entered in Table 2.1 (columns 1 and 2, respectively). The 4,096 studies initially identified by the search were reduced in number on the basis of the relevance of their titles, and all search results were filtered to show only studies written in English. Nine other databases were searched for relevant literature (Embase, PsycInfo, Web of Science, Medline, ComDisDome, Intute, Social Sciences Citation Index, ASSIA and Sociological Abstracts). These databases were selected from the University of Manchester John Ryland s Library website, and placed a focus on medicine, psychology and sociology subject areas. However, any studies obtained from searches of these databases were either irrelevant to the review or duplicates of studies identified through PubMed. This resulted in a total of 150 studies which were selected on the basis of being potentially useful to the review. Table 2.1: Results of PubMed search for systematic literature review Search term Studies identified (all studies) Studies selected for review on Older workers health Later life employment Work quality later life 53 1 Retirement health the basis of relevant titles In the second stage of the literature search, abstracts were read in order to further select the studies which were relevant to the research questions of interest. It was essential here to ensure a set of inclusion criteria were adhered to. The inclusion criteria were as follows: i. Study populations had to be inclusive of older workers, defined in this instance as those aged 60 and over.

30 29 ii. iii. iv. An exposure group of older workers had to be examined in relation to a control group of retirees. Studies were also included if they focused on an exposure group of retirees in strong relation to a control group of later-life workers. Studies were excluded if non-workers consisted solely of individuals who were not retired, such as those out of work due to disability or unemployment. Study outcomes had to be relevant to the outcomes of interest to the research questions of the thesis 2. Although the range of study outcomes examined was extensive, inclusion was based on measures which covered general mental health, subjective health and cognitive function. Studies were excluded if they strongly focused only on specific physical disorders or diseases. Studies had to be based on the statistical analysis of large-scale datasets. Studies were excluded if they had a very low sample size (n<100), were qualitative in nature or were based on case studies. Exclusions were also made where studies did not include some form of statistical analysis to examine the impact of working on health in later life. Reading the abstracts of the studies initially deemed relevant for the review lead to exclusion of a further 101 studies. This resulted in a final number of 48 studies satisfactorily meeting the inclusion criteria for the systematic review. The third step in the literature review was a search of the bibliographies of the selected studies. Any studies with titles which indicated a relevance to the review were subsequently obtained. Again, a reading of abstracts allowed elimination of any research which did not adhere to the list of inclusion criteria. This lead to the inclusion of a further eight studies, leaving the final number for further analysis at The outcomes of interest to this study are depressive symptoms, self-rated health and cognitive function. Detailed descriptions of these can be found in Chapter 4, alongside justification of their use in this research.

31 An overview of the studies included in the review Table 2.2 presents an overview of the 58 studies selected for the final systematic review. Here, details are provided of the study authors, year of study publication, the data used for the research, basic sample descriptives (age and gender distribution), the outcomes examined and covariates included in the fully adjusted statistical analyses. A discussion of the key studies features is presented thereafter.

32 31 Table 2.2: Overview of the studies included in the systematic review Author (year) Study type Data used Sample size Age Gender Outcome measure(s) Covariates Adam et al. (2007) Longitudinal Survey on Health, Ageing and Retirement in Europe, Health and SHARE 25,916 obs. HRS SHARE 64.4 HRS 66.9 SHARE 54.3% female HRS Episodic memory Age, education, country, gender, ethnicity, morbidity and illness, physical activity. Retirement Study, English Longitudinal Study of Ageing 17,097 obs ELSA 8,431 obs ELSA 66.1 (means) 59.2% female ELSA 55.2% female Adelmann et al. Cross-sectional American Changing Lives 1, (mean) 67.2% female Life satisfaction, depression and selfefficacy Age, education, health problems, ethnicity, gender. Behnke (2012) Longitudinal English Longitudinal Study of Ageing 1,439 Mean employed age Mean retired age Employed 50% female Retired 54% female Chronic conditions, self-rated health, depression, underlying health stock. Gender, age, marital status, people in household, children, grandchildren, country of birth, education, socio-economic classification, job characteristics, region characteristics, pension characteristics, health characteristics and habits, job and health expectations.

33 32 Table 2.2: Overview of the studies included in the systematic review Author (year) Study type Data used Sample size Age Gender Outcome measure(s) Covariates Bonsang et al. (2012) Longitudinal Health and Retirement Study 14, Male and female Cognitive function IV analysis using eligibility age for social security as instrument. Bosse et al. Cross-sectional Normative Aging 1, and older Male only Psychological Age, occupational grade (1987) Study (USA) symptoms Broerson et al. (2006) Cross-sectional Periodical Occupational Health Surveys (Netherlands) 14,138 All working ages 27.4% female Self-reported specific health measures. Age category, gender, occupational grade, social group. Butterworth et al. (2006) Cross-sectional National Survey of Mental Health and Well-Being (Australia) 4, % female Depression Age, gender, living arrangement, marital status, housing tenure, income source, socio-economic status. Buxton et al. (2005) Cross-sectional 2000 Psychiatric Morbidity Survey (UK) 1, % female Depression Current working status, age of leaving education, occupation, social class, household tenure, retirement type. Calvo et al. (2013) Longitudinal Health and Retirement Study 6, (mean across waves) 47.3% female Self-reported health, depression IV analysis using early retirement windows as instruments. Controls for

34 33 Table 2.2: Overview of the studies included in the systematic review Author (year) Study type Data used Sample size Age Gender Outcome measure(s) Covariates wealth, income, spousal employment status, gender, ethnicity, education, employment type. Chandola et al. (2007) Longitudinal Whitehall II Study 8,637 at baseline 56 (mean across waves) Male and female SF-36 Scores (physical and mental health components) Occupational class, employment grade, employment status, age, gender. Charles (2002) Longitudinal Health and Retirement Study, National Longitudinal Survey of NA HRS 62.6 NLS-MM 66 Male only Life satisfaction Ethnicity, age, education, marital status, health status, depression, loneliness. Mature Men (means) Choi et al. (2007) Longitudinal Health and Retirement Study 8, (mean) 60.7% female Depression Age, education, ethnicity, medical conditions, ADL impairments, marital status, caregiving status, volunteering status, physical activity, smoking status, drinking activity. Christ et al. (2007) Longitudinal National Health Interview Survey 23, and older 56.9% female K6 Nonspecific Psychological Distress Gender, age, ethnicity, educational level,

35 34 Table 2.2: Overview of the studies included in the systematic review Author (year) Study type Data used Sample size Age Gender Outcome measure(s) Covariates (USA) Scale occupational type. Coe and Lindeboom (2008) Longitudinal Health and Retirement Study 3, (mean) Male only Depression, ADL difficulties index Education, marital status, number of children, wealth, age, occupational type, race, ethnicity. Coe and Zamarro (2011) Longitudinal Survey of Health, Ageing and Retirement in Europe 5, (mean) Male only Self-rated health, depression IV analysis using countryspecific retirement ages as instruments. Controls for education, marital status, health and ADL limitations, weight, type of employment. Coursolle et al. (2010) Longitudinal Wisconsin Longitudinal Study 5, Male and female Depression, psychological wellbeing (functioning) Family circumstances, working hours and conditions, gender, income, educational attainment. Crowley et al. Longitudinal National Longitudinal 1, at Male only Affect balance, Covariates not specified. (1986) Surveys of Labour Market Experience (USA) follow-up happiness index, subjective health Dave et al. (2006) Longitudinal Health and Retirement Study 77,194 obs Male and female Depression, ADL Difficulties Index, Self- Health insurance status, parental status, ethnicity,

36 35 Table 2.2: Overview of the studies included in the systematic review Author (year) Study type Data used Sample size Age Gender Outcome measure(s) Covariates reported health religion, income, marital status. Drentea et al. (2002) Cross-sectional Aging, Status, and Sense of Control; National Survey of Families and Households ASOC 2,592 NSFH 12,897 ASOC NSFH (means) ASOC 57% female NSFH 53% female Distress, depression Work and activity characteristics, age, gender, ethnicity, marital status, education, family income, work history. (USA) Ekerdt and Bosse Longitudinal Normative Ageing Male only Self-reported health Age, employment status. (1982) Study (USA) Fernandez et al. (1998) Longitudinal North Carolina Survey Male and female Depression Income, education, physical health, retirement status, social networks. Gall et al. (1997) Longitudinal Retirement Research Study (Ontario) (mean at follow-up) Male only Self-reported health, life satisfaction. Age, retirement income, marital status, ethnicity, retirement type, occupation type. Gill et al. (2006) Household, Income and Labour Dynamics (Australia) 1, Male only SF-36 questionnaire (mental functioning) Age, financial hardship, income, home ownership, socio-economic status, physical health, partner s

37 36 Table 2.2: Overview of the studies included in the systematic review Author (year) Study type Data used Sample size Age Gender Outcome measure(s) Covariates health. Hammerman- Rozenburg et al. (2005) Longitudinal Jerusalem Longitudinal Cohort Study and 77 Male and female Physical examinations Socio-economic status, lifestyle characteristics, illnesses, medical consumption. Hao (2008) Longitudinal Health and Retirement Study 27,341 obs Male and female Depression Work status, volunteering, marital status, general health, age, education, income. Herzog et al. (1991) Cross-sectional Americans Changing Lives Survey 1, and older Male and female Depression Age, gender, occupational type, occupational grade. Jokela et al. (2009) Longitudinal Whitehall II Study 7, % female 5 standard tests of cognitive performance, Socio-economic status, retirement type. SF-36 Questionnaire (physical and mental health) Jokela et al. (2010) Longitudinal Whitehall II Study 7, % female SF-36 Questionnaire (physical and mental health) Employment status, reason for retirement, socioeconomic status (salary, job grade), age, gender. Jorm et al. (1998) Longitudinal Respondents sampled from Canberra and (mean at baseline) Male only Cognitive function, dementia Age, education, ethnicity, occupational type

38 37 Table 2.2: Overview of the studies included in the systematic review Author (year) Study type Data used Sample size Age Gender Outcome measure(s) Covariates Queanbeyan electoral rolls Keyes and Reitzes (2007) Longitudinal North Carolina Survey Male and female Self esteem (Rosenberg 10-item scale), depression Age, gender, race, retirement status, marital status, educational level, general health. Kim and Moen (2002) Longitudinal Cornell Retirement and Well-Being Study (USA) 458 Male 60.9 female % female Morale, depression Age, income, subjective health, marital quality, personal control. (mean at baseline) Kumari et al. (2010) Cross-sectional Whitehall II Study 2, and older at baseline Male and female Cortisol level Age, gender, smoking status, waking up time, sleep duration, stress level, walking/gait speed. Kumari et al. (2010) Cross-sectional Whitehall II Study 3,992 Male 60 female 61.5 (means) Male and female Cortisol level Educational attainment, employment grade, income, subjective social position, age, gender, smoking status, sleep duration, CES-D scores, various biological

39 38 Table 2.2: Overview of the studies included in the systematic review Author (year) Study type Data used Sample size Age Gender Outcome measure(s) Covariates variables. Laaksonen et al. (2012) Longitudinal City of Helsinki employees 7, (mean normal Male and female Psychotropic drug purchases Gender, occupational social class, retirement type. retirement age) 53.5 (mean disability retirement age) Latif (2012) Longitudinal Canadian National Population Health Survey 12,403 Mean retired age = 66 Mean nonretired age = 54% female (retired sample) 51% female (nonretired Self-reported health Gender, age, marital status, education, income, tenure, geographical area type, place of residence. 65 sample) Lindeboom et al. (2002) Longitudinal CERRA Panel Study (Netherlands) 3, and older Male only Depression, cognitive function Age, partnership status, occupational class, educational level, length of working life. Luoh and Herzog (2002) Longitudinal Asset and Health Dynamics among the 4, (mean) 63% female Self-reported health, physical functioning Age, gender. Education, ethnicity, marital status,

40 39 Table 2.2: Overview of the studies included in the systematic review Author (year) Study type Data used Sample size Age Gender Outcome measure(s) Covariates Oldest Old (USA) limitations household income, major chronic conditions, smoking status. Mandal and Roe (2008) Longitudinal Health and Retirement Study 7, at baseline 51.3% female Depression Type of work exit, marital status, health and life insurance, ADL limitations, poor physical health, age, education, gender, ethnicity, employment type. Marmot et al. (1996) Cross sectional & longitudinal Whitehall II Study 18, at baseline Male only Mortality Age, length of follow up, car ownership, employment grade McMunn et al. (2009) Cross-sectional English Longitudinal Study of Ageing (ELSA) 5,384 Male 65 and older Female 60 and older 62.4% female Wellbeing (CASP19), Satisfaction With Life Scale, depression Socially productive activities, gender, age, wealth, health, partnership status. Mein et al. (2003) Longitudinal Whitehall II Study 1, at baseline 36.2% female SF-36 Questionnaire (mental and physical functioning) Employment grade, socioeconomic status, marital status, job control, gender, age. Mojon-Azzi et al. Longitudinal Swiss Household Panel % female Self-reported health Gender, baseline health,

41 40 Table 2.2: Overview of the studies included in the systematic review Author (year) Study type Data used Sample size Age Gender Outcome measure(s) Covariates (2007) educational level, occupational class, years from retirement, employment status. Moon et al. (2012) Longitudinal Health and Retirement Study 5, (mean at baseline) 54% female Cardiovascular disease Ethnicity, birth state, education, parental education, income, wealth, physical activity, weight, alcohol use, smoking habits, ADL limitations, cardiovascular conditions. Oksanen et al. (2011) Longitudinal Finnish Public Sector Study 11,019 41% >60 75% female Purchases of antidepressant medication Gender, age, socioeconomic status, employer, area of employment, health problems, reason for retirement. Reitzes et al. (1996) Longitudinal Carolina Health and Transitions Study at follow-up 52% female Self-esteem, depression Functional limitations, age, ethnicity, marital status, gender, household income, education, occupational prestige, measures of worker

42 41 Table 2.2: Overview of the studies included in the systematic review Author (year) Study type Data used Sample size Age Gender Outcome measure(s) Covariates commitment and identity measures. Roberts et al. (2011) Longitudinal Whitehall II Study 2, at follow-up 25.9% female Cognitive function Age, gender, education, occupational class, Mill Hill score, work characteristics, leisure activities, physical and mental health. Rohwedder and Willis (2009) Cross-sectional Health and Retirement Study; English Longitudinal Study of HRS c. 20,000 ELSA Male and female Cognitive function Variation in retirement behaviour explained by national policies. Ageing; Survey of Health, Ageing and Retirement in Europe c. 9,000 SHARE c.1,000-3,000 per country Seitsamo et al. (2007) Longitudinal Finnish Institute of Occupational Health Survey 3, (mean at baseline) 40% female Occupational Stress Questionnaire (wellbeing), self-rated functional capacity Activities carried out, work characteristics, diseases, retirement status, gender, age. Siegrist et al. (2012) Cross sectional & longitudinal Survey on Health, Ageing and Retirement SHARE/ ELSA SHARE/ ELSA SHARE/ ELSA Depression Gender, age, education. Employment status, working

43 42 Table 2.2: Overview of the studies included in the systematic review Author (year) Study type Data used Sample size Age Gender Outcome measure(s) Covariates in Europe, 10, % % female hours, self-rated health. English Longitudinal Study of Ageing, Health and Retirement Study, Japanese Study of Aging and Retirement HRS 1,739 JSTAR 1,512 HRS 31.5% JSTAR 31.6% HRS 55.3% female JSTAR 40.0% female Szinovacz and Davey (2004) Longitudinal Health and Retirement Study 2, and older at baseline 49.3% female Depression Ethnicity, age, ADL difficulties, Self-rated health, education, importance of work, household assets, years in last job, forced retirement, job stresses, grandchildren, spousal health, marital information, past psychiatric problems. Van Solinge et al. (2007) Longitudinal Unnamed survey of Dutch multinational companies and older 42% female Medical consumption, severity of health problems, self-rated Access to resources, job characteristics, retirement expectations, age, gender. health Villamil et al. Cross-sectional British Psychiatric 4, % female Depression, anxiety Age, gender, marital status,

44 43 Table 2.2: Overview of the studies included in the systematic review Author (year) Study type Data used Sample size Age Gender Outcome measure(s) Covariates (2006) Morbidity Survey 2000 ethnicity, education, social class, working status, life events, debt, geographical area, tenure, house moves, close friends, perceived social support, children at home. Volkoff et al. (2010) Cross-sectional SVP50 (France) 11, and older Male and female Several physical and mental health outcomes Work-related factors, gender, age. Wahrendorf et al. (2008) Longitudinal French GAZEL cohort study 14, (mean) 27.22% female Depression, perceived control Gender, age, relationship status, volunteering status, care giving status, selfperceived health, education, household income, household size, occupation. Warr et al. (2004) Cross-sectional British subjects contacted through employers/agencies 1, (mean) 51% female Affective state, life satisfaction Age, gender, marital status, qualifications, household financial information, selfrated health, health limitations, voluntary work, work quality indicators.

45 44 Table 2.2: Overview of the studies included in the systematic review Author (year) Study type Data used Sample size Age Gender Outcome measure(s) Covariates Westerlund et al. (2009) Longitudinal French GAZEL cohort study 14, (mean) 21.3% female Self-rated health Gender, year of birth, marital status, work-related factors, occupational category. Westerlund et al. (2010) Longitudinal French GAZEL cohort study 14,104 Mean age of retirement % female Chronic disease, depression, fatigue Gender, age at retirement, employment grade, marital status, pre-retirement health characteristics. Zucchelli et al. (2007) Longitudinal Household, Income and Labour Dynamics in Australia 1, and older 44.3% female Health limitations, selfassessed health Marital status, partner s health and job status, income, tenure, age, education, occupation type, geographical information, physical limitations.

46 45 Although the studies outlined in Table 2.2 share a common focus on the topic of later-life employment or retirement in relation to health outcomes, the immediate observation is the differences in study characteristics. The second column, depicting the general analysis type of each study, shows that 15 of the papers use crosssectional analysis, meaning 74.1% of the studies are longitudinal in nature (two papers use both cross-sectional and longitudinal research methods). Additionally, the table shows the data sources of the studies. Here, the HRS is the most commonly analysed study (13 studies) followed by the Whitehall II study (8 studies), ELSA (5 studies), SHARE (4 studies) and GAZEL (3 studies). The datasets used for each piece of research also indicate the country from which each study population is derived. Of these, 41.5% are from the United States, 24.6% are British and 23.1% are European 3. The remaining 10.8% consists of four Australian datasets, two Canadian datasets, one study taken from Israeli data and one from Japanese data 4. This initial large-scale difference in study populations is important to bear in mind due to both cultural differences as well as differences in employment, retirement and pension policies which are likely to affect individuals across the life course and into old age. Variations in the data sources used for the studies subsequently leads to variations in the sample sizes of the studies. Such variation is demonstrated in the fourth column of Table 2.2, showing 15.9% of the studies contain sample sizes of below 1,000 cases, 54.0% between 1,000 and 10,000, 17.5% between 10,000 and 15,000 and 11.1% over 15, The fifth and sixth columns of Table 2.2 show further sample details for each study by detailing age and gender distributions. Sample age details are particularly relevant as they reveal additional information on the type of retirees included in the analyses. While this particular work is interested only in those of or above SPA, the a large number of the studies include some information on those retiring as early as age Excluding the UK. 4 Values here correspond to a total of 64 datasets used within the 55 studies outlined in Table Values here correspond to a total of 63 sample sizes. Missing data here is due to reporting of observation numbers only or absence of sample size information altogether. 6 Some studies stratify and compare findings on the basis of age rather than generalise retirement as the same event regardless of its timing.

47 46 Where means data is provided, 63.3% of samples have a mean age of 60 or above 7. Where studies provide information on the range of sample ages only, 69.4% include respondents younger than 60, and 38.9% include respondents aged 50 or below 8. Such information is of key importance as type of retirement is often associated with certain health outcomes and subsequently sources of selection bias. For example, those retiring involuntarily in their fifties are more likely to be displaying poor health than those capable of continuing work into their sixties (Calvo et al. 2013). Additionally, regardless of employment status, age is strongly correlated with various aspects of health and wellbeing. Declines in physical and cognitive health are typically much greater among individuals in their sixties compared to those who are younger. With regards to gender, where relevant information is provided, a slightly larger number of studies contain a higher female population than male (56.8%) 9. Additionally, nine studies include males only, yet none are inclusive entirely of women. Differences in gender ratios are important due to the likelihood that roles concerning employment and work are different for men and women across the life course, and these differences in experiences of work are likely to persist into retirement (Kim & Moen 2002; Quick & Moen 1998; Moen 1996). Additionally, some evidence suggests differences exist in the health outcomes of men and women in later-life regardless of employment status, such as the propensity for women to suffer higher levels of mental health problems throughout life (Mirowsky, 1996; Nolen-Hoeksema, 1987). Another important difference in studies to be noted is the outcomes measured. A comparison of findings on the health effects of later-life employment is made more complex when the outcomes recorded differ. For example, mental health is often inclusive of scores of depression and anxiety, but also often of depression alone, and so the effect on mental health as a generalised term becomes impossible to justly review. Within the studies included in Table 2.2, a total of 85 health outcomes are specified. Of these, 28.2% relate to depressive symptomatology or depression, 15.3% relate to self-rated health, 8.2% relate to physical functioning, and 4.7% relate 7 Thirty studies provide mean age data. 8 Thirty-six studies provide age range data. 9 This value corresponds to a total of 36 reported gender percentages across all studies (some referring to subsamples of entire study populations).

48 47 to cognitive functioning. Of the remaining 43.6% of outcomes, 23.5% pertain to other mental health conditions or states, including anxiety, life satisfaction and negative affect. It is not, however, just differences in outcomes which might be problematic in comparing studies when searching for an overall answer as to whether a general effect of later-life employment on health exists. It is also important to note that the scales used for each general outcome type might vary. Among the studies presented in Table 2.2, this issue is most visible among studies concerning depression. Of the 24 studies examining depression as an outcome, 18 (75%) use the CES-D scale, but others use different scales such as CIS-R, CIDI and Euro-D. The final column of Table 2.2 looks at the factors for which the models analysed in each of the studies adjust. Some factors, such as age, gender and background health characteristics are universally accounted for. Other commonly included covariates include marital status, income and wealth, occupational type and grade, housing tenure and educational attainment. It is important to note that many of the studies focus on a particular aspect of later-life employment or retirement, such as specific job stressors, aspects of social support and effects of multiple roles. Subsequently, a high number of studies include covariates which are specific for their analysis of interest and might be therefore less relevant to overall older working populations.

49 Key outcomes of the systematic review Table 2.2 in the previous section of the chapter has highlighted the fact that, although a large amount of previous work on the effects of later-life employment on health outcomes has been carried out, the populations on which this work has been conducted are vastly different. Firstly, study samples vary on, among other aspects, the basis of their country of origin, sample size, ratio of male to female participants and mean ages. Secondly, definitions of retirement and later-life employment differ, usually on the basis of age and reason for exiting the workforce. Similarly, definitions of health vary according to the different outcomes of interest and, furthermore, the various scales used to depict these. Thirdly, the methodologies of the studies also differ, with varying types of statistical analysis offered, large differences in sample sizes and some studies cross-sectional in nature while others are longitudinal. Bearing the above in mind, it might already be expected that heterogeneity will be observable in relation to the findings of the studies presented in Table 2.2. The following sub-sections of the chapter focus on the findings taken from the systematic review in accordance with the outcomes of interest to this particular study. Depressive symptomatology and self-rated health are analysed in individual sections, with an overview of study findings and presentation of actual effects offered in table format, and discussions following thereafter. Although the forthcoming research within this study places an interest in cognitive function, there were not enough studies concerning this outcome to allocate a section to it in its entirety. Instead, studies which do concern cognitive function will be discussed alongside other key health outcomes measured, in section Where results are presented for findings concerning retirement, these are always in reference to those who are still in employment.

50 Depression Table 2.3 presents the findings of 22 studies concerning depression and its relationship with later-life employment. Across all studies, negative coefficients were indicative of fewer depressive symptoms, and positive coefficients indicated more. Unfortunately, the large majority of literature places its key focus on retirement rather than employment, and literature concerning theories of employment in later life again tends to focus on its exit rather than its continuation. However, studies are only included in the review on the basis that findings concerning retirement are given in direct comparison to reference groups of older workers. Table 2.3: Findings from literature regarding depressive symptomatology Study (first author) Adelmann (1994) Behnke (2012) Butterworth (2006) Buxton (2005) Calvo (2013) Outcome scale CES-D CES-D CIDI CIS-R CES-D 13 (reverse scale) Overview of findings Effect type Effect 10 Multiple roles (including employment) are associated with lower depression than no or single roles. Retirement has no significant effect on depression. Early retirement among men is associated with greater depression than working, but not normal retirement. Early retirement is associated with higher levels of depression than working. There is no significant effect of retiring after normal retirement age in terms of CES-D scores. Standardised regression coefficient Multiple roles vs. no roles IV ATET Retirement vs. employment Odds ratio Early retirement Normal retirement (both vs. employment) Odds ratio Early retirement vs. employment IV Fixed effects coefficient ** * * 12 Short term effect Coefficients are taken from the fully adjusted models presented in each analysis. 11 Specific p values were not given for this study, but findings were stated to be significant with 95% confidence intervals. 12 Specific p values were not given for this study, but findings were stated to be significant with 95% confidence intervals. 13 In the original paper, the CES-D scale has been reversed. Scores presented here use the non-reversed scale so as to be comparable to other values in the table.

51 50 Table 2.3: Findings from literature regarding depressive symptomatology Study (first Outcome Overview of findings Effect type Effect 10 author) scale Long term effect IV Random effects coefficient Short term effect Long term effect (all vs employment) Choi (2007) CES-D Working is associated with significantly lower CES-D scores among women, but not men (coefficients are part of a larger model concerning life changes in older age). The effect of work is greater when in conjunction with volunteering. Standardised regression coefficient Men Working Work & volunteering Women Working Work & * ** volunteering (all vs. no work or volunteering 14 ) Coe (2008) CES-D No significantly detrimental effect of retirement on depression scores among men. IV coefficient Short-term retirement (2 years) Long term retirement (4 years) (both vs. employment) Coe (2011) Euro-D and selfrated No significant effect of normal retirement on depression among men. IV coefficient Retirement Euro-D score Felt depressed in a month (both vs Although the reference group is not specified as retired, the sample used in this study is aged 65 and over and so this group is likely to be largely comprised of individuals who have retired.

52 51 Table 2.3: Findings from literature regarding depressive symptomatology Study (first author) Coursolle (2010) Dave (2006) Drentea (2002) Fernandez (1998) Hao (2008) Keyes (2007) Outcome Overview of findings Effect type Effect 10 scale employment) CES-D Retirement is associated with lower Fixed effects depression levels after employment coefficient with a high work-family conflict Retired (all) -0.97** occurrence. Retired (work -0.98*** interfered with family) Retired (family interfered with work) (all vs. employment) CES-D Complete retirement leads to a Fixed effects significant increase in symptoms of coefficient depression. Retirement ** (vs. employment) CES-D Retirement has no significant Non-significant associations with depression. coefficients (vs. employment) not provided. CES-D Retirement is associated with a ANOVA means decrease in depression. difference Retirement -0.6* vs. employment CES-D Work in later-life is protective of GCM coefficient increased depressive Work only symptomatology. Combined with Work and *** volunteering, this effect is even volunteering *** greater. (both vs. no activity ) CES-D Retirement is significantly OLS coefficient associated with lower levels of Retirement -0.17** depression. The focus on religion in Religious identity -0.31** this study suggests the interaction and retired between religious identity and Religious identity depression is greater among those and working -0.18* who retire, rather than those who (retirement is in

53 52 Table 2.3: Findings from literature regarding depressive symptomatology Study (first Outcome Overview of findings Effect type Effect 10 author) scale continue to work. reference to employment and employment is in reference to retirement) Kim (2002) CES-D There is no significant relationship between either continuous or new retirement on symptoms of depression. Unstandardised regression coefficient Men Continuously retired Newly retired Women Continuously retired Newly retired (all vs. employment) Mandal (2008) CES-D Voluntary retirement is significantly associated with a reduction in symptoms of depression for both men and women. Where workforce exit is not voluntary, depression is significantly increased in each instance. 2SLS coefficient All Voluntary work exit Involuntary work exit Men Voluntary work exit Involuntary work exit Women Voluntary work exit Involuntary work * 0.244* * 0.086** * 0.205* exit (all vs. employment) Oksanen Drugs Old-age retirement is associated with Prevalence

54 53 Table 2.3: Findings from literature regarding depressive symptomatology Study (first Outcome Overview of findings Effect type Effect 10 author) scale (2011) purchase less antidepressant usage compared to pre-retirement. Retirement due to poor mental and physical health is associated with an increase. difference 1 year prior to and 1 year after retirement 2 to 4 years -0.9* following retirement (both vs. employment) Reitzes (1996) CES-D Retirement is significantly associated with fewer symptoms of depression. Regression coefficient Retirement vs *** employment Szinovacz (2004) CES-D Being retired for one year or less is significantly associated with lower depression among men with a nonworking spouse. This is not true for women, or for men or women who OLS coefficient Men Spouse not employed Retired 1 year -0.54** have been retired for more than one year. vs. employment Villamil (2006) CIS-R Retirement is significantly associated with greater risk of depression for men, but not for women. Odds ratio Retirement (men) Retirement (women) 3.2** 1.4 (both vs. employment) Wahrendorf (2008) CES-D Although not relating to paid employment, the study shows a significant association between productive activity and lower levels of depression. There is no significant association with provision of care. Standardised regression coefficient Almost daily voluntary work Almost daily informal help 16 Almost daily care for a person -0.17* * Odds ratio is stated as significant, but p value is not shown. Informal help is defined as provision of help to family, friends or neighbours.

55 54 Table 2.3: Findings from literature regarding depressive symptomatology Study (first Outcome Overview of findings Effect type Effect 10 author) scale (vs. no activity) Westerlund CES-D Retirement is significantly Odds ratio (2010) associated with a decrease in 1 year after 0.60* 17 depressive symptoms. retirement vs. 1 year before retirement *p<0.05; **p<0.01; ***p<0.001 It is immediately clear, on looking at Table 2.3, that establishing whether or not an effect of later-life employment on depression exists is difficult. Although the CES-D scale is used in the vast majority of cases, and where it is not, other scales follow the same direction (with negative values indicating fewer symptoms of depression), the column describing effect type shows a varied list of outcomes, with many studies placing a focus on ideas outside just those concerning employment as opposed to retirement. The total number of studies which provides information concerning the direct association with later-life employment, as opposed to retirement, is 18. Of these, seven conclude that later-life employment has a detrimental effect on depression (Oksanen et al. 2011; Westerlund et al. 2010; Mandal & Roe 2008; Keyes & Reitzes 2007; Szinovacz & Davey 2004; Fernandez et al. 1998; Reitzes et al 1996), five state its effect is beneficial (Hao 2008; Choi & Bohman 2007; Dave et al. 2006; Villamil et al 2006; Adelmann 1994), and six find no significant association exists at all (Calvo et al. 2013; Behnke 2012; Coe & Zamarro 2011; Coe & Lindeboom 2008; Drentea 2002; Kim & Moen 2002). Unfortunately, due to the varied nature of effect types, statistical models employed and study methodologies, it is impossible to quantitatively compare the coefficient values provided in the final column of Table 2.3. However, some general findings can still be found. It should be noted that where studies find retirement, as opposed to working, to have a detrimental effect on depression, these findings are inclusive of all participants (for example, men and women and all work backgrounds). In contrast, the studies with evidence to show a beneficial effect of employment contain effects which are dependent on some individual or workplace characteristic. To elaborate, 17 Odds ratio is stated as significant, but p value is not shown.

56 55 the study effects in the papers by Choi & Bohman (2007) and Villamil et al. (2006) are gender-specific. Choi & Bohman find a significant reduction in the depression scores of women who work, but not of men 18, and Villamil et al. find retirement to be associated with increased depression in men yet not in women. Mechanisms through which men and women experience later-life work and retirement differently will be discussed in detail further on in this section of the chapter. It is also interesting to note that three of the four studies showing a positive effect of employment place a focus on multiple roles in later-life. The study by Adelmann (1994) finds a significant improvement in depressive symptoms on the basis of individuals participating in more than one socially productive role. The studies by Choi & Bohman (2007) and Hao (2008) demonstrate that, although working on its own is protective of mental wellbeing, in combination with volunteering, the protective effect is even greater. However, it should be noted that the study by Choi & Bohman uses a reference group comprised of individuals with no activity, rather than of those who are retired, and so results are not necessarily reliable in reference to the research questions presented here. The associations between productive activities and health evident here relate well to ideas arising from activity theory in later-life. As discussed in section 2.1 of the chapter, higher levels of involvement in socially meaningful activities, including volunteering as well as employment, are associated with higher levels of health and wellbeing outcomes. These include personal control (Hayward et al. 1998) and selfworth and meaning (Wethington et al. 2000; Herzog & House 1991). Additionally, participation in a higher number of productive activities is likely to lead to a greater number of social contacts, leading to better psychological wellbeing through increased sources of social support and personal support networks (Lin et al. 1999; Lemon et al. 1972). Finally, a paper by Siegrist et al. (2004) suggests the altruistic nature of voluntary work improves levels of self-esteem, and so might contribute additional benefits to mental wellbeing than participation in paid employment alone. Examining, on their own, the studies showing no significant effect of later-life employment (or of retirement in comparison to later-life employment), it is 18 The reference group in the study by Choi is those who do not participate in either work or voluntary work, not those who are specifically retired. However, the study focuses on participants aged 65 and over only so the majority of the sample will be in retirement.

57 56 worthwhile to note that four of them use analysis types which might be particularly useful in terms of reducing bias and establishing causal effects. The studies by Behnke (2012), Calvo et al. (2013), Coe and Lindeboom (2008) and Coe and Zamorro (2011) each use instrumental variable (IV) analysis, which aims to specifically seek causal relationships through eradication of selection bias present in sample characteristics (Greenland 2000; Angrist 1996). The study by Kim and Moen (2002) presents very well stratified results, by both retirement length and gender. These pieces of research are in contrast to many other studies which provide significant effects through some form of adjusted regression coefficient for those who are employed in direct contrast to those who are retired, or vice-versa. Differences in results such as these highlight the importance of using a method sufficiently able to deal with group differences and issues of selection bias, and may highlight where studies which have not used such methods may have subsequently estimated biased results. Table 2.3 clearly demonstrates that results concerning the effect of working in later life in general are heterogeneous. Studies can, however, also be broken down by the different study-specific aspects focused on in each. For example, some studies provide results on the basis of retirement type. As would be expected, where early retirees were included in a study, results demonstrated a definite detrimental effect on symptoms of depression, and often this was in contrast to normal retirement as well as employment. Butterworth et al. (2006) state depression significantly increases among early retired men, but not among those who retire at the normal age. Buxton et al. (2005) echo the findings of increased depression among early retirees, although do not include a group of normal retirees against which to compare this result. Oksanen et al. (2011) further this idea by comparing those who retire due to poor physical and mental health to those who retire normally and demonstrate that normal retirees see a decrease in depression (on the basis of antidepressant purchases) and those who retire due to poor health see a significant rise. Finally, the study by Mandal et al. (2008) examines the association between voluntary retirement and health, as opposed to involuntary workforce exit, and again finds a significant decrease in depression among those who retired voluntarily and an increase among those whose retirement was involuntary.

58 57 The findings concerning early retirement are not surprising as long-term poor health or onset of disability is likely to be a strong determinant of depression regardless of employment status (Mein et al. 2000; Kathol & Petty 1991). Early exit from the workforce, and the subsequent potential loss of meaningful roles, social networks and enjoyable activity are likely to exacerbate higher levels of depression arising from illness or disability alone (Butterworth et al. 2006). Additionally, when the workforce has been left due to functional limitations, the capacity to continue meaningful roles through other channels, such as voluntary work or participation in hobbies might also be lowered, again leading to higher levels of depression and poor mental wellbeing. Involuntary workforce exit due to non-health related reasons, such as organisational restructure, is also associated with higher levels of depression and poorer mental health (Van Solinge & Henkens 2007; Szinovacz & Davey 2005). Early retirement taken on the basis of having to care for a spouse or family member has been shown to bear significant detrimental effects on wellbeing (Szinovacz & Davey 2005; Szinovacz et al. 2001). This is likely to be due to the effects of forced work exit in combination with the stress associated with the provision of care. From a continuity theory perspective, not only is the individual forced to leave a role which provided a sense of fulfillment, but a new, less enjoyable mandatory role has been assigned. Another channel of involuntary retirement arises from job loss and redundancy, which again have been shown to lead to increased rates of depression (Gallo et al. 2006; Gallo et al. 2000). As well as a temporary loss of income, older workers might consider prospects for re-employment to be particularly poor, and may worry that any new roles acquired will have lower incomes and lower workplace statuses (Couch 1998). Such ideas again tie in with continuity and role theories, by which higher levels of wellbeing are associated with the individual s ability to maintain a sense of fulfillment through participation in roles which are meaningful and enjoyable. Another important aspect to be considered is length of retirement. Four of the studies included in Table 2.3 examine retirement on the basis of its length. Theories concerning transitions into retirement and beyond have focused on the idea of an initial honeymoon period, which is associated with an improvement in wellbeing, followed by various stages of reorientation and stability which are associated with declines in mental wellbeing (Atchley, 1976). Short-term retirement has also been

59 58 associated with better mental health by means of role strain reduction, by which exit from a stressful or demanding workforce provides an initial sense of relief (Kim & Moen 2002). Three of these studies (Calvo et al. 2013; Coe & Lindeboom, 2008 and Kim & Moen 2002) conclude that no significant effects exist on depression regardless of whether retirement has been relatively recently entered (one-to-two years) or if it is long-term. Again, it is difficult to compare study coefficient magnitude and direction here due to vast differences in methodology and analysis type employed. The final study with a focus on retirement length (Szinovacz & Davey 2004) finds a significant result on the basis of gender and spousal employment status, stating retirement of less than one year is associated with fewer depressive symptoms among men with non-working partners only. This ties in with ideas of gender conformity and role theory, whereby a husband s retirement before his spouse leads to feelings of a loss of power in the marital relationship, in turn leading to lower levels of self-esteem and mental wellbeing (Davey & Szinovacz 2004; Kim & Moen 2002). Although it has already been mentioned in reference to studies showing a beneficial effect of employment, the role gender appears to play across all studies included in the review can also be examined. Choi & Bohman (2007) find later-life employment to be good for women but not men. Studies by Szinovacz & Davey (2004) and Villamil et al. (2006) instead place a focus on retirement, with the first associating retirement with significantly less depression among men and the latter significantly more. It is important to note, however, that the finding by Szinovacz & Davey is that which, as mentioned beforehand, focuses on specifically short-term retirement among men with non-employed spouses, and so a fair comparison of results cannot truly be made on the basis of strong sample differences. There are several mechanisms through which later-life employment and retirement might be experienced differently on the basis of gender. Ideas stemming from both role and continuity theory lend to the idea that men might find effects of retirement on mental health particularly detrimental. Role theory promotes the idea of employment as a central role, and loss of this in turn leads to poorer mental health outcomes (George 1993). Ideas surrounding gender-role conformity (Moen et al. 2001) suggest loss of these roles might be particularly detrimental to the health of males, as men are more likely to view their position and role in the family as closely

60 59 linked to their ability to work and provide (Davey & Szinovacz 2004; Kim & Moen 2002; Myers & Booth 1996). This effect might be particularly pronounced if the male exits the workforce before his spouse (Davey & Szinovacz 2004; Moen et al. 2001). Such ideas might be particularly prevalent when researching older cohorts, where a stronger sense of traditional family structures is likely to exist. Additionally, employment roles might sometimes be more significant to men than women as they are more likely to have spent continuous time in set careers or workplaces, while women are more likely to have taken time out from employment to provide other roles, such as childcare or caring for family members (Kim & Moen 2002; Han & Moen 1999). Women are also more likely to work in lower paid, poorer quality and less rewarding roles, again making exit of the workforce an easier transition (Elder & Schmidt 2004; Bosma et al. 1998). Similarly, a paper by Szinowacz & Washo (1992) outlines the means by which a higher number of significant life events outside of the workplace experienced by women than men leads to an easier transition into retirement. However, when exit from the workforce is made on the basis of life events, such as in order to provide care-giving duties to an ill family member, mental wellbeing might be negatively impacted (Dentinger & Clarkberg 2002). It must also be remembered that employment roles are not necessarily beneficial to health and wellbeing. Ideas of role strain theory can also be applied to the different ways in which men and women experience later-life working and retirement. When retirement is taken from roles which are particularly stressful or physically demanding, a sense of relief may lead to increased mental wellbeing. As men are more likely to participate in manual and physical employment than women, this particular mechanism of work affecting health is likely to be more prevalent among male populations (Ekerdt et al.1983). There are two studies in Table 2.3 which present significant findings on the basis of specific individual characteristics. Firstly, Keyes & Reitzes (2007) find retirement is, overall, significantly associated with less depression, and that this finding is strengthened by having a greater religious identity. Secondly, Coursolle et al. (2010) find retirement is beneficial on the basis of the individual exiting a job which placed a high conflict on the balance between work and family. This latter finding is particularly interesting as it suggests work quality and demands play a role in mental wellbeing, which will be investigated in greater detail further in the study.

61 60 Finally, it is worth noting that, although not strictly concerning later-life employment, volunteering in older age appears to always show beneficial results in relation to depression. The aforementioned studies by Adelmann (1994), Choi & Bohman (2007) and Hao et al. (2008) state the positive impact of multiple roles, each of which includes volunteering (the latter two studies focus only on volunteering as a multiple role in conjunction with employment). Additionally, the study by Wahrendorf et al. (2008) shows almost daily volunteering or informal help to significantly lower depressive symptoms. This set of findings tie in with several of the aforementioned theories concerning older age and employment. As discussed previously, participation in voluntary work allows continuation of meaningful and purposeful roles which may have been lost through retirement (Herzog & House 1991), and additionally, the voluntary nature of the work carried out lends to positive effects on self-esteem (Siegrist et al. 2004). Voluntary work also aids in the acquisition of social support and networks, which again may be beneficial to mental wellbeing (Lin et al. 1999). These factors might be especially beneficial when the individual has retired from work which was stressful or demanding. As the ability to ascertain whether or not an overall effect of working later in life on depression is difficult, it is worth considering which of the studies included in the review might lend the greatest strength in answering the key research questions. Four of the studies presented in Table 2.3 use instrumental variable (IV) analysis (Calvo et al. 2013; Behnke 2012; Coe & Zamarro 2011; Coe & Lindeboom 2008), a method which controls for confounding through the use of an instrumental variable related to the exposure but not to an individual s background characteristics, therefore estimating causal effects of the exposure on outcomes. In each instance, no significant effect is found of later-life working on depression. However, it should be noted that these studies do not differentiate between types of work or work quality, which are likely to influence both the propensity to work for longer as well as health. These studies also include individuals who retire prior to reaching SPA, with no definition of whether early retirement was voluntary or forced. This differentiation of early retirement type is especially important in light of the paper by Mandal & Roe (2008), which uses 2-stage least squares IV analysis to establish causal effects, and finds voluntary work exit to be significantly associated with a reduction in symptoms

62 61 of depression, yet forced work exit to be significantly associated with a worsening of depressive symptoms. The majority of the remaining studies included in Table 2.3 use OLS or logistic regression techniques. Here, coefficients estimate associations between later-life working and depression, and not effects which are causal in nature and reliably demonstrate direction. Standard regression techniques assume no correlation exists between the model regressors and its error terms, but such an assumption might be considered problematic when considering that the association between later-life working and depression is likely to be influenced by many other factors. For example, greater levels of wealth might be associated with better employment types, which in turn lends to an increased likelihood of prolonged workforce participation, but it may also be associated with lower levels of depression through better levels of life satisfaction. As a result, it can be concluded that the majority of studies which have previously examined effects of later-life working on depression are not useful in establishing whether or not causal effects are present. Further to methodological issues arising from use of techniques which do not sufficiently control for confounding, 5 of the studies included in Table 2.3 are crosssectional in nature (Butterworth et al. 2006; Villamil et al. 2006; Buxton et al. 2005; Drentea 2002; Adelmann 1994). In these instances, the issue of reverse-causality is added to the potential problematic nature of inadequate methodology. A further 4 studies use samples with n<1,000 (Keyes & Reitzes 2007; Kim & Moen 2002; Fernandez et al. 1998; Reitzes et al. 1996). Of these, three establish a significant association between retirement, as opposed to later-life working, and lower levels of depression (Keyes & Reitzes 2007; Drentea 2002; Reitzes et al. 1996). However, the, low statistical power in comparison to studies with larger samples leads to an increased likelihood of errors and unreliability of results Self-rated health Table 2.4 presents the findings concerning self-rated health 19. Where the outcome scale used is described as poor health, this refers to a binary variable taken from a five-point Likert scale self-assessed health question (with good health as the 19 Some studies shown in Table 2.2 included and discussed means differences in self-rated health scores, but did not subsequently include the outcome in their final, adjusted, results. These studies have been omitted from the table.

63 62 reference category). Again, findings are only discussed here if a direct comparison between employment and retirement is established. Discrepancies between studies existed on the basis of whether a high Likert-scale score was indicative of better or worse health, and so for ease of interpretation, some results presented in Table 2.4 have been reversed so that, in all cases, a lower score is indicative of better self-rated health. Table 2.4: Findings from literature regarding self-rated health Study (first Outcome Overview of findings Effect type Effect 20 author) scale Behnke (2010) Poor health Retirement leads to a decline in self-rated health. ATET Retirement 0.043* vs. employment Calvo (2013) 5-point Likert scale Retirement at age 60 has a significant negative effect on selfrated health. IV Fixed effects coefficient Short term effect Long term effect IV Random effects coefficient Short term effect Long term effect (all vs. 0.37** 0.29* 0.38** 0.34** employment) Coe (2011) Poor health Early retirement is significantly associated with a temporary IV coefficient Retirement vs * reduced likelihood of poor selfrated health. employment Crowley (1986) Subjective self-rated health The initial years of retirement are not significantly associated with health. Regression coefficient Normal retirement vs. employment Dave (2006) Poor health Complete retirement significantly increases occurrence of poor selfrated Fixed effects coefficient 0.025*** health. Retirement vs. employment Kremer (1985) 5-point Likert scale There is no significant effect of retirement on self-rated health. Means difference Employed score Retired score Coefficients are taken from the fully adjusted models presented in each analysis.

64 63 Table 2.4: Findings from literature regarding self-rated health Study (first author) Outcome scale Overview of findings Effect type Effect Latif (2012) 5-point Likert scale There is no significant effect of retirement on self-rated health. Fixed effects IV coefficient Retirement vs employment Mojon-Azzi (2007) Change in health There is no significant short-term effect of retirement on health. However, effects show the Odds ratio Improvement in health following 1.9 (CI ) association to be beneficial. retirement Van Solinge (2007) 5-point Likert scale The most voluntary form of retirement (not health or organisation related) is not OLS coefficient Very voluntary retirement vs (t -1.11) significantly associated with selfrated health. employment Westerlund (2009) Change in health There is no significant increase in suboptimal self-rated health in the period one year prior to and one year following statutory retirement. GEE odds ratio for suboptimal health Statutory retirement (year before compared to year after) 0.57 (CI ) *p<0.05; **p<0.01; ***p<0.001 A clear definition of self-rated health is important if effects upon it are to be understood. When respondents are asked to rate their general health, responses are likely to incorporate aspects of physical, mental and functioning health (Pinquart 2001, Singh-Manoux et al. 2005). Research has found correlations between subjective health measures and aspects of mental health such as depression and anxiety (Pinquart 2001). However, correlates are stronger between subjective health measures and objective physical health (Kaplan & O Baron-Epel 2003; Pinquart 2001), and correlates between subjective health and mortality are consistently strong across all socio-demographic groups (Idler & Benyamini 1997; Kaplan et al. 1988) In the ELSA dataset itself, the measure of self-rated health correlates only slightly higher with a measure of physical functioning (ADL score) than a measure of depression (CES-D) r=0.268 and 0.265, respectively. This will be discussed in greater detail in Chapter 4.

65 64 Three of the four significant results in Table 2.4 are produced by means of IV analysis which again, through the use of an instrumental variable unrelated to error terms, seeks causal effects on outcomes of interest (Greenland 2000). Self-rated health might be particularly prone to issues of selection bias in studies of older populations. If it is considered that health deteriorates with age regardless of employment status, it is especially important to recognise the potential of confounding of results due to age within analyses of this outcome. It is also important to consider issues of selection bias arising from the increased likelihood of those with poorer levels of health to be exiting the workforce, therefore contributing to a lower health status of individuals who are retired overall. Therefore, interpretation of analyses using methods such as IV to directly deal with issues of confounding is especially important here. The lack of ability for other methods to successfully deal with bias arising from confounding should be particularly well considered here. The significant findings concerning self-rated health continue to show a variety of results. Two of the IV analyses conclude retirement, as opposed to continuation of work, to have a detrimental effect on self-rated health (Calvo et al. 2013; Behnke 2010), while the study by Coe & Zamarro (2011) shows retirement to be associated with a temporary improvement. Again, however, this difference in findings is likely to be due to differences inherent in the study designs. The study by Calvo et al. uses US data from the HRS, the study by Behnke et al. uses ELSA and the study by Coe & Zamarro uses European data from SHARE. Additionally, the instrumental variables used to assess relationships between working and health vary by study: that by Calvo et al. uses early retirement opportunity windows, that by Behnke et al. uses attainment of SPA and the study by Coe & Zamarro uses country-specific early and state pension retirement ages. On the basis of this, it is unsurprising that even studies using the same methods of analysis produce widely differing results. Theories of ageing and employment support the idea that self-rated health can be either negatively or positively impacted by continuation of work, and differentiation can be dependent on whether subjective health measures are deemed representative of physical or mental wellbeing. Activity, role and continuity theories suggest extended employment might be beneficial to health in that continuation of enjoyable roles, or roles which might enable greater physical wellbeing, lead to a protective

66 65 effect on health decline (Lemon et al. 1972). However, prolonged working, or cases of job-lock, in roles which are physically demanding might encourage faster decline, and instead lead to better health in retirement (Ekerdt et al. 1983). Where self-rated health reflects mental wellbeing, the same issues relating to longer periods of working are applicable as in the previous discussion of depression. Again, loss of socially-meaningful roles or forced continuation of employment due to financial pressures is likely to lead to a worsening in scores, while relief of stressful or demanding activities, ability to enjoy the retirement period or continuation of fulfilling employment roles is likely to enhance scores. However, retirement also brings the possibility of a more relaxed pace at which to enjoy activities, as well as increased time allowance for activities which promote health and wellbeing, such as exercise and physical activity (Jokela et al. 2010; Midanik et al. 2005; Drentea 2002) Other health outcomes Table 2.2 identified the variation in outcomes examined within research on health and later-life employment. Previous sections focused solely on an in-depth discussion of studies regarding depression and self-rated health. This section of the chapter provides a discussion of other health and wellbeing outcomes commonly discussed in literature in reference to working beyond normal retirement age. Cognitive functioning Four of the studies in Table 2.2 examine effects of working on cognitive function. Three of the studies (Bonsang et al. 2012; Rohwedder & Willis, 2010; Adam et al. 2007) use the same outcome of episodic memory, and use IV analysis to seek causal effects. The fourth study by Roberts et al. (2011) examines a cognitive function index comprised of verbal memory, semantic fluency and verbal fluency. Each study analyses the scores of retirees in reference to older workers. The three IV analyses find significant negative effects of retirement as opposed to work on cognitive outcomes, and the study by Roberts et al. finds retirement to be associated with

67 66 smaller increases in cognitive function than continuation of work across study waves 22. Overall findings that retirement damages cognitive function tie in with the use it or lose it hypothesis (Coyle 1994) which proposes continuation of mindful activities in later life prevents the onset of dementia, and further that a lack of mentally challenging activities might exacerbate loss of cognitive function (Hultsch et al. 1999). Verghese et al. (2003) find participation in leisure activities in older age (75 years onwards) to be associated with a reduced risk of dementia. A popular theory here concerns the protective effect of continuation of mental tasks on cognitive reserve (Scarmeas & Stern 2003; Stern 2002). This can be described as the brain s ability to maintain cognitive function when age-related pathological deterioration begins to take place. In relation to a life course perspective, Stern (2003) notes that individuals ability to maintain higher levels of functioning is heterogeneous and is likely to be affected by lifestyle as well as genetic factors, and so issues of selection become apparent. Research has shown higher occupational achievement is associated with a better cognitive reserve (Cagney & Lauderdale 2002; Schooler et al. 1999), as is greater educational attainment (Le Carret et al. 2003). However, evidence also suggests the individuals with a lower ability to maintain cognitive reserve in the first instance are the individuals who do not attain higher qualification and educational achievements earlier in the life course (Stern 1999). In turn, when examining the cognitive function of older workers, it must be considered that cases with the poorer cognitive reserve are the cases who retire earlier. Considering factors such as education and occupation to influence this, methods successfully accounting for selection from such factors may help to reduce bias in results. Physical functioning Four studies included in the review focus on effects of work, or retirement in comparison to work, on physical functioning. Measures of physical functioning include the SF-36 questionnaire (Jokela et al. 2010; Mein et al. 2003), difficulties with activities of daily living (ADL) (Dave et al. 2006) and self-reported measures (Luoh & Herzog 2002). Results here are, again, varied. Jokela et al. find statutory, 22 The increase in cognitive function scores with age is likely to be due to learning effects, whereby respondents become familiar over time with the cognitive function questions asked of them, therefore improving performance.

68 67 but not early, retirement is associated with better physical functioning than continuation of work. Dave et al., however, find complete retirement to be associated with a 61.6% increase in ADL difficulties. Mein et al. find physical functioning decreases with age regardless of employment status. The study by Luoh & Herzog finds participation in one-hundred or more annual hours of paid or voluntary work leads to subsequent lower levels of poor functioning. This study is particularly interesting as its focus is on people over retirement age only and confounding from selection into early retirement, which is strongly linked to levels of functioning and disability (Karpansalo et al. 2004; Mein et al. 2000), are not troublesome here. There are several mechanisms through which continuation of work and retirement might affect physical functioning. The first of these is the physical activity provided by each, and effects are likely to vary on the basis of sample characteristics. For some, for example those working in manual jobs, employment might be a key source of exercise, and longer periods of employment might therefore be protective of physical decline (Gall & Parkhouse 2004). Physical decline in retirement might be particularly pronounced if the individual does not continue physical activity. On the other hand, exit from the workforce might lead to an increased allowance for physical activity, leading to improvements in, or protection from decline of, physical functioning. Such ideas tie in with activity theories of successful ageing and role continuity, whereby the continuation of activities and tasks beneficial to health are adapted within the changing lifestyle brought about by retirement (Herzog & House 1991). Mental health (excluding depression) Several of the studies presented in Table 2.2 focus on mental health outcomes other than depression. Commonly reported ones include life satisfaction, anxiety, morale and affective state. There is generally a higher rate of significant findings on other mental health outcomes than was observed among the findings concerning depression in section Again, however, these findings are varied, and heavily dependent on factors outside of straightforward worker or retiree group membership. Many studies suggest retirement, in comparison to continued work, has a beneficial effect. Gall et al. (1997) and Kim & Moen (2002) find retirement to be associated with higher

69 68 levels of life satisfaction and morale, respectively. In both instances, evidence of the aforementioned honeymoon period of retirement persists. Gall et al. find life satisfaction to be at its highest rate in the one-year period following retirement, while Kim & Moen find levels of morale to be at their highest in the two years following workforce exit. In both instances, the relationship attenuates over time. This strengthens the argument of retirement as a liberating experience from stressful and demanding roles of employment, although in time adjustment to retirement forms its own stressors (Atchley, 1989). Coursolle et al. (2010) find retirement to be associated with better levels of a psychological functioning scale, which incorporates aspects of self-acceptance, autonomy and feelings of purpose in life. However, this finding is only significant among men who have retired from situations of high family-work conflict, whereby familial circumstances increased workplace stress. Women retiring from high family-work conflict situations did not see an association with improved psychological functioning. This links to aforementioned ideas of the centrality of the role of work being greater for men. Women are tied to family-related duties to a greater extent than men, resulting in less centrality of the employment role across the life course and subsequently smaller impacts of retirement (Kim & Moen 2002). Drentea (2002) found retirement to be associated with significantly lower levels of anxiety and distress among both men and women. However, this relationship was dependent on the activities carried out in retirement, and was only present among retirees who described their daily pursuits as fulfilling, non-routine and inclusive of social interaction. This again supports the idea that workforce exit leads to relief from stressful and routine roles. However, the ability to enjoy personal pursuits in retirement relies on adequate financial resources and functioning ability (Scherger et al. 2011). Studies which exclusively found later-life employment to be beneficial to mental wellbeing were those by Villamil et al. (2006), Gill et al. (2006) and Bosse et al. (1987). The study by Villamil et al. found employment, as opposed to retirement, to signifincantly reduce anxiety among men, although this was not the case for women. The study by Bosse et al. focussed only on men and found working at age 66 and beyond was associated with reporting the fewest psychological symptoms. Gill et al.

70 69 found working men to have better mental health than those who had retired. This relationship was strongest among early retirees and attenuated as age increased. These findings again relate to the increased centrality of the role of employment for men. Maintaining employment might be particularly protective of mental health among males who have experienced more continuous workforce participation across the life course than women. Additionally, adherence to gender-role conformity through workforce participation may also lead to reductions in mental health problems for men. As discussed previously in the chapter, continuity and activity theories maintain that the ability to prolong participation in activities which are rewarding and lead to a greater sense of self-fulfillment are also likely to lead to better mental wellbeing. Finally, the study by Warr et al. (2004) shows later-life employment was significantly associated with better life satisfaction and affective state than retirement, but this was dependent on the respondent preferring to work in later life than retire. Outcomes were also better among early retirees on the same premise: that their early retirement was voluntary and preferred. In each instance, mental wellbeing was a function of role quality rather than of role membership alone The role of work quality Work quality is an important predictor of health and wellbeing across the life course (Siegrist et al. 2004; de Jonge et al. 2000; Adelmann 1987). Several studies identified through the systematic review procedure focus on the importance of occupational type and quality. Although the definition of good quality work varies across studies, as do the outcomes measured, findings are typically similar, in that indicators of better quality employment are associated with better health and wellbeing. Warr et al. (2004) found life satisfaction and affective balance are highest among older workers with greater levels of control and autonomy at work. Similarly, Wahrendorf et al. (2008) found lower levels of depression among older voluntary workers with high levels of autonomy. McMunn et al. (2009) found workers, volunteers and care-givers with a good effort-reward balance had better wellbeing scores than those without. Conversely, a study focusing on poor occupational characteristics found higher levels of both mental and physical health problems

71 70 among older workers with intense time pressures within the workplace (Volkoff et al. 2010). Employment type and grade are also commonly used measures of workplace characteristics. Although these are not necessarily representative of the quality of employment, clear associations between these factors and health outcomes emerge. Broerson et al. (1996) find a lower prevalence of a range of general health complaints among white collar workers than blue collar workers, Chandola et al. (2007) find a faster decline in physical health among those in lower status occupations than high, Seitsamo et al. (2006) find manual older workers display the lowest levels of general mental and functional wellbeing, and Jorm et al. (1998) find realistic occupations, including skilled trades, technical and some service work, to be associated with poorer cognitive performance and a higher risk of certain types of dementia in reference to all other types of work. A large number of studies examining effects of employment quality and type consider the lasting effects of workplace characteristics into retirement. Results generally follow the same pattern, with retirement from good quality employment or better occupational grades associated with better health outcomes in retirement (Christ et al. 2007; Mein et al. 2003; Charles et al. 2002; Tuomi et al. 1991). However, retirement from poor quality work and lower occupational grades also appears to be beneficial to health. While poor working environments and high levels of workplace stress were associated with poorer health before retirement, levels of depression (Coursolle et al. 2010; Fernandez et al. 1998), suboptimal self-reported health (Westerlund et al. 2009) and mortality risk (Marmot et al. 1996) improved once retirement started, although relationships generally attenuated over time. Although the relationship between work quality and health might seem straightforward, issues of selection and confounding must again be considered. While it might be assumed that participation in good quality employment leads to better health and wellbeing, it is likely that factors affecting the propensity to have good quality employment are also associated with a propensity to have good health. The typical traits of good quality employment, such as better financial rewards, low stress environments and lack of physically demanding labour, are those which are also often characteristic of higher grade occupations (Stansfeld et al. 1999). Higher

72 71 grade occupations are likely to be more commonly observed among individuals with higher levels of education, social class and wealth, all of which are characteristics likely to be associated with better health across the life course (Chandola et al. 2007). Subsequently, the magnitude of the effect of the good quality work on health in itself is uncertain, as it is likely that these additional factors also play a role in determining health outcomes in later life. Similarly, the traits of poor quality work and low grade occupations are likely to be similar and so lower levels of health among these workers might be attributed to a range of social factors predicting poorer health in addition to the work type itself. Such ideas tie in with the human capital theory proposed in section 2.1, whereby cumulative beneficial or detrimental effects of work type throughout the life-course lead to increased likelihoods of workforce participation or exit which may subsequently impact health (Dannefer 2003). In terms of continuity theory, working for longer in good quality, purposeful roles might lead to prevention of health deterioration. This might be particularly applicable to areas such as depression (Atchley, 1989) and cognitive function, whereby continuation of employment might be protective of decline (Bonsang et al. 2012; Rohwedder & Willis 2009; Adam et al. 2007). This relates to aforementioned ideas of compression of morbidity, whereby extending working lives leads to longer periods of healthy living and shorter periods of morbidity thereafter (Fries 2002). Similarly, continuation of poor quality roles might lead to higher levels of depression, especially if the continuation is due to lack of affordability of retirement (Ekerdt, 1983). Physical health might also be damaged by continuation of work which is overly physically demanding or heavily manual in nature. Again, Atchley (1982) writes of the honeymoon period of retirement, in which the initial relief from workforce stressors leads to a temporary increase in health and wellbeing. This might be particularly true for individuals retiring from poor quality work, which is reflected in the aforementioned studies showing improvements in levels of health following retirement from stressful roles and low grade employment. However, ideas surrounding theories of successful ageing suggest the positive effect of retirement is only sustainable if the individual conducts activities which maintain purposefulness and meaning (Scherger et al. 2011; Bound & Waidman 2007; Drentea 2002; Herzog & House 1991). When this is not the case, a feeling of letdown with retirement might eventually lead to negative effects on health (Atchley, 1989).

73 Overview and conclusions This chapter has provided an overview of research previously carried out on the effects of working in later-life on health. A careful systematic review procedure provided a total of 58 studies which examined the effect of later-life employment on a variety of health and wellbeing outcomes. Additionally, studies with a focus on effects of retirement were included on the basis of containing a strong reference group of later-life workers. Results were broken down by type, with a key focus placed on depression and self-rated health due to the interests of this particular study on these outcomes, as well the good availability of studies which had examined these outcomes. The overall conclusion of the literature review was the non-comparability of the heterogeneity of findings. Table 2.2 provided key characteristics of each of the studies included, and the between-study differences were immediately obvious. Study samples were highly heterogeneous. Many studies were inclusive of individuals aged fifty and over, rather than of those who had reached, or were nearing, traditional retirement ages. Alongside this, many studies compared workers to a heterogeneous group of retirees, often inclusive of those who had retired early, or were temporarily absent from the workforce, for a variety of reasons. While some studies used methods to account for this, few specifically analysed respondents who were of retirement age or above. Definitions of work and retirement also varied, subsequently leading to sample differences. The type of work carried out differed, with some studies focusing on stress in the workplace and demanding employment types, in comparison to better quality work. A variety of data sources from different countries was used. This might be particularly worthy of note where differences in normal retirement ages and policies exist, leading to further differences in sample and retirement types. Some studies focused on specific populations, such as male only, or people from specific types of employment, such as those with high levels of workplace stress or conflicts with family life. The methods used to analyse data differed too. While the majority of studies were longitudinal, a large number were cross-sectional in nature, leading to differences in the interpretation of results. A broad set of analytical techniques were used, and only a few employed methods, such as IV analysis, from which causal mechanisms can be

74 73 assumed. Each study accounted for a different list of covariates, although the majority of analyses accounted for age, gender and indicators of socio-economic position, and the list of different outcomes examined across studies was extensive. The majority of studies included in this review have too great a variation in their samples, methods and definitions of employment and retirement to be reasonably compared. However, when drawing main conclusions from previous research, particular strength should be added to the findings of those studies which employed methodology specifically to deal with issues of heterogeneity and selection. These were discussed in greater detail in the relevant outcome sections. Studies which used IV techniques to examine effects on depression of retirement, in direct contrast to later-life working, found no significant effect existed. Of the studies which used similar techniques to look at self-rated health, two found a significant detrimental effect of retirement in contrast to working (Calvo et al. 2013; Behnke 2010), and one found a temporary beneficial effect (Coe & Zamarro 2011). However, both the lack of significant results produced in the analyses of depression as well as the varied results in the analyses of self-rated health can again be attributed to study design, as in each instance, either work quality was not accounted for or those retiring prior to reaching SPA were included in the analyses. The findings of the literature review further demonstrate the rationale for the research questions provided in Chapter 1. Heterogeneous associations between laterlife employment and health suggest an investigation into working, as opposed to retirement, is needed with a focus placed on issues of selection bias, using a method such a propensity score matching explicitly designed to deal with this. Studies which have used methods specifically to deal with selection bias when considering effects of work, in relation to retirement, have consistently found varied results, many of which were non-significant. However, the lack of stratification by characteristics of work within these studies lends strength to further examination of health differences between workers and retirees on the basis of work quality, as outlined in the second research question of interest to this study. The following chapter further examines the literature detailed in the review provided here by means of a meta-analysis. Extractable data taken from relevant studies is used to examine whether calculation of pooled effects is possible, as well as to

75 74 provide a measure of the extent of between-study heterogeneity present. The findings will highlight the importance of stratifying older workers on the basis of their characteristics when researching effects of employment in later-life.

76 75 Chapter 3: Meta-analysis of studies concerning the effects of later-life employment on health outcomes Chapter 2 provided a systematic review of literature on the topic of the health effects of later-life employment. Comprehensive details of the procedure of this review are discussed there. Results from the systematic review of literature concluded that key findings of studies previously carried out on the topic provide contradictory evidence with regards to whether or not later-life employment is either beneficial or detrimental to health. An overall finding was hard to establish, due to the fact that almost all of the studies included a wide range of factors deemed likely to affect the effects of employment or retirement on older populations, such as previous workplace stressors, and quality of employment. This chapter will detail the procedure and results of two meta-analyses of data taken from a selection of the studies on the outcomes of depressive symptomatology and self-rated health. The chapter will begin with the underlying rationale for, and principles of, meta-analysis. A description of how the relevant data from the studies included in the review was extracted is presented, alongside details of the strict inclusion and exclusion criteria used from the outset. The pooled effects of the analysis will be examined, along with indicators of heterogeneity present between the studies involved. Following this, a subgroup analysis will be carried out in order to determine the effect of different study-level covariates on any heterogeneity present. Finally, a discussion of evidence of, and reasons for, potential biases affecting the pooled study estimates will be presented. It is important to note that the number of studies possible for use in this analysis is very small, and this issue will be taken into consideration throughout the chapter, with the best possible means of overcoming any problems likely to be encountered in such instances put into practice. The chapter will finish with a discussion of the overall results, and consideration of how this analysis paves the way for further research to be carried out on the topic within this thesis.

77 Calculating the pooled estimate in meta-analysis A meta-analysis aims to synthesise the results of several comparable studies into one pooled effect in order to establish an estimate of the overall finding taken from already-published studies. This pooled effect can be calculated as part of either a fixed- or random-effects model, dependent on the assumptions underlying the analysis (Borenstein et al. 2007). A fixed-effects meta-analysis assumes that one true effect exists across all studies, and that the only variation existent is that within each study, around its own calculated effect. A random-effects model assumes that variation exists between study populations in addition to within studies, and that, due to this, it is not possible to assume one true effect exists across all (Field 2003). These issues of heterogeneity and choice of model will be discussed in greater detail further in this chapter. Derivation of the pooled estimate in a meta-analysis is not simply a case of calculating the average across all studies to be included. Instead, studies are weighted according to the inverse of the variance of their effect coefficient. This weight is then taken into account when calculating the pooled effect (Sánchez-Meca & Marín-Martínez 1998). In a fixed-effects model, where the effect size is assumed to be the same across studies, the value of the weight attached to the ith study is dependent only on the inverse of the variance within that study, and so larger studies will be assigned greater amounts of weight and smaller studies can often be assigned only a tiny value in comparison. In a random-effects model, weights are more evenly distributed across studies, as the pooled estimate is assumed to be the average of a range of effects (Borenstein 2007). Equations denoting study weights, pooled estimates, and the variance of the pooled estimates are provided in Appendix 1. Values of pooled estimates along with individual study estimates and their assigned weighted value will be presented in the final meta-analysis section.

78 Measuring heterogeneity in meta-analysis As suggested in the chapter introduction, one of the key aims of a meta-analysis, following the establishment of a weighted average effect of exposure, is to ascertain whether or not heterogeneity is present within the analysis, as well as then attempting to explain the sources of its presence and why the various studies included display such differing results (Berman & Parker 2002). Pereira et al. (2010) describe this assessment of heterogeneity in meta-analysis as the evaluation of the between-study variance, and also state that this variation may reflect biases arising from a variety of sources, including publication bias as well as the bias occurring from differences in study designs, populations and outcome measures. The latter sources of bias are often referred to as methodological heterogeneity (Higgins & Thompson 2002). As selection effects are likely to be particularly prominent in research concerning health and employment in later-life, this methodological heterogeneity will be the main focus of this study, and issues associated with it will be covered in detail further on in the chapter. The same paper notes that statistical heterogeneity, which exists when the variation between studies is greater than would be expected to occur by chance, is important to take into consideration, and may affect the choice of final model to be used for the analysis, as well as conclusions drawn from it. Heterogeneity is a central issue of meta-analysis, and its examination should be expected in any such use of the method. Differences in study designs, populations, methodologies and locations can only be expected to render different outcomes (Higgins 2008). However, the question must be asked as to how much heterogeneity is acceptable, and how it should be handled. Higgins 2008 paper proposes any level of heterogeneity present within meta-analysis is acceptable, providing that the method and data used for completion of the meta-analysis are correct and some attempt is made at explaining its source. This study will use subgroup analysis to enable a deeper understanding of any substantial heterogeneity present. As the studies used here are observational in design, a high level of heterogeneity is to already be expected. Observational studies likely include a wide variety of participants, and both their personal characteristics and placement in either the exposure or control group is out of the researcher s control. Within a meta-analysis

79 78 of observational studies, in which each uses different participants from differently sampled datasets and pays consideration to a variety of differing confounding factors, we can subsequently expect large population differences to exist Heterogeneity and the choice of final model: fixed-effects versus random-effects Both Higgins (2008) and Huedo-Medina et al. (2006) state the level of heterogeneity present in a meta-analysis may lead to the researcher choosing to employ one particular model over another, which is an important decision as both are interpreted differently. In general, statistical software will provide an output for meta-analysis of both fixed-effects and random-effects models. Field (2003), states that these two different model types differ in their underlying assumptions, as well as in their calculations of the effect size and significance. Fixed-effects models assume all studies included in the analysis are taken from a population with a fixed overall effect size. In other words, within this particular research, a fixed-effects model would assume that the effect of later-life employment is the same for everyone, and the explanation for the differing effect sizes presented in the different studies is due to sampling error alone. When a random-effects model is employed, the researcher is assuming that not only has variation arisen from sampling error, but also that error exists within the different populations sampled by each study. Additionally, a random-effects model assumes that each of these separate populations differ among a so-called super-population, encompassing them all (Field 2003). Therefore, statistically speaking, the key difference between the two models can be explained by the level of error incorporated in each: in a fixed-effects model there is only one source (sampling error), and in a random-effects model there are two (sampling and population errors). The choice of model in a meta-analysis can be partly based on the inferences the researcher wishes to make. Many researchers opt for a fixed-effects approach to meta-analysis, due to its simplicity (Field 2003). However, where a considerable amount of variation appears to exist between studies, a random-effects model is more than likely the correct model to employ (Spector and Thompson 1991). One can speculate as to whether there is much purpose in assuming a random-effects

80 79 model if the key aim of the analysis is to obtain an average effect size. This is because the very nature of the selection of a random-effects model over a fixedeffects model implies that the researcher does not believe there to be an overall average effect, as there is, in fact, a set of various effects from various subpopulations of the overall population. In instances such as these, the researcher may choose to place a key focus on testing the heterogeneity between the different study effect sizes, explaining the sources of such variation, rather than placing a focus on the pooled estimates alone (Field 2003). Choice of the correct model within meta-analysis is crucial, and assumption of invalid estimation of results may occur if an incorrect decision is made. The statistic on which this decision usually depends is Cochran s Q homogeneity test (Cochran, 1954), an indicator of whether heterogeneity is present within a meta-analysis or not, with its p-value denoting its significance. A significant p-value, suggesting significant variability exists between the studies included in the analysis, indicates that a random-effects model should be preferred (Field 2003). Alternatively, if the p- value of Q is not significant, then the homogeneity hypothesis is retained, and a fixed-effects model should be employed. Berman (2002), states that the value of the Q statistic is often considered conservative, and suggests the significance level should be set to 0.10, rather than the typical However, as the rest of the work presented throughout this thesis uses the 5% significance level, this is the value which will also be considered here Choice of statistic to measure heterogeneity Q, τ 2 and I 2 The previous paragraph refers to Cochran s Q statistic, which is normally provided in any computational meta-analysis output. The Q statistic is, essentially, dichotomous in nature, telling the researcher only if heterogeneity is or is not present within the meta-analysis (Pereira et al. 2010). Again, if Q is not significant, a fixedeffects model is applied to the data, implying that a single, true effect, in actuality, exists across all studies. Appendix 1 provides formulae for the calculation of the Q statistic. Despite its essentially dichotomous nature, Berman and Parker (2002) note that a larger value of Q suggests significant heterogeneity exists between studies. However,

81 80 there are several limitations to the Q statistic, some of which are of particular importance to this specific study. Pereira et al. (2010) note that one should be cautious of assuming heterogeneity does or does not exist on the strength of the Q test alone, as it lacks power to detect true heterogeneity in instances where the number of studies included in meta-analysis is small, such as in this instance, and has too great a power for detecting heterogeneity when a very high number of studies are included in analysis. Details of how this study will overcome this issue are discussed further in this section. Another limitation of Q is that it can only be used to measure existence of heterogeneity in a meta-analysis, not the extent of it (Huedo-Medina et al. 2006; Higgins & Thompson 2002). In addition to this, the specific value of Q is meaningless and cannot be compared across meta-analyses which include differing numbers of studies. This is because Q follows a chi-square distribution with k 1 degrees of freedom, and analyses with differences in k are therefore not comparable. Again, these points are of relevance here as this study will use subgroup analysis to compare levels of heterogeneity after stratification of the studies included in the main analyses. One possible means of dealing with the shortcomings of Q is to examine the quantity of the between-study variance statistic, τ 2, estimated for random-effects models in meta-analysis (Higgins 2008). τ 2 is the estimate of the standard deviation of underlying effects across studies, and Equation 9 in Appendix 1 demonstrates calculation of the statistic. However, τ 2, again, is not a particularly reliable measure of variance where only a small number of studies are included in the analysis as its value is also dependent on the number of studies analysed (Higgins & Thompson 2002). In order to overcome the impracticalities of working with Q and τ 2 in meta-analysis, Higgins and Thompson (2002) suggest two additional statistics of beneficial use in reporting meta-analyses, H 2 and I 2, the latter of which will be of particular importance to this particular analysis. A brief overview of the H 2 statistic is provided in Appendix 1, with its formula demonstrated in Equation 10. In addition to H 2, Higgins & Thompson (2002) propose the use of the I 2 statistic in order to overcome the difficulties of using Q and τ 2 in meta-analysis. Unlike the H 2

82 81 statistic, which informs us only if heterogeneity is present, I 2 can be used to assess both the existence and extent of heterogeneity, using the precisions of the individual study estimates. Essentially, it is a measure of the extent of the variability in effect sizes which is not simply down to random error within studies (Pereira et al. 2010). A value of I 2 above 0 demonstrates that this variation does exist between study populations, and where the value of I 2 is high, a random-effects model should be chosen in preference to a fixed-effects model. The value of the statistic can be interpreted as a percentage of total variability which is due to statistical heterogeneity. I 2 is particularly useful as, unlike Q, its power is not dependent on the number of studies included in meta-analysis, and can be used successfully with analyses containing only a few studies (Higgins et al. 2003). The I 2 statistic is also not dependent on the effect type analysed, and so the I 2 of differently conducted meta-analyses can be compared simultaneously. The formula for the I 2 statistic is denoted in Equation 12 in Appendix 1. An additional benefit to use of the I 2 statistic is its ease of interpretation. If I 2 = 0, the researcher can assume that all variability in the effect size estimates is due to sampling error, and a fixed-effects model can be assumed. If I 2 = 50, one can assume that half of the effect size variability is due to sampling error, and the other half is due to true heterogeneity between studies. Higgins & Thompson (2002) propose a cautious guideline for assessing the extent of heterogeneity by use of the I 2 statistic, with a value of around 25% or less depicting low heterogeneity, around 50% depicting moderate heterogeneity, and around 75% depicting high levels of heterogeneity. The value of I 2 is related to that of the aforementioned τ 2, in that the two should increase in accordance with one another (Higgins & Thompson, 2002). The results of the meta-analysis carried out here, on the effects of later-life employment on depression and self-rated health, will use the I 2 statistic to complement both the values of Q and τ 2.

83 Rationale and data for present meta-analysis In its simplest form, the meta-analysis method consists of statistically integrating the findings of several comparable studies in order to obtain an average effect of treatment or exposure on participants across several pieces of research. This then allows a summary of the totality of evidence relating to a particular issue (Spector and Thompson 1991). By providing a weighted average of the effects of each study, an accurate idea of the true effect of interest can be obtained (Field 2003). Berman and Parker (2002) state that meta-analysis should aim to complete at least one of two goals: the summary of data available on the topic of interest, and the explanation of between-study variation (heterogeneity). Huedo-Medina et al. (2006) take this proposal one step further by stating a third possible aim of meta-analysis: the identification of variables or study characteristics that may explain heterogeneity where it is present. This particular analysis will examine each of these three aspects of the method in turn. This analysis will provide pooled estimates of the effects of later-life employment on two health outcomes: depressive symptomatology and selfrated health. Existence, as well as the extent, of any heterogeneity will be considered and, finally, explanations for why this possible heterogeneity exists will be put forward through use of subgroup analysis. Spector and Thompson (1991) stress the importance of establishing a set of reliable inclusion criteria prior to undertaking any meta-analysis. As shall be outlined further in the chapter, the validity of any meta-analysis is heavily influenced by the presence of bias, one possible source of which can be differences in basic study quality and design. In order to combat such possible sources of bias before analysis, a strict set of inclusion criteria was established prior to data extraction, ensuring each study to be included in the final meta-analyses consisted of relevant methodology, participants, research questions, comparisons and outcomes. The studies initially considered for use in the meta-analysis are the same as those examined in the systematic review of literature on the topic of the health effects of later life employment. Details of these studies can be found in Table 2.2 in Chapter 2. Just as is the case for a systematic review, a meta-analysis should only be undertaken if the researcher is certain that all possible studies on the topic have been researched and their suitability for the analysis ascertained (Berman & Parker 2002). Details of the

84 83 initial literature search of the topic and the ways in which studies were obtained can also be found in Chapter Inclusion criteria Due to the fact that, for the meta-analysis, data was to be extracted from the suitable articles, a stricter set of inclusion criteria had to be adhered to than that which was proposed for the systematic review in Chapter 2. The purpose of inclusion criteria is to eliminate, wherever possible, biases that may arise from conducting synthesis with widely varying study characteristics. In this instance, studies were deemed suitable for inclusion in the analysis on the basis that they i) were able to answer the research question, ii) contained extractable data, iii) were methodologically similar and iv) had comparable exposure and control groups. The remainder of this section examines these aspects in greater detail. A table providing specific details on each of the studies excluded from the analysis can be found in Appendix 2. Ability to answer the research question The studies included in the analysis needed to be able to answer the general research question regarding the health effects of later-life employment. Therefore, data was needed which could allow for a comparison of a health effect between those working in later life and those who were retired. Studies excluded on this basis were those which had a focus on a particular aspect of employment (n=6), such as those working in stressful environments only, or those working in a specific type of job, as inclusion here would be likely to create biases when compared to samples of people taken from all employment and workplace types. Additionally, studies were deemed suitable for answering the research question on the basis that they examined effects of working on either depression or self-rated health. A total of 15 studies were excluded on the basis that they examined other health outcomes. Extractable data Data needed to be presented in a format which was useable to the researcher. Where types of data were not strictly comparable, the data needed to be in a format which could be transformed to a comparable measure (in this analysis Cohen s d). This included mean health scores, odds ratios, regression coefficients with standard errors,

85 84 or a correlation coefficient. In each instance, this data needed to be available for both an employed and retired group. A total of 13 studies were eliminated on the basis of non-extractable data. Methodological similarity The methods used in each of the studies needed to be similar enough to be deemed comparable. In this instance, studies were included in the analysis on the basis that they were observational analyses of large-scale datasets. All studies presented in the systematic review were eligible for inclusion in the meta-analysis on this basis. Comparable exposure and control groups It was important that definitions of the employed and retired groups of individuals remained as similar as possible, in order to ensure that the synthesis of results was based on a similar population. There were 10 studies which did not provide comparable groups, and were subsequently excluded from the analysis. Further details of these definitions can be found in section 2 of the chapter. Table 3.1 lists the studies chosen for inclusion in the final analyses, along with a brief description of the ways in which inclusion criteria were met. It is important to note that where studies included extractable data regarding both depression and selfrated health, studies have been included in both analyses separately. Looking at Table 3.1 shows the final analysis will be comprised of eleven effects from eight different studies in the case of depression, and eight effects from seven different studies in the case of self-rated health.

86 85 Table 3.1: Studies included in meta-analysis and details of how inclusion criteria are met. Author Year Study design Data used Participant age/ Gender/ Country Sample size/ number of observations Scale of measurement Depressive symptomatology Reitzes 1996 Cross-sectional National Institute of Aging Data years 48% male USA Gall 23 Longitudinal years 1997 Retirement Research Male only Study USA Butterworth Cross-sectional years 2006 National Survey of Male only Mental Health and Australia Well-being Butterworth Cross-sectional years 2006 National Survey of Female only Mental Health and Australia Well-being Villamil Cross-sectional years 2006 British Psychiatric Male only Morbidity Survey UK Villamil Cross-sectional years 2006 British Psychiatric Female only Morbidity Survey UK Dave Longitudinal years 2006 Health and Male & Retirement Study female USA Wahrendorf Cross-sectional years 2008 GAZEL 73% male France Coursolle Cross-sectional years 2010 Wisconsin Male & Longitudinal Study female Employed = 427 CES-D Retired = 291 Employed = 224 SCL-90 Retired = 117 Total = 1928 CIDI (group-specific n not reported) Total = 2261 CIDI (group-specific n not reported) Total = 1,875 CIS-R (group-specific n not reported) Total = 2,253 CIS-R (group-specific n not reported) Employed = CES-D 44,799 Retired = 31,411 Employed = 2,817 CES-D Retired = 11,660 Employed = 667 CES-D Retired = 1360 Effect type Effect size (SE or 95% CI) Cohen s d (0.076) Cohen s d (0.137) Odds ratio 1.23 ( ) Odds ratio 1.03 ( ) Odds ratio 3.2 ( ) Odds ratio 1.4 ( ) Cohen s d (0.007) Pearson s r Regression coef The study by Gall et al (1997) was not included in the sections concerning depression and selfrated health in the systematic review. This is due to the fact that the means difference used here was presented only in a table of descriptive information, and was not included in any of the adjusted analyses from which study conclusions were drawn.

87 86 Table 3.1: Studies included in meta-analysis and details of how inclusion criteria are met. Author Year Study design Data used McMunn Cross-sectional 2009 English Longitudinal Study of Ageing McMunn Cross-sectional 2009 English Longitudinal Study of Ageing Self-rated health Kremer Cross-sectional 1985 Retirement Planning Society of Haifa Data Crowley Cross-sectional 1986 NLS of the Labour Market Experience of Mature Men Gall Longitudinal 1997 Retirement Research Study Van Solinge Longitudinal 2007 NIDI Data Zucchelli Longitudinal 2007 Household, Income and Labour Dynamics in Australia Zucchelli 24 Longitudinal 2007 Household, Income and Labour Dynamics in Australia Dave Longitudinal 2006 Health and Retirement 65+ years Male only UK 60+ years Female only UK years Male only Israel years Male only USA years Male only USA 55+ years 58% male Netherlands 50+ years Male only Australia 50+ years Female only Australia years Male & Employed = 185 Retired = 1,757 CES-D Odds ratio 2.14 ( ) Employed = 347 CES-D Odds ratio Retired = 2, ( ) Employed = 310 Likert scale Cohen s d Retired = (0.081) Employed = 699 Likert scale Odds ratio Retired = ( ) Employed = 224 Likert scale Cohen s d Retired = (0.118) Employed = 778 Likert scale Odds ratio Retired = ( ) Employed = 533 Likert scale Odds ratio Retired = ( ) Employed = 407 Likert scale Odds ratio Retired = ( ) Participant age/ Gender/ Sample size/ number of observations Scale of measurement Effect type Effect size (SE or 95% CI) Country USA (0.31) Employed = Dichotomous Odds ratio 44,799 question The study by Zucchelli et al (2007) was not included in the section concerning self-rated health in the systematic review. Again, the means difference used here was presented only in a table of descriptive information, and was not included in any of the adjusted analyses from which study conclusions were drawn.

88 87 Table 3.1: Studies included in meta-analysis and details of how inclusion criteria are met. Author Year Study design Data used Participant age/ Gender/ Country Sample size/ number of observations Scale of measurement Study female Retired = 31,411 (good/ USA poor) Roberts Longitudinal years Employed = 1,561 Likert scale 2011 Whitehall II Cohort 67% male Retired = 470 Study UK Effect type Effect size (SE or 95% CI) ( ) Odds ratio 0.77 ( ) As discussed beforehand, a key limitation of a meta-analysis of observational studies is that it is almost impossible to eliminate differences between study participants. In this instance, the exposure and control groups are strictly defined, as are the outcome measures, but beyond that, it is difficult to ensure that any other similarities remain between studies. It is apparent, especially by looking at the third column of Table 3.1, that the age range, gender specifics and countries of studies are widely varied: the only means by which the studies are similar is in that they all provide an effect for either a retired or older employed group of people on either depression or self-rated health. It can also be noted that every study within the proposed meta-analysis uses a different study dataset. Further details of these can be found in Chapter 2 on the systematic review of this literature. In the studies of depression, different scales to measure the outcome have been used. However, most are scaled so that lower scores are indicative of improved mental health, and where this was not originally the case, scales have been reversed to make it so. Studies using a questionnaire that can be used to measure other mental disorders alongside depression have either been reduced in the original study in order to measure depression alone, or results were broken down by disorder type in the study results themselves. For self-rated health, a Likert-scale questionnaire was used in each case, with participants asked to rate their health as either very good, good, fair, poor or very poor. Again, all scales have been transformed so that their direction is the same and, as with the scales of depression, lower scores here are indicative of better self-rated health.

89 Meta-analysis and repeated measures design Most of the studies included in the meta-analysis are cross-sectional in design, or use raw data taken from only one point in time in order to compare means or create odds ratios. However, the studies by Gall (1997), Dave (2006), Van Solinge (2006) and Roberts (2011) provided only the crude data for means analysis or odds ratio calculation across all waves of their research. This can be problematic in that with cross-sectional analysis, variation may exist only between the scores of different participants within each group being studied, whereas with longitudinal analysis, variation in scores for the same people over different points in time are included in one overall effect. Methodological differences such as these, when analysed simultaneously, are a potential source of heterogeneity in meta-analyses, which must be dealt with and explained. There are various perspectives on how best to deal with repeated measures data in meta-analyses. Ideally, individual case data would be provided for each wave so that an effect, such as Cohen s d, could be easily calculated from two specific time-points. However, this is rarely the case, and the only option often left for consideration is to decide whether to approximately estimate or ignore such covariance (Ishak et al. 2007). Where estimates of this covariance are approximated, evidence suggests that miscalculation of within-subject covariance has little influence on the effect estimate (Ishak et al. 2007). The alternative to approximation of the covariance is ignorance of it, by which its value is assumed to be 0, and the exposure and control groups to be compared are then assumed to be completely independent of one another. However, this assumes that the longitudinal nature of the data and the within-subject variance is not important to the study outcomes, and so caution should be taken when interpreting these results (Ishak et al. 2008). In this particular instance, covariance will be ignored. As the coefficient of interest is the effect of later-life employment on health outcomes, and two distinct sets of data exist for both the retired and employed groups under study, it will be assumed that no significant correlations exist between the groups, despite the fact that, where someone has retired during the study period, some individuals will be included in both groups. Becker (2000) notes there are two potential options when calculating values of Cohen s d in studies with longitudinal designs. Either the original standard deviations of the employed and retired group means can be used for pooling, as

90 89 occurs normally with the calculation of d, or a paired t-test can be used, as this takes into account the correlations between the employed and retired groups. In this analysis, as the potential covariance is to be ignored, the original standard deviations of the means will be used in calculation of d, in accordance with calculation of d for the cross-sectional studies.

91 Definition of the variables analysed in the meta-analysis Although meta-analyses are susceptible to various sources of bias, some of which shall be covered in greater detail further on in the chapter, one possible means of ensuring bias is reduced to some extent is to ensure that the outcome effects to be included in the analysis, as well as populations and exposures, are all strictly comparable. The five sections below indicate how these aspects of each study to be used here are comparable to one another. Such precision in defining these areas of interest is essential. It is already known that large differences exist between the basic characteristics of the studies, which is likely to create high levels of heterogeneity, and so a strong similarity between characteristics chosen as inclusion criteria for the analysis, such as comparable depression measurement scales, is key to reducing these levels of heterogeneity as much as is possible. Also, a key focus of a metaanalysis is the provision of pooled effects, which is only justifiable if the individual study effects can be fairly combined. An outlier with regards to study aspects such as measurement scales is likely to produce immediately biased analysis results Depression The different scales used to measure symptoms of depression have previously been listed in Table 3.1. The most common scale used in the studies was the Center for Epidemiologic Studies Depression (CES-D) scale, by Reitzes (1996) Gall (1997), Dave (2006) and McMunn (2009). The CES-D scale is comprised of questions relating to symptoms of depression, with a lower score representative of less depressive symptomatology than a higher score. Each of the remaining studies in the meta-analysis of depression used different scales. Butterworth et al. (2006) use the Composite International Diagnostic Interview (CIDI) to measure depressive symptomatology. This scale can be used to measure various mental disorders (Butterworth 2006), but the focus of the analysis from which data has been extracted here is labelled depression in the study, and therefore is suitable for comparison to other depressive scales. The original effect type taken from this study is an odds ratio, so scale differences can be ignored, as a cut-off point to distinguish the difference between either showing symptoms of depression or not can be assumed to be the same as those used in the other scales.

92 91 Finally, Villamil et al. (2006) use the Revised Clinical Interview Schedule (CIS-R) in order to determine symptoms of depression in participants. Again, the CIS-R can be used to diagnose symptoms of various common mental disorders, but the authors distinguish depression from any other disorder in their work and label this clearly in the presentation of their data. Odds ratios already calculated by the authors were extracted for the meta-analysis, and again we can assume a cut-off point for showing signs of depression to be in line with the cut-off points of other scales measuring the same outcome Self-rated health In six of the seven studies of self-rated health, a Likert-scale questionnaire was used in order to measure participants outcomes. Typically, there were five points on the scale, with participants asked to rate their health as either very good, good, fair, poor or very poor. However, the study by Crowley (1986) used only a four-point question, with the equivalent of very poor omitted, and the study by Van Solinge (2007) used just a three-point scale with the responses (very) good, not good/not bad and (very) bad. These differences are not problematic in this meta-analysis, as where odds ratios are constructed initially, the categories provided still allow for a measure of sub-optimal health to be created in line with the studies using the fivepoint scale. When Cohen s d is calculated, the same continuous structure can be assumed and effects can then indicate either a negative or positive effect on health. The only study which used a different method to discriminate between poor and good self-rated health was that by Dave et al., which simply asked participants whether they considered themselves to be in good or poor health, the results of which were included in the study as dichotomous variables. Again, odds ratios here look at sub-optimal health, and so the proportion of people who had responded with poor health was used as the group of interest, and an odds ratio was thereby constructed for them Employment In all studies included in the meta-analysis, a clear definition of both employment and retirement was essential, as these comprised the two key groups of interest.

93 92 Where studies did not provide a clear-cut definition of employment, for example, where the employed group of participants included those working very few hours per year or semi-retired people, exclusion from the meta-analysis was considered appropriate and the studies were removed (further information provided in section and Appendix 2). Where studies provided several definitions of employment status, data was taken only from those classed as employed and retired, so the comparisons remained strictly between the two. Coursolle s study (2010) provided data for a group of partially retired individuals in addition to the typical two groups of employed and retired. With regards to other studies in the analysis, some include partially retired as a part of the workforce, while others do not. None of the studies include these as part of the fully retired group. In this instance, the presentation of the available data does not make it possible to include the partially retired people in the employed group, as the original effect is a regression coefficient showing the effect of retirement on health, and no coefficient exists for the group of people classed as employed. Therefore, the partially retired individuals are hereby excluded from the meta-analysis Retirement As with the definition of employment, it was essential to ensure that the definition of retirement was strict. As the studies included in this analysis are observational in nature, a high number of differences are likely to exist between them, and in order to justify synthesis it is important to ensure the reference groups are also wholly comparable to one another. Definition of retirement can be more complex than that of employment. Butterworth et al. (2006) define their retired population as those who were not engaged in either full-time or part-time employment, and exclude those who were classed as simply older and unemployed. The distinction between older unemployed and retired is essential to this analysis, as the two populations are likely to be very different from one another and the results for both cannot be justifiably synthesised. Any studies which presented comparable groups as employed and unemployed, including retired were omitted from the present analysis.

94 93 The majority of studies clearly labelled participants as either retired or not, and some included results for the same participants both before and following retirement. One exception to this is the study by McMunn et al. (2009), whereby participants are classed as either in paid work or not in paid work. However, McMunn s study only provides data for those who are of SPA and above, and so it can be assumed that a very large majority of those in the not in paid work category will be retired. In addition to this, there is a benefit to McMunn s study in that early retired are excluded from the analysis. This is an important factor to be considered as will be discussed further in the chapter. One final point to be made is that the original crude data provided in Crowley s (1986) study covered information for people who had retired for a number of reasons, including ill health. In order to retain a sample as similar as possible to the others included in this meta-analysis, data was only taken, and combined, for those who had retired mandatorily or voluntarily at the SPA. Those who had retired for health reasons, were discouraged or took voluntary early retirement were excluded, as these groups of people are likely to share characteristics different in nature to those of the retired group in general. Again, this idea will be discussed in further detail later in the chapter Study populations One of the initial tasks of a meta-analysis is to ensure the studies to be synthesised are as similar as possible. The majority of meta-analyses use randomised controlled trials from clinical settings to examine health-related data. A greater discussion of the benefits of randomisation in studies is provided in Chapter 4. As this particular analysis uses observational studies, there are immediately likely to be large differences between the different populations used for each of the studies. Tables 2.2 in Chapter 2, and 3.1 earlier in this chapter, have already outlined some important sample characteristics within each study. In many ways, some study similarities do exist. Each study included has a comparable outcome measure, a clearly defined employed and retired group of participants, and the samples are always inclusive of older employees. However, there are also large discrepancies. Some studies include males only, some include participants only of retirement age and above, every study uses a different dataset, and the total of 20 effects on both depression and self-rated health arise from a total of 6 countries. It is also notable that the sample sizes within

95 94 each of the studies vary. Additionally, studies vary by the extent to which they distinguish between types of retirement, with 10 of the effects from 7 studies providing a retirement effect inclusive of respondents from all types of retirement (Roberts et al. 2011; Wahrendorf et al. 2008; Van Solinge 2007; Zucchelli et al. 2007; Dave et al. 2006; Gall et al. 1997; Reitzes et al. 1996). This already observable presence of heterogeneity justifies future retention of random-effects meta-analysis models, as well as subsequent subgroup analysis in order to further examine the extent and impact of these sources of heterogeneity. Such aspects of the analyses will be considered further in the chapter.

96 Meta-analysis of Cohen s d from different effect types Conversion of all effects to Cohen s d Meta-Analysis can be carried out on various effect types. Table 3.1 provided information on the effect types taken from each study. Six studies concerning depression, and five concerning self-rated health, provide odds ratios, or data from which odds ratios could be calculated. Three studies on depression, and two for selfrated health, provided unadjusted data from which means differences (Cohen s d) could be calculated directly. Additionally, one value of Pearson s r and one standardized regression coefficient is provided for the analysis of depression. In instances such as these, it is possible to transform the effects which are not already a means difference to means differences which can then be used for synthesis with others. This is usually a value of Cohen s d (Chinn 2000). It is important, wherever possible, to transform and include as many available effects as possible. Omission of effects due to their type differing from the majority available, or analysis only of the same effect types separately, can result in a loss of information and validity that could otherwise be captured and a meta-analysis, wherever possible, should include as many studies as possible simultaneously (Whitehead et al. 1999). No further data manipulation was required for the already calculated values of Cohen s d, and so the initial step of the analysis was to calculate Cohen s d for each of the other effect types. Equation 13 in Appendix 1C denotes straightforward calculation of Cohen s d using unadjusted means data, and Equation 14 denotes calculation of its standard error. Theoretically, calculation of the standard error of transformed values of d can also be obtained by this method, provided the relevant information is extractable from each study. Transformation of odds ratios to values of Cohen s d is more challenging. Odds ratios predict the likelihood of an event, such as the occurrence of depression or suboptimal health. Cohen s d, on the other hand, measures the size of the impact of an intervention, such as later-life employment, on an outcome such as depression. In addition to this there are different assumptions underlying the construction of each effect type, and to combine them in one analysis simultaneously would override these and create invalid inferences. The key difference in assumptions lies in the type

97 96 of distribution that each follows: Cohen s d is assumed to follow a Normal distribution, whereas the distribution of odds ratio data is assumed to be skewed to the right of a Normal distribution. This is because the value of an odds ratio cannot be negative and so the left-hand side of its distribution is restricted (Bland & Altman 2000). However, Chinn (2000) notes that the transformation of dichotomous data into continuous data can be justified if the scale used to initially measure the variable of interest was continuous in the first place. This was the case here: depressive symptoms were measured using continuous scales in all relevant studies, and cut-off points on these scales were selected in order to dichotomise the data (either showing symptoms of depression or not) 25. If we assume that the distribution of dichotomous variables, as provided for odds ratios, is Normal, as we assume when modelling continuous data, and that the only real difference in the distributions lies in the tails, as observed in the distributions assumed by construction of odds ratios, transformation of these dichotomous variables into ones which can be analysed on a continuous scale can be justified. Stockute et al. (2006) write that the only difference between the variance of a Normal distribution, N (μ, σ 2 ), and the variance of the logistic distribution is by the scaling of π 2 /3. Therefore, an odds ratio can be converted to a continuous effect size (d) by log transformation of the original effect and division of this value by π/, which is The formula, alongside the formula for the standard error of the transformed d, is shown in Equations 16 and 17 in Appendix 1. A transformation was made of a regression coefficient (Coursolle et al. 2010) to Cohen s d, which is possible with use of the t-statistic. Equations 18, 19 and 20, in Appendix 1 demonstrate calculation of the t-statistic, alongside two possible subsequent calculations of d from this. Finally, the Pearson s r coefficient was transformed into a value of d using the formula presented in Equation 21 in Appendix 1 (Rosenthal 1991). 25 One study of self-rated health used an already dichotomised variable of suboptimal-versusoptimal self-rated health (Dave et al. 2006). Again, however, the scale from which this was originally derived was a five-point Likert scale question.

98 Results of the meta-analysis of combined effects: later-life employment on depression The following analyses discuss tables of results and key statistics, and results are visualised with forest plots. The forest plots presented here are for random-effects models only. Although the actual weighted values of the study estimates take account of heterogeneity between studies and are therefore relatively evenly distributed in a random-effects meta-analysis, the boxes representing the individual study effects on the plots are sized according to their calculated within-study random error alone, as provided by the fixed-effects models. As a result, although it might offer some insight into study variation, this representation of study size can be ignored when assuming a random effects model is the model of use to the study. When interpreting values of d, it is useful to refer to Cohen s effect size threshold, which suggests an effect of 0.2 is considered small, 0.5 is considered moderate, and 0.8 is considered large (Cohen, 1988). This scale will be referred to when interpreting the results of the meta-analyses here. Following the conversion of all relevant effect types to Cohen s d, along with calculation of their standard errors, an overall meta-analysis was run in order to examine pooled effects and heterogeneity. Table 3.2 shows the results of the overall meta-analysis model of later-life employment and depression. Studies here are inclusive of effects from both mixed-gender studies and single-gender studies (effects for both genders taken from the same study are presented separately where possible). Although, where two effects are taken from the same study for males and females, the study design is identical, the different gender samples are completely independent of one another, and variations in their sample sizes, and therefore standard errors of effects will exist. On this basis, it is deemed appropriate for these effects to be included in the meta-analyses as their own study estimates. In some instances, the original data provided for the analysis of depression focused on the effects of retirement, rather than employment. As effects of later-life employment are of primary interest to us here, values have been reversed on their mean of 0, so that effects which were originally positive are now negative, and viceversa. Negative scores are indicative of better mental health.

99 98 Table 3.2: Meta-analysis of combined effects (d) later-life employment on depression final model. Q p(q) τ 2 I 2 Fixed-effects pooled estimate (95% CI) Random-effects pooled estimate (95% CI) % (p = 0.000)*** (-0.217, ) (p = 0.103) (-0.376, 0.034) ***p<0.001 Firstly, the pooled estimates shown in Table 3.2 can be considered. The fixed-effects estimate here is and the random-effects estimate is Both effects are small, according to Cohen s aforementioned effect size threshold. Also, both values are negative, suggesting an association between later-life employment and improved mental health when compared to retirement. The fixed-effects estimate is significant at the 5% level, with a p-value of The random-effects estimate is not significant, with a p-value of This suggests the differences between the amalgamated study populations are too great to establish any true effects of employment in later-life on levels of depression. Table 3.3: Study estimates and calculated weight values Study Weights (%) Study 95% confidence interval Fixed Random estimate Lower Upper Reitzes (0.7) 8.85 (9.7) Gall (0.2) 7.94 (8.7) Butterworth (0.4) 8.56 (9.4) Butterworth (0.8) 8.91 (9.8) Villamil (0.07) 5.91 (6.5) Villamil (0.2) 7.51 (8.2) Dave (85.6) 9.32 (10.2) Wahrendorf (9.5) 9.29 (10.2) Coursolle (1.9) 9.14 (10.0) McMunn (0.1) 7.35 (8.0) McMunn (0.4) 8.57 (9.4) The pooled results can also be examined by means of the forest plot provided in Figure 3.1. Here, it is notable that only four of the eleven study estimate confidence intervals contain the pooled effect. This suggests a high level of heterogeneity is present once again, as might be expected. The confidence interval for the pooled estimate itself is reasonably small, although appears somewhat average in

100 99 comparison to the others in the analysis. Its width can be calculated as 0.410, whereas the confidence interval of the fixed-effects estimate is far smaller at The forest plot in Figure 3.1 also shows that in the fixed-effects analysis, a great deal of weight is placed on the study by Dave et al. (2006), with only tiny amounts, in comparison, attributed to the other studies. Table 3.3 shows the original study estimates along with the weight value given to each. In the fixed-effects model, 85.6% of the total weight value is attributed to Dave s study, with the remaining 14.4% distributed among the total of ten remaining studies. It is possible that this is creating bias within the results, and this shall be discussed in further detail shortly. The study by Wahrendorf et al. (2008) holds 9.5% of the remaining weight. In the random-effects model, although the within-study error still contributes to the visual weight on the forest plot, the actual weighted estimates in Table 3.3 show the weights to be much more evenly distributed, with the study by Dave et al. accounting for only 10.2% of the total weight. The remaining total weight value is distributed evenly among the remaining studies. This occurs due to the fact that, although the within-study error is accounted for in a random-effects model, measured by the inverse of the variance which is very small in the case of Dave s study, the between-study error is also taken into account which therefore balances out the large within-study differences. In order to establish whether results are strongly biased by the inclusion of such a highly weighted study, a sensitivity analysis can be carried out whereby the analysis is re-run with the problematic study removed. A sensitivity analysis was carried out here with the omission of the study by Dave et al. The direction of the pooled effect remained the same and the magnitude was not greatly affected, and so the results presented in Table 3.2 can be assumed to be robust to the exclusion of the study. A full report of the sensitivity analysis is provided in Appendix 3.

101 100 Figure 3.1 Forest plot of the effects of later-life employment on depression, from analysis of combined effects (d) (male and female combined and separate). Reitzes 1996 Gall 1997 Butterworth 2006 Butterworth 2006 Villamil 2006 Villamil 2006 Dave 2008 Wahrendorf 2008 Coursolle 2009 McMunn 2009 McMunn 2009 Combined Cohen's d The Q statistic for the final meta-analysis is very large at , and the estimate of the standard deviation of the pooled studies τ 2 is However, both Q and τ 2 have low power when only a small number of studies are included in a meta-analysis and is dependent on the number of studies involved making comparisons with forthcoming subgroup analyses difficult. Instead, we can calculate the I 2 statistic on 10 degrees of freedom, which is both more reliable and comparable. The value of I 2 shows 98.5% of the heterogeneity present in the analysis is due to variation beyond that which exists within studies only, and so a random-effects model should be assumed. The very high value of I 2 here will be examined and discussed further, by means of subgroup analyses, later in the chapter.

102 Results of the meta-analysis of combined effects: later-life employment on self-rated health Just as for depression, an analysis of all effects transformed to d was run for selfrated health. The odds ratios included in the preliminary analyses indicated an association between suboptimal self-rated health and continued employment, rather than retirement. Although these odds ratios were calculated as the likelihood of suffering sub-optimal health as opposed to good health, the calculations were made from Likert scale type questions, just as were used to discern the mean self-rated health scores in the studies from which Cohen s d was originally calculated. As with the analysis of depression, whereby an odds ratio suggestive of increased likelihood of depression was created by use of a cut-off point on a continuous scale, the same occurred for scores of self-rated health, satisfying the assumption that the odds ratios to be converted originally belonged to a continuous scale of measurement rather than one which is completely dichotomous 26. Again, all original effects (Table 3.1) were transformed to values of d using equations 13 to 21 in Appendix 1. Just as in the preliminary analyses on self-rated health, the effects of later-life employment on the outcome are to be examined here, in comparison to those of retirement. The final meta-analysis of combined effects consisted of eight studies, including both mixed-gender and single-gender groups (Kremer, Crowley and Gall presented effects for males only, and Zucchelli presented a separate effect for males and females). Table 3.4 shows the key statistics from this final meta-analysis, with the individual study values of d and corresponding weight values shown in Table 3.5. The forest plot of the random-effects meta-analysis model is shown in Figure Again, one study (Dave et al. 2006) provided an already dichotomous version of self-rated health (suboptimal-versus-optimal). However, this was originally derived from the same 5-point Likert scale survey question as the other measures of self-rated health.

103 102 Table 3.4: Meta-analysis of combined effects (d) later-life employment on suboptimal self-rated health. Q p(q) τ 2 I 2 Fixed-effects pooled estimate (95% CI) Random-effects pooled estimate (95% CI) % (p = 0.000)*** (-0.811, ) (p = 0.303) (-0.696, 0.216) *** p<0.001 The pooled estimates of the analysis show the effect from the fixed-effects model is , which borders the moderate-to-large threshold size, and from the random effects model is , which is again classed as reasonably small. In both cases, the effect is negative, which suggests later-life employment has a beneficial effect on self-rated health. This is in line with the findings of the analysis on depression, also linking later-life employment with improved mental health. The fixed-effects estimate is significant at the 5% level, with a p-value of 0.000, whereas the randomeffects estimate is again non-significant, with a p-value of Table 3.5: Study estimates and calculated weight values Study Weights (%) Study estimate 95% confidence interval Fixed Random Lower Upper Kremer (4.7) 2.33 (12.6) Crowley (3.7) 2.32 (12.6) Gall (2.2) 2.29 (12.4) Van Solinge (5.2) 2.33 (12.6) Zucchelli (1.6) 2.26 (12.3) Zucchelli (1.1) 2.22 (12.1) Dave (77.2) 2.36 (12.8) Roberts (4.2) 2.32 (12.6) Figure 3.2 shows the forest plot of the results. Three of the eight study confidence intervals contain the pooled estimate, and the confidence interval of the study estimate by Kremer at al. (1985) finishes precisely on the pooled estimate s vertical line (both values are -0.24). In other words, half of the studies within this metaanalysis have results that are in accordance with the pooled estimate. It is also noticeable that two of the studies have findings which contradict the trend of effects, with study estimates from Gall et al. (1997) and Van Solinge et al. (2007) showing positive values, associating later-life employment and suboptimal health, with values of 0.73 and 0.17, respectively.

104 103 The forest plot in Figure 3.2 again shows a higher proportion of the total weight value of all studies is attributed to the study by Dave et al. (2006). Table 3.5 shows this proportion to be 77.2% in the fixed effects model (although only the highest by a value of 0.2 in the random-effects model, with 12.8%). Again, this is due to the small standard error of the study effect (0.02), the inverse of which calculates the weight. This, in turn, is due to the fact that the study by Dave et al. has a far greater number of observations within it (Table 3.1). Again, a sensitivity analysis carried out with the omission of this study shows results to be robust (Appendix 3). Figure 3.2 Forest plot of the effects of later-life employment on suboptimal self-rated health, from analysis of combined effects (d) (male and female combined and separate). Kremer 1985 Crowley 1986 Gall 1997 Dave 2008 Van Solinge 2007 Zucchelli 2007 Zucchelli 2007 Roberts 2010 Combined Cohen's d With only half of the study estimate confidence intervals containing the pooled estimate, a high level of heterogeneity is expected within this analysis. The value of Q is This is smaller than that of the analysis of depression, but as Q is related to the number of studies k included in the analysis this is the anticipated result. The value of τ 2 is 0.423, but again, its value is of little use to the meta-analysis

105 104 presented here due to the small number of studies included within it. From Q, we can calculate I 2 on 7 degrees of freedom, which indicates 98.7% of the heterogeneity present exists between the studies, rather than in only the random error within them. Again, this issue will be further dealt with in the subgroup analysis section of the chapter.

106 Subgroup analysis Identification of potential sources of heterogeneity Tables 3.2 and 3.4 showed the main results of meta-analyses for depression and selfrated health, respectively. The tables show the amount of heterogeneity in both final analyses is very high, with an I 2 statistic of 98.5% in the analysis of later-life employment on depression, and 98.7% in the analysis of later-life employment on self-rated health. As stated previously, the I 2 statistic shows the proportion of heterogeneity existing between the studies included in the analysis, and so we can conclude from these values that the studies selected in each analysis are very different from one another and unsuitable for producing reliable pooled effects. Prior to the running of any analyses, it was ensured, to as great an extent as possible, that the definitions of various study aspects were comparable across studies (Tables 2.2 and 3.1). However, the high levels of heterogeneity present in the meta-analyses suggest heterogeneity exists somewhere beyond the identifiable level, and so must be further examined. Not only will this enable some understanding of the impact of certain between-study discrepancies, it will additionally allow some insight into how further research on the topic can be conducted and controlled. The key step in assessing the possible sources of heterogeneity in the meta-analyses was to run a series of subgroup analyses. Subgroup analysis consists of dividing the studies included in the original meta-analysis according to some study-level characteristic which is evident in some studies and not in others. Subsequently, findings can be examined in comparison of one group of studies to the other. The subgroup analyses here will be carried out according to the areas of potential bias outlined below in points 1 to The use of effects calculated from crude data provided in the studies, in comparison to study-adjusted effects (for example, author-calculated odds ratios). 2. The inclusion of individuals of early retirement age in both the retired and later-life employed groups, in comparison to studies of individuals only of SPA and above.

107 The use of longitudinal data, whereby the same participants are included in both the employed and retired groups, and the use of cross-sectional data, in which the two groups are entirely distinct. 4. The effects of employment on men and women individually, alongside an analysis of both single- and mixed-gender groups in general. 5. Two country-specific subgroup analyses will be run: one for studies conducted using UK data and one for those which used data generated in the USA. The following section of the chapter provides a description of how each of these areas may affect the analysis here Crude data versus study-provided estimates When extracting data from studies with regards to analysing the effect of later-life employment on depression, seven of the study outcomes were calculated from crude data. Studies by Reitzes, Gall and Dave contained mean scores for both retired and employed groups of individuals from which a standardised means difference was calculated; the study by McMunn contained proportional data from which odds ratios were calculated, and the study by Wahrendorf contained a value of Pearson s r, showing only the correlation between two variables, from which Cohen s d was calculated. In the analysis of self-rated health, all of the studies provided crude data only. It is important to note that where crude data has been used, no controlling takes place for any other effects on the outcome. This is of importance as there are several individual characteristics which might be particularly likely to affect results when researching effects of employment in older age on health, such as age, type of work and gender. Previous studies on the topic have found that physical health deteriorates, in general, with age, particularly after the age of 60 (Jokela 2010; Banks et al. 2006; Case and Deaton 2005) and, conversely, depression becomes less likely in older age (Schieman et al. 2002; Charles et al. 2001; Feinson 1985), and so lack of adjustment on this basis is likely to be particularly problematic. In the analysis of depression, some studies provided author-derived estimates. Coursolle s study included an unstandardised regression coefficient along with its standard error, from which a t statistic was derived in order to finally calculate Cohen s d for inclusion in analysis, and studies by Butterworth and Villamil

108 107 provided odds ratios. In each of these studies, estimates were derived from models already controlling for age, and the studies by Coursolle and Villamil also controlled for a number of other factors, including occupational type and income. It can be hypothesised that the effect size will be greater where the effect size, d, has been derived from crude data, as age and other potentially confounding variables have not been already controlled for here and are likely to be unobserved sources of bias affecting the results. In other words, as well as any effect of later-life employment or retirement alone, effects of age are also likely to be leading the estimates. It can also be hypothesised that the value of I 2 will be smaller in the analysis of adjusted estimates, as the controls for will reduce some of the betweenstudy population heterogeneity. As only crude data was provided for the studies on self-rated health, comparisons of data type will only be made using the studies of depression Inclusion versus exclusion of early retirees Another consideration is whether or not a study has included early retirees in its analysis. Early retirement is substantially different from statutory retirement. The reasons for early retirement are likely to create bias in the final estimates of the retirement group. While optional early retirement may be associated with positive experiences and suitable wealth to enjoy retirement, when early retirement has been taken on the grounds of poor health, there is likely to be a negative effect on health scores. Jokela (2010) writes of the health selection involved in later-life employment, whereby those who continue to work do so because they are able to, and that suboptimal health is a leading cause of early retirement. The study by Buxton et al. (2005) finds that prevalence of mental health disorders is significantly higher in early retirees than in those who continue to work. Early retirement in UK studies is classed as below age 65 for males and 60 for women 27. As mentioned previously, it is from age 60 onwards when physical health begins to decline the most rapidly. We would therefore expect to see a stronger negative effect on self- 27 Retirement ages in countries of the other studies included in the analyses are 65 in Australia and the Netherlands, but age 60 in France and age 67 for men and 65 for women in Israel. The USA has no statutory retirement age, and a tendency to work until an older age than in the UK (the normal retirement age is from 65 to 67 years, depending on year of birth in relation to pension entitlement).

109 108 rated health where a study has included only people of state pension age or older. If a study has included early retirees, it has also included younger workers in its employed group (this age is never younger than 50, except in the study by Roberts (2011) where the first wave of data (pre-retirement) consists of individuals aged 38-60). Subsequently, the inclusion of these younger older workers who, if theories of health selection hold true, are fit and able to work, may lessen the size of the effect on self-rated health where studies have not controlled for age (all of the studies analysing self-rated health). As referred to beforehand, older age is generally associated with lower levels of depression. Where studies include early retirees a smaller positive effect on depression scores might be observed than in studies where they are not due to age effects alone. However, associations are also likely to exist between reasons for early retirement and depression, for example, workforce exit due to poor health associated with a greater number of depression symptoms Repeated measures versus cross-sectional design The third subgroup analysis focuses on whether the design of the studies included in the original meta-analyses was longitudinal or cross sectional. The definition of the study design here is based only on that which is applicable to the type of data extracted for the meta-analysis. For example, studies with adjusted analyses which could not answer the direct question of whether later-life working affected health might have also provided descriptive sample statistics which showed extractable means differences and their standard errors for working and retired groups. In the analysis of self-rated health, four of the eight studies use a repeated measures design: Gall (1997) provide data on the same sample one year prior to retirement and then six to seven years afterwards; Dave (2006) provide one average score across all seven waves of the Health and Retirement Study used in their work, and Roberts (2011) provide mean scores for individuals both prior to retirement and then following. Here, unlike the aforementioned study by Gall, in which all participants in the employed group became retired, ensuring the exact same sample was used for both the employed and retired groups, in the study by Roberts et al. only 30.1% of the original employed sample retired, and so the two groups are not identical, with some participants included in both and others not. Finally, Van Solinge (2007)

110 109 provides data for participants in 1995, before any had retired, and again for 2001, when all participants had retired. In the analysis of depression, only two studies use longitudinal data, and so subgroup analysis for this outcome is not possible. As previously discussed, meta-analyses of studies containing repeated measures data can be analysed alongside those containing cross-sectional data. Research by Ishak et al. (2007) suggests calculation of within-subject covariance is often not possible from the extractable data provided by studies, and where it is, its calculation offers little bearing on the magnitude of pooled effects. In a standard meta-analysis, it is permissible to ignore within-subject correlations and treat the two groups under study as completely separate from one another. It can, however, be hypothesised that where longitudinal studies are analysed separately from those using cross-sectional data, an increase in levels of heterogeneity may be observed, as there is likely to be more between-subject variation in the error of the calculated effect sizes (Yee & Niemeier 1996) Single-gender versus mixed-gender groups Another possible source of heterogeneity in the meta-analyses arises from gender differences captured within the original study estimates. The original meta-analyses used a model inclusive of separate male and female samples, alongside mixedgender samples. One hypothesis might be that the inclusion of single and mixed gender groups in the same analysis contributes to the high levels of heterogeneity demonstrated in the main analyses. A second hypothesis might be that higher values of I 2 are likely to be observed among mixed gender groups analysed together when compared to synthesised studies of just men or just women. Pooled effect sizes might also differ according to stratified gender analyses, as evidence exists that the relationship between employment, retirement and ageing is experienced differently by males and females. Looking at the individual original study effects from the meta-analysis on depression (Table 3.3), substantial differences in effect size exist between male and female groups even when taken from the same study. Butterworth et al. (2006) present an odds ratio of 1.23 for retired men to suffer depression, but a smaller odds ratio of 1.03 for women. Likewise, the study by Villiamil et al. (2006) states retired men are 3.2 times more likely to suffer depression than men continuing to work, whilst

111 110 retired women are only 1.4 times more likely. Finally, the study by McMunn et al. (2009) finds retired men are 2.14 times more likely to suffer depression than those working beyond SPA, while women are 2.11 time more likely. The study by Zucchelli et al. (2007) on self-rated health finds the odds ratio for men employed in later life to suffer sub-optimal health to be 0.31 (3.2 times more likely that those who are retired, if the inverse of the value is calculated), and for women to be 0.38 (2.6 times more likely to suffer sub-optimal health). In each case, for both depression and suboptimal self-rated health, it can be concluded, from this particular set of evidence, that effects are greater among men than women, and this is likely to be reflected again in the subgroup analysis. Due to small study numbers, it was only possible to carry out subgroup analysis to compare mixed gender groups to male-only groups in the case of self-rated health. The analysis of depression compares mixed groups to male- and female-only groups. As examination of the original study estimates alone indicates differences in effects for gender, it can be hypothesised that there will be less heterogeneity among analyses of single-gender samples alone, as those incorporating both males and females within one estimate are already contributing a heterogeneous effect Country-specific studies (UK and USA) Table 3.1 shows the studies of depression to use data from a total of four countries, and those of self-rated health use data from a total of five. It is possible that such variation in the nationalities of the data is a significant contributor to the high levels of heterogeneity present in the final analyses of both health outcomes. There are country-specific differences regarding several aspects of later-life employment, including the aforementioned variation in statutory retirement age, pension schemes, preparation for retirement and workplace characteristics. Due to small study numbers, the subgroup analysis here will focus only on those studies from the UK and the USA, with the hypothesis that an analysis of study effects derived from data originating from one country only will show lower levels of heterogeneity than when effects from all countries are analysed simultaneously. It can also be hypothesised that, due to the differences underlying retirement plans and policies in the two countries, there is likely to be a difference in the pooled effect sizes of studies using UK data in comparison to those using data from the USA, with the USA possibly showing more detrimental effects on health due to the lack of a statutory retirement

112 111 age and likelihood of increased pressure to continue working for financial reasons into later years. At the time of publication of the studies included in the analyses here, the UK state pension age, when state pensions may first be received, is 65 for men and 60 for women. In the US, however, no such set retirement age exists, and the normal retirement, when individual is eligible for receipt of state benefits, occurs between the ages of 62 and 70 for both men and women Results of the subgroup analysis of later-life employment on depression Prior to running subgroup analysis, study-level dummy variables were created in accordance with the hypothesised areas of potential heterogeneity outlined in the previous section. Table 4.1 in Appendix 4 shows the binary responses for each variable for each study. These may be referred in order to ascertain the specific studies included in each subgroup analysis. After creating the seven dummy variables, a total of ten analyses were run, with an analysis for the binary response 1 run in comparison to the response 0 for each variable. Table 3.6 shows the results of the subgroup analyses of later-life employment on depression. Table 3.6: Key statistics from final meta-analysis and subgroup analyses later-life employment on depression Analysis Q p(q) τ 2 I 2 Fixed-effects Results of overall meta-analysis (Table 2.3) Overall metaanalysis Results from subgroup analyses pooled estimate (95% CI) % *** (-0.217, ) Crude data % ** Study estimates % (-0.226, ) (-0.026, 0.111) Early retirees % ** (-0.225, ) Random-effects pooled estimate (95% CI) (-0.376, 0.034) (-0.504, 0.085) (-0.257, 0.102) No early % (-0.473, 0.073)

113 112 Table 3.6: Key statistics from final meta-analysis and subgroup analyses later-life employment on depression Analysis Q p(q) τ 2 I 2 Fixed-effects pooled estimate (95% CI) Random-effects pooled estimate (95% retirees (-0.034, 0.107) (-0.402, 0.163) Male groups % ** Female groups % Mixed-gender groups Single-gender groups (-0.473, ) (-0.211, 0.027) % ** (-0.217, ) % ** (-0.309, ) UK studies % ** (-0.529, ) USA studies % 0.255** (0.241, 0.268) CI) ** (-0.711, ) (-0.409, 0.069) (-0.284, 0.410) ** (-0.504, ) Fixed-effects model (no heterogeneity present) (-0.425, 0.144) * signifies estimate is significant at the 10% level. ** signifies estimate is significant at the 5% level Subgroup analysis of later-life employment and depression: heterogeneity The rationale behind a subgroup analysis is an investigation of potential sources of heterogeneity and so, for that reason, attention will be paid to this first. The first row of Table 3.6 shows the results of the initial meta-analysis, the value of I 2 for which was very high at 98.5%. The column of I 2 values shows many of the subgroup analyses provide an estimate lower than this, suggesting the analyses may have proved useful in explaining some sources of heterogeneity. The initial subgroup analysis was performed on studies providing crude data only in comparison to those which had calculated their own adjusted effects, such as odds ratios or regression coefficients. The analysis of crude data effects was comprised of six studies, and adjusted effects of five. Here, it was expected that heterogeneity would be reduced where study estimates had been provided as the authors had controlled for potentially confounding effects which might impact health outcomes. The key factor controlled for here was age, which is already known to have a

114 113 substantial effect on health outcomes in that, regardless of employment status, health deteriorates with age (Banks et al. 2006). This is particularly true beyond the age of 60 (Case & Deaton 2005). Where studies were using mixed-gender groups, gender was also controlled for, and again it is known that continuation of employment and retirement are experienced differently for men and women. Table 3.6 shows the hypothesis regarding heterogeneity to be correct. The value of Q is reduced from in the analysis of crude data only, to in the analysis of adjusted estimates only. The value of τ 2 is also much smaller in the analysis of adjusted estimates, at 0.028, compared to in the analysis of crude data. Again, however, these statistics have low power when an analysis is comprised of relatively few studies, and are dependent on the number of studies included in the analyses. Subsequently, calculation of the I 2 statistic is necessary to enable a fair comparison between the two. This demonstrates 99.2% of the variance present in the analysis of crude data is due to heterogeneity between studies rather than chance, but that this value is 77.6% in the analysis of adjusted estimates. The proportion of unexplained heterogeneity is 21.6% lower among studies which had calculated adjusted effects than those providing only crude data for effects to be created for the purpose of this meta-analysis alone. This is probably due to the controlling for confounding effects, such as age and gender, which are likely to bias results if their potential impacts are ignored. The second subgroup analysis focused on the inclusion of early retirees in the study populations, in comparison to their exclusion. Studies including early retirees were those who used people under the age of 65 for males and 60 for females in both or either of their employed and retired comparison groups. A total of seven studies included early retirees, compared to four which did not. Again, the hypothesis regarding early retirees is proved correct according to Table 3.6. The value of Q for the studies in which early retirees are included is high, at , whereas its value is in the analysis of studies including only those of retirement age and over. The value of τ 2 is also smaller in the analysis of older populations only at 0.073, compared to However, once more, it is the value of I 2 which is of particular interest, and in this instance its value shows 99.0% of the heterogeneity is at the between-studies level when early retirees are included and

115 % when they are not. Although these levels are both still high, the reduction in the analysis excluding early retirees is likely to be because removal of the younger population has removed some of the heterogeneity arising from a larger age range. Studies including early retirees use participants aged 50 and over and those which do not use only participants who, at minimum, are at least ten years older. Additionally, early retirees are likely to be a heterogeneous group alone, as reasons for early workforce exit are likely to vary and subsequently likely to have differing impacts on health outcomes. The next variable to be considered in the subgroup analysis was gender. The original study effects presented in Table 3.3 have already suggested differences exist in the impact of employment in later-life on the basis of gender. Table 3.6 shows the Q statistic in the analyses of males and females individually is very small in comparison to any other subgroup analysis, at for men, and for women. The value of τ 2 is also relatively small at in the male analysis and in the female analysis. Once again, we can calculate a value of I 2, which tells us that 73.6% of the variance is due to between-study differences in the analysis of males only and 61.6% in that of females only. These are the two smallest values of I 2 in all subgroup analyses of depression suggesting that, if heterogeneity is at its lowest levels when men and women are analysed separately, gender differences are large and future analysis should place a focus on the different roles of later-life employment and retirement for the two groups. In addition to the analyses of solely males and females, two more analyses were run. One was comprised of study effects from mixed-gender groups only, and the other of effects from single-gender groups only. The hypothesis here was that heterogeneity would be smaller in the single-gender groups. It is worth noting, however, that a large proportion of any heterogeneity explained away here will possibly be due, in part, to the fact that there are 7 different effects from only 4 studies and so a smaller number of different populations from which effects have been extracted. Table 3.6 shows this hypothesis to be correct. The Q value for the analysis of single-gender groups is , compared to in the analysis of mixed-gender groups. The value of τ 2 is also much smaller, at 0.048, compared to Again, the I 2 statistic can be calculated, which shows 77.3% (a moderate-to-large amount) of the variance in the analysis of single-gender groups exists between study populations, compared

116 115 to 99.5% in the analysis of mixed-gender groups. In fact, the level of heterogeneity in the mixed-gender groups is the highest of all subgroup analyses run, again indicative of the large impact gender bears. The final subgroup analysis was by country, with studies using data from the UK and the US analysed separately. Here, the suggested level of heterogeneity present among the UK studies is very low, with a value of Q at The value of I 2 is 0, suggesting the studies are completely homogenous. Subsequently, the fixed-effects model can be retained in this instance, and it can be assumed that one common effect is shared across all study populations. The forest plot in Figure 3.3 shows the confidence intervals of all study estimates contain the pooled estimate. The original hypothesis predicted that analysing studies by country would reduce levels of heterogeneity, as there is a possibility of country-specific effects on health outcomes due to differences in work and retirement policies. The lower level of heterogeneity among UK studies may be due to set pension ages leading to greater similarities between those who work and those who retire, unlike the US where retirement ages are not as well defined. Not only are studies conducted in the US likely to contain a wider variety of retirement ages, there are also likely to be a higher number of people remaining in the workforce longer due to the fact that they cannot afford to retire at the same age as is possible in the UK, leading to further differences in impacts of work on health. However, it must also be noted that the four UK effects presented here are taken from only two separate studies, so again a smaller number of different study populations is also likely to be contributing to the strong comparability of results.

117 116 Figure 3.3 Fixed-effects forest plot of analysis of UK studies only. The confidence intervals of all study effects contain the pooled estimate, and homogeneity is therefore assumed (later-life employment on depression). Villamil 2006 Villamil 2006 McMunn 2009 McMunn 2009 Combined Cohen's d Subgroup analysis of later-life employment and depression: pooled estimates Although the key focus of this subgroup analysis is the explanation of heterogeneity, it is also interesting to look at the pooled estimates taken from the analyses. Table 3.6 shows both the fixed- and random-effects pooled estimates from the subgroup analyses, along with the pooled effects from the original meta-analysis, which was in the fixed-effects model, and in the random-effects model. Returning to the first analysis, of effects drawn from crude data in comparison to study-adjusted effects, the pooled effect from the crude data studies is in the fixed-effects model, and in the random-effects model. Both of these estimates are larger than in the original analysis. This would be expected here, due to the fact that the crude data estimates do not control for anything else and so are likely to be

118 117 larger due to uncontrolled-for bias. In the model of adjusted estimates only, the fixed-effect estimate actually changes direction and becomes positive, with a value of However, looking at a forest plot of the results (Figure 3.4), shows this is due to the presence of only one positive study estimate which happens to bear the greatest amount of weight (Coursolle et al., with 55.9% of the total weight value attributed to all studies. Further details of this and all specific study estimates can be found in Table 4.3 in Appendix 4).The fixed-effects pooled estimate for the adjusted effects is also not significant at the 95% level, as in both the original model and the crude data model, and its reduction in magnitude suggests the effects of employment are smaller after controlling for confounding factors such as age. Figure 3.4 Fixed-effects forest plot of analysis of studies of adjusted estimates only, demonstrating the large heterogeneity present due to the small number of studies holding very different effects and weighted values (later-life employment on depression). Butterworth 2006 Butterworth 2006 Villamil 2006 Villamil 2006 Coursolle 2009 Combined Cohen's d The second subgroup analysis focuses on inclusion versus exclusion of early retirees, and provides a pooled estimate of in the fixed-effects model and in the

119 118 random-effects model where early retirees are included. These values are and , respectively, where early retirees are not included. Again, the estimates where early retirees are included are larger than in the original analysis, suggesting early retirees add strength to the detrimental effects of retirement on depression. This would be expected once again, as early retirement is often associated with poorer wellbeing, due to many early retirees exiting the workforce on the basis of poor health. The finding suggests that prolonging employment among this group might lead to smaller detrimental effects on depression. The fixed-effects estimate turns positive where only an older population is included in studies. Again, this is due to the small number of studies included in the analysis, and the heavy weight attributed to the studies by Reitzes and Coursolle (a total of 80.6% of the entire weight added to all studies). Table 4.3 in Appendix 4 presents further information on specific study estimates and attribution of weight. The random-effects model provides a pooledestimate of when only an older population is included. It would be expected that this estimate would be smaller than that of the early retirees as, again, depressive symptomatology traditionally decreases with age. Table 3.6 shows larger pooled effects exist among men than women, with pooled estimates of and in the fixed- and random-effects models respectively for men, and and , respectively, for women. According to Cohen s aforementioned threshold for effect size, we can conclude the effects for males to be moderate in size, whereas those for females are only small, suggesting a greater impact of later-life employment on mental health for men. The pooled estimates of the single-gender only analysis complement this finding, with the fixed-effects model providing an estimate of and the random-effects model providing one of , both of which are significant at the 5% level. It would be expected that these results should lie somewhere in the middle of those for males and females analysed separately, as they do.

120 119 Figure 3.5 Fixed-effects forest plot of analysis of studies not including early retirees, demonstrating the large heterogeneity present due to the small number of studies holding very different effects and weighted values (later-life employment on depression). Reitzes 1996 Coursolle 2009 McMunn 2009 McMunn 2009 Combined Cohen's d Finally, studies conducted with UK data were analysed in comparison to those using data from the USA. The hypothesis here was that levels of depression would be lower among the later-life employed population in the UK, due to that fact that those working beyond retirement age are more likely to be doing so through choice than in America, where no set statutory retirement age exists and therefore a higher number of people may feel pressured to work for as long as they can before entering retirement. This hypothesis was correct, with a pooled estimate of in the UK, which is a moderately sized coefficient according to Cohen s scale, and suggesting an association between later-life employment and lower levels of depression. However, both the fixed- and random-effects estimates from the USA data models are positive (0.255 and 0.141, respectively), suggesting a link between later-life employment and higher levels of depression, albeit non-significant. Results taken

121 120 from the US studies might be particularly useful, in that retirement policies focus on keeping individuals in the workforce for longer, the effects of which are central to this research Results of the subgroup analysis of later-life employment on self-rated health Similar to the subgroup analysis on depressive symptomatology, study-level binary variables of specific study characteristics were created. There are some differences here from the subgroup analysis of depression. Firstly, all studies used in the analysis of self-rated health provided crude data rather than study-adjusted effects. The implications of this will be considered further in the chapter. However, for the time being, it means that no subgroup analysis was carried out according to data type. In addition to this, only one study (Zucchelli et al. 2007) provided a female-only effect, and so no comparison of male and female analyses could be carried out. Instead, gender is considered with regards to male-only group analysis in comparison to one of mixed-gender groups, and whether analyses of males and females separately provide differing levels of heterogeneity to those analyses of them simultaneously. Finally, only one study provided an effect from UK data, and so no analysis was run on UK-only studies, although three studies used US datasets, and so there is an analysis of US data compared to other those from all other countries. Table A4.2 in Appendix 4 shows the study-level binary variables that were created for the subgroup analysis of self-rated health and again can be used as a reference in order to ascertain which studies were included in each of the analyses. Table 3.7 shows the results of a subgroup analysis of studies regarding later-life employment and self-rated health. The original meta-analysis had a very high level of between-study heterogeneity present, at 98.7%.

122 121 Table 3.7: Key statistics from final meta-analysis and subgroup analyses later-life employment on suboptimal self-rated health. Analysis Q p(q) τ 2 I 2 Fixed-effects pooled estimate (95% CI) Random-effects pooled estimate (95% CI) Results of overall meta-analysis (Table 2.4) Overall metaanalysis % *** (-0.811, ) (-0.696, 0.216) Results from subgroup analyses Early retirees % ** (-0.859, ) (-0.834, 0.363) No early retirees % ** (-0.391, ) * (-0.549, 0.037) Longitudinal design % ** (-0.867, ) (-0.087, 0.744) Cross-sectional design % ** (-0.457, ) Male groups % ** (-0.239, ) Mixed-sex % ** groups (-0.907, ) ** (-0.637, ) (-0.624, 0.412) (-1.127, 0.484) Single-sex groups % ** (-0.269, ) (-0.634, 0.259) USA studies % ** (-0.933, ) (-1.125, 0.878) Non-USA studies % ** (-0.280, ) ** (-0.588, ) * signifies estimate is significant at the 10% level. ** signifies estimate is significant at the 5% level Subgroup analysis of later-life employment and self-rated health: heterogeneity The first subgroup analysis was of studies inclusive of early retirees in comparison to those not. Table 3.7 shows that, as was hypothesised, heterogeneity is reduced when early retirees are excluded from the analysis, with the I 2 statistic showing the proportion of heterogeneity between study populations to be 83.2%, rather than 98.9% when they are included. The value of τ 2 is also much smaller where early

123 122 retirees are excluded, with a value of 0.037, compared to when they are included. Again, this is likely, in part, to be due to the fact that removal of studies including early retirees results in the analysis of a sample in which participants are more similar. All are above the SPA and the reason for retirement, in the majority of cases, has been older age alone, rather than factors such as illness or disability, as is likely to be the case among the studies in which early retirees are included. The second subgroup analysis concerning self-rated health was of study design, with an analysis of studies using a longitudinal design, and therefore the same participants in both the retired and employed groups, in contrast to cross-sectional studies, in which the participants in the two groups are completely different. The hypothesis here was that heterogeneity would be lower in the group of studies using repeated measures design, as the between-participant variation would be reduced. However, the results in Table 3.7 show this is not the case, with the heterogeneity lower in the analysis of cross-sectional studies (I 2 =77.9%, compared to 99.3%). This could be due to the fact that longitudinal data has, potentially, two sources of heterogeneity: both within and between subject, while cross-sectional data has only the between subject heterogeneity present. This finding must be treated with caution, however, due to the small number of studies in the analysis. Also, all of the cross-sectional studies are gender-specific in nature, which may also be contributing to the reduction in heterogeneity present. The third subgroup analysis of self-rated health studies concerned gender. As stated beforehand, there was only one female-specific effect among these studies (Zucchelli 2007), and so, in this instance, segregated male and female analyses could not be run. However, an analysis was run on the male-only studies, and Table 3.7 shows heterogeneity here is reduced, although only slightly. The value of I 2 is 96.2%, compared to 98.7% in the main analysis. When an analysis was run on single-gender studies in comparison to mixed-gender studies, heterogeneity is reduced again, although this difference, too, is small. The proportion of between-study heterogeneity where only single-gender study effects are analysed is 95.2%, and where mixed-gender study effects are analysed, 99.3%. Although both values of I 2 here are very large, it is notable that the prevalence of heterogeneity among the mixed-gender only studies is actually slightly larger than when both groups are

124 123 included in the main analysis, suggesting there is some reduction in heterogeneity on the basis of gender. Finally, an analysis was run on studies conducted using data from the USA in comparison to all other countries included in the main analysis. The hypothesis here was that heterogeneity would be lower in the analysis of American studies only when compared to all countries combined, as experiences of later-life employment in America are likely to be more similar among Americans alone. However, the level of heterogeneity among studies using data from the USA is actually higher than in the main analysis, with an I 2 statistic of 99.3%, compared to 98.7%. If we look at the I 2 statistic for all countries analysed simultaneously without the inclusion of the USA studies, the value is 90.9%, which suggests the experiences of later-life employment in the USA alone is more varied than in all other countries combined Subgroup analysis of later-life employment and self-rated health: pooled estimates Table 3.7 also shows the pooled estimates from the subgroup analyses. The first analysis was of studies including early retirees, in comparison to studies excluding early retirees. While, on the one hand, early retirement might be associated with poorer health and later-life employment might be carried out on the basis of physical ability to do so, older age is also associated with a decline in health regardless of employment status, particularly beyond the age of 60 (Jokela et al. 2010; Banks et al. 2006; Case & Deaton 2005). Here, the random-effects pooled estimate for the analysis including early retirees is , compared to in the analysis excluding early retirees. There is not much difference between the two, perhaps showing the two contrasting effects at play. However, looking at a forest plot of the analysis of studies inclusive of younger retirees (Figure 3.5), it is noticeable that the fixed-effects model is influenced by the high level of weight attributed to the study by Dave et al. (2006). Figure 3.6 shows a forest plot of the same analysis, but with the study by Dave et al. removed. Here, the broken line representing the pooled effect has moved closer to 0, indicating the effect size has become smaller.

125 124 Figure 3.6 Random-effects forest plot of analysis of studies not including early retirees (laterlife employment on suboptimal self-rated health). The high level of weight attributed to the study by Dave et al. (2006) is strongly influencing the results. Gall 1997 Van Solinge 2007 Zuchelli 2007 Zuchelli 2007 Dave 2008 Roberts 2010 Combined d The output of the sensitivity analysis without Dave et al. can be found in Appendix 3, and shows the random-effects pooled estimate has reduced, from in the analysis including the study by Dave et al., to This effect is much smaller in size, although retains its negative direction, suggesting that later-life employment is beneficial for self-rated health. This effect is also much smaller than that of the analysis of studies excluding early retirees, suggesting that health is poorer among the younger working group than the working group comprised of only people of retirement age and above. It is also interesting to note that heterogeneity is reduced after removal of the study by Dave et al., suggesting large differences exist between this particular piece of research and the others in the analysis.

126 125 Figure 3.6 Random-effects forest plot of analysis of studies not including early retirees (laterlife employment on suboptimal self-rated health), with the study by Dave et al. (2006) removed. The pooled estimate has moved closer to the 0 mark, signifying a smaller effect. Gall 1997 Van Solinge 2007 Zuchelli 2007 Zuchelli 2007 Roberts 2010 Combined d The third subgroup analysis on self-rated health was by gender. In the analyses of depression, the effect for males was shown to be larger than the effects for both women alone, and men and women combined. However, such a result is not replicated here. Table 3.7 shows the random-effects pooled estimate for males alone is , which is small, according to Cohen s effect size threshold. This is also smaller than the effect size for mixed-gender groups (-0.322) as well as the original analysis effect size (-0.240). It is not much smaller than the effect for single-gender groups (-0.188), but that would be expected here, as four of the five single-gender groups are of males, with only one effect for females (Zucchelli et al. 2007). This suggests that the effect may be greater for females than males, and this idea will be discussed further in the following section. The same differences in effect size

127 126 between the groups analysed are also reflected in the fixed-effects estimates presented in Table 3.7. Finally, an analysis of studies conducted using data only from the USA provided a pooled estimate of in the fixed-effects model, and in the randomeffects model. As it is the random-effects model that is of interest here, we can compare this to the main analysis pooled estimate of , and see that it is almost half its size. This suggests that later-life employment is associated with a decreased likelihood of suboptimal health, although less so than other countries. Looking at the pooled estimate of all countries excluding the USA, its value can be noted as in the fixed-effects model, and in the random-effects model. Both of these results are significant at the 5% level, which is particularly interesting as the randomeffects estimate is not significant when all countries are analysed jointly in the main analysis. This might suggest that the effects of later-life employment in the USA contribute to the heterogeneity present within the results.

128 Conclusions, limitations and rationale for future research Conclusions of the meta-analysis Two final meta-analysis models were presented in tables 3.2 and 3.3, for depressive symptomatology and self-rated health, respectively. The overall trend in results points in the same direction for both analyses and suggested later-life employment is associated with improved well-being for both outcomes. The meta-analyses also demonstrate that, in accordance with Cohen s scale of effect size, the effects of laterlife employment on health outcomes are small, with a pooled effect for depression of , and for sub-optimal self-rated health of From these coefficients, it can be observed that levels of depression among those remaining in employment in later-life are slightly lower than among those of the same age who are retired, the same can also be concluded for suboptimal self-rated health, with those who are employed are slightly less likely to suffer suboptimal self-rated health than those who are retired. However, these results were not significant and so caution must be taken when drawing conclusions from them. Subgroup analysis allowed an examination of the effects of certain study-level covariates on the results of the meta-analyses. With regards to depression, it was found that an analysis of study effects comprised of older populations only (SPA and above) provided a smaller effect (-0.120), and an analysis inclusive of early retirees, and therefore a slightly younger population overall, provided a larger effect (-0.147). This suggests the inclusion of early retirees, who are already associated with poorer levels of mental health than those who retire later, add to the negative effect of retirement when compared to a retired group of statutory retirement age and above. At the same time, prevalence of depression tends to decrease with age, which may also help to explain the smaller pooled estimate in the group older people only. With regards to self-rated health, only a small difference exists in the pooled estimates of the studies including early retirees compared to those which do not. Although the findings suggest slightly better self-rated health among the populations of older people only in comparison to populations including younger people, both early retirement and older age, regardless of employment status, can be associated with poorer health. However, reasons for early retirement are likely to be varied, making

129 128 early retirees a highly heterogeneous group to being with, and creating difficulties in interpreting any true effects which may underlie the results. The subgroup analysis also demonstrated gender differences in the outcomes, particularly with depression. The pooled estimate for male-only groups is larger than that for females ( compared to ), with the results of studies providing a combined male and female effect lying between these two values. Therefore, we can conclude that the link between later-life employment and improved mental health is greater in males than in females. This finding ties in with the ideas surrounding role theory presented in Chapter 2. Especially among already older populations, it might be the case that men see their role in employment as particularly important in feeling valued and worthwhile, and continuation of work into later-life might therefore lead to improved levels of mental health. This importance of work might be less marked among women of the generations included in the studies analysed due to fulfillment of other familial and societal roles across the life course (Kim & Moen 2002). A distinction between gender effects in the analysis of self-rated health was harder to discern, as only one study provided a female-only effect. However, the pooled estimate for male-only studies was smaller than the pooled estimate for all studies ( , compared to ), suggesting that the beneficial effect of later-life employment on self-rated health may be greater for women than it is for men. This could be explained by differences in retirement age. As, due to retirement policies currently in place, men typically retire later in life than women, the average age of males across studies is likely to be higher and therefore health is likely to be poorer than that of women on this basis alone. Finally, the subgroup analysis suggested differences exist between countries with regards to impacts of later-life employment. In the analysis of depression, studies using data from the USA showed an association between poorer mental health and later-life employment, which was in contrast to both the overall analysis of all countries combined and studies from the UK only, which linked later-life employment to lesser depressive symptomatology. This lack of specific SPA in the USA, and the subsequent encouragement for people to work longer because they cannot afford to retire securely, ties in with ideas surround effects of job-lock. It would be expected that levels of depression would be higher when individuals are

130 129 working when they might prefer to be retiring, and these effects might be particularly pronounced among cases working in poor quality employment. In the analysis of self-rated health, although the association between later-life employment and improved health remained for studies using data from the USA only, the effect was approximately half of its value for all other countries, suggesting levels of self-rated health are lower among older employees in the United States than in other countries analysed in this particular research. Again, this may tie in with ideas of job-lock, with cases forced to continue employment for longer with poor physical health because they cannot afford to exit the workforce earlier Limitations and rationale for further research on the topic of later-life employment and health outcomes One of the key reasons for conducting the meta-analysis was to detail the heterogeneous effects of later-life employment on health outcomes presented within the systematic review. Although levels of heterogeneity remained very high within the meta-analysis, subgroup analysis allowed some insight into aspects of the studies which produce high levels of heterogeneity. These aspects will be taken into consideration in further analysis of the topic, and a focus will be placed on examining the extent to which confounding factors might influence any analysis concerning effects of later-life employment on health. Despite the heterogeneous results, general trends were established among the results of the meta-analysis. The systematic review presented in Chapter 2 provided too great a variety of results to discern whether or not any overall effects existed. This was with regards to many aspects of the studies, including design and methodology, samples used, outcomes examined and covariates adjusted for. As was to be expected, the value of I 2 was too high to produce significant results, but an association was demonstrated between working in later life and both better levels of depression and self-rated health. Although the number of studies with extractable data was small, some attempt was made to homogenise studies, and data were only taken from studies which did not focus on specific aspects of employment, such as manual occupations or high levels of workplace stress. The subgroup analysis

131 130 attempted to further reduce study differences by stratifying on the basis of some measurable study-level characteristics. Although meta-analysis can be conducted with few studies, as with any statistical analysis, precision is increased with larger sample sizes. This particular analysis focuses on only a small number of studies, with only eleven analysed for depression, and eight for self-rated health. However, methods have been implemented to work alongside this limitation throughout the chapter, and justification for use of the methods employed is made wherever necessary. Subgroup analyses contained even lower numbers of studies, which again indicates caution must be taken when examining results. However, although pooled effects are unlikely to be highly reliable, the subgroup analyses have proved useful in demonstrating the importance of accounting for various study-level characteristics that have contributed to the variation in findings presented in the systematic review. Publication biases are particularly hard to detect in an analysis comprised of a small number of studies. This has not been a focal point of the meta-analysis here. All of the studies included in the analyses have been published in journals specialising in the fields of health and social sciences, and all have been published within a reasonably similar timeframe. Additionally, when levels of heterogeneity arising from study variation are so high, the likelihood of detecting any bias which might be due to issues of publication is small, and any estimation would be hard to establish. Another limitation arises from the fact that, by chance, all eight of the studies included in the analysis of self-rated health have effects calculated from crude data only. This is an important point to remember, as no controls for possible confounding factors, such as age, which is already known to have an effect on health outcomes in itself, are in place here. This means the effects used in the analysis are not only of employment but also of these potential confounders. The main body of analysis in this research will control for possible confounding effects, including age, and it will be interesting to compare the outcome for self-rated health from the forthcoming analysis with that of the result from the meta-analysis in order to gain a measurement of the possible extent to which these confounding variables distort precise estimates.

132 131 The observational design of the studies included within this meta-analysis leads to another of its limitations. Meta-analyses are most often performed using studies comprised of randomised controlled trials in clinical or psychological settings, in which the researcher decides eligibility for inclusion in the study and also controls the exposure for each participant (Stolberg et al. 2004). Clearly, in observational studies, this is not an option and, in this instance, those who retire and those who continue employment beyond retirement age do so entirely for reasons beyond control of the researcher. This is a key contributor to heterogeneity, as the lack of control over the exposed populations means they are likely to be inherently different, both within studies and between studies. The very high values of I 2 throughout are demonstrative of heterogeneity arising from this issue. This expected variation is an extension of the heterogeneous findings of the systematic review presented in the first chapter. The analysis of self-rated health has its own limitation in that the data extracted for inclusion were always crude estimates and never taken from adjusted models. Therefore, it should be expected that the magnitude of the pooled effects in all analyses of self-rated health are larger than they would be had then been extracted from models accounting for sources of bias. The high level of heterogeneity observed within this analysis has important implications for the forthcoming analyses of the topic of working in later-life. Its presence creates difficulty in finding any reliable common effect, as the varied populations themselves, both between the studies in general and between the exposure groups within those studies, mean effects of later-life employment are expected to differ because they are looking at different people (hence the retention of the random-effects model). The analysis highlights the importance of ensuring the focus is not placed on the significant results estimated by the fixed effects models, which might lead the researcher to infer that later-life working has strong beneficial effects on depression and self-rated health regardless of population characteristics, such as type of employment or underlying levels of health. Indeed, two recent systematic reviews of literature on later-life working find it to be largely beneficial to health outcomes, but the authors fail to successfully account for population differences and generalise their findings to individuals with an extensive list of different characteristics (van der Noordt et al. 2014; Maimaris et al. 2010). This

133 132 conclusion is unlikely to be the case, and the large value of I 2 in both analyses presented here directs the researcher to consider the potential impact of an individual s socio-demographic and health characteristics on whether or not continuation of employment in later life would really offer beneficial outcomes. This particular study aims to further the current research discussed in this chapter and Chapter 2 by using propensity score matching, which works to align the observational design as closely as possible with that of randomised controlled trials. In this instance, the method will work by ranked matching of individuals in the employed and retired groups, based on similarities in the distribution of their observed variables (Rosenbaum & Rubin 1983), and ensuring as high a level of comparability between the two groups of interest as possible. The need for a method to successfully control for bias in attempting to determine causal relationships in the subject area is further demonstrated in the subgroup analysis of the chapter, where heterogeneity is reduced by analysing studies with certain study-level covariates separately. The subgroup analysis also proved such stratification affects the pooled estimates provided by the analyses. This again suggests different effects exist for different groups of people. The meta-analysis here showed substantial differences in the effects for males and females, as well as a difference in effects and reduction of heterogeneity based on inclusion and exclusion of early retirees. This further justifies the decision to retain only respondents of SPA and above in the forthcoming work produced here. Examining further covariates which may impact overall results was beyond the scope of this meta-analysis, due to the small number of studies with different covariates included within them, but it has proved useful in stressing the importance of taking these covariates into consideration and, once again, the potential benefits of using a matched-participant design in the following chapters of the thesis.

134 133 Chapter 4: Data and Research Methodology In light of the findings of the systematic literature review and meta-analysis presented in Chapters 2 and 3, respectively, this chapter will outline the key methodological ideas behind the research carried out for the purpose of this thesis. Firstly, an overview of the methodological issues potentially creating heterogeneity among previous research on the topic will be provided, with a focus on how confounding and selection bias might be particularly relevant when looking at wellbeing and employment in later-life. Secondly, there will be a description of the data and methods to be employed within this particular research. The means by which the study sample was selected will be detailed, and definitions of key concepts will be outlined. Furthermore, propensity score matching as a method to eliminate confounding by selection bias will be discussed.

135 Key Methodological aims of the research The systematic review and meta-analysis carried out in chapters 2 and 3 of the thesis have highlighted the problematic issue of heterogeneity and bias when searching for causal effects within observational research. The overall conclusion of the systematic review was that previous research concerning the effects of later-life employment and wellbeing provides a variety of results, with no clear direction of effects distinguishable. Examination of the methods of previous work on the topic suggests this may be due to the inclusion of heterogeneous populations, with characteristics so diverse that they cannot be fairly compared without appropriate statistical techniques. The meta-analysis provided a useful means of quantifying this heterogeneity. This confirmed the main conclusion of the systematic review: the level of heterogeneity within and between populations studied was too high to establish any real effect of later-life employment on outcomes of wellbeing. Following the meta-analysis, subgroup analysis was carried out to assess whether or not focussing on clearer definitions of the population of interest, such as segregated male and female analysis and exclusion of early retirees, could reduce these high levels of heterogeneity. The value of the I 2 statistic, which was very high in the original meta-analysis, did reduce in magnitude, suggesting that a proportion of the heterogeneity present was due to the incomparable samples. However, subgroup analysis can only provide a limited estimate of how large an effect such a factor might truly have, as already estimated study effects and samples defined in previously carried out research are not viable for heavy manipulation. The methods used within this study aim to overcome these issues of heterogeneity, in order to establish whether a true causal effect of working beyond retirement age on health really exists. In other words, this research poses the question: if all potential bias due to individual characteristics is removed, does the health of those who continue employment beyond statutory retirement age, and those who retire, really differ? This question will firstly be asked of employment in comparison to retirement alone. Secondly, the question will be asked of how work quality affects this relationship, and how continuation of poor and good quality work impacts health in relation to retirement. In light of the systematic review and meta-analysis conducted so far, the study hypothesises that workers will need to be examined as a

136 135 stratified group in order to understand the true effects of continued employment, and that different effects will be present on the basis of the stratified work quality.

137 Bias and causality in observational studies Randomisation: the Gold Standard When seeking to establish causal effects of a particular variable (in this instance, later-life employment, as opposed to retirement) on an outcome (in this instance, health), the ideal methodology to be employed is randomisation. Randomisation ensures every individual involved in the study has an equal chance of belonging to the exposure or control group, regardless of personal characteristics. A randomised study of the effects of later-life employment on wellbeing would ensure that every individual within the study population had an equal chance of either continuing employment or entering retirement once SPA was reached, and would be assigned to either group entirely at random. Factors such as employment type, work quality, social class and baseline health measures, which might influence group membership in reality, would have no bearing on whether an individual is selected into the exposure or control group. Such total random allocation of group membership leads to the assumption that these potentially confounding personal characteristics are therefore evenly distributed among each of the study groups, and cannot lead to biases in measured group differences. However, in reality, selection of group membership by the researcher is not a viable option and would present substantial ethical challenges, and it is in these instances that pre-collected observational data is particularly useful to research. Where correct methods are used, analyses can be designed to simulate the assumptions of randomisation to as great an extent as possible. The key issue separating randomised trials from observational studies is that of confounding and bias due to the non-random selection of individuals into either the treated (exposure) or control response groups. Any trial includes a sample of individuals i, each with a vector of background characteristics X, which is descriptive of i prior to belonging to either the treated (r Ti ) or control (r Ci ) response group. To use this example, this vector of background characteristics X, which may include factors such as gender, age, socio-demographics, workplace characteristics and baseline health, describes each study participant before they reach retirement age and enter either later-life employment or retirement. In a randomised trial, each participant i is assigned to their response group entirely at random, and it can

138 137 subsequently be assumed therefore, that the vectors of characteristics X will have the same distributions among both groups (X it = X ic ). In other words, it can be assumed that there are no differences between those who continue employment and those who do not in terms of their socio-demographic, workplace and health characteristics. In a randomised trial, where the assumption of X Ti = X Ci holds, the causal effect of group membership r Tj r Cj can be estimated. To use this example again, this would mean that, as long as the background characteristics of the employed and retired people were exactly the same, the value of r Tj r Cj will be solely due to the effect of belonging to the group of employed individuals compared to the group of retired individuals, and therefore a causal effect of later-life working can be been assumed (Rosenbaum 2002). Again, the main pitfall of using observational data is that randomisation is not possible. The data has been collected beforehand, and whether the individual belongs to the exposure or control group is pre-determined. This pre-determination of group membership is problematic. In the case of this study, it is highly likely that an individual s background characteristics increases the probability of entering either later-life employment or retirement. Some of the factors influencing the retirement decision were outlined in Chapter 1, and will be discussed in greater detail further on in the work. For example, those individuals with physical functioning difficulties might be more likely to exit the workforce earlier than those without, immediately creating a retired group with a likelihood to demonstrate poorer physical functioning than their better-functioning working counterparts. In other words, the vector of background characteristics of the employed and retired groups differ (X Ti X Ci ), and the key assumption underlying randomisation and subsequent ability to estimate a true causal effect cannot be satisfied. The effect estimated by r Tj r Cj will be biased by the poorer mean baseline health of the retirees. This bias is likely to exist on the basis of many individual characteristics, and will be discussed in further detail shortly. These characteristics affecting both probability of group membership and the outcome of interest are referred to as confounders, and their effect on the magnitude or direction of outcomes is referred to as selection bias.

139 Selection bias and confounding in observational studies Bias occurs as the result of errors inherent in the design or analysis of a study. This subsequently provides the researcher with an inaccurate estimation of effects, with a magnitude or direction of effects which is not truly representative of the actual effect. When the aim of the study is to generalise findings to wider populations, biasing of results can have serious consequences. In this instance, biased results may lead to the assumption that an effect of working beyond SPA has an effect on health, when in fact it does not. It is particularly important to consider selection bias when conducting observational research. Selection bias arises due to insufficient controlling for variables which confound results by affecting both the chances of belonging to a particular group as well as the outcomes of belonging to a particular group. In this instance, a confounding factor would influence the probability of working in later-life as well as health, such as the aforementioned example of poorer baseline physical functioning leading to an increased likelihood of workforce exit as well as of subsequent poorer health. Where confounding is present, an association may be apparent when in actual fact there is none which truly exists (Grimes & Schulz 2002). If, in this instance, the effect of retirement is tangled up with the effects of poor physical functioning, it is not possible to estimate its effect alone. Instead, the effect of retirement is estimated in addition to the effect of poor baseline physical functioning, and the biased results will be unreliable. As demonstrated in the previous section, totally random allocation to the exposure or control group is the only means by which the effects of confounding can be overcome, as this ensures any confounding factors are evenly distributed between the two groups. In turn, this leads to a null effect of any individual background characteristics (Jepsen 2004). In observational studies however, where randomisation of group membership is not viable, the research design and statistical analysis can be conducted with a view to ensuring effects of confounding are controlled for and subsequently minimal (Rosenbaum 2002). This particular research attempts to control for potential confounding and selection bias in both the design and execution stages of the study.

140 139 The potential effects of confounding in previous studies on the topic of later-life employment and health are highlighted by the mixed findings of the systematic review, and the high levels of heterogeneity demonstrated by the meta-analysis. Both are likely to be the result of attempting to examine studies with populations too dissimilar to be fairly compared. Drawing from this conclusion, the analysis to be presented throughout this thesis will focus on a restricted group of individuals who are selected on the basis of key similarities to one another. In turn, this will lead to minimal effects of confounding which might otherwise occur if inclusion of individuals with highly variant background characteristics was permitted. Such characteristics to be considered here are age and type of economic activity, and will be discussed in greater detail shortly. The analysis of the data uses two techniques to control for confounding factors. Firstly, multivariate spline regression analysis will allow an examination of the adjusted pattern of health outcomes across the transition to retirement age and beyond. Secondly, propensity score matching will be used as a form of analysis to ensure comparisons of wellbeing scores between the groups of employed and retired individuals are only made among individuals who, on the basis of their background characteristics, are deemed comparable. Again, a further explanation of these methods will be presented later in the chapter.

141 Data and methodology for the current research Overview of the data: The English Longitudinal Study of Ageing (ELSA) The data to be used within this study are taken from the first five waves of the English Longitudinal Study of Ageing (ELSA). The study is comprised of 12,099 individuals at baseline (11,392 core members), and data is collected by interview every two years. Biomedical information is collected by means of a nurse visit every four years. Respondents were initially selected from either the 1998, 1999 or 2001 wave of the Health Survey for England (HSE), a cross-sectional, annually collected, nationally representative household survey. The HSE employs equal probability sampling, and so the core baseline ELSA sample taken from this is a nationally representative cohort of people aged 50 years and older (born on or before 29 th February 1952), living in private households within England (Marmot et al. 2002). The ELSA datasets are particularly useful to this research as they are multidisciplinary in nature and contain a rich set of variables which are relevant to the research questions of interest. Detailed socio-demographic information is provided for all cases, alongside detailed workplace, pension and financial information. Such factors are crucial in providing a sufficient set of covariates on which to compare individuals when considering issues of potential selection bias and confounding. Additionally, ELSA provides detailed information on various aspects of health and wellbeing. This work will focus on just three of the health variables provided: depression, self-rated health and cognitive function, all of which will be discussed in greater detail further in the chapter Selecting the sample As has been highlighted throughout this work so far, sample characteristics are a highly important consideration in research seeking to establish causal effects. It has already been considered that the heterogeneity observable in the systematic review and meta-analysis is partially the result of attempting to compare essentially noncomparable populations. As a result, it has not been possible to establish whether a general direction of findings concerning later-life employment and health exists. The effectiveness of subgroup meta-analysis in removing some of the heterogeneity

142 141 indicated by the I 2 statistic on the basis of simple factors such as gender and retirement age further strengthens the argument for greater comparability of sample individuals in the search for causality. The first step of this particular research is to deal with potential heterogeneity within the data to be used by means of sample restriction. Table 2.2 in Chapter 2 outlines the key sample characteristics for each of the studies included in the systematic review, and their diversity is immediately obvious. The remainder of this section describes the selection process for the sample to be used throughout this work and the means by which it aims to overcome some of the potential issues of bias. One of the key concerns arising from previous studies on the topic of later-life employment and health is that of the inclusion of non-comparable age groups. As this study is interested only in the effects of working beyond SPA, it is essential that only those of statutory SPA and above are included in the analyses. Of the studies included in the systematic review, a large proportion included participants from age 50 onwards (Table 2.2). Age is, in itself, a potential source of confounding. Increased age is associated with a lower likelihood of being in work, alongside an increased likelihood of poorer physical wellbeing and cognitive function. Additionally, some studies also suggest an increase in age is associated with a decrease in symptoms of depression, although it is likely this finding is also the result of complex relationships between individual circumstances and wellbeing (Jorm et al. 2000). As mentioned beforehand, the research carried out here uses the first five waves of the ELSA data. This provides a total of 16,761 cases who respond to at least one of these waves 28. As the interest lies in observing individuals as they make the transition to retirement age, it is necessary to only include those who actually do so at one of the waves of data. In other words, the study is only interested in men who reach age 65 and women who reach age 60 while belonging to ELSA. The study is also only interested in participants who either continue to work or retire at SPA directly from employment. Therefore, information must be gathered concerning economic activity at the wave prior to the one at which the individual reaches SPA. As a result, the sample must only contain older adults who retire from wave 2 onwards. Subsequently, the oldest participants are those who reach 28 There are 16,765 cases in total but four contain no information whatsoever and so are immediately discounted from the study.

143 142 retirement age by wave 2, and the youngest are those who do so by wave Removing all cases who do not meet this age criteria, as well as those who do not provide any employment information at the wave prior to reaching retirement age, leaves a sample of 2,490 cases. The next step of selecting the sample is to ensure inclusion of only those individuals who make a pure transition into either later-life employment or retirement from either employment or self-employment at the previous wave. Any respondents not participating in employment or self-employment at the previous wave are removed from the sample. This leaves a dataset of 1,227 cases. Finally, it must also be ensured that information is captured on employment activity at the wave at which each respondent reaches retirement age. There are 60 cases with missing data here, and removal of these leaves a final dataset for analysis of 1,167 cases. The large reduction in sample size, from 16,761 initially, is due to the exclusion of cases who do not cross SPA at any wave of data. As a result, the analyses to be discussed further in the study are relevant only to individuals of SPA at the time of writing (65 for men and 60 for women). Additionally, the sample size is reduced by the exclusion of cases who do not work directly until SPA, as well as of those who do not either continue to work or take normal retirement upon reaching SPA. Although this eliminates a high number of cases from the data, potential issues of selection bias arising from the inclusion of early retirees are reduced. The first section of the following chapter will discuss some of the key differences between the overall ELSA sample, which is representative of the overall English population, and the sample selected for the analyses presented within this particular study The exposure and control variables In the attempt to find a causal effect of later-life employment on health outcomes, the first stage of this research was to ensure definitions of the key themes of the research questions were precise and therefore unlikely to produce bias within the 29 Due to the biennial nature of the data, some individuals will reach retirement age at the year between waves. These cases are classed as reaching retirement age at the wave subsequent to this.

144 143 results themselves. The work looks at three key concepts throughout: later-life employment, retirement and work quality. Later-life employment and retirement The two main concepts are those of later-life employment and retirement, which are used to define the exposure and control groups in the final propensity score matching analyses. For the purpose of this study, later-life employment is defined only as employment at and beyond the UK SPA, and so later-life employees, in this instance, are classed only as those individuals who are working at and beyond age 65 for men and age 60 for women. Employment itself encapsulates individuals who state they are either employed or self-employed, and those who partake in both full-time and part-time work. Two variables in the ELSA dataset enable a good measure of this status. Firstly, participants are asked to describe their current economic activity, with optional categories of retired, employed, self-employed, unemployed, permanently sick or disabled, looking after home or family, other and semiretired. A second binary variable asks respondents if they took part in paid employment within the past month. Those with negative responses here were excluded from the sample, regardless of whether or not they described themselves as in employment initially. To demonstrate the importance of retaining clearly defined economic groups, Table 4.1 shows mean health scores by economic activity at the wave prior to reaching retirement age. Lower scores for depression and self-rated health indicate better outcomes, as do higher scores for cognitive function. Here, the poorer health outcomes among those who are unemployed, caring for family members or long-term ill can be observed in comparison to those who are of interest to this particular study. Table 4.1: Wellbeing and employment activity at the wave prior to reaching retirement age. Mean (S.D.) CES-D Poor self-rated health Cognitive function Employed (1.57) (1.11) (5.39) Retired (1.68) (1.12) (5.66) Unemployed (2.07) (1.42) (5.65) Permanently sick or disabled (2.38) (1.53) (6.05) Caring for home or family (2.07) (1.07) (6.08)

145 144 Table 4.1: Wellbeing and employment activity at the wave prior to reaching retirement age. Mean (S.D.) CES-D Poor self-rated health Cognitive function Semi-retired (1.35) (0.96) (4.90) As important as ensuring a clearly defined exposure group is determined for the study is ensuring an equally as well-defined control group is also established. In this instance retirement is classed as occurring at age 65 for men and age 60 for women, in line with the UK SPA in effect at the time of writing. Cases were classed as retired on the basis of describing their economic activity as retired in the same variable used in construction of the exposure group. Additionally, to ensure the group of retirees remained as similar as possible, respondents who stated they were retired but who had also taken part in paid employment over the past month were excluded from the analysis. As a result, the control sample used throughout this thesis consists solely of men and women who do not enter any form of retirement prior to reaching these ages. Examination of economic activity at the wave of data prior to that at which SPA is reached shows these cases to have been involved in either paid or selfemployment directly beforehand. This again ensures no early retirees will be included in the control group. Throughout this study, later-life employment is measured by means of a binary variable constructed from the above information, where 1 denotes those who enter later-life employment on reaching retirement age, and 0 denotes those who enter retirement. Work quality The second key concept considered throughout this work is that of work quality. Following the initial analyses of effects of overall later-life employment on health outcomes, those who do continue to work will be stratified according to whether they consider their employment to be of poor or good quality. Propensity score matching will again be employed to examine whether causal effects of later-life working exist 30 Due to the need for clearly defined exposure and control groups, respondents who class themselves as semi-retired, and whose precise economic activity is not clear, are removed from the study.

146 145 on the basis of work quality, and how the effects of these stratified measures of work compare to those of retirement. A binary measure of work quality is used throughout the research presented here, with 1 denoting poor quality work and 0 denoting good. The variable is derived from an effort-reward imbalance (ERI) scale (Siegrist 1996), which measures reciprocity within the workplace, the underlying theory of which was discussed in Chapter 2 (section 2.1). The original scale is constructed from two questions concerning effort in the workplace (whether work is physically demanding and whether the respondent feels under pressure due to a heavy workload) and five questions concerning reward (whether recognition is received, accuracy of salary, whether job prospects are good, whether job security is good and whether adequate support is received in difficult situations). A ratio score of effort and reward is calculated for each respondent, and a binary variable is derived from this score with 1 denoting a low level of balance (ratio score 1) and 0 denoting a high level of balance (ratio score >1). Research on the scale has found the binary ratio version of the variable to be as reliable in analyses as all other forms of the variable, including its continuous, quartile and log-transformed ratio forms (Niedhammer et al. 2004). ELSA also provides questions regarding workplace control and demand. However, these variables will not be examined here as the work presented within this study aims to further existing research which has shown significant associations to exist across the working life course between reciprocity and both depression (Siegrist 2008; Godin et al. 2005; Pikhart et al. 2004; Niedhammer et al. 1998) and self-rated health (Marmot et al. 2006; Siegrist 2005; Niedhammer et al. 2004). Additionally, while demand-control models are often restricted to measures of workplace characteristics, effort-reward models allow inclusion of macro-economic labour market characteristics, such as salary and wealth (Siegrist et al. 2004), which are important factors in this study. Analyses of the effects of work quality will be applied to workers of SPA and over only, and the impact of work quality will be compared to that of retirement. Further examples of research conducted using the effort-reward imbalance scale can be found in Siegrist et al. (2012) and de Jonge et al. (2000).

147 The outcome variables The research here centres on three key measures of health. The first of these is depression, used here as an insight into the potential effects of later-life working and mental health and wellbeing. Depression is measured using a shortened version of the CES-D scale, which has been shown to be an effective measure of mental health among older populations (Lewinsohn et al. 1997). The second of these is self-rated health. Self-rated health is generally deemed a good measure of an individual s general health and wellbeing, incorporating aspects of both physical and mental functioning. Previous research on the validity of self-rated health as an outcome variable finds it to correlate well with objective health measures (Miilunpalo et al. 1997; Bound, 1989) and that this suitable correlation persists into older age (Hunt et al. 1980; Fillenbaum, 1979). The third outcome examined is cognitive function. This study will examine whether or not continuation of employment attenuates the decline in functioning ability which is traditionally observed with increasing age Depression The ELSA datasets include the CES-D scale (Radloff, 1977) as a measure of depressive symptomatology. ELSA uses an eight-point version of the original twenty-point questionnaire on symptoms of depression, and responses to each of the questions are binary, Responses are recoded to follow the same direction where the original data does not, so that a positive response always indicates presence of the symptom. The eight symptoms identified by the data, relating to the week prior to the interview being held, are feelings of depression, finding everything an effort, restless sleep, unhappiness, loneliness, non-enjoyment of life, feelings of sadness and lack of motivation. The highest possible score is 8, whereby the respondent would be suffering all eight symptoms, and the lowest possible score is 0, whereby no depressive symptoms are presented Self-rated health Self-rated health is measured using a five-point Likert scale response to the question of how a respondent rates their general health. In waves 1, 2, 4 and 5 of the data, this variable corresponds to the Health and Retirement Study version, and possible responses are excellent, very good, good, fair and poor. Wave 3, however, uses a different set of potential responses, in accordance with HSE coding: very good, good, fair, bad and very bad. This difference in scales is problematic,

148 147 and research conducted on the matter suggests the different wording is likely to promote different participant definitions of the terms when being responded to (Bowling & Windsor 2008). One potential solution was to create a four-point scale response, combining the categories of excellent and very good and poor and very poor. However, wave 1 contains both versions of the variable, and a crosstab of each demonstrated doing so created a poor match of categories, with some individuals classing themselves as in very good or good health in one variable yet fair or poor in the other. Subsequently, the decision was made to remove wave 3 data from further analyses of self-rated health. As the data is based on information at retirement age, the only consequence of this will be a reduction in sample size. This will be discussed in greater detail in the relevant analysis chapters Cognitive function The cognitive function variable to be used within this analysis is an index comprised of two separate executive and memory function indices. Executive function scores are generated through two key tasks: a letter cancellation task and a verbal fluency task. The letter cancellation task, originally developed for the 1946 Birth Cohort Study (Richards et al. 1999), presents the respondent with a 26-row by 30-column grid of random letters, and requests that they cross out as many of 65 pre-specified letters (in this case, P s and W s) as possible within a one minute timeframe, reading across the grid from left to right, and following rows from top to bottom (Huppert et al. 2006). The total number of letters reached by the respondent can then be used to calculate a speed of processing and a measure of accuracy. The verbal fluency task asks respondents to list as many words as possible from a pre-defined category (in this instance, animals) within one minute. The combined executive function index has a potential score range of 0 (poorest executive function) to 20 (highest executive function). The memory function index comprises of asking the respondent to correctly identify the day and date, a prospective memory task, and a word recall task. Prospective memory is measured by asking respondents early on in the interview to remember to write their initials at the top of a page attached to a clipboard when the interviewer hands the clipboard to them later on in the interview. If, after five seconds of being handed the clipboard, the respondent did not write their initials, the interviewer would prompt them. Finally, the word recall task asked respondents to remember a

149 148 spoken list of commonplace words, and to recall them both immediately afterwards and then following a delay, during which further cognitive tests are performed. The index presents a potential score range of 0 (poorest memory function) to 24 (highest memory function). The total cognitive function scale included in the analysis consists of the combined executive and memory function indices. Therefore, the highest possible index score, representing high levels of cognitive function, is 44, and the lowest is Justification for use of the outcome variables The outcome variables examined in this work were chosen on the basis of their ability to reflect three key domains of health: emotional and mental health, overall health and physical functioning, and cognitive functioning. The CES-D scale provides an overview of the emotional health of individuals, and the scale has been shown to correlate well with both clinical ratings of depression and other scales designed to measure levels of depression (Radloff 1977). Depression is one of the most commonly examined outcomes in research into the effects of later-life working, and its association with the topic has been shown to vary with the heterogeneous samples and definitions of working studied (see chapters 2 and 3). Therefore, examination of the scale in this work, which aims to account for these differences in seeking associations, will be useful in establishing both whether unbiased effects exist, as well as whether they prove later-life working to be beneficial or detrimental to mental wellbeing. Self-rated health is used as an outcome as it may provide an overall picture of health and wellbeing. This is preferable to using objective information concerning specific diseases, which is likely to lead to an incomplete overview of health. Investigations into the validity of self-rated health as an outcome variable have found it to correlate well with objective physical health (Kaplan & Baron-Epel 2003; Pinquart 2001), physical functioning (Dwyer & Mitchell 1999), mortality (Idler & Benyamini 1997; Kaplan et al. 1988; Mossey & Shapiro 1982) and mental health (Pinquart 2001). Again, self-rated health is another commonly examined outcome in studies of working in later-life, with the effects on it shown to vary, and so establishment of bias-free effects will add to previous research. Additionally, self-rated health might be particularly strongly linked to issues of selection into later-life working, in that

150 149 poor physical functioning and high rates of morbidity, both of which can be reflected by one s own overview of health, are likely to lead to an increased probability of workforce exit (Karpansalo et al. 2004; Humphrey et al. 2003; Dwyer & Mitchell 1999). Finally, a scale measuring levels of cognitive function complements analyses of mental and general health and provides a means of testing the use it or lose it hypothesis (Coyle 1994), which suggests continuation of mindful activities, such as employment, leads to a lesser decline in cognitive ability. Although previous studies concerning effects of working later in life on cognition have always shown beneficial effects (Bonsang et al. 2012; Roberts et al. 2011; Rohwedder & Willis 2010; Adam et al. 2007), the analyses presented within this work will add to previous research by demonstrating the effects of work quality on the outcome. Alongside the three key outcomes discussed above, the ELSA data provide a rich set of variables concerning health and wellbeing outcomes, including objective measures of health as well as measures of quality of life, life satisfaction, physical function and morbidity. This research uses CES-D, self-rated health and cognitive function only for two key reasons. Firstly, as mentioned beforehand, these outcomes represent three distinct areas of health which may be affected by extending the working life, and self-rated health especially might be viewed as a useful tool in understanding various elements of an individual s health status. Secondly, these three outcomes are commonly examined and discussed in previous research on the topic. The heterogeneous nature of findings concerning depressive symptoms and self-rated health among older workers was discussed in Chapter 2, and the research carried out within this study provides an opportunity to add to earlier work on laterlife working and these outcomes The covariates An earlier section of this chapter discussed the benefits of using randomised trials in research. Random allocation of group membership ensures all observable and unobservable individual characteristics are evenly distributed and on the basis of this, no bias can be present. The greatest pitfall of observational studies is the inability to randomly assign individuals to groups. In this instance, each individual s decision to

151 150 either continue employment or retire at retirement age is most probably linked to various personal characteristics, including factors concerning wealth, the workplace, and baseline health. Accounting for the diversity of these characteristics is central to distinguishing causal effects from selection effects. This section of the chapter outlines the covariates which are used in the analyses focusing on such issues throughout this work. These covariates are both those which appear often in the literature detailed in the systematic review, as well as those which might be less documented yet are deemed likely to have some association with either the exposure or outcomes of interest. Detailed descriptive statistics concerning the covariates can be found in the following chapter. Gender The subgroup meta-analysis presented in Chapter 3 demonstrated a substantial reduction in the I 2 statistic when analysing men and women separately, suggesting the relationship between employment and health is likely to differ according to gender. This was particularly pronounced in the analysis of depression. Such a finding is interesting if aspects of role theory are considered in conjunction with a life-course perspective approach. The role of employment is likely to differ across the life course for women to a greater extent than for men, with familial constraints leading to intermittent periods of part-time employment, absence from the workforce and a higher frequency of changes in job (Kim & Moen 2002). Therefore, the impact of retirement on women might be smaller due to prior adjustment to alterations in work patterns. This idea ties in again with the larger effect on depression seen among male only studies in Chapter 3, where a stronger importance of the continuous role of work is constructed over the life course. Additionally, evidence suggests that factors pertaining to successful retirement differ according to gender. Research by Quick & Moen (1998) finds positive experiences of retirement among men to be based on good prior workplace characteristics and adequate retirement planning, but for women to be based on sufficient retirement income and the financial ability to exit the workforce early. Additionally, being female leads to a higher likelihood of extended working as the SPA for women is lower than for men, and women are more likely to delay retirement so that it coincides with that of their parter (Humphrey et al. 2003). Marital status Prevous research has suggested older people who are married or cohabiting are more likely to continue to work than those who are divorced,

152 151 separated or widowed (Whiting 2005; Szinowacz & Davey 2004). This is likely to be because partners may wait for their spouse to retire before doing so themselves, especially among women (Humphrey 2003). Here, marital status is a binary variable, where married represents those who are in a first or secondary marriage, or a civil partnership. The reference group incorporates all individuals who are not categorically married, including those who are single, separated, divorced and widowed. Previous research has linked good spousal support with less stressfully experienced transitions into retirement and therefore better mental wellbeing (Lee 1978). Research also indicates this finding is particularly pronounced among women (Dehle & Weiss 1998). Moen et al. (2001) consider the importance of gender-role conformity among married couples, and link continuation of work among women with retired husbands with poorer marital satisfaction and mental wellbeing among both genders. Wealth As was discussed in the systematic review, there is much evidence to suggest that strong associations exist between wealth, employment and health in later life (Chandola et al. 2007; Marmot et al. 2002; Breeze et al. 2001). The paper by Chandola et al. (2007) on health and inequality in older age demonstrated accelerated decreases in wellbeing later in life on the basis of low grade employment. It might be expected that wealth is particularly troublesome in terms of confounding of results. On the one hand, wealth is likely to be associated with the health outcomes of interest, with higher levels of wealth associated with higher levels of health. Additionally, wealth is likely to be linked with timing of retirement in that those who occupy better-paid and therefore better quality jobs are more likely to want to remain in the workforce for longer than those who do not. The analysis here uses a pre-defined quintiles of wealth variable, where a value of 1 represents the poorest individuals and 5 represents the wealthiest. Socio-economic classification In addition to wealth as a measure of social inequality, a three-category version of the NS-SEC socio-economic classification system is used to identify socio-economic standing throughout this work. The variable was originally commissioned by the Office for National Statistics (ONS) as a means of representing individuals social class on the basis of employment conditions (Rose & O Reilly 1998). It is important to include a measure of social class as lower social statuses are associated with poorer health and earlier workforce

153 152 exit (Banks & Casanova 2003; Mein et al. 2000) and higher socio-economic positions with better health and longer workforce participation (Whiting 2005). The variable has been demonstrated to show significant health inequalities on the basis of its categorisation (Chandola & Jenkinson 2000). The long version of the variable, also included in the ELSA data, is collapsed to create three key economic categories: managerial and professional (ranging from high managerial to high supervisory positions), intermediate (ranging from intermediate clerical and administrative to non-professional and agricultural own account positions) and routine and manual (ranging from lower supervisory to routine operative or agricultural positions). Qualification level Although no previous research discussed here placed a large focus on qualifications and later-life employment in relation to health, it might be expected that qualification level plays a role in determining the likelihood of an individual remaining in the workforce. Previous research has shown those with higher qualifications work for longer (Whiting 2005) and that those without qualifications are less likely to (Disney & Hawkes 2003) The variable to be used here is binary in nature, with those who have a qualification above O-levels compared to those who have only O-levels or lower. Property ownership Property ownership is included in the forthcoming analyses as a binary variable, with those who own their property outright compared to a reference group of those who do not, which includes those who are still making mortgage repayments, those who are renting, and those who are living rent-free. Housing tenure has been shown to be an important predictor of both health outcomes and of likelihood to work for longer (Grundy & Holt 2001). Indeed, a study by Smeaton & Mckay (2003) determined the highest probability of continuing work in later-life was observable among individuals who had outstanding mortgage payments on their properties. This variable might display an interesting relationship with later-life employment and health. On the one hand, home-ownership is likely to be associated with higher levels of wealth, social class and work quality, which might also be associated with an increased likelihood of longer workforce participation as well as better health outcomes. However, individuals who do not own their property outright might also be more likely to remain in the workforce for longer due to financial constraints arising from continuous mortgage or rental payments. This forced continuation of work ties in with theories of job-lock which

154 153 propose non-preferential working beyond retirement age are likely to be associated with poorer health outcomes, particularly in terms of mental wellbeing. These ideas are considered in further detail in the following chapter. Partner s employment status This binary variable, of those with a working partner versus those without, is included as having a partner in the workforce may make participation in employment more likely. Previous research has suggested the decision to retire is largely driven by the desire to retire at the same time as one s spouse (Whiting 2005; Humphrey et al. 2003; Szinovacz & DeViney 2000). Women, who reach SPA five years after men, might be more likely to wait until their partner leaves the workforce before doing so themselves so that the transition to retirement can be taken together, which is associated with better mental health outcomes (Szinovacz & Schaffer 2000; Lee, 1978). Conversely, this again relates to the aforementioned theories of gender role conformity suggested by Moen et al. (2001) whereby retired men with working partners view their non-traditional role as detrimental to mental wellbeing. Caring and volunteering The analyses throughout this study control for whether an individual provides caring duties or voluntary work in addition to their role as a worker or retiree. Previous studies have found older people in part-time employment are more likely to take part in voluntary work, and that opportunities to volunteer may affect patterns of employment (Warburton et al. 1998). Research discussed in the systematic review showed significant beneficial effects of participation in voluntary work alongside paid employment (Hao 2008; Choi & Bohman 2007; Adelmann 1994). Ideas surrounding activity and role theories suggest volunteering in retirement allows the individual to maintain socially productive activities and a sense of meaning, therefore leading to better psychological wellbeing (Wethington et al. 2000; Hayward et al. 1998). Care-giving responsibilities, however, are linked to poorer health outcomes as well as earlier workforce exits, especially among older women (Evandrou & Glaser 2004; Dentinger & Clarkberg 2002). Again, binary indicators are used for both caring and volunteering. Difficulties with activities of daily living (ADL) score The ADL score (Dwyer & Mitchell 1999) is a measure of physical functioning which is controlled for throughout the analysis. Participants are asked thirteen questions regarding whether

155 154 or not they have difficulty in completing certain physical tasks 31. The set of tasks outlined in the survey include walking, sitting for prolonged periods of time, climbing stairs, kneeling and crouching and management of heavy objects. The variable here is continuous, with a score of 13 indicating the highest possible number of difficulties, and 0 indicating none. Physical work The ELSA data provides a measure of employment type at each wave, with three recoded binary categories sedentary, standing and manual. Much of the literature discussed in Chapter 2 detailed the ways in which different types of employment affect different aspects of wellbeing, with those in manual occupations often displaying poorer health than those in sedentary work. Alongside expected differences in wellbeing, a relationship between employment type and propensity to remain in the workforce might exist, with those in manual work less inclined to continue working due to stressful and physically demanding roles (Banks & Casanova 2003). Those in less physically demanding occupations may be more likely to delay retirement. Self-employment This covariate is included in the analyses as a binary variable. Self-employment is likely to be less stressful and demanding than working as an employee which in turn, is likely to be associated with prolonged workforce participation and better health outcomes (McNair et al. 2004). Working for the same employer This binary measure is accounted for as continuation of employment at retirement age is likely to be more common among people who can remain with the same employer they worked with prior to retirement age. This may be especially true among cases in good quality employment with no incentive to change employers. Subsequently, there are likely to also be associations with better levels of wealth and health. Private pension status The final variable included across analyses depicts whether or not an individual belongs to a private pension scheme (including occupational pension schemes). Membership of such schemes is likely to be associated with an increased propensity of longer workforce participation, with the incentive of increased savings over extended periods of time working. Again, these links may be 31 Some waves include two further questions regarding ADL, but the scale here is comprised only of those which are included across all waves.

156 155 particularly apparent among cases in good quality employment who are happy and healthy enough to delay retirement Statistical analysis The following section outlines the rationale for, and key ideas of, the methods to be used within this thesis. A more detailed description of the methods is provided at the start of each relevant analysis chapter. Removal of cases bearing the potential to skew results, such as those who differ greatly in age or type of retirement, is the initial step in removing selection bias from an observational study, and paves the way to conducting statistical analysis which might establish whether or not causal effects are present. Details of how this was managed for the forthcoming analyses presented within this work were provided in an earlier section of this chapter. Detailed descriptive statistics for the selected sample are provided in Chapter 5. Prior to running the final propensity score matching analysis, this study examines trajectories of health over the retirement age period. Longitudinal piecewise spline models are run in order to examine differences in the slopes of health outcomes of those who continue to work and those who retire. Additionally, piecewise splines allow an examination of whether the point of crossing retirement age and making the transition into either later-life employment or retirement is associated with an immediate change in outcomes when compared to the period immediately before retirement age. One of the popular theories of retirement is that of the aforementioned honeymoon period (Atchley 1982), whereby the months immediately following workforce exit are associated with improvements in health due to the loss of the stressful roles of employment. However, the discussion of previous literature on the topic found diverse effects of immediate retirement, again due to heterogeneous sample characteristics. The analyses here will provide a useful insight into the effects of crossing SPA for the sample of interest to this study, which is those who work directly until reaching retirement age. Piecewise spline regression analysis is a form of multivariate analysis, and although its main purpose here is to describe health trajectories, it is also a means by which

157 156 confounding can be controlled for. The vector of individual background characteristics, X, is included in the analysis as coefficients and the overall effect of the exposure is subsequently modified by their inclusion. The models are run with age centred on the UK SPA (65 for men and 60 for women), and the trajectories span the five year period before and following attainment of the normal retirement age. A spline model with one knot point (k = 1) at time t k, as will be implemented in Chapter 6, can be denoted for individual j at time t ij as in Equation The propensity score matching analysis offers an extension to the presentation of health trajectories by specifically controlling for selection biases and subsequently estimating the effects of working beyond retirement age on the three health outcomes of interest. Here, results focus specifically on the effects of group membership, rather than of crossing SPA as in the spline models. By ensuring an even distribution of covariates X across both the employed and retired groups, the technique aims to seek causal effects which are free from bias, and can be denoted as in Equation 2, where D is the binary exposure group response and X is the vector of background covariates. 2. A propensity score is calculated for each sample member by means of a probit model, representing the likelihood of each individual belonging to the exposure group on the basis of their background characteristics. Each individual belonging to the exposure group is then matched to the individual, or set of individuals, in the control group whose propensity scores are the closest in value to their own. In other words, each individual is matched with its counterpart who theoretically has the most similar set of background covariates. Subsequently, if it can be assumed that no differences exist between the full set of matched individuals on the basis of their covariates, the assumption of randomisation holds true and the background characteristics are deemed irrelevant to group membership. Therefore, the difference in the health outcomes of the individuals in the exposure and control groups is the result of the exposure alone (later-life employment as opposed to retirement) rather than of other

158 157 factors such as, in this instance, gender, wealth or baseline wellbeing, and this effect can be considered causal. The propensity score method provides a useful extension to the spline models used to examine trajectories of health in relation to employment. Not only does it use the vector of covariates X in the calculation of the propensity score, it then further adjusts for potential bias by comparing only individuals who are similar on this basis of this. Additionally, propensity score matching allows the estimation of three key effects of interest. Firstly, the average treatment effect on the treated (ATT) calculates the actual effect of the exposure (later-life employment) on the outcome (health) within the study-specific sample. Secondly, propensity score matching allows estimation of the counterfactual effect, or the average treatment effect on the untreated (ATU). This allows an examination of what the outcome might have been for the untreated individuals had they instead been treated. In this instance, this estimates the effects of continued working on the retired counterparts had they, instead, decided to remain within the workforce. Finally, calculation of these two effects allows estimation of a weighted average treatment effect (ATE), which is the effect later-life employment would have on any randomly selected individual outside of the study-specific population. Further details of these effects are discussed prior to the running of the propensity score analyses in Chapter 7. Another advantage of matching techniques following regression is the assurance of common support which further enables eradication of potential selection bias within the data. The assumption of common support holds when there are individuals in both the exposure and control groups who are similar enough of the basis of their characteristics X to be fairly compared. The same assumption does not underlie regression techniques. Subsequently, although the treatment effect adjusts for X, there are no means of controlling for the fact that higher values of a particular covariate are associated to a greater extent with the exposure group, or lower with the control group, therefore leading to subsequently biased extrapolations (Angrist and Pischke 2009). The ability of matching to compare only cases who are similar means unreliable extrapolation of the data in this manner is not possible. Furthermore, the precise level at which individuals must be matched can be manipulated by the researcher, by means of restricted caliper distances or degrees of stratification. Therefore, propensity score matching is more reliable than regression

159 158 techniques where there is a strong lack of covariate balance between the exposure and control groups. Following implementation of matching, tests of balancing are carried out on the covariates X entered into the models. Significant covariate imbalance (bias) prior to matching which is subsequently nullified through implementation of the technique demonstrates usefulness of the matching method in reducing selection bias to a greater degree than is possible by means of regression analysis (Dehejia & Wahba 2002). Alternative methods which aim to specifically deal with confounding were considered for analysis here, but propensity score matching was deemed the most appropriate method for this particular work. A popular method which aims to deal with confounding and seek causal effects is instrumental variable (IV) analysis. The technique works by use of an instrument, which is a variable related to the exposure and related to the outcome only through the effect of the exposure, and not through the vector of an individual s background characteristics. This subsequently leads to the assumption that treatment assignment is ignorable, on the basis of both observable and unobservable sources of bias (Greenland 2000; Angrist et al. 1996). To use this study as an example, an instrument would be related to an individual s chance of working beyond SPA, but would not be related to confounding factors, such as wealth and health. Such a variable might be a change in retirement policy. Considering current changes to the UK SPA, whether or not an individual has been affected by the rise in retirement age, according to their date of birth, could potentially be used as an instrumental variable. However, the main disadvantage of IV analysis is that an instrument with a strong association with the exposure is difficult to ascertain, and failure to do so will immediately result in biased estimates and an increased sensitivity to sources of unobserved bias (Martens et al. 2006). At the time of writing, instrumental analysis using changes to UK retirement age policy was difficult to implement as the numbers of respondents affected by such changes were small and limited to females only. A more reliable analysis of working past SPA using IV analysis might be an interesting extension to this research in the future. A second popular method which aims to overcome issues of confounding and seek causal effects is regression discontinuity design (RDD). To use the example of this study again, RDD would determine the effect of continuing work at SPA on health outcomes by estimating discontinuity of effects at the point of reaching retirement

160 159 age caused by entering later-life employment. The method assumes a smooth (for example, linear) association between outcomes and an individual s background characteristics, and so if this association becomes disjointed at the point at which later-life employment begins (the cutoff point), this will be the impact of exposure (Hahn et al. 2001). In other words, the average outcome of individuals at the point immediately before reaching retirement age is the counterfactual effect of those who retire at SPA. The measured treatment effect is the magnitude of the difference before and after the point at which the exposure is reached. However, a key limitation of RDD is its assumption that the sample s background characteristics are balanced by means of the exposure variable, and that at the point of reaching the cutoff, individuals who are about to enter later-life employment are exchangeable with those who are about to retire (Imbens & Lemieux 2008). Earlier chapters have highlighted the heterogeneous nature of older workers, and as this particular work focuses on selection into later-life working, it would be inappropriate to make the assumption that these individuals are the same as those who take retirement, and so propensity score matching, which can successfully account for these group differences, is the preferred method here.

161 Dealing with unobserved confounding in observational studies This chapter has so far outlined the problems that arise from, and means of dealing with, confounding in observational studies. However, these ways of overcoming confounding can only be applied where it is observable in nature. Unfortunately, it is often not the case that all confounding factors can be captured in the covariates X provided by the data, and observational studies are particularly prone to confounding which arises from some individual characteristic which remains unobserved within the dataset available to the researcher (Rosenbaum 2002). As with observable forms of bias, the only means by which the researcher can be assured unobserved bias is not affecting results is through randomisation. It can logically be assumed that if randomisation ensures an even distribution of all observable background characteristics X, it must also ensure an even distribution of those characteristics which are not observable u i (Rosenbaum 2002). While techniques such as matching might be useful in adjusting for identifiable biases, they are not efficient in the case of factors which are unobservable and therefore have no means of being controlled for. Indeed, it is this inability to successfully account for all potential confounding factors which remains the major drawback to using observational data within research: individuals who might believe be comparable on the basis of their background covariates X are, in fact, not comparable on the basis of their unobservable covariates u i. Although it is not possible to control for unobserved bias within observational studies, there are means by which it can be estimated and subsequently taken into account when making final conclusions and generalisations. Following the final analysis using propensity score matching, a sensitivity analysis will be carried out on the results, in this instance using Rosenbaum bounds (Rosenbaum 2002). This allows consideration of how great in magnitude the effect of any unobserved bias would need to be in order to alter the findings, and subsequently how much caution should be exercised when generalising these findings outside of the study-specific population. Again, further details on the mechanics of the sensitivity analysis will be provided in the relevant chapter.

162 Summary of the chapter The first section of this chapter outlined the disadvantages of observational studies with regards to how non-randomisation can lead to biasing of results, placing a focus on the ways in which the topic covered in this particular piece of work might be especially prone to issues of selection bias and confounding. In turn, these issue lead to difficulty in establishing causal effects. The chapter also provided a description of the methodology to be used within this study. An overview of the ELSA data was presented, alongside a description of how the sample of interest was selected for use. The key concepts of later-life employment, retirement and work quality, as they are to be used in the forthcoming analyses, were outlined, and a description of the variables was provided, both in terms of how each variable is to be included in the analysis alongside the rationale for their inclusion. The chapter then outlined the means by which the main statistical analysis will further control for bias and confounding by use of propensity score matching, a method which measures differences in outcomes between individuals who are as similar as possible on the basis of the values of their background covariates. Finally, the issue of potential unobserved bias has been discussed, and a sensitivity analysis following the implementation of propensity score matching has been proposed in order to ascertain how great an issue for this particular study unobserved confounders are likely to be.

163 162 Chapter 5: Sample and Descriptive Analysis Chapter 4 provided an overview of why strict sample selection is important when seeking to establish whether or not a causal effect of working beyond the statutory retirement age truly exists. Additionally, the conflicting results of the systematic review and the high levels of heterogeneity due to incomparability of study populations within the meta-analysis further highlight the importance of sample selection as a means of reducing selection bias in observational research. This chapter provides descriptive statistics for the selected sample of older adults. The strict selection process for the sample, as well estimation of study-specific treatment effects by propensity score matching, means external validity in reference to the wider population is low. Although this is not necessarily problematic, as will be discussed in following chapters, the first section of this chapter compares the sample selected for this particular research to the overall, nationally representative, ELSA sample in order to ascertain key differences. Following this, sociodemographic and wellbeing statistics are provided for the cases to be included in the forthcoming analyses, in order to examine the most basic relationships between employment, health and the variables of interest to the study. Initially, statistics are provided for all later-life workers in comparison to retirees. Subsequently, they are provided for later-life work stratified on the basis of its quality, and findings are compared between poor and good quality employment as well as retirement.

164 Comparing the selected sample to the ELSA population As outlined in Chapter 5, research which aims to reduce potential selection bias by using methods such as strict sample selection and propensity score matching are particularly prone to suffering low external validity. In this instance, it is unlikely that the selected sample of individuals of interest to this study, who continue to work at least until the point at which they reach retirement age, shares the same characteristics as the overall population, where individuals leave the workforce for various reasons at a diverse range of ages. This section will outline how the sample to be utilised in the forthcoming analyses differs from the general ELSA population from which is it drawn. A brief overview of the ELSA dataset was presented in Chapter 4, whereby it was noted that ELSA is deemed representative of the population of those aged 50 and over living in English private households. With participants originally selected from the Health Survey for England, wave 1 of the ELSA data provides a set of 11,392 baseline core members who are therefore representative of the wider English population aged 50 and over. Additionally, the dataset includes the younger cohabiting partners of this core sample, and any new partners of the core sample as time progresses. However, these individuals are not necessarily representative of the target population, and so are removed from the general ELSA analysis carried out within this section of the chapter. Tables 5.1 and 5.2 present a set of wave-specific descriptive socio-demographic and wellbeing statistics for both the overall ELSA population of core members and the sample selected as described in Chapter 4. Figures presented are either percentages or means and standard deviations are presented in parentheses. Table 5.1: Characteristics of the core ELSA sample (S.E.) Wave 1 Wave 2 Wave 3 Wave 4 Wave 5 N=11,391 N=8,780 N=7,535 N=6,623 N=6,242 ELSA sample (core members only) Female 54.47% 54.47% 54.95% 54.53% 54.59% Age (10.44) (10.08) (9.926) (9.57) (8.14) Male employed (5.02) (4.64) (4.41) (4.33) (4.35)

165 164 Table 5.1: Characteristics of the core ELSA sample (S.E.) Wave 1 Wave 2 Wave 3 Wave 4 Wave 5 N=11,391 N=8,780 N=7,535 N=6,623 N=6,242 age Female employed age (4.45) (4.33) (3.98) (4.00) (6.86) Married 66.47% 65.35% 63.88% 63.32% 63.14% Wealth (mean quintile) NS-SEC professional/ (1.42) (1.41) (1.41) (1.40) (1.39) 30.37% 31.16% 32.59% 34.07% 34.59% managerial NS-SEC intermediate NS-SEC routine/manual Qualification(s) above o-level 23.67% 24.80% 25.41% 25.46% 25.66% 45.96% 44.05% 42.00% 40.46% 39.74% 28.89% 31.21% 39.47% 38.81% 38.38% Owns property % % % % % Partner works 36.76% 33.89% 32.26% 28.13% 23.56% Volunteer 11.74% 13.55% 14.48% 14.51% 15.44% Carer 9.46% 14.05% 12.65% 11.28% 11.91% Sedentary work 37.29% 38.71% 39.80% 42.96% 42.80% Standing work 32.06% 31.41% 31.67% 29.96% 29.47% Manual work 30.64% 29.88% 28.53% 27.08% 27.73% Self-employed % 18.56% 18.62% 21.88% 24.84% Working in same 85.01% 86.52% 82.07% 87.94% 89.43% job as last wave 33 Poor work quality % 29.68% 26.82% 27.86% Has a private pension 62.30% 66.65% 68.13% 69.82% 70.16% ADL score (1.80) (1.94) (2.08) (2.09) (2.14) CES-D score (1.99) (1.96) (1.96) (1.90) (1.95) Self-rated health (1.12) (1.12) (1.09) (1.10) 32 Percentages here are of cases classed as in paid employment only. 33 The variables relating to work quality were not collected at wave 1 of the data.

166 165 Table 5.1: Characteristics of the core ELSA sample (S.E.) Wave 1 Wave 2 Wave 3 Wave 4 Wave 5 N=11,391 N=8,780 N=7,535 N=6,623 N=6,242 Cognitive function (6.69) (6.68) (6.85) (6.86) (7.21) Table 5.2: Characteristics of the sample selected for analysis (S.E.) Wave 1 Wave 2 Wave 3 Wave 4 Wave 5 N=1,062 N=1,051 N=1,013 N=1,074 N=1,034 Female 66.01% 66.03% 66.04% 64.9% 66.05% Age (3.34) (3.34) (3.29) (3.36) (3.28) Male employed age Female employed age (2.50) (2.27) (2.01) (2.10) (2.21) (2.30) (2.15) (2.06) (2.02) (1.98) Married 74.58% 73.64% 74.14% 74.30% 73.02% Wealth (mean quintile) NS-SEC professional/ managerial NS-SEC intermediate NS-SEC routine/manual Qualification(s) above o-level (1.41) (1.29) (1.31) (1.32) (1.32) 35.39% 33.02% 32.64% 31.96% 33.11% 27.39% 28.24% 30.17% 31.12% 31.65% 37.22% 38.47% 37.19% 36.93% 35.24% 39.45% 39.37% 46.50% 44.16% 43.23% Owns property 43.57% 52.57% 57.34% 66.51% 72.74% Partner works 51.88% 43.96% 40.38% 33.99% 26.4% Volunteer 11.02% 10.47% 11.55% 11.82% 16.54% Carer 10.73% 17.03% 13.62% 13.78% 16.44% Sedentary work 38.11% 41.14% 41.71% 42.59% 43.31% Standing work 34.68% 31.68% 33.17% 33.38% 33.76% Manual work 27.21% 27.17% 25.12% 24.03% 22.93%

167 166 Self-employed % 16.42% 17.45% 20.14% 22.12% Working in same job as last wave 84.13% 87.32% 83.43% 89.58% 90.38% Poor work quality % 27.43% 25.91% 22.53% Has a private pension 79.94% 80.69% 80.45% 79.42% 78.34% ADL score (0.62) (0.85) (0.77) (0.74) (0.73) CES-D score (1.65) (1.63) (1.56) (1.55) (1.59) Self-rated health (0.94) (0.97) (0.96) (0.94) Cognitive function (5.28) (5.39) (5.47) (5.36) (5.46) Tables 5.1 and 5.2 demonstrate the ways in which the sample selected for analysis differs from the main, representative, ELSA sample on the basis of the variables of interest to this study. The remainder of this section discusses some of the key sample differences. Gender The percentage of females included in the selected sample is consistently around 12% higher than in the main ELSA sample. This increased percentage of women is likely to be due to differences in SPA. As the selected sample includes specifically those who retire at normal retirement age or work beyond this, and as the normal retirement age is five years lower for women than for men, there is an increased chance that women are more capable of working beyond their retirement age than men on the basis of better health. Additionally, women might continue to work to some extent beyond their normal retirement age due to their male partners remaining in the workforce for a longer period of time. Age One of the key factors on which the sample for analysis was selected was attainment of SPA at one wave of the data. For this reason, the selected sample will differ in age from the general ELSA sample. Tables 5.1 and 5.2 show the mean age at each wave is considerably younger among the selected sample due to elimination of a large number of older ELSA members as well as younger. 34 Percentages here are of cases classed as in paid employment only.

168 167 Mean working age As would be expected, the mean working age increases across waves as the ELSA cohort ages. The selected sample has a higher mean working age than the overall sample. This is due to the age restrictions implemented in the sample selection process leading to the elimination of cases who did not cross retirement age in the data period. Marital status The sample selected for this research also has a higher rate of married individuals than the original ELSA sample, by around 10% across waves. Again, this may, in part, be due to differences in age: the second sample in the table is considerably younger than the general ELSA population, and so there are fewer individuals likely to have been affected by the death of partners. Additionally, if it is considered that the decision to either continue work or retire may be associated with partner s working status at the time of reaching retirement age, the selected sample might be biased to include a higher number of people with partners due to the fact that all individuals either work or retire and do not leave the workforce for any other reason over which they might have less control, such as poor health. Wealth Tables 5.1 and 5.2 show the selected sample is, on average, slightly wealthier than the core ELSA sample across all waves. This could be in part due to continuous income received through working later in life, or due to the fact that individuals who work in better quality and subsequently higher paid jobs have an increased propensity to continue employment later into life. Additionally, the relationship between higher wealth and better health might contribute to an ability for wealthier individuals to have the physical capacity to work longer than their poorer counterparts. Socio-economic classification Within the whole ELSA sample, the highest percentage of workers are those in routine and manual occupations, followed by those in managerial and professional roles. Longitudinally, the proportion of routine workers decreases, while that of managerial and professional workers increases. This might be expected when considering functional limitations is likely to lead to greater exit of manual work among older populations (Dwyer & Mitchell 1999; Karpansalo et al. 2004). The middle group of intermediate workers also increases in proportion as the sample ages. In the selected sample, although there are always more manual workers than managerial, the difference in proportion is always much smaller. The

169 168 longitudinal differences are also smaller, and the only group with a continuous pattern is that of intermediate workers, which increases slightly at each wave. Finally, the percentage of manual workers in the selected sample is always smaller than that in the entire ELSA sample, suggesting that specific later-life employment has some association with intermediate and professional occupational classes. This is to be expected, as routine employment is particularly prone to selection effects of physical wellbeing, and as the population ages, ability to work in manual roles is likely to decrease. Qualification status The percentages shown here correspond to individuals who have qualifications higher than O-levels. The selected sample consistently has a higher percentage of cases with more qualifications than within the ELSA sample. Again, this may be expected considering higher qualifications might lead to better quality work. In turn, this increases the likelihood of an individual remaining in the workforce until retirement age or beyond due to both better physical capability of doing so, alongside greater enjoyment of the job. Property ownership As would be expected, the percentage of cases who own their property outright increases across waves, as higher numbers of individuals complete paying off mortgages. Until breaking even at wave 5, there is a higher percentage of property owners among the overall ELSA sample than the selected sample. Typically, it might be expected that the observed higher levels of wealth, qualifications and social status would be followed by a higher percentage of property owners. However, the actual results may reflect that, alongside those who work later due to the fact they enjoy their work, some work due to outstanding debts. Additionally, the younger average age of the selected sample may contribute to the results in that people have not had time as yet to complete mortgage repayments. Volunteering and Caring Although no strong longitudinal patterns exist with regards to the percentages of individuals who volunteer or provide care, differences do exist between the two samples. The selected sample consistently has a higher proportion of caregivers than the overall ELSA sample, and potential reasons for this will be explored further on in the chapter. Until wave 5, there are always fewer volunteers in the selected sample, which might be expected considering there are higher numbers of people still in employment than in the overall sample.

170 169 Employment type With the exception of wave 4, there is always a higher proportion of individuals in sedentary work in the selected sample than in the main ELSA sample. Longitudinally, the percentage of people working in sedentary employment increases with time. Again, this complements the idea that people continue employment which is enjoyable, better-paid and not physically difficult. The percentage of people participating in standing employment is always higher in the selected sample, and this gap widens in waves 4 and 5. This could be due to the fact that people are leaving their main employment as they get older and taking on lesser roles in standing jobs, such as shop work. As might be expected, the percentage of people working in manual employment is always smaller in the selected sample, and the gap between the percentage of manual and other workers is always much larger in the selected sample, again tying into ideas concerning the association between better employment conditions and later-life employment. Self-employment The proportion of self-employed individuals (from the sample of all employed cases) in both datasets increases across waves, suggesting those who are self-employed might be more likely to continue into later-life than employees. However, the percentage is always slightly lower in the selected sample. Same job With the exception of wave 3, the number of people working in the same job recorded at the previous interview increases longitudinally. Additionally, the percentage of people in the same job is always higher in the selected sample than the general ELSA population. Both factors suggest being able to work with the same employer increases likelihood of working later in life. Work quality In the main ELSA sample, the proportion of people in employment with a poor work quality generally decreases across waves. However, within the selected sample, there is a strong longitudinal trend for the percentage to decrease by each wave. This strongly suggests that poor work quality leads to the decision to exit employment at earlier ages. This idea is strengthened by the fact that the proportion of people in poor quality work is consistently lower in the selected sample and between-wave differences generally increase over time. This suggests the relationship between exit of poor quality employment and age is intensified among older populations.

171 170 Private pension status The selected sample consistently has a much higher percentage of individuals who have a private pension than the main sample. This could be due to the type of work those with private pensions participate in, in that those in better quality and higher paid jobs are more likely to be able to afford pension contributions. In turn, those in better forms of work have a greater likelihood of remaining in the workforce once retirement age is attained. Additionally, the ability to continuously contribute to private pensions might act as an incentive for those who are healthy enough to continue to work to do so for the sake of better financial status once retirement is finally realised. ADL score One of the key factors concerning selection bias in the field of later-life employment is wellbeing. Tables 5.1 and 5.2 demonstrate mean ADL scores among the main ELSA sample are consistently higher than among the selected sample, which has a higher proportion of workers. This suggests employment in later-life is often due to good health and an ability to continue working. Additionally, while there is a definite pattern of increasing scores across waves in the main sample, this pattern is much weaker within the selected sample. Although the mean ADL score is slightly lower at the start of the study here, its increase is much more gradual than among the population overall, again providing evidence of healthy worker effects. CES-D score While CES-D scores remain reasonably constant across waves among both samples, the selected sample always demonstrates lower scores, and therefore better mental health, at each wave than the main sample. Here, issues of selection bias are particularly evident as there are likely to be many factors affecting this difference, including the higher levels of wealth, perceived status, qualifications and physical wellbeing, all of which may confound any potential effects of the higher employment rate also observed among this sample. Self-rated health Again, self-rated health is always better among the selected sample. Unlike CES-D scores, self-rated health worsens over time. However, although still reasonably small, the decline is greatest among the selected sample. Relationships between self-rated health and other sample characteristics will be considered in greater detail shortly. Cognitive function Cognitive function scores are always considerably higher in the selected sample than the main sample. At this stage, it is difficult to tell if better

172 171 cognitive functioning is strongly associated with other favourable characteristics which encourage working into later-life, or whether the use it or lose it hypothesis presented in earlier in the work leads to a subsequent increase in cognitive function. It should be noted that longitudinal patterns in cognitive functioning are difficult to establish at this point due to potential learning effects arising from individuals taking the same test over multiple waves. Greater attention will be paid to this issue further on in the analysis. Tables 5.1 and 5.2 demonstrate definite differences exist between the sample of interest and the overall ELSA sample, highlighting the fact that, although internal validity arising from the methods of interest is likely to be very good, external validity would be poor if findings were to be generalised to the overall older English population. Assuming the main ELSA sample to be representative of the general English population aged 50 and over, we can state that the sample of interest to this study is, on average, younger, wealthier, more likely to be in managerial or intermediate employment, of higher perceived social status, better educated, less likely to volunteer, more likely to provide care, and are working in sedentary or standing employment with good work quality and private pensions. In terms of health, the selected sample presents fewer difficulties with activities of daily living, fewer symptoms of depression, better self-rated health and better cognitive function. The sample selection process ensures the results to arise from this research can only be applied to a small subsection of older workers and retirees, who are notably different in their characteristics than the older working population as a whole. However, this is not problematic, as this particular research is interested in the impact of working specifically beyond SPA. If changes to policy are to be aimed at encouraging workers to remain in the workforce until reaching SPA, it is important to understand how this will impact the health outcomes of this specific group of older workers.

173 Associations between later-life employment, sociodemographic characteristics and health This section of the chapter considers the selected sample only, examining some of the relationships between its socio-demographic characteristics in relation to laterlife employment and the three key health outcomes of interest to the study: depression, self-rated health and cognitive function. The methods to be used in forthcoming analyses place a focus on removal of selection bias, and so this chapter will consider where such biases may lie. The socio-demographic variables selected for use are both those which were frequently reported in earlier research on the topic, alongside variables from which bias is likely to arise within this particular study Associations between employment status and health Table 5.3 presents mean health scores by employment activity at retirement age. The three health outcomes of interest are depression, as measured by CES-D score, selfrated health and cognitive function. The CES-D score has a range of 0 to 8, with a higher score indicative of greater depressive symptomatology. Self-rated health has a score range of 1 to 5, again with a higher score indicating poorer health. Cognitive function scores have a possible range of 0 to 44, and in this instance a higher score represents better function. T-tests are used to examine the significance of group means differences. Table 5.3: Health at retirement age by employment activity at SPA (S.E.) Continues employment Retires CES-D (1.59) (1.52) Self-rated health (0.93)* (0.99) Cognitive function (5.41)*** (5.70) ***p<0.001 **p<0.01 *p<0.05 Table 5.3 tells us that, without controlling for any other factors, self-rated health and cognitive function are better among those who continue employment once statutory retirement age is attained. This is potentially due to selection effects, where those with already better health and cognitive function may be remaining in the workforce

174 173 for the reason alone that they are able to do so, rather than there being some protective effect of later-life employment on aspects of health. Conversely, levels of depression among retirees are slightly lower than among employees, although the difference in scores is very small. In fact, t-tests carried out to indicate significant differences between those who continue to work and those who retire demonstrate the only significant differences here to be for the scores of self-rated health and cognitive function. The following descriptive analyses will consider these relationships in further detail, and examine how they might vary among different groups of individuals Associations between employment status, gender and health One key feature of the data to consider throughout the analyses hereby presented is that of gender. The subgroup analysis which followed the meta-analysis presented in Chapter 3 showed a large reduction in levels of heterogeneity, measured by the I 2 statistic, when studies of men and women analysed separately were examined, and Chapter 2 outlines means by which both health and employment status are likely to differ in later-life with regards to gender. Table 5.4 shows the frequency and percentage of men and women within the selected sample who are employed or retired across the data. Where figures are shown for employment status at retirement age, all men are aged 65 or 66, and all women are aged 60 or Table 5.4: Employment activity by gender and wave Wave Employment activity Male Female Wave of reaching retirement age Enters later-life employment 182 (44.39%) Enters retirement 228 (55.61%) 503 (66.45%) 254 (33.55%) One wave following retirement age Remains in later-life employment 67 (30.73%) 238 (52.77%) 35 Men aged 66 and women aged 61 are those who crossed normal retirement age between specific waves of data collection.

175 174 Table 5.4: Employment activity by gender and wave Wave Employment activity Male Female Remains in retirement 151 (69.27%) 213 (47.23%) Two waves following retirement age Remains in later-life employment 28 (21.88%) 95 (37.70%) Remains in retirement 100 (78.12%) 157 (62.30%) (column %) Table 5.4 demonstrates that the selected sample has a much higher proportion of women in later-life employment than men. At reaching retirement age, 44.39% of men continue employment, compared to 66.45% of women. At the wave following attainment of normal retirement age, only 30.73% of the men who continued working previously were still in employment, compared to 52.77% of women. Again, at the wave subsequent to this, only 21.88% of men are still in employment, compared to 37.7% of women. There are two factors possibly playing a part in the differing values here. Firstly, it is important to note that the women in each section of Table 5.4 are consistently younger than the males, by roughly five years (the difference in SPA), and so may continuously be in better health and more capable of continuing employment than their male counterparts. Secondly, as men reach normal retirement age at an older age than women, it is possible that women of a similar age to their partners are continuing to work until they can retire simultaneously. Table 5.5 presents employment activity for men and women at the wave of reaching normal retirement age, according to whether or not they still have a partner in the workforce.

176 175 Table 5.5: Employment activity by gender and partner s employment status at SPA Continues employment Retires Male Working partner 77 (42.31%) 47 (20.66%) No working partner 105 (57.69%) 181 (79.34%) Female Working partner 227 (45.13%) 60 (23.62%) No working partner 276 (54.87%) 194 (76.38%) Table 5.5 shows 42.31% of men who continue employment at retirement age have a working partner, alongside 45.13% of women. While these numbers are not particularly insightful, it is worthy to note that a clear pattern appears to exist among retirees, with 79.34% of men and 76.38% of women who retire at normal retirement age having a partner who does not belong to the workforce. Although it cannot be concluded that a working partner encourages one s own continuation of employment, it seems reasonably likely that an effect may exist on the choice to retire. Table 5.6: Health at SPA by gender and employment activity (S.E.) Continues employment Retires Male CES-D (1.48) (1.19) Self-rated health (0.93) (1.00) Cognitive function (5.23) (5.47) Female CES-D (1.62) (1.72) Self-rated health (0.93) (0.99) Cognitive function (5.25) (5.60) ***p<0.001 **p<0.01 *p<0.05 Table 5.6 shows associations between health, gender and employment activity at retirement age, all of which are non-significant. Here, health is better across all outcomes among those who continue employment, with the exception of depression in men. Men who retire at normal retirement age display lower CES-D scores than those who remain in the workforce. This relationship is not true for women, whereby

177 176 retirement is associated with raised levels of depression. This finding can be considered in relation to Table 5.3, in which CES-D scores are very slightly lower among retired than employed individuals, and it is immediately apparent how heterogeneous results may subsequently arise from simply failing to account for differences in gender alone. The role of employment and the transition to retirement is potentially completely different for men than it is for women, and theories concerning this idea have been presented in earlier chapters. CES-D scores are always higher among women than men. For both men and women, self-rated health is better among the group of individuals who continue employment. These differences are likely, in part, to be due to selection effects, with those who are physically capable of remaining in the workforce later in life doing so, and those who are not opting for retirement when they are eligible to receive a pension. The better health of women than men in general is probably due to the aforementioned age effects. The difference in scores between men and women is notably greater among the employed that the retired. Again, this is possibly an effect of age. The very similar self-rated health scores of retired men and women might be further evidence of selection effects in that cases who are retiring may be doing so on the basis of poorer health and therefore display similar levels to one another. Cognitive function follows a similar pattern to that of self-rated health, with scores higher among both men and women who continue to work rather than retire. Scores are noticeably higher among women than men in both the employed and retired groups. The five year difference in age between genders here might imply a relationship exists between decline in cognitive function and age, regardless of employment activity. The results in Table 5.6 complement the results of the subgroup meta-analysis in Chapter 3 by demonstrating that, aside from differences in health existing in accordance with employment activity, differences exist additionally on the basis of gender alone. Analyses of men and women simultaneously in relation to retirement age activity are always going to be particularly prone to confounding from bias as, not only do characteristics of health differ between men and women, the effect of the five year difference in retirement ages will also bear an impact.

178 Associations between employment status, wealth and health The second socio-demographic characteristic considered in relation to health is total net non-pension wealth. In the forthcoming analyses this variable is presented as a quintile, but for ease of interpretation, descriptive statistics in Table 5.7 use a binary optimal-versus-suboptimal version 36. Table 5.7: Health at SPA by wealth and employment activity (S.E.) Continues employment Retires Optimal wealth CES-D (1.45) (1.30) Self-rated health (0.88) (0.87) Cognitive function (5.26) (5.54) Suboptimal wealth CES-D (1.70) (1.65) Self-rated health (0.96)* (1.02) Cognitive function (5.27)* (5.43) ***p<0.001 **p<0.01 *p<0.05 Table 5.7 shows there are relationships of some nature between wealth and health. Substantial differences exist between those with optimal and suboptimal wealth across all outcomes, with the wealthier cases always faring better. CES-D scores are marginally better among retired cases than employed among those with optimal wealth, but among those with suboptimal wealth, the difference appears much greater. This complements the idea of enjoyment of retirement being dependent on financial sufficiency (Ekerdt 1983). Additionally, the literature discussed in Chapter 2 proposed retirement from poor quality employment, which is likely to be associated with lower levels of wealth, was often associated with better mental health outcomes (Coursolle et al. 2010; Fernandez et al. 1998). This was due to ideas surrounding role strain reduction, whereby relief of physically or mentally demanding roles leads to a greater appreciation of workforce exit in the initial retirement period (Moen et al. 2001). However, none of the results regarding depression are significant. Figure 5.1 shows mean CES-D score as above but for men 36 Optimal wealth is defined as the highest two quintiles of wealth, compared to the lowest three.

179 178 and women separately, as differentiation in the relationship between employment and health has already been established on the basis of gender. The reverse story for men and women regarding later-life employment and depression can be observed again in Figure 5.1, with employment associated with greater depression in females and lower depression in males, regardless of wealth. As would be expected, optimal wealth is associated with less depression for both genders, although it is notable that the depression scores for wealthier women are actually very slightly higher than the depression scores of the poorer men, again highlighting the gender difference in depressive symptomatology in general. Self-rated health is always poorer among the less wealthy, although here it is better among those who work than those who retire. This may suggest the opposite effect to that presented by the case of depression: that people are still working because they are physically capable, rather than due to financial constraints. Cognitive function is, again, always better among the wealthier group, and always higher among workers than retirees, although the mean function scores of the poorer workers are still lower than those of the wealthier retirees.

180 179 T-tests were carried out on the basis of wealth to assess whether differences between the working and retired groups were significantly different and, as demonstrated by Table 5.7, although health might differ substantially on the basis of wealth category, exposure and control group differences are only significant among those with suboptimal wealth on the basis of self-rated health and cognitive function. This suggests the relationship between employment status in later life and health may be stronger among poorer individuals. Table 5.8: Optimal wealth by employment activity at SPA Continues employment Retires Optimal wealth 331 (50.53%) [61.30%] 209 (45.83%) [38.70%] Suboptimal wealth 324 (49.47%) [56.74%] 247 (54.17%) [43.26%] (column %) [row %] To consider these results from another angle, Table 5.8 breaks down employment activity at retirement age by wealth. Continuing employment appears to occur equally for both wealthier and poorer individuals, yet the decision to retire occurs at a greater rate among those who are poorer (54.17 %), possibly suggesting a link between retirement and poorer job quality. Looking at the row percentages in the table shows a much higher proportion of those in the optimal wealth group continue to work instead of retiring (61.3%). This could be due to better quality employment leading to increased likelihood of longer workforce participation, or of higher wealth due to continuation of employment. Among the suboptimal group, the percentage of workers is lower at 56.74%. Again, such a finding strengthens the idea of possible links between wealth, employment quality and activity at retirement age Associations between employment status, property ownership and health For the purpose of this analysis, the original tenure variable has been recoded into a binary response, separating those who own their property outright (owners) from those who are making mortgage or rental payments (non-owners).

181 180 Table 5.9: Property ownership by employment activity at SPA Continues employment Retires Owns property outright 417 (61.14%) [54.65%] 346 (71.78%) [45.35%] Pays rent or mortgage 265 (38.86%) [66.08%] 136 (28.22%) [33.92%] (column %) [row %] Table 5.9 demonstrates expected patterns with regards to ownership and the choice to participate in later-life employment % of retirees in the sample own their property outright, compared to 61.14% of workers. Additionally, 66.08% of people who are still making mortgage or rental payments continue to work rather than retire at normal retirement age. This strongly suggests the decision to work or retire later in life is made on the basis of financial constraints and commitments, as has been shown in previous research (Smeaton & Mckay 2003). Table 5.10 presents mean health scores by employment activity and ownership status. Table 5.10: Health at SPA by property ownership and employment activity (S.E.) Continues employment Retires Property owner CES-D (1.48) (1.51) Self-rated health (0.92) (0.96) Cognitive function (5.44)** (5.76) Non-property owner CES-D (1.70) (1.53) Self-rated health (0.93)** (1.04) Cognitive function (5.35)* (5.54) ***p<0.001 **p<0.01 *p<0.05 CES-D scores of property owners are low and very similar for both workers and retirees, and a t-test of scores between the two groups proves non-significant. However, among non-owners, scores are markedly higher, and the highest scores of all are present among the working non-owners, although again, no significant different exists between non-owner workers and retirees. This corresponds again with ideas concerning financial constraints leading to forced continuation of work

182 181 when it may not be the preferred option, subsequently resulting in poorer mental health scores. Self-rated health is always better among workers than retirees, and here it is better among working property owners than non-owners. The difference in scores between non-owner workers and retirees is significant, with p<0.01, and again suggests a deeper relationship may exist between wealth and wellbeing among poorer individuals than among those who might be wealthier. Cognitive function is always significantly better among workers than retirees, with working owners showing the best scores and retired non-owners showing the worst. This may be indicative again of relationships between wealth and employment confounding outcomes on the basis of employment activity. Again, due to the different nature of relationships between work and depression for men and women, Figure 5.2 illustrates how these relate to tenure for the two groups. The difference in scores is not too different on the basis of property ownership for women, with retirement again associated with poorer scores than employment. Among men, however, while scores are very similar for both workers and retirees who own their property, there is a large discrepancy in the scores of those who do

183 182 not, with working non-owners displaying much higher levels of depression than any other male group. This suggests again that there is a group of male workers who continue employment due to financial pressures, leading to an increase in depression. The level scores of male workers and retirees who own their property might reflect this idea further: if financial constraints do not exist, and individuals either retire or continue employment because it is simply their preferred option, there would potentially be no bearing on levels of depression Associations between employment status, socio-economic classification and health Table 5.11 considers whether socio-economic classification, measured by use of the three-class NS-SEC variable, differs among those who work beyond retirement age and those who do not. Table 5.11: Employment activity at SPA by socio-economic classification at the previous wave Continues employment Retires Professional/managerial 235 (34.56%) 150 (31.85%) Intermediate 200 (29.41%) 131 (27.81%) Routine/manual 245 (36.03%) 193 (40.98%) (column %) Although the differences in post-retirement age employment status are not overly large, it is worth noting that the highest proportion of retirees exit the workforce from manual jobs. However, the highest proportion of those who continue to work are also from this category, although the difference between this and those who continue in professional type employment is much smaller (1.44% less, compared to 9.13%). This may, once again, suggest associations exist between both preferred workforce exit from demanding roles, as well as forced continuation of them due to lower levels of wealth and financial constraints. Table 5.12 examines wealth and house ownership status on the basis of NS-SEC category to further investigate this pattern.

184 183 Table 5.12: House ownership and wealth by employment status and NS-SEC category at SPA (S.E.) Continues employment Retires Professional/managerial Mean wealth quintile (1.31) (1.14) Owns property outright 60.99% 77.48% Does not own property outright 39.01% 22.52% Intermediate Mean wealth quintile (1.23) (1.27) Owns property outright 61.57% 75.00% Does not own property outright 38.43% 25.00% Routine/manual Mean wealth quintile (1.29) (1.19) Owns property outright 60.74% 65.31% Does not own property outright 39.26% 34.69% (column %) The results in Table 5.12 show linear patterns among those who retire at normal retirement age, as occurs also in Table 5.11 with regards to NS-SEC category itself. As expected, those who retire from professional type jobs have the highest percentage of property owners, followed by those from intermediate work and finally by those from the manual category. The mean wealth scores follow the same pattern, with the highest wealth seen among those retiring from the professional category and the lowest among those from the manual group. Comparing later-life workers and retirees shows, as might be expected, a considerably higher percentage of retirees than workers own property outright, which suggests financial constraints due to property payments play some role in the decision to leave the workforce. The difference in the percentages of workers and retirees who own property also follows a linear pattern. A difference of 16.49% between property owners among professional workers is observable, 13.43% among intermediate workers, and 4.57% among manual workers. The findings here could be indicative of potential selection

185 184 bias in that the declining percentage of property owners by NS-SEC category among retirees may suggest factors alongside financial issues play a role in the decision to leave the workforce. Such factors may include poor health, forcing the worker into retirement when financial demands are still prevalent. Among both workers and retirees, wealth declines by NS-SEC category. Additionally, those who retire consistently have slightly lower mean wealth scores than those who work. This is likely to be due to the continuation of income among those who do not stop working once retirement age has been reached. Table 5.13: Health at SPA by NS-SEC category and employment activity (S.E.) Continues employment Retires Professional/managerial CES-D (1.42) (1.42) Self-rated health (0.93)** (0.97) Cognitive function (5.04) (5.52) Intermediate CES-D (1.44) (1.34) Self-rated health (0.89) (0.92) Cognitive function (5.31) (5.03) Routine/manual CES-D (1.80) (1.67) Self-rated health (0.93) (1.04) Cognitive function (5.36)** (5.52) ***p<0.001 **p<0.01 *p<0.05 Table 5.13 shows mean health scores by NS-SEC category. Firstly, it can be noted that the poorest scores are always among manual workers, within both working and retired categories. The best scores are consistently among those working in professional and managerial roles. In the case of cognitive function, a linear pattern can be observed in scores by NS-SEC category. The scores of those in both professional and intermediate employment are considerably higher than of those in manual work. This may be evidence of the use-it-or-lose-it hypothesis, with people participating in mindful work retaining higher cognitive function than those in work

186 185 which is manual and mentally non-taxing, yet it could also be the case that those individuals with the higher cognitive scores to begin with are those who are participating in the higher paid roles because they are able to Associations between employment status, qualification level and health The next consideration is of whether an individual s qualification level appears to be linked to health and employment activity at retirement age. Qualification level here is a binary variable, with O-levels or below compared to qualifications higher than O-levels. Table 5.14 presents employment status at retirement age by qualification level, and Table 5.15 shows health by both qualification and employment status. Table 5.14: Employment activity at SPA by qualification status Continues employment Retires Above o-level 289 (42.31%) [58.15%] 208 (43.24%) [41.85%] O-level or below 394 (57.69%) [59.07%] 273 (56.76%) [40.93%] (column %) [row %] Table 5.14 shows that among the sample overall, the percentage of individuals without qualifications higher than O-level standard is higher than those with. However, differences between the working and retired groups are small. A very slightly higher percentage of cases with higher qualifications continue to work than retire, and rates of later-life employment are similar within both qualification groups, with a higher percentage of workers in each instance. Again, two contrasting ideas may be at play here, with longer workforce participation linked to both higher qualification status by means of better roles, and lower qualification status by means of poorer quality work leading to lower levels of wealth and then less affordability of retirement. 37 Although the use-it-or-lose it hypothesis has commonly been referenced in literature concerning cognitive function and activity in later-life, recent research using the Lothian study has discredited its influence and has suggested that physical exercise may play a more important role in dementia protection (Gow et al. 2012; Deary et al. 2007).

187 186 Table 5.15: Health at SPA by qualification level and employment activity (S.E.) Continues employment Retires Above o-level CES-D (1.41) (1.34) Self-rated health (0.89) (0.95) Cognitive function (5.14) (5.61) O-level or below CES-D (1.71) (1.63) Self-rated health (0.95) (1.03) Cognitive function (5.42)*** (5.48) ***p<0.001 **p<0.01 *p<0.05 Table 5.15 shows health scores by qualification and employment status at retirement age. Here, patterns are visible. Depression is shown to be lower among those who retire than those who continue work, regardless of qualification level, although CES- D scores are noticeably lower among those with qualifications than without, which might be expected on the assumption that better qualifications are associated with better wealth, which in turn may link to better health. Figure 5.3 breaks this finding down by gender in order to visually observe how the association might differ here. As was seen previously, among men it is working which bears the greatest link with depression, while among women this changes to retirement. However, the differences in CES-D score are much greater among those who attained only O-level or lower qualifications. The CES-D scores of workers and retirees with higher qualifications are very similar among both men and women, and the scores of those with better qualifications are always lower than of those with poorer qualifications. This, again, complements ideas concerning a link between lower wealth and depression.

188 187 Table 5.15 shows self-rated health to follow the same pattern displayed in previous analyses, with workers always faring better than their retired counterparts. Better qualifications are also associated with better health, with even the retired cases here showing better scores than the working cases with lesser qualifications. Finally, cognitive function also follows previous patterns, with workers scoring higher than retirees. In this instance, the difference is significant. Scores are considerably higher among those with qualifications, regardless of employment activity. This may be suggestive of the aforementioned use-it-or-lose-it hypothesis, with scores declining once work has been exited. However, it is again unclear if working preserves cognitive function, or if those with higher cognitive function to begin, which may be reflected by qualification level, leads to continuation of higher cognitive scores across the life course.

189 Associations between employment status, other socially productive activities and health As well as considering the potential socio-demographic influences on health in accordance with employment status in later-life, the role of other socially productive activities can be examined. Studies outlined in the systematic review by McMunn (2009), Hao (2008), Choi (2007) and Adelmann (1994) demonstrate the ways in which participation in more than one socially productive activity, namely volunteering in conjunction with employment, can have a protective effect on health outcomes, particularly depression. Table 5.16 demonstrates numbers of later-life workers and retirees who participate in the two activities of interest to this study, volunteering and caring. Table 5.16: Other activities by employment status at SPA Continues employment Retires Volunteers 63 (9.20%) [45.32%] 76 (15.77%) [54.68%] Does not volunteer 622 (90.80%) [60.51%] 406 (84.23%) [39.49%] Provides care 97 (14.16%) [55.43%] 78 (16.18%) [44.57%] Does not provide care 588 (85.84%) [59.27%] 404 (83.82%) [40.73%] (column %) [row %] As might be expected, a greater proportion of retirees participate in voluntary work than non-retirees. This finding ties in with ideas surrounding continuity theory, discussed in Chapter 2, with participation in voluntary work providing a continuation of many of the aspects of life which would have existed within the workplace, such as social networks, feeling valued and carrying out meaningful tasks (Atchley, 1989). The proportion of caregivers among retirees is only slightly greater than that among workers. Again, this may be expected if providing care is unlikely to be a choice, but rather a necessary commitment in the familial setting. This idea is further strengthened by the higher proportion of people who remain in the workforce and provide care as opposed to voluntary work. Table 5.17 shows mean health scores according to activity status. Bearing in mind the significant protective effect of volunteering in conjunction with paid employment

190 189 demonstrated in the systematic review (Chapter 2), it is interesting to note that the lowest depression scores here are among the cases for who this is the case. Looking at depression scores on the basis of employment or retirement alone (Table 5.6), employment has an association with higher levels of depression than retirement, yet the opposite is true when voluntary work is added into the equation. CES-D scores of individuals who do not volunteer are always higher than those who do. This could be, yet again, suggestive of a protective effect of volunteering on mental health, although this assumption must be treated with caution as it could also be the better baseline mental health score leading to an increased likelihood of participation in voluntary work in the first place. Self-rated health scores are also always better among volunteers, and again, this finding may be particularly subject to selection bias in that those who have the better health to begin with are more likely to volunteer due to the fact alone that they are able to. Cognitive function scores are very similar among all cases apart from those who retire and do not volunteer (so take part in no socially productive activities). Again, however, it is difficult at this stage to ascertain whether this is due to a protective effect on cognition of socially productive roles, or if higher cognitive ability leads to an increased likelihood of participating in any work, whether it is paid or voluntary in nature. Table 5.17: Health at SPA by other activities and employment status (S.E.) Continues employment Retires Volunteers CES-D (1.00) (1.18) Self-rated health (0.77) (0.74) Cognitive function (4.50) (5.88) Does not volunteer CES-D (1.63) (1.57) Self-rated health (0.93)** (1.02) Cognitive function (5.49)*** (5.61) Provides care CES-D (1.38) (1.52) Self-rated health (0.87) (0.99) Cognitive function (4.97) (5.62)

191 190 Table 5.17: Health at SPA by other activities and employment status (S.E.) Continues employment Retires Does not provide care CES-D (1.62) (1.51) Self-rated health (0.94)* (0.99) Cognitive function (5.45)*** (5.69) ***p<0.001 **p<0.01 *p<0.05 T-tests of mean scores between workers and retirees are only significant for selfrated health and cognitive function among non-volunteers, so again it is the people who might have the poorer background and baseline characteristics who appear to have the stronger association between these factors and health scores. Table 5.17 also shows an association between caring responsibilities and increased rates of depression. Again, such a result might be expected on the assumption that caring duties are mandatory and a burden, rather than preferential. The worst scores are among those who retire and provide care, and it is this group who might have the least respite from caring responsibilities in the form of workplace activity. Among self-rated health scores, the worst are given to those who are retired and do not provide caring duties. This, once again, could be due to selection bias in part, in that individuals who were physically unable to continue employment may be also incapable of caring duties. Cognitive function is again always higher among workers than retirees, and the difference in scores is significant among non-caregivers Associations between employment status, ADL score and health The final subsection here considers ADL score in relation to employment status and health. Table 5.18 shows information concerning ADL score on the basis of employment status. Mean ADL scores by employment status are provided, alongside frequencies of workers and retirees on the basis of a binary ADL variable, where a value of 1 depicts those with one or more difficulty, and 0 those with none.

192 191 Table 5.18: ADL score by employment status at SPA (S.E.) Continues employment Retires Mean ADL score (0.58)* (0.859) Binary ADL variable 38 No ADL difficulties 626 (91.39%) [59.96%] 418 (86.72%) [40.04%] 1 ADL difficulty 59 (8.61%) [47.97%] 64 (13.28%) [52.03%] ***p<0.001 **p<0.01 *p<0.05 As might be expected, the first row of Table 5.18 shows the mean ADL score of retirees to be almost twice that of those who continue employment at retirement age. This is likely to be reflecting the increased likelihood of people with poorer physical functioning ability to exit the workforce once they are able. However, it could also be reflective of a protective effect on basic physical functioning of continuing employment in contrast to exiting employment. The second part of the table shows that, although numbers of individuals with one or more ADL difficulty are always low, only 8.61% of workers fit into this category, compared to 13.28% of retirees. This, again, may reflect both the increased likelihood of leaving work if physical functioning is poorer, as well as potential protective effects of work on physical functioning. Table 5.19: Health at SPA by ADL score and employment activity (S.E.) Continues employment Retires No ADL difficulties CES-D (1.56) (1.35) Self-rated health (0.89) (0.94) Cognitive function (5.35)** (5.59) 1 ADL difficulty CES-D (1.85) (2.05) Self-rated health (1.01) (1.00)

193 192 Table 5.19: Health at SPA by ADL score and employment activity (S.E.) Continues employment Retires Cognitive function (6.02) (6.28) ***p<0.001 **p<0.01 *p<0.05 Mean CES-D scores are always lower among individuals without ADL difficulties than with. The worst depression scores are observable among those who retire and suffer ADL difficulties. This is interesting in reference to Table 5.6, which suggests the general association between retirement and depression, without consideration of any potential confounders, is for better scores than employment. This may suggest a beneficial link between work and mental health among those with poorer physical ability. The strongest association with ADL score is with self-rated health, where scores are always considerably worse among those who do experience ADL difficulties. Again, the absolute worst scores can be found among retired people with one or more ADL difficulty, which may suggest the same beneficial effect of employment for people with poorer physical functioning, but is more likely suggestive of the fact that the people with the poorest health are leaving the workforce and entering retirement as soon as they are financially able to. Finally, cognitive function scores follow a similar pattern, with scores always better among workers, and the poorest scores given to those who are retired alongside suffering from ADL difficulties.

194 Work Characteristics of the selected sample The previous section of the chapter considered differences in the socio-demographic characteristics of workers and retirees. The following section examines work-related characteristics of the selected sample, with regards to how such factors might affect propensity to enter later-life employment, as well as how these characteristics might interact with health Associations between self-employment and health Tables 5.1 and 5.2 showed both the ELSA and selected samples see an increase in self-employed individuals across time. Table 5.20 presents information on whether individuals retire or continue employment on the basis of whether or not they are self-employed. A slightly higher percentage of those who retire are self-employed at the wave prior to reaching normal retirement age (86.3%, compared to 81.3% of later-life workers). Table 5.20: Employment activity by self-employment status Continues employment Retires Employed at wave prior to reaching retirement age Self-employed at wave prior to reaching retirement age 557 (81.31%) 416 (86.31%) 128 (18.69%) 66 (13.69%) Employed at retirement age 552 (80.58%) - Self-employed at retirement age 133 (19.42%) - (column %) Table 5.21 shows health by self-employment status. Although those who are selfemployed show slightly better outcomes across all three measures, a t-test of the results show none are significantly different. Therefore, we can conclude that, although self-employment seems to have some association with a slightly increased likelihood of retirement, it has no real association with improvement in health within this particular sample.

195 194 Table 5.21: Health at SPA by self-employment status (S.E.) CES-D Self-rated health Cognitive function Employed (1.60) (0.95) (5.43) Self-employed (1.54) (0.84) (5.29) ***p<0.001 **p<0.01 *p< Associations between employment type and health In addition to the NS-SEC categorization of employment type, the ELSA dataset provides alternative information regarding the type of employment, with a threecategory variable denoting sedentary, standing and manual type of work. Table 5.22 shows the numbers and percentages of individuals working in each type of employment at both the wave prior to, and the wave of reaching retirement age by employment activity at retirement age. Table 5.22: Employment activity at SPA by employment type Continues employment Retires Sedentary employment at wave prior to reaching retirement age Standing employment at wave prior to reaching retirement age Manual employment at wave prior to reaching retirement age Sedentary employment at retirement age Standing employment at retirement age Manual employment at retirement age 280 (41.73%) 155 (33.41%) 215 (32.04%) 165 (35.56%) 176 (26.23%) 144 (31.03%) 283 (42.81%) (33.28%) (23.90%) - (column %)

196 195 The results in Table 5.22 suggest differences exist in the types of employment of those who continue working and those who retire once retirement age is reached. Among retirees, type of employment is reasonably evenly distributed, whereas among those who continue to work, both before and after retirement age, there are clear differences in the percentages of people across groups. As might be expected, the highest proportion of workers participate in sedentary employment, and the fewest in manual work. Table 5.23 shows health by employment type. Table 5.23: Health at SPA by employment type (S.E.) CES-D Self-rated health Cognitive function Sedentary (1.33) (0.90) (5.24) Standing (1.70) (0.97) (5.37) Manual (1.71) (0.87) (5.72) ***p<0.001 **p<0.01 *p<0.05 CES-D score is markedly better among those in sedentary employment, with very similar scores demonstrated between those in standing and manual employment. Again, self-rated health is better among those in sedentary work, with a smaller difference in the scores of those who participate in standing and manual types of employment. Cognitive function is also better among sedentary workers, and the lowest scores can be found among those in manual work Associations between hours worked, employer and health The ELSA data provide information regarding the number of hours worked by respondents, as well as whether or not they work for the same employer as they did at the wave prior to interview. Both pieces of such information are important to consider with regards to understand the working patterns of the sample of interest to the study. A change in working hours or employer may signify a move from an individual s main career into bridge employment, as was discussed in Chapter 2. Due to low sample numbers, missing data and collinearity issues, it was not possible to include these two variables in the later adjusted models. However, they provide a useful insight into the working behaviour of the sample.

197 196 ELSA provides a variable regarding the number of hours worked. The original variable is continuous, but has been recoded for the purpose of this study into a binary variable where 1 depicts full-time work and 0 depicts part-time work. Here, 30 or more hours per week constitutes full-time employment, and less than 30 constitutes part-time employment. Table 5.24 shows whether or not an individual continues work or retires at SPA according to the number of hours worked at the wave prior to reaching retirement age. Table 5.24: Employment status by hours worked prior to SPA and hours worked at SPA Continues work Retires Works full-time at the wave prior to reaching SPA Works part-time at the wave prior to reaching SPA Works full-time at the wave of reaching SPA Works part-time at the wave of reaching SPA 357 (65.18%) [59.30%] 245 (61.71%) [40.70%] 184 (34.014%) [54.76%] 152 (38.29%) [45.24%] 265 (50.00%) (50.00%) - (column %) [row %] The column percentages in Table 5.24 show 65.18% of respondents who enter laterlife employment work full-time hours at the wave prior to reaching SPA. However, at the wave of entering retirement age, this has reduced to 50%. The percentage of retirees who leave the workforce from full-time work is similar to that of those who continue employment at 61.71%. The row percentages show 59.3% of respondents in full-time work prior to reaching SPA continue work, compared to 54.76% of those previously participating in part-time work. Table 5.25 shows mean health scores according to whether an individual works fullor part-time. The self-rated health and cognitive function scores of those in part-time work are slightly better than those in full-time work, but the opposite is true for depression. However, in each instance, the difference in scores are small and not significant.

198 197 Table 5.25: Mean health scores at SPA by hours worked (S.E.) Works full-time Works part-time CES-D (1.61) (1.55) Self-rated health (0.97) (0.91) Cognitive function (5.38) (5.74) One of the key ideas surrounding continuity and activity theories of ageing is that of bridge employment; whereby an individual may leave their career in order to participate in lesser paid roles before leaving the workforce entirely (Feldman 1994). Here, we might expect to see these individuals moving from full-time work to parttime work between waves. Table 5.26 shows hours worked before SPA by hours worked at SPA. Results are confined to only those cases who continue employment in this instance. Table 5.26: Hours worked prior to SPA by hours worked at SPA Full-time hours at SPA Part-time hours at SPA Full-time hours prior to SPA 253 (97.31%) [74.41%] 87 (33.72%) [25.59%] Part-time hours prior to SPA 7 (2.69%) [3.93%] 171 (66.28%) [96.07%] (column %) [row %] As would be expected, Table 5.26 demonstrates a very small number of individuals move from part-time work prior to retirement age into full-time work at retirement age (2.69%). Among the 253 cases working full-time in later-life employment, the large majority were working the same hours previously. The row percentages show around a quarter of the sample who are working full-time at the wave prior to reaching retirement age move into [part-time work thereafter. The ELSA data also provides information concerning whether or not the respondent works for the same employer as they did at the wave prior to interview. While continuation of work might be more likely among individuals who can remain with the same employer, a transition to bridge employment prior to retirement is more likely to be associated with a change of employer. Table 5.27 shows whether or not

199 198 respondents work for the same employer at the wave of reaching retirement age as they did at the previous wave. Table 5.27: Working for the same employer at the wave prior to reaching SPA Works for same employer as at the wave prior to reaching SPA Works for a different employer to the wave prior to reaching SPA 478 (88.68%) 61 (11.32%) Table 5.27 shows 88.68% of the sample continuing to work at SPA do so with the same employer they worked for at the wave previously. Table 5.28 shows health outcomes according to change in employment. Table 5.28: Mean health scores at SPA by employer at the previous wave (S.E.) Works for the same employer Works for a different employer CES-D (1.61) (1.39) Self-rated health (0.94)* (0.86) Cognitive function (5.47) (5.96) ***p<0.001 **p<0.01 *p<0.05 In each instance, health scores are better among the workers who have changed employer at reaching retirement age, although this difference is only significant for self-rated health. If these respondents are considered likely to be participating in bridge employment, the pattern of health scores relates to aforementioned ideas that bridge employment is often taken on the basis of positive retirement and workplace attributes (Feldman 1994), and subsequently leads to better health outcomes (Zhan et al. 2009).

200 Associations between work quality and retirement, sociodemographic characteristics and health This chapter has already examined some of the basic relationships between adjusted health scores and socio-demographic characteristics on the basis of whether an individual continues to work or retires at SPA. However, a second key focus of this study concerns work quality post-retirement age. The analysis presented in Chapter 7 will use propensity score matching to examine differences in health scores between those in poor and good quality employment. Subsequently, effects of poor and good quality employment will be compared to effects of retirement. Firstly, however, this section of the chapter provides some descriptive statistical information on the same health and socio-demographic variables discussed in section 5.2, in order to examine associations between these and an individual s quality of work. Work quality is defined by the recoded effort-reward ratio variable which was outlined in Chapter 4. The sample used within this section of the chapter focuses on 578 individuals who provide information regarding work quality, with 109 in poor quality employment and 469 in good. The retired group to which they are compared consists of the same 482 individuals included in the previous descriptive analyses, who exited the workforce at normal retirement age. To test the group means for significant differences, one-way ANOVA tests were performed across the three categories of interest (poor quality work, good quality work and retirement) Associations between work quality and retirement and health Table 5.29 shows rates of poor and good quality later-life employment and retirement by work quality measured at the wave prior to that at which each individual reaches SPA. Work quality was not recorded at wave 1 in the ELSA data, and so those who reach retirement age at wave 2 here are not included in the frequencies and percentages.

201 200 Table 5.29: Employment activity at SPA by work quality Poor quality Good quality Retirement employment employment Poor work quality at 49 (54.44%) 45 (16.98%) 93 (32.07%) wave prior to reaching SPA [26.20%] [24.06%] [49.73%] Good work quality at 41 (45.56%) 220 (83.02%) 197 (67.93%) wave prior to reaching SPA [8.95%] [48.03%] [43.01%] (column %) [row %] As might be expected, Table 5.29 shows an apparent link between work quality and the decision to remain in work or retire once retirement age is attained. The row percentages show 49.73% of people who retire were in poor quality employment at the wave prior to reaching retirement age, compared to 43.01% who were in good quality work. In terms of those who are still working, 74.65% in total are working in good quality employment, and 73.52% were in good quality employment before reaching retirement age 39. Only 35.47% of all workers belonged to poor quality employment at the previous wave. It is also notable that, by retirement age, 24.06% of people who were in poor quality work at the previous wave had managed to move into good quality work (47.87% of all cases who continued employment). As might be expected, the percentage of people shifting into poor quality work from good is much lower, at 8.95% of all cases belonging to good quality work prior to retirement age (15.71% of all cases who continued employment). However, the majority of people who continue into poor quality later-life employment do so from poor quality employment at the previous wave (54.44%), and 83.02% of people who continue into good quality work do so from good quality work previously. Table 5.30 shows the basic relationship between work quality and the three key health outcomes of interest to the study. P-values for the F statistic on three degrees of freedom, produced by means of one-way ANOVA, are presented in the final 39 Percentages here are of the total of the 355 workers included in table 5.24 (those currently in poor or good quality employment from both poor and good quality employment at the wave prior to reaching retirement age).

202 201 column of the table, and demonstrate whether significant differences in health scores exist on the basis of group membership. Table 5.30: Health at SPA by work quality/retirement (S.E.) Poor quality work Good quality Retirement P-value (one- work way ANOVA) CES-D (1.97) (1.23) (1.52) Self-rated health (1.05) (0.86) (0.99) Cognitive function (5.64) (4.97) (5.70) Table 5.30 shows a hierarchy of scores with regards to depression and self-rated health, with the poorest scores among poor quality workers, the middle scores among retirees and the highest among good quality workers. Mean scores are significantly different on the basis of group membership for both outcomes. The closer mean self-rated health scores of poor quality workers and retirees is interesting in that it might suggest this group is comprised of people with poorer health who can either afford to leave the workplace and retire, or who suffer poor health but cannot afford to leave. These ideas are discussed in greater detail further on in the chapter. The pattern differs for cognitive function scores. Here, retirement is associated with the lowest scores, which might lend to the use it or lose it hypothesis in that it is those who have left the potentially protective sphere of the workforce who have performed worst in the tests. As would be expected, the best scores are among the good quality workers, who may be engaged in a higher level of mentally challenging activity, but who also might have a higher cognitive ability to begin with, which has subsequently driven the individual to the better work throughout the life course. Again, differences in scores are significantly different on the basis of group membership.

203 Associations between work quality and retirement, gender and health Section 5.2 of the chapter established that, when looking at later-life employment in comparison to retirement, large differences exist on the basis of gender. This is particularly apparent with regards to CES-D score, with later-life employment among men associated with increased depression, but among women with less. Table 5.31 shows frequencies of work quality by gender. Table 5.31: Work quality/retirement by gender Poor work quality Good work quality Retirement Male 34 (24.29%) 105 (25.80%) 228 (47.30%) Female 106 (75.71%) 302 (74.20%) 254 (52.70%) (column %) Stratifying work quality by gender shows the experiences of later-life employment differs very little between men and women on the basis of work quality, and as yet does not offer any explanation as to why health, particularly depression, appears to vary so widely by gender. As might be expected, the vast majority of both poor and good quality workers are female, which is due to the higher proportion of females in the sample overall (Tables 5.1 and 5.2) and the higher number of women who continue to work than retire (Table 5.4). Table 5.32 examines mean health scores by both work quality and gender. Table 5.32: Health at SPA by gender and work quality/retirement (S.E.) Poor work Good work Retirement P-value quality quality (one-way ANOVA) Male CES-D (1.88) (1.04) (1.19) Self-rated health (1.24) (0.81) (1.00) Cognitive function (4.71) (4.87) (5.47) Female CES-D (2.01) (1.28) (1.72) 0.000

204 203 Table 5.32: Health at SPA by gender and work quality/retirement (S.E.) Poor work Good work Retirement P-value quality quality (one-way ANOVA) Self-rated health (0.99) (0.87) (0.99) Cognitive function (5.36) (4.89) (5.60) Table 5.32 shows that, for all three health outcomes, the hierarchy of scores observed previously still exists, with those in retirement seeing the middle scores in each instance, those in poor quality work seeing the worst and those in good quality work seeing the best. Differences in the cognitive function scores of men are particularly defined by such a manner, which may suggest the relationship between employment and retaining cognitive function might be more prominent among males than the population overall. Again, significant differences exist consistently across all outcomes with the exception of cognitive function among women. The same hierarchy of scores persists among women, but this time the differences in scores are smaller. In each instance, the scores of retirees and poor quality workers are much closer than in the case of males. It is also important to note that the initial relationship observed between depression, working and retirement on the basis of gender in section 5.2, with later-life working associated with greater depression than retirement for men, but the opposite to be true for women, disappears once work is stratified by quality. Instead, the same pattern hold for both genders, with good quality work associated with the fewest symptoms of depression, retirement in the middle, and poor quality work associated with the most. Finally, it is worthwhile to note that self-rated health and cognitive function scores are consistently better among women than men, across all three comparison groups. Here, it is again important to remember the age effect which will be confounding scores, as all women contributing to the data hereby presented are five years younger than the men.

205 Associations between work quality and retirement, sociodemographic characteristics and health This section of the chapter looks at the basic relationships that appear to exist between quality of work and retirement, health and the socio-demographic characteristics which seemed to play an important role in determining group membership or health scores in section 5.2 of the chapter: wealth, property ownership, socio-economic classification and qualification level. Table 5.33 shows mean scores and percentages for each variable at the wave of reaching retirement age by the two work quality groups and retirement. Table 5.33: Socio-demographic characteristics by work quality/retirement at SPA (S.E.) Poor work quality Good work quality Retirement Mean wealth (1.24) (1.33) (1.28) Owns property 59.29% 63.70% 71.78% NS-SEC professional/managerial 23.57% 36.36% 31.52% NS-SEC intermediate 32.86% 29.48% 27.56% NS-SEC routine/manual 43.57% 34.15% 40.92% Qualification(s) above 35.00% 44.83% 43.24% O-levels The mean wealth scores show the highest wealth exists among those in good quality work, and the lowest among those in poor. Those who are retired are in the middle, and this possibly represents a group of individuals who no longer have optimal wealth due to exiting the workforce, but who also have enough money to be able to afford retirement. As might be expected, the highest proportion of property owners is among the retirees, again suggestive of the fact that this group is able to afford to retire and no longer have the financial burden of mortgage or rent payments. Additionally, the lowest percentage of property owners is among those in poor quality work, again suggesting this group is remaining in poor quality employment out of necessity. Although the percentages of those in poor quality work follow a

206 205 linear pattern across NS-SEC groups, there is a far more even distribution among those in good quality work. The higher number of people retiring from professional and managerial type employment may be reflective of an ability to be able to afford to leave the workforce. Finally, as would again be expected, a much higher percentage of people in good quality work have higher qualifications than those in poor quality work. Once more, the similar percentage of retirees with higher qualifications may be reflecting affordable departure from higher paid work. Table 5.34 considers each of the socio-demographic variables in relation to health alongside work quality. Wealth has been divided into optimal and sub-optimal measures as in section 5.2, with optimal wealth denoted by those in the wealthiest two quintiles of the scale and suboptimal by those in the lowest three. Appendix 5 again presents the tables of ANOVA tests of between group significance, which will hereby be discussed. Table 5.34: Health at SPA by socio-demographic characteristics and work quality/retirement (S.E.) Poor quality work Good quality work Retirement P-value (one-way ANOVA) CES-D Optimal wealth (1.19) (1.30) score (1.30) Suboptimal wealth (2.21) (1.30) (1.65) 0.000*** Owns property (1.63) Does not own property (2.32) Professional/managerial (1.64) Intermediate (2.11) Routine/manual (2.03) (1.23) (1.51) 0.000*** (1.17) (1.53) 0.000*** (1.17) (1.42) 0.013* (0.85) (1.34) 0.000*** (1.51) (1.67) 0.018* Above O-levels (1.03) (1.34) 0.000***

207 206 Table 5.34: Health at SPA by socio-demographic characteristics and work quality/retirement (S.E.) Poor quality work Good quality work Retirement P-value (one-way ANOVA) (1.64) Self-rated health O-levels or lower (2.13) Optimal wealth (0.91) Suboptimal wealth (1.05) Owns property (0.95) Does not own property (1.17) Professional/managerial (0.95) Intermediate (1.08) Routine/manual (1.06) Above O-levels (0.96) O-levels or lower (1.11) (1.38) (1.63) 0.000*** (0.84) (0.87) (0.89) (1.02) 0.000*** (0.88) (0.96) (0.82) (1.04) 0.000*** (0.90) (0.97) 0.009** (0.81) (0.92) (0.82) (1.04) (0.81) (0.95) 0.006** (0.89) (1.03) 0.039* Cognitive Optimal wealth (5.54) function (5.90) (5.23) Suboptimal wealth (5.43) 0.000*** (5.27) (4.50) Owns property (5.76) 0.002** (5.82) (5.15)

208 207 Table 5.34: Health at SPA by socio-demographic characteristics and work quality/retirement (S.E.) Poor quality work Good quality work Retirement P-value (one-way ANOVA) Does not own property (5.37) (4.61) (5.54) 0.001** Professional/managerial (5.82) Intermediate (5.39) Routine/manual (5.23) Above O-levels (5.55) O-levels or lower (5.56) ***p<0.001 **p<0.01 *p< (4.87) (5.13) (4.69) (5.06) (4.77) (5.52) (5.03) (5.52) 0.000*** (5.61) (5.48) 0.000*** The health scores presented in Table 5.34 show that, in all but one instance, those in good quality work fare better than both those in poor quality work and retirees. The exception to this is the case of self-rated among those with optimal wealth, where the best score is among those in poor quality employment. This could be due to this group remaining in work because they want to, rather than because they cannot afford not to, and if under the assumption that poor quality work is more likely to be manual in nature, this could be evidence of a protective effect of physical employment on health. In the cases of depression and self-rated health, in almost all instances the poorest scores are among those in poor quality employment. The first exception here is among those with optimal wealth, where retirement is associated with the highest levels of depression. This may reflect a group of people who have left good jobs without wanting to, perhaps for reasons of poor health or having to provide caring duties. The second exception is among those in good quality employment and optimal wealth, where self-rated health is at its worst. If it is considered that good

209 208 quality work is mainly sedentary in nature, this may hereby be observation of a group of individuals who continue to enjoy employment and are able to do so despite poorer health, when their counterparts in poorer quality work would be forced to exit the workforce. Although the worst depression and self-rated health scores are almost always observed among those in poor quality employment, it is interesting to note that the poorest cognitive function scores are always observed among retirees, with the exception of those who retired from managerial roles. It should also be noted that those who retired from managerial positions have the best cognitive function scores of all retirees, which may suggest baseline cognitive function had driven these individuals into the higher status employment to begin with. The poorer cognitive function among retirees in general may reflect the absence of a protective effect of work once it has been left. However, it must also be remembered that the less favourable types of employment, participation in which may be driven by lower cognitive function to begin with, are likely to be associated with an increased likelihood to retire anyway. Considering differences in health between socio-demographic groups, it is notable that scores are almost always better among those with favourable characteristics. Those with optimal wealth always have better scores than those with suboptimal wealth across all comparison groups. Health is better among property owners in all instances, other than the cases of depression and self-rated health scores among those in good quality work. Here, non-owners fare slightly better. Routine and manual employment is consistently associated with poorer health than both intermediate and managerial. In most instances, managerial work shows an association with the best health outcomes. Where intermediate work sees better scores, this may be attributable to an absence of increased role responsibility leading to lower levels of stress. Finally, higher qualifications are also consistently associated with better health outcomes, a fact which again suggests a complex relationship exists between wealth, job type and health. Finally, p-values of the F statistic produced by means of one-way ANOVA are presented in the final column of the table. Unlike the analyses of workers in general compared to retirees in section 5.2, whereby few significant differences in mean

210 209 health scores were observed, scores are frequently significantly different on the basis of group membership. This is particularly noticeable with regards to depression, where group mean scores are always significantly different with the exception of respondents with optimal wealth Associations between work quality and retirement and other activities As in section 5.2 of the chapter, it is also useful to look at other activities in relation to work quality, retirement and health outcomes. Table 5.35 shows frequencies of those who volunteer and provide care by employment quality and retirement. Table 5.35: Other activities by work quality/retirement at SPA Poor work quality Good work quality Retirement Volunteers 9 (6.43%) [7.09%] Does not volunteer 131 (93.57%) [12.78] Provides care 17 (12.14%) [10.69%] Does not provide care 123 (87.86%) [14.14] (column %) [row %] 42 (10.32%) [33.07%] 365 (89.68%) [41.29] 64 (15.72%) [40.25%] 343 (84.28%) [39.43] 76 (15.77%) [59.84%] 406 (84.23%) [45.93] 78 (16.18%) [49.06%] 404 (83.82%) [46.44] The figures in Table 5.35 show, as might be expected, higher rates of volunteering and care-giving take place among retirees than either those in poor or good quality employment. Additionally, much higher rates of other socially productive activities are carried out among those in good quality work than poor % of people in good quality work volunteer, compared to just 6.43% of people in poor quality work. A higher percentage of individuals in good quality work provide care, although the difference between the groups here is smaller (15.72%, compared to 12.14%). It could be assumed that, among workers in general, caring duties are carried out due to necessity, which may be reflected by the higher proportions of caregivers than volunteers. Subsequently, the lower percentage of caregivers among poor quality workers may be reflective of poorer health among this group, with demanding and

211 210 poor quality work being exited in order to sufficiently provide those care duties. The row percentages show the same pattern, with the highest proportion of both volunteers and caregivers in retirement, followed by good quality work and then poor. Due to the low sample numbers in the volunteering and caring categories, examination of health by participation in other activities and work quality is not feasible. However, section provided an overview of health and participation in voluntary and care work among workers, as well as retirees, in general Associations between work quality and retirement, ADL score and health The final descriptive section will again consider ADL score and its relationship with work quality and health. Table 5.36 shows mean ADL score by work quality and retirement, alongside a binary measure of ADL which allows examination of those with no ADL difficulties in comparison to those who classify themselves as having one or more. Table 5.36: ADL score by work quality/retirement at SPA (S.E.) Poor work quality Good work quality Retirement Mean ADL score (0.68)** (0.51) (0.86) Binary ADL variable 40 No ADL difficulties 119 (85.00%) [12.93%] 1 ADL difficulty 21 (15.00%) [19.27%] **p<0.01 (column %) [row %] 383 (94.10%) [32.93%] 24 (5.90%) [22.02%] 418 (86.72%) [45.43%] 64 (13.28%) [58.72%] The mean ADL score among those in poor quality work is over twice that of the mean among those in good quality work. The mean score of those who retire is very similar to, but slightly lower than, those in poor quality work. Considering the binary ADL score variable, the column percentages demonstrate that among good quality workers, there is a much smaller percentage of individuals with at least one ADL 40 (column %); [row %]

212 211 difficulty than among poor quality (5.90% compared to 15%). Again, a higher percentage of poor quality workers suffer an ADL difficulty than retirees. Assuming retirees might exit the workforce due to a reduced ability to remain within it, it might be expected that the opposite would be true. However, this finding may be evidence of the aforementioned idea of job lock, with individuals remaining in work because financially they have to, rather than because they can do so due to better health. Looking at the row percentages, it is noticeable that the highest proportion of individuals with one or more ADL difficulty are retired, while the lowest is in poor quality work, again suggesting selection effects are present. The distribution of people in poor and good quality work is more evenly spread among cases with one or more ADL difficulty than among cases with no difficulties, where a much higher proportion remain in good quality work than poor. Again, this is likely to be evidence of confounding. If poorer health is associated with lower levels of wealth, which in turn is associated with poorer work quality, it might be expected that a number of cases with lower physical functioning ability cannot afford to retire from less favourable jobs. Table 5.37 provides further evidence of such an idea. Mean wealth scores are consistently higher among those in good quality work, regardless of ADL score, and the lowest wealth is observable among those in poor quality work with poorer physical functioning. Table 5.37: Mean wealth at SPA by ADL score and work quality/retirement (S.E.) Poor work quality Good work quality Retirement No ADL difficulties (1.19) (1.33) (1.28) 1 ADL difficulty (1.47) (1.44) (1.18) Again, due to low numbers of cases with ADL difficulties in this sub-section of the analysis, it is not possible to carry out reliable analysis of health scores on the basis of this variable in relation to work quality.

213 Overview of chapter Overview of the sample characteristics The results presented throughout this chapter are unadjusted and caution must be taken when assuming any associations exist between the variables of interest. However, they provide a useful overview of the sample selected for the forthcoming analyses, and the early observation of potentially confounding factors is an important aspect of this particular study, The first section of the chapter outlined the importance of understanding the key differences between the sample selected for the forthcoming analyses and the overall ELSA population, which is representative of the general English population of people aged 50 and over. It is especially important to consider these differences in light of the low external validity produced by use of matching techniques within strictly confined samples. Through doing so, a greater understanding of the populations to which findings can be generalised can be achieved. Tables 5.1 and 5.2 showed several important differences exist between the overall ELSA sample and the sample of interest to the study. The proportion of females is higher, and the mean age is lower, due to inclusion of cases on the basis of reaching retirement age in the study period. Additionally, the selected sample has a higher proportion of married individuals, is slightly wealthier, has a lower percentage of manual workers than the overall population, a higher number of people with qualifications above O-Level and a lower number of people who own their property outright. In terms of health, the selected sample has fewer ADL difficulties, fewer symptoms of depression, better self-rated health and better cognitive function than the main ELSA sample. Tables throughout the chapter provided information on characteristics of workers and retirees from retirement age onwards. In comparison to retirees, later-life workers were more likely to be female, and women were more likely to work further beyond retirement age than men. Compared to later-life workers, both men and women who retired were more likely to have a partner who no longer belongs to the workforce, but there was little evidence of an association between having a working partner and opting to continue to work. Compared to those who continued work, those who retired were also more likely to be poorer, and individuals with greater wealth were more likely to continue to work than retire. However, workers were less

214 213 likely than retirees to be property owners, and a higher proportion of non-property owners continued to work than retired. As expected, those who retired were most likely to be from manual and routine occupations, less likely to have qualifications higher than O-Levels and more likely to be involved in caring duties or voluntary work than those who continued to work. Such findings suggest later-life employment to be associated with better wealth and better working conditions However, they also suggest retirement occurs on the basis of financial stability, as well as on the basis of leaving unfavourable working conditions. When work was stratified by quality and then compared to retirement, the complex nature of relationships continued to be observed. Good quality work was consistently associated with better wealth, better qualifications and professional and managerial occupation types, and poor quality work was associated with the lowest levels of wealth, the lowest rates of home ownership and the highest rate of respondents in manual jobs. Characteristics of retirees were typically closer to those of the poor quality workers than the good quality workers. Again, a much higher proportion of retirees owned their own homes than people in either good or poor quality work. This is perhaps the strongest indication of the potential of financial pressure to lead to continued employment in later-life. Section 2.3 considered the workplace characteristics of the selected sample with regards to self-employment, type of employment, hours worked and whether or not respondents continued to work for the same employer at retirement age as the at the previous wave. Findings here demonstrated later-life workers were slightly more likely than retirees to have been self-employed at the wave prior to reaching SPA and slightly more likely to have been working full time prior to reaching SPA. Approximately a quarter of later-life workers moved from full-time employment before SPA to part-time employment at SPA. The number of respondents who moved from part-time to full-time work at SPA was very small. Differences in health scores on the basis of these characteristics were not significant except in the case of self-rated health, where scores were worse among those who did not change employer at SPA compared to those who did. As stated beforehand, low sample numbers as well as issues of collinearity and missing data means the forthcoming analyses cannot account for these aspects of employment. These low numbers are

215 214 due to the strict sample selection process which was enforced in order to ensure only those who worked directly until SPA were to be included in this particular study Overview of the relationships between health and employment Initial crosstabs of health scores by employment status showed self-rated health and cognitive function to be significantly better among later-life workers, and depression to be non-significantly better among those who retired. A breakdown of scores by gender showed self-rated health and cognitive function to be consistently better among women than men, which was most likely due to the younger age of women included in the sample of the basis of SPA. The relationship with depression was more intricate, and higher rates were associated with continued employment for men yet retirement for women. Patterns of health on the basis of socio-demographic characteristics were as had been expected, with scores always better among those who were wealthier, property owners, in higher status occupations and who held higher qualifications. Differences in scores on the basis of these characteristics between workers and retirees often suggested that issues of job-lock were present, with the highest levels of depression evident among those who continued to work in poorer conditions, with lower wealth and non-property ownership. Conversely, depression scores were low among workers with characteristics suggestive of better wealth and working conditions. When the relationship between health and employment was stratified by work quality, a hierarchical effect became apparent. Depression and self-rated health scores were best among those in good quality work and worst among those in poor. Scores of retirees were between the two. Cognitive function was poorest among retirees, which again may be due to loss of protective effects of work after workforce exit, or to intricate relationships between socio-demographic factors associated with lower cognitive function leading to an increased likelihood of workforce exit. In terms of self-rated health and cognitive function, the scores of cases in poor quality work and retirees were closer than those of people in good quality work, suggesting again a close relationship exists between socio-demographic factors, employment type and health outcomes. Finally, cognitive function displayed a much stronger pattern with employment for men than for women, with males in poor quality work showing substantially lower scores than any other group.

216 Heterogeneity, selection bias and the forthcoming analyses Across all health outcomes, and especially in the case of depression, health scores were consistently worse among workers with poorer socio-demographic characteristics. Scores were better among working cases who displayed more favourable socio-demographic characteristics. Stratifying employment by quality strengthened these relationships, with poor quality work associated with the poorest outcomes and good quality with the best. Throughout the chapter, complex relationships between work quality, socio-demographic characteristics and health were evident. Although only an initial analysis, the emergence of significant differences between work quality groups, and between these separate groups and retirement, tells us that employment as a comparison to retirement in later-life is too heterogeneous a concept. The fact that mean health scores of those in retirement lie between those of poor and good quality workers in almost all instances and across all socio-demographic groups demonstrates this further, and provides insight into why very few significant differences were observable when all workers were grouped together into one category. The work carried out in this chapter is important in that it supports the ideas presented by the systematic review and meta-analysis. Not only are there strong differences within the sample on the basis of their characteristics determining the decision to continue to work or retire, but the key definition of employment in itself, is too heterogeneous a concept to focus on in one generalised analysis. The following analyses will employ methodology which aims to examine these potential sources of selection bias in greater detail, ultimately enabling examination of whether or not any causal effects of employment, and quality of employment, really exist.

217 216 Chapter 6: Health Trajectories and Later-Life Employment using Spline Models This chapter uses piecewise spline regression techniques to examine trajectories of health outcomes over the normal retirement age period. The chapter asks the key research question of whether or not patterns of health differ in the periods before and after attainment of SPA, and whether variations in patterns differ according to whether an individual works or retires, as well as on the basis of the quality of their work. Following from the descriptive statistics provided in Chapter 5, hypotheses hold that optimal patterns of health will be observed among those working in better quality work, and that better health will be observable among retirees than those who continue to work in poor quality employment. One of the theories outlined in Chapter 2 focussed on the idea of a honeymoon period following retirement (Atchley 1982), whereby the initial relief of workplace stress results in a period of better health and wellbeing. Studies by Westerlund et al. (2010) and Jokela et al. (2010), both of which specifically used forms of growth curve modelling, provided strength to this theory by showing trajectories of health did, indeed, improve in the period initially following retirement. However, other studies which analysed short- and long-term effects of retirement produced diverse results, with one showing significant effects attentuating after one year (Oksanen et al. 2011), one showing a persistent detrimental effect using IV analysis (Calvo et al. 2013) and two showing no significant effects of either short- or long-term retirement, both of which also used IV analysis (Calvo et al. 2013; Coe & Lindeboom 2008). The spline models used here will allow an examination of whether or not a honeymoon period is observable among the group of older workers included in this study, that is, those who work directly until reaching SPA and then enter normal retirement. The intercepts of the spline models are allowed to differ in order to ascertain whether leaving the workforce and entering statutory retirement is associated with a sudden improvement in health. Addtionally, slopes of health can be examined in the period prior to- and following SPA. Atchley (1982) notes that although the period immediately following retirement leads to improvements in

218 217 health, these effects attenuate over time, and within as short a period as six months following workforce exit, retired individuals become accustomed to their new lives. As tasks become increasingly routine, the health of retirees becomes similar to those who are working, and in some cases worsens, especially where initial expectations of retirement become unrealistic (Gall et al. 1997; Atchley 1982). After a brief overview of the data, models will be run which allow observation of whether the intercepts and slope gradients of the three key outcomes of interest to this work depression, self-rated health and cognitive function change once retirement age has been reached and individuals enter either later-life employment or retirement. Results are further stratified to compare later-life workers alone on the basis of whether they participate in poor or good quality work. The model specifications also enable examination of whether the time of reaching retirement age and making the transition into later-life employment or retirement has an immediate effect on health outcomes by means of assessing differences in intercepts at this specific time-point. Again, differences can be assessed between the groups of interest. The spline models are an interesting addition to the study as they allow an insight into how trajectories of health differ according to employment status and type and, additionally, they allow consideration of the impacts of specific individual sociodemographic characteristics on health outcomes. The spline models allow an observation of whether leaving pre-retirement age work and entering either later-life employment or retirement leads to a change in patterns of health. They provide a complementary analysis to the propensity score matching presented in the following chapter, which instead focuses on the impact of socio-demographic characteristics on the likelihood of membership of the working and retired groups, and which subsequently seeks differences in health scores on the basis of group membership.

219 Data used in this chapter The analysis within this chapter uses waves 1 to 5 of the ELSA datasets as described already. The sample used is that which is outlined in Chapter 5, with all members responding to at least two waves of the survey, one of which is the wave at which they reach retirement age, and another the wave prior to this. Due to the nature of the analyses carried out here, the data are set longitudinally, so that individuals can be measured according to each time point of the data they responded to. This yields a strongly balanced dataset, with the majority of cases responding to all waves of data, containing a total of 5,595 observations. Detailed descriptive statistics relating to the sample were provided in Chapter 5. The data is centred at retirement age, at which point all men are aged 65 and all women are aged 60. As a result, there is very little age variation at each wave included in the analyses. The ELSA datasets provide weights for longitudinal analyses to deal with issues of attrition and non-response, but these were not used in this instance. This was because the strict sample selection process undertaken at the beginning of the analysis created a dataset of individuals who are not representative of the overall population, and therefore weighting in these circumstances is unneccesary. The chapter provides analyses for the three health outcomes of interest already outlined in previous chapters. To recap, depression is measured using the 8-point version of the CES-D scale, with a score of 0 indicating no depression and of 8 indicating the highest possible level. Self-rated health is measured using the fivepoint Likert scale response, with 1 indicating excellent health and 5 indicating poor, and again excludes findings at wave 3 due to an incompatibility in scores inherent in the original ELSA datasets. Cognitive function uses the index described in Chapter 4, with scores ranging from 0 to 44, and here a higher score is indicative of better function. Table 6.1: Sample size at each centred wave, according to employment status and type. Wave (centred) Employed Retired Poor quality employment Good quality employment

220 219 Table 6.1: Sample size at each centred wave, according to employment status and type. Wave (centred) Employed Retired Poor quality employment Good quality employment , As the analysis here focuses on trajectories of health, with all cases responding to, and centred on, retirement age, there is a large variation in sample size at each wave. Table 6.1 demonstrates the number of centred observations (cases) at each wave. The first two columns of frequencies correspond to whether cases continue employment or retire once retirement age is reached, and the last two correspond to whether those who do continue to work do so in poor or good quality employment. This latter section is comprised only of 535 individuals who work and who also provide data concerning work quality. The highest number of individuals is at waves 0 and -1, where all cases reach retirement age and state whether they continue employment or retire. Information is also recorded for each respondent at the wave prior to this, in order to ensure that all cases are in the workforce before reaching retirement age. The reduction in retired cases from wave 1 onwards is the result of non-response or attrition. Where the number of employed cases begins to reduce at the same time, this is the result of both missing data as well as exit from the workforce into retirement. There is a lower number of respondents prior to wave -1, which is due to the fact that respondents joined the sample anytime between waves 1 to 4, as their inclusion in the analysis was dependent on reaching SPA at any point. As no work quality information was collected at wave 1 of the ELSA data, there is no information regarding it at its equivalent, wave -4.

221 Results of the piecewise spline regression analysis This section of the chapter uses piecewise regression in order to examine two key pieces of information. Firstly, the question can be asked of whether reaching retirement age and entering either later-life employment or retirement has an immediate effect on health outcomes. This can be assessed by segmenting the regression at the centred wave of retirement age and observing any differences which might lie in the intercepts of the before-and-after outcome slopes. Secondly, it can be asked whether trajectories of health follow different paths prior to- and following- retirement age by examining differences in the slopes of the models segmented at the centred retirement age. Again, slopes can be examined for those who work and those who retire individually, to assess whether differences occur on the basis of employment. In addition to examining intercept and slope differences for those who continue work and those who retire, piecewise regression models are used to assess whether differences in health trajectories exist on the basis of work quality. Later-life employment will be stratified on the basis of poor or good quality, and any effects of work quality on trajectories of health when participating in later-life employment can be observed. The spline models presented in this chapter provide a useful insight into patterns of health across the retirement age period and complement the final set of analyses to be presented in Chapter 7. Chapter 5 presented unadjusted health scores on the basis of employment and socio-demographic variables. While the forthcoming analyses in Chapter 7 will use these variables in the calculation of the propensity score, their focus will be centred on issues of predicting group membership. The spline models here allow, instead, an examination of the ways in which these variables directly impact health outcomes. Trajectories of health can be compared on the basis of group membership, and the impact of various background characteristics on these trajectories can be examined for groups separately. Models are produced using Stata s generalised estimating equations command (xtgee). An autoregressive correlation structure is used in order to account for the correlation between the outcomes of individuals over close periods of time. Splines are constructed using Stata s mkspline command, and for ease of interpretation, the

222 221 marginal option is implemented. This allows examination of the coefficients for changes in slopes and intercepts at retirement age, rather than of the slope and intercept coefficients themselves. While spline regression allows any number of time-points k, or knots, on which slopes and intercepts may vary, the analyses here uses only one knot at the point of centred retirement age. The first reason for this is the fact that this is the point in time in which the research is most interested. Secondly, sample numbers are considerably smaller at other points in time, and results of additional knots on the basis of this would be unreliable. Tables 6.2 to 6.4 show the results of the piecewise regression analyses for outcomes of depression, self-rated health and cognitive function, respectively. Each of the models control for gender, age (centred), wave of response, marital status, wealth (quintile), whether qualifications over O-Levels are held, housing tenure, whether or not the respondent s partner is in employment, NS-SEC category (professional and intermediate, with manual as the reference category), whether the participant volunteers, whether the participant provides care, private pension membership, and ADL difficulties. An interaction term between gender and age is also included in each of the models. Additional interaction terms included in preliminary runs were showed no significant effects, and so were excluded from the final models. In the first instance, the tables present unadjusted slope and intercept coefficients. Fully adjusted coefficients are presented thereafter. Two coefficients concerning the slope are presented within each analysis. The preretirement age slope is the gradient of change in health scores until retirement age is reached. In all instances, respondents are in employment prior to retirement age. The second consideration of slope values is the change in slope which occurs at retirement age. By the time this change is measured, respondents have entered either continued employment or retirement, and so this coefficient is useful in highlighting whether a change in employment status bears any immediate effect on trajectories of health. A coefficient for change in intercepts is also presented. Again, this allows a demonstration of whether the mean health scores of individuals show an immediate change as retirement age is attained. In the fully adjusted models, covariate values and their standard errors are also presented. Models have been run separately for the same four groups of interest detailed in the previous chapter: all workers, retirees, people in poor quality work and people in good quality work. Groups are not

223 222 compared directly in this instance, but instead differences in patterns of health are compared within groups over time. Group comparisons will form a key focus of the following chapter Depression Table 6.2 presents the results of four piecewise regression models carried out on depression. The unadjusted model shows slope and intercept coefficients, and it is only in the case of retirees that a significant change in slope occurs once retirement age is reached. Here, an increase in the gradient of the slope of is observed, suggesting depression worsens over time once retirement has been entered. Continuation of work has no significant effect. Prior to retirement age, the gradient of slopes is consistently negative, suggesting a link with lower levels of depression while in the workforce. Using the coefficients for the pre-retirement age slope and the subsequent change in slope, it is possible to calculate the precise slope gradient following retirement age. Across all three working groups, the later slope remains negative, and the slope for retirement is the only one suggestive of worsening health. Table 6.2: Results of piecewise regression on depression (CES-D score) Later-life employment versus retirement Poor versus good quality work Covariate Continues employment (n=2779) Retires (n=1814) Poor quality laterlife employment (n=575) Good quality later-life employment (n=1698) Unadjusted model Pre-retirement age slope (0.04) (0.05) (0.11) (0.05) Change in slope (0.06) (0.07)* (0.17) (0.07) Change in (0.11) (0.12) (0.29) (0.29) intercept Fully adjusted model Pre-retirement age slope (0.10) (0.12) (0.27) (0.11) Change in slope (0.06) (0.07)* (0.16) (0.07) Change in (0.11) (0.13) (0.29) (0.12)

224 223 Table 6.2: Results of piecewise regression on depression (CES-D score) Later-life employment versus retirement Poor versus good quality work Covariate Continues employment (n=2779) Retires (n=1814) Poor quality laterlife employment (n=575) Good quality later-life employment (n=1698) intercept Female (0.08)*** (0.08)*** (0.21)* (0.09)*** Wave (0.03) (0.04) (0.09) (0.04) Married (0.08)*** (0.09) * (0.20) (0.09)* Wealth (0.02)** (0.03) (0.07)** (0.03) Qualifications (0.07)* (0.08) (0.19)* (0.08)* above O-Levels Tenure (0.07) (0.08) (0.18) (0.07) Partner works (0.07) (0.08) (0.19) (0.08)* NS-SEC (0.08) (0.10) (0.23) (0.09) professional NS-SEC (0.08)** (0.09)*** (0.19) (0.09)** intermediate Volunteers (0.10) (0.11)* (0.28) (0.11) Provides care (0.09) (0.10) (0.22) (0.10) Private pension (0.08) (0.10) (0.21) (0.09) ADL difficulties (0.05)*** (0.04)*** (0.12)*** (0.06)*** Gender*age (0.02) (0.02) (0.05) (0.02) Constant (0.28)** (0.31)*** (0.72)* (0.02)* ***p<0.001 **p>0.01 *p<0.5 The second part of the table shows results of the adjusted model. After accounting for all covariates, the change in slope at retirement age for those who retire remains significant. Additionally, the pre-retirement age slope coefficient for those who retire has changed from negative to positive, suggesting an association with the lead up to retirement and higher levels of depression. The value of this change in slope is 0.152,

225 224 which suggests depression scores increase after entering retirement when compared to the period before. Tale 6.2 also presents coefficients for changes in intercept at the point of reaching retirement age. The small yet positive values for all such coefficients in the adjusted model suggests that crossing retirement age is associated with a slight worsening of mean depression scores. However, none of the reported intercept changes are significant. Figure 6.1 presents a visual demonstration of the findings presented in Table 6.2. The x axis shows time to, and then time following, attainment of normal retirement age. The value of 0, and the centred line placed in the middle of each chart, represents age 65 for men and age 60 for women. Although the majority of slope and intercept differences within groups are not significant, the graphs clearly demonstrate overall trajectories of depression to differ. Retirees appear to have the highest mean scores, with those who continue into poor-quality later-life employment showing similar pre-retirement age scores, despite the differences in the directions of trajectories. This complements the observation made in Chapter 5, that the characteristics of retirees and people in poor quality work were commonly far more similar than those of retirees ad respondents in good quality work. Figure 6.1 also shows the scores of all working groups to be, on average, lower than those of retirees and the lowest scores of all exist among those who participate in good quality employment. As the changes in intercepts in Table 6.2 were not significant, the slopes in Figure 6.1 are unbroken at the centred point of reaching retirement age.

226 225 The spline models are also useful in that, like ordinary regression, they provide an insight into factors which significantly affect the outcome, and the magnitude and direction of these effects can be compared across groups. The covariate which appears to bear the most impact on depression scores is the number of ADL difficulties an individual has, and is significant with p<0.001 in each of the models run. The greatest impact on depression is seen among those who continue into poor quality work, and again ties in with aforementioned ideas of job lock, whereby individuals continue to work, often in stressful or demanding occupations, or when they physically struggle, due to financial constraints. Those in poor quality work see the strongest impact of ADL difficulties on depression, which is again likely to reflect issues of job-lock which may be particularly prevalent among individuals in poorer quality work who cannot afford to leave the workforce despite suffering poorer health, and who subsequently suffer poorer mental health. Being female is significantly associated with an increase in depression in all groups. Chapter 5 provided much evidence that the relationship between work in later-life varied on the basis of gender, with women seeing an association with higher levels