UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

The labour market behaviour of older individuals Schirle, Tammy 2006

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

Item Metadata

Download

Media
831-ubc_2007-267912.pdf [ 9.64MB ]
Metadata
JSON: 831-1.0100419.json
JSON-LD: 831-1.0100419-ld.json
RDF/XML (Pretty): 831-1.0100419-rdf.xml
RDF/JSON: 831-1.0100419-rdf.json
Turtle: 831-1.0100419-turtle.txt
N-Triples: 831-1.0100419-rdf-ntriples.txt
Original Record: 831-1.0100419-source.json
Full Text
831-1.0100419-fulltext.txt
Citation
831-1.0100419.ris

Full Text

T h e Labour Market Behaviour of Older Individuals by T A M M Y SCHIRLE B.A., The University of Manitoba, 1999 M.A. , Dalhousie University, 2000 A THESIS S U B M I T T E D IN PARTIAL F U L F I L M E N T O F T H E R E Q U I R E M E N T S F O R T H E D E G R E E O F D O C T O R OF PHILOSOPHY in T H E F A C U L T Y O F G R A D U A T E STUDIES (Economics) T H E UNIVERSITY O F BRITISH C O L U M B I A November 2006 © T a m m y Schirle, 2006 Abstract This dissertation investigates several aspects of the labour force par t ic ipat ion and re-tirement decisions of older individuals , introduced i n Chapter 1. Chapter 2 examines how several components of Canada 's income security system could affect individuals ' incentives to retire. The components of Canada 's income security system are documented and we show how they act to change the incentives to retire through a series of simulations. T h i s chapter also provides a thorough survey and cr i t ica l review of the international evidence on public pensions and retirement, w i t h the broad weight of the evidence suggesting that the structure of public pensions contributes to the decision to retire. In Chapter 3 I f i l l some of the gaps i n the Canad ian literature on retirement decisions, which has focused almost exclusively on the role of public pensions. In this chapter I extend the analysis of Baker et al . (2003, 2004a) to examine not only the effects of public pensions, but also the effects of health and employer-provided pensions on individuals ' decisions to enter retirement. Us ing data from the Survey of Labour and Income Dynamics , my main finding is that having poor health, or the occurrence of health events such as the onset of a disability, significantly increases an individual ' s l ikel ihood of entering retirement. Another key contr ibut ion to the Canad ian literature is the finding that individuals are responsive to the financial incentives found in employer-provided pension plans. Addi t ional ly , my estimates indicate that individuals consider their entire financial picture when making their retirement decisions. Chapter 4 seeks to explain the substantial increases i n older men's labour force part ici-pat ion rates that have been observed since the mid-1990s. Us ing data from the U . S . M a r c h Current Popu la t ion Survey, the Canadian Labour Force Survey, and the Un i t ed K i n g d o m Labour Force Survey, I investigate the hypothesis that husbands treat the leisure t ime of their wives as complementary to their own leisure at older ages. G i v e n this complementar-ity, a large por t ion of the increase i n older men's par t ic ipat ion rates may be explained as a response to the recent increases i n older women's par t ic ipat ion i n the labour force, which are largely driven by cohort effects. The methodology of Dinardo , For t in , and Lemieux (1996) is used to decompose the changes in older married men's par t ic ipat ion rates, demonstrating that increases i n wives' par t ic ipat ion in the labour force can explain roughly one quarter of the recent increase i n par t ic ipat ion in the U . S . , up to one half of the recent increase in par t ic ipat ion i n Canada , and up to two fifths of the recent increase i n the U . K . Older men's educational at tainment is also an important factor explaining recent increases in part icipa-t ion, yet cannot be expected to drive further increases i n par t ic ipat ion rates. In contrast, expected increases i n older wives' part ic ipat ion over the next decade are expected to drive further increases i n older men's part icipat ion rates. n Table of Contents Abstrac t i i Table of Contents i i i L i s t of Tables v i L i s t of Figures v i i i Acknowledgements i x Co-Author sh ip Statement x 1 Introduction 1 2 Public Pensions and Retirement: International Evidence in the Canadian Context 5 2.1 Int roduct ion 5 2.2 Canada 's Retirement Income Security System 7 2.2.1 Canada Pension P l a n - Quebec Pension P l a n 8 2.2.2 O l d Age Security 9 2.2.3 Guaranteed Income Supplement 10 2.2.4 T h e Allowance 10 2.2.5 Summary 11 2.3 Simulat ions 11 2.3.1 Methodology 11 2.3.2 Base Case Results 14 2.3.3 Extended Simulat ion Results 16 2.3.4 Po l i cy Simulations 19 2.3.5 Summary of Simulations 20 2.4 International Evidence 21 2.4.1 Research on related topics 22 2.4.2 E a r l y evidence on pension wealth and retirement 23 2.4.3 S t ruc tura l models of retirement 26 2.4.4 Es t ima t ion of accrual and level effects of pensions 27 2.4.5 N a t u r a l Experiments 33 i i i 2.5 Conclusions 36 3 The Effects of Health and Financial Incentives on Retirement Decisions in Canada 50 3.1 Introduct ion 50 3.2 Previous Li tera ture . 51 3.3 Canada 's Retirement Income System 54 3.3.1 Income Security Programs 54 3.3.2 Employer -Prov ided Pension Plans 55 3.3.3 Other Sources of Income 55 3.4 M o d e l l i n g the Retirement Decision 56 3.5 Es t ima t ing the Effects of Hea l th and F inanc ia l Incentives 58 3.5.1 D a t a 59 3.5.2 Measurement of K e y Variables 59 3.5.3 Identification of Weal th and A c c r u a l Effects 64 3.6 Results 66 3.6.1 T h e Effects of Income Security Programs 66 3.6.2 T h e Effects of Employer-Provided Pensions and Other Income . . . 68 3.6.3 The Effects of Hea l th 69 3.7 Conclusions 70 4 Why Have the Labour Force Participation Rates of Older Men Increased Since the Mid-1990s? 87 4.1 Introduct ion 87 4.2 D a t a and Recent Trends i n Par t ic ipa t ion 89 4.2.1 D a t a 90 4.2.2 Trends i n Par t ic ipa t ion 90 4.3 Leisure Complementar i ty and Par t ic ipa t ion Decisions 92 4.3.1 A Simple M o d e l of Shared Leisure and Income Effects 92 4.3.2 Es t ima ted Effect of Wives ' Par t ic ipa t ion on Husbands ' Decisions . . 95 4.3.3 Robustness Checks - Other Factors Important For Par t i c ipa t ion De-cisions 98 4.4 Decomposing the Changes i n Par t ic ipa t ion 99 4.4.1 P r o b i t / D F L Decomposi t ion of Changes i n Par t i c ipa t ion 100 4.4.2 P r o b i t / D F L Decomposi t ion Results 104 4.4.3 A d d i t i o n a l Evidence 105 4.4.4 L P M / O a x a c a Decompositions 107 4.5 Future Trends 109 4.6 Conclusions . 110 iv 5 Conclusions 144 Bibliography 147 Appendices 153 A Financial Incentives in Income Security Programs for Chapter 3 . . . 153 B Sample Selection and Construction of Key Variables for Chapter 4 . . 161 B . l U n i t e d States 161 B . 2 Canada . . . 162 B . 3 U n i t e d K i n g d o m 163 B .4 Other D a t a 164 v List of Tables 2.1 Basic statistics on simulated individuals 38 2.2 Base case simulations 39 2.3 Pr iva te pension simulations 40 2.4 Range of Earnings Simulations 41 2.5 Work interrupt ion simulations 42 2.6 Il lustrative Po l i cy Simulations 43 2.7 Summary of Retirement Studies 44 3.1 Importance of Various Income Sources 71 3.2 Rate of E x i t F r o m Retirement 72 3.3 Characteris t ics of Retirees and Non-Retirees 73 3.4 Hea l th Measures by Age 74 3.5 The Di s t r ibu t ion of Imputed and A c t u a l Incomes 75 3.6 The Di s t r ibu t ion of Income Security Measures 76 3.7 The Di s t r ibu t ion of Income Security + Pension Measures 77 3.8 The Di s t r ibu t ion of To ta l Income Measures 78 3.9 Retirement P r o b i t Results I 79 3.10 Retirement (F ixed Effects) Prob i t Results II 80 3.11 Retirement P r o b i t Results III 81 3.12 Retirement P r o b i t Results I V 82 3.13 Retirement P rob i t Results V 83 3.14 Retirement P rob i t Results V I 83 4.1 Characteris t ics of M a r r i e d M e n Age 55-64, Un i t ed States 112 4.2 Characteris t ics of M a r r i e d M e n Age 55-64, Canada 113 4.3 Characteris t ics of M a r r i e d M e n Age 55-64, U n i t e d K i n g d o m 114 4.4 M o d e l Est imates, Pooled Samples, Un i t ed States 115 4.5 M o d e l Est imates, Pooled Samples, Canada 116 4.6 M o d e l Est imates , Pooled Samples, Un i t ed K i n g d o m 117 4.7 Es t ima ted Effect of Wives ' Par t ic ipat ion, Pooled Samples, U n i t e d States . 118 v i 4.8 Weights and Coefficients used in the Decomposit ions 119 4.9 Decomposi t ion Results (Using Pooled Est imates) , U n i t e d States 120 4.10 Decomposi t ion Results (Using Pooled Est imates) , C a n a d a 121 4.11 Decomposi t ion Results (Using Pooled Est imates) , U n i t e d K i n g d o m . . . . 122 4.12 M o d e l Est imates, A n n u a l , Un i t ed States 123 4.13 M o d e l Est imates, A n n u a l , Canada 124 4.14 M o d e l Est imates, A n n u a l , Un i t ed K i n g d o m 125 4.15 Decomposi t ion Results (Using A n n u a l Est imates) , U n i t e d States 126 4.16 Decomposi t ion Results (Using A n n u a l Est imates) , C a n a d a 127 4.17 Decomposi t ion Results (Using A n n u a l Est imates) , U n i t e d K i n g d o m . . . . 128 4.18 Robustness checks - M o d e l Estimates, Pooled Samples, U n i t e d States . . . 129 4.19 Robustness checks - M o d e l Estimates, Pooled Samples, Canada 130 4.20 Robustness checks - M o d e l Estimates, Pooled Samples, U n i t e d K i n g d o m . 131 4.21 Robustness checks - M o d e l Estimates, Pooled Samples, U n i t e d States . . . 132 4.21 Robustness checks - M o d e l Estimates, Pooled Samples, U n i t e d States . . . 133 4.22 Effect of W i v e s ' Par t ic ipa t ion , by Age Group, U n i t e d States 134 4.23 Characterist ics of Mar r i ed M e n Age 45-54 135 4.24 Forecasted Changes i n Men ' s Par t ic ipa t ion Rates 136 A . l S imula t ion - Var i a t ion i n I S W Accruals Across Gender and M a r i t a l Status 157 A . 2 S imula t ion - Var i a t ion in I S W Accruals Across other Retirement Income Levels 158 A . 3 S imula t ion - Var ia t ion in I S W Accruals Across Earned Income 159 A . 4 S imula t ion - Var i a t ion i n I S W Accruals Across Years Worked 160 vii List of Figures 2.1 M a l e and Female Surv iva l Probabi l i t ies 49 3.1 P rov inc i a l Hea l th Care Expenditures per Person by Age and Sex, 2000 . . . 84 3.2 M a l e Pa r t i c ipa t ion Rates by Age Group, 1976-2004 85 3.3 Female Par t i c ipa t ion Rates by Age Group , 1976-2004 85 3.4 Cond i t iona l P robab i l i ty of Retirement at Different Ages 86 4.1 Par t i c ipa t ion Rates of Individuals Age 55-64, by Sex 137 4.2 Par t i c ipa t ion Rates of Individuals Age 55-64, by Sex 138 4.3 Age-Par t i c ipa t ion Profiles for Selected B i r t h Cohorts of Women 139 4.4 Par t i c ipa t ion Rates of Individuals Age 60-64, by Sex and M a r i t a l Status . . 140 4.5 Shared Leisure and Income Effects 141 4.6 Par t i c ipa t ion Rates of Individuals Age 25-54, by Sex 141 4.7 Oaxaca Decomposit ions of His tor ical Par t ic ipa t ion Rates, U . S . and Canada 142 4.8 Oaxaca Decomposit ions of His tor ical Par t ic ipa t ion Rates, U . K 143 v i i i A c k n o w l e d g e m e n t s I sincerely thank Nicole For t in , my thesis supervisor, for her advice, support, and pa-tience since the first stages of my thesis research. I am also heavily indebted to Thomas Lemieux and K e v i n M i l l i g a n , my thesis committee members, for their input and guidance. I thank D a v i d Green and Cra i g R i d d e l l for their advice and suggestions. I would also like to thank T A R G E T , Nicole For t in , and K e v i n M i l l i g a n for providing me wi th invaluable research opportunit ies and financial support while complet ing my P h D . I am also grateful for the continual assistance of the Department 's staff, notably Maureen C h i n . I would also like to thank my fellow P h D students, family, and friends who have con-t r ibuted in a variety of ways to the completion of this thesis. In part icular, I would like to thank Pierre Brochu , A n e t a Bonikowska, Doris Poon , Stephanie M c W h i n n i e , A n a Ferrer, and Jacob W o n g for their friendship, support, cr i t ic ism, and ideas. I thank my M o m for her support on more difficult days and for tel l ing me the answer is five. Mos t of a l l I thank my husband Pa t r ick for his love and patience and his belief that a good set of instruments should always include a t rumpet . ix C o - A u t h o r s h i p Statement Chapter 2 (Publ ic Pensions and Retirement: International Evidence in the Canadian Context) was co-writ ten w i t h Professor K e v i n M i l l i g a n ( U B C ) . M y contr ibut ion to the product ion of this piece of research is outl ined below. • Identification of research program - a smal l contr ibut ion, as the in i t i a l idea was gen-erated by my co-author. • Design of research program - a moderate contr ibut ion, in consultat ion wi th my co-author. • Performing the research - a large contribution, inc luding the description of Canada's system and the review of the relevant international li terature. M y co-author was p r imar i ly responsible for producing the simulations i n this chapter. • Manuscr ip t preparation - a large contribution, as drafts were prepared and revised jo in t ly w i t h my co-author. x C h a p t e r 1 Introduction The potential social and economic consequences of popula t ion aging have sparked consider-able interest in the labour market behaviour of older individuals . Concerns about populat ion aging generally fall into two broad categories. F i r s t , popula t ion aging is expected to result in higher expenditures on health care, public pensions and other publ ic ly funded programs used by our seniors. G i v e n that many of these programs are funded on a 'pay-as-you-go' ba-sis, the fiscal sustainabil i ty of these programs has been called into question. 1 Second, there are concerns that the retirement of the baby boom generation could lead to sk i l l shortages wi th the loss of experienced people from the labour force. To address these concerns some countries have expressed a desire to alter the structure of ret irement. 2 Developing a solid understanding of the determinants of older individuals ' labour market behaviour is therefore a necessary first step i n addressing these concerns. In this thesis, I investigate several factors that influence the labour force part icipat ion and retirement decisions of older individuals. It is useful to begin by clarifying how I conceptualize retirement i n this thesis, as the term 'retirement' can have several meanings. I have characterized retirement as an individual 's permanent wi thdrawal from labour market activities after par t ic ipat ing in the labour force through most of his or her adult life, as this appears to be the ac t iv i ty of greatest interest to pol icy makers. Individuals who have permanently wi thdrawn from the labour market are not likely, for example, to delay receipt of public pension benefits and are not offering their skills to employers. 3 W h e n examining the retirement decision using survey data, I have defined entry to retirement as a departure from the labour force for more than one 1 Note that Canada's public pensions ( C P P / Q P P ) appear to be sustainable following several reforms in 1997 that led to the creation of a reserve fund to cover future pension payments. 2 F o r example, Canada's 2005 Budget P lan (Department of Finance, 2005) states that "Wi th the upcoming labour scarcity, there is a need... to ensure that older Canadians do not face disincentives to work" and that "minimizing institutional and financial disincentives to work has the potential to raise the labour force attachment of older Canadians." 3 Another common characterization, for example, defines entry to retirement as the point when an in-dividual leaves a career job. Such individuals, however, may continue working in a post-career job, either full-time or part-time, and would not qualify for many income-tested public pension benefits. 1 year. The window of observation is restricted for pract ical reasons, but is adequate given that very few individuals over the age of 55 who wi thdraw from the labour force for such an extended per iod of t ime actually return to the labour force. To note, I have chosen to not make use of self-reports of retirement for several reasons. F i r s t , the use of self-reports leaves the concept of retirement very ambiguous. Second, al though self-reported retirees w i l l tend to fit an expected profile of retirement, i n many data sets the measurement of self-reported retirement is based on reasons why individuals left their last job and w i l l miss any individuals who were la id off or left jobs for health reasons and then entered retirement (see Gower (2004)). I should also note that my chosen definition of retirement (and the examinat ion of retirement i n this thesis) does not address the various paths that individuals may take into retirement. However, as Gus tman and Steinmeier (1983, 1984) demonstrate, the majori ty of workers face hours constraints that would prevent them from gradually phasing out of full-time jobs into retirement and it is most common for individuals to move directly from full t ime employment to full retirement (see Gus tman and Steinmeier (1986)). I begin i n Chapter 2 (co-authored w i t h K e v i n Mi l l igan) w i t h an examinat ion of evidence on the impact of Canada 's public pensions on the retirement decisions of the elderly. The components of Canada 's income security system are documented and we show how they act to change the incentives to retire through a series of simulations. For example, on one hand we demonstrate that C P P / Q P P actuarial adjustments do not adequately compen-sate individuals for foregone years of pension receipt and reduce eventual GIS payments, thereby creating disincentives to remain in the workforce at older ages. O n the other hand, individuals that have experienced several work interruptions over their lifetime may have the incentive to continue working as this would allow individuals to drop some low-earnings years from their work history when calculating the average income that determines the level of C P P / Q P P benefits they are eligible for. We then provide a thorough survey and cr i t ical review of the international evidence on public pensions and retirement. The broad weight of the evidence, inc luding several recent Canadian studies of retirement behaviour, suggests that the structure of publ ic pensions contributes to the decision to retire. W h i l e there is a fairly extensive international literature examining retirement decisions, the literature i n C a n a d a has focused almost exclusively on the role of income security programs. In Chapter 3, I extend the analysis of Baker et a l . (2003, 2004a) to examine not only the effects of public pensions, but also the effects of health and employer-provided pensions on individuals ' decisions to enter retirement. Us ing panel data from the Survey of Labour and Income Dynamics (1995-2002) I am able to observe individuals ' labour market transitions, health status, job characteristics, and income from various sources. I use an option value framework for the analysis of financial incentives, creating two variables to 4 T h i s definition of retirement, its implications, and some alternatives are discussed at greater length in Chapter 3, where this definition is used. 2 capture the financial incentives associated wi th public and employer-provided pensions. Fi rs t , a wealth measure is created representing the discounted present value of income from various sources. Increases i n wealth from public or employer-provided pensions are expected to reduce the number of years a person is in the labour force. Second, I measure an individual ' s incentive to immediately enter retirement as the amount of wealth an indiv idual could accrue by delaying retirement unt i l a future op t imal date (referred to as a peak accrual value). Hea l th is treated as a preference shifter, w i t h the expectation that poor health increases the d isu t i l i ty of work and therefore increases the l ikel ihood of entering retirement. A probit model is used to estimate the effects of financial incentives and health on entry to retirement, w i t h specifications controll ing for ind iv idua l fixed effects, spousal and family characteristics, and the endogeneity of health reports. M y ma in finding in this chapter is that having poor health, or the occurrence of health events such as the onset of a disability, significantly increases individuals ' l ikel ihood of entering retirement. I address identification issues associated w i t h using self-assessed health measures and find that having poor health raises the l ikel ihood of entering retirement by more than twenty percentage points. Another important contr ibut ion of this chapter is the finding that employer-provided pensions have significant wealth and accrual effects, which had not been accounted for i n previous Canad ian studies. I also find that the financial incentives in Canada 's income security programs have significant accrual effects on the retirement decision. Overa l l , the results presented i n this chapter suggest that reforming Canada's retirement income policies could address many concerns about populat ion aging if designed to affect individuals ' t iming of retirement. Augment ing the concerns about populat ion aging, i n many countries the part icipat ion rates of older men had fallen for several decades. In the mid-1990s, however, a clear reversal in the labour force par t ic ipat ion rates of men age 55-64 occurred i n Canada , the Uni t ed States, the U n i t e d K i n g d o m , and several other European countries. In Chapter 4, I seek a common explanat ion for the recent increases in older men's par t ic ipat ion. Us ing data for Canada, the U . S . , and U . K . , I investigate the hypothesis that husbands treat the leisure t ime of their wives as complementary to their own leisure at older ages and that the recent increase i n older men's par t ic ipat ion rates is largely a response to recent increases in the par t ic ipat ion of older wives. Mode l l i ng the husband's and wife's par t ic ipat ion decisions as a system of simultaneous probit equations, I am able to identify the effect of a wife's par t ic ipat ion i n the labour force on the husband's par t ic ipat ion decision using a measure of cohort effects as an instrument for wives' part ic ipat ion. T h e results show that in a l l three countries, husbands have clear preferences for sharing leisure t ime w i t h their wives as a wife's par t ic ipat ion i n the labour force has a positive and significant effect on the l ikel ihood of husbands to participate. Us ing the decomposit ion methodology pioneered by D i N a r d o et a l . (1996), known as the D F L methodology, a decomposit ion of older married 3 men's par t ic ipat ion rates is then undertaken, demonstrat ing that a substantial por t ion of the recent increases i n older marr ied men's par t ic ipat ion can be explained as a response to the higher l ikel ihood of wives to participate in the labour force. In this chapter I also investigate the role played by changes i n the age structure and educational attainment of marr ied men age 55-64. T h i s group has become relatively younger (as the baby boom cohort enters this group) and more educated over the past decade, dr iv ing a substantial por t ion of the recent increase in part icipat ion, especially i n the Un i t ed States. Look ing forward to how these factors w i l l affect future trends i n older men's par t ic ipat ion rates, expected increases i n older wives' par t ic ipat ion w i l l continue to place upward pressure on older marr ied men's par t ic ipat ion rates. However, it appears that the effects of education have been exhausted as the education levels of upcoming cohorts of older men are not substantially higher than the current cohort of older men, and therefore cannot be relied upon to drive further increases i n older men's par t ic ipat ion rates. In Chapter 5 I provide some concluding remarks and outline some areas for future research related to this thesis. 4 C h a p t e r 2 Public Pensions and Retirement: International Evidence in the Canadian Context 2.1 Introduction The engagement of governments i n pensions is internationally pervasive. M u l l i g a n and Sa la - i -Mar t in (2004) observe that 166 countries have some type of publ ic pension program. Given this ubiquity, great interest has arisen in developing an understanding of the eco-nomics of public pensions. One branch of this inquiry asks how pensions affect the labour market decisions of the elderly. The motivat ion may lie i n a desire to expand our knowledge of how the exist ing or future structure of public pensions might affect retirement decisions. Moreover, i n some countries there may be an explicit desire to alter the structure of retire-ment through reforms to public pensions. In either thorough investigation of the effects of pensions on retirement becomes a necessary first step. A n understanding of the effects of public pension programs on labour supply begins wi th a basic lifecycle model of labour supply. In the simplest model , an ind iv idua l chooses a path of lifetime consumption and labour supply to maximize u t i l i ty subject to the constraint that the discounted present value of lifetime income equals the discounted present value of lifetime consumption. T h e fundamental tradeoff that must be contemplated is between higher consumption (afforded through more work) and higher leisure. If one works more, the higher income allows one to consume more. However, more work implies less t ime available for leisure. Every worker therefore chooses a lifetime path for work that balances the desire for consumption and leisure. Pub l i c pensions potential ly change a worker's decision i n two ways. The first is through changing the to ta l lifetime income of the worker (which is equivalent to his or her wealth). 5 The discounted present value of benefits net of contributions made to the program is part of the lifetime budget constraint. If the discounted flow of benefits equals the discounted flow of contributions, then public pensions w i l l have no effect on ind iv idua l behaviour. However, if benefits exceed contributions, then a person's lifetime income is increased by the presence of the program. A s s u m i n g leisure is a normal good, this increase i n wealth induces a person to reduce labour supply and enjoy more leisure. A l t h o u g h i n theory this reduction i n labour supply could be spread over an individual ' s lifespan (i.e. a reduction i n the number of hours worked i n each period), it is more l ikely to reduce the number of years that an individuals works . 1 Th i s mechanism is called the wealth effect. Another way publ ic pensions can affect retirement decisions is through the accrual of rights to future pension income. If working an addi t ional year raises the discounted sum of the future benefits, a worker w i l l have a stronger incentive to continue working for the addi t ional year, when comparing the advantages of retirement (more leisure) to the advantages of more work (even higher retirement income when she or he does retire). For example, benefits i n most countries are based on some function of average lifetime earnings. More work w i l l increase lifetime earnings, which may translate into higher future public pension benefits. Other features of public pensions that change benefits depending on the t iming of retirement, such as actuarial adjustments, delayed retirement credits, and means-tested programs, can also influence how extra years of work translate into higher (or lower) future benefits. T h i s channel is called the accrual effect. If pensions were paid based on contributions, then the accrual effect can be made to disappear. T h i s occurs i n employer-provided defined contr ibut ion plans or i n public pension plans such as Sweden's new public system of 'not ional ' accounts. 2 T h e level of the explicit or impl ic i t pension wealth does not depend on the t iming of the retirement decision i n a contributions-based system, so the accrual effect disappears. W i t h o u t an accrual effect, the structure of the pension can be said to be 'neutral ' w i t h respect to the retirement decision. Tha t is, the decision to retire does not depend on the structure of the system, but instead reflects the individual ' s undistorted choice about the tradeoff between extra leisure and extra retirement income. W i t h non-zero accrual effects, the retirement decision w i l l be distorted, w i t h a different and subopt imal mix of leisure and income. Th i s non-neutrality generates costs ak in to the standard efficiency losses of taxat ion. More recent model ing of the effects of retirement benefits on labour supply has focused on the accrual effect. T h e canonical model comes from Stock and Wise (1990). U t i l i t y is derived from income (which affords consumption), w i t h d isut i l i ty from work. In each period an ind iv idua l compares the expected present value of lifetime u t i l i ty from retiring 1Most workers face hours constraints in that employers typically offer jobs only at standard hours of work. For example, Gustman and Steinmeier (1983, 1984) show that the majority of workers face hours constraints that would prevent them from gradually phasing out of full time jobs into retirement. 2See Palmer (2000) for a description of Sweden's system. 6 immediately to the expected present value of ret ir ing at each future age, t rading off income and work. T h e m a x i m u m of the difference in expected present values of ret ir ing at each future age and immediate retirement is called the."option value" of postponing retirement. If the opt ion value is negative, the ind iv idua l w i l l choose to retire immediately. If the option value is positive, the ind iv idua l w i l l choose to continue working and retains the option of ret ir ing at a future date. In the next period, any ind iv idua l who continued to work w i l l determine the opt ion value of postponing retirement again, given any new information. The key insights of the opt ion value model are the forward-looking nature of the decision and the tradeoff between earlier retirement and higher retirement income. Beyond the narrow economic variables, retirement takes place i n a social context. The behaviour of one's spouse and peers could influence the retirement decision. In addit ion, health may affect retirement either because current work becomes impossible or because future health affects the t ime per iod over which pension benefits may flow. The focus of much of the economics literature on the financial motivations for retirement in no way precludes the impact of other factors. Our focus in this paper on the economics of the decision should therefore be interpreted in the broader context of social science research. We begin by describing Canada 's retirement income security system. We pay part icular attention to how each component of the system contributes to both the wealth and the accrual effects described above. We then proceed to simulations that lay out the strength and the magnitude of the retirement incentives present i n Canada 's system, and show how it varies across different individuals . The next question we address is how important these incentives may be for retirement decisions. To do so, we present a comprehensive survey of the international l i terature on public pensions and retirement. We finish w i t h a summary of the major findings of our study. 2.2 Canada's Retirement Income Security System Canada's retirement income security system includes four dist inct components. In this section, we provide the inst i tut ional detail on each component, describing how it might affect the incentives to retire. T h e descriptions are not meant to be exhaustive listings of the rules governing benefits. Instead, the focus is on the parts of the rules that have the greatest impact on retirement incentives. Before beginning the description of the system, we w i l l clarify our use of certain terms. We use income security generically to refer to public pension programs for the elderly i n any country. W h e n referring to the pr imary income security program i n the Un i t ed States, we capitalize it and cal l it by its name of Social Security. 7 2.2.1 Canada Pension P lan - Quebec Pension P lan The largest component of the income security system is the Canada Pension P l a n and Quebec Pension P l a n ( C P P / Q P P ) . The C P P and Q P P are earnings-related pensions funded by payrol l taxes on employees and employers. The two plans are administered separately by the federal government for the C P P and the Quebec government for the Q P P . Mos t details across the two programs are similar . The calculat ion of the benefit is the product of three parts. T h e first part is determined by earnings histories. The contr ibutory period is the window of t ime between 1966 or age 18 (which ever is later) and age 60. If retirement occurs after age 60, the contr ibutory period is extended, up to a m a x i m u m of age 65. Months i n which a disabi l i ty benefit was received, or were spent caring for a chi ld under age 7, are dropped from the contr ibutory period. The worker may also drop the lowest-earning 15 percent of the months i n the contributory period. For work after age 65, the earnings are only included in the calculat ion i f it results in an increased benefit. In each month i n the contr ibutory period, the ratio of earnings to 1/12 of the Year 's M a x i m u m Pensionable Earnings ( Y M P E ) is calculated. T h e Y M P E is set annually, and equaled $40,500 i n 2004. These ratios are capped at one, so that earnings in excess of the Y M P E are not considered for the pension benefit calculat ion. T h e final step in the earnings-rated part of the pension formula is to take the average of the ratios over a l l of the months i n the contr ibutory period. The second part of the benefit calculation aims to update the earnings history to the level of earnings prevail ing at the t ime of retirement. T h i s is accomplished by taking the average Y M P E i n the five years preceding the t ime of retirement (the five years includes the year of retirement). We cal l this the pension adjustment factor. The th i rd part of the benefit calculation adjusts the pension for the age of retirement. The ' fu l l ' pension is received i f retirement is at age 65. For every month before age 65, an actuarial adjustment of 0.5 percent is deducted from the full benefit. Symmetrically, retirement after age 65 receives a bonus of 0.5 percent per month of delay. These actuarial adjustments are capped at 5 years, meaning that the earliest one can c la im regular benefits is at age 60, at a 30 percent (30 percent is 60 months times 0.5) reduction from the full benefit level. The product of these three parts is then mul t ip l ied by the C P P / Q P P replacement rate of 25 percent and d iv ided by 12 to arrive at the monthly benefit. T h i s is summarized in the following formula. M o n t h l y Benefit = (earnings rating) • (pension adjustment factor) • (actuarial adjustment) • 0.25 • (1/12) (2.1) 8 The monthly benefit, once ini t ia ted, is updated quarterly for changes i n the consumer price index. U p o n the death of the recipient, any surviving spouse may be eligible for survivor benefits. 3 How does the C P P / Q P P affect retirement incentives? F i r s t , there is a wealth effect embodying the to ta l discounted amount of future benefit flows. T h i s encompasses both the regular benefits and the spousal benefits. Higher wealth (or equivalently, a higher annual flow of retirement income) is predicted by theory to lead to earlier retirement. In addi t ion, the C P P / Q P P pensions have many channels of influence on the accrual incentive to retire. F i r s t , i f the extra periods at work have high enough earnings so that they are included i n the pension calculation, then the retirement pension w i l l be larger when it is eventually taken. T h i s means that more work leads to a higher pension once it is ini t iated. The 15 percent ' throw-out ' rule and the earnings averaging rules help to determine the strength of this impact . Second, the actuarial adjustment depends specifically on the age of retirement. If retirement is delayed one month past age 60, then one month of pension receipt is foregone. However, the actuarial adjustment leads to a higher pension benefit once benefits are eventually init iated. The actuarial adjustment attempts to balance these amounts. T h r o u g h this actuarial mechanism, the t iming of retirement has an effect on the net present value of pension benefits received. 2.2.2 Old Age Security The O l d Age Security ( O A S ) pension is a uniform demogrant w i t h a m a x i m u m benefit of $466.63 per month i n September 2004. The pension amount is updated quarterly for changes i n the Consumer Pr ice Index, and the income is taxable as regular income. It is available to a l l individuals over the age of 65 meeting residency requirements. 4 There is a clawback of O A S benefits from very high income individuals : the O A S for an individual is reduced by 15 cents per dollar of personal net income exceeding $59,790 (in 2004). A s such the full O A S pension is el iminated when an individual ' s net income exceeds $96,972 (in 2004). The effect of the O A S pension on retirement incentives occurs main ly through the wealth effect. The O A S benefit does not depend on the date of retirement directly, so there is no direct accrual effect from working extra years. For those who are subject to the O A S ^Survivor benefits are paid at a rate of 60 percent of regular benefits if the survivor is age 65 or more, and 37.5 percent plus a fixed amount for survivors under age 65. These amounts differ in the Canada and Quebec Pension Plans. 4When first introduced in 1952 OAS was only available to individuals over the age of 70. The eligibility age was reduced to 65 over the last half of the 1960s. To be eligible for benefits, individuals must have been a Canadian citizen or legal resident of Canada at some point before application and must have resided in Canada for at least 10 years after reaching age 18 (if currently in Canada) or twenty years (if currently outside Canada). The benefit is prorated for pensioners with fewer than forty years of Canadian residence (after the age of 18), unless they are "grandfathered" under rules that apply to the persons who were over age 25 and had established attachment to Canada prior to July 1977. 9 clawback, however, there w i l l be some accrual effect. The accrual effect for them arises because extra work increases the C P P / Q P P benefit which then serves to decrease the O A S benefit through the clawback. However, the clawback affects relatively few seniors so this interaction between the C P P / Q P P and the O A S is of less general impor tance . 5 2.2.3 Guaranteed Income Supplement The Guaranteed Income Supplement (GIS) is paid to Canadians from age 65. It is also indexed to prices, but is not taxable income. The pension benefit was set in September 2004 at $560.69 for single individuals and $365.21 for each member of a couple. The unique feature of the GIS is the income test. For each dollar of family income (excepting O A S income), the G I S benefit is reduced by 50 cents for singles and by 25 cents each for married couples. For 2004, 34.5 percent of O A S recipients also received GIS benefits. The GIS affects retirement incentives in two strong yet dist inct ways. F i r s t , for those who are age 65 or more and would receive the GIS if they retired, labour market earnings w i l l reduce GIS payments by 50 cents on the dollar. Th i s is in addi t ion to the income taxes that would be payable on the labour market earnings, so continued work past age 65 is strongly discouraged by the G I S . The second channel through which the GIS affects retirement incentives is more subtle but perhaps even more important . E x t r a work after age 60 leads to a higher C P P / Q P P pension through the actuarial adjustment. However, each dollar of extra C P P / Q P P income that is earned w i l l lead to a decrease of 50 cents in GIS income, for those who receive GIS . Essentially, for GIS recipients, the value of the actuarial adjustment is cut in half. For this reason, ext ra work past age 60 can have a strong impact on the retirement income received in the future. T h e simulations later in the paper explore this mechanism i n more detail . 2.2.4 T h e Allowance The Allowance is pa id in two circumstances. F i r s t , it is pa id to the 60-64 year o ld spouses of current O A S recipients. Second, it is paid to 60-64 year o ld widows or widowers. The amount pa id is equal to the O A S pension plus the marr ied component of the GIS pension. L ike the G I S , it is clawed back on family income. However, the clawback rates are 75 cents on the dollar for the ' O A S ' por t ion of the Allowance, and 50 cents on the dollar for the ' G I S ' por t ion of the Allowance. The Al lowance affects retirement through the same two channels as described above for the G I S . However, the direct channel of the clawback on labour market earnings is stronger here because of the 75 percent clawback. In addi t ion, the more subtle channel of 5 According to Myles (2004), in 1996 3.1 percent of OAS recipients were subject to the clawback but still received partial OAS benefits while 1.6 were not eligible for OAS because their benefits were fully clawed back. 10 the interaction w i t h C P P / Q P P benefits is much less important for the Allowance because the Allowance can only be received for a m a x i m u m of five years. T h i s means that only five years' wor th of C P P / Q P P actuarial adjustments w i l l be effectively reduced, in contrast to the GIS which reduces them for a l l ages past age 65. 2.2.5 Summary The four components of Canada 's retirement income system each separately embody inter-esting features that influence the decision to retire. However, when the four components are combined, the interactions among the indiv idual components provide some of the sharpest incentives to retire. Descr ibing these interactions is made easier by reference to numerical examples, so we tu rn next to some simulations. 2.3 Simulations The goal of this section of the paper is to quantify the strength of the incentives to retire described i n the previous section. To do so, we take a ' typ ica l ' ind iv idua l and calculate his or her income from al l four components of Canada 's income security system. We then compare the differences i n the incentives when we vary his or her private pension income, amount of lifetime earnings, and continuity of lifetime earnings. F ina l ly , we show some policy simulations to demonstrate the sensitivity of the incentive measures to smal l changes in pol icy parameters. We do not a i m to provide a comprehensive analysis of the incentives to retire, for that is beyond the scope of the paper. Instead, we use the simulations as an i l lustrative tool to point out how the components of Canada 's retirement income system work individual ly and interactively to influence the decision to retire. Because of the i l lustrative nature of the simulations, no at tempt should or can be made to infer nat ional ly representative results from the results presented here. The section begins w i t h a description of the methodology that underlies our calculations. Th i s is followed w i t h the presentation of the s imulat ion results first for the base case, then for several alternative scenarios. 2.3.1 Methodology In order to calculate an individual ' s pension entitlement, we require several pieces of infor-mat ion. We need a complete earnings history back to 1966 (or age 18), sex, age, mar i ta l status, province of residence, and information on private pensions or other income. These pieces of information can then be combined using a pension income calculator to arrive at public pension income in any given year. B y recalculating the pension income for a l l ages after retirement and discounting for time preference and for mor ta l i ty probabilities, 11 we arrive at a measure of the expected net present value of publ ic pension income. We cal l this the Income Security Wea l th ( ISW) corresponding to a part icular retirement age. W h e n this calculat ion is repeated for a l l potential retirement ages, an age profile for I S W can be described and the rate of I S W accrual from year to,year can be derived. B o t h the level of I S W and its rate of accrual are the objects of our attention. We use the pension income calculator developed for and described in Baker, Gruber and M i l l i g a n (2003, 2004a) for our calculations. The calculator first derives the C P P / Q P P benefit, given a lifetime earnings history. Next , it calculates the retirement income for each age during retirement, by assigning the C P P / Q P P benefit, O A S , G I S , and the Allowance both to the worker and his or her spouse. We project benefits into the future assuming they remain constant in real terms. A l l clawbacks are accounted for. The calculator then takes the taxable components of income and applies provincia l and federal taxes to arrive at an after-tax measure of retirement income at a given age. 6 T h e flow of retirement income across ages is discounted using an assumed rate of t ime preference (three percent real) and sex-specific mor ta l i ty probabili t ies (taken from Statistics Canada (2002b)). The output of the calculator is an age-profile of I S W for al l potential retirement ages under consideration. For our calculations, we seek to define a ' typ ica l ' i nd iv idua l in order to characterize retirement incentives. We consider someone in 2002 who is 55 years o ld and lives i n Ontario. Th i s implies that the year of b i r th was 1947, and that the first year of work eligible for the C P P / Q P P is 1966 at age 19. T h e worker is contemplating retirement at some age between 55 and 70. For the earnings history, we take a series of average weekly earnings and annualize i t . 7 In our base case, we assume that the worker earned i n every year from age 22 to the present, w i t h no interruptions. F r o m 18 to 21 we assume zero earnings (proxying for years in school). T h i s means that there are three zeros i n the earnings history, from ages 19 (in 1966) to 21 (in 1968). W h e n projecting earnings into the future from 2002, we assume that earnings stayed constant i n real terms at the 2002 level. We also assume i n our base case that the worker has no income outside of earned income and publ ic pension income -this means no Registered Retirement Savings Plans, employer-provided pensions, or other sources of income. F ina l ly , we assume that the C P P pension is not taken unt i l retirement - no work occurs after the C P P pension is taken. 8 We simulate our base case for married and single males and females. T h e married couples are assumed to each have the same b i r th year and earnings history. W h e n considering the retirement age of the husband, we hold constant the wife's retirement age at 60. Similarly, when considering the wife's retirement age, we hold constant the husband's retirement age at 6 To clarify, the current year tax policy is used when calculating tax payable. 7rThere is no consistent series covering the entire time period necessary for our analysis. We build our series from three CANSIM II series: V78310 for 1965 to 1983, V250810 for 1984 to 2000, and V1597104 for 2001 and 2002. 8Under the CPP and the QPP, you must have stopped work in the month the pension is taken. After that, work may begin again and the pension is not changed. 12 60. Th i s base case is not meant to produce results that are representative for the Canadian economy. Instead, the a im here is to demonstrate how the incentives vary in one simple case. A more complete and representative analysis featuring the fullness of heterogeneity we observe i n the Canad ian labour force is beyond the scope of this paper. In addi t ion to the base case, we conducted three sets of simulations i n which we varied the base scenario i n different dimensions. In the first, we t ry adding sequentially higher amounts of private pension income to examine the effects of the GIS and Allowance clawbacks. In the second, we look at differences across workers of different wage levels by running simulations wi th an earnings history comprised of earnings that are only a certain percentage of the average weekly earnings. F ina l ly , we twist the earnings history i n a different way by studying the effect of ' incomplete ' earnings histories in which the worker had absences from the labour market. These ext ra simulations w i l l help to provide more information on how the retirement incentives vary across individuals . Table 2.1 presents a basic description of our base simulated individuals . We consider the case of a single m a n or woman, w i t h no income aside from publ ic pensions. The first two rows show the probabi l i ty of l iv ing to a certain age, given that the ind iv idua l is currently age 60. Females display greater longevity, wi th the probabi l i ty of surviv ing unt i l age 95 at more than twice that for males, 0.113 to 0.039. Average life expectancy from age 55 (the age at which the condi t ional probabil i ty of l iv ing is 0.50) is 84 for females, and 79 for males. The full survival curves, condit ional on surviving to age 55, are shown in Figure 1. Not only are females different because they have a higher probabi l i ty of survival , but the shape of the survival curve is also different. For example, after age 84, the drop in probabil i ty of survival is greater for women than for men. Because the lifetime pension measures we use w i l l compare positive and negative flows across ages, both the level and the shape of the survival curves w i l l play a role. The rest of Table 2.1 shows pension flows at a part icular age. Because the earnings for our simulated male and female are assumed to be the same, these pension flows could be for a single person of either sex. The th i rd row displays the O A S entitlement, expressed in 2002 dollars. It pays $5,328 per year, starting at age 65. T h e next 4 rows of the table show the C P P entitlement (the simulated ind iv idua l is from Ontario) and the GIS entitlement i f the worker retires at age 60 (in 2007) or age 65 (in 2012). If taken at age 60, the C P P pays $6,335 annually. T h e full G IS amount in 2002 is $6,336, so the C P P payments reduce the GIS payments by $3,167.50 ($6335*0.50), leaving $3,169 in G I S payments start ing at age 65. If the same ind iv idua l continues to work unt i l age 65, the C P P entitlement grows to $9,501. 9 T h i s supplemental $3,166 i n C P P leads to a reduction i n the annual GIS payment of $1,583 ($3,166*0.50), which leaves GIS payments of $1,586 annually. T h i s example gives 9Note that this is greater than the $9,465 maximum pension available in 2002. The pension for our simulated individual is higher because he or she will reach age 65 in 2012, when the maximum pension will be larger. 13 some prel iminary indica t ion of how the GIS and C P P interact w i t h each other to change retirement incentives. T h e extra C P P benefit received for delayed retirement from age 60 to 65 is reduced by half through the G I S . How this change i n annual pension flows changes the lifetime totals is the subject of the simulations that follow. 2.3.2 Base Case Results The results for the base case are presented in Table 2.2. T h e first co lumn shows the level of I S W in 2002 dollars for each case at age 55. T h e columns across the table contain the year-to-year accrual of I S W across different potential ages of retirement, from the point of view of the 55 year old i n 2002. So, for example, the age 57 co lumn contains the difference in I S W for retirement at age 57 and at age 58. A t the right end of the column, we report the final level of I S W at age 70. D o w n the table we consider four family types: single and married males, and single and married females. We start w i t h the single male. For ages 55 and 56, the rate of accrual is $1,269 and $1,073. T h e dropout provision plays an important role here. Fifteen percent of months may be dropped from the C P P / Q P P calculation. F r o m ages 19 to 59, there are 41 years, which generates just over six years of dropouts. Since we assumed no earnings between ages 19 and 21, this means that someone ret ir ing at age 55 has three 'zero' earnings years from ages 19 to 21, then 5 more from ages 55 to 59 before c la iming the C P P / Q P P at age 60. The six dropout years cancel six of the zero years, but two zero years remain in the calculation. A n extra year of work at age 55 therefore replaces one of the zero years. T h i s generates the positive accrual . A t age 57, however, the accrual changes to being very close to zero. T h e reason again is driven by the dropout provision. Retirement at age 57 means that there are three zero years before c la iming at 60. W h e n added to the three zero years from ages 19 to 21, the three years from ages 58 to 60 combine to total six years of zero earnings, which is equal to the number of dropout years. If retirement is delayed one year, therefore, the extra year of work no longer crowds out a zero year from the calculat ion but instead crowds out a high earnings year. Because the difference in earnings between the extra year of work and the year that is replaced is small , the benefit to continued work drops sharply at this age. Simi lar explanations underlie the smal l positive accruals at ages 58 and 59. It is important to stress that the drop in the accrual at age 57 is specific to the setup of this s imulat ion. If a different number of low earnings years were i n the earnings history, the drop would be elsewhere in the profile. The main thrust to be learned here is that the benefit to working an ext ra year between ages 55 and 59 depends heavily on the difference between the extra year of earnings and the year that it replaces i n the C P P / Q P P calculations. Later i n this section, we demonstrate this more direct ly by showing the accrual paths for simulated individuals w i t h several years of work interruptions. 14 A t age 60, the accrual turns negative, reaching -$1,544. There is s t i l l a benefit from extra work through the replacement of a bad earnings year in the dropout mechanism. However, there are negative thrusts that dominate in this s imulat ion which come from two intertwined sources. F i r s t , at age 60, extra t ime at work means that a year of C P P / Q P P benefit receipt is foregone. In compensation, the year of delayed retirement leads to an actuarial adjustment of the C P P / Q P P benefit when it is taken. Ideally, the actuarial adjustment w i l l compensate the worker for the foregone pension income in that year. The second influence on the accrual rate after age 60 is the effect of the C P P / Q P P actuarial adjustment on the GIS benefit. The extra six percent of the pension that is awarded for delaying retirement is counted as income when calculat ing the GIS benefit. T h i s means that 50 cents on the dollar for the actuarial adjustment disappears from the GIS payment. Th i s affects every GIS payment from age 65 unt i l death. Effectively, this shrinks by half the benefit of the actuar ial adjustment and ti l ts the incentives toward negative values. We explore this further below i n simulations where the ind iv idua l is not in receipt of the GIS to separate out the effects of the GIS and the C P P / Q P P actuarial adjustment. A t age 65, the accrual becomes more sharply negative at -$6,503. The reason for the j ump down at age 65 is again the GIS . The extra year of work at age 65 produces an increase in the C P P / Q P P benefit through the actuarial adjustment, which would decrease the GIS as discussed above. However, extra work at age 65 also produces earned income which directly decreases the G I S payment. In fact, at the assumed level of earnings, equal to the average weekly wage i n Canada , the GIS is pushed to zero as it decreases by 50 cents on the dollar of earned income. F r o m ages 65 to 69, the accrual stays negative, but diminishes i n absolute value. T h e decrease is driven by the fact that there are ever fewer years over which the pension w i l l be received. So the extra year of work may change future pension flows, but there are fewer years over which those flows are received. The second row shows the same simulations, but for a marr ied man. A t ages 55 and 56, the accrual is higher than i n the single man case. T h i s occurs because the higher C P P / Q P P benefit that is earned w i t h extra work pays off not only i n a higher benefit for the husband, but also i n a higher survivor pension for his wife after he dies. T h i s amplifies the accrual effect seen for the single man. After reaching age 65, the accrual is more negative for the married man, reaching -$7,904. Earned income after reaching age 65 direct ly claws back the G I S , and a marr ied family has a higher GIS payment if retired. For this reason, extra work after age 65 hurts the marr ied man more because both his and his wife's GIS payments are reduced as he earns income. The th i rd and fourth rows repeat the exercise for females. Because the same earnings profile is used as for the males, the observed differences are dr iven solely by differences i n morta l i ty probabili t ies. W h i l e it is certainly not typ ica l for females to have the same earnings history as males, forcing them to be equal in this s imulat ion allows us to isolate the 15 influence of non-earnings factors by holding the earnings history constant. Because females live longer on average, changes in pensions have longer last ing effects on women. A t ages before 60, the female accrual is higher than the corresponding male s imulat ion because the increment to the C P P / Q P P pension earned by an extra year of work is received over more years, on average. T h i s makes the payoff to extra work higher. F r o m age 60 on, working an extra year means forgoing a year of C P P / Q P P receipt, but gaining a higher C P P / Q P P pension i n every subsequent year. T h e net present value of the ' investment' i n an extra year of work w i l l be different for men and women not only because women are more l ikely to live longer, but also because the shape of their survival curve differs from men. T h i s leads to the results for ages 60 plus that are observed in the table. The accrual for females is less negative from 60 to 64, but more negative after age 6 5 . 1 0 T h i s result is dr iven solely by differences i n morta l i ty across males and females. The difference between single and married females is less pronounced than was the case for single and marr ied men. T h i s is a result of the survivor pension. For males, it is more l ikely that the female w i l l out-survive h im. Th i s means that any increment to his C P P / Q P P pension w i l l be reflected in her survivor pension. However, for females, it is less l ikely that the husband w i l l out-survive her. W h i l e in expectation there w i l l be some survivor benefit received by her husband, i t is smaller in expected value than the survivor benefit of a surviving wife. T h i s serves to shrink the difference between single and married females, as the extra work by the marr ied female does not lead to a big boost in her husband's expected future survivor benefits. 2.3.3 Extended Simulation Results / To complement the simulations i n the base case, we present several extended simulation results to i l luminate and clarify some of the pathways through which Canada 's retirement income security system affects retirement incentives. In a l l cases, we performed the extended simulations on the single male. We made this choice to t ry to simplify the environment, allowing us to focus more easily on the factors under consideration i n a part icular simulation. W i t h o u t spouses, there are fewer 'moving parts' in the simulations. T h e male-female and married-single differences i n the extended simulations look very s imilar to the corresponding difference in Table 2.2. The first extended s imulat ion considers single men w i t h different amounts of private pen-sion income. Because private pension income is included as income for the GIS calculation, 1 0 This occurs because the probability of surviving to very old ages drops more quickly for females than for males. Thus, the positive returns to extra work at very distant ages is discounted very heavily for both males and females, while the foregone pension benefit is discounted more heavily for males than for females. This results in a more negative accrual for females. At age 69, males again become less negative, but this is driven by the exhaustion of GIS benefits - the CPP pension for retirement at age 70 is so large that the entire GIS pension is clawed back. 16 those w i t h higher private pension income w i l l receive less G I S . Over some threshold, the private pension w i l l be sufficient to completely crowd out the GIS payments. The no-GIS case is of great interest when compared to the base case, because w i t h no GIS payments the pure effect of the C P P / Q P P can be seen in isolation from its interaction effect w i th the GIS . Th i s gives a better picture of the channels through which the incentives are generated. The private pension simulations are presented in Table 2.3. The same columns appear as were seen i n Table 2.2, w i t h I S W at age 55, the accrual rates by age, and finally the age 70 I S W level. In the first row, we repeat the base case results for the single male for comparison. In the subsequent four rows we show the results for differing levels of annual private pension income, from $2,000 i n the 2nd row to $8,000 in the 4th r o w . 1 1 In a l l cases, we assign the private pension income to the man start ing at the age of retirement. In the rows for $2,000 to $6,000, the single male s t i l l receives GIS upon retirement. Th is means that the higher C P P / Q P P pension resulting from the extra work between ages 55 and 59 w i l l d imin ish the G I S payment. T h i s serves to attenuate the gain from extra work in this age range relative to the no-GIS case i n the last row. A s one moves from the base case of $0 of private pension income to $6,000, the gain from an extra year of work is actually slightly smaller for higher private pension income. T h i s occurs because the higher private pension income pushes the worker into a higher tax bracket dur ing his retirement years in this s imulat ion. T h i s means that his gain i n C P P / Q P P income from working an extra year is taxed at a higher rate when he has a private pension income, thus diminishing his accrual rate between ages 55 and 59. For the $8,000 row, the accrual is much higher than the other rows, reaching $1,603 at age 55. T h i s occurs because at $8,000 of annual private pension income in these simulations, the single male receives no GIS payments. Thus, the differences we see in row 5 are driven by the removal of the impact of the GIS on retirement incentives. A t ages 55 to 59, the worker no longer sees half of his C P P / Q P P gain from continued work taken away from his GIS payment. T h i s generates the stronger work incentives at a l l ages for the $8,000 case. A t age 61, the ordering of the magnitude of the incentives i n the base case compared to the $6,000 case is reversed. A t age 60, a delay in retirement leads to an increased actuarial adjustment to the C P P / Q P P pension. For those w i t h higher private pension income, there is less GIS income to be affected by the C P P / Q P P actuar ial adjustment, so the accrual is less negative for those w i t h higher private pension income. A t age 61 however, the increased C P P / Q P P payment the ind iv idua l receives leads to the person no longer being eligible for a GIS payment. A s such, the indiv idual receives the full actuarial adjustment by delaying retirement one more year, e l iminat ing the effect of the GIS and resulting in a positive 11Obviously, many Canadians earn more than $8,000 in private pension income. We do not show higher amounts because they show very little difference when compared with the $8,000 case. This is because the GIS payments are already at zero in the $8,000 case, so extra pension income only affects ISW through possibly higher income taxes, and through the OAS clawback for the top few percent of seniors. 17 accrual. F r o m age 65, the negative accrual becomes monotonical ly smaller w i t h increasing private pension income. T h i s results from the direct impact of the private pension income and earnings on the G I S . W i t h a larger private pension income, there is less GIS to be clawed back by earned income. T h i s diminishes the negative effect of the GIS on work incentives. The second set of extended simulations varies the earnings history of the single man. Instead of assigning h i m the full average weekly earnings i n each year, we study cases in which he earned 80 percent, 60 percent, 40 percent, and 20 percent of the national average. Importantly, the smaller wages i n each earnings history are applied only to years prior to age 55; from age 55 we assume that earnings are at the full nat ional average. We make this perhaps odd assumption i n order to hold as much constant as possible to isolate the effect we wish to consider: W h a t is the impact of having a low earnings history compared to a high earnings history on work incentives at older ages? The results of this s imulat ion appear in Table 2.4. A g a i n , i n the first row we have reproduced the single man base case results from Table 2.2. Rows 2 through 5 show the earnings histories of 80 percent through 20 percent of the average earnings. Several factors combine here to generate the observed patterns. A t ages before 60, extra work generates higher C P P / Q P P income at retirement through the earnings formula. A n extra year of work w i l l replace a low-earnings year in the formula, resulting i n a higher C P P / Q P P benefit. The differences across rows i n Table 2.4, therefore, are driven by differences in the level of the low-earnings years that are being replaced. For the s imulat ion w i t h 20% of the average earnings i n each year, the gain to continued work is positive up to age 61, because the extra year of work is replacing a year of very low earnings. For the s imulat ion wi th earnings at 80% of the average earnings level, an extra year of work generates higher accruals than in the base case, but s t i l l turns negative at age 60 as the negative effect of the C P P / Q P P actuarial adjustment and the GIS interaction s t i l l dominate. The final simulations appear i n Table 2.5. In this set of simulations we examine the impact of having different amounts of work interruptions on the accrual of I S W . A t the top of the table we reproduce the base case. In the subsequent four rows we substitute increasing numbers of zero earnings years into the earnings history at ages before 55. In the 2nd row, we replace earnings wi th a zero i n a l l years ending i n a 4 or 9. In the th i rd row we then replace a l l earnings in years ending i n a 3 or 8 w i t h a zero. We continue this pattern down to the last row i n which earnings in four out of every five years have been replaced w i t h a zero. For ages from 55 on, we do not replace earnings w i t h zeros so that the potential earnings from continued work are the same across a l l five rows. Th i s isolates the effect of interruptions in the earnings histories from having low earnings years after age 55. For ages 55 to 59, the benefit of continued work is very similar across a l l four rows 18 wi th work interruptions. T h i s occurs because there are so many zero years i n the earnings history that an extra year of work post-55 always replaces a zero year in the C P P / Q P P formula. T h i s highlights the important impact of work histories featuring interruptions -they tend to increase the work incentives because continued work brings a larger boost in the C P P / Q P P earnings rat ing. F r o m age 60 to age 64, the benefit of continued work varies down the table, w i t h less negative (and more positive) accruals for the simulations wi th more earnings interruptions. Aga in , this makes sense because these workers are more likely to be replacing zeros i n their C P P / Q P P calculation than workers w i t h complete earnings histories. After age 65, the direct effect of earnings on the G I S takes a l l of the accruals to be negative. T h e accruals for the simulations w i t h more interrupted histories are less negative because they continue to be able to use their extra years of earnings to replace zero earnings years in their C P P / Q P P calculation. 2.3.4 Policy Simulations The final set of simulations we present a im to il lustrate the sensit ivity of the incentive measures to smal l changes i n pol icy parameters. It is beyond the scope of this paper to analyze or recommend any pol icy alternatives, so the interpretation of these results should remain narrowly focused on their i l lustrative power. We examine the effect of four separate policy changes. We describe each policy change briefly: • Po l i cy S imula t ion A : Change the actuarial adjustment in the C P P / Q P P from 0.5 percent per month to 0.7 percent per month. • Po l i cy Simula t ion B : In the C P P / Q P P , grant a full ' throw-out ' year for every year of work start ing at age 60. • Po l i cy S imula t ion C : For the GIS clawback calculat ion, use the C P P / Q P P pension entitlement from age 60 rather than the actual C P P / Q P P income. • Po l i cy S imula t ion D : For the GIS clawback calculat ion, exempt labour market earnings from the income measure. The results of the simulations are presented in Table 2.6. Po l i cy A changes the actuarial adjustment in the C P P / Q P P . Th i s should be expected to increase the annual accruals because delayed retirement results i n a larger actuarial adjustment than under the status quo case. Indeed, the simulations show that accruals at every age from the age of entitlement at age 60 are higher than under the status quo system, averaging $1,307 higher. Po l i cy B increases throw out months 'earned' from work after turn ing 60 from 0.15 under the status quo up to 1 full month for every month worked. Because the base case 19 worker has almost the m a x i m u m amount of C P P / Q P P , the abi l i ty to throw out extra years does not have a substantial impact . In simulations not shown here, for workers w i th more incomplete earnings histories the impact of this policy is greater. The th i rd pol icy change is policy C . Th i s change aims to counteract the interaction be-tween the GIS and the C P P / Q P P actuarial adjustment by making G I S payments no longer depend on the age the C P P / Q P P is claimed. This is achieved by using the C P P / Q P P enti-tlement from age 60 i n the clawback calculation for the G I S . T h a t is, the actual C P P / Q P P income received is not used but instead a 'Active' amount calculated as though the ind i -v idua l had retired and claimed C P P / Q P P at age 60 is used instead. A s expected, this policy change has a substantial impact on the key age 60-64 range. N o longer does extra C P P / Q P P earned through the actuarial adjustment have a negative impact on future GIS receipts. Over the five years from 60 to 64, the average increase in the accrual is $2,582. Po l i cy D is the final pol icy change we consider. Th i s pol icy exempts earned income from the GIS clawback calculat ion. For those age 65 and older, the clawback of the GIS on earned incomes provides a strong disincentive to stay in the labour force. W i t h the exemption, the final row of Table 2.6 makes clear that for those aged 65 and over there is a substantial improvement i n the accrual w i t h earned income exempted from the GIS calculation. The improvement averages to $3,755 over the ages 65 to 69. These pol icy simulations have demonstrated that the incentive measures discussed in this paper are sensitive to smal l changes in policy. A full evaluation of various policy alternatives potent ial ly would be very informative, but is beyond the scope of this paper. 2.3.5 Summary of Simulations The simulations i n this section have attempted to demonstrate how Canada 's income secu-r i ty system generates disincentives to remain in the work force at older ages. The profiles of the simulated individuals are not meant to be par t icular ly representative of Canadian older workers overall, but instead were chosen to br ing forward different interesting fea-tures of the retirement income system that affect work incentives. In summary, there are several factors that account for the patterns of accruals across different individuals in the simulations. T h e y are: • Accrua ls increase when extra work replaces a low earnings year through the C P P / Q P P formula. • Accrua ls are larger (in absolute value) for marr ied individuals because extra C P P / Q P P benefits also increase survivor pensions. • Accruals are different for women because any change i n the flow of pension income is received over more years of life and because of differences i n the shape of the survival curve. 20 • Accruals decrease because the C P P / Q P P actuarial adjustment does not sufficiently compensate for the foregone year of pension receipt. • Accruals decrease because the actuarial adjustment of the C P P / Q P P decreases even-tual GIS payments. • Accruals decrease because earnings directly reduce the GIS and Allowance benefits received. It is important to stress that these simulated accruals are not overly large compared to many other countries. One way to compare the accruals across countries is to calculate the impl ic i t tax on (or subsidy to) continued work at each age as the ratio of the accrual to earnings. In our base case simulat ion for single males, continued work at age 55 implies a subsidy rate of 3.6% while continued work at age 65 implies a tax rate of 18%. These impl ic i t taxes are s imilar for individuals in the Uni t ed States (Diamond and Gruber , 1999). However, i n France the subsidy rate to continued work at 55 for a relatively comparable ind iv idua l is 75% while continued work at age 60 implies a tax rate of 66% (Blanchet and Pele, 1999). In Be lg ium, the subsidy rate for continued work at 55 is only 0.2% while the tax rate for continued work at age 60 is 59% (Pestieau and Stijns, 1999). W i t h i n our Canad ian results, the most s t r iking feature is that the disincentives to con-tinue working are strongest among GIS recipients, who represent the bo t tom one th i rd of the income dis t r ibut ion among individuals age 65 and over. T h i s is exemplified by the simulations presented i n Table 2.3, whereby accruals are most negative among individuals w i th the lowest private pension income. These findings suggest that the work disincentives are strongest among the worst off retirees; perhaps those who would benefit most from a few extra years of work to increase subsequent retirement income. 2.4 International Evidence In order to best understand the role played by the retirement incentives uncovered in the simulations, we provide i n this section a detailed review of the empir ica l evidence on public pensions and retirement. T h e scope of the review covers the ma in international evidence in order to provide context for the Canadian evidence. B r o a d surveys on Social Security are provided by Feldstein and L iebman (2002) and specifically on the labour market impact of social insurance programs by Krueger and Meyer (2002). E m p i r i c a l work est imating the effects of income security programs on labour supply can be roughly d iv ided into four groups. The first set of studies t r ied to estimate the retirement impact of pension wealth without focusing on the subst i tut ion effect of pension accruals. Contemporaneous w i t h the first group of studies, several papers take a more 's tructural ' approach by est imating parameters from an explicit model of behaviour. A sharp break 21 in the nature of the research occurred i n the early 1990s as a result of the confluence of three factors. F i r s t the Hea l th and Retirement Study became available for researchers in the Un i t ed States, which provided much richer data to study retirement. Second, the Stock and Wise (1990) 'opt ion value' framework introduced the dynamic and forward looking nature of the retirement decision. F ina l ly , expanded comput ing power facilitated analysis of vast arrays of micro data in the Heal th and Retirement S tudy and from other sources. Fol lowing this break, the th i rd set of studies buil t on the earlier work by incorporating dynamic measures of pension accruals into the analysis. F ina l ly , the fourth set of studies has at tempted to use natural experiments - pol icy changes - to estimate the sensitivity of retirement decisions to changes in incentives. To summarize the results of the body of research, we have provided a l is t ing of each study and its core result i n Table 2.7. 2.4.1 Research on related topics Before embarking on a tour through the evidence on public pensions and retirement, we briefly review some Canad ian research on two related topics i n order to provide more con-text. T h e first is the effect of private pensions. Pesando and Gunderson (1988, 1991) map out pension wealth profiles for common employer-provided pensions in Canada (flat benefit and final earnings plans) w i t h the goal of identifying the incentives to work created by the structure of pensions plans. Unl ike Lazear (1983), who finds that pension wealth peaks at the date that an ind iv idua l first qualifies for early retirement, Pesando and Gunderson (1991) find that there is no clear peak age for pension wealth, and i n fact the pension wealth profiles exhibit discontinuities. G i v e n these profiles, Pesando and Gunderson (1988) argue that mandatory retirement bans and related legislation l imi t the abi l i ty of employers to de-sign pension plans that create work disincentives through postponed retirement provisions that reduce pension wealth for retirement at older ages. The second related topic is mandatory retirement. Kesse lman (2004) and Gunderson (2003) provide thorough overviews of mandatory retirement practices in Canada. Kessel-man (2004) argues against contractual mandatory retirement (ie. w i th in an agreement between employers and employees in the form of a pension or collective agreement) as this often forces workers to leave their jobs earlier than desired and that banning mandatory retirement could help reduce the pressures associated w i t h earlier retirement (eg. fiscal pressures or potential sk i l l shortages). Gunderson (2003) argues against age discr iminat ion but does not oppose contractual mandatory retirement as it may be preferred by workers and employers. Gr ierson and Shannon (2004) provide evidence that banning mandatory retirement i n Canada would have l i t t le effect on the share of older people working, using the implementat ion of mandatory retirement bans i n M a n i t o b a and Quebec to identify this effect. 22 2.4.2 Early evidence on pension wealth and retirement The first body of research on public pensions and retirement we examine is characterized by a focus on the level of pension wealth rather than the dynamic incentives that are featured in later work. Below, we review the main findings and provide a cr i t ica l analysis of the key papers i n the literature. We start w i t h research that uses t ime series data, and therefore relies on var ia t ion i n income security parameters or benefits over t ime to identify the effect of income security programs on labour supply. For example, Pellechio (1979) uses Canadian t ime series data from 1946 to 1975 to determine the effect of I S W on retirement using O A S , as wel l as the in t roduct ion of C P P / Q P P , on non-part ic ipat ion rates of individuals age 65 and over. H e finds that I S W has a positive and marginal ly significant impact on non-part icipat ion and that the introduct ion of C P P / Q P P had a positive but insignificant impact on non-part ic ipat ion rates. However, use of t ime series data can be misleading i n determining the impact of income security programs. Over this period, there was a general tendency for par t ic ipat ion rates of the elderly to decline and for benefit generosity to increase. T h e positive association found between I S W and non-part ic ipat ion may spuriously reflect the coincidence of the two trends which may have been changing for unrelated causes. 1 2 In addi t ion to this, the introduct ion of C P P / Q P P coincides w i t h a reduction i n O A S el igibi l i ty age to 65 and the introduction of GIS . The effects C P P / Q P P relative to these other programs cannot be easily disentangled. Studies that use cross-section data covering individuals i n one year of data use differ-ences in income security benefits across individuals to identify the effects of income security programs. Since policies are the same for everyone at a given point in t ime, cross-sectional variat ion i n income security benefits actually reflects differences i n ind iv idua l characteristics - such as earnings histories and mar i ta l status - that determine benefits. A s such, these studies may actually be identifying the effects of these characteristics on labour supply rather than the effect of income security programs. To overcome this problem, many studies have t r ied using panel data to identify the parameters from changes i n income security programs over t ime that influence the benefits individuals receive. The use of panel data, however, w i l l not necessarily overcome the identification issues associated wi th cross-sectional or t ime series data. Bosk in (1977), for example, uses a sample of white married males i n their sixties from panel data to estimate the effects of Socia l Security benefits on the probabi l i ty of retirement and finds that benefits have a positive and significant effect on retirement. However, when time effects are controlled for, the Social Security benefits are found to have a much smaller or insignificant effect, indicat ing that the original estimates are largely picking up general trends in benefits and retirement over t ime, and that identification of the effects of Social 1 2 A linear time trend is included in Pellechio's model. The coefficient on the time trend is positive but not significantly different from zero. 23 Security benefits is relying on differences between individuals . A n influential paper by B o s k i n and H u r d (1984) uses U . S . panel data from the Retire-ment His tory Survey ( R H S ) from 1969-1973, a t ime when Social Security benefits grew rapidly due to ad hoc changes to U . S . Social Security legislation and the over-indexation of benefits. T h e rising benefits would have created an unexpected increase i n Social Security wealth which could be expected to induce early retirement. Us ing a sample of white married men, B o s k i n and H u r d estimate the probabil i ty of retirement by age (59-65). They c la im their results strongly support the hypothesis that unexpected changes i n Social Security wealth have a positive effect on retirement and that younger individuals too young to c la im benefit and w i t h low assets would be l i t t le affected by the changes i n S S W . However the positive estimates of the marginal effects of Social Security wealth on retirement are not statist ically significant. In fact, only negative effects are s tat is t ical ly different from zero. Furthermore, since their econometric model controls for cohort effects and estimates the re-tirement probabi l i ty by age, they effectively absorb any changes over t ime in benefits across individuals , and therefore rely on variat ion in benefits related to cross-sectional differences in ind iv idua l characteristics for identification. Bosk in and H u r d (1978) also use U . S . panel data from the R H S (1969-1971) to estimate the probabi l i ty of making the transi t ion from work to retirement. Us ing a sample of white married males age 62-65, they find that higher Social Security benefits imply a higher probabil i ty of retirement. However there appear to be several identification issues for their econometric model . T h e ma in identification issue is a lack of clar i ty about the identification of the effect of Socia l Security benefits. F i rs t , they include gross wages, Social Security benefit levels, and net wages in the model. G iven that the latter two variables are actually functions of the gross wage, it is not clear how they can disentangle the effects of each variable separately. Second, they use instrumental variables to control for the fact that tastes for work may influence both Social Security benefits (through work habits and the marginal tax rate) and the probabi l i ty of retirement. However, the instruments are merely a nonlinear combinat ion of past wages and other income which implies they are not picking up any exogenous var ia t ion i n benefits. D i a m o n d and Hausman (1984) use the U . S . Na t iona l Long i tud ina l Survey of Older M e n (1966-1978) to determine the effects of bad health, unemployment, and permanent income on retirement i n two stages. F i rs t , they use a sample of males age 45-69 to estimate hazard models for the t ransi t ion into retirement. The find that Socia l Security benefits, interacted wi th age indicators to account for age-related provisions, have large positive effects on the probabi l i ty of retirement. However, since age indicators do not enter the hazard model separately, the Socia l Security variables may s imply be picking up spikes i n the retirement hazards not related to Socia l Security programs. Furthermore, the key finding of this study is that simulations based on these results imply roughly half of a l l retirements of men age 24 62-64 are due to the availabil i ty of reduced Social Security benefits. However, this is based on a comparison of estimated retirement rates to estimated retirement rates setting the Social Security benefit to zero, effectively el iminat ing any general age effects of retirement decisions which should not be at t r ibuted to Social Security. It also appears to be the case that their model is largely picking up the general t rend of increasing Social Security benefits and decreasing par t ic ipat ion rates. For example, when the model is estimated using a sample from 1973-1978 (a period w i t h relatively large drops i n par t ic ipat ion rates and relatively large increases i n Social Security benefits) the effect of Social Security is much higher than i n the full 1966-1978 sample. In the second stage, D i a m o n d and Hausman use a probit model and a competing risks model to estimate the probabi l i ty of older unemployed workers entering either retirement or new employment. S imi lar to the first stage, they find that higher Socia l Security benefits (interacted w i t h age) lead to a higher probabil i ty of retirement. Burtless (1986) uses U . S . panel data from the R H S to estimate the age at which an ind iv idua l retires. He uses an econometric model that describes the u t i l i ty maximiz ing behaviour of individuals when choosing their retirement age and allows for anticipated increases in Socia l Security benefits to affect individuals differently than unanticipated in-creases i n benefits. Based on a sample of men aged 58 to 63 i n 1969, (excluding farmers, disabled men, men who retired before age 54, and men receiving substantial income from welfare programs, federal c iv i l service pensions and rai l road retirement benefits,) Burtless finds that increasing Socia l Security benefits by 20 percent (10 percent in 1969 and 1972) have a short run effect of reducing the expected retirement age by 0.09 years (roughly 1 month) and a long run effect of reducing the retirement age by 0.17 years (roughly 2 months) and increasing the probabi l i ty of retirement at ages 62 and 65 by about 2 percent. G iven these estimates, Burtless concludes that the observed decline in retirement ages and employment of older males over this period cannot be explained by Social Security alone. A s w i t h other studies i n this era, however, Burtless 's reliance on changes i n Social Secu-r i ty over t ime should make us question whether the results found merely reflect a spurious correlation between rising benefits and falling retirement ages. A few recent studies follow methodologies comparable to this early set of studies. Comp-ton (2000) attempts to determine the effect of Canada 's income security programs on re-tirement decisions using data from the Survey of Labour and Income Dynamics 1993-1996. Compton uses a sample of individuals age 50 and over to estimate a hazard model for en-t ry to retirement and uses a sample of individuals age 55 and over to estimate an ordered probit for exit from full t ime employment to full t ime employment, par t ia l retirement, or full retirement. A weakness of this approach is the lack of sufficient historical earnings information to accurately construct the earnings history and pension entitlement of each indiv idual . 25 The key covariate used to identify the effect of income security programs in these mod-els is the individual ' s expected C P P benefits. Overal l , C o m p t o n finds that income security programs have no effect on retirement decisions. It is not clear, however, whether an in -dividual ' s benefits are measured appropriately. To measure expected benefits, Compton estimates the C P P / Q P P benefit level that an ind iv idua l could expect at age 60 using re-ported benefits from S L I D , and then imputes this amount for individuals in the sample (even i f they are over age 60). W h i l e the estimation procedure does account for the fact that she can only observe C P P / Q P P benefits among recipients, the variables used in es-t imat ing C P P / Q P P are not exogenous to the retirement decision and are also included in the econometric models. T h i s leaves the identification of the retirement effect to depend on the shapes of the assumed dis tr ibut ion, which is a fairly tenuous base for inference. Furthermore, C o m p t o n attempts to separate wealth and subst i tut ion effects by including variables for investment income and home ownership i n the model . It is not clear that in -vestment income can capture wealth effects, nor is it clear that these variables can capture the wealth effects of income security programs. F ina l ly , the results show insignificant effects of financial incentives on retirement only because the standard errors are very large. The lack of precision means that large effects cannot be ruled out based on this evidence. T o m p a (1999) investigates various determinants of indiv iduals ' decisions to take up C P P / Q P P retirement benefits. T o m p a uses a sample of individuals turn ing age 60 between 1987 and 1994 from the Longi tud ina l Adminis t ra t ive Databank at Statistics Canada, a subset of the T l F a m i l y F i l e ( T I F F is a yearly cross-sectional file of a l l tax filers and their families) to estimate a hazard model for the durat ion from age 59 to the age of first take-up of benefits. T o m p a finds that for women, higher levels of C P P / Q P P income actually reduce the l ikel ihood of tak ing up C P P / Q P P , while for men the level of C P P / Q P P income is insignificant for the take-up decision. It is not clear, however, that Tompa's analysis provides us w i t h much information about how levels of C P P / Q P P benefits affect the take-up decision, and especially retirement, since the C P P / Q P P income variable included i n the model is the observed benefit an indiv idual is collecting, not the benefit that a person would be entit led to at each age. T o m p a does find that poor job prospects, health and joint retirement decisions are important determinants of C P P / Q P P takeup. 2.4.3 Structural models of retirement Structura l est imation involves a more explicit use of an economic model of behaviour. The advantage of s t ructural est imation is that one may study the long run equi l ibr ium effects of changes i n pol icy and perform policy experiments on the model , once the structural parameters are estimated. T h e disadvantages of this approach are a sometimes less rigor-ous attention to the identification of parameters and the requirement of sometimes strong assumptions about the 'correct ' specification of the model . 26 Fields and M i t c h e l l (1984) estimate a s tructural model to determine the effects of several changes to Socia l Security in the U S introduced i n 1983. T h e reforms included raising the normal retirement age, delaying cost of l iv ing adjustments, lowering early retirement benefits and increasing late retirement payments. To obtain the parameter values of the structural model, they use a sample of white, married, male private sector employees between age 59 and 61 i n 1969, following individuals to 1979, to estimate a condi t ional ordered logit model of retirement ages. Retirement age is defined as the age at which a worker left his 1969 job. To make the da ta relevant to the early 1980s, they inflate earnings profiles to reflect the increase i n average wages between the 1970s and 1982, inflate private pension benefits for inflation, apply tax formulas for 1982, and use 1982 Social Security benefit rules to calculate benefits. T h e y find that the reforms mentioned above would have a positive impact on the average retirement age. For example, increasing the normal retirement age from 65 to 68 would on average increase retirement ages by 1.6 months. W h i l e it is not clear that the data used to estimate this model is appropriate, the interesting point to take from this s tudy is that a three year increase in the normal retirement age is not expected to increase retirement ages by three years, rather it may only increase retirement ages by a couple months. Simi lar to Fie lds and M i t c h e l l (1984), Gus tman and Steinmeier (1985) also estimate a s tructural model to examine the effects of the U . S . program reforms i n 1983. Thei r model is more parsimonious i n that they use a modified life-cycle model w i t h a C E S ut i l i ty function over consumption and leisure ( individual characteristics w i l l affect the weight placed on leisure) and the s tandard lifetime budget constraint, but allow individuals to work full t ime or part t ime (at a lower wage). T h e y use a sample of white males who are not self-employed when working full-time from the R H S (1969-1975) to estimate the parameters of this model. T h e y find that increasing the normal retirement age from age 65 to age 67 would reduce the probabil i ty of entering retirement at age 65 by roughly 4 percentage points. A t age 67, the probabil i ty of entering retirement would increase by roughly 1 percentage point, indicat ing results s imilar to Fie lds and M i t c h e l l (1984) whereby increasing the normal retirement age w i l l not increase the average retirement ages by the same amount. To summarize, the s t ructural modeling line of research has yielded significant insights into what would happen under proposed 'counterfactuaP reforms. B y estimating deep structural parameters governing behaviour, a good base is made for inference about reforms such as extending the retirement age. 2.4.4 Estimation of accrual and level effects of pensions A newer generation of studies attempts to estimate the effects of income security programs on labour market behaviour by estimating reduced form models that incorporate both the incentives for an ind iv idua l to continue working and measures of income security wealth 27 using cross section or panel data. Th is literature follows the work of Stock and Wise (1990) whereby indiv iduals ' decision to retire depends on the opt ion value of continuing to work. Tha t is, an ind iv idua l w i l l compare the expected present value of ret ir ing to the value of continued work w i t h the opt ion to retire i n the future. Recent studies have several features in common; these features be discussed here prior to discussing the ind iv idua l papers. The reduced form models typical ly take the form of a probit model i n which the depen-dent variable is d u m m y variable indicat ing whether an ind iv idua l enters retirement. The key explanatory variable is a forward looking measure intended to capture the incentives of income security programs and is measured i n one of three ways. F i r s t , the simplest measure is referred to as a single year accrual which captures the effect of another year of work on future income security benefits. Th i s is defined as the expected present discounted value of income security benefits i f a person were to retire i n the next year, less the expected present discounted value of benefits i f a person were to retire in the current year. The single year accrual measure w i l l be positive i f continuing to work for an addi t ional year increases the future benefits from income security benefits. A second incentives measure, peak value, accounts for the income security benefits accrual possible i f the ind iv idua l retires many years into the future. For this measure, the expected present discounted value of benefits for a l l possible future retirement ages is evaluated and an op t imal date of retirement based on these benefits is determined. The peak value is then the expected present discounted value of benefits i f the ind iv idua l retires at this op t imal date less the value i f the ind iv idua l retires immediately. The th i rd incentives measure is similar to the peak value but for this measure the expected present discounted value of wage and non-labour income i n addi t ion to benefits for al l possible future retirement ages is evaluated to determine an op t imal date of retirement. Furthermore, an indirect u t i l i ty function is placed over wages and benefits, often using behavioural parameters estimated by Stock and Wise (1990). T h e option value is then the expected present discounted value of indirect u t i l i ty over wages and benefits i f the ind iv idua l retires at an op t imal date less the value if the ind iv idua l retires immediately. W h i l e this th i rd measure is more parsimonious than the first two, it is computat ional ly cumbersome and requires relatively more assumptions regarding individuals ' expectations of future income. In these retirement models there are several identification issues. F i rs t , individuals are more l ikely to prefer retirement as they age. A linear age variable w i l l potentially capture this effect if preferences for leisure evolve l inearly w i t h age. Wage earnings may also proxy for differences i n the preference for work. However, bo th age and wages enter in the calculat ion of income security benefits i n the incentives measures. A s such, including age and wage measures as covariates may make it more difficult to isolate the effects of program incentives from worker heterogeneity. We can expect that the inclusion of such 28 variables w i l l result i n understating the effect of program incentives. Similar ly, the option value measure, as it captures the full financial incentive on retirement of both future wage earnings and income security benefits combined, may reflect i n part this wage proxy for heterogeneity, rather than the financial retirement incentives. A second issue arises because it is common to find that the retirement rate at the 'normal ' retirement age (or age of first el igibi l i ty for benefits) is much larger than predicted rate based on financial incentives alone. T h i s l ikely reflects a l iqu id i ty constraint - many employees have not saved enough to retire without receiving Social Security or employer-provided pension benefits. Inclusion of indicator variables for each age allows for such jumps i n retirement rates, but we might not be able to isolate the effect of financial incentive measures from plan el igibi l i ty ages. Gruber and Wise (2004) provide a country-by-country analysis of retirement behaviour that follows the s tandard approach described here. The goal of the Gruber and Wise volume is to provide comparable estimates of the effect that income security programs have on retirement behaviour across countries, following up on the work of Gruber and Wise (1999) which identified the income security program incentives to retire early in several countries. In each chapter, the authors from the country estimate the probabi l i ty of entering re-tirement as it depends on the incentives found in each country's income security programs. The retirement probits for each country include a measure of the incentives that income security programs provide (ie. the one year accrual, peak value or opt ion value measure), a control variable for income security wealth, and controls for age, earnings, industry sector, and demographics such as sex and education. The common finding among most of the country analyses (some of which are described i n more detai l below) is that the retirement incentives inherent i n most income security programs are strongly related to retirement and this finding rarely depends on whether age indicators or a linear age variable is used i n the specification. In a few cases, however, the estimated effect of incentives is not statistically significant and of the wrong sign. In many cases, the effect of income security wealth is not statist ically significant from zero and is often of the wrong sign. These results may part ly be due to relatively l i t t le variat ion in income security wealth i n some countries, while there is more variat ion i n the incentives measures. However, the overwhelming impression from the twelve country studies is the consistency of a positive impact of I S W levels on retirement and a negative impact of higher accruals. Baker et a l . (2004a) provide the Canadian analysis of the effect of income security programs on retirement behaviour. The pr imary data source used in this analysis is the Longi tud ina l Worker F i l e developed by the Business and Labour Marke t Analys is D i v i -sion of Statistics Canada . The data set combines information from three administrative data files: the T-4 file of Revenue Canada, the Record of Employment file of H R D C and 29 the Long i tud ina l Employment Analys is P rogram of Statistics Canada . T h e Longi tud ina l Worker F i l e provides information on individuals ' wages and salaries, 3-digit industry codes, province and size of establishment for each job the ind iv idua l holds i n a given year, their age, sex and job tenure. The focus of the analysis is the per iod 1985-1995. Separate samples of males and females aged 55-69 i n 1985 are drawn, and then younger cohorts of ind iv id -uals are added as they tu rn 55 i n 1986-1995. The sample excludes agricul tural workers, individuals in other p r imary industries and individuals w i t h missing age, sex, or province variables. The sample is selected condit ional on working (defined as positive T-4 earnings). If an ind iv idua l has positive earnings in one year and zero earnings i n the next, the year of positive earnings is considered the retirement year . 1 3 In the empir ica l analysis, incentives measures are constructed as described above. There are a few things to note about the income profiles required i n the construction of these mea-sures. W h i l e earnings histories are available for each ind iv idua l back to 1978, information on earnings back to 1966 and in the future must be imputed for the purposes of calculat-ing C P P / Q P P entitlements. Similar ly, several assumptions regarding spouses are required to impute non-labour income profiles. For O A S entitlement, the authors are not able to deal w i th residency requirement. For C P P / Q P P entitlements, the authors are not able to account for years spent using disabi l i ty benefits or years spent caring for children. Baker et a l . (2004a) provide several specifications for the retirement probit , estimated separately for men and women, using either the one year accrual , peak value or option value incentives measures and either a linear age variable or age indicator variables. In al l specifications they find that income security wealth has a positive and significant effect on retirement and that incentives to continue working have a negative and significant effect on entry to retirement. T h e largest estimates for men are found when using the one-year accrual measure w i t h a linear age variable, whereby a US$1,000 increase i n accrual is associated wi th a 2.21 percentage point decline in retirement rates. A US$1,000 increase in accrual is associated w i t h a 1.52 percentage point decline i n retirement when age indicator variables are included in the model . The reduced effect of these incentives when age dummies are included i n the model may be associated wi th these variables p icking up the effect of the income security program's el igibi l i ty ages. For example, the size of the estimate for the age 60 indicator is consistently and substantially larger than the age 59 indicator, imply ing that the avai labi l i ty of C P P / Q P P benefits may have some impact on retirement decisions beyond what can be explained by the changes i n financial incentives alone. A s discussed above, however, it is impossible to separate the effect of program el igibi l i ty from general tastes and trends i n retirement behaviour. In a complementary analysis, Baker et al . (2003) find similar , al though slightly smaller 1 3Baker et al. (2003) test other definitions of entry to retirement that include EI earnings with labour market earnings, or included earnings below a minimum threshold with zero earnings, and do not find that the different definition significantly affects the direction of the results. 30 effects of income security programs on retirement behaviour when using a sample of men age 55-64. For example, they find that a $1,000 increase in the one-year accrual results in a 0.98 percentage point decrease i n retirement rates among men, compared to the 1.52 percentage point decrease associated wi th a US$1,000 increase i n the one-year accrual in the comparable specification described above. T h e smaller estimates result in part from a much richer set of controls for the earnings of each ind iv idua l . T h i s provides one of the main findings of the paper - that richer earnings controls may attenuate some of the estimated parameters observed i n the literature. The 2003 piece also extends their 2004 analysis by checking the results i n subsamples of the main data to see if the pattern of results across samples conforms w i t h the patterns predicted by theory and economic intui t ion. T h e y find that the incentives measures have a larger effect among individuals w i t h a lower probabi l i ty of being covered by an employer-provided pension plan ( R P P ) than individuals w i t h a high probabi l i ty of R P P coverage. Furthermore, the incentives measures have the largest effect on individuals in the lowest average lifetime income quartile. In contrast, among men i n higher income quartiles, the incentives measures have a positive or insignificant effect. These estimates indicate that individuals l ikely to be more dependent on income from income security programs are more sensitive to program rules. It also suggests that the work disincentives bite hardest among those who may most need extra income in their retirement years. Coi le and Gruber (2004) find similar results to Baker et al . (2003, 2004a) for the Uni t ed States using da ta from the Hea l th and Retirement Survey (1992-1998). The H R S is a survey of individuals age 51-61 i n 1992 and their spouses. T h e y construct person-year observations for each year between 1980 and 1997 i n which the ind iv idua l is between the ages of 55 and 69 and working at the beginning of the year, using information from the earnings histories available i n the H R S . To note, they exclude any individuals that would have retired prior to age 55 or appear to have re-entered the labour force following retirement. Coi le and Gruber (2004) also provide several specifications for their probit model, com-parable to Baker et a l . (2004a), but find that Social Security has a much smaller effect. In al l specifications, Socia l Security wealth has a positive estimated effect, however the effect is not always significantly different from zero. The results for the incentives measures are more ambiguous. W h e n using a linear age variable in the regression, a $1000 increase i n the one year accrual variable is associated wi th a positive effect on retirement rates (0.0015 percentage point increase) among m e n . 1 4 The negative effect of the peak value and option value incentives measures are significant in most specifications, however the peak value does not significantly affect retirement rates among women. Interestingly, probits using age indi -cator variables consistently demonstrate relatively large increases i n retirement rates at age 60, similar to those i n Canada , yet the individuals i n this s tudy are not eligible for Social 1 4There is also a positive but insignificant effect when age indicator variables are used in the regression. 31 Security benefits un t i l age 62. 5 Coile and Gruber (2000) estimate similar retirement models that incorporate both Social Security and private (employer-provided) pension incentives, and find that the results differ from those for Socia l Security alone. The incentives variables used i n their probit models i n -clude income from private pensions, derived from the pension determination information in the H R S . 1 6 In comparison to the results using only Social Security, incorporat ing pensions results i n a negative and significant coefficient on the one-year accrual variable. Further-more, the effect of the peak value incentives is half as large and the coefficient on Social Security wealth is significant when pensions are included in the incentives measure, sug-gesting that people are less responsive to changes in pension incentives than Social Security wealth. Coi le and Gruber (2000) also provide specifications w i t h addi t ional control variables including an indicator of poor health, affecting the magnitude and significance of their estimates. L i m i t e d at tention has been paid to the role of health i n Canada . Campol ie t i (2002) estimates probit models for the part icipat ion of older workers i n Canada as it depends on an indicator of disabi l i ty status, but does not control for publ ic pensions or any.other form of income i n models. Magee (2002) uses information i n S L I D to determine the effects of self-reported health and disabil i ty on several reasons for job separation and finds that health or work-related disabi l i ty does not have a significant effect on the probabil i ty of job separation due to retirement. These results may not te l l us much about retirement, however, since job separation due to illness and disabil i ty is also a possible response to the job separation question and those who separate from a job due to illness and simultaneously retire may not be associated w i t h retirement i n this study. Several studies using European data have found results s imilar to those for the U . S . and Canada. Borsch-Supan (2000) uses data from the G e r m a n Socio-Economic Panel (1984-1996) to estimate retirement probits similar to those found i n Gruber and Wise (2004). Us ing a sample of 5 5 ^ 0 70 year old men and women, they find that the incentives to delay retirement created by income security programs (measured using the option value) have negative and significant effects on the probabil i ty of retirement. Furthermore, their simulations suggest that if Germany were to move to an actuarial ly fair benefit formula, early retirement would occur less frequently. The model does not separately control for income security wealth. Interestingly, there are no spikes i n the pat tern of estimated coefficients reported for age indicators despite spikes in Germany 's d is t r ibut ion of retirement ages at age 60 and 65. T h i s would indicate that observed increases in retirement at these ages are due solely to the incentives found i n income security programs and i n the absence of these 1 5There are much larger spikes in the estimated retirement rates at age 62 and 65 than at age 60. 1 6 They omit observations that are missing pension data, thereby dropping 40% of the observations used in the income security analysis. Most of these omitted observations were individuals from smaller firms with lower retirement rates, earnings, education and tenure. 32 programs there would merely be a pattern that reflects older workers being more likely to retire than younger workers. Borsch-Supan (2000) also uses country level data to provide a quali tat ive investigation into the relationship between the incentives found in income security programs and labour force par t ic ipat ion i n Europe. Borsch-Supan finds that i n almost a l l cases spikes in the dis t r ibut ion of retirement ages can be identified w i t h ages i n which certain pension rules start of cease to apply, as these rules often create kinks i n the income security accrual profiles. (Here, income security accrual is measured as the percentage change i n income security wealth for one year of delayed retirement for the average worker w i th in a country.) B l o n d a l and Scarpet ta (1998) attempt to provide a quanti tat ive cross-country t ime se-ries analysis of the relationship between income security program incentives and retirement behaviour i n their work for the O E C D . They estimate labour force par t ic ipat ion rates of older men (55-64) using an accrual rate (measured as the percent change i n old-age pen-sion benefits for a 55 year o ld male by working for 10 more years), replacement rates of unemployment, disabili ty, old-age pension and special early retirement benefits, the unem-ployment rate of prime age males, the share of the prime age popula t ion in the total working age populat ion, union density, and the standard age of entitlement to old-age pensions as explanatory variables. T h e y find that higher accrual rates are significantly associated wi th higher par t ic ipat ion rates. W h i l e i n theory this type of cross-country analysis could be useful as it relies on cross-country var ia t ion i n income security programs to identify the effects of these programs, in practice it is not clear that this type of analysis is informative as it requires extreme simplifications of very complex programs i n order to measure variables i n a way that would be comparable across countries. For example, i n this study, the accrual rate is measured for a specific type of ind iv idua l who delays retirement for a long per iod of t ime. In many countries, accrual profiles are highly non-linear. A s such, unless accrual profiles in a l l countries are linear it is not clear that the analysis w i l l adequately measure the incentives provided by income security programs. A s an example, based on calculations of the accrual rate using income security wealth calculations by Blanchet and Pele (1999, Table 3.5) the accrual rate for a 55 year o ld postponing retirement un t i l 65 is -25% while the accrual rate for postponing retirement unt i l 60 is 7% and postponement un t i l age 57 is 16%. In this case, the accrual rate used by B l o n d a l and Scarpetta would not properly capture the retirement incentives contained i n France's income security programs. 2.4.5 Natural Experiments A s discussed earlier i n this section of the paper, the key problem w i t h using cross-sectional and t ime series da ta for the study of retirement is the difficulty i n ensuring that the es-t imated effects of income security programs are not merely picking up differences across 33 individuals or general trends i n retirement and benefits over t ime. One potential solution to this problem is to identify natural experiments - situations i n which program changes affect one group i n the populat ion (the treatment group) but not a different yet similar group (control group) - and see how the treatment group behaves differently in response to program changes. In general, these studies have shown mixed evidence on the effect of income security programs on retirement behaviour. One of the first examples of natural experiments being used to identify the effects of income security programs on labour supply is a study by Krueger and Pischke (1992) in which they rely on a change to U . S . Social Security provisions in 1977 that reduced ben-efits for some individuals based only on their year of b i r th . Specifically, prior to 1977, a s i tuat ion referred to as double indexation existed because average monthly earnings ( A M E ) were increasing w i t h inflat ion and under the benefit calculat ion rules, the replacement rate attached to each bracket i n the benefits formula was set to increase wi th inflation. A s a result, workers who postponed retirement could increase their benefits at a rate greater than what would be actuarial ly fair. To eliminate this double indexat ion, the amendment introduced a new benefit formula i n which average indexed month ly earnings ( A I M E ) are used instead of A M E ( A I M E is indexed to average wage growth) and the replacement rates were held constant while the brackets for each replacement rate were adjusted for changes in the average annual wage. These amendments were phased in over 5 years. The affected group, therefore, was individuals born 1917-1921. T h i s group became known as the 'notch generation.' In effect, these changes resulted in an exogenous and unexpected reduction in I S W for the notch generation that can act as an experiment for identifying the effects of I S W . Krueger and Pischke use a sample of 60-68 year o ld males from the M a r c h C P S (1976-1988) to estimate logit models for labour force par t ic ipat ion rates that control for both I S W and the growth i n I S W for delaying retirement one year, as well as age and time effects. The i r findings suggest that the negative relationship typica l ly found between I S W and labour supply is spurious. For example, when year effects are not controlled for, I S W is found to have a negative effect on part icipat ion rates and the growth in I S W for delayed retirement has a positive and significant effect. However, when t ime effects are accounted for, bo th I S W and growth i n I S W have an insignificant and positive effect on part icipation. Furthermore, they use logit models controll ing for coverage by the 1977 changes i n benefits (similar i n concept to a D D estimator) and find that the changes had a negative effect on labour force par t ic ipat ion rates and weeks worked and a positive effect on the proport ion of individuals report ing being retired. Several natural experiments have been found in Canada . Baker and Benjamin (1999a,b) rely on differences in the t im ing of program changes i n C P P and Q P P to identify the effects of income security programs on retirement behaviour. Baker and Benjamin (1999a) look 34 at the in t roduct ion of early retirement provisions in C P P / Q P P which were introduced earlier i n Quebec (1984) than the rest of Canada (1987). Us ing a standard Differences in Differences ( D D ) analysis (wi th the control group defined by geography), they found that while the in t roduct ion of these provisions significantly increased the rate of C P P take-up, it d id not have a significant impact on labour force par t ic ipat ion, indicat ing that those taking up the early pension benefits were only marginal ly attached to the labour force and the new provisions d id not affect labour force behaviour. One problem w i t h this type of analysis, however, is that i t is only able to capture the immediate effects of pol icy changes. In this case, many individuals age 60-64 at the t ime that early retirement provisions were introduced would have already planned their retirement based on their savings, employer-provided pensions, and the previously existing policy rules. A s such, only individuals w i t h l i t t le other income would have the immediate incentive to collect benefits. Younger cohorts would have t ime to adjust their lifetime leisure-consumption plans and employers may adjust their pension provisions in light of the changes to C P P / Q P P . 1 7 Thus , while no immediate effects on labour supply are found, a long-run analysis may find some effect. 1 8 Baker and Benjamin (1999b) consider the effect of the e l iminat ion of the earnings test in C P P (1975) and Q P P (1977) for individuals age 65-69, potential ly making work more attractive. T h e D D estimator used here again relies on geography to define the control group. A D D D . estimator, which uses individuals age 60-64 as an addi t ional control group, is also used. F r o m their analysis, it is clear that the removal of the earnings test resulted i n higher take-up rates but had no significant effect on retirement, employment or part icipat ion rates. F r o m the D D D analysis, they conclude that the removal of the test is associated w i t h large shifts from part year full t ime work to full year full t ime work among older men. However, it is not clear that the 60-64 year old group is a suitable control group because forward-looking 60-64 year olds would find their incentives affected by this policy change as well . Baker (2002) uses D D estimators to determine the effect that the introduct ion of the S P A i n 1975 had on the labour market behaviour of marr ied men (age 65-75 wi th spouses age 60-64) and women (age 60-64 wi th spouses age 65-75). For the control groups, it is assumed that any individuals who d id not immediately qualify for S P A benefits (i.e. men wi th younger or older spouses, single men, and women w i t h younger spouses or younger women w i t h spouses over 65) would not be affected by the pol icy change. Baker finds that men eligible for the S P A experienced a relative decline in their labour force part icipat ion wi th the in t roduct ion of the program. A m o n g eligible women, they find some negative effect 1 7Gruber (2000) tests whether early retirement provisions affected younger males and found no effect. 1 8 For example, Baker and Benjamin (1999a) demonstrate that after the early retirement provisions were introduced, retirement hazards at age 60-64 increased. However, it is not possible to disentangle general trends in retirement from the long-run effects of policy change. 35 of the S P A on women's par t ic ipat ion in that par t ic ipat ion rates of S P A eligible women d id not rise w i t h the par t ic ipat ion rates of other women. T h i s finding is not consistent across control groups, however, and the largest effects are found when women age 60-64 wi th spouses under age 65 are used as control groups. G iven that the spouses i n the treatment group (men 65-75) were affected by the el iminat ion of the earnings test i n 1975, these results could be biased by the contemporaneous reform. In summary, evidence based on natural experiments has been mixed and largely incon-clusive. In some cases, such as the study of the 'notch' generation i n the Un i t ed States by Krueger and Pischke, l i t t le effect of income security programs was found. In other cases, such as Baker 's s tudy of the Spouse's Allowance in Canada, a strong and significant effect was found. T h e weakness of this approach for the study of retirement may lie in the t ime necessary to respond to a change i n policy. The short run impact may differ from the long run impact . 2.5 Conclusions This paper has described Canada 's retirement income security system and provided an em-pir ical context in which to understand its impacts on the retirement decisions of elderly Canadians. In the simulations, we find that many components of the system act indepen-dently and in concert to change the incentives to retire. W h i l e these incentives are not large compared to some European countries, it is worth reflection that they are strongest for those who receive the G I S . Moreover, Baker et al . (2003) find that the reaction of these lower-income individuals to the work disincentives is stronger than i t is for higher-income individuals . Since G I S recipients are from the bo t tom one t h i rd of the senior income distr i-but ion, this means that the strongest disincentives are faced by those who perhaps might most benefit from some extra income in their retirement years. Look ing at the international empirical research record, we find a fairly consistent and robust pattern of evidence suggesting that financial incentives i n public pension programs affect retirement decisions. W h i l e other factors such as family, health, and community likely enter the decision to retire i n addi t ion to the financial motive, the clearest and most direct pol icy lever to affect retirement decisions is through the structure of public pension programs. T h i s means that decisions made in the presence of these work disincentives differ from those that would be made under a 'neutral ' system, imply ing a policy-induced inefficiency. We close by not ing that some inefficiency i n the provision of retirement income to seniors may be unavoidable i n a practice. A s wi th the provision of many public programs, there is a trade-off between equity and efficiency. For many low-income seniors, the retirement income security system i n Canada provides a significant por t ion of their total income and 36 contributes significantly to poverty alleviation. Sensible reforms will seek a balance between equity and efficiency improvements. 37 Table 2.1. Basic statistics on simulated individuals Ages 60 65- 75 85 95 M a l e P robab i l i ty of l iv ing to age, given alive at 55 0.959 0.895 0.662 0.299 0.039 Female P robab i l i ty of l iv ing to age, given alive at 55 0.975 0.938 0.796 0.493 0.113 A n n u a l O A S entitlement 0 5328 5328 5328 5328 A n n u a l C P P entitlement i f c l a im C P P at 60 6335 6335 6335 6335 6335 A n n u a l G I S entitlement i f c l a im C P P at 60 0 3169 3169 3169 3169 A n n u a l C P P entitlement if c la im C P P at 65 0 9501 9501 9501 9501 A n n u a l G I S entitlement i f c l a im C P P at 65 0 1586 1586 1586 1586 Note . — A l l dollar values in 2002 Canadian dollars. A l l reported entitlements are annual flows at 2002 rates for singles. Table 2.2. Base case simulations ISW accruals ISW at 55 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 at 70 Males Single 142109 1269 1073 71 23 16 -1544 -2737 -3430 -4050 -4488 -6503 -6328 -6147 -5955 -4794 98585 Married 249141 1637 1355 93 39 13 -1064 -1929 -2577 -3201 -3749 -7904 -7762 -7790 -7850 -7492 200960 Females Single 174951 1457 1233 81 28 17 -1084 -2452 -3202 -3886 -4392 -6573 -6465 -6352 -6232 -4787 132342 Married 250039 1654 1340 104 49 -1 -734 -1886 -2602 -3297 -3917 -8114 -8045 -8152 -8295 -7662 200481 Note. — A l l dollar values in 2002 Canadian dollars. Reported are the one-year accruals of ISW from ages 55 to 69, as well as the age 55 and age 70 ISW level amounts. Table 2.3. Pr ivate pension simulations ISW accruals ISW at 55 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 at 70 Base Case 142109 1269 1073 71 23 16 -1544 -2737 -3430 -4050 -4488 -6503 -6328 -6147 -5955 -4794 98585 Private Pension Amount $2,000 131118 1007 844 54 24 15 -2341 -2884 -3332 -3787 -4175 -5563 -4719 -3129 -3335 -3512 96285 $4,000 117926 902 756 47 20 15 -2025 -2531 -2999 -2769 -1194 -2324 -2609 -2866 -3087 -3282 93980 $6,000 104481 898 764 47 12 20 -544 350 -88 -516 -889 -2033 -2333 -2604 -2838 -3048 91679 $8,000 96950 1603 1360 84 24 32 1079 603 151 -291 -673 -1743 -2056 -2341 -2590 -2813 89379 Note. — A l l dollar values in 2002 Canadian dollars. Reported are the one-year accruals of ISW from ages 55 to 69. as well as the age 55 and age 70 ISW level amounts. The simulations are for single males. Table 2.4. Range of Earnings Simulations ISW accruals ISW at 55 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 at 70 Base Case 142109 1269 1073 71 23 16 -1544 -2737 -3430 -4050 -4488 -6503 -6328 -6147 -5955 -4794 98585 Earnings History 80% 137017 1359 1186 329 283 274 -994 -1943 -3031 -3652 -4152 -6378 -6196 -6009 -5815 -5624 96654 60% 127608 1358 1241 582 565 558 -318 -851 -1640 -2284 -3388 -6227 -6027 -5836 -5632 -5433 94276 40% 116777 1367 1272 848 825 832 132 -299 -723 -1176 -1579 -5271 -5293 -5641 -5433 -5218 91420 20% 105946 1367 1319 1100 1100 1091 582 260 -76 -450 -789 -4745 -4618 -4493 -4529 -4506 88559 Note. — A l l dollar values in 2002 Canadian dollars. Reported are the one-year accruals of ISW from ages 55 to 69, as well as the age 55 and age 70 ISW level amounts. The simulations are for single males. Table 2.5. Work interruption simulations ISW accruals ISW at 55 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 at 70 Base Case 142109 1269 1073 71 23 16 -1544 -2737 -3430 -4050 -4488 -6503 -6328 -6147 -5955 -4794 98585 Years Worked 80% 132187 1366 1359 1367 1358 1367 -713 -1496 -2838 -3529 -4008 -5965 -5799 -5630 -5881 -5157 97988 60% 123610 1367 1358 1367 1359 1366 -126 -662 -1363 -2052 -3140 -5910 -5726 -5545 -5365 -5182 95356 40% 115018 1358 1367 1359 1366 1359 226 -218 -656 -1152 -1539 -4974 -5280 -5460 -5263 -5068 92443 20% 105082 1367 1359 1366 1359 1367 626 301 -50 -438 -766 -4564 -4444 -4352 -4450 -4685 89078 Note. — A l l dollar values in 2002 Canadian dollars. Reported are the one-year accruals of ISW from ages 55 to 69, as well as the age 55 and age 70 ISW.level amounts. The simulations are for single males. Table 2.6. Illustrative Pol icy Simulations ISW accruals ISW at 55 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 at 70 Base Case 142109 1269 1073 71 23 16 -1544 -2737 -3430 -4050 -4488 -6503 -6328 -6147 -5955 -4794 98585 Policy A 134174 1131 950 55 24 15 1253 -249 -2116 -3072 -3853 -6027 -5917 -5799 -3985 -3133 103451 B 142109 1269 1073 71 23 16 -1484 -2688 -3403 -4035 -4484 -6516 -6333 -6155 -5963 -4582 98918 C 140747 1974 1669 107 36 28 944 -205 -846 -1428 -1805 -4983 -5139 -5268 -5362 -5431 115038 D 142109 1269 1073 71 23 16 -1544 -2737 -3430 -4050 -4488 -2347 -2375 -2395 -2401 -1436 117358 Note. — A l l dollar values in 2002 Canadian dollars. Reported are the one-year accruals of ISW from ages 55 to 69, as well as the age 55 and age 70 ISW level amounts. The simulations are for single males. Policies A through D are described in the text. Table 2.7. Summary of Retirement Studies Paper Primary- Analys is K e y Results D a t a Source Bosk in (1977) U . S . , P S I D (1968-72) Es t imate the effect of Social Security bene-fits on the probabil i ty of retirement using a logit model . • A n increase i n benefits from $3000 to $4000 per year raises the probabi l i ty of retirement from 7.5 to 16%. Pellechio (1979) Canad ian t ime series data (1946-1975), Economic C o u n c i l Est imates effect of I S W from O A S and introduct ion of C P P on nonpart icipat ion rates age 65+ using a 2SLS model • A $2300 increase in I S W (1971 dollars) raises the nonparticipa-t ion rate from 67.9 to 73.2 (6.1 percentage points). • Introduct ion of C P P / Q P P raised the nonpart ic ipat ion rate. Bosk in U . S . and H u r d (1969-(1978) 1971) R H S Est imate the effect of Social Security bene-fits on the probabil-i ty of retirement using a logit model wi th i n -strumental variables. • A $1000 increase i n benefits raises the probabi l i ty of retirement by 8 percentage points over two years. (The estimated probabil i ty of retirement is 11.5%.) D i a m o n d U . S . N L S and Haus- of Older man (1984) M e n , (1966-1978) Est imate hazard mod-els for the transit ion into retirement and probit models for the transi t ion from unem-ployment to either re-tirement or employ-ment. • A b o u t half of a l l retirements of men age 62-64 are due to the avail-abi l i ty of reduced Social Security benefits. Bosk in U . S . and H u r d (1969-(1984) 1972) R H S Est imate effect of I S W on the probabil i ty of retirement by age us-ing a logit model that captures cohort effects. • A US$10000 increase i n I S W (in 1969 dollars) raises the retirement rate by 7.8 percentage points. • The increase i n Social Security benefits can account for the en-tire 8.2 percentage point decline i n par t ic ipat ion rates of older men from 1968 to 1973. . 44 Table 2.7 - Cont inued Paper P r i m a r y Analys is K e y Results D a t a Source Fields and U . S . R H S Est imate the effect of K» Increasing the normal retirement M i t c h e l l (1969- Social Security pro- age from 65 to 68 would on aver-(1984) 1979) gram reforms using a age increase retirement ages by 1.6 s t ructural model. months. Gus tman U . S . and Stein- (1969 meier 1975) (1985) R H S Est imate the effect of Social Security pro-gram reforms using a s tructural model. • Increasing the normal retirement age from 65 to 67 would reduce the probabi l i ty of retirement at 65 by 4 percentage points while increas-ing the probabi l i ty of retirement at 67 by 1 percentage point. Burtless U . S . R H S Est imate short and (1986) (1969- long run effects of 1979) Social Security ben-efit increases on retirement age us-ing an econometric model that accounts for anticipated and unanticipated benefit increases. • In the short run, increasing ben-efits by 10% i n 1969 and 1972 re-duced the retirement age by 0.09 years • In the long run, increasing bene-fits above their 1969 level by 20% reduced the retirement age by 0.17 years (2 months) and raised the l ikel ihood of retirement at ages 62 and 65 by 2%. T o m p a Canad ian Estimates the effect of (1999) L A D C P P / Q P P benefits on (1987- C P P / Q P P take-up us-1994) ing durat ion models. • For women, higher levels of C P P income actual ly reduce the l ikel i -hood of tak ing up C P P . • For men, the level of C P P income has no effect on the C P P / Q P P take-up decision C o m p t o n Canad ian Estimates the effect of • F inds income security programs (2001) S L I D C P P / Q P P benefits on have no effect on retirement (1993- entry to retirement us-1996) ing hazard and ordered probit models B l o n d a l Various Es t imate labour force and Scar- sources part icipat ion rates petta across countries to (1998) determine the effects of income security program incentives on retirement behaviour • Increasing the pension accrual rate by 10 percentage points would increase labour force part icipat ion rates of men 55-64 byl .3-2.5 per-centage points. 45 Table 2.7 - Cont inued Paper P r i m a r y Analys is K e y Results D a t a Source Borsch- Germany Est imate retirement Supan G S O E P probit models to de-(2000) (1984- termine the effect of 1996) incentives to continue work based on income security programs, controls for random effect. • A n increase in incentives is as-sociated w i t h a decrease i n the re-tirement rate. • Introducing an actuarial ly fair benefit formula to Germanys pen-sion system would cause retire-ment at ages 59 and below to drop from 28.6% to 18.5%. Coi le and U S H R S Gruber (1992-(2000) 1998) Es t imate retirement probit models to de-termine the effects of Social Security wealth and incentives to continue work based on Social Security and private pensions. • A US$1000 increase i n incen-tives is associated w i t h a 0.00025-0.00044 percentage point decrease i n the retirement rate among men when pensions are included in the incentives measures and a 0.00065 to -0.00047 percentage point change when pensions are not included. • A US$10000 increase in Social Security wealth is associated wi th a 0.025-0.057 percentage point in -crease i n retirement rates among men when pensions are included in the incentives measures and a 0.32-0.41 percentage point in -crease when pensions are not in -cluded. Baker, Canad ian Gruber L W F and M i l l i - (1978-gan (2003) 1996) Es t imate retirement probit models to de-termine the effects of income security wealth and associated incentives to continue work. • A $1000 increase in income se-cur i ty wealth accrual is associ-ated w i t h a 0.19-2.43 percentage point decrease in the retirement rate among men and a 0.01-3.48 percentage point decrease among women. • A $10000 increase in income se-curi ty wealth is associated wi th a -0.32 to 0.69 percentage point change i n the retirement rate among men and a -0.12 to 0.83 percentage point change among women. 46 Table 2.7 - Cont inued Paper P r i m a r y Analys is K e y Results D a t a Source Coi le and U S H R S Est imate retirement Gruber (1992- probit models to de-(2004) 1998) termine the effects of Social Security wealth and associated incentives to continue work. • A US$1000 increase in Social Se-curi ty wealth accrual is associated w i t h a -0.0005 to 0.0015 percent-age point change in the retirement rate among men and a -0.00005 to 0.0006 percentage point change among women. • A US$10000 increase in Social Security wealth is associated wi th a 0.11 to 0.35 percentage point increase i n the retirement rate among men and a 0.18 to 0.26 percentage point increase among women. Baker, Canad ian Est imate retirement Gruber L W F probit models to de-and M i l l i - (1978- termine the effects gan (2004) 1996) of income security wealth arid associated incentives to continue work. • A US$1000 increase in income security wealth accrual is associ-ated w i t h a 0.67 2.21 percentage point decrease in the retirement rate among men and a 0.24-2.06 percentage point decrease among women. • A US$10000 increase in in -come security wealth is associ-ated w i t h a 0.01-0.09 percentage point increase i n the retirement rate among men and a 0.02-0.09 percentage point increase among women. Krueger and P i s -chke (1992) U . S . M a r c h C P S (1976-1988), co-hort level da ta Es t imate effect of I S W and one year accrual for delayed retirement on retire-ment indicators using a logit model. Iden-tification from Notch generation. • G r o w t h i n U . S . Social Security benefits i n the 1970s could explain less than 1/6 of the observed de-cline i n male labour force part ici-pat ion rates (based on the largest estimates). 47 Table 2.7 - Cont inued Paper P r i m a r y Analys is K e y Results D a t a Source Baker and Benjamin (1999a) Canad ian S C F Indi-v idua l files, (1982-83 and 1985-1990) D - D analysis, estimate effect of introducing early retirement provi-sions to C P P / Q P P on labour force part icipa-t ion of older men. • Introduct ion of early retirement provisions led to a significant in -crease i n benefit take-up among men age 60-64, but d id not in -crease the incidence of early retire-ment. Baker and Benjamin (1999b) Canad ian S C F Cen-sus family files, (1972, 74, 76, 78 and 80). D - D analysis, estimate effect of el iminat ing retirement test (tax-back of earnings) from C P P / Q P P on labour supply of older men. • E l i m i n a t i n g the retirement test led to large shifts from part year full t ime work to full year full t ime work among older men. Baker (2002) Canad ian S C F Cen-sus family files, (1972, 74, 76, 78 and 80), excludes Quebec. D - D analysis, estimate effect of introducing S P A on labour market attachment of older i n -dividuals • T h e par t ic ipat ion rates of S P A eligible men fell 7-11 percentage points relative to other men be-tween 1972 and 1980. 6-7 percent-age points of this decrease are at-t r ibutable to the introduction of the S P A . • The par t ic ipat ion rates of S P A eligible women remained fairly constant from 1972 to 1980 while the par t ic ipat ion rates of other women rose. 4-9 percentage points of the 5-10 percentage point diver-gence i n par t ic ipat ion rates may be a t t r ibuted to the introduction of the S P A . Note Data set acronyms used include: CPS Current Population Survey; HRS Health and Re-tirement Survey; LWF Longitudinal Worker File; NLS National Longitudinal Study; PSID Panel Study of Income Dynamics; RHS Retirement History Survey; SCF Survey of Consumer Finances; SLID Survey of Labour and Income Dynamics. 48 C h a p t e r 3 The Effects of Health and Financial Incentives on Retirement Decisions in Canada 3.1 Introduction Concerns about the social and economic consequences of popula t ion aging have sparked an interest i n the labour force behaviour of older workers. Current ly , nearly one th i rd of Canada 's adult popula t ion is at least 55 years o ld and this proport ion is expected to increase to 40% by 2026 (Habtu , 2002) as individuals are l iv ing longer and the baby boom generation is entering their s ixt ies . 1 Popula t ion aging is l ikely to lead to higher expenditures on health care, income security programs, and other publ ic ly funded programs used by our seniors. 2 The long-term viab i l i ty of these programs, however, is being called into question as most programs continue to be financed on a 'pay as you go' basis and there w i l l be relatively fewer young people available to finance these programs. Furthermore, labour force par t ic ipat ion rates of older workers reached record lows i n the mid-1990s. Between 1976 and 1996 par t ic ipat ion rates of older workers over age 55 fell by 25% and have only recently begun to increase. 3 G i v e n these long-term trends there are concerns that sk i l l "The statistical analysis in this chapter relies on Statistics Canada microdata, made available through the British Columbia Inter-University Research Data Centre. This study reflects the views of the author and does not reflect the opinions of Statistics Canada. 1I refer to the 'baby boom' generation as individuals born 1946-1959. The 'baby bust' refers to those born 1960-1973 and the 'echo of the baby boom' refers to individuals born after 1973. These definitions follow Beaudry and Parent (2000). J To note, older individuals' health care is much more costly than health care for younger individuals. See Figure 3.1 for some details. 3Calculated using CANSIM II series V2461251. The decline in participation rates is attributed to older men, as in Figure 3.2. The participation rates of older women increased over time, most recently for women age 60-64 as in Figure 3.3 50 shortages may occur as the baby boomers leave the labour force. These various concerns have spurred interest i n finding ways to ensure that the po-tential supply of labour by older workers is efficiently ut i l ized. Baker et a l . (2003, 2004a) demonstrate that Canada 's Income Security programs contain provisions that may distort the labour market decisions of older individuals , by providing incentives to enter early re-tirement. Morisset te et a l . (2004) find that over one quarter of retirees might have changed their decision to retire i f they had been able to reduce their work schedule without their employer-provided pension being affected. Furthermore, Morisset te et a l . (2004) find that one-third of recent retirees retired for health reasons. Ensur ing that the labour market for older workers runs as efficiently as possible, through changes to publ ic pol icy or other means, first requires a solid understanding of these factors that influence retirement decisions. 4 In this paper I investigate how several factors may influence the retirement decision wi th a focus on an individual ' s health and their financial incentives to retire associated wi th income security programs, employer-provided pensions, and other sources of income. W h i l e the effects of income security programs on retirement decisions have been examined in the Canad ian literature, l imi ted attention has been paid to the role of health and employer-provided pensions. Th i s paper contributes to the li terature on retirement decisions by providing a more comprehensive examination of how these various factors affect retirement decisions. A l so , I address several problems associated w i t h identifying the effects of each factor by exploi t ing the panel aspect of the Survey of Labou r and Income Dynamics , the pr imary source of data used i n this study. I find that the structure of Canada's income security programs have importance incentive effects for retirement. I also find that employer-provided pensions significantly contribute to ind iv idua l decisions to retire. Furthermore, I find that health has a clear and strong influence on retirement decisions. In section 3.2 I begin by providing some context for this s tudy w i t h a brief review of the li terature related to retirement incentives. Th i s is followed by an overview of the environment i n which older workers' decisions are made (in section 3.3). In section 3.4 I provide a theoretical framework for understanding the retirement decision. In section 3.5 I describe the econometric model that is estimated i n this paper, the data used, and measurement of key variables. In section 3.6 I present the results of the statist ical analysis. F ina l ly , I offer some conclusions. 3.2 Previous Literature In recent years the li terature on retirement behaviour has mushroomed, examining the impacts of income security programs, health, employer-provided pensions, family charac-401der workers who retire for health reasons may or may not be considered a potential source of labour supply depending on how or whether their health interferes with their ability to work or receive training for employment suited to their health. 51 teristics, and several other factors on the labour supply of older workers. Despite the large international literature, however, there remains only a handful of Canadian studies exam-ining retirement behaviour, and this literature has focussed almost exclusively on the role of income security programs. As several papers have reviewed different aspects of this lit-erature in detail, in this section I provide a brief review of some key studies with a focus on the Canadian literature.5 In Chapter 2 I provided a thorough survey of the international evidence on public pensions and retirement, the largest branch of research on retirement behaviour. Recent Canadian studies have provided conflicting evidence on the effects of Canada's income security programs on retirement. On one hand, Baker et al. (2003, 2004a) use Canadian panel data (1985-1995) and an option value framework (described later in this paper) to show that wealth from income security benefits have a positive and significant effect on entry to retirement and that the incentives to continue working in income security programs have a negative and significant effect on retirement. Baker (2002) finds that the introduction of the SPA in 1975 had a negative effect on the labour force participation of men eligible for the program. On the other hand, Baker (2002) also finds that the SPA did not have a clear effect on the labour supply of eligible women. Furthermore; Baker and Benjamin (1999a) find that while the introduction of early retirement provisions to CPP/QPP significantly increased the rate of CPP/QPP take-up, it did not have a significant impact on labour force participation. Their results indicate that those taking up the early pension benefits were only marginally attached to the labour force anyway. Baker and Benjamin (1999b) consider the effect of the elimination of the earnings test in CPP (1975) and QPP (1977), potentially making work more attractive. They find that elimination of the earnings test had no significant effect on retirement, employment or participation rates. Removal of the test, however, is associated with shifts from part year full time work to full year full time work among older men. Despite the mixed evidence in Canada, the bulk of the international evidence suggests the structure of public pensions contributes to the retirement decision. A second branch of research has examined the role of health in the retirement decision. Several U.S. studies have incorporated health into their analysis (including Coile and Gruber (2000) and Dwyer and Mitchell (1999)), however recent research in the U.S. finds that the availability of employer-provided retiree health insurance is an important factor that may act as a constraint for the retirement decision.6 In Canada, the availability of retiree health 5In Chapter 2, we reviewed the empirical evidence on public pensions and retirement. Feldstein and Liebman (2002) provide a broad survey of public pensions while Krueger and Meyer (2002) focus on the labour market impacts of social insurance programs. Lumsdaine and Mitchell (1999) provide an overview of the theoretical issues involved in analysing retirement behaviour and the empirical evidence on the labour supply effects of public and private pensions. Lazear (1986) provides an earlier review of the retirement literature with a focus on private pensions. Currie and Madrian (1999) provide an overview of the literature linking health and labour market behaviour. 6 A U.S. study by Gruber and Madrian (1995) suggest that one year of continued employer-provided post-retirement health insurance increases the retirement rate by 30%. Blau and Gilleskie (2001) and Blau and 52 insurance is not expected to be an important constraint for the retirement decision given Canada's universal health care system. Us ing Canad ian data from the Survey of Labour and Income Dynamics , Magee (2002) finds that poor health and work-related disabil i ty do not have a significant effect on the probabil i ty of job separation due to retirement. These results may not tel l us much about retirement, however, since job separation due to illness and disabi l i ty is also a possible response to the job separation question i n S L I D , and those who separate from a job due to illness and simultaneously retire may not be associated wi th retirement i n this study. Us ing Canadian data from the N P H S , Campol i e t i (2002) estimates probit models for the par t ic ipat ion of older men as it depends on an indicator of disabil i ty status, and finds that disabi l i ty has a large negative effect on labour force part icipation. Campol ie t i , however, does not control for public pensions or any other form of income in the models. In the li terature on health and retirement behaviour, there are concerns that the esti-mated effect of health on retirement behaviour may be biased. One potential source of bias is referred to as just if icat ion bias which arises if individuals are l ikely to rationalize their retirement by report ing poor health. Dwyer and M i t c h e l l (1999) test self-ratings of poor health against more objective measures and find that there is no evidence i n support of the justification hypothesis. Another related source of bias is measurement error i n self-reported measures of health, which may be more problematic when using self-assessed measures of general health. Baker et al . (2004b) use N P H S data i n conjunction w i t h Ontar io diag-nostic/ treatment records to examine the extent of this problem and find significantly high measurement error i n objective measures of health, inc luding false (positive and negative) reports of cancer. Furthermore, they find that for several conditions the misreport ing of a health condi t ion (false positive reporting) is negatively correlated w i t h work (as opposed to not work ing) . 7 A s a t h i rd potential source of bias, there is some evidence that health improves w i t h retirement, par t icular ly among blue collar workers (Marsha l l and Clarke, 1998). If this is the case, the endogeneity of health would lead to a downward bias i n the estimated effect of health i f health is reported after entry to retirement. In a th i rd branch of retirement research, a handful of Canad ian studies have exam-ined how employer-provided pension plans may affect an individual ' s retirement decision. Pesando and Gunderson (1988, 1991) map out pension wealth profiles for common employer-provided pensions i n Canada ( including flat benefit and final earnings plans) w i t h the goal of identifying the incentives to work created by the structure of pension plans. Unl ike Lazear (1983), who finds that pension wealth peaks at the date that an ind iv idua l first qualifies for early retirement, Pesando and Gunderson (1991) find that there is no clear peak age Gilleskie (2003) also find that the availability of retiree health insurance has an impact on the employment behaviour of older men in the U.S.. 7This result is only significant for conditions with higher potential for personal subjective assessment (eg. back problems). 53 for pension wealth, and in fact the pension wealth profiles exhibit discontinuities. Pesando and Gunderson, however, do not attempt to estimate the effects that the incentives created by pension plans have on observed retirement behaviour. Whether pension plans will have an effect is not obvious, as Morissette and Zhang (2004) have recently shown that many individuals are not aware of whether a pension plan is provided by their employer. 3.3 Canada's Retirement Income System Canada's retirement income system consists of several parts. First, Canada has a set of in-come security programs that provide retirement income to the elderly including the Canada Pension Plan and Quebec Pension Plan (CPP/QPP), Old Age Security (OAS), the Guaran-teed Income Supplement (GIS) and the (Spousal) Allowance (SPA). Second, the government provides tax assistance for savings through employer-provided pension plans (or Registered Pension Plans) and Registered Retirement Savings Plans. Many individuals also rely on other forms of private savings for their retirement. 3.3.1 Income Security Programs The various components of Canada's income security programs and the financial incentives they create to enter retirement were thoroughly examined in Chapter 2 and are only briefly outlined here.8 The largest component of the system is the CPP/QPP, through which indi-viduals over the age of 60 are eligible for an earnings-related pension upon retirement. Any individual over the age of 65 (who meets residency requirements) is also eligible for the OAS pension (a cash transfer that is clawed back from high income individuals). The GIS is an income-tested benefit also available to individuals over 65, with benefit amounts depending on marital status and family income. Finally, the Allowance (SPA) is another income-tested benefit available to 60-64 year old spouses of OAS recipients and widows/widowers. Canada's income security programs appear to be an important source of income for retirees. More than three quarters of Canadians age 60-68 who are not in the labour force receive CPP/QPP benefits (see Table 3.1). One quarter of these individuals receive GIS benefits and half receive OAS. Roughly half of individuals age 60-68 out of the labour force receive the majority of their income from income security programs. Given their importance as a source of retirement income and the retirement incentives built into the structure of these programs (demonstrated in Chapter 2) we would expect Canada's income security programs to affect retirement decisions. 8 The financial incentives found in Canada's income security programs are further discussed in Appendix A. 54 3.3.2 Employer-Provided Pension Plans In Canada, less than half of paid workers are covered by a registered pension plan (RPP). The proportion of female paid workers covered by pension plans remained fairly constant during the 1990s around 40%. For men however, the proportion covered by pension plans dropped from 49% in 1991 to 41% in 2001.9 The vast majority of employer-provided pension plans in Canada take the form of defined benefit plans which provide a monthly benefit that typically depends on the years a person has spent with the employer, the wages they earn and the individual's age of retirement. In 1996, 88% of pension plans were defined benefit plans while 10% were defined contribution plans. Over recent years a larger proportion of plans have taken the form of defined contribution plans, for which pension benefits vary depending on the contributions accumulated for each individual and the return on investment. In 2001, 14% of pension plans were defined contribution plans. Pension benefits are an important source of income for retirees. 44% of older individuals out of the labour force have income from employer provided pension plans and this is the major source of income for 29% of these individuals (see Table 3.1). 3.3.3 Other Sources of Income Private savings may also be held in the form of Registered Retirement Savings Plans which are not accounted for in this study. Although RRSPs have become much more commonly used by Canadians, it does not appear to be the case that RRSPs will account for a large portion of current retirees' incomes. Using a sample of economic families with at least one member between the ages of 45 and 64 from the 1999 Survey of Financial Security, while 74% of these families reported that they have ever had an RRSP, only 64% of families held positive RRSP investments. The median RRSP holding was only $11,500. When RRIFs are included with RRSP holdings, the median for these economic families rises to $13,000. The median value of RRIF and RRSP holding for economic families with at least one member over 65 was zero as only 27% of these families have positive RRSP holdings. Aside from the resources already discussed here, some retirees may rely on investment income, provincial social assistance, disability benefits, or other transfers to fund their retirement. While a large portion of older individuals receive some form of investment income, very few rely on this as their main source of income. Only 9% of individuals age 60-68 who are not in the labour force report investment income as their major source of income (see Table 3.1). Those age 50-59 who are not in the labour force rely more heavily on investment income. Of these individuals, 15% report investment income as their major source of income, although only 32% have any investment income. Reliance on other forms of income such as disability benefits or provincial social assistance is much more common 9Pension coverage from Statistics Canada (2002a) 55 among younger retirees. 3.4 Modelling the Retirement Decision There are several ways to model the retirement decision.10 In the simplest model, individuals choose a path of lifetime consumption and labour supply to maximize utility subject to the constraint that the discounted present value of lifetime income equals the discounted present value of lifetime consumption. Changes in total lifetime income are expected to have wealth effects that allow the individual to enjoy more leisure. Given hours constraints faced by many individuals, it is expected that such wealth effects will reduce the number of years that an individual works.11 More recent modelling of lifetime labour supply recognizes that many forms of income depend on the age at which an individual retires in that additional years of work may raise remaining lifetime income. To capture the effect that the accrual of lifetime income could have on the retirement decision, and following the work of Stock and Wise (1990), the retirement decision is viewed as one in which individuals will choose to retire in the current period if the expected present value of retiring immediately is greater than the expected present value of continuing to work and holding the option of retiring in the future. Each period, if the individual continues to work, this decision is re-evaluated. At age t, individuals will choose a retirement age (r) to maximize the expected present value of lifetime (indirect) utility Vt(r) given by r - 1 T EtVt{r) = ^ / 3 s - V ( S | i ) C / u ; ( y s , ^ , X s ) + ^ ^ - t 7 r ( S | t ) C / r ( y s , J B s ( r ) , X s ) (3.1) s=t s=r where Uw and Ur represent the indirect utility of future income while working and while retired respectively, ws is the wage earned at age s, Bs(r) are retirement benefits at age s that depend on the age of retirement, ys is non-labour income at age s, and Xs represents individual characteristics. Future utility is discounted for the probability of survival to age s given survival to age t (7r(s|£)) and discounted for preferences at B — 1/(1 + 5). Assuming individual characteristics act as preference shifters that are additively sepa-rable from the utility gained from income sources and separating health status from other preference shifters, (3.1) becomes r - 1 EtVt(r) = YlBa~t^\t)[UV)(ys,Ws) + 'ywHa + va] s=t 10Several models are described in Lumsdaine and Mitchell (1999). 1 1 Gustman and Steinmeier (1983, 1984) show that the majority of workers face hours constraints that would prevent them from gradually phasing out of full time jobs into retirement. 56 + £ Bs-^{s\t)[Ur{ys, Bs(r)) + lrHs + Q (3.2) where the preference shifter Hs represents the health status of an individual at time s, and vs and £ 3 are other preference shifters. The decision to retire is based on comparing the utility of entering retirement at a future optimal age to the utility of retiring immediately, represented by Gt(r*) = EtVt(r*) ~ EtVt(t) (3.3) where r* = arg max EtVt(r). (3.4) r€t+l,t+2,...,T The retirement decision is made such that if Gt(r*) < 0 the individual will enter retire-ment immediately. If Gt(r*) > 0 the individual will delay retirement, hold the option of retiring in the future, and re-evaluate this decision in the next period. Assuming the individual specific preference shifters follow a first order autoregressive process (va = pvs-\ + evs and £s = p£a-i + a n d Ha = pHHs-\ + eHs), equation (3.3) may be written as Gt(r*) = f ] T B'-tirisWviy,,*,,) + ^ /3'-*7r( S | t ) t /r(jfc, Bs(r*))\ \ s=t s=r* / T s=t r * - l r * - l + )s~*[(7t«, - lr)Ht] + E 7r(S|t)(^)s_t[vt - it] (3-5) s=t s=t Gt(r*) = gtiO + jHt + Xm (3.6) Here, 5t(f*) represents the financial incentives that an individual has to delay retirement and hold the option of retiring at a future optimal date, Ht represents the individual's current health status, and rjt represents other current preference shifters. The theory underlying retirement behaviour therefore implies that wealth, the accrual of wealth, and individual characteristics will matter for the retirement decision. Capturing wealth and accrual in practice will require choosing a functional form for the indirect utility function in (3.5). The general forms used for wealth and accrual variables, respectively, are T Wit = £/r-^ S|i)(ys+ £,(£)) (3.7) 57 / r * - l T \ ft* = E P"~t<s\t)Uvl(ya,wa) + £ ^ _ t T ( s | * ) ^ r ( ^ , 5 s(r*)) J \ s=t s = r * / T - Y ^ F ' ^ W W r i y s + Bsit)) (3.8) The wealth variable Wu captures the present value of remaining lifetime income given that the individual retires in the current period. The accrual variables is typically defined in one of three ways. First, a one-year accrual variable is defined by placing a linear utility function over income and setting r* — t +1, capturing the gains in expected lifetime income by delaying retirement for one year. More appropriate given the underlying theory is a peak value accrual variable, which also places a linear utility function over income but chooses r* as the retirement age that maximizes the present value of future income. Stock and Wise (1990) parameterize the utility function as Uw = (ys + wsy (3.9) Ur = (kys + kBs(r)y. (3.10) and then estimate the parameters. While several papers have used their estimates of k and 7 in application of this theory, these parameters may only reflect the preferences of the individuals represented in their sample and are therefore not used in this paper.12 3.5 Estimating the Effects of Health and Financial Incentives The preceding sections have suggested that Canada's retirement income system creates several incentives to enter retirement. The econometric model used to estimate the effect of these incentives as well as the effects of health and other characteristics is stated as R*t = Po + Pi Wit + fait + PzHit + 8AXit + eit (3.11) where • Rn = 1 if i enters retirement {R*t > 0), and • Rit = 0, if i continues to work (R*t < 0). R*t is a latent variable representing the gains in lifetime utility if the person retires in the current period relative to some future period. Although the difference in utility is unobserved, I observe individuals make the decision to either enter retirement or continue 1 2 Stock and Wise (1990) use a sample of 1.500 salesmen, 50 years of age or older in 1980, within a large firm and measure retirement benefits as the benefits these individuals would receive from a defined benefit pension plan. Income from public pensions is not considered in their procedure. 58 to work. As such, estimating equation (3.11) using a probit model is most appropriate. Wu is a measure of wealth, ga is an accrual measure of the financial incentives to enter retirement, Ha represents the individuals current health, and Xu represents other factors we want to control for that would influence retirement decisions. 3.5.1 Data The primary source of data used to estimate equation (3.11) is the Survey of Labour and Income Dynamics (SLID). SLID is a longitudinal survey following individuals over the course of 6 years, with three panels currently available (1993-1998, 1996-2001, and 1999-2002). From each year 1996-2001, I take a sample of individuals who spent at least part of that year in the labour force, are age 50-68, and are flagged as paid workers during the year.13 I further exclude individuals whose labour force status or health information is missing. Since questions regarding health status were not asked until 1996, earlier years of the survey cannot be used here. Given the panel aspect of this survey, I am often able to use individuals' future and past information when defining variables.14 Two other sources of data are used in the construction of variables. First, I use esti-mates from the public use Survey of Consumer Finances Census Family files (1973, 1975, 1977, and 1979) and Individuals files (1981-82, and 1984-1997) to create the wage profiles discussed below. Second, I use the public use Census 1996 to match investment income to individuals in my sample, also discussed in more detail below. Third, survival probabilities are calculated using Statistics Canada's complete life tables (1995-1997). 3.5.2 Measurement of K e y Variables Retirement Ra The definition of retirement used in this study is meant to capture individuals who depart from and remain out of the labour force for an extended period of time. A person is defined as entering retirement during the observation year if they report being in the labour force for at least part of the observation year and not at all in the labour force in the following year. A person is defined as not entering retirement if they continued in the labour force in the following year. 1 3 T o note, a self-employed worker wi l l still be included in the sample if they also held a paid-worker job during the year. Using this flag rather than dropping all workers who were self-employed in the year allows me to keep individuals for whom self-employment is a secondary activity. Robustness checks discussed in the next section demonstrate that dropping all self-employed individuals wi l l not change the final results. 1 4 I do not have access to past (future) information when the individual is observed at the beginning (end) of the panel. As such, individuals observed at the end of each panel have to be dropped from the sample because I rely on future information to define their retirement status (see below). Similarly, when I rely on past health information to define my health variables (see below) 1 have to drop observations at the beginning of the panel. Furthermore, I am not able to observe past self-reported health status for years prior to 1996, so when relying on this information to define the health variable I have to drop all observations in 1996. 59 This definition of retirement results in an expected retirement hazard, presented in Fig-ure 3.4. The retirement hazard represents the probability of entering retirement at each age given that the individual was in the labour force at that age. Small spikes in the probability of entering retirement occur at age 55 (when many employer-provided pension plans allow early retirement) and at age 60 (when individuals are first eligible for CPP/QPP). A large spike occurs at age 65 which is when individuals become eligible for several income security benefits and may be subject to mandatory retirement. Individuals who enter retirement (by this definition) are also observed to have the ex-pected income and labour market outcomes. There are few individuals who return to the labour force after entering retirement by this definition. Less than 10% of individuals aged 60-64 who would have been defined as retired would re-enter the labour force within four years of retiring (see Table 3.2). Many of these individuals may have spent only a short time unemployed within the four years. Individuals who have entered retirement also be-come much more likely to report pensions or government transfers as their major source of income, while those defined as continuing to work continue to report wages as their main source of income (see Table 3.3). Health Hit The health measure is intended to capture how a person's general health affects their de-cision to retire, expecting that poor health increases the disutility of work and therefore increases the likelihood of entering retirement. To measure health, I use individuals' self-reported, current health status which may be rated excellent, very good, good, fair, or poor. Dummy variables are used to indicate reported health. Using current health status, however, has its problems. First, respondents are asked in January following the survey year about their current health. As such, the current health variable is actually describing the individual's health at the end of the year rather than describing their health throughout the year. This could contribute to some identification issues since this health measure may be reported 'post-retirement'. As discussed earlier, the estimated effect of health may be biased upward if justification bias is a problem, or downward if health tends to improve after retirement. To address these concerns, I have included specifications that use the individual's 'pre-vious year' or past reports of health instead of the current health report. Effectively, this will be the individual's reported health at the beginning of the observation year rather than their health at the end of the year, hopefully picking up the health status that affected their decision to retire. Health reported at the beginning of the year, however, will miss events that happen during the year to worsen the person's health and therefore push them into retirement. For example, this could be a general worsening of health or the onset of a disability. To address 60 this I will also provide specifications that use health measures reflecting a change in health status. For example, I can measure whether the person reports not having a disability at the beginning of the year, but reports having a disability at the end of the year (using a disability status variable in SLID, I will later refer to this variable as 'New Disability'). I can also measure whether the person reports a worsening of health by comparing their reported health at the beginning of the year to their reported health at the end of the year (I refer to a 'small shock ' as measuring any worsening of health, and 'large shock' as measuring a worsening of health from excellent, very good or good to fair or poor). The various measures that I use are summarized in Table 3.4. Financial Incentives gu and Wu Financial incentives are measured by including an accrual variable and a wealth variable in the econometric model, as described in section 3.4. This requires estimating expected future incomes from each source of income, contingent on all possible dates of retirement. When constructing the income profiles, I allow an individual to live up to the age of 102 and retire up to the age of 69. With respect to income security program benefits, I determine the amount of benefits that a person will be eligible for upon retirement at each possible age. The calculation of CPP/QPP benefits requires a full wage history. I construct a wage history for each indi-vidual in several steps. First, I use the SCF and SLID to estimate annual wage regressions for men and women separately, with experience, education, province, and marital status as explanatory variables and then use these estimates to impute a wage history.15 For years prior to 1973, I use the 1973 estimates and adjust the earnings for inflation. I have a full marital history for individuals and use this when imputing wages.16 In the second step, I use the individuals' years of full time full year equivalent experience (provided in SLID) to determine how many years of this history should be filled with zeros. I then fill the appropri-ate number of years at the beginning of the wage history with zeros. Finally, to determine benefits when a person does not immediately enter retirement, I assume that wage earnings will grow with expected inflation with no real wage growth. Given this wage history, I use the policy parameters in place in the observation year to calculate the CPP/QPP benefits the individual is eligible for.17 l f l The highest level of education report in SLID is recoded into 6 categories (grade 8 or less, grade 9-10, high school (not making a distinction for graduation), some post-secondary education, post-secondary degree (less than Bachelor's degree), and university degree (Bachelor's degree or higher) to match the coding of education in the SCF. I also add a random component to their annual earnings based on the variance in unexplained earnings in each year. Adding the random component adds variation to the wage distribution but does not affect results. 1 6There are a handful of observations with no marital history. For these individuals I use their current marital status and date of first marriage to create a marital history. 1 7 I am not able to account for the drop-out provisions related to child care and disability. 61 Calculation of OAS, GIS, and SPA is much simpler. OAS is adjusted for the residency requirements when the individual is an immigrant according to the year the person immi-grated. All amounts take into consideration the expected income from other sources when calculating the amount the person is eligible for. For employer-provided pension benefits, I am able to observe whether a person is covered by a pension plan with their employer, whether they contribute to a pension plan, and any income from registered pension plans. Unfortunately, I am not able to observe specific provisions of individuals' pension plans. I assume that if a person reports they are not covered by a pension plan, report zero contributions to an RPP, and report zero pension income in the case that they do retire, that they will not have access to a pension plan when they retire. For these individuals I assume that the entire future stream of income from pension benefits is zero. For individuals who appear to be covered by a pension plan with their employer, I impute a future pension for their stream of future income from pensions contingent on retirement age. To do this I estimate a heckman selection model for pension income (using maximum likelihood), which can be represented by where p* is the pension amount they would receive if they retired in year t and is only observed among individuals who retire (R — l). 1 8 X includes the number of years an individual is with their employer, age indicators, union status, public or private sector, occupation indicators (10 groups), hourly wage, and size of the employer. Z includes X plus indicators for health, marital status, whether a spouse is in the labour force, number of children in the census family, and non-linear combinations of years with the employer and wages. When imputing the pension amount for future retirement ages, I assume that the indi-vidual stays with their employer, thereby increasing the number of years with the employer, and that other job characteristics remain the same. I allow wages to increase with inflation with zero real wage growth. Investment income is an important source of retirement income for some individu-als, however estimating expected investment income is difficult. To impute future invest-ment incomes, I match individuals in my sample to individuals in the Census and as-1 8 I f the potential pension benefits of all individuals covered by a pension plan were observable, OLS would be sufficient for estimating the parameters used to impute potential pension benefits. However, 1 am only able to observe pension benefits if the individual enters retirement and therefore a sample selection problem arises if there are unobservable characteristics determining the pension that, are related to whether a person enters retirement. The Heckman selection model accounts for this. If sample selection is not a problem, the estimates will be equivalent to those obtained using OLS. p* =XB + u R* = Z~f + v (3.12) (3.13) 62 sign investment income as the cell-specific expected median investment income (Prob(I > 0)c*(MedianI\I > 0)c).19 I calculate the spouse's income as I do for the individuals I observe, except that some assumptions are made to simplify calculations. First, I assume that if a spouse is not currently in the labour force, they will remain out of the labour force for the rest of their lives. For spouses who are in the labour force, I assume they will stay in the labour force and then retire at age 60 or immediately if they are already over the age of 60. Survivor benefits under CPP/QPP are also calculated using this assumption Finally, tax payable is calculated given each source of expected income and spouse's income, contingent on the date of retirement considered. The projections of future income that I create approximate incomes actually observed fairly well. In Table 3.5, comparisons of the distributions of imputed and actual incomes are presented. The distribution of imputed wages, CPP/QPP, and OAS match the actual dis-tributions very well. The amount of GIS a person is eligible for is typically underestimated, however, given overestimation of investment income. For pension income, a few individ-uals have been assigned a pension income although I don't see them collecting a pension when they retire.20 This occurs because some individuals may not have been vested in the employer-provided pension plan upon retirement. When constructing the wealth and accrual measures, incomes are discounted at a dis-count rate of 3% and for the probability of survival for the individual and their spouse. Finally, all financial incentives measures are stated in real terms. Other Explanatory Variables Xa In all specifications of equation (3.11), I include a full set of age indicators, province in-dicators, indicators for sex, marital status, whether a spouse enters retirement, whether a spouse has poor health, as well and the number of children under the age of 18 in the census family. In some specifications, controls for lifetime earnings, experience, and current wages are included. Lifetime earnings represents the total imputed earnings of the individual since 1966 in real terms. This set of controls includes a cubic in lifetime earnings of the individual as well as a cubic in lifetime earnings of the spouse. Similarly, wage controls include a cubic "'individuals are place in the following cells: 1. Not in the Labour Force: (a) 3 regions (East, Ontario, West), (b) 5 age groups (50-54, 55-59, 60-64, 65-75, 75+), (c) 2 marital status (married, not), (d) sex 2. In the Labour Force: (a) 3 regions (East, Ontario, West), (b) 4 age groups (50-54, 55-59, 60-64, 65-68), (c) 2 marital status (married, not), (d) sex, (e) 4 occupation groups (i. SOC91 (A-C) managers, professional, clerical, sciences; ii. SOC91 (D-F) health, government, teachers, art; iii. SOC91 (G) services; iv. SOC91 (H-J) trades, transport, manufacturing ) 2 0Imputed pensions are zero below the 40th percentile among those who retired. 63 in the individuals' current wage and a cubic in the spouses' wage. Experience controls include a linear term for the years of full time full year equivalent experience of the individual and their spouse. These variables are included to control for individual heterogeneity that reflects differences in tastes for leisure. For example, we can expect that individuals with a higher preference for work will also have longer work histories and hence potentially higher wealth and accrual measures. If this heterogeneity is not controlled for, the estimated effects of wealth and accruals may be biased downward. 3.5.3 Identification of Wealth and Accrual Effects A crucial issue in the analysis of the financial incentives in income security programs and employer-provided pensions is identification. The main concern here is that several individ-ual attributes that we expect to influence the decision to retire - such as age or preferences for leisure (which may be proxied for with wages) - also determine in part the value of the incentives measures.21 In this section, I exemplify the variation that is used to identify the incentive effects of income security programs and employer-provided pensions. As demon-strated in" the following examples, there is very little exogenous variation in the income security wealth and accrual measures and this limitation of the data should be kept in mind when evaluating the results in subsequent sections. In contrast, measures of wealth and accrual that include employer-provided pension benefits have more exogenous variation available to identify their effects. The distribution of the wealth and accrual measures are summarized in Tables 3.6-3.8. Three variations of the measure are used - the first only includes income from income security benefits, the second adds employer-provided pension income and the third includes all forms of income in the construction of the variable. From these tables, we can see that within each age group there is a considerable amount of variation in the wealth and accrual measures. Furthermore, we can see there is a great deal more variation in the wealth and accrual measures that include employer-provided pensions than those that only include income security benefits. Within each age group at one point in time, the main source of variation in the income security wealth measures reflects differences in individuals' wage histories. The identifying variation for income security wealth and accruals largely relies on changes in wealth and accruals over time - both across individuals and for individuals that are repeated in the sample. For example, between 1997 and 1999, the CPP/QPP benefit formula changed so that the pension adjustment factor (see section 2.2.1) was reduced.22 For an individual who 2 1See Gruber and Wise (2004, p. 12-13) for more discussion of this issue in the context of analyzing the incentives in income security programs. 2 2 Unti l 1997, the pension adjustment factor was based on a 3 year moving average of the year's maximum pensionable earnings. In 1998, this was based on a 4 year moving average and beginning in 1999 this was based on a 5 year moving average. 64 did not retire in 1997, subsequent values of income security wealth would be lower given the new policy and his revised expectations (holding all else constant). A look at the median wealth and accrual by age in each of Tables 3.6-3.8 indicates that the individuals' age-wealth profiles that include employer-provided pensions are also more non-linear than the income security wealth profiles, and may include several local maxima in the profile (unlike the income security wealth profiles that appear likely to have a global maximum that is also the only local maximum in the profile). These non-linearities relate to the early and special retirement provisions of employer-provided pensions whereby employees are commonly offered bonuses for retirement at specific ages.23 Effectively, these provisions will create some variation across age groups that does not merely reflect changes in an individuals' preferences for retirement over work but rather truly reflects differences in financial incentives. Within an age group, variation in the employer-provided pension benefits will reflect differences in union status, public or private sector employment, and job tenure. For an individual that is repeated in the sample, variation in their employer-provided wealth and peak accrual measures may have several sources. Primarily, the profiles will change for each year of delayed retirement given that the employee will observe a new average pension offer and this gets built into the employee's expected future pension benefits.24 The individual will also see an increase in their pension offer relating to increases in their seniority and wages. For a few individuals in the sample, the choice not to retire may involve a change in employment that may entirely change the pension benefits they are eligible for. Finally, suppose we held an individual's age-wealth profile constant over time. In this case, we will still see variation in the individual's peak accrual values if he or she remains in the sample. The nature of that variation depends on the shape of each individual's profile and the age they are first seen in the sample. For example, consider an individual with a fairly simple concave wealth profile with only one maximum. If the individual's peak wealth remains to be seen at a future age, delaying retirement will imply a smaller peak accrual value when the individual reconsiders their retirement decision the next period. As another example, consider an individual with a more complex age-wealth profile that involves several local maxima and the individual's preferences are such that he has already passed the age where the global maximum occurred (without retiring).25 At an age where the individual is at or has just passed the global maximum, it is likely that peak accruals will be negative, reflecting the smallest amount a person has to lose by delaying retirement. 23Pesc.arus and Rivard (2005) indicate that such provisions are most common at ages 55 and 60 in Canadian pension plans, which corresponds to the median peak accrual values in Table 3.7. Such non-linearities are reflected in the coefficients on age when predicting an individual's pension benefits. 2 4 This is accounted for by the inclusion of a set of year indicators when estimating the employee's pension offer. 2 5 T o note, we have to consider that the financial incentives discussed here are not the only factor affecting retirement decisions and might be balanced by the effects of other retirement determinants. 65 We would expect that a larger negative amount would lead to a higher likelihood of entering retirement. At another later age, the individual's peak accrual could turn positive again as they approach a local maximum. To summarize, after age and individuals' wage histories (proxying for individuals' leisure preferences) are controlled for, there is little variation left in the measures of income security wealth and accrual. As such, any estimated incentive effects of income security programs may not be well-identified. In contrast, there is more identifying variation found in the incentive measures for employer provided pension plans. 3.6 Results Results for various specifications of the retirement probit are presented in Tables 3.9 - 3.14. In each table, the wealth and accrual variables used in each specification vary according the income included in the calculation of these measures. In Table 3.9, the first set of spec-ifications include only income from income security programs in the calculation of wealth and peak accrual. The second set adds pension income to income from income security programs and the third set adds investment income, wages, and taxes to the calculation of wealth and peak accrual variables. Within each block of the table, different specifications of the retirement probit are provided. 3.6.1 T h e Effects of Income Security Programs Beginning with the impact that income security programs have on the decision to retire, results presented in the first column of Table 3.9 indicate that wealth from income security does not have a significant effect. As discussed in the previous section, however, the lack of significance here may in part be due to a lack of variation in income security wealth.26 As Baker et al. (2003) also point out, there is a need to control for unobserved heterogeneity in work preferences. Here, individuals with higher incomes and longer work histories are likely to have higher income security wealth as well as higher preferences for work and a lower likelihood of entering retirement, resulting in a downward bias in the estimated effect of income security wealth. To control for this heterogeneity, I add the controls described earlier for lifetime earnings, current wages, and years of full time full year experience. As a result, the coefficient on income security wealth does increase substantially but is still not significant. To check whether these extra controls are sufficient, I use a fixed effects probit estimator, the results of which are presented in Table 3.10.27 The fixed effects estimates 2 ( >The size of the effects are just as large as the effects of employer-provided pensions. 2 7 The fixed effects probit involves an incidental parameters problem due to the connection between sample size and parameter space. As such, the fixed effects probit estimates are not consistent. Wooldridge (2002) conjectures that the estimators of the marginal effects have reasonable properties. Fernandez-Val (2005) finds the bias in estimators of marginal effects to be negligible. Note that the marginal effects presented in 66 are not substantially different from the estimates obtained with the retirement probits. The accrual of income security wealth, measured using the peak accrual, is significant across specifications, even after we control for individual specific fixed effects. This would indicate that individuals pay attention to the amounts they can gain from income security programs by delaying retirement. Considering the results presented in Table 3.11, where past health reports are used to measure poor health rather than current reported health, it appears that the endogeneity of health may be an issue when identifying the effects of income security programs. When using past health reports, the magnitude of the coefficients on income security wealth and accrual increase substantially, and the effect of income security wealth becomes marginally significant. However, when using other measures of health, as presented in Table 3.12, the income security wealth measure is not significant. The endogeneity and identification of the effects of health are discussed in more detail below. Similar specifications using the one year accrual of income security wealth are presented in Table 3.13. Once the appropriate controls are included in the specifications, the effects of income security wealth are insignificant and the estimated effect of the one year accrual is insignificant and of the wrong sign. This is not surprising given the expectation that individuals are forward looking when making retirement decisions and wealth profiles may be non-linear, making the peak accrual measure more appropriate. Finally, in Table 3.14 I check the robustness of these estimates to different sampling choices. In the first block of estimates in the table, I have additionally excluded any in-dividuals who had no wage income in the observation year, despite reporting having been in the labour force. Effectively this excludes individuals who were unemployed during the observation year, but not employed, implying a slightly different measure of retirement. There are no substantial changes to the estimated effect of income security programs when compared to the estimates previously discussed. In the third block of the table, I further address the concern that the self-employed act very differently than paid workers by ex-cluding individuals who spent any time self-employed in the observation year.28 Again, the results are nearly identical to those presented in Table 3.9. The collection of estimates imply that income security programs have important accrual effects, but may not have the significant wealth effects suggested by earlier studies. There are several potential explanations for the discrepancy in results. On one hand, early studies used data covering a time period in which benefit generosity was generally increasing while retirement age and labour force participation of older workers were falling and did not Table 3.10, which are evaluated at the mean, are not directly comparable to the marginal effects presented in Table 3.9 as those are evaluated for a particular type of person. 2 8 That is, the original sample used for the retirement probits kept any individual who was a paid worker at some point during the observation year. They may have also been self-employed at some point during the observation year. 67 always control for time effects. As such, these studies may have been picking up a spurious correlation between wealth from income security and the likelihood of entering retirement.29 The time period used in this study, however, captures a period in which the labour force participation rates of older workers had levelled off and actually increased. Furthermore, in comparison to Baker et al. (2003) (who control for time effects), when constructing income and benefit measures I have more complete information regarding the length of time individuals have spent employed in their lifetime, improving my estimates of income security wealth and capturing the non-linear nature of wealth profiles. On the other hand, Baker et al. (2003) have access to more wage history information for the period they can observe and do not have to rely entirely on imputed wage histories as I do to create their measures of income security benefits. More importantly, as discussed in section 3.5.3 it may be the case that there is simply not enough exogenous variation in income security benefits formulas over the time period I study to identify any wealth effects.30 3.6.2 T h e Effects of Employer-Provided Pensions and Other Income To see how employer provided pensions may affect retirement decisions differently than income security programs, I repeat the specifications used to test for the effects of income security program, but add pension income to the measures of wealth and accrual. In Table 3.9 these results are presented in the second block of the table. In each specification the wealth effects are positive and significant and accrual effects are negative and significant as expected. A large portion of the variation identifying these effects comes from differences across individuals' access to pension plans.31 Concerns that there may be something spe-cific about individuals that leads them into jobs with or without pensions and that this characteristics is also related to their preferences for work and therefore retirement would be addressed with the use of a fixed effects estimator. When the fixed effects estimator is used, results do not change substantially, as presented in Table 3.10. The last set of retirement probits in Table 3.9 use wealth and accrual measures that include not only income security benefits and pension income, but wage income, investment income, and taxes as well. The estimated wealth and accrual effects differ only slightly from the wealth effects of employer-provided pensions. 2 9 I have estimated models that include year effects. The size and significance of the coefficient on income security wealth diminishes when year indicators are included, but other results do not change significantly. 3 0Baker et al. (2003) were able to use more substantial policy changes, specifically the introduction of early retirement provisions in the mid-1980s, as a source of exogenous variation in their study. , n T o note, one possible reason for the insignificance of income security wealth effects relative to the significance of wealth effects associated with employer-provided pensions is that there exists less variation in income security wealth (which can be accessed by virtually all retirees). 68 3.6.3 T h e Effects of Health The estimates provided in Tables 3.9-3.14 consistently demonstrate that health status has a significant effect on the likelihood of entering retirement. The effect is substantial, as demonstrated in Table 3.9 where estimates imply that having poor health raises the likeli-hood of entering retirement by more than twenty percentage points.32 Given the subjective nature of this health measure and the potential identification issues discussed earlier in this paper, it is useful to consider variety of measures to check the robustness of this result. One concern was that individuals will mis-report their health in order to justify their retirement, creating an upward bias in the estimated effect of poor health. Using the past reports of poor health to create the health measure indicates this may be a concern. Comparing results in Tables 3.9 and 3.11, the magnitudes of the probit coefficients on health drop slightly when past health reports are used.33 Similarly, estimates in the first block of Table 3.12 in which a complete set of current health indicators are used suggest poor health has a higher impact on retirement than when the past health reports are used as in the second block of Table 3.12. These estimates also suggests that any justification bias is greater than any bias associated with improved health with retirement. While use of past health reports may correct for justification bias, estimates of the ef-fect of health using past health reports may be biased downward if health shocks contribute significantly to retirement rates. As was demonstrated in Table 3.4, several individuals ex-perience a worsening of their health over the year and this could push them into retirement. In the third, fourth and fifth blocks of Table 3.12 I have used indicators for the worsening of health and demonstrate that any worsening of health will raise the probability of entering retirement. The onset of a new disability will raise the likelihood of entering retirement by nine percentage points while experiencing a relatively large shock to health (such that you change your reported health from excellent, very good or good to fair or poor) raises the likelihood of entering retirement by roughly eight percentage points.34 To note, specifications of the probit that included indicators for access to health insur-ance and access to life and disability insurance through an employer, as well as interaction terms for poor health and access to insurance, were also estimated to check whether access to health insurance might act as a constraint as it appears to in the United States. Although the results are not presented here, the effect of insurance on the likelihood of retirement was completely insignificant. Furthermore, the effect of poor health did not differ between individuals with and without health or disability insurance. 3 2 The marginal effect is slightly lower for individuals under the age of 60 and slightly higher for individuals over age 60. 3 3Note that the coefficients on wealth and accrual also increase. 34Bffects of health shocks are slightly higher (lower) for older (younger) individuals. 69 3.7 Conclusions This study provides estimates of the effects of health, Canada's income security programs, and employer-provided pensions on individuals' decisions to enter retirement. The results indicate that individual health has a significant and strong influence on retirement be-haviour. Identification issues associated with using self-assessed health measures have been addressed in this study, suggesting that justification bias may be an important problem that should be taken into consideration. Examining the effects of income security programs on retirement behaviour, while the results suggest that income security programs do not have significant wealth effects, the accrual of income security wealth that can be achieved with delayed retirement appears significant. The estimated effects of income security wealth do not correspond to those found in several earlier studies that also estimate reduced form models of the decision to enter retirement. The insignificance of wealth effects, however, correspond to the results found in several studies using natural experiments to identify the effects of income security programs. It is likely the discrepancy between earlier studies and this study is due to a relative lack of exogenous variation available to identify the effect of income security wealth over the time period studied here. The estimates provided in this paper suggest that employer-provided pensions have significant wealth and accrual effects on the decision to enter retirement. Wealth effects appear to be stronger than accrual effects, which in part captures the differences in wealth between those with and without pension plans. The addition of other forms of income to the measures of wealth and accrual do not significantly change the results, except that the accrual of wealth becomes more significant suggesting that individuals consider their entire financial picture when making the decision to retire. These results potentially have important implications for policies affecting retirement income. Primarily, changes to the structure of benefit formulas that affect accruals in wealth may influence the timing of retirement. Furthermore, the interaction of different policies (such as the interaction between CPP/QPP and GIS) should be an important consideration. Finally, it is important to consider that the importance of health and other forms of income may trump the effects of financial incentives found in any income security program. 70 Table 3.1. Importance of Various Income Sources In the Labour Force Not in Labour Force 50-59 60-68 50-59 60-68 Proportion Reporting Positive Income Wage .93 .86 .13 .09 CPP/QPP .03 .46 .2 . .76 GIS .07 .27 OAS .2 .49 Investment .4 .45 .32 .46 Pension .06 .26 .21 .44 Proportion Reporting Majority of Income From Each Source Wage .91 .71 .08 .02 . CPP/QPP .01 .09 .15 .26 OAS+GIS .04 .25 Investment .02 .04 .15 .09 Pension .02 .1 .18 .29 No Income .05 .02 .45 .09 Note. — SLID sample of individuals 1996-2001, age 50-68, who are not self-employed. The top panel describes the propor-tion of individuals in this sample reporting positive income from each source. The bottom panel describes the proportion of in-dividuals for whom the given income source is the major source of income out of the sources listed here. The proportion with 'No Income' are those not reporting any income from wages, CPP/QPP, OAS, GIS, investments or pensions. 71 Table 3.2. Rate of Exit From Retirement Within 1 Year 2 Years 3 Years 4 Years Age 50-64 0 0.15 0.23 0.26 Age 50-54 0 0.21 0.40 0.45 Age 55-59 0 0.22 0.29 0.31 Age 60-64 0 0.05 0.07 0.09 Note. — Using annual labour force status to define entry to retirement. Exit from retirement refers to re-entry to the labour force in the first, second, third, or fourth year after entry to retirement. This sam-ple uses individuals defined as retiring in 1994 (first panel of SLID) or in 1997 (second panel of SLID), age 50-64 in 1994 or 1997 72 Table 3.3. Characteristics of Retirees and Non-Retirees t- 1 t t+1 Annual Labour Force Participation Rate R = l .91 1 0 R = 0 .99 1 1 vlajor Source of Income Wages, R = 1 .76 .60 .09 Wages, R = 0 .89 .90 .85 Pension, R = 1 .06 .11 .35 Pension, R = 0 .02 .02 .03 Transfers, R. = 1 .13 .19 .41 Transfers, R = 0 .05 .04 .06 Average Wages and Salaries R=l 27780 19219 3320 R = 0 33631 34729 33577 Note. — R — 1 when the individual en-ters retirement in year t, R, = 0 otherwise. Annual labour force status is used to define entry to retirement, as described in the text. SLID sample of panel 1 individuals age 50-68 in 1994 and panel 2 individuals age 50-68 in 1997. 73 Table 3.4. Health Measures by Age Age 50-54 55-59 60-64 65-68 Current Health Poor 0 01 0 02 0 03 0 02 Fair 0 07 0 10 0 10 0 11 Good 0 26 0 29 0 26 0 27 Very Good 0 40 0 35 0 39 0 38 Excellent •o 25 0 24 0 23 0 22 Past Health Poor 0 01 0 02 0 02 0 02 Fair 0 07 0 10 0 09 0 10 Good 0 24 0 28 0 24 0 25 Very Good 0 41 0 37 0 40 0 37 Excellent 0 28 0 24 0 25 0 26 New Disability 0 09 0 10 0 12 0 14 Small Shock 0 28 0 28 0 29 0 29 Large Shock 0 05 0 07 0 07 0 09 Note. — Using full sample of 25810 observations described in section 3.5. For past health informa-tion, only 17618 observations available. See text for definitions. 74 Table 3.5. The Distribution of Imputed and Actual Incomes Mean Median 1st Dec. 9th Dec. Std. Dev Waget, (Rt = 0) 32284 25313 8890 65431 24292 Waget, (Rt = 0) 37331 32215 7000 68350 39121 Waget_1 33509 25552 8452 69988 26538 Waget-i 35778 31453 4834 66781 30860 CPPt+i (Rt = l) 2935 199 0 7963 3399 CPPt+i, (Rt = 1) 2944 0 0 8059 3636 CPPt+i (Rt = l,Aget+l > 60) 5568 6060 1443 8876 2693 CPPt+u (Rt = l,Aget+i > 60) 4809 5238 0 8637 3234 OASt+l, (Rt = i) 1023 0 0 5049 1998 OASt+u (Rt = i) 901 0 0 4901 1877 OASt+1, (Rt = l,Aget+1 > 65) 4776 5049 3660 5232 839 OASt+i, (Rt = 1, i4ge t + i > 65) 4209 4901 1286 5232 1588 GISt+i, (Rt = i) 97 0 0 0 558 GISt+i, = 1) 240 0 0 15 940 GISt+i, = l , ^ e t + i > 65) 451 0 0 0 1138 GISt+i, (Rt = 1, ^ e ( + 1 > 65) 878 0 0 3504 1618 Pensiorit+i, (Rt = 1) 9533 6870 0 24898 10189 Pensiont+i, (Rt = 1) 9982 0 0 33956 14551 Pension (i?t = l,Aget+i >60) 8880 6870 0 22501 9318 Pensiont+i, (Rt — l,Aget i > 60) 9266 2374 0 31212 12956 Investment^, (Rt = 1) 909 920 467 1370 338 Investrnentt, (Rt = 1) 1494 8 0 3848 7064 Investmentt, (Rt = 0) 523 437 229 886 322 Investmentt, (Rt — 0) 802 0 0 2041 3499 Note. — The sample used is the same as that described in section 3.5. Imputed incomes are denoted with^. Rt ~ 1 indicates the individual entered retirement during the observation year t. 75 Table 3.6. The Distribution of Income Security Measures Age N Wealth One Year Accrual Peak Accrual Median 1st Dec. 9th Dec. Std.Dev. Median 1st Dec. 9th Dec. Std.Dev. Median 1st Dec. 9th Dec. Std.Dev. 50 3,131 110147 80330 154600 29857 1775 588 6511 2665 12225 2961 23646 8121 51 2,870 114943 85363 160127 30417 897 -9625 1931 5934 9948 0 21219 9349 52 2,667 118252 85022 163287 32366 1865 599 4542 2152 10179 2542 22195 7847 53 2,483 123718 92480 172575 32565 1367 -3963 2235 3515 8608 493 20141 8474 54 2,288 128154 94965 177454 33533 1126 -1237 2219 3758 6956 828 19960 7651 55 1,986 134940 96968 183699 35556 1140 92 2737 1017 6794 852 20237 7812 56 1,726 136859 101748 189949 36073 1091 87 2527 914 5738 757 17586 7320 57 1,525 140724 105190 199882 37578 856 0 2330 1111 4200 346 14920 6371 58 1,333 147949 112786 209772 38981 976 12 2996 1779 2629 51 12340 6184 59 1,184 153417 119760 215285 39274 662 0 2224 1589 1066 0 8913 4442 60 1,045 159617 122410 225085 40926 -184 -2039 1446 1420 -88 -2036 7066 4689 61 866 165749 129239 234392 41521 -1039 -2739 1500 2154 -1039 -2739 5972 4344 62 726 171261 133587 245980 42464 -1876 -3938 117 2880 -1759 -3644 2608 3595 63 597 174123 137878 240763 40287 -2504 -4101 491 3232 -2504 -4101 3557 5335 64 461 178320 136963 250170 42932 -3177 -4625 803 2816 -3177 -4625 2988 3750 65 394 185119 135861 254223 43499 -3061 -5376 275 2352 -3061 -5376 1095 3183 66 231 186514 148680 249314 40753 -4415 -8158 -734 6546 ' -4165 -6238 177 3254 67 167 183368 129779 244861 43328 -4830 -7102 401 9558 -4830 -7102 401 9681 68 130 191543 146445 255467 45092 -5014 -7381 -934 2747 -5014 -7381 -934 2747 Note. — Description of sample and definitions of variables in section 3.5. Table 3.7. The Distribution of Income Security + Pension Measures Age N Wealth One Year Accrual Peak Accrual Median 1st Dec. 9th Dec. Std.Dev. Median 1st Dec. 9th Dec. Std.Dev. Median 1st Dec. 9th Dec. Std.Dev. 50 3,131 204072 96346 448996 147318 -9187 -65267 1732 29955 107896 5506 149898 59563 51 2,870 174181 100609 405799 129435 79562 715 132084 59179 155502 5075 205586 86032 52 2,667 266419 108314 507459 166529 -41754 -54436 1747 24828 59395 3890 89134 34098 53 2,483 225413 111290 472029 154192 46883 733 58369 25608 112735 3118 136583 57288 54 2,288 273574 118562 518201 164770 67073 631 81835 35733 68328 2603 83652 34314 55 1,986 344029 124555 590725 184606 -23148 -32707 1499 .14378 -18907 -32707 12558 19159 56 1,726 305070 126361 546680 169383 -12853 -23696 1632 10440 439 -23696 17257 17615 57 1,525 284707 131189 536750 163862 1452 -2011 8625 4277 3114 -2011 25486 14816 58 1,333 274819 137719 543600 163252 -30311 -43267 1527 19815 137 -37957 15834 22819 59 1,184 254931 144166 506597 151043 2821 25 13954 5596 • 4708 79 48252 20946 60 1,045 257537 147919 504418 149748 -8486 -20787 334 8668 -857 -20787 36154 21088 61 866 255119 153223 479915 139655 -21232 -39290 -252 17017 -551 -13865 50751 24665 62 726 255196 155808 439787 120979 16828 -3325 30847 14507 17351 -3325 77402 33570 63 597 261815 156082 474867 137022 -34308 -50525 -1223 22059 1314 -4511 51211 25007 64 461 244065 154429 404648 100542 -21150 -35532 -528 14742 57558 -4198 99085 42637 65 394 233722 166686 402048 104203 -4221 -12131 101 5078 91029 -4692 128191 56882 66 231 233819 166382 413200 101780 110729 -5073 141875 61747 110729 -5073 141875 61721 67 167 279393 162572 547821 147559 -111457 -164307 -3297 74175 -11.1457 -164307 -3297 74199 68 130 242735 163157 387834 85649 -7041 -20889 -2354 6963 -7041 -20889 -2354 6963 Note. — Description of sample and definitions of variables in section 3.5. Table 3.8. The Distribution of Total Income Measures Age Wealth One Year Accrual Peak Accrual Median 1st Dec. 9th Dec. Std.Dev. Median 1st Dec. 9th Dec. Std.Dev. Median 1st Dec. 9th Dec. Std.Dev. 50 345756 117000 648103 213159 6463 -31334 30802 23824 246840 106091 530620 172074 51 310689 116820 590137 190069 79578 10406 124609 46028 268327 122890 546207 175369 52 348481 132017 660388 204002 2891 -24252 27242 20207 200291 64552 452717 153966 53 315852 130935 601562 190263 45136 10245 74013 24009 213006 87386 452537 147400 54 332711 162186 605065 178136 55489 10382 82301 26909 167131 55744 399394 142256 55 357393 174242 619612 189181 7195 -11606 28875 16158 121748 3929 329215 133265 56 332781 151671 569456 162844 10081 -4967 33097 14973 122161 21955 335245 125708 57 305566 157194 540584 156123 17667 6397 38940 13428 119338 27653 291594 113036 58 294550 157220 542708 153544 6112 -17325 27218 16780 105526 20646 262407 97173 59 273454 158444 496643 138511 17593 7988 39783 12976 95066 30554 239281 86375 60 269954 162061 462364 124681 7879 -4025 28477 13100 77591 19892 210320 80823 61 262551 164402 453921 115145 3559 -16846 23987 15727 66291 17007 185457 69782 62 257069 171184 406569 98030 23674 5029 41377 14498 78254 26880 158540 56428 63 263139 172500 410151 107960 • 1057 -25248 21127 17768 53477 15068 132156 48625 64 246642 173820 373170 77955 1462 -15708 19693 14064 66869 14221 111811 41281 65 238145 178931 359654 78668 6792 278 22735 9246 74276 17054 111485 35382 66 238218 171175 353347 73025 74424 3501 100336 38583 78904 9386 100367 34423 67 277250 162724 419173 99197 -59606 -95549 19483 47640 -36549 -89995 37467 52174 68 241141 172003 344467 64019 5395 -2626 22676 9613 5395 -2626 22676 9613 Note. — Description of sample and definitions of variables in section 3.5. Table 3.9. Retirement Probit Results I Income Security IS + Pensions All Income Poor Health 0.216 0.238 0.239 0.230 0.239 0.248 0.249 0.245 0.252 (.036) (.039) (.039) (.037) (.04) (.039) (.038) (.04) (.04) [.864] [.865] [.861] [.901] [.896] [.897] [.900] [.895] [.895] Wealth 0.003 0.012 0.014 0.014 0.018 0.017 0.014 0.016 0.015 (.010) (.018) (.018) (.002) (.003) (.003) (.003) (.004) (.004) [.018] [.070] [.077] [.090] [.109] [.097] [.082] [.094] [.085] Peak Accrual -0.100 -0.140 -0.155 -0.017 -0.017 -0.018 -0.011 -0.016 -0.017 (.040) (.064) (.062) (.007) (.007) (.008) (.003) (.006) (.006) [-.666] [-.798] [-.872] [-.111] [-.107] [-.107] [-.063] [-.091] [-.094] Lifetime earnings no yes no no yes no no yes no Controls Experience no no yes no no yes no no yes Controls Wage no yes yes no yes yes no yes yes Controls Note. — 25810 observations. Marginal effects are presented representing a 60 year old single male in Ontario, standard errors of the marginal effect in parentheses, probit coefficients are in square brackets. All specifications include the basic set of controls described in the text. Table 3.10. Retirement (Fixed Effects) Probit Results II Income Security IS + Pensions All Income Poor Health 0.165 0.148 0.166 (.097) (.100) (.098) [.565] [.523] [.571] Wealth 0.049 0.081 0.056 (.110) (.020) (.015) [.213] [.360] [.245] Peak Accrual -0.489 -0.101 -0.021 (.244) (.056) (.012) [-2.109] [-.449] [-.092] Note. — 3195 observations, 1131 individuals. Marginal ef-fects evaluated at the mean are presented, standard errors of the marginal effect are in parentheses, probit coefficients are in square brackets. All specifications include controls for age, whether the spouse works, spouse's health, and number of children under 18. 80 Table 3.11. Retirement Probit Results III Income Security IS + Pensions A l l Income Past Poor Health 0.201 0.226 0.231 0.213 0.233 0.241 0.232 0.240 0.245 (.050) (.055) (.055) (.051) (.055) (.055) (.053) (.056) (.056) [.794] [.804] [.799] [.830] [.824] [.820] [.834] [.833] [.824] Wealth 0.023 0.041 0.037 0.015 0.020 0.019 0.015 0.017 0.016 (.013) (.022) (.024) (.003) (.004) (.004) (.003) (.005) (.005) [.143] [.220] [.190] [.093] [.108] [.097] [.081] [.087] [.080] Peak Accrual -0.123 -0.179 -0.223 -0.015 -0.017 -0.018 -0.010 -0.018 -0.020 (.045) (.073) (.076) (.008) (.010) (.010) (.004) (.008) (.008) [-.761] [-.956] [-1.135] [-.096] [-.091] [-.091] [-.056] [-.092] [-.096] Lifetime earnings no yes no no yes no no yes no Controls Experience no no yes no no yes no no yes Controls Wage no yes yes no yes yes no yes yes Controls Note. — 17618 observations. Marginal effects are presented representing a 60 year old single male in Ontario, standard errors of the marginal effect are in parentheses, probit coefficients are in square brackets. A l l specifications include the basic set of controls. Table 3.12. Retirement Probit Results IV (1) (2) (3) (4) (5) IS IS+Pension IS IS+Pension IS IS+Pension IS IS+Pension IS IS+Pension Poor 0.258 0.271 0.248 0.262 - - - - -(.04) (.041) (.056) (.056) Faii- 0.082 0.091 0.050 0.058 - - - - - -(.017) (.017) (.017) (.018) Good 0.019 0.023 0.039 0.045 - - - - - -(.009) (.009) (.013) (.013) Very Good 0.010 0.012 0.007 0.008 - - - - - -(.008) (.007) (.009) (.01) New Disability - - - - 0.091 0.093 - - - -(.016) (.016) Small Shock - - - - - - 0.017 0.020 - -(.009) (.01) Large Shock - - - - - - - - 0.071 0.080 (.021) (.022) Wealth 0.014 0.016 0.035 0.018 0.015 0.016 0.035 0.019 0.035 0.020 (.016) (.003) (.021) (.004) (.017) (.003) (.024) (.004) (.024) (.004) Peak Accrual -0.135 -0.016 -0.195 -0.016 -0.150 -0.021 -0.216 -0.019 -0.216 -0.019 (.055) (.007) (.069) (.009) (.06) (.008) (.076) (.011) (.076) (.011) N 25810 25810 17618 17618 25810 25810 17618 17618 17618 17618 Note. — (2) uses past health variables. Marginal effects are presented, standard errors in parentheses. A l l specifications include basic set of controls plus controls for experience and current wages. Table 3 .13. Retirement P rob i t Results V Income Security IS + Pension Wealth 0.098 0.010 0.032 0.034 0.010 0.014 0.019 0.018 (.012) (•01) (.016) (.018) (.002) (.003) (.003) (.003) One Year -0.204 -0.005 0.005 0.004 -0.011 0.004 0.005 0.005 Accrual (.069) (.09) (.104) (.106) (.004) (.012) (.013) (.013) Age Controls no yes yes yes no yes yes yes Lifetime Earnings no no yes no no no yes no Controls Experience no no no yes no no no yes Controls Wage no no yes yes no no yes yes Controls Note. — Marginal effects are presented, standard errors in parentheses. A l l specifications include b&sicset of controls and an indicator for poor health unless otherwise indicated. Table 3.14. Retirement P rob i t Results V I IS IS + Pension IS IS + Pension Poor Health 0.204 0.212 0.206 0.220 0.243 0.244 0.245 0.253 (.04) (.04) (.04) (.041) (.04) (.039) (.041) (.04) Wealth 0.023 0.017 0.018 0.017 0.014 0.018 0.018 0.017 (.017) (.019) (.003) (.003) (.018) (.018) (.003) (.003) Peak Accrual -0.093 -0.126 -0.013 -0.014 -0.148 -0.163 -0.018 -0.019 (.061) (.063) (.007) (.008) (.065) (.064) (.007) (.008) Lifetime Earnings Controls yes no yes no yes no yes no Experience Controls no yes ' no yes no yes no yes Wage Controls yes yes yes yes yes yes yes yes N 25265 25265 25265 25265 24601 24601 24601 24601 Note. — Marginal effects are presented, standard errors in parentheses. A l l specifications include basic set of controls. 83 18 000 1-4 5-14 ; 15-44 45-64 65-74'. 75-84 ,' 85+ A g e G r o u p Figure 3.1 Provincial Health Care Expenditures per Person by Age and Sex, 2000 Source: The Canadian Institute for Health Information (2002). (Table E.1.3) 84 40 - I — i — i — i — , — , — , — , — , — , — , — , — , — , — , — i — , — i — i — , — , — , — , — , — , — , — , , , , 1976 1980 1984 1988 1992 1996 2000 2004 Figure 3.2 M a l e Par t i c ipa t ion Rates by Age Group, 1976-2004 Source: Statistics Canada, CANSIM II Series V2461460, V2461470, V2461471, V2461472 1976 1980 1984 1988 1992 1996 2000 2004 Figure 3.3 Female Par t i c ipa t ion Rates by Age Group , 1976-2004 Source: Statistics Canada, CANSIM II Series V2461670, V2461.680, V2461681, V2461682 85 CD CO Conditional Probability of Retirement C O C h a p t e r 4 Why Have the Labour Force Participation Rates of Older Men Increased Since the Mid-1990s? 4.1 Introduction Since the mid-1990s there has been a clear reversal in the decline in labour force participation rates of older men in Canada, the United States, the United Kingdom, and several other European countries. Through most of the 20th century the participation rates of men age 55-64 had fallen steadily, elevating concerns that population aging will place pressure on public pensions and many other publicly funded programs. Over the past decade, however, there has been a remarkable increase in the participation rates of older men, potentially alleviating many of these concerns. What might explain this striking reversal in participation rates? Despite having very different policy environments and macroeconomic performance in recent years, increases in the participation rates of older married men in the U.S., Canada, and the U.K., have coincided with recent increases in older women's participation rates, while single men appear to follow a different trend. This paper exploits the cohort effects driving recent increases in older women's participation rates to identify the extent to which older married men's participation decisions are influenced by their wives' participation decisions. The response of husbands to their wives' participation decisions then has the potential to explain some of the recent increases in older men's participation rates. There are two main routes through which we can expect wives' participation decisions to affect the participation decisions of husbands. First, we can expect an income effect °Portions of the statistical analysis in this chapter rely on Statistics Canada microdata, made available through the British Columbia Inter-University Research Data Centre. This study reflects the views of the author and does not reflect the opinions of Statistics Canada. 87 whereby the husband is able to enjoy more leisure given the extra family income earned by a participating wife. This income effect will reduce a husbands' likelihood of participating in the labour force. Second, we can expect that couples have some preferences for shared leisure time, especially at older ages. Husbands may not enjoy their leisure time as much when their spouse is participating in the labour force. This shared leisure effect will result in a higher likelihood of participation among husbands whose wives are participating in the labour force.1 If shared leisure effects dominate any income effects associated with wives' participation, we would expect to see increases in the participation rates of older men in response to increases in the participation rates of older women. The existing literature provides considerable evidence that shared leisure effects play an important role in older men's retirement and participation decisions. The notion that a wife's leisure is complementary to her husband's is evidenced by the fact that in the United States 62% of older men report that they look forward to retiring only if their spouse can retire as well (Coile (2004b), Maestas (2001)).2 In Canada, older couples tend to rate the quality of their relationship higher when both spouses are not in the labour force than when they are both in the labour force or have retired alone (Chalmers and Milan (2005)). The prevalence of joint retirement suggests that this preference for shared leisure is an important determinant of participation and retirement decisions. Gustman and Steinmeier (2000) find that a wife's retirement status significantly affects the retirement preferences of a husband, although the husband's retirement does not appear to affect the retirement preferences of the wife. Coile (2004b) examined how the retirement incentives of husbands and wives are influenced by their own financial incentives to retire and the incentives of their spouse, and finds that husbands are very responsive to their wives' financial incentives to enter retirement. Coile also provides evidence that husbands are even more responsive to wives' incentives when they indicate that they enjoy sharing leisure time with their spouse. Blau (1998) provides further evidence that preferences for shared leisure affect retirement decisions, showing that the high incidence of joint retirements cannot be explained by financial incentives and that the non-employment of one spouse has a positive effect on the labour force exit rates of the other spouse. In the current analysis I am primarily concerned with the extent to which this preference for shared leisure, in conjunction with increased participation of wives, may help explain recent increases in the participation rates of men age 55-64. I begin by estimating the effect of a wife's participation decision on the participation decision of a husband. Here, I model 1A positive response to wives' participation may also reflect husbands' reluctance to abandon their role as the primary 'bread-winner'. 2This is based on responses to the question in the Health and Retirement Study: How much do you agree or disagree with the statement: "I look forward to retiring only if my wife can retire at about the same time." Responses include Strongly agree (14%), agree (48%), disagree (35%) and Strongly disagree (3%). Both authors are using samples of married men age 5.1-61 in 1992, restricted to couples in which neither spouse had retired by 1992. 88 the participation decisions of husbands and wives as a system of simultaneous equations and rely on a measure of cohort effects as an instrument for the wife's participation decision. To examine the explanatory power of wives' participation, a decomposition of older married men's participation rates is undertaken using the methodology of DiNardo, Fortin and Lemieux (1996) and data from the March Current Population Survey (United States), the master files of the Labour Force Survey (Canada), and the Quarterly Labour Force Survey (United Kingdom). The results demonstrate that a substantial portion of the recent increase in married men's participation can be attributed to shared leisure effects. I also investigate the role played by changes in the age structure and educational attainment of this group of men and show that these factors can explain a substantial portion of the recent increase in participation, especially in the United States. Looking forward to how these factors will affect future trends in older men's participation rates, changes in older wives' participation will continue to place upward pressure on older married men's participation rates. However, the education levels of the upcoming cohort of men do not differ substantially from the current cohort. As such, it appears that the effects of education have been exhausted and cannot be relied upon to further increase older men's participation rates. The remainder of this paper is organized in the following manner. I begin with a description of recent trends in the participation rates of older individuals, demonstrating the potential explanatory power of changes in wives' participation for recent increases in men's participation rates. In section 4.3,1 provide a simple framework to analyze the shared leisure and income effects associated with wives' participation and provide empirical evidence that shared leisure effects dominate income effects in the participation decisions of older men. In this section I also address other factors that may be important for retirement or participation decisions, such as health, macroeconomic effects, and stock market fluctuations. In section 4.4, I describe the methodology used to decompose the changes in participation rates, discuss the results of the-decomposition, and provide additional evidence of the importance of wives' participation in the participation decisions of husbands. In section 4.5 I present some evidence suggesting we may expect further increases in participation over the next decade. Finally, I offer some conclusions in section 4.6. 4.2 Data and Recent Trends in Participation In this section, I begin by introducing the data used in this study. I then document trends in older men's and women's participation rates and demonstrate the potential explanatory power of wives' participation for the recent increases in older men's participation rates. 89 4.2.1 Data Several micro data files are used to analyze older men's labour force participation in the U.S., Canada, and the U.K.. U.S. samples are taken from the Current Population Survey Annual Demographic Files (March CPS) while Canadian samples are taken from the master files of the Canadian Labour Force Survey (CLFS) and samples for the U.K. are taken from the U.K.'s Labour Force Survey (UKLFS). Each of these data sources provide information regarding the labour force status, education, and demographic characteristics of individuals and their spouses. The details of sample selection, the construction of key variables, and other data sources used in this paper are described in Appendix B. Generally speaking, the samples exclude individuals whose information is missing, and for the purposes of the decomposition exercise that follows I further exclude from the samples any husband who is more than 15 years younger or 15 years older than his wife. The analysis in this paper focuses on married men age 55-64 in each country. Descrip-tions of these samples are provided for years representing low and high points in older men's participation rates in Tables 4.1, 4.2, and 4.3 for the U.S., Canada, and the U.K. respectively. In each of these countries, the observed increase in labour force participation coincides with substantial increases in educational attainment for this group of men.3 Inter-estingly, the average age of these men fell slightly over the decade, reflecting a slight change in the age structure of this group.4 Perhaps surprisingly, the likelihood of having children at home has not increased in recent years. Notably, the likelihood of wives' participation in these samples increased substantially over the past decade. The increase is most striking for the Canadian wives of men age 55-64, whose participation rates increased by twelve percentage points between 1995 and 2005. Men whose wives are in the labour force are more likely to participate in the labour force. However, these men also tend to be slightly younger and more educated than men whose wives are not in the labour force. 4.2.2 Trends in Participation The participation rates of older men in Canada, the U.S., and the U.K. had fallen steadily through most of the 20th century.5 Figure 4.1 shows that the participation rate of men 3Note that educational attainment is not directly comparable across the different countries. For example, Canadian men appear to have much higher attainment of post-secondary education than American men in part due to the fact that Canadian surveys include trade certificates as post-secondary education while U.S. surveys do not. Also, U.K. categories of education are defined by an individuals' qualifications, which may differ slightly from the concept of education in the Canadian and U.S. surveys. 4There are clear changes across the distribution of age indicating the group became 'younger' over time in each country as the baby boom cohort begins to enter this group. 5Ransom and Sutch (1986) find that the participation rates of men age 60 and over in the United States decline after 1930. Johnson (1994) finds there was no clear trend in the participation rates of men age 60-64 in England and Wales up to the 1960s. 90 age 55-64 in Canada had fallen by nineteen percentage points between 1976 and 1995. The participation rate of these men also declined substantially in the U.S., by ten percentage points between 1976 and 1994. In the U.K., the participation rate of men age 55-64 fell by 23 percentage points between 1977 and 1995.6 Since the mid-1990s, however, the participation rates of older men in each of these countries has been increasing. In Canada the participation rate of men age 55-64 increased by 8 points between 1995 and 2005. In the U.S. the participation rate of men age 55-64 rebounded by more than half of the earlier decline, increasing by 6 percentage points between 1994 and 2005. A similar pattern is observed in the U.K. where the participation rate of these men increased by 5 percentage points between 1995 and 2005. Similar trends in the participation rate of men age 55-64 can be found in several other European countries, some of which are shown in Figure 4.2.7 The participation rate of older women has also increased substantially in recent years, as shown in Figures 4.1 and 4.2, although earlier trends in their participation rates were often very different from their male counterparts. In Canada, the participation rate of women age 55-64 increased by 13 percentage points between 1995 and 2005 after increasing by less than 5 percentage points over the previous two decades. In the U.S., the participation rate for these women increased by 7 percentage points between 1994 and 2005. In the U.K., their participation rate increased by 8 percentage points between 1995 and 2005. Again, similar trends are found in other European countries. The recent increases in older women's participation may be largely explained by cohort effects. The age-participation profiles for selected birth cohorts of women in the U.S., Canada, and the U.K. are presented in Figure 4.3. Consider, for example, women born in the U.S. in 1940 - one of the earliest cohorts of women to have legal access to the pill in the 1960s.8 The large increases in women's participation rates observed in the late 1970s can in part be attributed to this 1940 cohort, and women in this cohort turned 55 in the mid-1990s. The recent increases in older women's participation can be attributed to the entry of this 1940 cohort and later cohorts to this older age group.9 In Figure 4.4 the participation rates of men and women age 60-64 by marital status are plotted for the U.S., Canada, and the U.K.. In all three countries, the recent increases in women's participation are common to single and married women. Notably, however, the participation rates of married and single men in this age group have followed different 6Note that the definition of unemployment changed over the early years of the survey and is only consis-tently defined after 1984. 7Similar trends are also found in Sweden and Denmark. See Gruber and Wise (2005) for a brief discussion of recent trends in these countries. 8Goldin and Katz (2002) find that for college women, access to oral contraception led to a later age at first marriage and greater representation of women in professional occupations. Bailey (2005) shows that legal access to the pill increased the number of women in the paid labour force. 9See Beaudry and Lemieux (1999) for a cohort analysis of female labour force participation rates in Canada. 91 trends. In both the U.S. and the U.K., the participation rate of single men age 60-64 has remained fairly constant since the mid-1980s. Similarly, in Canada the participation rates of older single men remained fairly constant through the 1990s and did not increase until after 2002. As such, the recent increases seen in the participation rates of older men has largely been driven by an increase in the participation of older married men. The similar trends in married men's and women's participation rates suggest that hus-bands may be responding to the higher likelihood of participation among wives. The follow-ing sections examine the decisions of individuals directly, evaluating how the labour force participation decisions of married men have been affected by the participation decisions of their wives. 4.3 Leisure Complementa r i ty and P a r t i c i p a t i o n Decisions In this section I present a simple static model of married men's labour supply that high-lights the competing income and shared leisure effects of a wife's participation decision on her husband's participation decision.10 I then provide estimates of the effects of wives' participation on husbands' labour force status, demonstrating that shared leisure effects dominate any income effects associated with having a wife in the labour force. I then consider other potentially important determinants of the participation decision. 4.3.1 A Simple Model of Shared Leisure and Income Effects Consider a model in which the husband and wife maximize utility independently and, unlike the standard traditional family labour supply model, each spouse treats some share of the other's income as non-labour income.uThe husband's utility is defined over current consumption (Ct) and leisure time (£^) where 0 < £f < 1. When lH = 1, the husband is not participating in the labour force. The wife's leisure time ( £ w ) is similarly defined. At any point in time then, the husband chooses his leisure time in order to maximize The model presented here only provides a simple framework for understanding the role of leisure com-plementarity at one point in time. A dynamic model would be more appropriate in the context of examining retirement transitions which is not the focus of this paper. 1 1 The standard traditional model is specified such that the husband's labour supply decisions are in-dependent of his wife's behaviour and income. Lundberg (1988) notes that joint utility models of labour supply may be inappropriate given the revocability of the marriage decision. • Lundberg (1988) also notes that bargaining models can be considered a general alternative within which joint utility and traditional models may be nested and "[rjegardless of the particular behavioural model chosen, we can treat the labor supply of husband and wife as being jointly determined, and specify a pair of simultaneous equations". s.t. Ct = X [(1 - l?)w? + i(£? = \)B? + YtH' (4.1) +K (i - 4>r + i('r = w + v w (4.2) 92 where wf and are the wage rates of the husband and wife respectively, = 1) is an indicator function equal to one when the husband is not in the labour force and zero otherwise, Bf and B^ are benefits the husband and wife may receive when they are not participating in the labour force, and and Ytw represent the non-labour income of each spouse that does not depend on their labour force status.12 The parameters A and re represent the portion of the husband's and wife's income, respectively, that the husband may use for consumption. If A = re = 1/2 the husband and wife share income equally. The husband's characteristics (X^) and the wife's leisure choice (£™) enter the utility function as preference shifters, such that they affect the husband's preferences over leisure and consumption. For example, we can expect that the marginal utility of leisure is increasing with age. Similarly, if there exists some complementarity in leisure then non-participation of a wife should also increase the marginal utility of leisure.13 The husband will choose to participate in the labour force as long as UH{Cut?,X?,lY\e? < 1) - UH(Ct,£?,X?,£?\£? = 1) > 0 (4.3) and the standard solution to the utility maximization problem implies that MRStc = ^ = w ? if i?<\ MRSic = > w? if I? = 1 where MRSec is the marginal rate of substitution between consumption and leisure. In either of the above cases, if husbands prefer to share leisure time with their wives, then we would expect that for any given value of if, Xf1, and Ct, the marginal rate of substitution between leisure and consumption will be larger when the wife is not participating in the labour force relative to when the wife is working. That is, MRSUlY = 1) > MRStdiY < 1) (4.5) The decisions of the husband can easily be depicted graphically, as in Figure 4.5. The equilibrium En represents a situation in which both the husband and wife are in the labour force. As non-labour income, the husband considers the share re of wage earnings of his wife as well as the benefits he could receive if not in the labour force. When the wife leaves the labour force, the husband's decision is affected in two ways. First, the husband's budget constraint shifts down by the difference between the wife's wage earnings and the retirement benefits she can receive. In isolation, this income effect may reduce the amount of leisure 1 2 The indicator functions $.(£%r = l)and = .1) are used here so that benefits do not affect the slope of the budget constraint. Alternatively, a 100% negative income tax could be applied to benefits, this would not significantly change the implications of this model. "Alternatively this can be stated as {dU2/dtHdPw) > 0. 93 he consumes, leaving him at the equilibrium E i . Second, the husband's relative preference for leisure may be affected such that the indifference curves are steeper when the wife is not in the labour force, represented by the indifference curve U'^. This shared leisure effect may result in the equilibrium E' : in which both the husband and wife are out of the labour force. Overall the effects of a wife's participation on the husband's participation decision will depend on the relative magnitudes of the income and shared leisure effects. To further exemplify the role of leisure complementarity, this model can be solved using a functional form for the husband's utility similar to one used in Gustman and Steinmeier (2000) , UtH = -Cr + exp(XtHpH + ^1H)£f (4.6) a where the term exp(XfJ BH+1^ 1H) determines the relative value of leisure and will depend on the wife's leisure choices. The equilibrium choice of leisure (assuming an interior solution) for the husband may be written as exp(Xf0H+£^-fH) (4.7) and KW, 7 - - ^ \ (4.8) The first term captures the effects of preferences for shared leisure on the husband's leisure choices. That is, as long as jH > 0, indicating the wife's leisure is complementary to the husband's leisure, this term will be positive. The second term can be viewed as capturing the income effects associated with the wife's leisure time. As the wife takes on more leisure, the husband loses some income associated with the wife's earnings, thereby reducing the amount of leisure he will consume. The overall effect of a wife's leisure choices on a husband's leisure decisions will depend on the relative magnitudes of shared leisure and income effects. The wife's leisure and consumption choices may be similarly described, however there is no reason to assume that the wife faces a utility maximization problem that is perfectly symmetric to the husband's. Current evidence suggests that the structure underlying the participation decision of the wife is significantly different from the decisions of husbands. For example, in their analysis of joint retirement decisions, Gustman and Steinmeier (2000) find that although the wife's retirement status has substantial impact on a husband's re-tirement preferences, the husband's retirement status does not affect the wife's preferences 94 for leisure. For the functional form presented in equation (4.6), this suggests we would have a parameter j w = 0. Furthermore, Coile (2004b) provides evidence suggesting that wives are not responsive to the financial incentives of their husbands to retire although wives respond to their own financial incentives to enter retirement, suggesting that income effects associated with the husband's labour supply choices are not as significant for the wife's de-cision. Although women's labour supply has commonly been modelled as depending on the husband's income and the husband's labour supply as not depending on the wife's income, evidence from Blau and Kahn (2005) suggests that this is no longer appropriate as women have become much less responsive to their husbands' wages.14 As women in this cohort may have only entered career employment in the 1980s and are typically younger than their husbands, women's participation decisions may be relatively independent of their husbands as their primary concern may be the attainment of their own pensions and benefits upon retirement. To note, within this simple framework we can also see that pensions, stock market returns, and other individual characteristics such as health may be important determinants of the husband's participation decision. The next section focuses on estimating the effects of wives' participation on husbands' participation decisions. This is followed by a discussion of other potential determinants, however these other factors will not be incorporated into the decomposition of older men's participation rates as they are not expected to help explain recent increases in participation. 4.3.2 Estimated Effect of Wives' Participation on Husbands' Decisions From the simple model of husbands' and wives' labour supply described above, each spouse's participation decisions may described by the latent variables LH* and Lw*, where LH* = UH(Ct^,X^,if\£? <l)-UH(Cu£?,XiH,e™\e? = 1) (4.9) represents the husband's utility associated with participating in the labour force relative to non-participation. Similarly, Lw* represents the wife's utility associated with participating in the labour force relative to non-participation. The econometric model representing the couple's participation decision may be stated as Lf = lHLT + X?tf + xY$ + e? (4.10) I?* = iwL? + X?BY + X?B™ + Z?5 + eT (4.11) where we observe = 1 if the husband participates in the labour force (Lf* > 0) and Lf — 0 otherwise. Similarly, = 1 if the wife participates in the labour force. Here, the 1 4 Blau and Kahn (2005) find that married women's cross wage elasticity fell by 38-4.7 percent between 1980 and 2000. Estimates of women's own wage elasticity fell by 50-56 percent over this period. 95 husband's participation decision depends on the participation decision of his wife {Ly), the husband's own characteristics (Xf), and possibly characteristics of his wife (X^)- The wife's participation decision may depend on the participation decision of her husband, each spouse's characteristics, as well as characteristics of the wife (Z™) that would not affect the participation decisions of her husband. In this model, if the shared leisure effects associated with a wife's participation dominate any income effects we would expect jH to be positive. Using data from the past decade for each country, I begin by estimating the husband's participation decision in isolation with a simple probit model.15 For a set of baseline estimates (that are later used in the decompositions), I include the husband's education, age, and the number of own children in the household as characteristics that would affect his participation decision. A full set of year indicators are also included. Education and age will capture the husband's earnings potential. The husband's age will also affect his preferences for leisure, with the expectation that an individual's marginal utility of leisure is increasing in age. Results for the United States, Canada, and the United Kingdom are presented in the first two columns of Tables 4.4, 4.5, and 4.6, respectively. In each of the countries, the large positive and significant marginal effects of a wife's participation demonstrate that shared leisure effects dominate any income effects associated with the wife's participation in the labour force. In the U.S., the presence of a wife in the labour force is expected to increase a husband's likelihood of participating by 19 percentage points while in Canada the presence of a wife in the labour force increases the husband's likelihood of participating by 23 percentage points. The effect appears to be slightly larger in the U.K., where the presence of a wife in the labour force increases the husband's likelihood of participating by 27 percentage points.16 The results also show that in all three countries educational attainment is also an im-portant determinant of the participation decision, as more educated men are more likely to participate in the labour force. As expected, age also has a significant effect as married men are less likely to participate in the labour force as they get older.17 I then estimate the simultaneous probit model described by equations (4.10) and (4.11) using a reduced form model, which can be stated as 1 5 The years 1995-2005 are used in the Canadian and U.K. estimates, the years 1994-2005 are used in the U.S. estimates. 1 6These marginal effects are evaluated for a 60 year old high school graduate (A level or equivalent in the U.K.) in 2005 whose wife is in the labour force and no children are at home. 1 7 In the retirement literature it is suggested that a coefficient on age may also be picking up policy effects as many retirement income programs use age as the key variable defining eligibility. For this reason, models estimating retirement hazards will use a more flexible functional form for age (ie. a set of indicator variables). Such spikes do not appear as clearly for the participation decision and a linear function of age in the equation is reasonable. H (4.12) (4.13) w* w 96 and thought of as analogous to the reduced form equations used in a two-stage least squares model.18 As the instrument (Zw) for the wife's labour force status, I use the cohort-specific labour force participation rates of women at age forty.19 Here, cohorts are defined by the wife's year of birth. This instrument is designed to capture the step function observed across cohorts in the age participation profiles of women, indicated in Figure 4.3, with each successive cohort being more likely to participate in the labour force at all ages.20 As such, there is substantial variation in the participation rates of each cohort of women at age forty and we can expect the cohort's participation rates to be strongly correlated with wives' current participation decisions. Use of this instrument also requires that the past participation rates of a wife's cohort do not directly affect a husband's current participation decision. Given the joint nature of husbands' and wives' lifetime consumption and leisure choices, we would expect that a 'young' husband's plans for retirement will depend on his expectations of his wife's future labour market activity. These expectations may depend on his wife's labour market activity at the time (or in the past) and her characteristics. However, it is reasonable to assume that the labour market activity of other women (in his wife's cohort) will not play a separate and direct role in forming those expectations or his retirement plans.21 The model represented by equations (4.12) and (4.13) is estimated as a bivariate probit model, again using data from the past decade for each country. The results for the U.S., Canada, and the U.K. are presented in the last two columns of Tables 4.4, 4.5, and 4.6, re-spectively.22 In Canada and the U.S., the resulting marginal effects of a wife's participation are not significantly different from the probit model estimates. In the U.K., however, the resulting marginal effect is much smaller than the probit estimate, suggesting the probit es-timates were biased upward. This could, for example, be a reflection of assortative matching in the marriage market whereby individuals with similar preferences over work and leisure are more likely to match. To note, the estimated marginal effects of other variables are not significantly affected by the use of the bivariate probit model. 1 8 The reduced form model is used here because valid instruments are not available for the husband's participation decision within the data sets used in this paper. As such, I am unable to identify the effect of a husband's participation decision on the participation decision of the wife. Furthermore, the simulta-neous probit model described by equations (4.10) and (4.11) suffers from logical inconsistency, whereby the relationship between (eH,ew) and (LH,LW) is not one to one. 1 9See the appendix B for details on the construction of this measure. 2 0 Although it would be preferable to measure the intercept in the age-participation profiles (ie. partici-pation rates at age 20 for each cohort), reliable estimates are not available for earlier time periods. 2 1 Also note that we do not see similar cohort effects playing a role in the participation decisions of men. Between the ages of 25 and 54, the age-participation profiles of men do not vary substantially by cohort. 2 2Note that a full set of year dummies are also included in these specifications. 97 4.3.3 Robustness Checks - Other Factors Important For Participation Decisions The estimates presented thus far have not addressed several important factors that could potentially be important for explaining changes in the participation decisions of older men. For example, the results of the previous chapter demonstrated that health is a significant determinant of retirement decisions.23 However, it appears that the health status of Cana-dian men has not changed in recent years. In 1996 and 2003, 7% of Canadian men age 55-64 described their health as poor. The portion of these men reporting good to excellent health had not changed either, remaining stable at 78%.24 Using data from the March CPS, a variable indicating poor health was added to the baseline specification. While poor health was a significant determinant of husbands' participation, the addition of this variable did not significantly change the estimated effect of a wife's participation (see Table 4.7). The decline in employer-provided defined-benefit pension plans has been cited as part of the explanation for increased participation of older men in the U.S. (Gruber and Wise, 2005). However, the percentage of workers participating in defined benefit plans was 21% in both 1999 and 2004.25 Furthermore, between 1999 and 2004 the percentage of private sector employees participating in retirement plans actually increased from 48% to 50%.26 In Canada, the portion of men in the labour force covered by a registered pension plan fell only slightly since the mid 1990s, from 35% in 1995 to 33% in 20 02.27 As such, it is unlikely that changes in pension coverage could explain the substantial increases in participation. It has been suggested that the performance of the stock market in the U.S. has had an impact on older individuals' participation decisions. For example, Coronado and Perozek (2003) find that individuals who received unanticipated equity gains during the market boom of the late 1990s retired earlier than they had anticipated. The stock market bust of 2000 seems to be an unlikely explanation for changes in participation rates, however, given that a fair portion of the increase in older men's participation occurred prior to the stock market bust in the fall of 2000. Furthermore, Coile and Levine (2004) demonstrate that the stock market bust could not feasibly explain the recent increases in participation in the U.S. given that very few households have substantial stock holdings. Similarly, only 11% of Canadian households held stocks in 1999 and the median value of stocks (conditional on 23Several studies in the United States have found similar results including Coile (2004a), Coile and Gruber (2000) and Dwyer and Mitchell (1999). 2 4Based on responses in the Survey of Labour and Income Dynamics of a sample of men age 55-64. 2 SBureau of Labour Statistics series EBUDBINC0O0000AP. Percent of employees participating in defined benefit pension plans, all private industry. 2 6 Bureau of Labour Statistics series EBUALLRET00000AP. Percent of employees participating in all retirement plans, all private industry. 2 7 T h e portion of male paid workers covered by a RPP fell from 44% in 1995 to 41% in 2001, however most of this decrease occured between 1995 and 1998 and the coverage rates remained stable thereafter. Source: Statistics Canada, Canada's Retirement Income Programs: a statistical overview. Catalogue no. 74-507-XCB. 98 holding some positive value) was merely $10000 (Milligan, 2005). In the current sample of married men age 55-64 in the U.S., only 28% had any positive stock dividends and among those with dividends, the median value was only $600. When dividend income is added to the baseline specifications for the United States, the coefficient on stock dividend income is significant, but incredibly small in size.28 Furthermore, the addition of this variable does not change the effect of a wife's participation on her husband's participation decision. Macroeconomic effects are also important to consider here. To address this, specifica-tions of the baseline model for the United States that included state specific participation rates of 25-54 year olds were also estimated. In the U.S., the macroeconomic effects were barely significant and would be unlikely to explain any increase in participation given that the participation rates of younger workers have been falling in recent years (see Figure 4.6). Again, and more important for the current analysis, the addition of macroeconomic effects to the baseline model did not change the marginal effect of wives' participation. The esti-mates from similar specifications using Canadian data demonstrated that macroeconomic effects were a significant factor, but again did not change the marginal effect of wives' participation.29 Overall, the results of the baseline models reasonably describe the effects of wives' participation on the participation decision of husbands and are robust to the inclusion of other determinants of older men's participation decisions. The baseline model's estimates will be used in the following sections in the decomposition of older men's participation rates and other robustness checks on these estimates are discussed in section 4.4.3. 4.4 Decompos ing the Changes i n P a r t i c i p a t i o n The purpose of the decompositions in this section is to establish what portion of the increase in older married men's participation rate can be attributed to changes in wives' likelihood of participating in the labour force. I begin by using the estimates from the probit and bi-variate probit models of the previous section in a decomposition of the total change in older men's participation rates since the mid-1990s. Here, I use a decomposition methodology similar to that of DiNardo et al. (1996) which easily allows for the use of non-linear func-tions of covariates in the estimation of the participation decision. In this decomposition, I investigate how (1) changes in older men's characteristics, and (2) changes in the likelihood of married women to participate in the labour force can explain the observed increases in older married men's participation rates. Later in this section, I also consider how changes over time in the parameters describing the husbands participation decision can help explain 2 8 From the model estimates, a $1000 increase in stock dividends would reduce the likelihood of participa-tion by less than two tenths of a percentage point. 2 9 The baseline model adding the unemployment rate of 25-54 year olds in each economic region as a covariate was estimated. 99 the recent increases in older men's participation rates and provide additional evidence of the explanatory power of wives' participation for husbands' participation decisions. At the end of this section, I also present a simple Oaxaca decomposition of historical participation rates to further illustrate how changes in men's characteristics and older wives' participation have influenced the participation decisions of husbands. 4.4.1 Probit/DFL Decomposition of Changes in Participation The procedure for decomposing the total change in older married men's participation rates since the mid-1990s follows the work of DiNardo et al. (1996). The procedure is similar in spirit to the familiar Oaxaca (1973) decomposition of changes in means, however is easily generalized to allow for the use of non-linear functions of covariates in the estimation of participation rates. In each stage of the decomposition, counterfactual participation rates are created rep-resenting what the participation rate in 2005 (time t) would be if each factor had remained at its mid-1990s levels. For the decomposition of U.S. participation rates, the comparison year (time s) is 1994. For Canada and the U.K., the comparison year is 1995. The decomposition is sequential in that once the 2005 participation rate has been ad-justed for a factor, that factor remains adjusted in the next stage of the decomposition. In the primary order decomposition of older married men's participation rates described below, I begin by adjusting the 2005 participation rate for changes in older men's characteristics (including educational attainment, age structure, and whether there are children at home) followed by adjusting this participation rate for changes in the likelihood of married women to participate in the labour force. To understand the estimation procedure, it is useful to view each individual observation as a vector (LH,XH,Lw,t) made up of the husband's labour force participation status (LH), the husband's characteristics (XH), the wife's labour force participation status (Lw), and a date t, all of which are discrete random variables.30 Each observation belongs to a joint distribution F(LH,XH, Lw, t; 3, 7 , p), where B, 7 , and p are the population parameters that characterize the distribution.31 The joint distribution at one point in time is the conditional distribution F(LH, XH, Lw\t; 8,7, p). The probability function of LH at one point at time t may be written as ft(LH) = ^ ^ / ( i V ^ ^ l ^ ^ ^ f ; ^ ^ ) (4.14) X H LW 3 0 For ease of notation, in the following I use the same notation to represent the random variables LH, XH and Lw and their set of values. I also omit any subscripts that would identify each function F and / . 3 1Where p = Corr(tH,ew) is the correlation coefficient between the error terms of the latent variable model of the husband's and wife's participation, presented in equations (4.12) and (4.13). See also footnote 33. 100 E £ [f(LH\XH, L w , t L H { X H , L w = t; B, 7 , p) XH LW • f(XH\Lw,txH]Lw = t) • f(Lw\tLw = t)] . (4.15) To obtain the participation rate of husbands, the probability function above is evaluated for LH = l . 3 2 The decomposition then involves different 'datings' for the different explanatory factors. For the first stage of the decomposition, a hypothetical probability function is created to represent the participation decisions of husbands that would have prevailed in time t had the distribution of husbands' characteristics remained as it was in time s. That is, the counterfactual probability function fclt(LH) = YY[f(LH\XH,Lw,tLHlXHtLw =t;0,7,p) X H L W • f(XH\Lw,txnlLw = s) • f(Lw\tLw = t)] (4.16) = E E [f(LH\XH, Lw,tLHlXHiLw = t; (3,7, p) X H L W • TPXH{Lwf(XH\Lw,tXHlLw = t) • f(Lw\tLw = tj\ (4.17) is created, where f(XH\Lw,tXHlLw=s) ^ W = f(X»\L",tx„]LW=t) ( 4 - 1 8 ) is a reweighting function that captures the changes that have occurred in the distribution of older married men's characteristics between the years s and t. Note that by applying Bayes' rule, this reweighting function may be written as f(txH[Lw = s\XH,Lw)/f(tXHlLw = s\Lw) ^XH{LW ~ f{txH,LW = t\x»^)/ntxH{LW = t\Lwy ( 4 1 9 ) For the second stage of the decomposition, a second counterfactual probability function is created to also account for changes in older wives' likelihood to participate in the labour force. That is, the counterfactual probability function fc2t(LH) = EE[/(L"I*" =*;&7,P) x" Lw • ipXHlLwf(XH\Lw,tXHlLw = t) • f(Lw\tLw = s)] (4.20) = E E [f(LN\XH, Lw,tLH]XHtLw = t; B, 7  p) XH LW 3 2 That is, / t ( l ) = Pt{LH = 1). 101 •^XHlLwf(XH\Lw,tXHlLw =t)-TPLwf(Lw\tLw=t)\ (4.21) where the reweighting function ^ = im^=T) • (4-22) captures changes in older wives' participation decisions. This second counterfactual proba-bility function then represents the participation decisions of husbands had the distribution of their characteristics and the likelihood of their wives not changed since time s. Once estimates of the reweighting functions in equations (4.19) and (4.22) are found, they can be used to estimate counterfactual participation rates. To begin, an estimate of the participation rate at any point in time is simply a weighted average of individual's predicted likelihood to participate in the labour force. That is, where uu are sample weights for individuals in time t, G(-) represents the cumulative normal distribution function, and the predicted likelihood of participation is based on the latent variable model described in section 4.3.2 where Lft = 1 when tft > — (X-f 3 + jiff) and L|f = 1 when eft > -(X%3W + Z%5). The assumption is made that p = Corr(eH, ew) = 0 when the husband's decision is estimated using the probit model and — 1 < p < 1 when the husband's decision is estimated using the bivariate probit model.33 Note that p does not enter the calculation of any reweighting functions. To obtain estimates of the reweighting function in equation (4.19), ipxH\Lwi estimates are needed for the conditional probability of being in each year t and s. The probability of being in the year t, given men's characteristics XH and wives' participation Lw, can be estimated using the probit model P(tXHlLw = t\XH,Lw) = P(e > -SH(XH, Lw)) = G(8H(XH, Lw)) (4.24) where H(XH, Lw) is a vector of covariates that is a function of XH and Lw, G is again the cumulative normal distribution, and the model is estimated by pooling observations from the years t and s and using a dummy variable indicating that the individual is observed in year t or s as the dependent variable. Here, H(XH, Lw) is a set of dummy variables indicating each possible combination of individual's education-age-child-wife's participation status. The conditional probability P{tXH\Lw = t\Lw) is similarly estimated using a probit model with wive's participation as the only covariate. 3 3Note also that the assumption is made that the parameters describing the participation decision (/?, 7 , p) do not change over time. This assumption is relaxed and such changes accounted for in section 4 . 4 . 3 . 102 To obtain estimates for the second reweighting function in equation (4.22), TJJLW, first note that the function may be defined as lpLw = f(Lw\tLw = s) f(LW\tLw=t) = < P(Lw = \\tLW=s) P{Lw=l\tLW=t) if L w = 1 n T T W r T F l I T 1L ±J W (4.25) P(Lw=0\tLW=t) since the wife's participation status L w can only take values of 1 or 0. The probability P(LW = l\tLw = s) is then estimated as the participation rate of wives in the year s. The other probabilities are similarly estimated. These estimated reweighting functions are then multiplied by the sample weights used in equation (4.23) to create counterfactual participation rates. That is, the t = 2005 participation rate that would have prevailed had the distribution of men's characteristics remained as they were in the mid-1990s is P c l t { L a = l) = (4.26) Similarly, the 2005 participation rate that would have prevailed had the distribution of men's characteristics and wive's likelihood of participation remained as they were in the mid-1990s is estimated as PC2t(LH = l) = ZiUit$x»\LW$LwG(X»0 + 7iff) (4.27) One drawback of the sequential decomposition is that the effect of a given factor may depend on the order of the decomposition. A reverse order decomposition is therefore un-dertaken, where changes in wives' participation is considered first and followed by changes in men's characteristics, to check whether the impact of wives' participation is overstated in the primary order decomposition. In the reverse order decomposition, the relevant reweight-ing functions are defined as 1pXH f(tLw]XH = s\XH,Lw)/f(tLwlXH = s\XH) f(tLW[XH = t\X»,LW)/f(tLW\XH = t\X*l) f(tXH = S\XH)/f{tXH = S) f(tXH = t\X**)/f(tXH = t) (4.28) (4.29) and can be estimated in a similar manner to that described above. Note that the reweighting function tpLw^XH can also be stated as 1/>LW\; •^XH\LWlpLW ibXH (4.30) 103 A summary of the reweighting functions used in the decompositions is provided in Table 4.8. 4.4.2 P r o b i t / D F L Decomposition Results The results of the primary and reverse order decompositions are presented in Tables 4.9, 4.10, and 4.11 for the United States, Canada, and the United Kingdom respectively. The first column in each table are the results of the decomposition when the probit model esti-mates are used. The second column in each table uses the bivariate probit model estimates in the decomposition. In the United States, the participation rates of married men age 55-64 increased by 6.4 percentage points. Considering the results of the primary order decomposition, if men's characteristics had not changed since 1994 their participation rates would have been 2.4 percentage points lower in 2005. As such, the changes we've observed in men's character-istics can explain 37% of the total change in older married men's participation rates. If the likelihood of wives to participate in the labour force had not changed since 1994, the participation rate of married men age 55-64 would have been 1.5 percentage points lower in 2005, explaining 23-27% of the total change.34 Results are similar for Canada and the United Kingdom. In both countries, however, the effect of changes in men's characteristics is much smaller, explaining between 13% and 17% of the total change in the participation rates of married men age 55-64. In part, this is due to a slightly larger change in the age structure of men in the United States relative to the same groups of men in Canada and the United Kingdom. Although changes in educational attainment are difficult to compare across the countries, it also appears that educational attainment increased more in the United States than in the other two countries. In Canada, the portion of the total change in married men's participation rates explained by changes in wives' labour force participation is much larger than in the United States, explaining between 42% and 46% of the total change. The increase in wives' participation rates was also much larger in Canada (12 percentage points between 1995 and 2005) than in the United States (7 percentage points between 1994 and 2005). In the United Kingdom, the portion explained by changes in wives' participation is slightly lower than in Canada, explaining between 28% and 34% of the total change in married men's participation rates. While the U.K. saw increases in wives' participation comparable to the increases in the U.S., the probit estimates of the marginal effect of wives' participation on husbands' participation decisions was slightly larger for the U.K. than the U.S.. 3 4 T h e unexplained portion (0.026 or 40%) represents the the portion that is not explained by changes in characteristics or wives' participation. Here, the omitted year dummy is for 2005, so the unexplained portion is effectively the coefficient on the 1994 year dummy variable. 104 4.4.3 Additional Evidence Qualitatively, the reverse order decomposition results are similar to the primary order de-composition results. However, the results suggest that the order of the decomposition matters here. In all three countries, the effect of changes in wives' labour force participa-tion is much smaller when accounted for in the first stage of the decomposition. The effect is still fairly large, however, explaining more than 12% in the U.S., 28% in Canada, and 24% in the U.K.. Furthermore, the effect of changes in men's characteristics is much larger in the reverse order decompositions. This is largest in the United States where changes in men's characteristics explain 48% of the total change in older married men's participation rates. In the decompositions presented above, it is assumed that the parameters describing the husband's participation decision do not change over time. Relaxing this assumption and accounting for changes in the parameters over time does not substantially change the results of the decompositions. In Tables 4.12, 4.13, and 4.14 the model estimates using single years of data for the U.S., Canada, and the U.K. are presented. In all three countries, the magnitudes of the marginal effects of wives' participation on husbands' decisions change substantially over time but are not significantly different over time for the U.S. or Canada.35 Although not significantly different, the change in marginal effects has a moderate impact on the reverse order decomposition results, presented in Tables 4.15, 4.16, and 4.17. Consider, for example, the results for the United States. In the primary order decomposition, the change in coefficients is accounted for in the first stage and the results of the decomposition are similar to the results in Table 4.9. In the reverse order decomposition, however, the portion of the total change in older married men's participation explained by changes in wives' participation is much larger, at 22%, since the larger marginal effect in 2005 is used in the creation of the counterfactual participation rate. Overall, however, even when potential changes in parameters over time is accounted for in these decompositions, a large portion of the total change in the participation rate of married men age 55-64 can be explained as husbands responding to the higher likelihood of wives to participation in the labour force. Several specification tests have been done to check the robustness of the probit and bivariate probit model estimates, focussing on the effect of a wife's participation on the husband's participation decision. In Tables 4.18-4.21, the wife's age is added as a covariate in these models. The addition of this covariate increases the estimated effect of wives' participation, as expected, since the wife's age is effectively controlling for age differences between the husband and wife.36 In a dynamic context, the age difference between husbands and wives may matter for the retirement decision. For example, if a husband is significantly 3 5Testing the hypothesis H0 •• 72005 — 7i994 results in a z-score=0.8 for the U.S.. Testing the hypothesis H0 • 72005 = 7i995 results in a z-score=1.55 for Canada and 2.12 for the U.K.. 3 6 Adding the wife's age as a covariate is also going to capture some cohort effect since, with the inclusion of year dummy variables, entering the wife's age is practically equivalent to entering her year of birth. 105 older than his wife he can expect his wife to be able to support his retirement longer. If retirement is a relatively permanent action, a younger wife provides the husband with some financial security. As such, controlling for age differences would capture some of the income effect associated with a wife's participation. Similarly, the inclusion of the wife's wage income captures some of the income effect associated with a wife's participation in the labour force. A specification for the U.S. that included the wife's wage income (not presented here) resulted in the wife's participation having a slightly larger marginal effect. The inclusion of the wife's hourly wages in Canada had the same effect. The inclusion of a potential wage (predicted based on the wife's age and education) did not, however, have a substantial effect on the magnitude of the effect of a wife's participation. As noted in earlier sections, the econometric model used here does not allow us to simul-taneously estimate the effect of a husband's participation decision on his wife's participation decision. To address this, I estimated a model similar to that proposed by Lewbel (2005) which uses the spouse's labour force status interacted with an indicator variable for which spouse acts first in the bivariate probit model.37 Making the assumption that the older spouse makes the participation decision first, the results using U.S. data support the hy-pothesis that husbands' and wives' leisure time are complementary to each other. Although the use of this model is informative and addresses the problem of logical inconsistency in the simultaneous equations model, the estimates are not useful in the decomposition of participation rates. Finally, if the framework for understanding older men's participation decisions is ap-propriate we would expect to see different effects for different age groups. For example, among younger married nien the income effects associated with a spouse's participation in the labour force may be stronger given expected lower levels of wives non-labour income at younger ages. As well, given the different focus on leisure activities when younger, the preferences for shared leisure may not affect the relative value of leisure as much. When the husband's participation equation is estimated using a sample of younger married men age 25-34, the marginal effect of a wife's participation is relatively small as presented in Table 4.22. I would expect this positive effect is in part capturing assortative matching in the marriage market. Among older married men, income effects may be smaller as older individuals may face a lower earnings capacity and the non-labour income of each spouse may actually be higher when both the husband and wife are not working.38 When the hus-3 7 That is, the model is stated as where D-i = 1 when the husband is older and zero otherwise. See Lewbel (2005), section 3.4. 3 8 For example, joint social security benefits in the US depend on each spouse's retirement status. In i f = 7 (1 - Dt) * ^ + Xi Px + Xi 02 + Ci (4.31) (4.32) 106 band's participation equation is estimated for married men age 65-74 the marginal effect of a spouse's participation is significantly larger. 4.4.4 LPM/Oaxaca Decompositions To further illustrate the role of changes in men's characteristics and the likelihood of wives to participate in the labour force, it is useful to consider how these factors played a role in older men's participation rates prior to the mid-1990s. In this section, I present the results of Oaxaca decompositions of historical participation rates based on the results of a simple linear probability model. Here, I illustrate how the participation rates of married men age 55-64 may have been different in Canada and the United States had their age structure, education, and wives' participation remained as it was in 1980. This exercise is also repeated for the United Kingdom, showing how participation rates would have been different had the distribution of these factors remained as it was in 1992.39 The decomposition begins with the estimation of the linear probability model l Lit = A> + £ PtDtit + liL% + J2 dkEkit + £ aeAeu + vit (4.33) t j=o fc e in which a full set of dummy variables for each categorical variable is included in the model and a zero-sum restriction on the estimated coefficients of each categorical variable is imposed.40 Here, Lw has two categories where Lff is equal to one when the wife is not in the labour force and zero otherwise, while is equal to one when the wife is in the labour force and zero otherwise. Dt represents a full set of year dummies, Ek represents education categories, and Ag represents a full set of age dummies.41 The predicted participation rates for each year are then constructed using the means of Canada, income-tested benefits such as the Guaranteed Income Supplement and Spousal Allowance are available after age 65. 3 9 Unfortunately, reliable and comparable household data is not available prior to 1992. Attempts were made to use the U K L F S 1983-1991, but husbands and wives could not be uniquely identified in several years and the resulting predicted values of participation rates were unreasonably inaccurate. 4 0 The restricted linear probability model (LPM) used here is described in Gardeazabal and Ugidos (2004) and Fortin (2005). The advantage of this model is that the results of the Oaxaca decomposition will be invariant to which categories of each categorical variable are left out of the regression. The linear probability model provides a convenient approximation to the underlying response probabilities. For example, from the U.S. model, less than one percent of observations have predicted values greater than one. The L P M is used here because the Oaxaca methodology can not easily allow for the use of non-linear models such as a probit model. 4 1 T h e coding of education is slightly different in this specification than those.in the previous section in order to be historically consistent, see Appendix B for details. Also, I have excluded from this specification the number of own children in the household given difficulties in consistently identifying children in earlier years of the CPS. 107 each categorical variable in each year. For. example, l £ • ^ . 4 , 2 0 0 5 = Po + 02005 + 77-^^,2005 + X $kEk,2005 + XI ^ ^ , 2 0 0 5 -i=o k e (4.34) is the predicted participation rate for 2005. The construction of counterfactual participation rates is then done sequentially. I begin by constructing the historical series of participation rates of older men that would have prevailed had the sample's age structure not changed since 1980 (and 1992 for the U.K.). For example, in this first stage the counterfactual participation rate for 2005, holding age structure constant at its 1980 levels is Counterfactual historical series of participation rates also holding educational attainment and wives' participation rates constant since 1980 (and 1992 for the U.K.) are similarly constructed. The resulting predicted and counterfactual participation rates are presented in Figures 4.7 and 4.8, demonstrating how age structure, educational attainment, and wives' participa-tion in the labour force have a substantial influence on participation rates. Considering how the age structure of men age 55-64 has played a role, we can see that in the United States, participation rates would have been slightly higher through the 1980s had this group not become 'older' over this period. Furthermore, participation rates would have been much lower in all three countries after the mid-1990s had this group not become 'younger' over this period. Comparing the historical series holding age structure constant (B) and that holding age structure and educational attainment constant over time (C), we can see that educational attainment has played an important role throughout the 1980s and 1990s. The results suggest that by 2005, the participation rates of older men in the United States would have been four percentage points lower than observed, had the educational attainment of these men not increased since the 1980s. The last series of counterfactual participation rates (series D) in Figures 4.7 and 4.8 il-lustrate the effect of changes in wives' likelihood of labour force participation on husband's decisions to participate in the labour force. Interestingly, in the United States wives partic-ipation does not appear to have an impact on husbands' participation until the late 1980s when we first see the participation rates of older women increase. Similarly, in Canada only small increases in older women's participation occured between the late 1980s and mid-1990s. After the mid-1990s, the effect of wives' participation is much larger given the much 4 2 This is derived by taking the expectation of equation (4.33) conditional on D2005 = 1 and assumes that E(vit\D2005it = 1) = 0. (4.35) 108 larger increases in the participation of older women in Canada over this period. This is also seen in the United Kingdom, where the participation rates of older women had not changed substantially between 1992 and 1998, thereby having very little impact on the participation rates of older men. Only after 1998, when larger increases in older wives' participation occur do we see a substantial impact on husbands' participation. 4.5 Fu ture Trends Can we expect the participation rates of older men to continue to increase? This will largely depend on the characteristics of the cohort of men that will be approaching retirement ages in the near future. Consider the characteristics of men who are currently age 45-54 (presented in Table 4.23) relative to those men who were 45-54 in previous cohorts. While the group of married men age 55-64 became slightly younger between 1995 and 2005 in all three countries considered here (i.e. 45-54 year olds became younger between 1985 and 1995), it appears that by 2015 this group of 55-64 year olds will become a bit older. As older individuals are less likely to participate, this will act to reduce participation rates. Working in the opposite direction, we can expect the education levels of married men age 55-64 to increase over the next decade in Canada and the United Kingdom. In these countries, the increase in educational attainment will counteract the effects of aging. In the United States, however, the education levels of the upcoming cohort do not differ substantially from the current cohort and cannot be expected to influence participation rates. To place some magnitudes on the future effects of changes in the age structure and education of this group, projected participation rates are created using the probit estimates presented in Tables 4.4, 4.5, and 4.6, and a reweighting function similar to that in equation (4.29) based on the changes we observe in the distribution of characteristics of 45 to 54 year olds between 1995 and 2005. More specifically, I create the reweighting function where XH includes a full set of age-education interaction dummy variables. Use of this reweighting function effectively adjusts the distribution of age and education among older married men to reflect the distribution we expect to see in 2015.43 The resulting partici-pation rates indicate that in the United States and Canada, the changes we can expect in 4 d The reweighting function created here does not account for the fact that survival probabilities are lower at higher ages or that survival probabilities could be different across education groups. For similar reasons, it would be inappropriate to to use the function f{tx = 2005\XH, Age € [45, 54})/f(tx — 2 0 0 5 | X H , Age e [55,64]). f(tx = 2005\XH, Age e [45,54]) f(tx = 1995|X", Age e [45,54]) fjtx = 2015\XH, Age € [55,64]) f(tx = 2005\XH, Age € [55,64]) (4.36) (4-37) 109 the age structure and educational attainment of this group may have the effect of slightly reducing participation rates (by only half a percentage point, as summarized in Table 4.24). In the United Kingdom, the effect of expected increases in education attainment on partic-ipation rates appear to completely offset the effect of changes in the age distribution.44 Further increases in wives' participation could drive further increases in the participation rates of older men in all three countries, however it is difficult to forecast the participation rates of next cohort of wives. Although we do not see large increases in the participation rates of wives (of men age 45-54) between 1995 and 2005, the age participation profiles of these women have flattened at younger ages for recent cohorts (as was shown in Figure 4.3). If these profiles also flatten at older ages, we may see the increases in older women's participation continue over the next decade. To predict the potential effect of increases in wives' participation, the reweighting function ijjLw = { P r ( £ w = 0 | t f , w = 2 0 1 5 ) • f r W _ r ) { ' Pr{Lw=0\tLw =2005) 1 1 ^ ~ U is applied in a similar manner as the reweighting function in equation (4.25) assuming a participation rate for wives in 2015. Assuming increases in older wives' participation rates comparable to those seen over the past decade, we would expect to see a 1.5 percentage point increase in the participation rates of older men in the United States over the next decade, attributable to husbands' preferences for sharing leisure time with their wives. Similarly, we would expect to see a 2.5 percentage point increase in older men's participation rates in the United Kingdom and a 3 percentage point increase in Canada as husbands respond to the higher likelihood of their wives to participate in the labour force. 4.6 Conclus ions This study's main finding is that a substantial portion of the recent increases in older mar-ried men's participation rates may be explained by the recent increase in the participation rates of their wives. Wives' participation in the labour force has this substantial explanatory power because wives leisure time is complementary to the leisure time of their husbands, as evidenced by the positive effect of wives' participation on husbands' participation decisions. This leisure complementarity has led to husbands participating more in the labour force as their wives have become more likely to participate. The creation of counterfactual partic-ipation rates and the resulting decompositions of older married men's participation rates demonstrates that in the United States, up to one third of the increase can be explained as a response to changes in wives' labour force participation. As much as one half of the 4 4Note that similar results are found when projections are based on the estimates in section 4.4.4 and the means in Table 4.23. 110 observed increase in Canada and two fifths of the increase in the United Kingdom can also be explained as a response to changes in wives' participation. Increases in the educational attainment of older men and changes in the age structure of men age 55-64 are also shown to have significant explanatory power for the recent increase in older men's participation rates. Changes in their characteristics account for up to one half of the increase in older married men's participation rates in the United states, and up to one third of the increase in Canada and the United Kingdom. Considering future trends in older men's participation rates, we might expect the upward trend in older men's participation rates to continue if the participation rates of older women continue to increase. Such trends would alleviate some concerns associated with our aging populations. To keep in mind, however, this study has left a large portion of the increase in older men's participation 'unexplained'. If, for instance, the performance of the labour market in the United States continues to worsen or the health of Canadians declines, this upward trend may not materialize. I l l Table 4.1. Characteristics of Married Men Age 55-64, United States All Married Men Wife in LF Wife nol in LF 1994 2005 1994 2005 1994 2005 Participation Rate 0.67 0.74 0.76 0.81 0.56 0.61 Age 59.39 59.12 58.87 58.69 60.02 59.81 Education Grade 8 or less 0.11 0.05 0.08 0.03 0.14 0.07 HS dropout 0.11 0.07 0.11 0.06 0.12 0.09 HS grad 0.34 0.30 0.35 0.28 0.32 0.33 Some PS 0.15 0.17 0.15 0.17 0.14 0.15 PS 0.05 0.08 0.05 0.08 0.04 0.07 University 0.14 0.19 0.14 0.21 0.13 0.16 Grad/Prof. 0.11 0.16 0.12 0.17 0.11 0.14 Spouse In Labour Force 0.55 0.62 1 1 0 0 Age 56.35 56.22 55.03 55.27 57.94 57.77 Age Difference 3.05 2.89 3.84 3.42 2.08 2.04 Child at Home 0.33 0.31 0.34 0.34 0.31 0.28 Number of Kids 0.47 0.46 0.50 0.50 0.44 0.40 Note. — See Appendix B for a description of the March CPS sample and variables. 112 Table 4.2. Characteristics of Married Men Age 55-64, Canada All Married Men Wife in LF Wife noi in LF 1995 2005 1995 2005 1995 2005 Participation Rate 0.61 0.68 0.76 0.79 0.47 0.53 Age 59.32 59.09 58.62 58.56 59.94 59.87 Education Grade 8 or less 0.22 0.10 0.18 0.08 0.27 0.13 HS dropout 0.20 0.12 0.20 0.12 0.20 0.13 HS grad 0.15 0.17 0.15 0.18 0.15 0.16 Some PS 0.04 0.05 0.05 0.05 0.04 0.05 PS 0.25 0.33 0.26 0.34 0.23 0.32 University 0.07 0.13 0.09 0.13 0.06 0.11 Grad/Prof. 0.06 0.10 0.08 0.11 0.05 0.09 Spouse In Labour Force 0.47 0.59 1 1 0 0 Age 56.09 56.06 54.25 54.84 57.72 57.86 Age Difference 3.23 3.02 4.37 3.72 2.22 2.01 Child at Home 0.37 0.32 0.43 0.37 0.32 0.24 Number of Kids 0.56 0.49 0.64 0.58 0.48 0.36 Note. — See Appendix B for a description of the CLFS sample and variables. 113 Table 4.3. Characteristics of Married Men Age 55-64, United Kingdom All Married Men Wife in LF Wife not in LF 1995 2005 1995 2005 1995 2005 Participation Rate 0.65 0.71 0.80 0.85 0.48 0.52 Age 59.33 59.16 58.66 58.55 60.05 60.06 Education No qualifications 0.28 0.19 0.27 0.16 0.30 0.23 Other qualifications 0.12 0.11 0.12 0.11 0.12 0.12 CSE below grade 1 0.01 0.01 0.01 0.01 0.01 0.01 GCSE A-C or equivalent 0.07 0.11 0.07 0.11 0.06 0.10 A level or equivalent 0.34 0.32 0.34 0.33 0.33 0.30 Higher educ, below degree 0.07 0.09 0.07 0.09 0.06 0.08 Degree or higher 0.11 0.17 0.11 0.18 0.11 0.15 Spouse In Labour Force 0.52 0.60 1 1 0 0 Age 56.53 56.72 55.07 55.33 58.12 58.77 Age Difference 2.80 2.44 3.59 3.22 1.93 1.28 Child at Home 0.30 0.29 0.34 0.35 0.27 0.21 Number of Kids 0.41 0.42 0.45 0.49 0.37 0.31 Note. — See Appendix B for a description of the UKLFS sample and variables. 114 Table 4.4. Model Estimates, Pooled Samples, United States Dependent Variable: Husband's Labour Force Participation Probit Bivariate Probit Coefficient Marginal Effect Coefficient Marginal Effect Wife Participates 0.504 0.186 0.592 0.219 (.014) (.005) (.076) (.028) Education Grade 8 or less -0.205 -0.072 -0.191 -0.065 (.026) (.009) (.028) (.011) HS dropout -0.134 -0.046 -0.130 -0.044 (.024) (.009) (.025) (.009) Some PS 0.124 0.039 0.121 0.037 (.021) (.006) (.021) (.007) PS degree 0.109 0.035 0.105 0.033 (.030) (.009) (.030) (.009) University degree 0.309 0.091 0.304 0.088 (.021) (.006) (.022) (.007) Grad/Prof. degree 0.448 0.125 0.444 0.121 (.023) (.006) (.023) (.007) Age -0.121 -0.040 -0.118 -0.038 (.002) (.001) (.003) (.002) Number of Kids 0.086 0.028 0.085 0.027 (.009) (.003) (.009) (.003) Constant 7.374 7.158 (.151) (.242) Note. — Standard errors are in parentheses. Sample includes married men age 55-64 in 1994-2005. See Appendix B for details regarding sample selection and explanatory variables. The models include a full set of year dummies. Marginal effects are evaluated for a 60 year old high school graduate in 2005 whose wife is in the labour force and there are no children at home. 115 Table 4 .5 . Model Estimates, Pooled Samples, Canada Dependent Variable: Husband's Labour Force Participation Probit Bivariate Probit Coefficient Marginal Effect, Coefficient Marginal Effect Wife Participates 0.604 0.228 0.522 0.198 (.011) (.004) (.044) (.016) Education Grade 8 or less -0.138 -0.048 -0.148 -0.053 (.018) (.006) (.019) (.007) HS dropout -0.052 -0.018 -0.055 -0.019 (.018) (.006) (.018) (.006) Some PS 0.007 0.002 0.007 0.003 (.026) (.009) (.026) (.009) PS degree 0.066 0.022 0.065 0.022 (.016) . (.005) (.016) (.006) University degree 0.008 0.003 0.009 0.003 (.022) (.007) (.022) (.008) Grad/Prof, degree 0.241 0.075 0.244 0.077 (.024) (.007) (.024) (.008) Age -0.111 -0.037 -0.114 -0.039 (.002) (.001) (.002) (.001) Number of Kids 0.128 0.043 0.131 0.045 (.007) (.003) (.008) (.003) Constant 6.663 (.115) 6.880 (.161) Note. — Standard errors are in parentheses. Sample includes married men age 55-64 in 1995-2005. See Appendix B for details regarding sample selection and explanatory variables. The models include a full set of year dummies. Marginal effects are evaluated for a 60 year old high school graduate in 2005 whose wife is in the labour force and there are no children at home. 116 Table 4 .6 . Model Estimates, Pooled Samples, United Kingdom Dependent Variable: Husband's Labour Force Participation Probit Bivariate Probit Coefficient Marginal Effect Coefficient Marginal Effect Wife Participates 0.785 0.268 0.560 0.191 (.011) (.005) (.059) (.020) Education No qualifications -0.284 -0.084 -0.319 -0.102 (.015) (.005) (.017) (.006) Other qualifications -0.006 -0.002 -0.018 -0.005 (.019) (.005) (.021) (.006) C S E below grade 1 -0.095 -0.026 -0.103 -0.030 (.056) (.016) (.062) (.019) G C S E A - C or equivalent -0.043 -0.011 -0.052 -0.015 (.022) (.006) (.024) (.007) Higher educ, below degree -0.057 -0.015 -0.050 -0.014 (.022) (.006) (.025) (.007) Degree or higher 0.051 0.013 0.048 0.013 (.019) (.005) (.021) (.006) Age -0.105 -0.027 -0.114 -0.032 (.002) (.001) (.003) (.002) Number of Kids 0.041 0.011 0.045 0.013 (.008) (.002) (.009) (.003) Constant 6.422 7.098 (.125) (.210) Note. — Standard errors axe in parentheses. Sample includes married men age 55-64 in 1995-2005. See Appendix B for details regarding sample selection and explanatory variables. The models include a full set of year dummies. Marginal effects are evaluated for a 60 year old with A-level qualifications in 2005 whose wife is in the labour force and there are no children at home. 117 Table 4.7. Estimated Effect of Wives' Participation, Pooled Samples, United States Dependent Variable: Husband's Labour Force Participation Probit Bivariate Probit Coefficient Marginal Effect Coefficient Marginal Effect Baseline 0.504 0.186 0.592 0.219 (.014) (.005) (.076) (.028) Specification Includes Poor Healtha 0.505 0.178 0.613 0.216 (.015) (.006) (.087) (.031) Stock Market Dividends1" 0.503 0.186 0.594 0.220 (.014) (.005) (.076) (.028) Macroeconomic Effects0 0.504 0.187 0.590 0.219 (.014) (.005) (.075) (.028) Note. — Baseline estimates refer to the estimates presented in Table 4.4. aOnly data from 1996-2005 is used here. An indicator for poor health is used, set to 0 for the calculation of marginal effects. bThe value of stock dividends is used here, set to $555 for the calculation of marginal effects Specifications include year-state specific unemployment rates of 25-54 year olds, set at 4% for the calculation of marginal effects. 118 Table 4.8. Weights and Coefficients used in the Decompositions Primary Order Decompositions Sample Used? Pooled 1995-2005 Annual 1995&2005 Weights Coef. b Weights Coef. 2005 participation rate PP Accounting' for: A coefficients - - P06 A men's characteristics ^i05"4>X"\Lw Por, A wives' participation PP Wiaz,lbX»\LW'4>Lw 095 1995 participation rate PP Wj95 p9S Reverse Order Decompositions Sample Used: Pooled 1995-2005 Annual 1995&2005 Weights Coef. Weights Coef. 2005 participation rate PP Po5 Accounting for: A wives' participation ^>i0^i>Lw\X" PP ^>i05'>pLw\X" P05 A men's characteristics UM?,ljJT,w\x"1pX" PP Ui05'4>Lw\X'1'llJXn A coefficients - ^i05^Lw[XH4'XH 096 1995 participation rate W(95 PP W i 9 5 p95 a U . S . Decompositions are using 1994, not 1995. hpp refers to the estimates presented in Tables 4.4, 4.5, and 4.6, noting the models include a set of dummy variables indicating each year. 119 Table 4.9. Decomposi t ion Results (Using Pooled Est imates) , U n i t e d States ( U a (2)b Predicted participation rate 2005 0.734 0.732 Predicted participation rate 1994 0.670 0.668 Total change 0.064 0.064 Primary Order Decomposition Effect of: Change in men's characteristics 0.024 0.023 (37%) (36%) Change in wives participation 0.015 0.017 (23%) (27%) Unexplained 0.026 0.024 (40%) (38%) Reverse Order Decomposition Effect of: Change in wives participation 0.008 0.009 (12%) (14%) Change in men's characteristics 0.031 0.030 (48%) (48%) Unexplained 0.026 0.024 (40%) (38%) aEstimates using the probit model estimates of the hus-band's labour force participation, presented in Table 4.4. ''Estimates using the bivariate probit model estimates of the husband's labour force participation, presented in Table 4.4. 120 4.10. Decomposition Results (Using Pooled Estimates), Canada ( l ) a (2) b Predicted participation rate 2005 0.681 0.683 Predicted participation rate 1995 0.609 0.610 Total change 0.072 0.073 Primary Order Decomposition Effect of: Change in men's characteristics 0.010 0.011 (14%) (15%) Change in wives participation 0.033 0.030 (46%) (42%) Unexplained 0.029 0.032 (40%) (44%) Reverse Order Decomposition Effect of: Change in wives participation 0.020 0.017 (28%) (23%) Change in men's characteristics 0.024 0.024 (33%) (32%) Unexplained 0.029 0.032 (40%) (44%) aEstimates using the probit model estimates of the hus-band's labour force participation, presented in Table 4.5. ''Estimates using the bivariate probit model estimates of the husband's labour force participation, presented in Table 4.5. 121 Table 4.11. Decomposition Results (Using Pooled Estimates), United Kingdom ( l ) a (2) b Predicted participation rate 2005 0.714 0.719 Predicted participation rate 1995 0.648 0.656 Total change 0.066 0.062 Primary Order Decomposition Effect of: Change in men's characteristics 0.009 0.010 (13%) (17%) Change in wives participation 0.023 0.017 (34%) (28%) Unexplained 0.034 0.034 (52%) (55%) Reverse Order Decomposition Effect of: Change in wives participation 0.016 .011 (24%) (17%) Change in men's characteristics 0.016 0.017 (24%) (27%) Unexplained 0.034 0.034 (52%) (55%) "Estimates using the probit model estimates of the hus-band's labour force participation, presented in Table 4.6. ''Estimates using the bivariate probit model estimates of the husband's labour force participation, presented in Table 4.6. 122 Table 4.12. Model Estimates, Annual, United States Dependent Variable: Husband's Labour Force Participation Probit Bivaxiate Probit 1994 2005 1994 2005 Coef. M . E . Coef. M . E . Coef. M . E . Coef. M . E . Wife Participates 0.421 0.159 0.486 0.181 0.505 0.190 0.791 0.293 (.047) (.018) (.043) (.016) (.240) (.089) (.258) (.093) Education Grade 8 or less -0.303 -0.112 -0.210 -0.074 -0.290 -0.106 -0.166 -0.054 (.079) (.030) (.090) (.033) (.087) (.035) (.098) (.035) HS dropout -0.032 -0.011 -0.178 -0.063 -0.028 -0.009 -0.152 -0.049 (.081) (.028) (.084) (.030) (.082) (.028) (.087) (.030) Some PS 0.111 0.037 0.138 0.044 0.112 0.037 0.121 0.036 (.073) (.024) (.064) (.020) (.073) (.024) (.066) (.020) PS degree 0.026 0.009 0.042 0.014 0.026 0.009 0.020 0.006 (.113) (.039) (.086) (.028) (.113) (.038) (.086) (.026) University degree 0.202 0.066 0.439 0.126 0.202 0.065 0.406 0.106 (.076) (.024) (.064) (.018) (.076) (.024) (.071) (.023) Grad/Prof. degree 0.287 0.091 0.548 0.150 0.288 0.089 0.518 0.128 (.088) (.026) (.072) (.018) (.088) (.026) (.079) (.025) Age -0.133 -0.046 -0.094 -0.031 -0.130 -0.044 -0.083 -0.026 (.009) (.003) (.008) (.003) (.012) (.006) (.013) (.006) Number of Kids 0.083 0.029 0.101 0.034 0.082 0.028 0.094 0.029 (.030) (.011) (.028) (.010) (.030) (.011) (.029) (.010) Constant 8.058 5.723 7.837 4.909 (.524) (.475) (.824) (.879) Note. — Standard errors are in parentheses. Samples includes married men age 55-64 in 1994 or 2005. See Appendix B for details regarding sample selection and explanatory variables. The models include a full set of year dummies. Marginal effects are evaluated for a 60 year old high school graduate whose wife is in the labour force and there are no children at home. 123 Table 4.13. Model Estimates, Annual, Canada Dependent Variable: Husband's Labour Force Participation Probit Bivariate Probit 1995 2005 1995 2005 Coef. M . E . Coef. M . E . Coef. M . E . Coef. M . E . Wife Participates 0.633 0.244 0.588 0.220 0.701 0.269 0.375 0.141 (.035) (.013) (.034) (.013) (.130) (.048) (.179) (.067) Education Grade 8 or less -0.160 -0.058 -0.178 -0.062 -0.155 -0.056 -0.204 -0.074 (.057) (.021) (.065) (.023) (.058) (.021) (.069) (.026) HS dropout -0.053 -0.019 -0.131 -0.045 -0.054 -0.019 -0.137 -0.050 (.057) (.020) (.062) (.021) (.057) (.020) (.062) (.023) Some PS 0.107 0.036 -0.133 -0.046 0.103 0.034 -0.141 -0.051 (.095) (.032) (.084) (.029) (.096) (.032) (.083) (.031) PS degree 0.075 0.026 0.038 0.012 0.071 0.024 0.035 0.012 (.056) (.019) (.051) (.017) (.057) (.019) (.051) (.018) University degree 0.140 0.047 -0.075 -0.025 0.134 0.044 -0.076 -0.027 (.082) (.027) (.066) (.023) (.083) (.028) (.066) (.024) Grad/Prof. degree 0.387 0.121 0.166 0.052 0.378 0.115 0.167 0.056 (.087) (.025) (.070) (.022) (.088) (.027) (.070) (.023) Age -0.124 -0.044 -0.109 -0.036 -0.122 -0.042 -0.117 -0.041 (.006) (.003) (.006) (.002) (.008) (.004) (.009) (.004) Number of Kids 0.108 0.038 0.177 0.059 0.106 0.037 0.186 0.065 (.022) (.008) (.026) (.009) (.023) (.008) (.027) (.011) Constant 7.331 6.580 7.151 7.135 (.378) (.381) (.515) (.582) Note. — Standard errors are in parentheses. Samples includes married men age 55-64 in 1995 or 2005. See Appendix B for details regarding sample selection and explanatory variables. The models include a full set of year dummies. Marginal effects are evaluated for a 60 year old high school graduate whose wife is in the labour force and there are no children at home. 124 Table 4.14. Model Estimates, Annual, United Kingdom Dependent Variable: Husband's Labour Force Participation Probit Bivariate Probit 1995 2005 1995 2005 Coef. M . E . Coef. M . E . Coef. M . E . Coef. M . E . Wife Participates 0.769 0.282 0.829 0.277 0.357 0.133 1.002 0.334 (.036) (.013) (.039) (.013) (.178) (.066) (.203) (.068) Education No qualifications -0.190 -0.062 -0.353 -0.102 -0.155 -0.056 -0.354 -0.096 (.045) (.015) (.053) (.016) (.054) (.020) (.059) (.020) Other qualifications 0.023 0.007 -0.014 -0.004 -0.005 -0.002 -0.015 -0.003 (.061) (.018) (.066) (.017) (.071) (.025) (.069) (.016) C S E below grade 1 0.182 0.051 -0.340 -0.098 0.273 0.087 -0.407 -0.113 (.184) (.048) (.180) (.058) (.226) (.066) (.185) (.059) G C S E A - C or -0.035 -0.011 -0.045 -0.011 0.020 0.007 -0.042 -0.010 equivalent (.075) (.023) (.068) (.017) (.088) (.031) (.070) (.017) Higher educ, 0.009 0.003 0.008 0.002 0.057 0.020 -0.009 -0.002 below degree (.075) (.022) (.074) (.018) (.087) (.030) (.076) (.017) Degree or higher 0.139 0.040 0.016 0.004 0.176 0.058 0.006 0.001 (.063) (.018) (.059) (.014) (.075) (.024) (.061) (.014) Age -0.103 -0.031 -0.100 -0.025 -0.120 -0.042 -0.093 -0.021 (.007) (.002) (.007) (.002) (.010) (.005) (.012) (.004) Number of Kids 0.070 0.021 0.005 0.001 0.062 0.022 0.004 0.001 (.027) (.008) (.028) (.007) (.032) (.011) (.030) (.007) Constant 6.146 6.171 7.352 5.608 (.394) (.440) (.646) (.803) Note. — Standard errors are in parentheses. Samples includes married men age 55-64 in 1995 or 2005. See Appendix B for details regarding sample selection and explanatory variables. The models include a full set of year dummies. Marginal effects are evaluated for a 60 year old with A-level qualifications whose wife is in the labour force and there are no children at home. 125 Table 4.15. Decomposition Results (Using Annual Estimates), United States (l)a (2)b Predicted participation rate 2005 0.735 0.726 Predicted participation rate 1994 0.670 0.669 Total change 0.065 0.057 Primary Order Decomposition Effect of: Change in coefficients 0.029 0.020 (45%) (35%) Change in men's characteristics 0.022 0.021 (34%) (37%) Change in wives participation 0.014 0.016 (21%) (28%) Reverse Order Decomposition Effect of: Change in wives participation 0.008 0.012 (12%) (22%) Change in men's characteristics 0.032 0.030 (49%) (52%) Change in coefficients 0.026 0.015 (39%) (27%) aEstimates using the probit model estimates of the hus-band's labour force participation, presented in Table 4.12. bEstimates using the bivariate probit model estimates of the husband's labour force participation, presented in Table 4.12. 126 Table 4.16. Decomposition Results (Using Annual Estimates), Canada (l)a (2)b Predicted participation rate 2005 0.682 0.684 Predicted participation rate 1995 0.609 0.607 Total change 0.073 0.076 Primary Order Decomposition Effect of: Change in coefficients 0.021 0.022 (29%) (29%) Change in men's characteristics 0.016 0.015 (22%) (20%) Change in wives participation 0.036 0.039 (50%) (51%) Reverse Order Decomposition Effect of: Change in wives participation 0.020 0.012 (27%) (16%) Change in men's characteristics 0.022 0.022 (30%) (29%) Change in coefficients 0.031 0.042 (43%) (55%) aEstimates using the probit model estimates of the hus-band's labour force participation, presented in Table 4.13. bEstimates using the bivariate probit model estimates of the husband's labour force participation, presented in Table 4.13. 127 Table 4.17. Decomposition Results (Using Animal Estimates), United Kingdom (l)a (2)b Predicted participation rate 2005 0.714 0.706 Predicted participation rate 1995 0.648 0.656 Total change 0.065 0.050 Primary Order Decomposition Effect of: Change in coefficients 0.033 0.026 (51%) (51%) Change in men's characteristics 0.008 0.010 (13%) (21%) Change in wives participation 0.023 0.014 (36%) (27%) Reverse Order Decomposition Effect of: Change in wives participation 0.016 0.020 (25%) (40%) Change in men's characteristics 0.018 0.017 (27%) (34%) Change in coefficients 0.031 0.012 (47%) (24%) aEstimates using the probit model estimates of the hus-band's labour force participation, presented in Table 4.14. bEstimates using the bivariate probit model estimates of the husband's labour force participation, presented in Table 4.14. 128 Table 4.18. Robustness checks - Model Estimates, Pooled Samples, United States Dependent Variable: Husband's Labour Force Participation Probit Bivariate Probit Coefficient Marginal Effect Coefficient Marginal Effect Wife Participates 0.506 0.187 1.329 0.484 (.014) (.005) (.098) (.033) Grade 8 or less -0.205 -0.072 -0.067 -0.019 (.026) (.009) (.032) (.009) HS dropout -0.134 -0.046 -0.080 -0.022 (.024) (.009) (.025) (.007) Some PS 0.124 0.039 0.086 0.022 (.021) (.006) (.021) (.005) PS degree 0.109 0.035 0.067 0.017 (.030) (.009) (.030) (.008) University degree 0.309 0.092 0.247 0.060 (.021) (.006) (.023) (.006) Grad/Prof. degree 0.448 0.126 0.379 0.085 (.023) (.006) (.026) (.007) Age -0.122 -0.040 -0.105 -0.028 (.003) (.001) (.004) (.002) Wife's age 0.001 0.0003 0.020 0.006 (.002) (.001) (.003) (.001) Number of Kids 0.087 0.029 0.092 0.025 (.009) (.003) (.009) (.003) Constant 7.368 4.728 (.151) (.433) Note. — Standard errors are in parentheses. Sample includes married men age 55-64 in 1994-2005. See Appendix B for details regarding sample selection and ex-planatory variables. The models include a full set of year dummies. Marginal effects are evaluated for a 60 year old high school graduate in 2005 whose 55 year old wife is in the labour force and there are no children at home. 129 Table 4.19. Robustness checks - Model Estimates, Pooled Samples, Canada Dependent Variable: Husband's Labour Force Participation Probit Bivariate Probit Coefficient Marginal Effect Coefficient Marginal Effect Wife Participates 0.612 0.231 0.695 0.262 (.011) (.004) (.287) (.106) Grade 8 or less -0.137 -0.048 -0.128 -0.044 (.019) (.006) (.039) (.016) HS dropout -0.052 -0.018 -0.049 -0.016 (.018) (.006) (.021) (.008) Some PS 0.006 0.002 0.005 0.002 (.026) (.009) (.026) (.009) PS degree 0.066 0.022 0.066 0.021 (.016) (.005) (.017). (.005) University degree 0.008 0.003 0.007 0.002 (.022) (.007) (.022) (.007) Grad/Prof. degree 0.240 0.075 0.237 0.072 (.024) (.007) (.027) (.012) Age -0.114 -0.038 -0.114 -0.037 (.002) (.001) (.004) (.004) Wife's age 0.004 0.001 0.006 0.002 (.001) (.000) (.009) (.003) Number of Kids 0.132 0.044 0.131 0.043 (.008) (.003) (.008) (.005) Constant 6.634 6.396 (.115) (.850) Note. — Standard errors are in parentheses. Sample includes married men age 55-64 in 1995-2005. See Appendix B for details regarding sample selection and ex-planatory variables. The models include a full set of year dummies. Marginal effects are evaluated for a 60 year old high school graduate in 2005 whose 55 year old wife is in the labour force and there are no children at home. 130 Table 4.20. Robustness checks - Model Estimates, Pooled Samples, United Kingdom Dependent Variable: Husband's Labour Force Participation Probit Bivariate Probit Coefficient Marginal Effect Coefficient Marginal Effect Wife Participates 0.806 0.279 0.809 0.279 (.012) (.005) (.344) (.122) No qualifications -0.283 -0.085 -0.301 -0.090 (.015) (.005) (.033) (.018) Other qualifications -0.003 -0.001 -0.008 -0.002 (.019) (.005) (.024) (.007) C S E below grade 1 -0.095 -0.026 -0.097 -0.027 (.056) (.016) (.063) (.019) G C S E A - C or equivalent -0.041 -0.011 -0.049 -0.013 (.022) (.006) (.025) (.007) Higher educ., below degree -0.057 -0.016 -0.052 -0.014 (.022) (.006) (.025) (.007) Degree or higher 0.052 0.014 0.050 0.013 (.019) (.005) (.021) (.005) Age -0.113 -0.030 -0.113 -0.030 (.002) (.001) (.006) (.005) Wife's age 0.011 0.003 0.010 0.003 (.001) (.000) (.010) (.002) Number of Kids 0.057 0.015 0.054 0.014 (.009) (.002) (.011) (.003) Constant 6.314 (.126) 6.346 (1.106) Note. — Standard errors are in parentheses. Sample includes married men age 55-64 in 1995-2005. See Appendix B for details regarding sample selection and explanatory variables. The models include a full set of year dummies. Marginal effects are evaluated for a 60 year with A-level qualifications in 2005 whose 55 year old wife is in the labour force and there are no children at home. 131 Table 4.21. Robustness checks - Model Estimates, Pooled Samples, United States Dependent Variable: Husband's Labour Force Participation Probit Bivariate Probit Coefficient Marginal Effect Coefficient Marginal Effect Wife Participates 0.501 0.167 0.810 0.288 (.014) (.005) (.391) (.140) Education Grade 8 or less -0.204 -0.070 -0.159 -0.048 (.026) (.009) (.067) (.026) HS dropout -0.134 -0.045 -0.117 -0.035 (.024) (.008) (.034) (.013) Some PS 0.127 0.040 0.115 0.031 (.021) (.006) (.027) (.010) PS degree 0.109 0.034 0.095 0.026 (.030) (.009) (.035) (.011) University degree 0.311 0.094 0.295 0.073 (.021) (.006) (.033) (.016) Grad/Prof. degree 0.449 0.130 0.434 0.100 (.023) (.006) (.035) (.019) Age 55 0.419 0.122 0.404 0.094 (.032) (.008) (.040) (.019) 56 0.306 0.092 0.292 0.072 (.031) (.008) (.038) (.016) 57 0.265 0.080 0.256 0.064 (.031) (.009) (.034) (.014) 58 0.203 0.063 0.192 0.050 (.031) (.009) (.035) (.013) 59 0.098 0.031 0.093 0.025 (.030) (.009) (.031) (.010) 61 -0.120 -0.040 -0.118 -0.035 (.030) (.010) (.030) (.010) 62 -0.401 -0.142 -0.391 -0.127 (.030) (.011) (.035) (.020) 63 -0.551 -0.200 -0.535 -0.181 (.030) (.012) (.042) (.026) 64 -0.627 -0.230 -0.604 -0.207 (.031) (.012) (.050) (.032) 1 3 2 Table 4.21—Continued Dependent Variable: Husband's Labour Force Participation Probit Bivariate Probit Coefficient Marginal Effect Coefficient Marginal Effect Wife's Age 40-44 -0.064 -0.021 t -0.102 -0.030 (.056) (.019) (.074) (.021) 45-49 -0.084 -0.028 -0.125 -0.037 (.028) (.010) (.057) (.014) 50-54 0.010 0.003 -0.016 -0.004 (.019) (.006) (.037) (.010) 60-64 -0.013 -0.004 0.033 0.009 (.019) (.006) (.063) (.016) 65+ -0.054 -0.018 0.063 0.017 (.034) (.012) (.155) (.039) Number of kids 0.092 0.030 0.093 0.026 (.009) (.003) (.009) (.004) Constant 0.219 0.022 (.033) (.259) Note. — Standard errors are in parentheses. Sample includes married men age 55-64 in 1994-2005. See Appendix B for details regarding sample selection and explanatory variables. The models include a full set of year dummies. Marginal effects are evaluated for a 60 year old high school graduate in 2005 whose 55 year old wife is in the labour force and there are no children at home. 133 Table 4.22. Effect of Wives' Participation, by Age Group, United States Dependent Variable: Husband's Labour Force Participation Age Group Coefficient Marginal Effect 25-34 0.166 0.019 (.024) (.003) 35-44 0.252 0.036 (.019) (.004) 45-54 0.340 0.069 (.016) (.004) 65-74 0.842 0.272 (.018) (.008) Note. - The probit models include covariates for age, edu-cation, number of children, and year. Sample includes married men in each age group, excluding husbands who are more than 15 years older or younger than their wives, years 1994-2005. Marginal effects are evaluated for a 30/40/50/70 year old high school graduate in 2005 with no children and whose wife is in the labour force. 134 Table 4.23 Characteristics of Married Men Age 45-54 A. Age Structure United States Canada United Kingdom 1985 1995 2005 1985 1995 2005 1985 1995 2005 Age (mean) 49.42 49.15 49.39 49.36 49.09 49.26 49.40 49.21 49.38 Age 45 0.11 0.12 0.11 0.11 0.12 0.11 0.11 0.10 0.11 Age 46 0.10 0.11 0.10 0.10 0.11 0.11 0.11 0.11 0.11 Age 47 0.10 0.13 0.11 0.10 0.12 0.11 0.10 0.12 0.10 Age 48 0.10 0.11 0.11 0.10 0.11 0.10 0.11 0.12 0.11 Age 49 0.10 0.08 0.10 0.10 0.10 0.10 0.10 0.10 0.10 Age 50 0.10 0.09 0.11 0.10 0.10 0.11 0.10 0.09 0.09 Age 51 0.10 0.09 0.09 0.09 0.08 0.09 0.09 0.09 0.10 Age 52 0.10 0.10 0.10 0.10 0.09 0.83 0.09 0.10 0.09 Age 53 0.09 0.08 0.09 0.09 0.08 0.09 0.10 0.08 0.10 Age 54 0.10 0.08 0.10 0.09 0.08 0.08 0.10 0.08 0.10 B Educational Attainment United Kingdom Canada 1985 1995 2005 1985 1995 2005 No qualifications 0.48 0.19 0.12 Grade 8 or less 0.26 0.11 0.04 Other qualifications 0.08 0.13 0.10 HS dropout 0.19 0.11 0.06 CSE below grade 1 0.00 0.02 0.03 HS 0.29 0.26 0.28 GCSE A-C or Equiv. 0.09 0.09 0.13 Some PS 0.06 0.06 0.06 A level 0.17 0.33 0.33 PS 0.10 0.29 0.34 Higher ed., < degree 0.06 0.08 0.09 University 0.10 0.18 0.22 Degree or higher 0.12 0.15 0.20 United States 1985 1995 2005 Grade 8 or less 0.10 0.05 0.04 HS dropout 0.13 0.06 0.05 HS 0.36 0.29 0.31 Some College 0.17 0.24 0.26 College Grad 0.24 0.35 0.34 C. Wives' Participation in the Labour Force United States Canada United Kingdom 1985 1995 2005 1985 1995 2005 1985 1995 2005 Rate 0.65 0.75 0.76 0.75 0.80 0.84 0.69 0.78 0.80 135 Table 4.24. Forecasted Changes in Men's Participation Rates U.S. Canada U.K. Predicted 2005 Participation Rate 0.734 0.681 0.714 Effect of 2005-2015 Change in Age and Education -0.004 -0.005 -0.001 Change in Wives Participation3, 0.015 0.032 0.025 Unexplained15 0.026 0.029 0.034 Forecasted 2015 Participation Rate 0.771 0.737 0.772 aAssumes the same increase in wives participation as observed in the previous decade. b Assumes the same unexplained increase in participation ob-served over the previous decade, an obviously unfounded assump-tion. 136 0.30 "I 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 r 1976 1980 1984 1988 1992 1996 2000 2004 Figure 4.1 Participation Rates of Individuals Age 55-64, by Sex, 1976-2005 137 0.70 The Netherlands 0.75 Germany 0.75 0.65 0.55 0.45 0.35 0.25 1983 1986 1989 1992 1995 1998 2001 2004 Spain 0.15 Women 0.55 0.45 0.35 1983 1986 1989 1992 1995 1998 2001 2004 France 0.25 1983 1986 1989 1992 1995 1998 2001 2004 1983 1986 1989 1992 1995 1998 2001 2004 Figure 4.2 Participation Rates of Individuals Age 55-64. by Sex, 1983-2005 Source: Eurostat 138 0.9 0.2 I , • I 25 28 31 34 37 40 43 46 49 52 55 58 61 64 Age A. United States 0.85 0.15 I i i i i i i i i i j i i i i i j i i i i i i i i 25 28 31 34 37 40 43 46 49 52 55 58 61 64 Age B. Canada 0.15 I , , i i i , T I 25 28 31 34 37 40 43 46 49 52 55 58 61 64 Age C. United Kingdom Figure 4.3 Age-Participation Profiles for Selected Birth Cohorts of Women 139 0 .25 M a r r i e d W o m e n :—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i— 1976 1980 1984 1988 1992 1996 2 0 0 0 2 0 0 4 A. United States 0 .65 0 .55 0 .45 0 .35 0 .25 0 .15 S i n g l e W o m e n s Marr ied W o m e n 1976 1980 1984 1988 1992 1996 2 0 0 0 2 0 0 4 B. Canada 0.15 Mar r i ed W o r n e n ^ S i n g l e W o m e n 1976 1980 1984 1988 1992 1996 2 0 0 0 2 0 0 4 C. United Kingdom Figure 4.4 Participation Rates of Individuals Age 60-64, by Sex and Marital Status, 1976-2005 140 1 Figure 4.5 Shared Leisure and Income Effects of a Wife 's Depar ture F r o m the Labour Force 0.50 ~i 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 I r— 1976 1980 1984 1988 1992 1996 2000 2004 Figure 4.6 Par t i c ipa t ion Rates of Individuals Age 25-54, by Sex, 1976-2005 141 0.75 0.73 0.71 0.69 0.67 0.65 0.63 4 1980 1985 1990 1995 A. United States 2000 2005 0.80 0.75 0.70 4 0.65 0.60 0.55 v- •" ^ . *''*\\ *\ . — ( A ) ^ — / / " ( B ) . // " " ~ ~ / ,.-'(C) x - - _ \ 1980 1985 ~\ 1 T 1990 1995 B. Canada 2000 2005 Figure 4.7 Oaxaca Decompositions of Historical Participation Rates, Married Men Age 55-64, 1980-2005 Note - (A) Predicted values (B) Holding age structure constant at 1980, (C) Holding age structure and education constant at 1980, (D) Holding age structure, education and wives' participation constant at 1980. 142 0.72 0.63 -I 1 1 1 —i 1 1 1 1 . 1 1 1 1992 1994 1996 1998 2000 2002 2004 United Kingdom Figure 4.8 Oaxaca Decompositions of Historical Participation Rates, Married Men Age 55-64, 1992-2005 Note - (A) Predicted values (B) Holding age structure constant at 1992, (C) Holding age structure and education constant at 1992, (D) Holding age structure, education and wives' participation constant at 1992. 143 Chapter 5 Conclusions The main objective of this thesis was to examine the role of various factors in the labour force participation and retirement decisions of older individuals. The research presented in this thesis will be of interest to those who wish to address concerns about population aging through policy reforms or other means as it provides a greater foundation for understanding the labour market behaviour of older individuals. Chapter 2 described Canada's retirement income security system and provided a series of simulations that illustrate how many components of the system act to change the in-centives to retire. Key components affecting retirement incentives included the actuarial adjustments in CPP/QPP that do not sufficiently compensate individuals for years of fore-gone pension receipt and clawbacks to the GIS, which act independently and in concert to reduce incentives to continue working at older ages. The international evidence on public pension programs suggests that the structure of these programs affect retirement decisions. In Chapter 3, I estimated the effects of Canada's income security programs on individuals' retirement decisions. The results confirm that Canada's income security programs have significant accrual effects on retirement decisions, indicating that individuals consider the amount of wealth they can gain or lose by delaying retirement. However, unlike some earlier studies I find that Canada's income security programs do not have significant wealth effects. The discrepancy between the results in this chapter and those of previous similar studies is likely due to a relative lack of exogenous variation available to identify the effect of income security wealth over the time period studied here. The key contribution of Chapter 3 to the literature on retirement behaviour is the ex-amination of health and the incentives in employer-provided pensions,as determinants of retirement decisions in Canada. Previous Canadian studies were either not able to measure these determinants, examined health in isolation of other retirement incentives, or did not have micro-data available to estimate the effects of pension incentives. The results of this chapter indicate that individual health has a significant and substantial effect on retire-144 merit decisions. Having addressed identification issues associated with using self-assessed health measures, the results also suggest that justification bias may be an important is-sue when assessing the effects of health. The estimates in this chapter also suggest that employer-provided pensions have significant wealth and accrual effects on the decision to enter retirement. Wealth effects appear to be stronger than accrual effects, which in part captures the differences in wealth between those with and without pension plans. The ad-dition of other forms of income to the measures of wealth and accrual do not significantly change the results, except that the accrual of wealth becomes more significant, indicating that individuals consider their entire financial picture when making the decision to retire. These results should be an important consideration in policy circles - although changing the structure of income security programs may affect individuals' retirement decisions, the ef-fects of health and employer-provided pensions may trump the effects of financial incentives found in any income security programs. While the analysis presented in Chapters 2 and 3 improve our understanding of the overall effects of public and employer-provided pensions, future research is needed to clarify the potential behavioural effects of changes to specific retirement income policy parame-ters. The clearest evidence would come from studies using natural experiment approaches to identifying the effects of public or private pensions. However, thus far these studies ex-amining public pension changes have had mixed results and are not able to identify the more complex long run effects of policy change. In the absence of lengthy policy experiments, it is not clear this gap can be filled. There is potential, however, for exploring the effects of changes to employer-provided pension plans if data sets become available that directly link the provisions of employer-provided pension plans to linked employer-employee data sets (such as the Canadian Workplace Employee Survey). Chapter 4 examined recent trends in the participation rates of older men in Canada, the U.S. and the U.K., finding that a substantial portion of the recent increases in older married men's participation rates may be explained by the recent increase in the participation rates of their wives. Wives' participation in the labour force has this substantial explanatory power because wives leisure time is complementary to the leisure time of their husbands, as evidenced by the positive effect of wives' participation on husbands' participation deci-sions. This leisure complementarity has led to husbands participating more in the labour force as their wives have become more likely to participate. The creation of counterfactual participation rates and the resulting decompositions of older married men's participation rates demonstrates that in the United States, roughly one quarter of the increase can be explained as a response to changes in wives' labour force participation. As much as one half of the observed increase in Canada and two fifths of the increase in the United Kingdom can also be explained as a response to changes in wives' participation. Increases in the educational attainment of older men and changes in the age structure of men age 55-64 145 are also shown to have significant explanatory power for the recent increase in older men's participation rates. The potential for education to drive further increases in participation rates, however, appears to have been exhausted while expected increases in older wives' participation are expected to drive further increases in older men's participation. The evolution of women's labour market behaviour and its impact on the structure of retirement behaviour among men and women is an important area for future research. Not only will the higher participation rates of older wives have an impact on their husbands' participation decisions, women have come to form a larger portion of older labour force participants and the nature of their retirement decisions may differ slightly from men's. For example, the CPP/QPP provisions for time spent caring for children will take on greater importance. Maternity leave provisions in employment contracts and employment insur-ance schemes have important implications for women's employer-provided pensions and ultimately the timing of their retirement, especially as women appear to have become more career-oriented. Such policies and their potential impact on retirement merit further inves-tigation. Overall, this thesis has investigated the complex nature of older individuals' labour market behaviour and the research contributes to our understanding of retirement and labour force participation decisions among the elderly. 146 Bibliography Bailey, Martha J. (2005) 'More power to the pill: The impact of contraceptive freedom on women's labor force participation.' Vanderbilt University Economics Working Paper No. 04-WG01R Baker, Michael (2002) 'The retirement behaviour of married couples: Evidence from the spouse's allowance.' Journal of Human Resources 37, 1-34 Baker, Michael, and Dwayne Benjamin (1999a) 'Early retirement provisions and the labor force behavior of older men: Evidence from Canada.' Journal of Labor Economics Part 1, 17(4), 724-56 _ (1999b) 'How do retirement tests affect the labor supply of older men?' Journal of Public Economics 71, 27-51 Baker, Michael, Jonathan Gruber, and Kevin Milligan (2003) 'The retirement incentive ef-fects of Canada's income security programs.' Canadian Journal of Economics 36(2), 261-290 _ (2004a) 'Income security programs and retirement in Canada.' In Social Security Programs and Retirement Around the World: Micro Estimation, ed. Jonathan Gruber and David Wise (Chicago: University of Chicago Press) Baker, Michael, Mark Stabile, and Catherine Deri (2004b) 'What do self-reported, objective, measures of health measure?' Journal of Human Resources 39(4), 1067-1093 Beaudry, Paul, and Thomas Lemieux (1999) 'Evolution of the female labour force partic-ipation rate in Canada, 1976-1994: A cohort analysis.' Canadian Business Economics Summer, 1-14 Beaudry, Paul, Thomas Lemieux, and Daniel Parent (2000) 'What is happening in the youth labour market in Canada?' Canadian Public Policy XXVI(l), S59-S83 Blanchet, Didier, and Louis-Paul Pele (1999) 'Social security and retirement in France.' In Social Security and Retirement Around the World, ed. Jonathan Gruber and David Wise (Chicago: University of Chicago Press) pp. 101-133 Blau, David M. (1998) 'Labor force dynamics of older married couples.' Journal of Labor Economics 16(3), 595-629 Blau, David M., and Donna B. Gilleskie (2001) 'Retiree health insurance and the labor force behavior of older men in the 1990s.' Review of Economics and Statistics 83(1), 64-80 147 _ (2003) 'The role of retiree health insurance in the employment behavior of older men.' NBER Working Paper 10100 Blau, Francine D., and Lawrence M. Kahn (2005) 'Changes in the labor supply behavior of married women: 1980-2000.' NBER Working Paper Blondal, Sveingjorn, and Stefano Scarpetta (1998) 'The retirement decision in OECD coun-tries.' OECD Economics Department Working Paper no. 202 Borsch-Supan, Axel (2000) 'Incentive effects of social security on labor force participation: Evidence in Germany and across Europe.' Journal of Public Economics Boskin, Michael J. (1977) 'Social security and retirement decisions.' Economic Inquiry 15, 1-25 Boskin, Michael J., and Michael D. Hurd (1978) 'The effect of social security on early retirement.' Journal of Public Economics 10, 361-377 _ (1984) 'The effect of social security on retirement in the early 1970s.' Quarterly Journal of Economics 99(4), 767-790 Burtless, Gary (1986) 'Social security, unanticipated benefit increases, and the timing of retirement.' Review of Economic Studies 53(5), 781-805 Campolieti, Michele (2002) 'Disability and the labor force participation of older men in Canada.' Labour Economics 9, 405-432 Canadian Institute for Health Information, . (2002) 'National health expenditure trends 1975-2002.' National Health Expenditures Database, CD-ROM Chalmers, Lee, and Anne Milan (2005) 'Marital satisfaction during the retirement years.' Canadian Social Trends 76 (Statistics Canada Catalogue no. 11-008), 14-17 Coile, Courtney (2004a) 'Health shocks and couples' labor supply decisions.' NBER Working Paper 10810 — (2004b) 'Retirement incentives and couples' retirement decisions.' Topics in Economic Analysis and Policy Coile, Courtney, and Jonathan Gruber (2000) 'Social security and retirement.' NBER Work-ing Paper No. w7830 — (2004) 'The effect of social security on retirement in the United States.' In 'Social Security Programs and Retirement Around the World: Micro-Estimation' (Chicago: University of Chicago Press) pp. 691-729 Coile, Courtney, and Phil Levine (2004) 'Bulls, bears and retirement behavior.' NBER Working Paper Compton, Janice (2000) 'Determinants of retirement: Does money really matter?' Depart-ment of Finance Working Paper 148 Coronado, Julia, and Maria Perozek (2003) 'Wealth effects and the consumption of leisure: Retirement decisions during the stock market boom of the 1990s.' Board of Governors of the Federal Reserve System, Finance and Economics Discussion Series Currie, J., and B.C. Madrian (1999) 'Health, health insurance and the labor market.' In Handbook of Labor Economics, ed. 0. Ashenfelter and D. Card, vol. 3c (New York: Else-vier) pp. 3309-3416 Diamond, Peter A., and Jerry A. Hausman (1984) 'The retirement and unemployment behaviour of older men.' In 'Retirement and Economic Behavior' (Washington D.C.: Brookings Institution) pp. 97-134 Diamond, Peter, and Jonathan Gruber (1999) 'Social security and retirement in the United States.' In Social Security and Retirement Around the World, ed. Jonathan Gruber and David Wise (Chicago: University of Chicago Press) pp. 437-473 DiNardo, John, Nicole M. Fortin, and Thomas Lemieux (1996) 'Labor market institutions and the distribution of wages, 1973-1992: A semiparametric approach.' Econometrica 64(5), 1001-44 Dwyer, Debra Sabatini, and Olivia S. Mitchell (1999) 'Health problems as determinants of retirement: Are self-rated measures endogenous?' Journal of Health Economics 18, 173-193 Feldstein, Martin, and Jeffrey B. Liebman (2002) 'Social security.' In Handbook of Public Economics, ed. A.J. Auerbach and M. Feldstein, vol. 4 (New York: Elsevier) pp. 2245-2324 Fernandez-Val, Ivan (2005) 'Estimation of structural parameters and marginal effects in binary choice panel data models with fixed effects.' Available http://econ-www.mit.edu / graduate / candidates/download_res. php?id=178 Fields, Gary S., and Olivia S. Mitchell (1984) 'The effects of social security reforms on retirement ages and retirement incomes.' Journal of Public Economics 25, 143-159 Fortin, Nicole (2005) 'Greed, altruism, and the gender wage gap.' Working Paper, University of British Columbia Gardeazabal, Javier, and Arantza Ugidos (2004) 'More on identification in detailed wage decompositions.' Review of Economics and Statistics 86(4), 1034-1036 Goldin, Claudia, and Lawrence Katz (2002) 'The power of the pill: Oral contraceptives and women's career and marriage decisions.' Journal of Political Economy 110(4), 730-770 Gower, Dave (2004) 'Measuring the age of retirement.' Perspectives on Labour and Income Statistics Canada, Catalogue no. 75-001-XPE, 9(2), 11-17 Grierson, D., and Mike Shannon (2004) 'Mandatory retirement and older worker employ-ment.' Canadian Journal of Economics Gruber, Jonathan (2000) 'Disability insurance benefits and labor supply.' Journal of Polit-ical Economy 108(6), 1162-83 149 Gruber, Jonathan, and Brigette C Madrian (1995) 'Health insurance availability and the retirement decision.' American Economic Review 85(4), 938-948 Gruber, Jonathan, and David A. Wise (1999) Social Security and Retirement Around the World (Chicago: University of Chicago Press) _ (2004) Social Security and Retirement Around the World: Micro-Estimation (Chicago: University of Chicago Press) Gruber, Jonathan, and David Wise (2005) 'Social security programs and retirement around the world: Fiscal implications, introduction and summary.' NBER Working Paper Gunderson, Morley (2003) 'Age discrimination in employment in Canada.' Contemporary Economic Policy 21(3), 318-28 Gustman, Alan L., and Thomas L. Steinmeier (1983) 'Minimum hours constraints and retirement behavior.' Economic Inquiry 3, 77-91 _ (1984) 'Partial retirement and the analysis of retirement behavior.' Industrial and Labor Relations Review 37, 403-15 _ (1985) 'The 1983 social security reforms and labour supply adjustments of older individ-uals in the long run.' Journal of Labor Economics 3(2), 237-253 _ (1986) 'A structural retirement model.' Econometrica 54(3), 555-584 _ (2000) 'Retirement in dual-career families: A structural model.' Journal of Labor Eco-nomics 18(3), 503-545 Habtu, Roman (2002) 'Men 55 and older: Work or retire?' Perspectives on Labour and Income (Statistics Canada, Catalogue no.75-001-XIE) 3(12), 27-34 Jaeger, David A. (1997) 'Reconciling the old and new census bureau education ques-tions: Recommendations for researchers.' Journal of Business and Economic Statistics 15(3), 300-309 Johnson, Paul (1994) 'The employment and retirement of older men in England and Wales, 1881-1981.' Economic History Review 48(1), 106-128 Kesselman, Jonathan R. (2004) 'Mandatory retirement and older workers: Encouraging longer working lives.' CD. Howe Institute Commentary Krueger, Alan B., and Bruce D. Meyer (2002) 'Labor supply effects of social insurance.' In Handbook of Public Economics, ed. A.J. Auerbach and M. Feldstein, vol. 4 (New York: Elsevier) pp. 2327-2392 Krueger, Alan B., and Jrn-Steffen Pischke (1992) 'The effect of social security on labor sup-ply: A cohort analysis of the notch generation.' Journal of Labor Economics 10(4), 412-437 Lazear, Edward P. (1983) 'Pensions as severance pay.' In Financial Aspects of the U.S. Pension System, ed. Zvi Bodie and John Shover (Chicago: University of Chicago Press) pp. 58-75 150 Lazear, E.P. (1986) 'Retirement from the labor force.' In Handbook of Labor Economics, ed. O. Ashenfelter and R. Layard, vol. 1 (Elsevier Science Publishers BV) pp. 305-355 Lewbel, Arthur (2005) 'Coherence of structural models containing a dummy endogenous variable.' Unpublished manuscript, http://www2.bc.edu/ lewbel/coherell.pdf Lumsdaine, Robin L., and Olivia S. Mitchell (1999) 'New developments in the economic analysis of retirement.' In Handbook of Labor Economics, ed. A.J. Auerbach and M. Feld-stein, vol. 3C (New York: Elsevier) pp. 3261-3307 Lundberg, Shelly (1988) 'Labor supply of husbands and wives: A simultaneous equations approach.' The Review of Economics and Statistics 70(2), 224-235 Maestas, Nicole (2001) 'Labor, love and leisure: Complementarity and the timing of retire-ment by working couples.' Berkeley: Xerox Magee, William (2002) 'Effects of self-rated disability and subjective health on job separa-tion.' Income Statistics Division, Statistics Canada Marshall, Victor .W., and Phillipa .J. Clarke (1998) 'Facilitating the transition from em-ployment to retirement.' In 'Canada Health Action: Building the Legacy. Papers Com-missioned by the National Forum on Health,' vol. 2 (Sainte-Foy, Quebec: Editions Mul-tiMondes) pp. 171-207 Milligan, Kevin (2005) 'Life-cycle asset accumulation and allocation in Canada.' Canadian Journal of Economics 38(3), 1057-1106 Morissette, Rene, and Xuelin Zhang (2004) 'Retirement plan awareness.' Perspectives on Labour and Income (Statistics Canada 75-001-XIE) 5(1), 11-18 Morissette, Rene, Grant Schellenberg, and Cynthia Silver (2004) 'Retaining older work-ers.' Perspectives on Labour and Income (Statistics Canada, Catalogue no. 75-001-XIE) 5(10), 15-20 Mulligan, Casey B., and Xavier Sala-i-Martin (2004) 'Internationally common features of public old-age pensions, and their implications for models of the public sector.' B.E. Press Advances in Economic Analysis & Policy Myles, John (2004) 'The maturation of Canada's retirement income system: Income levels, income inequality and low-income among the elderly.' Statistics Canada, Catalogue no. 11F0019MPE No. 147) Oaxaca, Ronald (1973) 'Male-female wage differentials in urban labor markets.' Interna-tional Economic Review 14(3), 693-709 Palmer, Edward (2000) 'The Swedish pension reform model: Framework and issues.' World Bank Social Protection Discussion Paper Series No. 0012 Pellechio, Anthony J. (1979) 'Social security and retirement: Evidence from Canadian time series.' NBER Working Paper 151 Pesando, James, and Morley Gunderson (1988) 'Retirement incentives contained in oc-cupational pension plans and their implications for the mandatory retirement debate.' Canadian Journal of Economics 21, 244-64 _ (1991) 'Does pension wealth peak at the age of early retirement?' Industrial Relations 30, 79-95 Pescarus, Cristiana, and Maud Rivard (2005) 'Regimes de retraite d'employeur et incitations a la retraite anticipee au Canada.' Department of Finance Working Paper 2005-02 Pestieau, Pierre, and Jean-Phillipe Stijns (1999) 'Social security and retirement in Canada.' In Social Security and Retirement Around the World, ed. J. Gruber and D. Wise (Chicago: University of Chicago Press) Ransom, Roger L., and Richard Sutch (1986) 'The labor of older Americans: Retirement of men on and off the job, 1870-1937.' The Journal of Economic History 46(1), 1-30 Sprague, Alison (1988) 'Post-war fertility and female labour force participation rates.' The Economic Journal 98(392), 682-700 Statistics Canada, . (2002a) 'Canada's retirement income programs (1991-2001).' CD-ROM. Catalogue No. 74-507-XCB _ (2002b) 'Life tables, Canada, provinces and territories 1995-1997.' Catalogue No. 84-537XIE Stock, James H., and David A. Wise (1990) 'Pensions, the option value of work, and retire-ment.' Econometrica 58, 1151-80 Tompa, Emile (1999) "Transitions to retirement: Determinants of age of social security take up.' SEDAP Research Paper Wooldridge, Jeffrey M. (2002) Econometric Analysis of Cross Section and Panel Data (Cam-bridge: The MIT Press) 152 Appendix A Financial Incentives in Income Security Programs for Chapter 3 This appendix provides further discussion of the retirement incentives found in Canada's income security programs. The simulations contained here are similar to those found in Chapter 2, but rely on slightly different assumptions in the creation of income security wealth and accrual measures. Unlike the simulations in Chapter 2, I assume that individuals experience no real wage growth if they continue to work - an assumption that is made in the creation of wealth and accrual measures in Chapter 3. Also, unlike the simulations in Chapter 2, taxes are not accounted for in this section. Canada's retirement income system contains many provisions that may have both wealth and accrual effects on Canadians retirement decisions. Using the wealth and one year accrual measures, this is demonstrated the simulations presented in Tables A.l - A.4. For each simulation, I calculate the lifetime benefits that a hypothetical individual at age 55 would receive from Canada's income security system for retirement at ages 55-69. Presented in the table is the income security wealth (ISW), or the present value of lifetime income from income security benefits, if the individual retired at age 55 and at age 69. Also presented is the one year accrual in ISW for delayed retirement. The one year accrual of ISW at age 56 should be interpreted as the wealth that the 55 year old could gain by delaying retirement from age 56 to age 57. In the base case simulation, repeated in each table, I consider a single male at age 55 in 2001 who has earned the average wage since 1966 and is considering retirement between ages 55 and 69.1 I assume that the individual has no income other than earnings or income security benefits. I assume that the individual will experience no real 'Average wages are based on CANSIM II series V78310 for 1965-1983, V250810 for 1984-2000, and V1597104 for 2001. 153 wage growth if they continue to work. The same assumptions are made for any other cases presented in these tables, unless otherwise stated. For married couples, I assume the spouse is the same age as the individual, with the same earnings history, and will retire at age 60.2 In the first row of Table A.l , the single male can accrue small amounts of wealth by delaying retirement at ages 55 and 56. The drop-out provisions play a role here. This individual begins with a 40 year contributory period (1966-2005), but is allowed to drop 15% of his lowest earning years, amounting to 6 years. For retirement at 55, the years of zero earnings from age 55-59 will be dropped. By delaying retirement, the individual can fill one of these zeros with positive earnings and drop an earlier low-earnings year. These provisions continue to work until age 59, however in this example the additional earnings years turn out to be lowest earnings years and do not result in higher CPP benefits upon collection at age 60. Hence, accruals from age 57-59 are zero. If it were the case that the individual earned much more than in earlier years, these accruals would be positive. At age 60, the accrual turns negative. Although an extra year of work may raise the level of CPP benefits the individual could receive through the addition of extra drop-out months and the actuarial adjustment made to CPP benefits, the worker has to give up one year of benefits. Furthermore, any extra CPP benefits he could receive would reduce the GIS benefits he is eligible for. At age 65, the accrual becomes more sharply negative. The reason for this jump down is again the GIS. The extra year of work at age 65 not only increases the level of the CPP benefit upon collection, decreasing the GIS payment, but it also produces earned income that pushes the immediate GIS payment to zero. The remaining rows of Table A.l demonstrate how gender and marital status matter for the accrual of wealth. Women have higher ISW and accruals largely because they have higher probabilities of survival after age 55 and therefore will be able to collect benefits for a longer period of time. Married individuals' accruals remain positive at ages 60 and 61. This is in part due to the provisions for survivor benefits. In the case that the spouse passes, the survivor can receive more in survivor benefits if he/she is not already collecting CPP benefits. This is also in part due to the assumptions that the spouse retires at age 60. Although the individual foregoes a year of benefits, the spouse is able to collect benefits for that year reducing the amount that the couple is giving up. In Table A.2, the simulations demonstrate how other retirement income such as income from employer-provided pensions may affect the accruals that an individual may achieve. I assume that the individual would receive the given pension for retirement at age 55 and that subsequent years the pension will increase at a rate of 3% per year, the expected rate 2When calculating ISW, I do not consider the drop-out provisions for child care or disability in CPP/QPP, I discount future incomes at a discount rate of 3% and for the probability of survival (based on Statistics Canada's Lifetables 1996), I assume each period there is some probability the individuals' spouse will die and assign survivor benefits appropriately, and I assume that if the individual does not retire they will experience zero real wage growth. 154 of inflation used in the simulation. Here, individuals with pension incomes between $2000 and $6000 are initially eligible for some GIS benefits. The benefit reduction rate is the same across individuals, however, so accruals for individuals age 55-59 are the same when pension income is below $6000. When pension income is $8000, however, the accrual in ISW is actually higher at ages 55 and 56. This is because the individual is no longer eligible for GIS benefits, hence increases in CPP benefits for delayed retirement are no longer mitigated by a reduction in GIS. The elimination of the effect of the GIS generates stronger incentives to delay retirement at all ages. At age 60, the accruals are the same for individuals with low pension income as the base case individual, however, an interesting case arises with pension income at $6000. In this case, the individual would have been eligible for a small amount of GIS at age 60. As such the accrual is similar to the higher pension case, but not as high given they are still partly affected by a reduction in GIS. Delaying retirement until age 61 makes the person ineligible for GIS and the individual's accruals become no different than an individual with higher pension income.3 Notice that the interaction of GIS and CPP changes for delayed retirement can lead to a non-linear age profile of income security wealth. If individuals are forward looking, using the simple one-year accrual may be inappropriate for measuring the gains one can achieve for delayed retirement. In Table A.3 the simulations demonstrate how differences in earnings levels throughout the individual's work history can affect accruals, with each row representing an individual who has always earned the given fraction of average earnings. An individual who has always earned more than average has nothing to gain by delaying retirement between ages 55 and 59 given that they will be receiving the maximum CPP pension whether or not they delay retirement. For lower income individuals, accruals at age 55 and 56 are lower with lower earnings since CPP benefits will be lower. After age 60, the accruals are higher with lower earnings since the lower income individual is giving up a year of relatively lower CPP benefits. In Table A.4 the simulations demonstrate how differences in work histories may matter for the accrual of income security wealth. In the base case, the individual has a full earnings history. In the remaining rows, some years of the earnings history are replaced with zeros such that the individual has only 80, 60, 40, or 20% of their earnings history with positive earnings.4 Interestingly, for retirement at ages 55-59, each of these individuals has the exact same accruals since with additional years of work only 6 years of the history may be dropped from the contribution period for CPP and these individuals effectively drop the same years. After age 60, however, the individuals can increase the number of years 3 The increase in accruals at later ages is merely a product of the assumptions I've made for pension growth. Although the pension and the GIS thresholds are set to grow at, the same rate, they do not adjust dollar for dollar. As such, even individuals with only $2000 will eventually become ineligible for GIS (at age 68) if they continue to work. 4 I first replace years ending in a 9 or 4 with zeros, then years with an 8 or 3, then years with a 7 or 2, and finally those ending in 1 or 6. 155 dropped with each extra year worked. This works to the greatest advantage of individuals with the shortest work history, who can drop zeros from their history, whereas individuals with longer histories are dropping low earnings years, but not zero earnings years, having less of an impact on the CPP benefits they would be eligible for. Overall, these simulations demonstrate how Canada's income security system creates disincentives to continued work, especially among GIS recipients. In general, there appears to be some incentive to continue working until age 60. After age 60, there appears to be two types of individuals who would benefit most by delaying retirement - those with high retirement incomes who are not eligible for GIS benefits and those with poor work histories. 156 Table A.l . Simulation - Variation in ISW Accruals Across Gender and Marital Status ISW One Year Accruals ISW at 55 55 56 57 58 59 60 61 62 63 64 65 66 67 68 at 69 Males single 144549 31 9 0 0 0 -911 -1582 -2216 -2818 -3394 -4649 -4677 -4689 -4684 114968 married 191186 43 12 0 0 0 876 92 -596 -1195 -1720 -2078 -2480 -2835 -3143 178162 Females single 173396 35 10 0 0 0 -366 -1066 -1738 -2386 -3019 -4358 -4442 -4516 -4578 146972 married 222447 46 13 0 0 0 1265 299 -549 -1270 -1863 -2280 -2711 -3096 -3434 208869 Note. —-All dollar values in 2001 Canadian dollars. Table A.2. Simulation - Variation in ISW Accruals Across other Retirement Income Levels ISW One Year Accruals ISW at 55 55 56 57 58 59 60 61 62 63 64 65 66 67 68 at 69 Base Case 144549 31 9 0 0 0 -911 -1582 -2216 -2818 -3394 -4649 -4677 -4689 -4684 114968 Dther income $2000 136262 31 9 0 0 0 -911 -1582 -2216 -2818 -3394 -3989 -3110 -1951 -2326 114004 $4000 127974 31 9 0 0 0 -911 -1582 -2216 -1646 -754 -1087 -1538 -1951 -2326 114004 $6000 119687 31 9 0 0 0 857 938 348 -211 -754 -1087 -1538 -1951 -2326 114004 $8000 118957 47 13 0 0 0 1567 938 348 -211 -754 -1087 -1538 -1951 -2326 114004 Note. — All dollar values in 2001 Canadian dollars. All simulations are for single males. The base case has zero other retirement income. r-Table A.3. Simulation - Variation in ISW Accruals Across Earned Income ISW at 55 55 56 57 58 59 60 O 61 ne Year Accruals 62 63 64 65 66 67 68 ISW at 69 Base Case 144549 31 9 0 0 0 -911 -1582 -2216 -2818 -3394 -4649 -4677 -4689 -4684 114968 Earned income 1.5 146710 0 0 0 0 0 -863 -1568 -2240 -2879 -3484 -4677 -4717 -4740 -3877 117664 0.8 139093 24 7 0 0 0 -909 -1495 -2046 -2566 -3066 -4584 -4586 -4573 -4545 110756 0.6 129196 18 5 0 0 0 -728 -1180 -1605 -2005 -2390 -4476 -4432 -4378 -4313 103713 0.4 117762 12 3 0 0 0 -485 -787 -1070 -1336 -1593 -4352 -4258 -4157 -4049 95691 0.2 106328 6 2 0 0 0 -243 -393 -535 -668 -797 - -2414 -2355 -2294 -2229 94409 Note. — All dollar values in 2001 Canadian dollars. All simulations are for single males. The base case has average income throughout the work history. Table A.4. Simulation - Variation in ISW Accruals Across Years Worked ISW One Year Accruals ISW at 55 55 56 57 58 59 60 61 62 63 64 65 66 67 68 at 69 Base Case 144549 31 9 0 0 0 -911 -1582 -2216 -2818 -3394 -4649 -4677 -4689 -4684 114968 fears worked 0.8 135655 1381 1366 1352 1337 1323 -653 -1327 -1972 -2587 -3252 -4420 -4661 -4668 -4659 114214 0.6 125438 1381 1366 1352 1337 1323 -243 -803 -1349 -1877 -2451 -3678 -3702 -3726 -3749 110620 0.4 115203 1381 1366 1352 1337 1323 169 -278 -724 -1165 -1649 -3577 -3560 -3546 -3534 104096 0.2 104994 1381 1366 1352 1337 1323 579 245 -101 -456 -848 -3477 -3418 -3367 -3320 97590 Note. — All dollar values in 2001 Canadian dollars. All simulations are for single males. The base case has worked each year from 1966-2001. Appendix B Sample Selection and Construction of K e y Variables for Chapter 4 B . l United States Current Population Survey Annual Demographic Files (March CPS) The March CPS provides fairly extensive information on individuals' labour force sta-tus, income, job characteristics, and demographics. Each individual in a household is inter-viewed, allowing us to match an individual to their spouse and other family members. All samples exclude individuals whose labour force status, or spouse's labour force sta-tus, is unknown or missing. This effectively excludes all individuals who are married but the spouse is not present in the household, individuals who are in the armed forces, and individuals whose spouses are in the armed forces. This exclusion drops roughly one percent of observations in each year. In 2005, for example, this results in samples of 6538 married men age 55-64 and 2056 single men age 55-64. For the descriptive statistics in Table 4.1 and the decompositions, I use samples of married men age 55-64 and I further exclude any husbands that are 15 years older or 15 years younger than their wives. On one hand, this will exclude some observations where individuals have been incorrectly coded as spouses. For legitimately married individuals that are coded properly, we can expect this to exclude husbands who are least likely to be influenced by their wives' participation decisions. These exclusions leave me with 6277 observations in 2005. For the pooled samples 1994-2005, I have 56754 observations. Marital Status - In the March CPS, only legally married spouses are coded as married. In earlier years, cohabiting spouses were coded the same as roommates and are only recently identified as 'unmarried partners'. To be consistent, I have only included those legally married in my married samples. 161 Education - Generally, individual's highest level of educational attainment has been recoded into the categories (i) completion of grade 8 or less, (ii) attended high school but did not graduate, (iii) graduated high school, (iv) some post-secondary education, (v) obtained post-secondary degree, (vi) bachelor's degree (vii) graduate level or professional degree. In the Oaxaca decompositions, this is recoded into five categories following Jaeger (1997) to be historically consistent - (i) completion of grade 8 or less, (ii) attended high school but did not graduate, (iii) graduated high school, (iv) some college, and (v) college graduate. Number of Kids - Any children that are the legal or adopted son or daughter of the husband or the wife are included in the count of children in the family. Children of any age may be included here. This count may include never married children living away from home in college dormitories. The variable is constructed by using the individual's line number, the spouse line number and parent line numbers within each household to match individuals with their children. The variable for whether there are children at home is simply an indicator variable equal to one if the number of kids is positive. Cohort-specific participation rates at age 40 - Estimates for women born 1926-1965 are obtained by estimating participation rates of 40 year old women using the March CPS files for 1966-2005. Only women whose labour force status is unknown are excluded from these samples. For women born 1920-1925, estimates of the participation rates of women age 40-45 in 1960-1965 from the OECD Statistics Compendium (series U16213291) are used. Reliable estimates of participation rates for women at the age of 40 for women born before 1920 are not available, so the participation rate of women born 1920 is used here. Less than 0.02% of wives in the sample of husbands 1994-2005 were born before 1920. Only 0.2% of wives in the 1994 sample were born before 1920. B.2 C a n a d a Labour Force Survey Master Files (CLFS) The CLFS is a monthly survey that provides information regarding individuals' labour force status, job characteristics and demographics. The master files (as opposed to the public use files) provide more detailed information, including an individual's age rather than an age group. The LFS surveys the same dwelling for 6 months. Only the incoming rotation group is used in the sample. Information for each member of a sampled dwelling is collected and spousal information is provided as part of an individual's observation. Exclusions to the samples are nearly identical to those made in the March CPS. In addition to the exclusions for the CPS, for several years of the CLFS there are a handful of 162 individuals who are coded as single (never married) yet spousal information is provided for the individual and are also excluded from the samples. Marital Status - Prior to 1999, no distinction is made in the CLFS between legally married and common-law couples. Although the distinction is made in the more recent files, for the purposes of being historically consistent the Canadian common-law couples will be treated the same as legally married couples. Education - Generally, education represents the highest level of schooling completed and is recoded into the categories (i) completion of grade 8 or less, (ii) attended high school but did not graduate, (iii) graduated high school, (iv) some post-secondary education, (v) obtained post-secondary degree, (vi) bachelor's degree, and (vii) graduate level or professional degree. As the education variable changes in 1990, for the Oaxaca decomposition education is recoded into the categories (i) completion of grade 8 or less, (ii) grade 9-10 (high school drop out), (iii) grades 11-13 (high school), (iv) some post-secondary education, (v) obtained post-secondary degree, (vi) university degree. Number of Kids - The number of children in the household is derived by adding up the number of "own children" in the household, already matched to the individual in the LFS. This includes children by birth, adoption or marriage of any age and may include children who are away at school. Cohort-specific participation rates at age 40 - Estimates for women born 1936-1965 are obtained by estimating participation rates of 40 year old women using the CLFS master files for 1976-2005. Only women whose labour force status is unknown are excluded from these samples. For women born 1915-1935, the participation rates of women age 35-44 from 1955-1975 are used as estimates, obtained from various issues of Canada, Women's Bureau, Women in the labour force: facts and figures, Ottawa : Labour Canada, Women's Bureau, 1965-1977. Note these estimates match the participation rates provided by the OECD Statistical Compendium (series 316212291) 1960-1975. B.3 U n i t e d K i n g d o m Labour Force Survey (UKLFS) The UKLFS has been conducted as a quarterly survey since 1992. For the years 1992-2005, the spring quarterly household files are used. The survey is conducted biannually from 1975-1983 and annually from 1984-1991 with interviews conducted March-May. Estimates of participation rates presented for 1976, 1978, 1980, and 1982 in this paper are averages of 163 adjoining years. Exclusions to the sample are the same as those for the March CPS. Marital Status - In this paper, common-law couples have been denned as married. Prior to 1989, however, cohabiting couples were not separately identified in the UKLFS. Education - The UKLFS reports an individual's highest qualification, the coding of which is expanded over the years to include more detailed categories. I have recoded this to (i) degree or higher (which includes university degrees), (ii) higher education, below degree, (iii) A level or equivalent, (iv) GCSE A-C or equivalent, (v) CSE below grade 1 or equivalent, (vi) other qualifications, and (vii) no qualifications. The qualifications are coded in much more detail and are easily categorized consistently after 1993. Number of Kids - The number of children represents all children of the individual in the same family unit. This will only include never-married children and children who are not parents themselves. This will include any other adult children in the household. Cohort-specific participation rates at age 40 - Estimates for women born 1935-1965 are obtained by estimating the participation rate of women age 40 in each available year 1975-2005. For women age 40 in the years 1976, 1978, 1980, and 1982, an average of previous and following cohorts is used. Participation rates for earlier years are not readily available. For women born 1916-1932, estimates of the participation rate of women age 40-44 in Great Britain are constructed using various issues of the Annual Abstract of Statistics which provide estimates of the number of employees (employed women and women registered as unemployed) and population estimates for each year 1956-1972. Note that this series closely matches that constructed by Sprague (1988), except that the Sprague (1988) estimates are slightly higher given that she attempts to estimate the number of unregistered unemployed women over this time period based on the number of men not registered 1971 onwards. For women born 1933 and 1934, the participation rate at age 40 is filled with linearly interpolated values based on values for the 1932 and 1935 cohorts. B.4 Other Data Eurostat The European participation rates presented in Table 4.2 were constructed using the series for "Inactive population as a percentage of the total population (of a given sex) by age groups (%)" for each country from Eurostat (http://epp.eurostat.cec.eu.int). The series is reported quarterly since 1983 and is based on the EU Labour Force Survey. 164 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.831.1-0100419/manifest

Comment

Related Items