Journal of Happiness Studies

, Volume 15, Issue 1, pp 125–144

Subjective Well-Being and Retirement: Analysis and Policy Recommendations

Authors

    • Goldman School of Public PolicyUniversity of California, Berkeley
    • General Medical DisciplinesStanford University School of Medicine
Research Paper

DOI: 10.1007/s10902-012-9399-2

Cite this article as:
Horner, E.M. J Happiness Stud (2014) 15: 125. doi:10.1007/s10902-012-9399-2

Abstract

This study examines the relationship between retirement and subjective well-being (SWB), utilizing international data from sixteen countries in Western Europe and the US. Differences in social security regimes are exploited to estimate the retirement decision such that it is exogenous to individual-level characteristics. Although results from traditional ordinary least squares suggest an ambiguous relationship between retirement and SWB, this is due to comparatively lower SWB among those who choose retirement. The removal of selection bias reveals a large, positive effect that fades over a few years, suggesting a multi-stage adjustment to retirement. Individuals facing formal retirement at age 65 or later experience an increase in SWB that is roughly equivalent in total value to that of individuals facing earlier retirement, and both groups return to trend by age 70. This suggests that raising the formal retirement age, which is widely discussed today by policymakers, is relatively neutral with regard to SWB in the long-term.

Keywords

RetirementLife satisfactionAgingInternationalSocial security

1 Background and Motivation

1.1 Introduction

Globally, our population is aging (Coggan 2011). As a result of both increased life expectancies (Kinsella and Phillips 2005) and lower birthrates (Grant et al. 2004), a growing proportion of the world’s population is over age 65. In the United States and across much of Europe, this is made salient by the Baby Boomers, who began to turn 65 in January of 2011 (Barry 2010). Current projections of Western Europe suggest that there will be nearly two times as many people over the age of 65 as under the age of 15 by 2050 (Cendrowicz 2010). Thus, under current retirement age regimes, the size of the retired population relative to the size of the tax-base is growing (Gruber and Wise 1999), creating mounting costs with dwindling resources. Despite between-country variation in public pension programs, the vast majority of current social security programs are financially unstable (Auerbach and Lee 2011) and unsustainable (Dang et al. 2001).

In response to this fiscally untenable situation, several countries have been steadily increasing their retirement age. In Europe, analysts predict that many countries will be forced to increase the retirement age to 70, and Britain has already announced plans to increase the retirement age to 68 (The Economist 2011). France recently announced a (violently protested) increase in the formal retirement age from 60 to 62 (Bennhold 2010).

Clearly, decisions regarding what should be done about the mounting cost of social security programs must be made based upon a variety of factors including cost and long-term feasibility. However, an additional important consideration is the well-being of those near the retirement age and in retirement. This paper investigates the relationship between retirement and subjective well-being (SWB), for individuals near retirement and afterward. Any policymaker considering changing the age at which people retire ought to take into account the full welfare implication of this decision on the entire population including both taxpayers and retirees. If it were to turn out that raising the age of retirement reduces well-being of those near the retirement age, this should be considered a meaningful cost. On the other hand, if raising the retirement age increases rather than reduces the welfare of retirees, this might make a the policy change easier to implement.

Because the decision to retire is not exogenous, an identification strategy is used in the current paper to eliminate bias. It is possible that those who decide to retire are particularly prone to physical and mental health problems. The opposite is also possible, as people may be more likely to choose retirement if they have a full and exciting life outside of work. Either way, traditional ordinary least squares (OLS) methods will be biased, understating or overstating SWB among those entering retirement. The current paper utilizes an Instrumental Variables (IV) methodology to disentangle selection into retirement from retirement itself, finding that although retirement initially improves SWB, there is a steep decline within a few years. In addition, the current study shows that individuals facing relatively later formal retirement appear to have relatively equivalent long-term SWB as those facing earlier formal retirement.

1.2 Theories and Evidence on SWB and Retirement

Economic theory predicts that people should be generally happier in voluntary retirement. People will only retire if they predict some improvement in their overall welfare (formore details, see Charles 2004). Since retirees generally face a drop in pay in retirement, it follows that they should receive some other improvement in their overall quality of life if they choose to retire. There is some historical evidence that retirement is a typical normal good (Costa 1998; Kohli et al. 1991), wherein people trade in money for leisure (Aguiar and Hurst 2005). This literature suggests that people will make rational, utility-based decisions to retire.

However, the retirement decision is made under uncertainty (i.e.: individuals do not know if they will enjoy retirement since they have most likely never retired before); the decision is bulky (i.e.: individuals cannot reduce their hours by 1 h, then another, until they optimize utility); and the decision is, for the most part, irreversible. Thus, if people are poor predictors of what will make them happy, they may miscalculate when deciding whether and when to retire. If individuals are less happy in retirement, economic theory suggests this may be the result of a suboptimal response to a non-convex problem under the condition of irreversibility.

The psychological literature on retirement and SWB offers mixed predictions. On the one hand, retirement may enhance well-being as it reduces stress, increases free time, and marks a major milestone of accomplishment. On the other hand, retirement may diminish well-being by reducing social contact and removing a large portion of one’s identity (Calasanti 1996; Kim and Moen 2002). Psychologists have suggested that retirement’s effects on SWB is heterogeneous, affected by the age of retirement, the importance of one’s job to one’s identity, and other individual psychological characteristics (Michinov et al. 2008; Mutran and Reitzes 1981; Reitzes and Mutran 2004).

Regardless of the theoretical predictions, there is a persistent notion that retirement is actually detrimental to the individual (for a review, see: Ekerdt 1987). Consistent with this view, retirement has been shown to relate to reduced physical health and expedited aging (Dave et al. 2008), and to be associated with mental health problems (MacBride 1976), severe stress (Eisdorfer and Wilkie 1977) and feelings of loneliness and obsolescence (Bradford 1979). However, other studies have found the experience of retirement to be (at least somewhat) positive (Gall et al. 1997; Jackson et al. 1993; Mindanik et al. 1995), particularly for those who retire voluntarily (Bender 2004).

If adjustment to retirement is a multi-stage process, as suggested by Atchley (1976), then it is possible that the conflicting findings are the result of evaluations made at different stages of adjustment to retirement. Atchley theorized that new retirees experience a honeymoon stage, followed by a steep reduction in SWB as the realities of retirement and aging become more salient, and finally a lower, stable, permanent phase. His theory has been well supported by data (Atchley and Robinson 1982; Borsch-Supan and Jurges 2006; Ekerdt et al. 1985; Ekerdt et al. 1983; LaRue et al. 1985; Palmore 1986; Reitzes and Mutran 2004; Theriault 1994).

While these studies contribute meaningfully to our understanding of retirement and SWB, they suffer from a common shortcoming: retirement decisions and SWB are endogenous. The majority of previous research on retirement has relied on correlation or compared individual well-being before and after retirement. However, even a multitude of control variables and robustness checks cannot overcome the fact that people who choose to retire are different from people who do not choose to retire (Bound and Waidmann 2007; Charles 2004).

A few relatively recent studies have addressed the endogeneity of retirement decisions either by exploiting changes in retirement policies (Charles 2004; Neuman 2008) or by exploring indicators of welfare just around the age that social security kicks in (Bound and Waidmann 2007). In the last 2 years, two studies have exploited country-level differences in early and normal retirement ages (Coe and Zamarro 2011; Rohwedder and Willis 2010), at which point there are discontinuities in retirement incentives that have been shown to increase the retirement rate (Gruber and Wise 1999, 2004, 2007). Thus, researchers have been able to instrument for retirement such that it is uncorrelated with individual characteristics.

The above mentioned studies that address the endogeneity of retirement find that retirement improved subjective and objective health in a US sample (Charles 2004; Neuman 2008), and at least temporarily in a UK sample (Bound and Waidmann 2007). Further, retirement reduced individuals’ reports of mental illness in an international European sample (Coe and Zamarro 2011). Thus, unlike studies utilizing selection-on-observables designs, studies utilizing a good identification strategy find that retirement at least temporarily improves well-being, with one notable exception: researchers have also found that retirement expedites cognitive decline (Bonsang et al. 2012; Coe et al. 2009).

1.3 Roadmap to This Study

Unlike any previous study currently published, this project explores SWB as it relates to the overall quality of life in an international context using a quasi-experimental model. As pointed out above, many studies have evaluated this relationship using selection-on-observables designs, but only a handful have utilized a quasi-experimental design. Further, the other studies that exploit a quasi-experimental design have evaluated physical and mental illness, rather than psychological health. Understanding retirement’s relationship to SWB is important as it will allow for retirement policies to be created and adapted with an eye on the overall subjective experience of the retirees. Although Charles (2004) did perform a causal inquiry into retirement and SWB, his project focused only on the US. Thus, his paper cannot evaluate heterogeneity as a function of different formal retirement ages. Further, the US is a special case for many reasons, including strong incentives to continue working as long as possible, and the availability of public insurance (Medicare) only at age 65.

Once the retirement decision is estimated exogenously through an IV methodology, this study finds that retirement improves SWB as measured by reported life satisfaction and CASP (a multi-dimensional measure of SWB named for its components: Control, Autonomy, Self-realization, and Pleasure) (Hyde et al. 2003; Sim et al. 2011). This SWB improvement is large, but temporary, and is followed by a steep decline. Overall, the total increase in SWB for those facing younger and older retirement incentives is approximately the same in total size even when adjusting for mortality. Thus, over this portion of the life-cycle, the age of retirement incentives is neutral with regard to SWB.

Thus, the current paper adds to the existing literature on retirement and overall well-being by providing a causal evaluation of the role of retirement on SWB in an international context. Further, it investigates differential treatment effects as a result of variation in retirement policies and thus can guide policymakers. The paper proceeds as follows: Sect. 2 introduces the data and terms used in this study, Sect. 3 discusses the empirical model, Sect. 4 presents results, and Sect. 5 discusses and concludes.

2 Data

This study combines cross-sectional data from the 2006 Survey of Health, Ageing, and Retirement in Europe (SHARE) for 14 EU countries, the 2006 English Longitudinal Study of Ageing (ELSA) for the UK, and the 2004 Health and Retirement Study (HRS) for the United States.1 These three surveys provide individual-level data on demographics, socioeconomic information, health history, mental health, and psycho-social measures of health. The surveys have a variety of modules that are similar by design, and in the years used, several compatible psycho-social variables were measured.

Similar to Rohwedder and Willis (2010), a simple set of control variables are used, namely: marital status, gender, age, and age polynomials, minimizing possibly biasing differences in the wording of other demographic information. Only men are evaluated, since this women in these cohorts had a weaker relationship with the labor market (follows approach of Coe and Zamarro 2011). Season of interview is also controlled for to capture mood changes in response to the weather, such as seasonal affective disorder.

These data are supplemented by information from the Social Security Administration’s (SSA 2004, 2006) reports of international early and normal retirement ages. For a summary of early and normal retirement ages by country, see Table 1. For summary statistics, see Table 2. The sample is limited to people ages 50–70, providing a total sample of 18,345 individuals. The average age is just over 60, and 81.3 % are married. As can be seen, there is a wide range in the overall retirement rate for each country among 50–70 year olds, suggesting that retirement incentives and norms are quite varied.
Table 1

Retirement incentives by country, 2006

Men’s retirement ages

Early retirement age

Normal retirement age

Austria

61.5

65

Belgium

60

63

Czechia

58.5

61.5

Denmark

65

65

France

60

60

Germany

60

65

Greece

60

65

Ireland

65

65

Italy

57

65

Netherlands

60

65

Poland

65

65

Spain

60

65

Sweden

61

65

Switzerland

62

65

UK

65

65

USAa

62

65.4

Source Social Security Administration, 2004 and 2006

aThe USA retirement age listed is from 2004 because the HRS data is from 2004

Table 2

Summary statistics

 

N

Age

Married (%)

Retired below early ret. age (%)

Retired between early and normal (%)

Retired after normal ret. age (%)

Total

18,345

60.4

81.3

15.7

67.6

78.9

Austria

373

61.9

83.4

34.7

91.8

100.0

Germany

854

61.2

88.3

6.7

62.7

94.6

Sweden

821

61.3

86.1

11.3

35.0

92.5

Netherlands

869

59.9

87.9

7.1

53.8

89.8

Spain

612

60.2

84.8

12.0

45.3

96.5

Italy

897

61.3

90.5

9.8

69.6

94.6

France

878

59.3

86.0

14.4

N/A

89.7

Denmark

858

59.5

84.4

15.9

N/A

84.5

Greece

972

60.0

87.8

13.1

47.3

86.0

Switzerland

451

59.9

82.7

5.1

47.1

87.2

Belgium

984

59.6

85.6

15.1

64.9

92.9

Czechia

870

59.5

85.9

11.4

25.2

90.2

Poland

765

59.4

84.4

29.3

N/A

94.0

Ireland

367

59.7

79.6

19.1

N/A

68.1

UK

2,767

60.6

70.5

25.0

N/A

88.0

USA

5,007

60.8

77.2

11.6

39.9

52.8

2.1 Measures of SWB

The SWB measures employed in this study are life satisfaction and the CASP. Overall life satisfaction is commonly used in happiness studies (e.g. Blanchflower and Oswald 2011; Stevenson and Wolfers 2008), in part because this single-item measure has the advantage of clarity and face validity (Hall et al. 2010). The multi-item CASP scale, a SWB measure designed for older people, is also used (Hyde et al. 2003). This scale captures the individual’s perception of their own life by asking for qualitative evaluations of life characteristics. The 12-item version of the scale used in this study has strong internal validity (Sim et al. 2011). CASP has four components: Control, Autonomy, Self-realization, and Pleasure. Control is freedom from constraints imposed by family or lack of money; autonomy can be defined as self-sufficiency; self-realization is focused on enthusiasm about the future; and pleasure refers to finding life enjoyable and fulfilling. Clearly, items like: “I look forward to each day”, capture something purely subjective and affective. Although other items such as: “Shortage of money stops me from doing what I want” have a more objective component, individuals must decide how much money is enough and whether they feel satisfied. Insufficient funds “to do what I want” is hardly an indication of poverty. As such, these measures provide a broad subjective account of well-being along specific domains. While life satisfaction measures one component of overall SWB, CASP measures a subjective account of quality-of-life. In concert, the two outcome measures give a more comprehensive view into the experience of the individual. A summary of the items used and their summary statistics can be found in Table 3.
Table 3

SWB indicators

Item name

Questions

Life satisfaction

I am satisfied with my life (scales differ; standardized by country)

Range −3 to 3

Unstandardized mean: 1.5 (SD = 1.1)

 

Country range from 0.9 (Poland) to 2.0 (Switzerland and Denmark)

CASP

1. I can do the things that I want to do

Control

2. Family responsibilities prevent me from doing what I want to do (R)

 Range 0–9

3. Shortage of money stops me from doing what I want (R)

CASP

1. My age prevents me from doing the things I would like to (R)

Autonomy

2. I feel what happens to me is out of my control (R)

 Range 0–9

3. I felt left out of things (R)

CASP

1. I feel full of energy these days

Self-Realization

2. I feel that life is full of opportunities

 Range 0–9

3. I feel that the future looks good for me

CASP

1. I look forward to each day

Pleasure

2. I feel my life has meaning

 Range 0–9

3. On balance, I look back at my life with a sense of happiness

CASP FULL

Add up all CASP items (all CASP domains standardized by country)

 Range 0–36

Unstandardized mean: 26.3 (SD = 5.8)

 

Country range from 23.5 (Italy) to 29 (Switzerland)

Different countries may have norms not only in terms of average reports of SWB (which would be controlled for in country fixed effects) but also in the distributions. For a purely fictitious but illustrative example, it is possible that the French have both high average SWB and high variation of socially acceptable affect, while the Swedes have high SWB and a narrow band, and the Greeks have lower average SWB and a medium band. To avoid forcing a distribution that might overestimate the SWB effect in a country that values strong presentations of affect or underestimate the SWB effect in a less expressive country, SWB measures were normalized in terms of standard deviations from the mean, within country, such that:
$${\text{SWB}}_{\text{ic}} = ({\text{SWBR}}_{\text{ic}} - \overline{SWB}_{\text{c}} )/\sigma_{\text{c}}$$
where SWBic is the SWB score for individual i in country c, calculated as the difference between their raw score (SWBRic) and the country mean \(\overline{SWB}_{\text{c}}\), divided by country level standard deviations. Regression results are reported in terms of standard deviations within countries, which, combined with country fixed effects, should account for differences in norms of expression.2

2.2 Measuring Retirement Status

There are two general approaches to measuring retirement. One is to consider all people over the age of 50 who are not in the labor market as retired (used by: Bonsang et al. 2012; Charles 2004; Coe and Zamarro 2011). The downside of this approach is that homemakers, the sick and disabled, and the unemployed are considered retired. The other approach is to consider all people who call themselves retired to be retired (used by: Coe et al. 2009). This approach has good face-validity, but may capture some people who consider themselves retired because they have left their primary profession but are not entirely out of the workforce. This research will rely upon the latter definition of retirement. To reduce the noise caused by individuals who call themselves retired but are actually still working or looking for work, individuals will only be considered retired if they report that they are fully retired. Otherwise, they will be considered to have not yet retired.

2.3 Instruments

Because all individuals pass through the formal early and normal retirement ages, being eligible for early and normal retirement is exogenous to the individual when controlling for age. People are considered eligible for early or normal retirement when they have reached the early or normal retirement age for their nationality and gender. Referring back to Table 1, utilizing differences between countries and between men and women provides more variation; still, the vast majority of normal retirement ages are either 60 or 65. Although, from a researcher’s perspective, it would be preferable if there existed more variation in the retirement ages, there is no reason to suspect that any trajectory breaks should occur at these retirement ages unless it is related to retirement. Further, some countries allow firms to have mandatory retirement at a certain age (often congruent with the early and normal retirement ages). Unfortunately, information on mandatory retirement is not available in these data; if individuals are subject to mandatory retirement, it will increase the power of the first stage.

3 Empirical Framework

3.1 Hypotheses

This paper examines the relationship between retirement and SWB. Specifically, it considers the estimating equation:
$$ {\text{SWB}}_{\text{i}} = \beta_{1} {\text{R}}_{\text{i}} + \beta_{2} {\text{Z}}_{\text{i}} + \beta_{3} {\text{A}}_{\text{i}} + \beta_{4} {\text{W}}_{\text{i}} + \beta_{5} \sum {\text{C}}_{\text{c}} + {\text{v}}_{\text{i}} $$
where SWBi is the individual’s SWB, Ri is whether that individual is retired or not, Zi is a vector of observed characteristics (e.g.: marital status and gender), Ai is a vector of age polynomials, Wi is season of interview, and Ci is a set of country dummies. Studies have shown that SWB trends may be U-shaped by age (Stone et al. 2010), so it is very important to include a flexible functional form of Age. Thus, age is modeled flexibly with both a squared and cubed term in every model. Due to recent research indicating that retirement—when estimated exogenously—improves physical and mental health outcomes along some measures, it is hypothesized herein that SWB will be improved by retirement at least in the short-term. Specifically, the coefficient of Ri will be positive and significant. In addition, it is likely that OLS will be biased downward due to the relationship between SWBi and Ri.

3.2 Estimating Retirement Exogenously

As discussed at length by Gruber and Wise (1999), countries differ in both the age that retirement is supported and the level of support offered. Generally, individuals first have access to retirement benefits at the early retirement age. In many countries, individuals receive additional benefits if they continue working until the normal retirement age. Gruber and Wise discuss retirement incentives in terms of the implicit tax rate, which takes into account the replacement rate (which is the proportion of income that is replaced by social security in retirement) and any monetary incentives to continue working. Gruber and Wise find that individuals respond to these incentives, and they are more likely to retire when facing higher replacement rates.

An individual’s replacement rate is determined based upon a variety of factors that vary by country including former income, years in the workforce, and family size. However, despite wide variation in the absolute benefits levels for individuals even within-country, there is generally a large and discontinuous increase in the retirement incentives at the early and normal retirement ages (Gruber and Wise 2004). Although the retirement response is somewhat imperfect, the result is a large increase in the retirement hazard rate in the years surrounding the early and normal retirement ages. Gruber and Wise note: “common to all countries, the concentration of retirement [occurs at the] early and normal retirement ages” (1999, p. 10), which occurs with “empirical regularity” (2004, p. 17).

While the replacement rate itself is a somewhat endogenous predictor of retirement (as it is related to individual characteristics), everyone receives some increased incentive to retire at the early or normal retirement ages as the replacement rate increases discontinuously. Ideally, it would be possible to instrument retirement through both the retirement ages and the replacement rate/pension accrual by country, but enough detailed information on income and work history is rarely available. In the absence of such information, the retirement ages alone provide a strong first-stage estimate that has been successfully implemented by several researchers (Bonsang et al. 2012; Coe and Zamarro 2011; Neuman 2008; Rohwedder and Willis 2010).

3.3 Estimation Methodology and Execution

As mentioned above, retirement will be estimated through discontinuities in retirement incentives occurring at country-level early and normal retirement ages. As such, this methodology can be looked at as an application of the Fuzzy Regression Discontinuity (FRD) design. Specifically, the results will indicate whether there are discontinuous trends in SWB around the retirement ages while flexibly controlling for a wide-range of smooth age-effects. The relationship between retirement and SWB should be captured in part by nonlinear and non-monotonic changes in SWB that are off-trend in the years surrounding estimated retirement. However, unlike a traditional FRD, one distinct jump in SWB at the retirement age is not expected. Changes in SWB may manifest prior to or post transition, rather than exactly at retirement. In other words, the retirement-response and the psychological response to retirement are unlikely to occur exactly at the retirement ages. Thus, any change in SWB directly at the retirement ages will represent a lower bound for the overall local relationship between SWB and retirement.

First, the effect of reaching retirement age on the probability of being retired is estimated. Then, the relationship between the probability of retiring (given the retirement age) and SWB is modeled. This methodology allows for an exogenously predicted retirement variable, which should be orthogonal to individual characteristics that might impact both retirement and SWB (such as individual disposition). The fewer variables used solely in the first stage (exclusion criteria) and the more flexible the second stage, the stricter the model is as a whole. A stricter model will only capture effects exactly at the discontinuity in incentives, increasing the risk for Type II errors, while a less strict model will capture effects happening around the incentives, increasing the risk of misinterpreting nonlinearities as meaningful effects.

Consider the following model:
$$ \begin{aligned} Stage\;1&:{\hat{\text{R}}}_{\text{i}} ={\text{a}}_{1} {\text{Z}}_{\text{i}} + {\text{a}}_{2}{\text{A}}_{\text{i}} + {\text{a}}_{3} {\text{E}}_{\text{i}} +{\text{a}}_{4} {\text{N}}_{\text{i}} + {\text{a}}_{5}{\text{DE}}_{\text{i}} + {\text{a}}_{6} {\text{DE}}_{\text{i}}^{2}+ {\text{a}}_{7} {\text{DN}}_{\text{i}} + {\text{a}}_{8}{\text{DN}}_{\text{i}}^{2} + {\text{a}}_{9} \sum{\text{C}}_{\text{c}} + {\text{v}}_{\text{i}} \hfill \\ Stage\;2&:{\text{SWB}}_{\text{i}} = {\text{b}}_{1}{\text{Z}}_{\text{i}} + {\text{b}}_{2} {\text{A}}_{\text{i}} +{\text{b}}_{3} \overbrace{R}_{\text{i}} + {\text{b}}_{4} \sum{\text{C}}_{\text{c}} + {\text{v}}_{\text{i}} \hfill \\\end{aligned} $$
where Ei and Ni are dummy variables for whether or not the individual has reached the early or normal retirement age for his or her gender and country, DEi is the distance to and from the early retirement age, DNi is the distance to and from the normal retirement age, Ai is a vector of age polynomials, Wi is season of interview, Cc is a set of country fixed effects, and \( \overbrace{R}_{\text{I}} \) is the predicted retirement variable from Stage 1. This is similar to the model used by Rohwedder and Willis (2010) in response to Gruber and Wise’s (1999) findings that distance to the retirement ages and their squares are powerful predictors of retirement behavior. This model has good face validity, as the retirement and the psychological response are expected to occur in the years surrounding the retirement ages, not exactly at the early or normal ages. However, this model may capture some age-related psychological effects occurring in the years surrounding retirement. This model (IV1) will thus be considered only in addition to more rigorous but less realistic empirical models.
It is possible to reduce the chances of Type I error by omitting trends from the first stage. Consider IV2, which will be the preferred specification:
$$ \begin{aligned} Stage\;1&:{\text{R}}_{\text{i}} = {\text{c}}_{1} {\text{Z}}_{\text{i}} + {\text{c}}_{2} {\text{A}}_{\text{i}} + {\text{c}}_{3} {\text{W}}_{\text{i}} + {\text{c}}_{4} {\text{E}}_{\text{i}} + {\text{c}}_{5} {\text{N}}_{\text{i}} + {\text{c}}_{6} \sum {\text{C}}_{\text{c}} + {\text{v}}_{\text{i}} \hfill \\ Stage\;2&:{\text{SWB}}_{\text{i}} = {\text{d}}_{1} \overbrace{R}_{\text{i}} + {\text{d}}_{2} {\text{Z}}_{\text{i}} + {\text{d}}_{3} {\text{A}}_{\text{i}} + {\text{d}}_{4} {\text{W}}_{\text{i}} + {\text{d}}_{5} \sum {\text{C}}_{\text{c}} + {\text{v}}_{\text{i}} \hfill \\ \end{aligned} $$
This model flexibly accounts for age, but the retirement estimation will capture only retirement that occurs as a function of the early and normal retirement ages.
An even stricter model might control for atypical trends surrounding the retirement ages. For example:
$$ \begin{aligned} Stage\;1&:{\text{R}}_{\text{i}} = {\text{e}}_{1} {\text{Z}}_{\text{i}} + {\text{e}}_{2} {\text{A}}_{\text{i}} + {\text{e}}_{3} {\text{DEB}}_{\text{i}} + {\text{e}}_{4} {\text{DEP}}_{\text{i}} + {\text{e}}_{5} {\text{DNB}}_{\text{i}} + {\text{e}}_{6} {\text{DNP}}_{\text{i}} + {\text{e}}_{7} {\text{E}}_{\text{i}} + {\text{e}}_{8} {\text{N}}_{\text{i}} + {\text{e}}_{9} \sum {\text{C}}_{\text{c}} + {\text{v}}_{\text{i}} \hfill \\ Stage\;2&:{\text{SWB}}_{\text{i}} = {\text{f}}_{1} {\text{Z}}_{\text{i}} + {\text{f}}_{2} {\text{A}}_{\text{i}} + {\text{f}}_{3} {\text{DEB}}_{\text{i}} + {\text{f}}_{4} {\text{DEP}}_{\text{i}} + {\text{f}}_{5} {\text{DNB}}_{\text{i}} + {\text{f}}_{6} {\text{DNP}}_{\text{i}} + {\text{f}}_{7} \overbrace{R}_{\text{i}} + {\text{f}}_{8} \sum {\text{C}}_{\text{c}} + {\text{v}}_{\text{i}} \hfill \\ \end{aligned} $$
where DEBi and DEPi are the years until and years since the early retirement age respectively, and DNBi and DNPi are the years until and since the normal retirement age. This model will constrain the coefficient on retirement, as the additional controls explain away variation occurring in the years surrounding the retirement age. This specification (IV3) is the strictest model considered.

3.4 Instrument and Model Validity

There are two key criteria required for instruments to be valid: (1) The instruments must be related to the endogenous variable that they are being used to estimate; and (2) The instruments cannot affect the outcome variable except through the estimated variable. Proving the former requirement is rather straightforward. Previous research has shown a jump in retirement rates at the retirement ages (e.g. Gruber and Wise 2004). Figure 1 demonstrates this graphically, illustrating an acceleration in retirement around the retirement ages. Further, as will be shown, the F-stats for the exclusion criteria are promisingly high, and all are above the threshold for weak instruments (Stock and Yogo 2005). In sum, there is ample evidence that actual retirement is correlated with the formal retirement ages.
https://static-content.springer.com/image/art%3A10.1007%2Fs10902-012-9399-2/MediaObjects/10902_2012_9399_Fig1_HTML.gif
Fig. 1

Percent retired as a function of retirement ages. Area between dashed lines show region surrounding the normal retirement age. The dotted line is a linear trend provided for comparison. All 16 countries are included and weighted by population (2004 and 2006). Adapted from Charles (2004, Figures 1 and 2), which suggests that without incentives, retirement rate should be smooth. The dotted lines are for comparison, to show deviations from trend

Proving the requirement that the instruments do not affect SWB except through the mechanism of retirement is a bit more complicated, and requires intuition (Angrist and Pischke 2009, p. 109). Eligibility for retirement incentives is a good instrument because there is no reason why there should be discontinuous age-effects on SWB exactly at the retirement ages for each country. In other words, even if country-level retirement policies are endogenous to social-climates, there is no reasonable reason why SWB should change at the retirement ages unless there is some other consistent event that occurs at the same time. The United States is the only country in the sample with something like Medicare, which is a confounder (Card et al. 2009). However, as will be shown, excluding the United States does not substantially change the results. Thus, this instrument is a valid tool for estimating retirement.

Further, as with any IV model, these results must be interpreted as a Local Average Treatment Effect (LATE); the coefficients represent the average change in SWB for those who choose to retire at the early and normal retirement ages. The instrumental variable technique is effective at exploring those at the margin, who are prompted into action by the exclusion criteria, but it cannot be used to estimate those who do not respond to the bump in incentives. Thus, the current results should be interpreted as the effect of retirement on those who were willing to retire as a function of their benefits at the early and normal retirement ages.

Despite the shortcomings of IVs, to borrow a phrase from Imbens: “Better LATE than nothing” (Imbens 2009). As Imbens argues, IVs are in many ways extremely powerful; a well-executed model behaves like random assignment. As such, omitted demographic variables are not a source of bias in an IV model, as long as there is no relationship between these variables and the exogenously measured explanatory variable. Although there is a long list of variables that could be examined for their relationship to retirement and SWB (such as individual wealth, health, the level of pension benefits/implicit tax rate that the individual faces, how much the individual enjoyed pre-retirement work, whether the individual’s spouse has retired, whether the individual lives close to family or friends), the omission of these variables is not a source of bias because of the IV methodology.

In the current study, individual-level heterogeneous treatment effects are not explored. Other variables, particularly replacement rate and health, could provide other interesting and policy relevant insights. However, previous research suggests that finances and health may have complicated feedback loops with retirement. Accordingly, controlling for income is clearly not an option with cross-sectional data on retirement since those who retire will generally have lower income than those who do not. In addition, pre-retirement income is not available in these data. Standardizing wealth across countries would be problematic as well, since countries differ in norms around how and how much individuals save (e.g. whether they invest in stocks, property, or jewelry, or expect to move in with their children). Further, while wealth could theoretically be standardized similarly to SWB, it may capture different things in different countries (e.g.: enduring social class vs career success). As for individual health, there is evidence that it too has a dynamic rather than a static relationship with retirement; Bound and Waidmann (2007) found that retirement improves subjective and objective health. Thus, these explorations are better suited for within country analysis using longitudinal data and should be the subject of future research.

3.5 Other Specification Notes

The regressions are clustered by country, which further increases standard errors. With sixteen countries, the number of clusters is quite small, which has the effect of attenuating findings (Bertrand et al. 2004). All models and figures are weighted by country population. Country fixed-effects are included in all models. Samples are limited to people 50–70 years of age unless otherwise noted. This is 5 years prior to the earliest retirement age and 5 years post the latest retirement age, ranging from 55 to 65 years (see Table 1).

4 Results

4.1 SWB and Retirement

Tables 4 and 5 provide the results for life satisfaction and CASP, respectively. In nearly every case, there is a positive, statistically significant relationship between retirement and SWB when the model is estimated through an IV technique. Each model is displayed side-by-side with a comparable OLS model, which is not statistically significant in any case. The second model, IV2, is the preferred model, as it flexibly models age and has only minimal exclusion criteria. SWB improves by 0.4 and 0.9 of a standard deviation for life satisfaction and CASP using the preferred specification.
Table 4

Retirement improves life satisfaction

 

IV1

OLS

IV2

OLS

IV3

OLS

Retired

0.117

[0.121]

−0.00061

[0.0502]

0.441*

[0.245]

−0.00061

[0.0502]

0.458**

[0.229]

0.0198

[0.0554]

Married

0.410***

[0.0601]

0.411***

[0.0625]

0.408***

[0.0593]

0.411***

[0.0625]

0.408***

[0.0589]

0.219***

[0.0510]

Other controls

Age, Age2, Age3, season of interview

Age, Age2, Age3, season of interview

Age, Age2, Age3, season of interview, EarlyDisB, NormalDisB, EarlyDisP, and NormalDisP

Excluded instruments

EarlyIV, NormalIV, EarlyDis, NormalDis, EarlyDis2, and NormalDis2

EaryIV, NormalIV

EarlyIV, NormalIV

Inst. F-Stat

623.80

N/A

8.534

N/A

8.027

N/A

Hansen’s J

5.487

N/A

0.0913

N/A

0.233

N/A

Over-ID test

Passes

Passes

Passes

N

13,326

13,326

13,326

13,326

13,326

13,326

EarlyIV/(NormalIV): dummy for having reached early (full) retirement age

EarlyDis/NormalDis: [age − early (full) retirement age]

EarlyDisB/(NormalDisB): [age − early (full) retirement age] if younger than cutoff

EarlyDisP/(NormalDisP): [age − early (full) retirement age] if older than cutoff

*p < .05, **p < .01, ***p < .001

Table 5

Retirement improves subjective well-being as measured by CASP

 

IV1

OLS

IV2

OLS

IV3

OLS

Retired

0.657***

[0.208]

−0.000605

[0.0502]

0.869***

[0.268]

0.0191

[0.0567]

0.873***

[0.288]

0.0198

[0.0554]

Married

0.212***

[0.0484]

0.411***

[0.0625]

0.211***

[0.0477]

0.219***

[0.0510]

0.408***

[0.0589]

0.219***

[0.0510]

Other controls

Age, Age2, Age3, season of interview

Age, Age2, Age3, season of interview

Age, Age2, Age3, season of interview, EarlyDisB, NormalDisB, EarlyDisP, and NormalDisP

Excluded instruments

EarlyIV, NormalIV, EarlyDis, NormalDis, EarlyDis2, and NormalDis2

EaryIV, NormalIV

EarlyIV, NormalIV

Inst. F-Stat

603.3

N/A

8.532

N/A

8.103

N/A

Hansen’s J

0.992

N/A

0.744

N/A

0.760

N/A

Over-ID test

Passes

 

Passes

 

Passes

 

N

13,038

13,038

13,038

13,038

13,038

13,038

EarlyIV/(NormalIV): dummy for having reached early (full) retirement age

EarlyDis/NormalDis: [age − early (full) retirement age]

EarlyDisB/(NormalDisB): [age − early (full) retirement age] if younger than cutoff

EarlyDisP/(NormalDisP): [age − early (full) retirement age] if older than cutoff

*p < .05, **p < .01, ***p < .001

Because these models are clustered by country, there are few degrees of freedom and low significance still connotes a large effect size. The Cohen’s D effect size on the preferred models for life satisfaction and CASP are 1.65 and 0.93 respectively, which is generally considered to be a large effect size (Cohen 1969). However, the question is whether this is large enough to be policy relevant. Perhaps a reasonable metric for calibration would be the coefficient on marriage. Many previous studies have found that married people are much happier than unmarried people (Blanchflower and Oswald 2011; Grove et al. 1983; Stack and Eshleman 1998; Stone et al. 2010). In the current sample, the coefficient on retirement ranges from about equivalent to quite a bit larger than the coefficient for married, depending on the specification. This suggests that the increase in SWB seen in the current study is not only significant, but substantial.

The results can be modeled graphically as well. However, while this provides an illustrative and intuitive presentation of the findings, it is an understatement of the effect as the graph can only model retirement as a function of the normal retirement age instead of both the early and normal retirement ages. This is because early and normal retirement ages differ in their distance from one another, and centering on, for example, the normal retirement age means that the early retirement age will be occurring at different times for each country. In Fig. 2, life satisfaction and CASP are modeled as a function of Age, Age2, Age3, gender, and marital status and then modeled quadratically prior to and after the normal retirement age. Large improvements to SWB can be seen at the retirement age, followed by a steep decline—even with a vector of age controls.
https://static-content.springer.com/image/art%3A10.1007%2Fs10902-012-9399-2/MediaObjects/10902_2012_9399_Fig2_HTML.gif
Fig. 2

a Predicted life satisfaction by distance to/from normal retirement age. b Predicted CASP by distance to/from normal retirement age. Yearly predicted values overlaid. Model controls for age, age2, age3, gender, marital status, and season of interview. Results are weighted by sample size relative to country population

4.2 Between-Country Heterogeneity and the Post-retirement Decline

In order to understand the size and variation in the trend, coefficients for individual countries are modeled in Fig. 3. The figure depicts the within country differences in CASP occurring around the retirement ages. This analysis focuses on CASP because there is less unexplained variation in the CASP measures than in the life satisfaction measures (see Fig. 2). Clearly, the relationship between CASP retirement is much stronger in some countries than in others. Some previous research has shown that retirement is better for people who retire later (e.g. Bound and Waidmann 2007; Coe and Zamarro 2011). Initial analysis suggests that the results in the current study are congruent with this; as can be seen in Table 6, men who face formal retirement at age 65 or older are made happier by retirement than men who face younger formal retirement ages.
https://static-content.springer.com/image/art%3A10.1007%2Fs10902-012-9399-2/MediaObjects/10902_2012_9399_Fig3_HTML.gif
Fig. 3

Coefficients fo individual countries. This table shows the coefficient on “retirement” for SWB when retirement is measured using a FRD method around the early and normal retirement age(s). This is implemented as IV2. Results are weighted by sample size relative to country population

Table 6

Younger versus older onset of retirement incentives

Coefficient: IV-retirement

Life satisfaction

CASP

Excluded instrument F-Stat*

Younger benefit onset

0.277a

[0.153]

0.552***

[0.123]

22.88

Older benefit onset

2.594***

[0.768]

1.699***

[0.401]

26.78

Results are weighted by sample size relative to country population. This is the preferred model (IV2). Younger formal retirement is defined as having an option to retire with benefits before age 65. Older formal retirement is defined as having no option to retire with benefits until age 65 or later

aThe excluded instrument F-Stats reported are for CASP, which is generally almost identical as for life satisfaction. The results only differ slightly because there are a small number of people who did not answer the a CASP question or the life satisfaction question

*p < .05, **p < .01, ***p < .001

Figure 4 compares happiness by age for men who face younger versus older incentives to retire. As can be seen from the figure, reported happiness is quite similar for individuals approaching age 60; this is consistent with the notion that retirement incentives are driving the differences in outcomes. However, during the early 60 s, men who face earlier retirement appear to be happier. There is a bump in SWB following the formal retirement age followed by a steady decline. When men facing later incentives can retire with benefits at age 65, the later incentives group catches up and surpasses the happiness experienced by those in the earlier incentives group for a few years. However, the bump in SWB appears to last only a few years and by the time men reach age 70, both groups return to similar levels of happiness.
https://static-content.springer.com/image/art%3A10.1007%2Fs10902-012-9399-2/MediaObjects/10902_2012_9399_Fig4_HTML.gif
Fig. 4

Comparison of happiness by age of retirement incentives. Controls for marital status, age, age2, age3, and season of interview are included in the model. Younger formal retirement is defined as having an option to retire with benefits before age 65. Older formal retirement is defined as having no option to retire with benefits until age 65 or later. Over this portion of the life cycle, there is little difference in the total SWB experienced by people as a function of the onset of their retirement incentives

The total differences in happiness between the ages of 55 and 74 for the younger and older incentives groups can be calculated as the difference in the area under the “younger incentives” curve and the area under the “older incentives” curve in Fig. 4. However, for this to be indicative of total expected happiness, this value should be adjusted to reflect mortality during this period. To do this, age-adjusted mortality rates (weighted by country population) as reported by the World Health Organization for 2004 (WHO 2012)3 are used. With age 55 as a baseline, every year thereafter is discounted by the proportion of the population that would have likely died since age 55. The total difference between the areas under the SWB curves associated with earlier and later incentives is found to be 3.3, which is minimal given that the difference in SWB between the two groups in the average year is 31.9.

Previous studies utilizing a quasi-experimental approach have suggested that those retirement is better for those who retire later. However, once considering the wellbeing of people a few years away from retirement, this is not supported by these data. Instead, compared to men facing later retirement incentives, men facing earlier retirement incentives are equally happy in their late 50 s, happier in their early 60 s when they retire, less happy in their late 60 s, and then about equally happy in their 70 s. As such, the age of retirement incentives is about neutral to total SWB over this part of the life cycle.

4.3 Robustness Checks and Falsification Tests

As a validation tool, a few robustness checks and falsification tests are presented in Table 7. First, the models are estimated excluding possible outliers. Specifically, potential problem countries include the USA and Ireland. As discussed above, the USA has a discreet jump in Medicare eligibility right around the normal retirement age. Previous research has shown a variety of positive outcome for health insurance in general (Finkelstein et al. 2011) and Medicare in particular (Card et al. 2009) in US samples. In addition, as shown in Fig. 3, Ireland is an outlier with a particularly large treatment effect. When the models are estimated excluding these countries, the results remain significant and positive.
Table 7

Robustness check and falsification test

Coefficient: IV-retirement

Without Ireland

Without USA

Retirement ages −10

Retirement ages +10

Life satisfaction

0.186a

[0.103]

0.225a

[0.127]

0.545

[0.570]

1.125

[1.210]

CASP

0.206**

[0.102]

0.267**

[0.121]

−0.0398

[0.558]

0.528

[0.679]

Excluded instrument F-Stat*

36.73

33.79

2.14

1.92

The table above shows the results from eight separate regressions, all relying on the preferred specification (IV2). The first two columns perform a robustness check by excluding Ireland and the USA. The results are still significant, indicating that the findings are not driven by these outliers. The second two columns perform a falsification test. These models utilize fake early and normal retirement ages that are 10 years prior or 10 years post the real ages. The sample is people 40–60 and 60–80, respectively. Sample weights are adjusted accordingly

aThe excluded instrument F-Stats reported are for CASP, which is generally almost identical as for life satisfaction. The results only differ slightly because there are a small number of people who did not answer the a CASP question or the life satisfaction question

*p < .05, **p < .01, ***p < .001

At first thought, an ideal falsification test might be some health biomarker, such as “hand grip” or “lung capacity,” known to decline smoothly with age around the early and normal retirement ages. However, some researchers have found improvements to individual’s overall health as a result of retirement (Bound and Waidmann 2007; Coe and Zamarro 2011). Thus, in the current study, health biomarkers are not used as a counterfactual because while they may be excluded from the current models of psychological adaptation to retirement, they may be part of a dynamic response to retirement. Instead, the models are estimated with two fake “retirement ages”. The first uses a retirement age that is 10 years prior to the actual retirement age for each individual’s country and gender, and the second uses a retirement age that is 10 years post the actual retirement age for each individual’s country and gender. Age bandwidths and weights are altered accordingly. If a statistically significant result was found on these falsely estimate retirement coefficients, it would be reason for concern. However, the results for these falsification tests are reassuringly null.

5 Conclusions and Policy Recommendations

This research project has three primary findings: (1) In the time surrounding retirement, people experience a large improvement in their SWB as measured by CASP and life satisfaction; (2) A few years after retirement, SWB declines rapidly; and (3) Later versus earlier retirement incentives is overall neutral in terms of SWB. The sharp decline in the years after retirement appears to be caused by retirement itself. This is consistent with Atchley’s (1976) suggestion of a multi-stage adjustment to retirement which starts with a honeymoon, followed by a steep decline, and then followed by a final stable period.

From these data alone, it is not possible to fully identify the reason for the decline in SWB in the years after retirement because IV methods are most reliable directly around the treatment. Recall that Bound and Waidmann (2007) also find that retirement improves health in a UK sample, but that the effect dwindles rapidly when using a similar identification strategy, and that several studies have found exogenous measures of retirement to be associated with cognitive decline (Bonsang et al. 2012; Coe and Zamarro 2011). Within this context, it is reasonable to envision that individuals retire and enjoy a brief celebration that is quickly depleted by the ruins of inactivity, including cognitive and physical decline.

Further, as the Local Average Treatment Effect for this identification strategy focuses on people who retire as a function of public pensions right when they become available, it represents the effect of retirement on people who are fairly eager to retire. Thus the positive effects of retirement may serve as an upward bound for the positive effects of retirement. Individuals who derive a lot of pleasure from their work may be less likely to comply with these incentives when they are first available. These people who particularly enjoy work may delay retirement for a few more years.

This paper provides a first comparative look into the role of retirement on SWB in an international context. The current results suggest that a later formal retirement simply delays the SWB benefits of retirement, and that age of formal retirement is relatively neutral with regard to overall SWB, even when adjusting for mortality. Given the growing fiscal pressures to adjust the age of retirement upwards, it can be inferred from this study that welfare as measured by SWB may be on balance affected only marginally (if at all) by such changes. Thus, if it is necessary to increase the retirement age a few years to increase financial stability, policymakers need not worry that they are making people psychologically worse off in the long-run.

Footnotes
1

SHARE contains data on: Austria, Belgium, the Czech Republic, Denmark, France, Germany, Greece, Ireland, Italy, the Netherlands, Poland, Spain, Sweden, and Switzerland. ELSA contains data on the UK, and HRS contains data on the USA.

 
2

The SWBic variables are standardized prior to limiting the bandwidth of the ages in order to base SWB indicators on overall country norms.

 
3

Age-adjusted mortality rate data was used for the year of the retirement survey; the US data is from 2002 and the majority of the European data is from 2004. However, a few countries did not have data as recent as 2004. For the following countries, the most recent year available was used: Belgium (1997), Denmark (2001), and Italy (2003).

 

Acknowledgments

I would like to thank Ronald Lee, Steven Raphael, Jack Glaser, and Rucker Johnson, for their feedback and advice on this project. This research was supported in part by the National Institute on Aging Training Grant (T32-AG000246).

Copyright information

© Springer Science+Business Media Dordrecht 2012