The Association Between Individual Income and Remaining Life Expectancy at the Age of 65 in the Netherlands

Abstract

Introduction

Significant socioeconomic inequalities in mortality risk over many populations and time periods have been identified in the literature (Attanasio and Hoynes 2000; Hurd et al. 1999; Huisman et al. 2004; Kunst et al. 2004; Marmot et al. 1991; Menchik 1993; Palme and Sandgren 2008; Sullivan and Von Wachter 2009). These socioeconomic inequalities in mortality, commonly termed “differential mortality,” give strong evidence for an inverse relationship between income and mortality risk. Estimates of this relationship for different populations with respect to country and age range indicate that the ratio of mortality risk for individuals in the lowest quartile of the income distribution to that of individuals in the highest quartile ranges from around 2 in Europe to 3 in the United States.¹

At the same time, there is ongoing debate from various disciplines on the causal interpretation of this relationship and the possible pathways through which socioeconomic position affects health and mortality risk (Lindahl 2005; Macintyre 1997; Marmot et al. 1991; Smith 1999; Snyder and Evans 2006). Although there is some evidence that a high social economic status is associated with earlier diagnosis and higher survival rates for cancer (Aarts et al. 2010), most literature appears to dismiss the explanation that differential mortality is a result of low-income individuals having access to less or lower-quality health services (Attanasio and Emmerson 2003; Smith 1999). More widely accepted is the cultural or behavioral explanation that low-income individuals have an unhealthy lifestyle; for example, they are more often smokers, consume more alcohol, and have diets that are linked to obesity (Huisman et al. 2005; Macintyre 1997). The influential Whitehall studies, however, have shown that behavioral risk is not the sole explanation for differential mortality (Marmot et al. 1991). This latter finding has sparked research that emphasizes the long-term impacts on health of childhood socioeconomic circumstances and health conditions (Barker 1995, 1997; Case et al. 2002, 2005; Flores and Kalwij 2011; Van den Berg et al. 2006), cognitive abilities (Huisman and Mackenbach 2007), and psychosocial factors such as prolonged exposure to stress (Macintyre 1997; Smith 1999).

Independent of the causes of differential mortality, an inverse relationship between income and mortality risk has important implications for pension policy (Whitehouse and Zaidi 2008).² Public pension policy in many countries, including the Netherlands, aims at redistributing income from financially better-off to financially worse-off individuals. This redistribution may be adversely affected by differential mortality because, on average, low-income individuals receive public pension benefits for a relatively shorter period if the statutory retirement age is the same for all individuals (Nelissen 1999). Likewise, for private pensions, an inverse relationship between income and mortality risk implies that compared with high-income individuals, low-income individuals’ internal rate of return from a uniformly priced private pension plan is, on average, lower because their lower life expectancy results in receiving the benefits for a relatively shorter period (Bonenkamp 2007; Brown 2000; Hári 2007; Menchik 1993; Simonovits 2006). These policy concerns are exemplified by proposed pension reforms in the Netherlands that explicitly take into account this disparity by making it possible for workers in low-income sectors to receive a pension no more than two years before the proposed increased statutory retirement age of 67 (Stichting van de Arbeid 2010).

To gain insights into the size of the abovementioned differences in life expectancy between low- and high-income individuals, we estimate the association between individual income and mortality risk after the statutory retirement age (65) in the Netherlands. Our article’s contribution to the empirical literature on differential mortality is threefold. First, in contrast to previous studies for the Netherlands that most often examined the association between education and mortality risk (van Kippersluis et al. 2011; Kunst and Mackenbach 1994), we examine the association between income and mortality risk. We also quantify the association between lifetime income and remaining life expectancy at age 65, using Monte Carlo simulations. We choose age 65 because it is the statutory retirement age in the Netherlands after which all individuals receive (1) a public pension benefit that depends only on the years of recorded residency in the Netherlands between the ages of 15 and 65 (2 % of the full public pension benefit for each year) and (2) an occupational pension that depends on earnings history (Nelissen 1999). Because this retirement income consists primarily of pension income—and is therefore closely related to individual earnings history—it serves as a good proxy for an individual’s lifetime income.

Second, in contrast to most previous studies that, depending on data availability, have used either individual income or household income (i.e., the individual’s and spouse’s income combined), we distinguish between individual income and spousal income. This distinction allows us to examine the theoretical explanations provided in the literature for an association between spouse’s socioeconomic position and individual mortality risk. One explanation is that material hardship, measured by household and not individual income, matters for health status and mortality risk (Martikainen et al. 2001). Indeed, for most women from the cohorts in our analysis, mortality risk is perhaps more likely to be negatively related to their spouse’s than their own income because many left the labor force at the time of marriage or the birth of a first child. An explanation provided by Grossman’s (1972) health stock (economic) model is that a spouse may improve the efficiency of an individual’s investment in health stock, and the higher the partner’s socioeconomic position, the greater the improvement and the lower the mortality risk.³ A sociological explanation is that households share a lifestyle that is influenced by both partners (Torssander and Erikson 2009). Although our empirical model cannot discriminate between these different explanations, it can test, for both men and women, the claim that individual mortality risk is (negatively) associated with spousal as well as individual income. If we find empirical evidence for this claim, we could interpret it as support for one of the preceding explanations. At present, however, the empirical evidence on this issue is limited, with the exception of McDonough et al. (1999) for the United States and Torssander and Erikson (2009) for Sweden, who reported that for women (but not for men), spouse’s income is negatively associated with individual mortality risk.

Third, because in any analysis of mortality risk, the sample inherently becomes more selective with age in terms of both observed and unobserved characteristics, our empirical model also controls for unobserved individual-specific characteristics (i.e., random effects). Failing to control for this so-called dynamic selection may bias the results (Cameron and Heckman 1998; Van den Berg 2001). Still, most of the studies cited to this point have used (pooled) cross-sectional data and do not control for dynamic selection. Such control is possible in studies that use duration data if they include random effects. However, Hupfeld’s (2009) model, although based on duration data, included no random effects, and Van den Berg et al. (2006) claimed that in their estimations, including random effects had no effect on the impact of economic conditions early in life on individual mortality risk. We have panel data and can therefore account for dynamic selection by including random effects in our empirical model. A model complication does arise, however, for individuals who enter the panel after age 65. That is, given the dynamic selection process, random effects at age 65 imply a dependency between these random effects and the covariates at later ages. To take this complication into account, we include dynamic sample selection correction terms that control for this dependency at age of entry for individuals who enter the sample after age 65.

The article is organized as follows. The next section describes the data. The third section outlines the empirical model for analyzing mortality risk and explains the estimation procedure. The fourth section reports the analytical results, and the fifth section summarizes the main findings and concludes the article.

Data

The data are taken from the 1996–2007 Income Panel Study of the Netherlands (IPO, Inkomens Panel Onderzoek, CBS 2009a) and the 1997–2008 Causes of Death registry (DO, Doodsoorzaken, CBS 2009b), both gathered by Statistics Netherlands. The IPO, a representative sample of the Dutch population, consists of an administrative panel data set of about 92,000 individuals, randomly selected in 1996, which increased to about 99,000 individuals in 2007 because of population growth. Sampling is based on individuals’ national security numbers, and the selected individuals are followed for as long as they are residing in the Netherlands on December 31 of the sample year. The data set also includes individuals living in institutions for the elderly, such as nursing homes. Individuals born in the Netherlands enter the panel for the first time in the year of their birth; immigrants to the Netherlands, in the year of their arrival. An individual exits the panel on death or emigration from the Netherlands. Hence, the only reasons for panel attrition are mortality and emigration.

IPO contains data on demographic characteristics and income of each member of a selected individual’s household. These data are obtained from official institutions—most particularly, the population registry and tax office. The DO, on the other hand, provides information on date and cause of death for all residents deceased during the 1997–2008 period. These data come from medical records provided by medical examiners, who are legally obliged to submit them to Statistics Netherlands. The DO data set also assigns a personal identifier that allows determination of whether an individual in the IPO has died by the next calendar year.

We select individuals aged 65 or older, who, because of population aging, make up 12.8 % of the sample in 1996 and 13.9 % in 2007. This raw data set consists of 151,120 observations for 21,159 individuals over the 1996–2007 period. We remove about 5 % of the observations because of missing income information, about 1 % because of missing values on marital status, and 0.04 % because of negative income. These exclusions affect relatively more men than women (11 % vs. 4 %). Nonetheless, after we control for gender, the hypothesis that the mortality rate among the individuals excluded is equal to the mortality rate among those included is not rejected (a p value of .261), which suggests that these exclusions do not yield an endogenous sample selection. Panel attrition for reasons other than mortality (i.e., emigration or missing next year’s information) is about 0.3 % per year. The resulting sample consists of 19,258 individuals: 11,601 females and 7,657 males. Of these, 58.5 % entered the panel in 1996 at any age, and the remaining 41.5 % entered the panel after 1996 upon reaching age 65.⁴ The panel contains 141,725 observations and is a representative sample of the Dutch population age 65 and older over the 1996–2007 period.

Variable Definitions and Descriptive Statistics

The analysis is based on the variables of gender, age, marital status, and income. We define age as the individual’s age on January 1 of each year because in the Netherlands, the calendar year is also the fiscal year for income measurement, meaning that this choice ensures that income at age 65 is measured over the first entire calendar year of retirement. Table 1 reports the number of observations and marital status by age and gender. The marital status variable distinguishes between a single adult household that includes divorcees (hereafter, “single”), a married or cohabiting couple (“married”), and a widowed individual. The differences in marital status across age and gender result from the recognized fact that, on average, women live longer than men. These differences result in, for instance, an increasing proportion of women with age (last column, Panel A) and, at a given age, relatively more widowed and fewer married women than men (Panel B).

The IPO income data are based primarily on tax records and contain detailed and accurate information on all income components. Here, income is gross of income tax and social insurance contributions, and is measured in 2005 euros using the consumer price index (CPI). Individual income is the sum of pension, labor, transfer, and capital income. Kalwij et al. (2009), in their overview of these components and their definitions, showed that for Dutch men and women, on average, about 90 % of retirement income is pension income. All income components are observed for the individual and, in couple households, also for the spouse. The analysis excludes any income from other household members.⁵

As shown in Table 2,⁶ the mean income of single and widowed men is higher than that of single and widowed women, and the distribution of income for single and widowed men is wider than the distribution of income for single and widowed women. The income distribution for men also shows a decrease in median income with age, most probably because of changes in the income distribution over birth cohorts (Knoef et al. 2009). In addition, the income distribution for married women is rather compressed, not only because many retired married women have no earnings history and receive only public pension benefits, but also because pre-1990, part-time work often came with no pension plan or a pension plan that had a relatively high threshold before contributions could be made.⁷

A comparison of the tables for married Dutch men and women reveals that women’s income accounts for, on average, one-third of household income. The rank correlation between the individual’s and the spouse’s income is about .10 (not shown here).

Differential Mortality

We define mortality as being deceased before January 1 of next year. About 38 % of the individuals died over the sample period. The results in Table 3 confirm the accepted patterns that mortality risk increases with age and that men have a higher mortality risk than women. Likewise, age-specific mortality risk is lower among married individuals than among single or widowed individuals. In addition, the sample statistics on mortality risk by gender and age compare favorably with the population statistics from the Human Mortality Database (HMD column); that is, the differences between the two by age and gender are small.

In Table 4, Panel A, we pool the results for single and widowed individuals because the mortality risk by income quartile differs little between these two groups. As Panel A shows, mortality risk decreases as income increases for single/widowed men and women, an effect that is strongest up to the third quartile. One measure of differential mortality is the ratio of mortality risk among individuals in the first quartile of the income distribution to mortality risk among individuals in the fourth quartile (columns Q1 / Q4). A comparison of Panels A and B reveals that differential mortality is stronger for married men than for single/widowed men (2.4 vs. 1.7) but weaker for married women than for single/widowed women (1.4 vs. 1.9). These statistics are in line with findings of other European studies but lower than those reported for the United States (see the Introduction). Overall, differential mortality appears to decline with age. Interestingly however, as illustrated in Panel C, when spouse’s income is considered, a differential mortality pattern emerges for both men and women: one that, although weaker for men, is similar to that produced for individual income. We are not aware of comparable statistics in the literature.

Mortality Risk Model

As previously explained, we define mortality based on whether an individual is deceased in the subsequent year. In our mortality risk analysis, we consider the following latent variable model that relates next year’s mortality status, at age (a + 1), to individual characteristics at age a:

(1)

In the context of this article, the latent variable H_a+1 is an individual’s stock of health, which indicates that the individual is deceased if it falls in the next year below a certain threshold, normalized to 0 in Eq. (1). The variable M_a+1 denotes observed mortality at age (a + 1), and M_a+1 is equal to 1 if an individual became a years old and died at age (a + 1), and 0 otherwise. α_a is an age-specific intercept, while X_a is a (1 × k) vector of an individual’s observed characteristics at age a, including marital status and income, with a corresponding (k × 1) parameter vector β. Λ denotes an individual’s unobserved characteristics (i.e., a random effect) and is assumed to be constant over time and independent of the covariates at age 65. This assumption does not, however, exclude dependency between the covariates and Λ at later ages. We also assume that the random effect is normally distributed with a 0 mean and variance and that the error term ε_a follows a logistic distribution and is independently distributed across individuals and time with a 0 mean and a variance normalized to π²/3.

Dynamic Selection

In any study of mortality risk over the life cycle, the population at risk inherently changes with age. Hence, our model, by explicitly accounting for random effects, allows for sample selection by age on the basis of both observed and unobserved characteristics. Not accounting for such dynamic selection might yield inconsistent estimates of α_a and β (Cameron and Heckman 1998; Van den Berg 2001). If mortality risk is negatively related to income and positively to the random effect Λ, then low-income (high-income) individuals with a low (high) Λ value are more likely to survive another year than low-income individuals with a high Λ value. Hence, in this example, dynamic selection results in a population at risk in which the correlation between Λ and income becomes increasingly positive with age up to a certain age and then decreases as the sample becomes more homogenous with respect to income and Λ.

When all individuals are observed from the age of 65, the model takes dynamic selection into account. However, 58 % of the individuals enter our sample after age 65, implying that, according to the dynamic selection principle, these individuals, having survived from age 65 to the age of panel entry (τ), are a selective sample (of their cohort) in terms of their covariates (including income) and Λ. A post-65 entry thus produces the empirical complication of a dependency between the random effect and the covariates: a dependency that, as already explained, changes with the age of entry. Ideally, this dependency could be controlled for by explicitly modeling the probability of survival up to the age of entry (Ridder 1984); however, doing so would require data on the covariates from the age of 65 up to the age one year before entry, which we do not have. As an alternative, therefore, we explicitly account for the change in dependency between random effects and income at the age of entry, thereby maintaining the random effects assumption at age 65. Essentially, using Eq. (2), we parameterize the change in dependency between the covariates and the random effect Λ at the age of first observation τ for individuals who enter the sample after the statutory retirement age of 65:

(2)

where , with a corresponding ((2 k + 2) × 1) parameter vector γ. Scaling the effects of the covariates at age τ with the factors (τ − 65) and (τ − 65)² takes into account, for instance, that the dependency between the random effect and income becomes more positive with age and at some age decreases as the sample becomes more homogeneous. θ is a random effect that is assumed to be independent of and normally distributed with a 0 mean and σ² variance. Consistent estimates of the α_as and β are obtained under the additional assumption formalized in Eq. (2). Testing the joint hypothesis σ² = 0 ∩ γ = 0 tests for the presence of random effects and the implied need to include dynamic sample selection correction terms (i.e., the additional covariates ).

Estimation, Identification, and Empirical Specification

Given the model outlined earlier, age-specific mortality risk conditional on observed and unobserved characteristics can be written as follows:

(3)

The condition M_a = 0 formalizes the fact that all individuals in the population at risk are alive at age a, and F(.) is the logistic cumulative distribution function. Equation (3) is the basis for a likelihood function in which we integrate out the unobserved individual specific effect θ. With i denoting the individual, τ(i) is the age of the individual when first observed in the sample, and A(i) is the age of the individual when last observed in the sample. The variable m(i) is equal to 1 if the individual is deceased by age A(i) + 1, and 0 otherwise. Maximum likelihood estimates of the model parameters are given by

(4)

where α = (α₆₅,…,α_T), with T equal to the maximum age an individual may reach. N is the number of individuals, and Φ is the cumulative normal distribution function. This type of model is often referred to as a random effects panel data logit model (Wooldridge 2001). Equation (4) imposes proportionality among the age pattern, the covariates, and the random effect to ensure identification of the random effects distribution (Cameron and Trivedi 2005). For individuals who enter the sample at age 65, no dynamic sample selection correction terms are included (because for this group, τ = 65; and hence, is a vector with 0s), and identification of the γ parameters is established solely based on individuals who enter the sample after age 65. We estimate the model separately for men and women.

To estimate the associations between individual and spousal income and mortality risk, we parameterize Eq. (1) as follows:

(5)

The associations between mortality risk and individual income (⁠⁠) and spousal income (⁠⁠) are given by the parameters β₃ – β₇, and spousal income is equal to 0 for a single or widowed individual. The main advantage of this specification is that it nests two empirical specifications used in the literature and discussed in the Introduction. In the first nested model, which refers to the hypothesis β₄ = β₆ = β₇ = 0, only the individual (not the spousal) income is associated with mortality risk. In the second, which refers to the hypothesis β₃ = β₄ ∩ β₅ = β₆ ∩ β₇ = 2β₅, household income rather than individual and/or spousal income is associated with mortality risk.

Married is a dummy variable equal to 1 if the individual is married (including cohabitation), and 0 otherwise; Widow is a dummy variable equal to 1 if the individual is widowed, and 0 otherwise. The reference category for marital status is a single adult household. Besides controlling for time effects, as briefly explained in the next section, we also test for age-specific intercepts instead of a linear age function.

Empirical Results

Estimation Results

Before discussing the estimation results for the mortality risk model, we need to clarify two points. First, statistical tests show that for both men and women, the age dependency of mortality risk can be restricted to a linear function of age (in the index).⁸ Second, interpreting the estimated associations between marital status and mortality risk requires that we take into account spousal income, which is at least equal to the public pension benefit, so we address this issue separately in a subsequent section.

The results of our estimations are reported in Table 5 and discussed here for a 5 % level of significance. As the table shows, individual income is negatively associated with mortality risk for both men and women, and the parameter estimates are roughly of the same magnitude. Based on these estimates, we calculate that the negative association between income and mortality risk diminishes after about €130,000 (which is around the top 0.2 percentile of the income distribution). Moreover, although the association between mortality risk and spouse’s income is insignificant for men and only significant at a 10 % level for women (see the second from last row of Table 5), it is of roughly the same magnitude for both. In addition, the test statistics for men and women indicate that we do not reject the hypothesis that it is household, rather than individual and/or spousal income, that is associated with mortality risk (see the last row of Table 5).

As shown at the bottom of Table 5, for men, we find a significant presence of random effects and the required dynamic sample selection correction terms; but for women, these are significant at only the 10 % level. For both men and women, the standard deviations of the random effect are, although individually significant, relatively small compared with the standard deviation of the error term in Eq. (1), which is equal to . In addition, Tables 7 (Model 5) and 8 (Model 10) in the appendix report the estimation results for the model that excludes random effects and the dynamic sample selection correction terms. We do not discuss these here in detail, but overall, the differences with the estimates given in Table 5 are relatively small. A notable exception, however, is that the association of spouse’s income with male mortality risk is significant when random effects are not controlled for (Table 7, Model 5) and turns insignificant when controlling for random effects (Table 5).

Although these estimation results show the direction and relative size of the associations between the covariates and mortality risk, they provide no clear insights into the quantitative association with remaining life expectancy at age 65. In the next section, we therefore use the estimation results and apply Monte Carlo simulations⁹ to quantify remaining life expectancy at age 65 and examine by gender and marital status its association with individual and spousal income.

Simulation Results

In the Netherlands, all residents receive a flat-rate public pension benefit from the statutory retirement age of 65 onward. Occupational pensions, on the other hand, are determined by the so-called pension-related gross yearly salaries that are a function of preretirement earnings and years of employment, and by the rules of the public and private pension systems. Occupational pensions are also transferable to the surviving spouse. Our simulations take these issues into account when calculating retirement income for several types of reference households (see Table 9 in the appendix). In the baseline we have an individual with a median pension-related gross yearly salary, which is about €29,500 for a full-time employee who has worked for 40 years. The simulation results are given in Table 6.

In the first column of Table 6, we quantify the association between income and remaining life expectancy at the age of 65 by gender for three different household types (and at median income).¹⁰ Remaining life expectancy is almost 12 years for men and about 18 years for single women (first column, first two rows of Table 6). Remaining life expectancy at 65 is considerably higher for married individuals: about 16 years for married men (first column, third row) and 20 years for married women who have been employed part-time (first column, last row). The estimated differences between single and married individuals are about 4 years for men and 2 years for women, and are statistically significant (the p values are, respectively, .00 and .01).

Next, for each household type, we consider the differences from the baseline that result from changes in (lifetime) income. The second and third columns report the change in remaining life expectancy associated with an income 10 % higher than the median for either the man or the woman in each of the three household types. As the table shows, this association differs little across gender and household type: the point estimates vary between 0.16 and 0.20, with standard errors equal to 0.04. These estimates imply that life expectancy at age 65 for individuals with an income 10 % above the median is about 2 to 2.5 months higher than that for individuals with median income.

The fourth and fifth columns report the differences in remaining life expectancy between low- and median-income individuals for the different household types. Individuals on minimum wage or with no earnings (low income) during their working lives receive only a public pension benefit during retirement. For both men and women, and irrespective of marital status at age 65, the difference in remaining life expectancy is less than one year (fourth and fifth columns). The differences in remaining life expectancy between high-income individuals (those who earned twice the median income during their working lives) and median-income individuals are reported in the sixth and seventh columns. For men and women, the point estimates of these differences fall between 1.53 (for married women) and 1.91 (for single men). A comparison of the differences between these two extremes reveals that the difference in remaining life expectancy at age 65 between low-income individuals with only a public pension and high-income individuals with twice the median income is about 2.5 years for both men and women.

Finally, we turn to the association between remaining life expectancy at 65 and spousal income. As shown in the fourth and sixth rows of Table 6, compared with married women whose spouses earned median income, remaining life expectancy at age 65 of a woman whose spouse earned an income 10 % above the median is 0.18–0.19 years higher (about 2.5 months). For married men (fifth line), however, this difference is small and insignificant. In contrast, the difference in remaining life expectancy at 65 between women with a low-income spouse and women with a high-income spouse (see columns four and six, last row) is more than two years (1.62 – (–0.79)). This difference can be explained in two ways. First, spousal income is negatively associated with women’s mortality risk, albeit at only the 10 % level (see Table 5). Second, and more important, marriage is negatively associated with mortality risk; women benefit, on average, from an extended duration of marriage when married to a high-income man. For men, however, we find no such indirect association because men—having, on average, a shorter life span than women—benefit relatively less from a spouse’s higher income.

Summary

This analysis quantifies the association between individuals’ income and remaining life expectancy at the statutory retirement age (65) in the Netherlands. For this purpose, we estimate a mortality risk model that explicitly controls for unobserved individual-specific heterogeneity (random effects) using administrative data taken from the 1996–2007 Income Panel Study of the Netherlands supplemented with data from the Causes of Death registry.

Our main empirical findings are threefold. First, in terms of model specification, we find a significant presence of unobserved individual-specific heterogeneity (random effects) and the required dynamic sample selection correction terms for men; but for women, such heterogeneity is significant at only the 10 % level. For both men and women, the effects of not controlling for unobserved individual-specific heterogeneity on the estimated associations between the observed characteristics and mortality risk are relatively small but with a notable exception concerning the association between spouse’s income and mortality risk for men, which becomes insignificant when we control for random effects. These findings underscore the importance of controlling for unobserved individual-specific heterogeneity if inconsistent estimates are to be avoided.

Second, in regard to the association between spousal income and individual mortality risk, we find that conditional on marital status, and only for women, spousal income is weakly (at the 10 % level of significance) associated with mortality risk. This finding provides little support for the suggestions that an association with spousal income may exist when health or mortality risk is dependent on material hardship, measured by household not individual financial resources (Martikainen et al. 2001), or when couples share a lifestyle that is influenced by both partners (Torssander and Erikson 2009). Rather, it suggests that own income is more relevant to health and mortality risk than spouse’s income.

Third, in terms of the association between individual income and mortality risk, we find that individual income is about equally strong and negatively associated with mortality risk for men and women. The difference in remaining life expectancy at age 65 between low-income individuals with only a public pension and high-income individuals with twice the median income is about 2.5 years for both men and women. This difference is close to the retirement window between the ages 65 and 67 that is part of the proposed pension reforms for the Netherlands (Stichting van de Arbeid 2010). Such a retirement window may mitigate the adverse income redistribution effects that result from differential mortality (as discussed in the Introduction) by allowing low-income individuals with a low life expectancy to retire relatively early.¹¹

Acknowledgments

Financial support from Stichting Instituut GAK (through Netspar) is gratefully acknowledged. We wish to thank Margherita Borella, Koen Caminada, Katherine Carman, Stefan Hochguertel, Arie Kapteyn, Pierre Carl Michaud, seminar participants at the Utrecht School of Economics, the Center for Research on Pensions and Welfare policies in Turin (CeRP’s 10th anniversary conference), University of Padova, the Netspar workshop “Labour Force Participation and Well-being of the 50+ Population” (September 2009), the European Society for Population Economics Conference 2010, the Econometric Society World Congress 2010, and two anonymous referees for valuable comments and discussions.

Appendix

Notes

References