Working Paper Series Can the longevity risk alleviate The annuitization puzzle? Empirical evidence from Dutch data Federica Teppa ECINEQ WP 2011 – 223 ECINEQ 2011 – 223 October 2011 www.ecineq.org Can the longevity risk alleviate The annuitization puzzle? Empirical evidence from Dutch data* † Federica Teppa De Nederlandsche Bank (DNB) and Netspar Abstract This paper provides new evidence on individual preferences over annuities and lump sum payments based on hypothetical questions posed in the DNB Household Survey in 2005. Contrary to the majority of papers in the annuitization puzzle literature, this study allows to control explicitly for the subjective survival probability (SSP), a key driver of the decision about whether to annuitize or not as a perceived measure of longevity risk. We find that people expecting to live longer do claim to prefer the annuity. This finding is very robust to controlling for bequest motives. The relevance of this paper is twofold. First, it delivers an important empirical result on the role of the SSP that is still not directly tested in the literature. Second and more important, combined with the empirical evidence that on average individuals tend to systematically underestimate their life expectancy, the findings have strong policy implications. The annuitization puzzle may be alleviated by helping individuals in better assessing their longevity risk, rather than forcing their actions. Keywords: Longevity Risk; Annuitization Puzzle; Survey Data; Hypothetical Choices JEL classification: C5; C8; D12; G11 * We thank Maarten van Rooij for providing us with the data on the choice between the annuity and the lump sum payment. We thank CentERdata at Tilburg University for supplying the data of the DNB Household Survey. The paper has benefited from useful comments at the 14th International Business Research conference (Dubai, UAE, April 2011 - Best Paper Award), 9th International Workshop on Pension, Insurance and Saving (Paris, May 2011), International Journal of Arts Sciences conference (Bad Hofgastein, Austria, May 2011), 17th conference of the Society of Computational Economics - Computing in Economics and Finance (San Francisco, June 2011), 4th Society for the Study of Economic Inequality Conference (Catania, July 2011), Singapore Economic Review Conference (Singapore, August 2011), 67th International Institute of Public Finance Congress (Ann Arbor, Michigan, August 2011) and at DNB seminars. The views expressed in this paper are those of the author and do not necessarily reflect those of the institutions she belongs to. Any remaining errors are our own responsibility. † Contact details: [email protected] 1 Introduction Life expectancy has improved substantially since the past decades and it has ac- celerated in the recent years in all developed countries. In the Netherlands this phenomenon is particularly strong for males. According to the most recent World Health Statistics, life expectancy at birth has gone from 74 years in 1990 to 78 years in 2008 for males, and from 80 years in 1990 to 82 years in 2008 for females. In the same period, adult mortality rate, defined as the probability of dying between 15 and 60 years, has decreased from 11.6 percent to 7.8 percent for males, and from 6.7 percent to 5.7 percentfor females. The decliningfemale advantage inlife expectancy is observed in the US as well (Vallin, 1991) and largely driven by behavioral factors (namely smoking) rather than biological factors (Pampel, 2002). In an increasingly ageing society the need to provide with adequate insurance for late-life consumption has become a high priority item in the agenda of the policy makers. As the only contract that acts as insurance against longevity risk, the annuity should always be chosen by risky individuals, even in presence of bequest motives (Yaari 1965; Davidoff et al. 2005). Yet the empirical evidence from several countries showsthatonlyaminorfractionofindividualsvoluntarilybuysannuities(Jamesand Song 2001; Johnson et al. 2004; Beatrice and Drinkwater 2004). The combination of these two facts is known as the “annuitization puzzle”. The annuitization puzzle is a well documented phenomenon in the literature. Several potential explanations have been discussed extensively in the literature. They include both supply side reasons, e.g. highly priced annuities due to adverse selection and administrative costs (Brown et al. 1999, 2001 for the US; Cannon and Tonks2004,FinkelsteinandPoterba2004fortheUK),anddemandsidemotives,e.g. intra-family risk sharing (Kotlikoff and Spivak 1981), liquidity constraints and large out-of-pocket health expenditures (Palumbo 1999; De Nardi et al. 2010), preference for bequests (Friedman and Warshawsky 1990; Vidal-Melia and Lejarraga-Garcia 2006). Morerecentlyalternativetypicallybehaviouralexplanationshavebeenfound, e.g. framing effects or default effects (Bu¨tler and Teppa 2007; Agnew (et al.) 2008; Brown et al. 2008). This paper follows a different approach in that it focuses on longevity risk, a driver that should be key in this type of choice and that has been missing in the analysis so far. There are several ways to elicit information about individual life expectancy, both indirectly, looking at parental longevity, or directly, by asking subjective survival probabilities (SSP from now on). Both measures suffer from several drawbacks (e.g. focal points, rounding effects) but overall they seem to 3 convey meaningful information on the individual longevity. There is evidence from the Health and Retirement Survey (HRS) that SSP contain useful information on survival expectations. They have been found to be correlated with known mortality riskfactors, topredictactualmortality, althoughlesswellonceself-assessedhealthis controlled for (Siegel et al. 2003), and are claimed to closely approximate actuarial survival probabilities on average (Hurd and McGarry 1995; Smith et al. 2001; Hurd and McGarry 2002). The English Longitudinal Study of Ageing (ELSA) data have been used to test the predictive power of SSP for actual mortality and a systematic underestimation of survival chances relative to those given in actuarial life tables has been noted (Banks et al. 2004; O’Donnell et al. 2008). More recently, SSP for the Netherlands have been used to analyze their impact on retirement intentions and actual behaviour (van Solinge and Henkens 2010). In this paper we use subjective survival probabilities as measures of perceived longevity risk in a simple model for individual preferences over annuities and lump sum payments based on hypothetical questions posed in the DNB Household Survey in 2005. We find that people expecting to live longer do claim to prefer the annuity. This finding is very robust to controlling for bequest motives, that turns out to be the other main determinant for the choice of lump sum payments. The relevance of this paper is twofold. First, it delivers an important empirical result on the role of the SSP that is still not directly tested in the literature. Second and more important, combined with the empirical evidence that on average individuals tend tosystematically underestimate theirlifeexpectancy, the findingshavestrongpolicy implications. The annuitization puzzle may be alleviated by helping individuals in better assessing their longevity risk, rather than forcing their final actions. Thepaperisorganizedasfollows. Section2describesthedatausedintheempir- ical analysis. Particular emphasis is devoted to the subjective survival probability, on how it has been elicited and on how it relates to the main individual background and socio-economic characteristics. Section 3 describes the empirical model with a focus on the dependent variable and the sample restrictions. Section 4 reports and discusses the empirical results. Section 5 concludes. 2 The data The empirical analysis is based on data collected from the households participating in the so-called DNB Household Survey (DHS). The DHS, formerly known as the CentER Savings Survey, is an annual panel survey of more than 2,000 households in the Netherlands that started in 1993. The panel is run at Tilburg University by 4 CentERdata. Panel members are aged 16 years and older. In case of attrition, Cen- tERdata recruits new participants to maintain the panel size and to keep the panel as representative as possible on a number of relevant background characteristics such as age, gender, income, education, and region of residence. The DHS dataset further contains detailed information on employment status, pension arrangements, accommodation, wealth, as well as health status, and psychological concepts. The dataset thus provides the opportunity to combine both economic and psychological aspects of financial behavior. 2.1 The subjective survival probability (SSP) Thispaperfocusesonlongevityriskanditsimpactonthechoicebetweenanannuity and a lump sum payment. In this study we use survey questions on subjective survival probabilities available for 2005. We then merge these data with the 2005 DHS wave in order to have all the relevant information present in the survey. The life-expectancy questions given to the respondents have the following format which strictly follows the one used in the HRS as well as in the ELSA: Please indicate your answer on a scale of 0 to 10, where 0 means “no chance at all” and 10 means “absolutely certain”. SSPXX : How likely is it that you will attain (at least) the age of XX? Thetargetage(denotedbyXX)dependsonthecurrentageoftherespondent. In particular, SSP75 is presented to people aged between 16 and 64; SSP80 is presented to people aged between 16 and 69; SSP85 is presented to people aged between 65 and 75; SSP90 is presented to people aged between 70 and 80; SSP95 is presented to people aged between 75 and 85; SSP100 is presented to people aged between 80 and 90. Since the answers are on a 0-10 scale, we can interpret value 1 as “1 to 10 percent likely to attain (at least) the age of XX”, value 2 as “11 to 20 percent likely to attain (at least) the age of XX”, and so forth. This format is very similar to that used by van Solinge and Henkens (2010), even if they ask this probability on a 1 to 5 scale, and they only ask for the target age of 75. It is also important to note that by question design these probabilities are conditional on being alive at a certain age. Table 1 presents the main summary statistics and Figure 1 shows the histograms for each subjective survival probability. A careful analysis of these statistics is needed in order to assess the informative content and to validate the overall quality of the various SSPs. Table 1 and Figure 1 about here 5 The number of observations decreases severely as the target age increases, as a consequence of the routing in the question design. However, we can infer that the severalSSPshaveaconsistentandinformativecontent. Wenotethatboththemean and the median value of the SSPs monotonically decline with respect to the target age. The standard deviation is highest for SSP90 and SSP95, lowest for SSP100 and rather stable for the remaining SSPs. Several dispersion measures, like the variance and the standard error of the mean, provide some evidence that the respondents report lower chances to attain higher target ages, but they are also more uncertain about that, except for reaching age 100. The distributions are all non-symmetric but differ with respect to their skewness, which is negative for the three lowest target ages and positive for the three highest target ages. The most left-skewed distribution, with relatively few low values, and the most right-skewed distribution, with relatively few high values, are those for the extreme target ages, namely SPP75 and SPP100, respectively. This means that it is most likely to attain age 75 and least likely to attain age 100. In addition, the skewness monotonically increases with the target age; for SSP85 the distribution has roughly zero skewness and is unimodal (mean = median = mode = 5). Finally, inordertoassesswhetherthedataarepeakedorflatrelativetoanormal distribution we report the kurtosis. We observe that the histogram with the highest kurtosis is that for SSP75, with a distinct peak near the mean value. 2.2 SSPs and socio-economic variables The DHS contains a great amount of information on several background as well as socio-economic characteristics, both at the individual and at the household level. In thissectionwemakeanoverviewofhowtheSSPsrelatetosomeofthesevariables, in particulartothoseforwhichitisreasonabletoexpectameaningfulrelationship. We know for example from mortality tables that females have a higher life expectancy than male, on average. Similarly, there is some empirical international evidence about a positive correlation between life expectancy and education level, as well as financial situation. We also expect SSP to be associated with health status, both subjectively reported and derived from more objective illnesses. With these ideas in mind, we select gender, education level, self-assessed health (SAH from now on), long-term illness, smoking behaviour, drinking habits, and household income. Table 2 reports the mean values of each SSP by background and socio-economic factors. Table 2 about here 6 The findings for gender are rather mixed. Women tend to report higher survival probabilities than man on average, but only in one case out of six this difference is statistically significant (at the 5-percent level). Moreover, in two cases (namely SSP85 and SSP90) this difference is negative, though not significant. This findings contrasts with international evidence of women living longer than men, on average. We thus devote a deeper thought on this in the next subsection. The evidence for education level is more consistent, as the respondents with better education tend to have higher survival probabilities on average for all target ages up to 90. This health protective role of education is in line with Cutler, Lleras- Muney and Vogl (2010). In addition, the difference for SSP75 is strongly significant (1-percent level) whereas that for SSP80 is less significant (10-percent level). For the two highest target ages, the difference turns out to be positive, and also significant at the 5-percent level for SSP95. This finding is rather counterintuitive, but could be (partly) explained by selective mortality. A much more consistent picture is found for self-assessed health. For all target ages the individuals reporting good or very good SAH systematically report higher average survival probabilities than those with fair, bad or very bad SAH. The differ- ences are always strongly significant. Similar evidence is found for long-term illness. The respondents who claim to suffer for LT illness significantly report lower survival probabilities than those who claim to have no LT illness, on average. Both smoking and drinking behaviour seems to be only weakly related to SSPs. In both cases higher survival probabilities are reported by the respondents who declare to be non-smokers and to drink no alcohol, but the difference is strongly significant (at the 1-percent level) for the two lowest target ages only. Finally, theSSPmeasuresdonotseemtoberelatedatallwithhouseholdincome. We experimented with several cut-off points in household income, but the findings of no correlation are rather robust. This finding seems to be in line with Deaton’s findings that as far as controllable vs. non-controllable diseases (e.g. cardiovascular vs. all cancer types) is concerned, among adults income is not important, but education is. In particular Deaton finds that education is health protective for controllable diseases only, whereas income is never health protective. 2.3 Subjective vs. actuarial survival probabilities Another aspect that should be taken into account in assessing the quality of the SSPs is to relate them to actuarial survival probabilities. Do individuals perceive their longevity risk (and consequently form their subjective probabilities) correctly? 7 To answer this question we compare the subjective survival probabilities from survey data to the actuarial survival probabilities from official mortality tables. Actuarial survival probabilities are computed from mortality rates provided by Statistics Netherlands (CBS, Centraal Bureau voor de Statistiek). Since the DHS datareferto2005, weconsiderthe2005actuarialmortalityrates, byageandgender. In order to make the two series of survival probabilities comparable, we construct the subjective survival probabilities implied by the SSPs by transforming the SSPs from the 1-10 scale into percentages. Figure 2 reports the two series of statistics for the survival probabilities of reach- ing (at least) age 75. We only consider individuals aged 50+, for whom this kind of comparison is not affected by potential cohort effects. The upper panel refers to females; the lower panel refers to males. The figure clearly shows that females underestimate their survival probabilities at all ages. For some ages this underestimation is quantitatively very strong (around 25 percentage points for age 52 and age 60). Similar evidence is found in the HRS data for United States by Perozek (2008). Though evidence of substantial misperception of longevity risk for males as well is there, males seems to assess their survival probabilities better than females. The fact that males have a better clue of their true survival probabilities explains the surprisingly mixed picture that emerges from Table 2 above. The demographic trend of women living longer than men, on average, is not mirrored in the reported subjective survival probabilities by gender mainly as a consequence of the stronger misperception of the actuarial survival probabilities by females than by males. Overall, the empirical evidence documented so far seems to point to the con- clusion that the SSPs, though neither perfect nor exempt from limitations, convey reasonably meaningful information on individual longevity, and relate relatively well with a number of background and socio-economic characteristics, on average. These findings are fully in line with van Solinge and Henkens (2010). At the same time, the comparison between subjective and actuarial survival probabilities shows that individuals systematically underestimate their longevity, in some cases very strongly, especially for females. These findings are again fully in line with international figures (e.g. O’Donnell et al. 2008 for UK). Figure 2 about here 8 3 The empirical model 3.1 The dependent variable The dependent variable in our models is derived from hypothetical questions on preferences over a full annuity or a partial lump sum payment upon retirement. The first question reads as follows: Imagine you are 65 years old, and you are receiving AC 1,000 per month in state pension. Suppose you were given the choice to lower that benefit by half, to AC 500 per month. This one-half benefit reduction would continue for as long as you live. In return you would be given a one-time, lump sum payment of [AC 87,000 (for females) / AC 72,000 (for males)]. Would you take the AC 1,000 monthly benefit for life, or the lower monthly benefit combined with the lump sum payment? This initial question is asked to all respondents in the sample, irrespective of their working status and for all ages. At this stage, the respondents are given a fair deal. The lump sum payment is computed to be actuarially fair and thus the amount differs by gender: Females are confronted with a payment of 87,000 euros, males with 72,000 euros. The choice is then between a full annuity and a partial lump sum payment. For simplicity, from now on we omit the words “full” and “partial” when referring to the annuity and the lump sum payment, respectively. However, it is important to keep in mind, especially when interpreting the empirical results, that the other polar case of full lump sum payment is never offered to the individuals in this exercise. Depending on the answer given to this question, the respondents are asked a follow-up question, where the lump sum payments is made more (less) attractive to those individuals who had preferred the annuity (the lump sum payment) in the first round. Figure 3 reports the structure of the question sequence. Table 3 reports the mean values of the choice between the annuity and the lump sum payments for the full sample, as well as by gender and by the presence of children. Figure 3 and Table 3 about here The annuity is preferred by slightly more than half of respondents (54 percent) in Question 1.1 Conditional on having chosen the annuity in Question 1, then the 1This is in line with Brown (2001) who finds that 48 percent of the HRS sample reports that they will annuitize their DC plan. 9 annuityisstilllargelypreferredtothelumpsumpaymentinQuestion2a(69percent vs. 31 percent, respectively). Similarly, conditional on having chosen the lump sum in Question 1, then the annuity is preferred only by 40 percent of individuals in Question 2b. There is evidence of persistent preferences as only 17 percent of individuals switch from the annuity to the lump sum payment (172 out of 1,027), andonly18percentofindividualsswitchfromthelumpsumpaymenttotheannuity (185 out of 1,027). The overall picture does not change when the choice is made by gender and by the presence of children. We notice however that males and respondents without children prefer the annuity the most (57 and 56 percent respectively) in Question 1. Both the difference with females and the difference with people with no children are significant at the 5-percent level in Question 1. No significant differences by gender or by having children is found for the follow-up questions. We also made the analysis (not reported in the table) by the presence of partner and household income: the differences are non significant. It is important to notice that the framing of this question is not fully “neutral” as it involves an explicit opting-out option (a lump sum payment in place of half annuitized pension wealth). This set up was used in the 2004 wave of the HRS. In the 2008 wave of the HRS a somewhat different wording was used in order to elicit the information about willingness to annuitize: Imagine you are 65 years old, and you are receiving $1,000 per month in Social Security benefits. Imagine that you are currently getting $1,000 per month in Social Security benefits. Suppose you had a choice: either you could keep that $1,000 monthly benefit for life, or you could exchange it for a monthly benefit half that size, $500 per month for life, plus youd get a one-time, lump sum payment. What is the smallest lump-sum that you would be willing to accept in exchange for reducing your lifetime benefit by $500 per month? $ .... Amount We model the choice between the annuity and the lump sum payment by a standard binary choice model, where the dependent variable takes value 1 if the annuity is chosen in Question 1, 0 if the lump sum payment is preferred in that same question. We then perform simple probit regressions. 3.2 Sample restrictions ContrarytoBu¨tlerandTeppa(2007), whoprovidewithempiricalevidenceonactual choices, this paper is based on purely hypothetical choices between the annuity and the lump sum. In order to make this choice as close to reality as possible, we restrict 10