Do Individual Accounts Postpone Retirement?: Evidence from Chile. Alejandra Cox Edwards and Estelle James*

Transcription

1 Do Individual Accounts Postpone Retirement?: Evidence from Chile by Alejandra Cox Edwards and Estelle James* *We appreciate support from the Social Security Administration through the Michigan Retirement Research Center. We received helpful comments and suggestions from Richard Disney and other participants of the NBER 2005 Summer Institute Social Security Workshop. The findings and conclusions expressed are solely those of the authors and do not represent the views of the Social Security Administration, any agency of the Federal government, or the Michigan Retirement Research Center. Contact numbers: and 0

2 Draft, October 31, 2005 Do Individual Accounts Postpone Retirement: Evidence from Chile Estelle James and Alejandra Cox Edwards Postponing retirement among ageing populations will become increasingly important as a means to increase the labor force, its output and old age security. Recent research has focused on incentives stemming from the social security system that influence the worker s decision to retire. Defined benefit systems (both public and private) often contain penalties for postponing access to pensions or continuing to work while receiving a pension. In contrast, the tight link between contributions and accumulations and the actuarial conversion of accumulations into pensions in defined contribution systems may lead workers to postpone pensions or to continue working after withdrawals begin. The experience of Chile, which adopted its new system in 1981, offers an opportunity to test if the change in incentives has indeed produced the expected change in retirement behavior. Using probit analysis of household survey data from 1960 to 2004, we estimate the impact of the pension reform on the probability of 1) becoming a pensioner and 2) dropping out of the labor force, for older workers. We find strong effects of the new system on both propensities, after controlling for individual and macro-economic variables. In particular, restricted access to early pensions and the exemption of pensioners from the pension payroll tax appear to exert a powerful effect on labor force participation rates. 1

3 Do Individual Accounts Postpone Retirement?: Evidence from Chile Postponing retirement will become increasingly important as a means to increase the labor force, its output and old age security as populations age. Yet, the labor force participation rates of older men declined in many countries during the last few decades of the twentieth century (Hurd 1990, Borsch Supan 1998, Peracchi and Welch 1994, Anderson et al 1999). We now have international evidence that workers respond to incentives stemming from social security systems in making their retirement decisions (Gruber and Wise 1999 and 2004). Most defined benefit (DB) plans contain incentives that encourage early retirement from the pension system and the labor force. In contrast, the tight link between contributions and accumulations and the actuarial conversion of accumulations into pensions in defined contribution (DC) systems may lead workers to postpone retirement. The experience of Chile, which changed from a traditional pay-asyou-go DB to a fully funded DC system in 1981, offers an opportunity to test if the change in incentives has indeed produced the expected changes in behavior. Although DC plans controlled by workers are becoming more common, Chile is the only country that has had over 20 years experience with a social security system based on these plans. We are not aware of any previous studies addressing this impact of the Chilean reform. 1 Prior to 1981 Chile also had a traditional pay-as-you DB system that included disincentives to work among older individuals, similar to those found in many European countries today. These include: a high contribution rate that was required for all workers; eligibility conditions that enabled retirement at an early age, beyond which the expected present value of accrued pension wealth would fall; and a failure of incremental benefits to keep up with incremental contributions. In the old system, workers had strong incentives to start their pensions and to stop working and contributing as soon as possible 2

4 because this maximized the present value of their lifetime net benefits. In contrast, in the new DC system these work disincentives were removed as payroll taxes were greatly reduced, benefits were linked to contributions on an actuarially fair basis, and preconditions for early retirement were tightened. We expect the new system to raise the labor force prticipation rates of older workers, through two channels: (1) Postponed pensioning. The restrictions on early withdrawal and the actuarial adjustments when withdrawals are postponed should decrease age-specific pension probabilities. This will indirectly increase work propensities, since non-pensioners are more likely to work than pensioners due to liquidity constraints. (2) Increased work propensities among pensioners. The absence of penalties for continued work, the reduced payroll tax for all contributors and the exemption of pensioners from the pension payroll tax should directly increase labor force participation rates, especially for pensioners. Part 1 reviews the literature on determinants of retirement age, in particular on the impact of system characteristics, and outlines in greater detail the work and pension incentives in the new Chilean scheme. Part II describes our data set and aggregate results on the pension probabilities and labor supply of older workers before and after the reform. Part III applies the probit model to estimate the impact of individual, family and system characteristics on the probability of becoming a pensioner and dropping out of the labor force. We use birth cohort as a proxy for membership in the new system and focus on changes in the behavior of cohorts that reached age 50 after the reform, emphasizing differences between pensioners and non-pensioners. The Conclusion highlights lessons for other countries. 3

5 Our main findings: Labor force participation rates of older men (age 50-70) rose substantially after the reform, in contrast to the decline observed before the reform. This effect, which is evident in the aggregate data, remains after we control for individual and macroeconomic characteristics. The increase in labor force participation is concentrated among pensioners probably a response to their exemption from the pension payroll tax. Simultaneously, the probability declines that older men will become pensioners prior to the normal age of 65 mainly due to tightened eligibility conditions for early retirement. Postponed pensioning further increases participation rates for liquidityconstrained older men. These effects are stronger for cohorts that were younger on the date of the reform--therefore more likely to have switched to the new system and to take its incentives into account in their life cycle planning. According to our calculations, the total increase in labor force so far is about 4.4%, implying an annual growth increment of.21%. We expect that a new steady state with higher participation rates will eventually be reached, but our data only allow us to observe the transition stage. I. What do Theory and Empirical Evidence Suggest about Retirement Choice? Studies from other countries During the past decade a large literature has developed to explain the falling labor force participation rate among older men and to suggest policy changes that might reverse this behavior. This research has investigated macroeconomic variables such as higher incomes that lead workers to afford longer periods of retirement, personal characteristics that determine who leaves the labor force early and pension system rules that discourage continued work. We focus on the latter question in this paper. The benchmark study of this topic is by Gruber, Wise and their colleagues (1999 and 2004). Their study develops three measures of the incentive to retire early in the 4

6 defined benefit systems of eleven industrialized countries, and analyzes the relationship between these incentives and the observed labor force participation rates of older men across countries and time. Regressions show that, both at the macro and micro levels, the work disincentives and responses to them are large; therefore most people retire as soon as they become eligible for a pension; the greater the disincentives facing an individual the less likely he is to postpone retirement; and people in different cultural settings have very similar responses. The simulated effects on retirement choice and labor supply of policy reforms that remove these work incentives are consistently large. Several studies using US data emphasize the impact on labor force participation of the income effect stemming from social security wealth and the timing effect stemming from early retirement rules, both of which have changed over the past forty years. Hurd (1990) underscores the peak in hazard rates at age 62 that appeared after the early retirement option was introduced in 1961 (also see Burtless and Moffitt 1986). Hurd also argues that increasing social security benefits in the early 1970 s contributed substantially to falling labor force participation rates in the 1980 s and 1990 s, through an income or wealth effect. However, Burtless (1986), Peracchi and Welch (1992) and Blau (1994) conclude that this impact was quite small. Krueger and Pischke (1992) maintain that the apparent negative impact of social security wealth on labor force participation rates is due to correlated time effects rather than real system effects. They avoid this correlation by analyzing the impact on participation for the notch generation that retired in the late 1970 s and early 1980 s. Benefits and social security wealth were suddenly and unexpectedly reduced when double-indexing was eliminated in 1977, yet participation rates continued to fall, suggesting that the downward trend comes from other forces. Gustman and Steinmeyer (1985) simulate the long run response to the 1983 social security reforms and find a positive effect on full time work beyond age 65, mainly 5

7 due to the increase in the delayed retirement credit. Anderson, Gustman and Steinmeyer (1999) simulate how behavior might have been different in the US with 1969 rules as compared with 1989 rules and find a modest effect. In general, the estimated effects in these US studies are modest but this is not surprising in view of the fact that the rule changes have been relatively modest and the US system has long been actuarially fair compared with systems in many other countries. As for other countries: Borsch-Supan (1998) and Borsch-Supan and Schnabel (1999) look at Germany and several other European countries during They find differing and changing rates of decline in labor force participation rates of older workers across countries and time, which they attribute largely to varying disincentives coming from their pension systems. Disney (2004) estimates the pure tax component (separate from the savings component) in public pension programs across several OECD countries and time periods. He finds that women s economic activity responds more negatively to the pure tax component than to the contribution as a whole, while work activity by men does not seem to be affected. However, this analysis does not focus on the behavior of older men, whose retirement decisions may be more responsive to taxes than are the work decisions of younger men. Baker and Benjamin (1999) study the impact of the introduction of early retirement options in Canada in the early 1980 s. They find that this option had little impact on labor force participation rates mainly because those who took early retirement would otherwise have had low labor force participation rates anyway and worked only intermittently prior to retirement. Access to early retirement pensions increased their income but changed their behavior minimally as selection rather than incentive effects dominated. In Switzerland (despite a public benefit that nominally does not permit early retirement and a private benefit that rewards continued work) the labor supply of older men has fallen over the past 15 years. Butler, 6

8 Huguenin and Teppa (2004) attribute this primarily to the maturing of generous DC plans, which increased the pension wealth of many workers beyond the point they would have chosen voluntarily. Workers spend this wealth on longer periods of leisure in old age. This paper analyzes the actual response to changes in the incentives embodied in Chile s new pension system. Most of the studies mentioned above have had the advantage of longitudinal data, which we do not have. But few studies have had the advantage that system incentives have changed so dramatically and that data are available for a long period of time to observe the consequences. Work and pension incentives in the new Chilean scheme Three potentially separable decisions are involved in the choice of retirement age. First is the decision about when to stop contributing to the social security system, second the decision about when to start withdrawing retirement benefits, and third the decision about when to stop work. In principle we could model three separate behavior rules for these three decisions: 1.) Contribute so long as the benefits from marginal contributions exceed the consumption or alternative investment benefits that could be achieved with the money; 2.) Postpone withdrawal so long as saving is desired and the incremental return to social security wealth exceeds the return that could be earned outside the system; 3.) Work so long as the incremental benefits from work (wages, pensions and other fringe benefits) exceed the marginal utility of leisure. In traditional DB systems, such as those analyzed by Gruber and Wise (1999 and 2004), these decisions were closely linked by incentives and constraints so a sharp distinction is usually not drawn. Contributions are required so long as the individual works, so the work-contribution decision is joint. Usually the net marginal benefit from 7

9 incremental contributions becomes negative in early old age, which leads workers to stop working and contributing at the first opportunity--as soon as access to an alternative source of income is established. Beyond that age, the rate of return earned on accumulated social security wealth is usually far less than the market rate of return, so withdrawals in the form of a pension start as soon as eligibility is established. The pension is the alternative income source that enables individuals to stop working and contributing, thereby linking all three decisions together. Sometimes this is reinforced by a requirement that benefits are cut for those who do work or, even more strongly, by a prohibition on work for pensioners. The linkage of these three decisions, together with an early and falling age of eligibility, helps explain the declining labor force participation rates of older men in many countries during the past three decades and the lower participation rates in countries with higher implicit tax rates. Prior to 1981 Chile also had a traditional pay-as-you DB system that included disincentives to work among older individuals, similar to those found in many European countries today. While the old social security system in Chile was very fragmented by occupation and sector, in the largest sub-system, Servicio Seguro Social (SSS), the payroll tax for pensions was 23%, with another 10% for other social insurance, bringing the total to 33%. 2 Workers were eligible for a generous defined benefit after only sixteen years of contributions. Taking the pension at the first age of eligibility (well before age 60) usually maximized the present value of lifetime social security benefits because the normal age pension was never decreased in an actuarially fair manner. Continued work in the public sector was not allowed while receiving a pension. Although private sector work was usually allowed, it was penalized. The full contribution rate had to be paid, but the incremental benefit was small and sometimes non-existent. For example, in the SSS the monthly benefit was 50% of the base wage after 10 years of service and only 1% 8

10 additional for each year thereafter, until a 70% ceiling was reached after 30 years, at which point the incremental benefit became 0. In fact, postponed retirement could actually reduce the pension, since the wage base was the average of the last 5 years of work and wages typically peaked between ages Incremental benefits were further dampened by the fact that pensions were not automatically indexed for inflation (although inflation at times exceeded 100% per year). After 1981 a new system based on defined contributions was instituted in Chile, with incentives that were much more pro-work. The new Chilean system delinks these three decisions. It sets a threshold replacement rate beyond which workers can become pensioners if they wish. This constraint on early pensioning is tighter than in the old system. It allows pensioners to work or not work, and if they work to contribute or not contribute, according to their preferences. It also rewards contributions and postponed withdrawals on an actuarially fair basis, using market rates. These forces should decrease the propensity to start the pension since early pensioning is more difficult and postponing is no longer penalized--and increase the propensity to work while receiving a pension since the net wage has been increased by removing the mandatory payroll tax for pensioners and by lower payroll taxes for everyone. Both these effects should increase the labor supply of older workers. Specifically, in the new system: (1) Payroll taxes for pensions were cut to %, in contrast to the old rate of 23%; 3 (2) Contributions accumulate in the individual s account, invested in a pension fund (AFP), and earn a market rate of return that exceeded 10% real per year, on average, for the first two decades of the system; 4 9

11 (3) At the normal retirement age (65 for men, 60 for women) workers can start withdrawing regardless of the amount in their accounts. But those who accumulate enough to purchase a pension that meets a specified threshold may start withdrawing and stop contributing earlier. (4) Upon retirement, the accumulation is turned into a price-indexed pension (an annuity or gradual withdrawal) on actuarially fair terms, so incremental contributions and postponed withdrawals yield commensurate benefits. (5) Pensioners can continue to work and are exempt from the pension payroll tax. At that point, their net wage rises by as much as 15% (13/87). (Because of actuarial fairness, only part of the contribution will be considered a tax and the increase in net wage will therefore be less than 15% for most pensioners. Some portion is nevertheless a tax due to time preference and inflexibilities discussed below). (6) The new system contained a minimum pension guarantee (MPG), which has several implications for work incentives. The MPG sets a pension floor that is guaranteed by the government. Due to periodic increases roughly linked to wage growth, this floor has been maintained at 25-27% of the average wage. If the individual s own accumulation is insufficient to meet the floor for his entire retired lifetime, the government provides a subsidy. The subsidy may take the form of a top-up to an annuity, or payment of a full pension at the MPG level if the individual s account has been depleted through gradual withdrawals. Twenty years of contributions are needed to qualify for the MPG. Thus, low earners who have not yet contributed for 20 years have an incentive to continue working and contributing until they reach that point. Moreover, the applicable MPG is reduced for workers who start their pensions early, thereby encouraging those with small accumulations to delay pensioning until age 65; this probably also means they must continue working. These two factors should increase 10

12 work among low earning non-pensioners. However, once they reach 20 years of contributions, any marginal contributions of their own will simply displace the public subsidy for those who are near the MPG, which should discourage incremental work by non-pensioners. Pensioners are exempt from such contributions, so this disincentive stops once they start their pension. But access to the MPG is supposedly means-tested. If the means-test is enforced, that will discourage work by pensioners on the MPG. Thus, the incentives faced by individuals in the vicinity of the MPG are complex, conflicting and may cancel each-other out. In any event, they apply to only a small subset of individuals those with low earnings and a partial history of contributions. To ascertain the impact of the MPG on their actual work and pension propensities would require detailed individual-level data on accumulations and pensions that our data set does not provide. We reserve this topic for further investigation using a different data set and focus on the broader incentives outlined above, in this paper. Workers newly entering the labor force were required to join the new system. Older workers were given the option to switch, with recognition bonds (bonos) compensating them for their contributions to the old system. Switching propensities were high and inversely correlated with age. By % of all contributing workers were in the new system, including most workers under age 50, and by 2002 the majority of pensioners under age 70 were in the new system (Palacios and Whitehouse 1998; Acuna and Iglesias 2001; SAFP 2003). Pension age was gradually raised for those in the old system and immediately raised for the new system. Initially workers could start their pensions at age 65 for men, 60 for women. In 1987 a new regulation was adopted that allowed workers to start accessing their retirement funds early, when their accumulations reached a specified threshold. For most of the period analyzed in this study, the threshold was 50% of their 11

13 own average wage and 110% of the minimum pension guarantee (MPG); but in 2004 the government decided to increase this gradually to 70% and 150%, respectively. 5 Once the pre-conditions are met, early withdrawal is feasible. At that point, keeping money in the account and making further contributions becomes voluntary, even if the individual continues working. There were no early retirees in the new system before 1988, but by 1992 payouts for early retirement exceeded payouts for normal retirement and by 2002 the number of early retirees was 50% greater than that of normal retirees. (James, Martinez and Iglesias, 2004). Thus, (1) new pre-conditions for early retirement were more restrictive in the new system but (2) early withdrawal continues to be feasible and (3) the decision about when to start the pension is separate from the decision about when to stop work. Hypotheses about worker retirement behavior Propensity to become a pensioner in the new system. We hypothesize that once workers become eligible for pension withdrawal they will continue to keep their money in their social security accounts only if the rate of return earned exceeds their marginal time preference (i.e. they wish to continue saving), also exceeds what they could earn on investments outside the accounts, and this differential is large enough to compensate them for the inflexibilities and non-liquidities that they face in the system. These inflexibilities limit the way money can be invested and the timing and mode of withdrawals. 6 This is a source of disutility, in particular, to individuals who have little savings outside the system and are credit-constrained. We would expect such individuals to become pensioners and stop contributing as early as they can, unless they can earn above-market returns in the social security system. The main source of above-market returns in Chile is the tax system: investment earnings in the accounts are not taxed and additional contributions are tax-deductible, but 12

14 withdrawals are taxed. These tax incentives are relevant mainly to high earners (who are also least likely to be liquidity-constrained). Low earners, too, may postpone pensioning until age 65 in order to qualify for the full MPG. Workers in the middle, who are not strongly influenced by tax incentives, are expected to start their pension as soon as possible to minimize inflexibilities and non-liquidities. Their age of withdrawal is mainly determined by eligibility rules. This means that the pension decision is exogenous to the work decision in the new system. Nevertheless, we expect age of pensioning to be delayed by the reform, mainly because of more restrictive eligibility conditions. This delay should come to an end around age 65, at which point everyone is eligible to start withdrawing. 7 Impact of pensions on labor supply: the liquidity effect. We expect that pensioners are more likely than non-pensioners to retire from the labor force, as pensioners have access to an alternative source of income from their social security wealth. In that case, the reduced propensity to become a pensioner just predicted should increase the aggregate labor force participation rate of older workers until age 65. Impact of payroll tax on work decisions: substitution effects. Both pensioners and non-pensioners should have a substitution effect toward work stemming from the closer link between benefits and contributions. But a much stronger substitution effect applies toward pensioners, who do not have to make any further contributions, in the new system. This exemption from the pension payroll tax will increase the net wage of pensioners. 8 (The implication is that actuarial equivalence of contributions and benefits still leaves a substantial tax component). The reform should therefore have a much larger impact on labor force participation rates of pensioners than non-pensioners. Summary of hypotheses. Therefore, we would expect the following impacts on work and pension probabilities: 13

15 1. The reform will postpone age of pensioning until age 65, partly because of actuarial adjustments but mainly because of tightened conditions for early retirement. 2. The reform will increase the labor supply of older workers, both because it a. decreases age-specific pension probabilities, therefore liquid wealth; and b. increases the net wage, especially among pensioners. 3. Non-pensioners will have higher work propensities than pensioners because of liquidity constraints. But the greatest increase in labor force participation will occur among pensioners, because of their exemption from the pension payroll tax. 4. These effects should be greater for cohorts who were younger on the date of the reform, since older workers were less likely to switch to the new system. The effects should grow larger for successive cohorts who have a larger proportion of new system members, more time to take the new incentives into account in their life cycle planning, and greater opportunity to learn from the behavior of their peers in previous cohorts who work longer. Eventually, with everyone in the new system and full adaptation of behavior to incentives, a new steady state should be reached with higher participation rates. However, Table 1 (based on sources outside our sample) shows that this new steady state has not yet been reached among workers or pensioners, so we are observing the transition stage during the time span covered by our data. The data turn out to be consistent with these predictions. II. Data and Aggregate Results The key data source for the analysis is the Universidad de Chile Greater Santiago Area Encuesta de Ocupación, a household survey representative of the capital city and its surrounding metro area, of ,000 households, that has been collected since 1957, 14

16 with the latest available data from This survey is collected four times a year to measure the rate of unemployment, but the June survey also contains data on income sources, both labor and non-labor. For all members of the household, we know their gender, schooling, marital status and number children, as well as their current wage, pension and other income. We link our data set to variables that measure the macroeconomic conditions for each year of the survey, such as unemployment or growth rates. The chief advantage of these data is that they allow us to construct synthetic cohorts and follow their employment and retirement trajectories, both before and after the pension reform. To construct these synthetic cohorts we use data on 31,547 males who were born between 1900 and 1950, and were 50 to 70 years of age at some time between 1960 and The first 21 years were pre- and the second 24 years post-reform. This data set, however, has several shortcomings. It is not a longitudinal data set. Nor does it give retrospective data on earnings, pension accumulations or age of retirement. 9 For workers and other non-pensioners we do not know whether they are contributing currently or in prior years or, indeed, whether they are covered by any system. At any given point in time, about 70% of the labor force contributes. Most individuals have contributed at some point during their working lives, but some of them only sporadically (Arenas, Behrman and Bravo 2004; Berstein, Larrain and Pino 2005). Taking the proportion of men age who collect pensions as an ultimate indicator of coverage, this percentage has been 70-75% since Thus some of our non-pensioners are barely in the social security system and will never collect a pension. Since individuals who don t belong to any system or have only marginal affiliation are unlikely to be affected by the reform, our inability to remove them from the data set leads to an underestimate of the impact of the reform among those who do belong. This understatement applies particularly to non-pensioners. We would find a 15

17 bigger impact if we could distinguish those who are regularly contributing affiliates, and if regular contribution rates were to increase in the future the total impact of the reform would be greater than our estimates, for these reasons. By definition all pensioners belong to some system, so this underestimate of labor supply effects is absent among pensioners. However, additional sources of bias apply to them. We do not know when they started their pension or if they did so under new or old system rules. Nor do we know whether the pensions are due to normal old age, early retirement, disability, survivors, minimum pension guarantee or social assistance. Until 1992 the majority of pensioners in the new system were disabled and survivors (but some of the disabled were under age 50 and most survivors were women, both of whom are excluded from this analysis). There were few normal old age pensioners because most workers over age 55 in 1981 stayed in the old system and there were few early retirees because early retirement was initially not permitted. After 1992 an increasing proportion of retirees were old age or early retirement pensioners from the new system (James, Martinez and Iglesias 2004). 10 Our inability to separate out individuals who are new system affiliates and not disabled means that we underestimate the impact of the reform on new system pensioners, although not as much as for new system non-pensioners, since many mon-pensioners do not meaningfully belong to any system. Fortunately, all these biases lead us to underestimate reform impacts. Any impact that we find is probably smaller than the actual behavioral effect on new system affiliates. Summary of aggregate data As a first descriptive step, we look at the aggregate data without controlling for individual or macroeconomic effects. We confine our analysis to older males (over age 50). Females behavior would be important so study, but females are likely to be receiving survivors benefits, for which a different model would apply. In addition, our 16

18 data do not allow us to distinguish between old age and survivors benefits. We (1) create synthetic cohorts of individuals born in different years, from and follow their labor force participation rates and pension probabilities as they age, comparing these profiles between pre- and post-reform cohorts; 2) observe labor force participation rates by age group, as these age groups are observed in different years before and after the reform; and 3) construct approximate hazard and survival rates curves for the pre- and post- reform groups. We define post-reform cohorts as those born after 1931, who therefore reached age 50 after 1981, and were likely to be affiliated with the new system. Figures 1-7 summarize the aggregate picture. 1. In prime age all these cohorts had very high labor force participation rates, exceeding 95%, and these rates fall after age 50 for all cohorts. But work propensities fall at a much slower rate for post-reform than for pre-reform cohorts. (Figure 1). 2. Consequently, labor force participation rates of older men gradually increased for successive cohorts after the reform, contrasting with their decline before the reform (Figure 2). 3. Participation rates are much lower among pensioners than non-pensioners, a response to the absence of liquid social security wealth among the latter. However, the postreform increase was concentrated in pensioners, so the gap in their work propensities relative to non-pensioners narrowed (Figure 3). For example, among pensioners age 55 to 59, participation rates trebled between 1982 and 2004, from 14% to 42%, while work propensities among non-pensioners were fairly stable, at 90-94%. As a result, non-pensioners in this age group were 6.5 times as likely to work as pensioners in , but by this ratio had narrowed to At the same time, the probability of becoming a pensioner before age 65 fell dramatically. Pension probabilities for age were at a 40-year low in

19 At the same time, pension probabilities over age 65 were close to their historic highs, indicating that this was an effect of postponement, not reduced coverage (Figure 4). 5. We calculated the approximate survival rates in the labor force for older workers, by age group, for pre- and post-reform cohorts (born before and after 1931) and found they were much higher post-reform, up to age 69 (Figure 5) Labor force survival rates are much higher for non-pensioners, but the rise in survival rates post-reform is concentrated in pensioners, so the two grow closer together (Figure 6). 7. The hazard of becoming a pensioner falls after the reform for workers in their 50 s and early 60 s, due to tightened eligibility conditions. Consequently, a higher proportion of individuals survive as non-pensioners until age 65, at which point everyone can start withdrawing his pension (Figure 7). The postponement of pensioning and the increased work propensities of pensioners are both strongly consistent with our predictions of the impact of the reform. They constitute two channels for the large increase in aggregate labor force participation rates already described. III. Probit Estimation of Retirement and Pension Probabilities Do other forces besides the pension reform account for these effects such as changing characteristics of the labor force or macro changes in the economy? Unemployment rates rose steadily in Chile from the early 1970 s to the mid 1980s, which may help account for falling participation rates before the pension reform. The restructuring of the economy that took place during the late 1970 s and early 80 s led to a period of prolonged growth that might have encouraged labor force participation after 18

20 the pension reform. At the same time, schooling levels rose sharply; this might increase wage rates, which could increase labor force participation, but could also increase wealth and pension eligibility, which might reduce participation among older individuals. To investigate the relative importance of these factors we use probit analysis to estimate retirement and pension probabilities for individuals aged 50-70, as a function of the social security reform, controlling for individual characteristics and macro-economic year effects. We start at age 50 because we are interested in the behavior of older workers, and we stop at age 70 because our sample, which ends in 2004, includes only a small number of new system affiliates over age 70. As discussed above, we deal with the stock of workers and retirees. We are unable to distinguish among people who never worked, those who worked but retired in past years and those who retired in the current year. We do not know whether individuals worked and retired under new or old system rules or were not affiliated with any system. Since we are unable to make precise distinctions at the individual level, the following results understate any behavioral effects of the reform on new system affiliates. Despite this understatement, we find that the reform decreases pension probabilities and increases labor force participation of older workers, even when individual characteristics and macro-economic variables are controlled. Variables and statistical model Our dependent variables are the pension probability (PEN i ) and the labor force participation rate (LFP i ), which we model as a function of individual characteristics (Xi), time-related macro-economic characteristics (E t ) and cohort-specific reform indicators (Rj) defined below. We use a probit model in each case, defined as Pr(PEN=1 X,E,R) = _(X_ L +E_ L +R_ L ) (1) 19

21 Pr(LFP=1 X,E,R) = _(X_ P +E_ P +R_ P ) (2) where _ is the standard cumulative normal probability distribution, X_ L +E_ L +R_ L is the probit score for the PEN probit model, and X_ P +E_ P +R_ P is the probit score for the LFP probit model. In both cases, the model estimates a set of parameters that measure the effect of covariates on the probit score, and thus on the probability of PEN or LFP. Among the estimated parameters, we are particularly interested in _ L and _ P which allow us to measure the effect of reform on the probability of pension or the probability of participation in the labor force, given the values of the other variables in the probit score index. See Table 3 for the list of covariates and their sample means. Individual characteristics (Xi) included in the reduced form analysis are: age of individual (continuous variable for workers age 50 and over, with the coefficient allowed to vary for those over age 60 and 65 in some specifications). Pension and participation probabilities are very stable in prime age. Older age (over age 50) is expected to increase pension probability due to greater likelihood that eligibility requirements will be met. It is expected to decrease labor force participation rates due to greater probability of pensioning (hence having an alternative income source) and lower productivity (hence lower wage offer). When pension status is controlled, the negative impact of old age should become smaller. education (dummies for <6 years and >12 years): This is used as a proxy for permanent income and wealth. We expect that low earners (as proxied by low education) are less likely to be covered by the social security system and, if covered, will find it more difficult to meet the pre-conditions for early pensioning. The illiquidity of their social security wealth and the absence of non-social security wealth should lead them to 20

22 be credit-constrained and to stay in the labor force up to age 65, at which point everyone is eligible to start their pension. In contrast, the non-pension wealth of those with high education and their easier access to a pension should lead them to pension early and to retire from the labor force due to a liquid wealth effect. But this effect on work is counteracted by their higher wage rates and, after the reform, by the exemption of pensioners from the pension payroll tax both inducing substitution effects in favor of participation. On balance, we expect that individuals with high education are more likely to become early pensioners. The net impact of higher education on participation rates should be more positive when pension is controlled. The positive impact of the reform on work propensities of pensioners should implicitly affect individuals with higher education, disproportionately. real household income minus the person s own wage and pension, per capita (we transform the nominal value pesos into constant values using the consumer price index): Larger household income may raise the individual s tax bracket, hence decrease the probability of early pensioning; but it enables the choice of greater leisure financed by other sources, hence it may increase the probability of retirement from the labor market. marital status and number children under age 18: married men, especially those with children, may continue working later because of larger family consumption demands. They may also pension later because (1) in the new system they are required to purchase a joint pension which reduces the amount by about 15% and therefore makes it more difficult for them to meet the eligibility conditions for early withdrawal; and (2) the expected value to them of disability and survivors insurance is greater than the premium, because of community rating. 12 difference between own age and age of spouse: a younger wife may lead the husband to choose a more extended period of wage income. Additionally, if she works, 21

23 the husband might postpone his own retirement to coincide with her s (Coile 2003, Butler, Huguenin and Teppa 2004). Pension status and amount. Pension status and amount play an important role in this analysis. In both the new and old system, we expect that most individuals will start their pensions as soon as they meet the eligibility conditions. We estimate this probability in our PEN equations. The assumption that workers take their pension as soon as they are able, regardless of work preferences, enables us to treat pension status as an exogenous independent variable in our LFP equations. We expect pension status and amount to have negative effects on participation due to the liquid wealth effect. We also interact pension status with age and reform. Because we expect pension status to sharply reduce participation rates, work propensities of non-pensioners should decrease with age faster than that of pensioners, given that practically everyone eventually retires from the labor force. We expect the reform to have a more positive effect on the participation rates of pensioners, because of the substitution effect stemming from their exemption from the pension payroll tax and also because we know that all pensioners are in the social security system while this is not true of all non-pensioners. Rising benefits have been given as one explanation for falling participation rates in the US (see earlier discussion). In Chile, real pension amounts plummeted during the early 1970 s due to hyper-inflation without automatic indexation. But they recovered steadily during the late 1970 s and more irregularly during the 1980 s and early 1990 s. Cohorts that were born later got larger real pensions, on average (see Figure 10). The rise in pension amounts might have been expected to decrease labor force participation rates during the last 20 years. In that case, the rising aggregate participation rates described above would be even more surprising and would increase the behavioral change that remains to be explained by the reform. 22

24 But pension size did not increase as rapidly as wages, so if workers reacted to a target wage replacement rate rather than absolute pension size this might have led them to continue working after starting their pensions. In that case, our inability to include wage replacement rates would lead to an overestimate of reform impact. Our sample does not include data on prior wages of individuals, so we could not include the individual s wage replacement rate as a variable. However, we created a pseudo-replacement rate that we use in some specifications, by taking the ratio of individual i s pension to mean wage for i s cohort and schooling group in the pre-pension age range On average, this measure moves downward for post-reform cohorts, although to different degrees for different schooling groups (Figure 10). Time-specific macroeconomic conditions (Et): A leading counter-explanation for changes in observed behavior during the 1980 s and 1990 s would be changes in macro-economic conditions in Chile, which were large during The country went through a period of economic chaos and hyperinflation during the early 1970 s, high unemployment during the late 1970 s, its financial system went through a major restructuring in the early 80 s, and it enjoyed a period of prolonged growth that included a compete business cycle from the mid-80 s through The existence of a full cycle is important as it reduces the correlation between the phase-in of the reform and economic growth, and allows us to separate out these effects. We control for these macro-economic changes using rate of unemployment, Hodrick- Prescott filter (HPtrend, a smoothed non-linear representation of long term GDP growth), and real lngdp growth minus HP filter (HPdev). The Hodrick-Prescott filter (Hptrend) is more sensitive to long-term growth trends than short-term fluctuations. The difference between HPtrend and the actual lngdp growth (Hpdev) is a measure of short-term business cycles during

25 (Figures 8A and 8B). On the one hand, a higher HPtrend could have a negative effect on participation rates because of the income effect. It has been argued in the US literature that long-term income growth is a major reason behind the decline in pension age and labor supply of older male workers. On the other hand, greater long-term growth means a higher wage rate, therefore a higher cost of leisure, so the substitution effect works to increase labor supply. Also, higher wage growth leads to a lower replacement rate of final salary in a defined contribution plan, and even in a defined benefit plan that depends on earnings over some long time period. If workers have a target replacement rate, a higher growth trend will therefore lead them to delay pensioning and to work longer, especially in a DC plan. Cyclical deviations from the long-term growth trend (HPdev) should behave much like the unemployment rate, but with the opposite sign because they are negatively correlated. Unemployment in Chile was 8% in 1960, fell to 3% by 1973 and then rose to 23% by During the post-reform period, unemployment went through an entire cycle, falling to 6% by 1995 and then rising to 14% by 2002 (Figure 9). We expect pension probabilities to rise and labor market participation to fall during cyclical downturns in the economy, due to the discouraged worker effect, the greater difficulty older workers may experience in obtaining new jobs when laid off from their existing jobs, and the greater ease of meeting early retirement pre-conditions when periods with 0 wages are averaged into the own-wage base. Both higher unemployment and lower (negative) HPdev signify cyclical downturns. We keep both variables in our probits to increase the controls for macroeconomic conditions, although the high correlation between them may lead to ambiguity about their individual effects. Reform indicators (Rj): 24

26 Reform indicators are the main focus of our empirical analysis. Since we do not know which individuals are covered by the new versus the old systems (versus no system at all), our basic strategy is to measure reform according to the birth cohort of the individual, which proxies the probability that the individual is in the new system (Table 1). Individuals born before 1931, who were over age 50 in 1981, are likely to be old system affiliates and therefore unlikely to be affected by the reform. Starting with the cohort born in 1931, we expect the proportion of new system affiliates to be rising and therefore the impact of the reform on pension and participation probabilities would be rising as well. This will continue until the probability of affiliation with the new system approaches 100% and social norms about retirement have fully adapted to the new incentives. At this point a new steady state is reached, with higher participation rates than the old steady states but no further marginal increases. Thus, we are essentially modeling a transition from one steady state to another. We are uncertain a priori how long the transition period will last. Affiliation to the new system was around 80% for workers born in the 1940 s, who were in their 50 s by By 2002 pensioners in the age group were all born after 1931, and about 60% were new system affiliates (see Table 1). Even when affiliation reaches 100% it is possible that the transition period will continue because learning, changes in work norms and peer group effects operate with a lag. Therefore, our specification allows for the possibility that the marginal impact of reform will disappear, but we do not expect that to happen within the period covered by our data. We use two alternative methods for measuring the effect of reform--one captures single-cohort continuous shifts in the probit score functions, and the other captures discrete shifts for cohort groups in the probit score functions. 25

27 Continuous cohort effects. We use two continuous variables in the Rj set, one defining an underlying cohort trend starting with the pre-reform cohort born in 1916 (Coh) and the other measuring a drift from the underlying trend starting with the post reform cohort--individuals born in 1931 or after (Coh31). Both measure a uniform marginal effect per birth cohort, so the total effect relative to the omitted cohort is obtained by multiplying the coefficient by the cohort year and adding both trends, if postreform.. For all the reasons given above, we expect that Coh31 will have a negative effect on pension probabilities and a positive effect on labor force participation rates. We include a quadratic term, Coh31sqd, which we expect to have a counteracting coefficient as the transition period draws to a close, allowing the marginal cohort effects to decline. In some specifications we interact Coh, Coh31 and Coh31sqd with pension status to test our hypothesis that the change in work incentives was greatest among pensioners. We also interact the cohort terms with a dummy for individuals over age 65, to separate out ages where all affiliates were eligible to start their pensions. Discrete cohort effects. In this case, Rj is a set of dummy variables that are nonzero for specific age-ranges and cohorts. One problem with using continuous cohorts is that we do not have a uniform age distribution for all cohorts. Due to our 2004 cut-off year, later cohorts (those born after 1940) are under-represented in their 60 s and this may bias our results. Secondly, the continuous specification does not allow us to isolate the cohorts and age-cohort combinations for which the reform effect begins and is strongest. Third, cohort effects may not closely follow a particular functional form, such as a linear or quadratic form. To address these issues, we set up seven broader cohort groups and break them into age categories. Table 1 reports the proportion of new system affiliates by cohort group (based on data source outside our sample), Table 2A records the age of the various 26