*Munich Center for the Economics of Aging (MEA) **Hamburgisches WeltWirtschaftsInstitut (HWWI)

Transcription

1 At what age do you expect to retire? How individuals adjust their expectations to increases in the statutory retirement age Michela Coppola* Christina Benita Wilke** *Munich Center for the Economics of Aging (MEA) **Hamburgisches WeltWirtschaftsInstitut (HWWI) Abstract. Population aging poses an evident threat to the financial sustainability of pension systems based on a pay-as-you-go scheme. To cope with this threat, pension systems have undergone numerous reforms in many countries in order to keep people longer at work. One important element of these reforms typically is an increase in the statutory retirement age at which workers are legally allowed to retire. Such a change is likely to not only affect actual retirement behaviour but also retirement plans of younger cohorts. Understanding how individuals adjust their retirement expectations to increases in the statutory retirement age is crucial since this has direct implications for other important decisions, such as on labour supply, engagement in (further) education and of course savings and investments. We consider the 2007 German pension reform that legislated a gradual increase in the statutory retirement age from age 65 to age 67. Our analysis is based on the SAVE survey, a longitudinal study of German households, that directly asks respondents at what age they expect to retire. Using a difference-in-differences approach, we look at the changes in subjective retirement expectations over time and estimate to what extent they can be attributed to the 2007 reform. Our results show that the reform has clearly shifted retirement expectations of the younger cohorts. We also find heterogeneity among respondents, with the better educated having revised their expectations to a greater extent. This result has important policy implications. As individuals with lower educational attainments tend to earn lower wages (thus acquiring lower pension claims) and save less for retirement, the fact that they have been slower in updating their retirement expectations causes concern regarding their income level after retirement. JEL-Code: D03, D1, D84, H55, J26 CONTACT: Michela Coppola Munich Center for the Economics of Ageing (MEA) Max-Planck-Institute for social law and social policy Amalienstr. 33 D Munich coppola@mea.mpisoc.mpg.de Acknowledgements: We thank Michele Belloni, Martin Brussig, Bernard Casey, Mike Hurd, Wolfgang Nagl, Steffen Reinhold, Joachim Winter for their useful comments on earlier drafts of this paper. The usual disclaimer applies. [1]

2 1. Introduction Population ageing poses an evident threat to the financial sustainability of pension systems based on a pay-as-you-go (PAYG) scheme where contributions by the young directly finance benefits for the old. The combination of low fertility rates and substantial gains in life expectancy lead to an increase in the so-called old age dependency ratio, the ratio of people aged 65 years and above to those of working age (15 to 64 years). In addition, many countries experienced a so-called baby boom in the 1950s and 1960s, followed by a so-called baby bust thereafter so that comparatively large cohorts are followed directly by comparatively small cohorts. This will worsen the ratio even further once the baby-boom cohorts reach retirement age, thus adding extra strain to public finances. Increasing the statutory retirement age (i.e. the age at which workers are legally allowed to retire) is one of the reform measures widely recommended in order to cope with these upcoming demographic challenges (e.g. OECD 2006, EU 2012). It is supposed to keep people at work longer, thus alleviating the pressure on public pension finances through an increase in revenues (as individuals keep on paying contributions to the pension system) and a decrease in expenditures (as individuals spend less years in retirement). Early retirement windows, however, are provided in most of the western countries so that an increase in the statutory retirement age does not necessarily imply an equivalent postponement of actual retirement. Thus, in terms of the effectiveness of such type of reform an important question is to what extent people adjust to it. While several studies analysed the effect of increases in the statutory retirement age on actual retirement behaviour (e.g. Berkel and Börsch-Supan, 2004; Hanel, 2010) very little is known about how such a reform affects individuals retirement expectations. 1 The present work sheds new light on the latter link. Focusing on retirement expectations rather than outcomes is important for at least two reasons. First, if public policies aimed at altering actual retirement patterns are to be successful, they have to alter workers' retirement expectations in the first instance. Second, subjective retirement expectations play a key role for other important economic decisions such as decisions on labour supply, engagement in (further) 1 Strictly speaking we should talk about retirement plans rather than expectations as individuals can actively determine their retirement entry, so that this is not an uncertain future outcome. However, retirement decisions may also be influenced by factors over which individuals in fact have no or little control, such as changes in institutional regulations, employment status or health. We thus follow the existing literature and refer to subjective retirement expectations instead of retirement plans. [2]

3 education and saving and investment decisions. Understanding how public policies influence subjective retirement expectations is therefore also crucial for better understanding labour supply and skills shortage developments as well as current saving behaviour and wealth accumulation. In this paper, we refer to the 2007 German pension reform. Germany is not the first country to have legislated an increase in the statutory retirement age. Many countries have already done so, but usually, this measure was only one part of a broader pension reform package. In Germany, on the contrary, the increase in the legal retirement age has been legislated separately about three years after another major package of pension reforms. This allows us to disentangle the effect of the increase in the statutory retirement age as distinct from other changes of the pension system. Furthermore, the German reform provides a quasi-experimental setting as it affected only the insured labour force, leaving other employment groups, such as the self-employed, unaffected. Our analysis is based on data from SAVE ( Saving and old-age provision in Germany 2 ) - a longitudinal survey of German households that directly asks respondents at what age they expect to retire. We look at subjective retirement expectations over time and use a difference-in-differences (DiD) approach to estimate to what extent they can be attributed to the 2007 German pension reform. In a second step we investigate if there is heterogeneity in the response to the reform focussing on the role played by education. Our results show that the retirement expectations of the younger cohorts increased by almost two years in response to the reform. If these expectations are realized in the future, the reform will have been quite successful. However there is some heterogeneity in the response, with better educated individuals revising their expectations to a greater extent. This result has important policy implications. As individuals with lower educational attainments are also those being more at risk of having lower state pensions in the future and also those who are less likely to save for the old-age, the fact that they have been slower in updating their retirement expectations causes concern regarding their well-being after retirement. This paper is structured as follows. Section 2 offers a short description of the institutional settings of the German pension system and the 2007 reform. Section 3 provides an overview of the existing literature on the effect of changes in institutional settings on subjective retirement expectations. In section 4 we present the SAVE data that our empirical analysis is based on and provide some first descriptive results. In section 5 we explain our methodological approach and present the results of our econometric analysis. Section 6 concludes. 2 SAVE stands for Sparen und Altersvorsorge. [3]

4 2. The German public pension system and the 2007 reform The German public pension system covers all private sector employees and parts of the public sector employees. 3 Civil servants as well as liberal professionals (like architects, accountants, lawyers or medical doctors) have their own separate pension system and are therefore not covered by the public pension system. Self-employed usually self-insure, although they may choose to participate in the public system instead. Until the late 1980s, the German public pension system was celebrated for providing a high and reliable level of retirement income at reasonable contribution rates, becoming a model for many social security systems worldwide. However, negative incentive effects of past reforms combined with projected demographic challenges eventually led to financial distress and a pension reform process that started in the 1990s. This process continued also in the 2000s and culminated in the 2004 reform based on the proposals of the Rürup Commission. 4 Being subject to a very controversial debate, the Commission s proposed measure of an increase in the statutory retirement age was legislated only three years later under the 2007 German pension reform. The new law (passed in March 2007) legislated a gradual increase in the legal retirement age from age 65 to age 67, adjusting the statutory retirement age each year by one month from age 65 to 66 for the cohorts born between 1947 and 1958, and then each year by two months for the birth cohorts born between 1959 and For cohorts born after 1963 the new statutory retirement age of 67 will fully apply. 5 Table 1 gives an overview of the new cohort-specific statutory retirement ages resulting from the reform. 3 The German law puts public employees into two categories (namely ordinary employees and civil servants), making a distinction that does not exist in most other countries. While the ordinary employees in the public sector are covered by the public pension scheme, civil servants are excluded. 4 For a detailed description of the German pension reform process, see Wilke (2009). 5 For a detailed description of the 2007 German pension reform and a first assessment, see Bucher- Koenen und Wilke (2009). [4]

5 Table 1: Statutory Retirement Age by birth cohort after the reform 2007 Legal Retirement Age Birth year Increase in months Years Months Source: German statutory pension scheme, ml?nn=46814 In terms of pension benefits, the increase in the statutory retirement age leads to higher pension deductions for individuals who do not adjust their retirement behaviour to the new rule. As the minimum retirement age of 63 persists, with the same adjustment factors applying as before 6,, early retirement remains possible but at higher costs. On the contrary, adjusting to the reform by postponing retirement may even imply higher pensions, as individuals stay longer at work, thus contributing longer to the pension system and accumulating higher pension entitlements. Although the reform directly affected only members of the German public pension system, the other pension systems covering civil servants and liberal professionals have followed and also increased the statutory retirement age for their members. As a consequence, it is only the self-employed for whom nothing changes. 6 For each month of early retirement pension benefits are reduced by 0.3%. [5]

6 3. Existing empirical evidence on retirement expectations Since the seminal works of Bernheim (1989, 1990), retirement expectations have been the subject of an increasing amount of research in recent years. There seems to be general consensus that retirement expectations are quite accurate with regard to the expected outcome. Chan and Stevens (2004), for example find a strong correlation between retirement expectations and realizations, while Benitez-Silva and Dwyer (2005) find that retirement expectations are consistent with the rational expectations hypothesis. Several studies have estimated the determinants of retirement expectations directly, finding that they vary with individual circumstances like health or occupational status in a plausible manner (e.g. Chan and Stevens, 2004; Cobb-Clark and Stillman, 2009). Very few papers, however, have looked at the effect of policy changes on retirement expectations. Michaud and van Soest (2008) analyse for the US how retirement expectations are affected by the repeal of the earnings test at ages above the normal retirement age that quasi taxed away earnings later in life. They find a substantial increase in the reported probability of working after age 62 for those workers whose marginal wage rate increased because of the repeal. The effect of pension reforms on retirement expectations has so far only been analysed in the context of the Italian pension reforms of the 1990 s. These studies are closely related to our research and we will therefore summarize them in more detail. The first study dealing with this issue is Brugiavini (1997), which looks at the shift in retirement expectations between the years 1991 and 1993 (after a major bill of reforms was passed in 1992) finding a surprising decline in the expected retirement age. She argues that the debate on early retirement that the reform initiated, shifted the attention of the respondents to this issue, so that after 1992 they started to think of their retirement age as the early retirement age and not any more as the normal retirement age, an effect known in the literature as recognition effect (Cagan, 1965). This interpretation is contested by Mastrogiacomo (2004), who points out two sources of bias not taken into account in Brugiavini's work: the bias due to sample attrition and that due to don't know answers. He finds that over the time span , the reforms indeed induced individuals to postpone their retirement plans by more than two years. He recognizes the existence of a recognition effect, but he finds that it can be ascribed only to those who were actually not involved in the reform. He also finds that the reforms, particularly the first one introduced in 1992, increased uncertainty among Italian workers, although the results remain unchanged after the [6]

7 model is corrected to take this into account. Finally, Bottazzi et al. (2006) estimate the effect of the Italian pension reforms on households' retirement expectations by comparing expectations before and after the whole reform process (that is, before 1992 and after 1999). They find that on average the expected retirement age increased by about two years for men and three years for women as a result of the whole set of reforms. While these studies reveal that individuals indeed seem to adjust their expected retirement ages in the anticipated direction, these effects cannot be attributed solely to increases in statutory retirement ages as various changes of the Italian pension system occurred quasi simultaneously Data and possible expectation measures The analysis in this paper is based on the survey saving and old-age provision in Germany (SAVE), a representative, longitudinal survey that concentrates on saving and asset allocation choices of German households. The panel started in 2001 and consists of about 3,000 households, which, since 2005, have been surveyed every year. For the purpose of this paper, we use the waves from 2005 to The SAVE dataset is particularly well-suited for the purposes of our study for three reasons. First, there is no other German survey that directly asks about retirement expectations in a longitudinal setting. 9 However, we need such a setting in order to observe how reported retirement expectations have evolved over time. Second, SAVE does not only collect extensive information on all aspects of the household's balance sheet, it also offers extensive information on actual health and other relevant social and psychological conditions. Third, in contrast to other surveys on retirement or aging that typically focus on older workers (like e.g. the HRS for the US, ELSA for the UK or SHARE for Europe), SAVE is characterized by the predominance of those birth cohorts that will be fully affected by the 2007 German pension reform. Of course, younger respondents face bigger uncertainty concerning their retirement, so that their answers are likely to undergo bigger changes over time and to be less representative of the actual behaviour. However, in this work we are not interested in 7 So if, for example, the legal retirement age increased after the reforms in the same way for all the workers who started working before 1995, irrespectively of their years of contributions, the change in the pension award formula makes a distinction between workers with more or less than 18 years of contributions in 1995 (for details see Bottazzi et al. 2006, Table 1). 8 See Börsch-Supan et al. (2008) for a detailed description of the dataset. Essig (2005) and Schunk (2006) provide further technical details. 9 The GSOEP, a German survey focused on households socio-economic conditions, merely asks its participant for the probability to retire within the next two years. These answers are thus not useful for the purpose of our analysis. The expected age of retirement has been asked only once, in the 1987 wave ( [7]

8 the match between expectations and outcomes, but rather in the expectations themselves, as we want to see how these adjust to increases in the statutory retirement age thereby affecting in return other important decisions in life such as on labour supply, education and of course savings and investments. In SAVE, respondents are asked to answer the following question: At which age do you expect to retire or to claim retirement benefits?. 10,11 Subjective retirement expectations in SAVE are therefore expressed as point estimates. How to correctly elicit expectations is still a matter of debate in the literature. A common critique to the approach of asking respondents to report a single number is that it is not clear which moment of the underlying distribution of possible outcomes is actually reported, making it difficult to correctly interpret the answers. Manski (2004) therefore strongly argues for asking individuals directly which are the chances of a certain event (e.g. to retire at a certain age) to happen. Although in the last years this way of eliciting expectations has been increasingly used in social surveys, the average respondent is not necessarily familiar with probabilistic concepts so that the reliability of the answers may be questioned. 12 In our specific case of retirement expectations, non-sophisticated individuals rather might have a specific retirement age in mind than a distribution of probabilities. Studies using point estimates for retirement expectations show that answers vary quite reasonably with individual characteristics and can serve as strong predictors of actual retirement behaviour (e.g. Disney and Tanner, 1999; Haider and Stephens, 2007). The first can also be shown for the SAVE data. 13 Thus, for the purpose of our study, point estimates provide a valuable source of information. 10 In German: In welchem Alter werden Sie voraussichtlich in Ruhestand gehen bzw. das Alterseinkommen beziehen? Respondents can express their age as an integer number only. There is no possibility to report years and month (e.g. 65 years and 7 months). We only use the answer by the respondent on his- or herself and not the answer on behalf of his/her spouse since we cannot treat this answer as if it was given directly by the partner. 11 As pointed out in Hanel (2010), retirement entry, labour force exit and the claiming of benefits are not necessarily interchangeable terms: indeed she finds a discrepancy between the age at which individuals leave the labour force and the age at which they start receiving pension benefits. Given the wording of the question, we argue that respondents in SAVE report the age at which they expect to claim their benefits. 12 Dominitz and Manski (1996) find for example that 40 percent of respondents provide answers incompatible with the (previously elicited) median of the subjective distribution of future income They also report a 21 percent item nonresponse to subjective income expectations (Dominitz and Manski, 1997). Similarly, van Santen et al. (2012) find that one third of the respondents are not able to provide consistent answers to probabilistic questions on retirement income replacement rates, and that removing these individuals from the sample leads to biased results. Another frequent problem with this kind of expectation measure is the tendency for respondents to provide rounded answers, typically at 0, 50 and 100% (Kleinjans and van Soest, 2010). Manski and Molinari (2010) investigate this problem in more detail and indeed find some heterogeneity in answering patterns. 13 The SAVE panel still is too short in order to check for realization effects. [8]

9 5. Methodological approach and descriptive statistics To single out the causal effect of the 2007 German pension reform on individual retirement expectations, we rely on a difference-in-differences (DiD) approach. Operationally, we thus pool together all the data for the periods before and after the reform and specify the following reduced form: y i = α 1,i + α 2 X i + β 1 TREAT + β 2 POST + β 3 TREAT*POST + ε i (1) where y i is the expected retirement age (ERA) for individual i, X i is a set of relevant variables which may affect individual s ERA, TREAT is a dummy variable equal to one if the individual belongs to the group affected by the treatment (in our case, by the increase in the statutory retirement age) and POST is a dummy variable identifying the period after the reform. The term of interest is the interaction between POST and TREAT: the coefficient β 3 measures any change in the ERA of the treated group in the period after the reform. The basic idea of the DiD estimator is to compare over time the changes in retirement expectations of individuals affected by the reform (the treated group) with those of individuals who are not affected (the untreated or control group): the change in the outcome of the untreated should identify any temporal variation that is not due to the policy. Therefore, once we control for all possible observable characteristics that may determine a difference in the outcome, any remaining difference in ERA between the two groups can be attributed to the reform. We further assume that any unobservable differences between the two groups remain constant over time for the period of our analysis. To put it differently, the critical assumption is that the control group is not affected by the reform and that its observed trend perfectly mirrors the evolution in the ERAs of the treated in the case the reform had not taken place (common trend assumption). The institutional setting of the 2007 German pension reform offers several possible control groups. Given the slow phasing-in of the process, one natural approach would be to compare the retirement expectations of employees of different birth cohorts that are affected differently by the reform. However, there are three problems with this approach. First and foremost, retirement expectations in SAVE are only reported as integer numbers (see footnote 9), thus making it impossible to record month wise changes in the ERA that may reflect the month wise phase-in of the reform. Second, since only those respondents that currently work are asked about their retirement expectations, there may be a selection bias in the control group towards individuals with a higher attachment to work that also report a higher expected retirement age. [9]

10 Finally, there is also likely to be an attrition bias since in the latest waves of the survey individuals born before 1947 already have reached their minimum retirement age and may thus drop out of the sample. The common trend assumption would thus be violated. We therefore choose a different approach: we restrict our analysis to the cohorts born after 1964, whose new statutory retirement age is exactly 67 years and we use selfemployed individuals who are not affected by the reform as a control group. Thanks to this very young sample, we can exclude both the selection bias towards individuals with a higher attachment to the labour force and thus higher subjective retirement expectations as well as the attrition bias because respondents retire. In addition, this young sample allows us to compare two groups that are at the same stage of their lifecycle, making the two groups more homogenous with respect to the time horizon they face. Furthermore, as self-employed are by definition not affected by the reform, the assumption that their expectations also remain unaffected by it is more likely to hold. This assumption is of course not testable and it could be argued that the selfemployed have been indirectly affected by the reform as well. 14 However, this would only be problematic in case of a downward revision of retirement expectations by the self-employed, for which we do not see any plausible reason. In contrast, an upward shift of retirement expectations by the self-employed would simply make our test less powerful in that it would bias downward the estimated effect for the employees In the following we provide some first descriptive statistics of our data in order to see whether these support our underlying assumption. First of all, we look at the composition of the samples of employees and self-employed. Their main characteristics should be quite stable over time.. Table 2 reports some summary statistics, separately for the two groups. The sample is made up of 2,824 observations, distributed over 5 years. The self-employed represent about 8% of the total sample. The distribution of the main characteristics shows that the structure of the samples of the treated and control group is fairly stable over time. 14 For example, as the employment biographies of the younger workers are becoming increasingly more discontinuous, with frequent transitions into and outside self-employment (Simonson et al. 2011a, 2011b), it could be argued that the self-employed incorporated nonetheless the increase in the legal retirement age of the employees in their own expectations, as they anticipate a transition into employed work in the future. Similarly, the new legal retirement age might affect the profitability of the businesses owned by the self-employed (if for example it is now more difficult to get rid of older, unproductive workers), thus affecting their retirement expectations as well. [10]

11 Table 2: Sample characteristics for self-employed and employees (birth cohort >1963) Self-Employed Employees Total Total Female RP 39.4% 36.8% 42.6% 37.1% 31.0% 37.7% 51.7% 51.7% 53.5% 51.6% 53.7% 52.4% Age < % 26.7% 20.5% 15.8% 11.5% 19.6% 36.1% 40.6% 38.4% 32.4% 26.3% 35.4% % 51.9% 53.2% 54.2% 57.1% 55.1% 54.7% 43.9% 41.1% 39.7% 39.2% 43.1% % 21.3% 26.2% 30.0% 31.4% 25.3% 9.2% 15.6% 20.5% 27.9% 34.5% 21.5% Mean Median RP with a partner 77.3% 82.5% 77.3% 85.7% 77.9% 80.4% 61.1% 62.8% 62.1% 64.3% 66.0% 63.2% Secondary Education Basic 24.4% 17.7% 22.2% 25.8% 24.0% 22.1% 24.8% 20.4% 21.3% 18.8% 18.6% 20.6% Middle 35.8% 39.4% 35.0% 41.9% 41.3% 38.5% 46.7% 44.3% 43.4% 44.5% 45.2% 44.6% High 39.7% 42.8% 42.9% 32.3% 34.7% 39.3% 28.6% 35.3% 35.3% 36.7% 36.1% 34.8% Post-secondary and tertiary education None 11.1% 10.0% 18.4% 8.1% 5.2% 11.1% 15.2% 9.2% 9.9% 12.2% 11.9% 11.3% Vocational training 76.9% 78.8% 71.3% 81.6% 82.9% 77.9% 73.3% 74.4% 74.3% 72.6% 74.9% 73.9% University 12.0% 11.1% 10.4% 10.3% 11.8% 11.0% 11.5% 16.4% 15.9% 15.2% 13.1% 14.8% Income and Wealth Net monthly income 3,032 2,650 2,436 2,783 2,908 2,718 2,225 2,105 2,147 2,218 2,337 2,193 Median 2,800 2,500 2,300 2,300 2,500 2,500 1,880 1,993 2,000 2,100 2,200 2,000 Net financial wealth (1,000) Median Obs Source: SAVE , imputed data, weighted results. Second, we investigate whether the common trend assumptions holds. There is no formal test for this but a graphical analysis of the time trend in the outcome variable for the treated and the controls can provide some useful insights. Figure 1 plots the average subjective retirement expectations for the self-employed and the employees in the years 2005 up to Subjective retirement expectations start from a very similar level for both groups in 2005, and develop in an almost parallel fashion in the years after the bill of reform was passed (2008 and 2009), thus providing some support for the underlying assumption at least for those years. Further support for the common trend assumption comes from the results of a placebo regression run over two years when there were no changes in the statutory retirement age. To this scope we use the years 2008 and 2009 respectively as the period before and after (see table A in the appendix for the results). As expected, the term β 3 in this regression is not significant, indicating that in absence of an institutional reform, the ERAs of employees and self-employed follow a common pattern. [11]

12 Figure 1: Subjective retirement expectations for self-employed and employees ( ) Source: SAVE Data , imputed data, weighted averages. Two further aspects are worth noting. First, the expectations of both treated and controls exhibit a peak in For the self-employed however this appears to be an isolate episode, as in the years after the reform, retirement expectations evolve back to their pre-reform level, while for the employees the increase seems to be longerlasting. To handle this temporary increase, the year 2007 is therefore omitted from the analysis. Second, the expectations of the employees start increasing already in 2006, probably because of an anticipation effect. Indeed, the policy change was introduced by a long debate so that when the law was approved in 2007 the increase in the legal retirement age was hardly a surprise, although it was not clear ex-ante which groups would be affected to what extent in detail Since, in the heads of the people, the reform cannot be solely attributed to a single year, we construct a special index for the intensity of reform discussions in the media. Figure 2 shows the number of Google web searches for two key terms retirement at 67 (Rente mit 67, 15 blue line on the graph) and retirement age (Renteneintrittsalter, red line on the graph) relative to the total number of Google searches over the period The interest of google users for the topic retirement at 67 emerges first at the end of 2005, peaks in the first quarter of 2006 and again between the end of 2006 and spring A similar trend (although on a higher level, especially for the period before 2005) emerges for the search term retirement age. 16 In the light of 15 Rente mit 67 is the term commonly used to refer to the 2007 German pension reform and more generally to the increase in the statutory retirement age. 16 It is interesting to note, that the number of searches for the term retirement age started increasing [12]

13 these considerations, we also omit the year 2006 from the analysis as belonging to the transitional phase and we compare individuals retirement expectations in the prereform period 2005 with those in the post-reform period of 2008 and Given that the data for 2005 has been collected in spring, it appears safe to assume no anticipation effects at that time. That leaves us with a sample of 1407 observations for the analysis. Figure 2: Number of Google web searches for the terms Rente mit 67 and Renteneintrittsalter relative to the total number of searches done on Google ( ) Source: Google insights for search Retrieved on Last, another important issue that has attracted considerable attention in previous analyses of subjective retirement expectations is how to interpret don t know or missing answers to the question on subjective retirement expectations. 17 In the SAVE survey, non-response rates to this question are quite low: on average, less than 5% of the respondents do not answer the question. 18 Some studies have found that these (missing) answers are still informative, representing rational responses by those facing greater uncertainty over their future behaviour (Disney and Tanner, 1999; Mastrogiacomo, 2004). Benítez-Silva and Dwyer (2002) find that those not reporting expectations are structurally different from other respondents and that the induced selection bias is significant. As a consequence, rather than throwing away these observations, we use a two-step Heckman selection model approach, which allows us to take care of possible biases induced by a non-random distribution of the missing answers. 19 As exclusion again in 2009, when the 1947 cohort (i.e. the first one being directly affected by the reform) approached the minimum retirement age of 63 year. 17 E.g. refer back to our brief scatch of the Brugiavini and Mastrogiacomo papers in section Mastrogiacomo (2004) reports similar nonresponse rates in the Italian Survey of Household Income and Wealth. 19 Missing information on the other variables is imputed using an iterative multiple imputation procedure based on a Markov-Chain Monte-Carlo method (for further details, Schunk, 2008; Ziegelmeyer, 2009). The inference based on such a multiply imputed dataset is more reliable, as the item-nonresponse bias is reduced. Furthermore, the increased number of available observations improves the efficiency of the estimates. Five imputed datasets are used for the analysis and the results are combined using Rubin s method (Rubin, 1987; 1996). [13]

14 restrictions (i.e. as variables affecting the probability of not answering the question on the ERA without directly affecting the reported ERA) we use several indicators taking value 1 if the respondent also did not answer one of the other expectation questions on the subjective economic situation, future unemployment or subjective life expectancy. The idea behind the choice of these variables is that individuals who are not able to formulate expectations in these areas are facing higher uncertainty and are therefore more likely to be unsure about their ERA as well. At the same time, not being able (or willing) to answer questions on expectations in other life domains should not affect the ERA if reported. 6. Results Our choice of explanatory variables is mainly driven by the consideration of the factors driving actual retirement behaviour and subjective retirement expectations. Besides controlling for the usual socio-demographic characteristics (gender, age, marital status and educational level 20 ) we include the household s economic situation 21, the employment status and employment history of the respondent as well as self-assessed current health status and expectations and satisfaction with the current job. In order to check for possible heterogeneity in the way individuals adjusted their retirement expectations, we estimate a second model where the interaction term TREAT*POST is further interacted with educational attainment. With this second specification it is possible to test if and to what extent the revision of the ERA varies with education. Furthermore, following the advice of Mastrogiacomo (2004), we also include selectivity dummies (i.e. variables taking value 1 if the respondent participates in year t and at least once more after year t) to check for possible biases in the estimates due to panel attrition. 22 As it is shown in the second last section of Table 3, the selectivity indicators are not jointly significant. We can therefore assume that there is no attrition. 20 In line with several studies focussed on Germany, we use two sets of variables to measure educational achievements. The first includes variables measuring the highest secondary school leaving certificate; the second set includes variables measuring post-secondary and tertiary school achievements. 21 Logarithm of the household s financial wealth; logarithm of the household s net monthly income; dummy variables if the household holds a state subsidized old-age insurance, a non-subsidized oldage insurance, an occupational pension or real estates. 22 In the following regressions, the reference group is therefore made up of all the individuals observed only in the survey 2005 (all the selectivity dummies equal to zero). [14]

15 The values of the Wald test reported at the bottom of Table 3 indicate selectivity due to missing answers to the ERA question, as the null hypothesis that no correlation exists between the selection equation and the equation of interest can be rejected at the 5% significance level. In order to estimate the size and the direction of the bias induced by omitting the missing answer we estimate a separate ordinary least square (OLS) model. Table A2 in the appendix shows the results. The differences in the coefficients are minimal and the estimated effect of the reform on the ERAs of the employees remains basically unchanged. 23 Before looking at the estimates of the effect of the reform 2007 on respondents ERAs, it is reassuring to notice that the reported values vary reasonably with individual covariates. In line with other studies on the determinants of subjective retirement expectations, we find that women expect to retire earlier than men, the difference being a bit more than half a year (7.1 months). We also find a positive effect of education: individuals with the German Abitur (the highest secondary school leaving certificate, earned after 13 years of schooling) expect on average to retire almost 0.6 year (about 7 months) later than individuals with the lowest certificate (Hauptschulabschluss, earned after 9 years of schooling). Similarly, having a university degree shifts the ERA by more than one year (1 year and 4 months). Finally, we find an inverse relation between the household s income and the ERA: an increase of 1% in the net monthly household s income is related with a reduction in the ERA of about 8 months. Coming to the financial situation of the households, it turns out that the composition of the financial assets rather than the amount accumulated is relevant in shaping individuals retirement expectations. In fact, while the coefficient on the logarithms of the financial assets is not significant, holding non-subsidized private old-age provision contracts significantly lowers the ERAs by about half a year. It is interesting to note that holding state subsidized private pensions (usually known in Germany as Riester pensions ) does not affect the reported ERA. That is probably due to the stricter regulation which applies to the subsidized products. In particular, the fact that the pension benefits cannot be paid before the age of 60 and that the savings must be disbursed in some type of annuity (with only a limited amount that can be paid out as a lump-sum payment) make the Riester pensions less suitable to finance an early retirement. On the contrary, non-subsidized contracts allow more flexible arrangements and might better suit early retirement plans. None of the indicators for the financial situation of the household is significant at 23 The predictions of the Heckman model return an average an ERA equal to years, the prediction of the OLS model is years. [15]

16 conventional levels. However, a 1% increase in the household s net monthly income is related with a decrease in the expected retirement age of 0.4 years (about 5 months). The coefficient of interest is that on the interaction term between the dummy for the period after the reform and the dummy identifying the treatment group. This coefficient is positive and significant indicating that the reform induced an increase of 1.6 years in the retirement expectations of the male employees born after The results in Model 2 reveal some heterogeneity in the way ERAs changed after the reform. While we find no significant differences among individuals with different types of post-secondary educational attainments, we find stark contrasts among individuals with different types of secondary school certificates. In particular, while individuals with the lowest attainment (Hauptschule) did not revise at all their retirement expectations, individuals with intermediate (Mittlere Reife) or higher attainments (Abitur) considerably increased their expectations. A possible explanation could be that lower-educated individuals failed to understand the reform. While a misunderstanding of the new legal retirement age seems to be quite unlikely (given that it enters the slogan Rente mit 67 which has been repeated quite often even in the popular press), an underestimation of the financial consequences of early retirement seems liekly. Alternatively, lower educated individuals may not have revised their ERAs as they expect to rely on social assistance anyway. [16]

17 Table 3: Estimated effect of the increase in the legal retirement age on ERAs. Heckman selection model. Variable Model 1 Model 2 Age Post Reform Employee Employee*Post Reform ** Female *** *** Partner East Germany Hauptschule Ref. Ref. Mittlere Reife (Fach-)Abitur * Post * Employee*Hauptschule Post * Employee*Mittlere Reife ** Post * (Fach-)Abitur ** No post-secondary education Ref. Ref. Vocational training ** * [17] University degree *** * Post * Employee*no post-secondary education Post * (Fach-)Abitur Never unemployed Log(household's financial wealth) Have occupational pensions? Have subsidized private old-age insurance? Have not-subsidized private old-age insurance? ** ** Have real estate? Log(household s net income) ***

18 Self-rated health: fair to bad Inheritance expected Worsening of health condition expected Improving income expected Unemployment expected Subjective life expectancy (years) Unsatisfied with current job s s Constant *** *** Exclusion Restrictions (step-one equation) Exp. development own economic situation: missing *** *** Prob. Become unemployed: missing *** *** Own life expectancy: missing Joint significance selectivity indicators Chi p-value Wald test of independent equations (rho = 0) p-value Legend: * p<.10; ** p<.05; *** p<.01, Standard errors under the coefficients. Note: Reference group: Self-Employed, Male, West Germany, Hauptschule, No vocational training, some unemployment spells in the past, good to very good self-rated health. Table 4 below reports the predicted ERAs calculated from model 1 for the different educational groups of the respondents in the period after the reform, setting the values of all the other covariates to the sample mean. While the average predicted ERA in the period after the reform is 66.9 years for a male employee with Abitur and university degree, that of an equivalent subject with the lowest secondary school certificate and no further education remains at 65 years even in the period after the reform. [18]

19 Table 4: Predicted ERAs in the period after the reform for different typologies of employees Men Women No voc. training Voc. training University No voc. training Voc. training University Hauptschule Mittlere Reife (Fach-) Abitur Average Source: SAVE , multiply imputed data. Sample restricted as described in text. Predictions from model 1, Table 3. Value of all the other covariates set at sample means. Standard errors under the coefficients. This result has important policy implications. As individuals with lower educational attainments are also those being more at risk of having lower state pensions in the future (due to lower incomes and more numerous and/or longer spells of unemployment) and also those who are less likely to save for the old-age (see Coppola and Gasche, 2011), the fact that they have been slower in updating their retirement expectations is a cause of concern regarding their well-being after retirement. The last row on table 4 reports the predicted average ERA for all employees in the period after the reform. It has to be noted that the average ERA is still below the new statutory retirement age for these cohorts, the gap being about 9 months for men and 1 year and 3 months for women. Still, as we have shown in our regression analyses, individuals have adjusted their expectations to the reform considerably. [19]

20 7. Summary and conclusion This paper contributes to the literature on how individuals adjust their retirement expectations to legislated policy changes. More specifically, we quantify the effect of an increase in the statutory retirement age on individual retirement expectations for Germany. We consider the 2007 German pension reform that offers a nice quasiexperimental setting to properly single out the effect of the policy on expectations. The analysis is focussed on the cohorts born after 1963, who according to the new rules will have to retire at the new statutory retirement age of 67. Using a difference-in-differences estimator and accounting for possible biases induced by panel attrition and non-response to the question on expected retirement age, we find that on average the ERAs of individuals born after 1963 increased in the period after the implementation of the reform by 1.6 years. Furthermore, we find heterogeneity regarding educational attainment, indicating that lower educated individuals have adjusted their retirement expectations not at all or far less than we can perceive for the higher educated. This result has important policy implications. As individuals with lower educational attainments tend to earn lower wages (thus acquiring lower pension claims) and save less for retirement, the fact that they have been slower in updating their retirement expectations causes concern regarding their income level after retirement. It is thus important that the reform implications are communicated well across different educational groups and that the right incentives for additional private old-age provision are set. [20]

21 APPENDIX Table A1: Results of a placebo regression (expectations in year 2008 vs. year 2009) (noch zu aktualisieren) Variable coeff. s.e. Age Post Reform Employee ** Employee*Post Reform Female ** partner East Germany Hauptschule Ref. Mittlere Reife ** (Fach-)Abitur *** No post-secondary education Ref. Vocational training ** University degree ** Never unemployed Log(household's financial wealth) Occupational pensions (Y/N) Have subsidized private old-age insurance? (Y/N) * Have not-subsidized private old-age insurance? (Y/N) ** Have real estate? (Y/N) Log(household s net income) Self-rated health: fair to bad Inheritance expected Worsening of health condition expected Improving income expected Unemployment expected Subjective life expectancy (years) Unsatisfied with current job ** s Constant *** Exclusion Restrictions (step-one equation) Exp. development own economic situation: missing *** Prob. Become unemployed: missing Own life expectancy: missing Test of independent equations (rho = 0) p-value Note: post reform indicates here the year 2009 as opposed to the year 2008 [21]

22 Table A2: Determinants of the ERAs and effect of the increae in the legal retirement age: OLS regression Variable Coeff. S.e. Age Post Reform Employee Employee*Post Reform ** Female *** partner East Germany Hauptschule Ref Mittlere Reife (Fach-)Abitur No post-secondary education Ref Vocational training ** University degree *** Never unemployed Log(household's financial wealth) Occupational pensions (Y/N) Have subsidized private old-age insurance? (Y/N) Have not-subsidized private old-age insurance? (Y/N) ** Have real estate? (Y/N) Log(household s net income) *** Self-rated health: fair to bad Inheritance expected Worsening of health condition expected Improving income expected Unemployment expected Subjective life expectancy (years) Unsatisfied with current job s s Constant *** [22]