What You Don t Know Can t Help You: Knowledge and Retirement Decision Making

Transcription

1 VERY PRELIMINARY PLEASE DO NOT QUOTE COMMENTS WELCOME What You Don t Know Can t Help You: Knowledge and Retirement Decision Making February 2003 Sewin Chan Wagner Graduate School of Public Service New York University sewin.chan@nyu.edu Ann Huff Stevens Department of Economics Yale University ann.stevens@yale.edu Abstract This paper examines the relationship between pension incentives, individuals knowledge about those incentives, and the retirement decision. Combining detailed self- and employer-reported data on private pensions, we construct measures of the accuracy of individuals self-reports of their pensions. We show that the minority of well-informed individuals display dramatically larger responses to financial incentives than indicated by average estimates. These results suggest that there is substantial heterogeneity in responsiveness to pension incentives across the population. Finally, we estimate a joint model of information acquisition and retirement decision-making. These results confirm the substantial differences in behavior between informed and uninformed segments of the population. The authors gratefully acknowledge financial support from Boston College, Center for Retirement Research, and the Social Security Administration. Views expressed in this paper do not reflect those of these institutions.

2 I. Introduction Recent empirical research has revealed that many workers lack knowledge about the details of their public and private pension plans. 1 However, these findings have not been systematically linked to empirical models of retirement behavior. Indeed, much of our current understanding of individual responsiveness to financial retirement incentives comes from studies that rely on administrative data about those incentives. 2 If many individuals are unaware of the relevant financial incentives, the average effects of incentives may mask important heterogeneity in individual behavior: large effects for the informed segment of the population and much smaller (or zero) effects for the uninformed segment. If true, this would also imply that the effects of changes in Social Security or private pension plans will depend critically on the extent of knowledge individuals possess about both the existing program and the changes. Further, if the extent of knowledge is correlated with other individual characteristics, it is important to understand which segments of the population are most likely to be well-informed about the relevant details of their retirement incomes. This paper uses self-reported, employer-reported and Social Security administrative data from the Health and Retirement Study (HRS) examine the relationship between knowledge of retirement benefits and individuals responsiveness to those benefits. We first examine differences in the extent to which retirement decisions of knowledgeable and uninformed individuals respond to their true pension and Social Security benefits. Second, we explore the 1 See Gustman and Steinmeier (2001a,b,c), Duflo and Saez (2002), and Choi et al. (2001). 2 For example, Stock and Wise (1990), Lumsdaine, Stock and Wise (1992), Samwick (1998), Coile and Gruber (2000) and Gustman and Steinmeier (2001b) all use information from employers or administrative sources that may not be fully known by individual workers. 1

3 determinants of individuals understanding of financial incentives. Finally, we estimate a model that combines these two components, with an emphasis on identifying the causal effect of enhanced information on individual retirement propensities and responsiveness to financial incentives. II. Relationship to Previous Literature Correlates of knowledge about pensions Several studies have examined the extent and accuracy of information that individuals possess about their future retirement benefits. The common finding across these studies is that such information is typically quite incomplete. Mitchell (1988) uses private pension information from the 1983 Survey of Consumer Finances (SCF) provided by employers and workers. She finds that the extent of knowledge varies depending on the exact measure used; knowledge of employer contributions to pensions and of early retirement rules and requirements is particularly incomplete. She also finds that membership in a union, job tenure, industry-level profitability, and the type of pension plan are related to the extent of workers knowledge, with the strength of this relationship varying substantially across alternative knowledge measures. In addition, minorities and males are less fully informed than white and female workers. Luchak and Gunderson (2000) examine pension information among employees of a unionized public utility industry in Canada. They find that individuals in worse health, those with higher earnings, those who have already worked beyond their pension maximizing age, and those in professional or semiskilled professions are more likely to know key components of their pension plans. 2

4 Finally, in a series of two articles, Gustman and Steinmeier (2001a, 2001c) examine determinants of knowledge concerning both Social Security and private pensions using the HRS, and the implications of this knowledge for retirement planning. They find that certain retirement planning activities, union membership and longer planning horizons signal individuals who are better informed about their pensions. In addition, knowledge seems to improve with age, or as workers approach typical retirement ages. Gustman and Steinmeier (2001c) also examine whether misinformed workers retire earlier or later than they had expected, and find that workers who initially overestimate their available benefits retire later than they had originally planned. However, they do not explore how the responsiveness to financial retirement incentives varies based on the extent of workers knowledge. Existing Research on Retirement Decision-Making Virtually all of the existing literature examining private pension incentives, and much of the recent literature on Social Security incentives, utilize employer- or administrative data about which the individuals themselves may be ignorant. Stock and Wise (1990), and Lumsdaine, Stock and Wise (1992) estimate both structural and reduced-form models of the retirement decision, using data provided to the researchers by a single large employer. Samwick (1998) uses the employer-provided data on private pensions from the SCF to estimate a regressionbased counterpart to existing structural models of retirement, focusing on role of forwardlooking financial retirement incentives, or option value in determining the probability of retirement. In a similar approach, both Coile and Gruber (2000) and Gustman and Steinmeier (2001b) use employer-provided pension data, along with administrative records of Social Security earnings histories to estimate the effect of both private pension and Social Security 3

5 benefits on retirement. None of these papers consider the effects of individual knowledge of these financial incentives on retirement behavior. In our own previous work, we have deviated from this tendency to rely on administrative pension reports, and focused on exploring the role of pensions in retirement behavior and retirement expectations using the self-reported pension data (Chan and Stevens, 2002, 2003). We find the self-reported financial incentives to be important determinants of retirement behavior, even after conditioning on earlier retirement expectations, or after controlling for individual-specific fixed effects. Our approach here will follow the recent literature in estimating regression-based retirement models that use a forward-looking summary measure of financial incentives (based on administrative and employer data) as the key dependent variable. In addition, we incorporate the self-reported pension and Social Security data to form measures of individuals understanding of forward-looking financial incentives, and explore how this understanding interacts with the incentives to affect retirement decisions. III. Data and Samples We use data from the first four waves of the Health and Retirement Study (HRS), including: (i) the self-reported, publicly released data on retirement income sources, retirement behavior, and retirement expectations; (ii) the employer-provided pension plan details, and (iii) the Social Security earnings histories, which allow us to calculate individuals Social Security 4

6 benefits at alternative future retirement dates. Several points should be made regarding these data sources. First, the employer-provided pension data are available only for jobs held at or before wave 1 of the survey, while self-reported benefit data are now available for four waves. Similarly, the administrative earnings histories include Social Security covered earnings only up to the initial wave. However, it is possible to use the employer-provided data to judge the accuracy of self-reported private pension information because job changes are not very frequent at this age range. Individuals who do change employers before retirement will be dropped from our samples, as will the small fraction of individuals who report changes to their pension plans between waves. 3 The wage histories from Social Security records can be supplemented with self-reported earnings in subsequent years so that Social Security benefits can be calculated for later waves as well. Because of this projection of the administrative data to subsequent waves, it will be important to control carefully for the survey wave in our analysis, and compare results across the initial and later waves. Second, we follow many previous researchers in assuming that the employer and Social Security administrative data represent the true benefit information. In some cases, particularly with respect to the employer reports, this assumption may not be accurate. For defined contribution pension accounts in particular, there are good reasons to expect that the employer reports could deviate from the truth (Uccello and Perese, 1999). Thus, in the case of defined benefit plans, we are more confident about relying on the employer reports as the true pension. 3 Individuals are asked whether there have been any changes to their pension plans since the previous survey. They are then asked about the details of their pensions, even if they report no change since the previous wave. 5

7 Throughout the analysis, we examine the sub-sample of workers with defined benefit plans to investigate whether our results differ systematically across pension plan types. Third, while the self-reported data are useful because they reflect what individuals actually know, many researchers have expressed concerns that such data may contain substantial measurement error. These pension reports may differ from the employer reports not only because individuals actually do not know the information, but also because they misreport this information. While this is certainly a concern, we believe that the potential importance of understanding how knowledge affects retirement decision-making overrides the concern about measurement error. Indeed, our preliminary results described below demonstrate that there are large and statistically significant differences between informed and uninformed individuals. These differences will be understated if our measures of knowledge are severely affected by measurement error. Using the self- and employer-reported data on pensions, our first step is to take the pension plan details and use them to construct the present value of pension wealth available to the individual at each possible future retirement age. Following previous literature, we construct financial incentive measures that are based on the difference between: (1) the present value of pension wealth if an individual retires today, and (2) the present value of pension wealth if an individual retires at the future date that maximizes his or her present value of pension wealth. This measure is referred to as an individual s pension gain. 4 4 This is similar to the option value concept used in Lumsdaine Stock and Wise (1992), and Samwick (1998), equivalent to the peak value used by Coile and Gruber (2000) and related to the premium value used by Gustman and Steinmeier (2001b). 6

8 In the case of the self-reported data, for those reporting defined benefit pensions, we have information on annual benefit amounts, the normal retirement (pension eligibility) age, the early retirement age, cost of living adjustments, and benefit reduction amounts (in the case of early retirement) from each survey wave. For individuals reporting defined contribution pensions, we have information on the current account balance, employer- and own-contributions, and basic portfolio allocation decisions for existing DC accounts. These pieces of data are combined to generate estimates of the present value of pension wealth at each possible retirement age. The employer-provided reports, while more detailed, contain similar information for each pension plan. In calculating the present discounted value of pension wealth at each retirement age, we assume a discount factor of 3%, and weight future values by age- and gender-specific survival probabilities taken from Social Security Administration actuarial tables. Table 1 shows summary statistics for our sample. To be included in the analysis, individuals must have an employer-reported pension plan, and cannot yet be retired at the time of the wave 1 survey. Individuals remain in the sample through wave 4, or until they transition to retirement. Roughly 7 percent of all observations reflect a retirement in the current year, while 36 percent of individuals in our sample have retired by the time of the wave 4 survey. Slightly more than half of the sample is male, and the average age is just over 57 years. The sample includes 5 percent hispanic respondents and 16 percent black respondents. More than threequarters of the sample members are married. The middle section of Table 1 provides means for key summary measures of pension wealth, based on both self- and employer-reports. The average level of pension wealth (assuming the individual retires in the current year) is approximately $65,000 (in 1992 dollars) 7

9 based on self-reports, but is more than $150,000 based on the employer-provided pension reports. The pension gain, calculated as the the present value of pension wealth if retirement occurs at the wealth-maximizing age minus the present value if retirement occurs today is $50,000 from the self-reports and $31,000 from the employer-reports. The mean incentive for continued work is somewhat higher in the self-reports. If we take the median values of pension gain, however, the relative values are reversed: the median pension gain based on self reports is $0, while the median pension gain from employer reports is $15,494. The final section of Table 1 presents means of some variables intended to describe the accuracy of individuals self-reported pension values. These measures are defined and discussed in the following section. IV. Information, Financial Incentives and the Retirement Decision Exploring Heterogeneity in Responsiveness to Financial Incentives Our first step in the empirical analysis is to document differences in the responsiveness of individuals to financial incentives based on a variety of measures of their understanding of these incentives. We want to capture the fact that there may be heterogeneity in the extent to which individuals respond to financial incentives, and that this heterogeneity may be related to the degree of information individuals possess about these incentives. Following much recent empirical work, we estimate a retirement equation described by: [1] Rit = α 0 + α1gainit + α 2INFORMEDit + α 3GAINit * INFORMEDit + X it β + ε it 8

10 where R it is a binary variable indicating that the individual i retires in year t, and ε is an error term. GAIN it represents the pension gain to continued work, as defined above, based on employer-reported pension data. Samwick (1998) and Coile and Gruber (2000) show that forward-looking measures such as pension gain have greater explanatory power than alternative, myopic, summary measures of financial status or incentives. X it is a vector that contains several important control variables, including age, demographic characteristics, calendar years, and measures of non-pension wealth and lifetime earnings. 5 With the exception of the terms associated with coefficients α 2 and α 3, this specification is identical to that used in previous work. Conceptually, INFORMED it is a variable that will reflect the extent of individuals knowledge regarding their pensions and social security benefits. In their earlier work, Gustman and Steinmeier (2001a) construct several measures of this variable. In particular they consider whether individuals correctly identify the type of their pension plan (defined benefit (DB) or defined contribution (DC)), whether individuals report that they do not know the value of their pensions, and various measures of the difference between the value of their pension based on self- and employer-reports. We use several similar measures in our analysis, and also focus on measures that come closer to capturing individuals knowledge of the actual gain to continuing work (or of the forward-looking financial incentives used in recent retirement models). First, following the basic strategy of Gustman and Steinmeier (2001a), we consider whether individuals report don t know when asked about the value of their pension. Specifically, we construct a variable, referred to as REPORT, that is equal to one for individuals 5 See Coile and Gruber (2000) for a discussion of the importance of carefully controlling for measures of lifetime earnings. It is also crucial to control fully for age since there are spikes and non-linearities in the age profile of retirement probabilities. Knowledge is also highly correlated with age. For these reasons we include a series of dummy variables for each year of age (51 to 70) in our sample. 9

11 who report enough of the key components of their pension plans to allow us to construct the present value of benefits at various possible retirement ages. If individuals report not knowing some major component of their pension, REPORT is set equal to 0. As noted by Gustman and Steinmeier, a large fraction of individuals report not knowing the value of their pension. As shown in Table 1, 59 percent of our sample have REPORT equal to 1. An alternative measure used in previous work is to compare the present value of an individual s level of pension wealth based on the self-report with that based on the administrative report. 6 Such a measure is necessarily limited to the subsample who actually report enough information on their pension to calculate the value of pension wealth. To implement this concept, we create a variable that is equal to one if (i) the ratio of the self-reported to employerreported pension wealth is between 0.5 and 2, or (ii) one of the reported pension wealth values is zero, and the other is less than $5,000 in absolute value. The second component of this definition is included to avoid misclassifying as uninformed those who report zero wealth rather than very small amounts of pension wealth. For the measure just described, Table 1 shows that 35 percent of our sample report levels of pension wealth are relatively close to those based on the employer reports. In many of the results below, we also set this, and similar, variables to 0 for individuals who do not report sufficient information (REPORT=0). Finally, because we are primarily interested in individuals knowledge of the financial incentive to retire inherent in their pension structure, we use a third measure of knowledge that captures the accuracy of the individuals pension gain based on comparisons between self- and employer- reports. We form each individual s pension gain using both self- and employerreported values of the present value of pension wealth at all future possible retirement ages. 6 Gustman and Steinmeier (2001), for example, create a variable equal to 1 if the self report is within 25% of the administrative report of pension wealth. 10

12 Then, the variable INFORMED is set equal to one if the ratio of pension gain based on self- and employer- reports is between 0.5 and 2, or if one is zero and the other is within $5,000. Table 1 shows that 30 percent of individuals in our sample are well-informed regarding this more complex measure of the incentive effects of pensions. This measure requires individuals to understand benefit eligibility, amounts and reductions at alternative hypothetical retirement ages. In Table 2, we combine these measures of knowledge about pension plans with a standard probit for retirement, in which the key financial incentives to retire are summarized through the pension gain measure. Other variables included in these specifications are: race, education, sex, self-reported health status, marital status, dummy variables for age, level of nonpension assets and a dummy for having less than $20,000 in total assests, availability of retiree health insurance from your employer (or Medicare), a measure of average lifetime earnings, and dummies for the survey wave. The full set of coefficients from these models are shown in Appendix Table 1. Standard errors have been corrected to account for repeated observations from the same individuals across survey waves. We present results for men and women pooled, although we have also obtained similar results when we estimate separate regressions for men and women. For comparison with previous literature, the first column of Table 2 shows the effect of pension gain and the level of pension wealth on the probability of retirement. As expected, a larger pension gain reduces the probability of retirement and is statistically significant. Individuals with larger financial gains from continuing work until a future optimal retirement date are significantly less likely to retire. The probit coefficient of implies that a onestandard deviation increase in pension gain ($52,881) increases the annual probability of retirement by 1.1 percentage points, or roughly 17 percent. Coile and Gruber (2000) report a 11

13 similar probit coefficient on the incentive term (combining both pensions and social security) of We have also included the level of pension wealth (based on retirement at age 65) in the retirement probit. This variable, as expected, is positive and statistically significant, suggesting that those with higher levels of pension wealth are more likely to retire. The magnitude of this coefficient is also very close to that reported in Coile and Gruber. We should note that some other previous work has not found a significant effect of the level of pension wealth on retirement probabilities. Samwick (1998), for example, finds that the level of pension wealth has no effect on retirement once a forward-looking incentive measure is included in the model. Overall, however, these basic results are consistent both with theoretical predictions and with previous empirical work. In the next columns of Table 2, we examine whether there is important heterogeneity in the responsiveness of individual retirement decisions to pension gain, and in particular whether such heterogeneity is related to the degree of information individuals possess about their pensions. As explained above, we begin by using a very simple variable (REPORT) that is equal to one for any individual who actually reports a value for each of the key components of his or her pension plan. In column 2, we show that knowing (or at least reporting some estimate of) these key values is associated with an increased probability of retirement. Reporting these values increases the annual probability of retirement by approximately 1 percentage point. This may reflect reverse causality: individuals who are actively considering retirement may be more likely to know enough to report some values for their pension plan. Also of interest is the interaction between REPORT and pension gain. This interaction, however, is very small and far from statistical significance. In column 3, we report results in which we use information on the accuracy of individual self-reported levels of pension wealth. Having self- and employer- reports 12

14 of pension wealth levels that are relatively close has no significant effect on retirement, and its interaction with pension gain is also insignificant. Given that our focus is on individual responsiveness to the financial incentives, we next focus on a measure that captures the accuracy of a self-reported version of pension gain. In column 4, we use INFORMED and its interaction with pension gain. This specification produces quite different results from the other information measures. The main coefficient on pension gain is very small and no longer statistically significant. This captures the effect of pension gain (from employer reports) for those individuals whose self-reports are substantially different than their employer reports (or who do not report enough information to construct it). Having reasonably accurate knowledge of one s pension gain significantly increases the probability of retirement, as indicated by the reported probit coefficient of This implies an increase of approximately 3 percentage points. Again, we note that this could reflect the fact that individuals particularly interested in retirement may be more likely to possess a high degree of knowledge about their pension plans. Perhaps most interesting, however, is the interaction between INFORMED and pension gain. For those individuals who accurately report the financial incentives implied by their pension structures, there is a very large negative effect of pension gain on retirement probabilities. The coefficient for this group is roughly 5 times the size of the average effect for the entire sample (shown in column 1 of the table). The probit coefficient of implies that a one standard deviation increase in pension gain would lead to a reduction in the retirement probability of 4.4 percentage points, or 65 percent. This highlights some dramatic heterogeneity underlying the average effect of pension gain reported in column 1, and the estimates reported in previous literature. The effect of pension-related financial 13

15 incentives appears to be driven by a minority of individuals (30 percent in our sample) who are relatively well-informed about their pensions. The final two columns of Table 2 confirm these results. In column 5, we include both the levels-based and the pension-gain based versions of the information variable. The results again show that individuals with detailed knowledge of their pension plans show a much greater responsiveness to pension-related incentives in making the retirement decision. Finally, in column 6, we drop from the sample individuals who do not report enough information to form a self-reported measure of pension gain (REPORT=0 and thus, INFORMED=0). Looking at the effects of the accuracy of information conditional on having reported all of the key pension components, allows us to distinguish between individuals who simply do not report information on their pensions and those who report values that are very different from their employer-reports. These coefficients are very similar to those reported in the previous columns. The results in Table 2 are robust to a number of variations in the basic sample. We obtain similar results when we restrict the sample to include only individuals with defined benefit pensions (for whom we have more confidence in the employer-reports as the truth.) Additionally, eliminating wave 1 from the analysis does not substantially alter the results. The use of a dichotomous variable to measure the accuracy of individual pension reports is somewhat restrictive. The idea behind this strategy, further developed below, is that individuals can be characterized as either informed or uninformed. However, there may exist a continuum of understanding about the relevant pension incentives. As an alternative to this summary measure, we have also considered measures that make use of the extent of correlation between the self-reported and employer-based series of pension values. That is, because an individual s ability to predict the pension gain measure properly depends upon their 14

16 knowledge of differences in pension wealth at different retirement ages, the shape of the retirement-age profile of pension wealth will be critical. Using the series of present values of pension wealth based on retirement from the current age to age 70, we calculate the correlation coefficient between the self- and employer reported series, and include this correlation coefficient, along with its interaction with pension gain, in the retirement probits. The use of this continuous measure of pension information held by individuals provides mixed results. One difficulty with using this measure is that, for individuals with no variation across retirement ages in their pension profiles (variance of their present values is zero), the correlation coefficient is missing. This is a fairly common situation among the self-reported pension values: many individuals (primarily those with defined contribution pensions) report zero pension wealth, or some other fixed amount, regardless of the age of retirement. Thus, we lose a large number of observations. Additionally, for those who do not report enough pension information to construct the present values, it is less clear for this measure what value to use. For our dichotomous measures used above, we have set the accuracy variables equal to zero (although this assumption is tested in column 6 of Table 2 for those with REPORT not equal to one). When using the correlation coefficient directly as our measure of knowledge, it is not obvious how to handle those who do not report pension values. If we enter the correlation coefficient directly as the measure of knowledge, eliminating those observations with missing values, the interaction between pension gain and this measure is not statistically significant. In another specification, we set the knowledge measure equal to the value of the correlation coefficient only if it is greater than zero. For negative correlation values and missing values, we set the variable to zero. This specification produces a negative interaction with pension gain, though the interaction itself is also not statistically significant. 15

17 The magnitude implies that an individual with a correlation of one between the self- and employer-reported pension wealth profile would have roughly twice the response to pension gain as a similar individual with a zero correlation. Finally, the interaction between knowledge measures and pension gain could simply reflect non-linearities in the effect of pension gain on retirement. If pension gain at or below zero (implying no gain, or a loss, in wealth from further work) is an especially strong indicator of retirement, and if having zero pension gain is correlated with the extent of an individual s information, the non-linearity could be driving the results in Table 2. This seems likely since many of the self reports, as noted above, have extremely flat retirement-age profiles of pension wealth. To address this concern, we add to the basic model a dummy variable indicating whether pension gain is less than or equal to zero. In addition, we interact our preferred knowledge measure, INFORMED, with this dummy variable. These results are summarized in Table 3. First, as expected, there are strong non-linearities in the effects of pension gain. In the first column, we include the dummy indicating non-positive pension gain values, but do not include the information measures and interactions. Having a zero or negative pension gain significantly increases the chances of retirement, above and beyond the linear effect of pension gain. The coefficient of on pension gain less than or equal to zero implies that having a zero pension gain raises the probability of retirement by approximately 1 percentage point. The second column of the table, however, shows that interactions between pension gain and INFORMED remain significant and of the expected signs. The coefficient on the interaction between INFORMED and pension gain is significant and equal to , implying that a one-standard deviation in pension gain reduces the probability of retirement by approximately one-third. The interaction between INFORMED and having a non-positive pension gain is also significant and 16

18 has the expected positive sign. This coefficient of implies that a well-informed individual with a zero pension gain is twice as likely to retire as a similar, but uninformed, individual. The results discussed so far offer important evidence on the potential range of responsiveness to pensions across the population, but they should be interpreted carefully. The fact that we find much larger effects of financial incentives among individuals who are better informed about their pensions does not necessarily suggest, for example, that providing better information to individuals will increase their responsiveness to those incentives. The decision to collect information about one s pension benefits is likely endogenous to the decision to retire or to begin considering retirement. Indeed, the positive and significant coefficients on INFORMED and REPORT in the retirement models suggest that we should not think of knowledge as exogenous. Individuals who are well-informed about their pension gain are 3 percentage points more likely to retire in a given year, regardless of what these financial incentives are. Being well-informed may simply be a signal that an individual has begun to seriously consider retirement. It is possible that this endogeneity of knowledge could also affect the interaction terms between INFORMED and pension gain. If, for example, those individuals whose retirement decisions are most sensitive to potential financial gains are more likely to understand the details of their pensions, the interaction terms of interest could also reflect this reverse causality. A Retirement Model with Selection into Information States Our next goal is to better understand the results presented in Tables 2 and 3. Does the powerful interaction effect between pension gain and information mean that the provision of information will substantially alter retirement behavior? Toward this goal, we estimate a two- 17

19 stage model based on our findings above. In the first stage we will estimate an equation of the following form: [2] INFORMED it = α 0 + β1gainit + β 2 X it + β3zit + vit where INFORMED it is a binary variable indicating whether or not an individual is informed about his or her financial options surrounding retirement, GAIN it is the pension gain as discussed above, X it and Z it are vectors of variables that affect information acquisition, and v is an error term. Note that both GAIN and X will appear in the second stage below, but that Z is unique to the first stage. 7 In the second stage, an individual s retirement decision can be modeled in the following way, depending upon which regime they are in: [3] a) b) R R it it U U U U = γ 0 + γ 1 GAIN it + γ 2 X it + ε it if uninformed (INFORMED it = 0) K K K K = γ 0 + γ 1 GAIN it + γ 2 X it + ε it if informed (INFORMED it =1) As shown by Heckman (1976) and Lee (1976), equation [3] can be consistently estimated by explicitly conditioning on selection into the two regimes (uninformed and knowledgeable). Inverse mills ratios are calculated from estimation of equation [2], and then included in the estimation of equation [3]. Assuming that vit U ε it, expectation of [3], conditional on selection into each regime is: K it, ε are jointly normally distributed, the U U U φ( δw / σ v ) [4] E( Rit I < 0) = γ 0 + γ 1 GAINit + γ 2 X it ρ1σ ε ( ) 1 1 Φ(( δw / σ ) v E( R it I > 0) = γ K 0 + γ K 1 GAIN it + γ K 2 X it + ρ σ 2 ε 2 φ( δw / σ v ) ( ) Φ(( δw / σ ) v 7 We allow for the possibility that GAIN affects the extent of knowledge, since models of information acquisition suggest that those with more potential benefit will be more likely to invest in information gathering. 18

20 W is a vector combing GAIN, X, and Z from equation [2] above; ρ1and ρ 2 are the correlation between the error terms in [2] and [3a] and [3b], respectively. The first point to note about this system of equations ([2] and [3]) is that it highlights the major challenge confronting estimation of this model: the separate identification of the switching equation and the retirement equations. To consistently estimate the parameters of the model, we need to identify factors that: (i) help predict the individual knowledge of pension incentives, but that, (ii) do not also affect retirement decisions. We propose and implement a preliminary identification strategy below. Before proceeding to estimation of equation [2], we briefly discuss which factors should affect the extent of information individual s possess about their pension plans. Mitchell (1988) proposes a general framework for thinking about information acquisition in this setting, identifying three broad sets of factors that should influence the extent of pension information individuals possess. First, the abilities of individuals to process complex financial information may affect retention and understanding of their pension plans. Educational attainment and other individual characteristics would fall into this category. Note that most of these characteristics should also affect retirement probabilities, and so will not help in the separate identification of equations [2] and [3]. Second, firm-based differences in the efficiency of information dissemination may affect the extent to which workers are well-informed. Employers have a role to play in providing information to the beneficiaries of their pension programs. Focusing on variables that may lead to information dissemination by pension providers is one way to identify factors that are more likely to be exogenous to workers retirement behavior. We use several variables that are 19

21 generated on the firm-side of the employment relationship as exclusion restrictions in the estimation of the selection model. Firm size, which is available in the HRS, is one such factor. Larger firms may be more likely to have formal information programs, or at least human resources departments able to systematically provide information to their workers. There is little reason to think that, after conditioning on financial incentives and lifetime earnings, the size of a worker s firm should influence retirement. 8 Union membership might similarly indicate the existence of an information-provision system for a given worker. Both Mitchell (1988) and Gustman and Steinmeier (2001c) find some evidence that unionized workers are better informed about their retirement incentives. Union status should not influence the retirement decision, as long as retirement incentives and other characteristics of the job that might alter the utility of leisure (physical demands of the job, for example) are controlled for. Fortunately, the HRS also contains a variety of measures of such job characteristics that might allow us to credibly exclude a worker s union status from the retirement regressions. We propose two other characteristics associated with the firm as potential determinants of an individuals information level. First, the HRS asks respondents whether there is a usual age of retirement at their firm. Employees in firms that do not have a usual age of retirement may be more likely to seek out information about their pension plans since they cannot simply follow what their co-workers are doing. A second potential identification strategy involves using information on the time until an individual s retirement-wealth maximizing age, based on the employer-reported pension gain. The idea here is that firms will have an incentive to inform individuals of upcoming pension benefits as these individuals approach kinks or important 20

22 vesting points in their pension benefit schedules. Conditional on the private pension gain associated with a given year and the individual s age (entered non-linearly), the number of years before or after their optimum should not influence retirement. 9 A third set of factors identified in earlier work as influencing individual knowledge of pensions involves the benefit of that information. A greater likelihood of benefiting from knowledge of pension details should lead individuals to acquire more information. The size of the gain itself maybe relevant here, as well as health status, job characteristics and the age and health of the individual and his or her spouse. As was the case with individual demographic characteristics, most of these factors will also affect the retirement decision. Based on the above considerations, as a first step, we estimate the model summarized by equations [2] and [3] using firm size, union status, no usual age of retirement, and the time left until the pension maximizing age from employer reports, as our excluded variables in estimating the selection model. We control for a host of additional firm and individual characteristics in estimating each of the two equations. Results from stage 1: predicting INFORMED Table 4 presents our estimates of equation [2]. The first set of explanatory variables are some basic individual characteristics that may reflect the ability of individuals to process complex financial information. While the education terms are jointly statistically significant, the negative coefficient on college graduate is somewhat surprising. Marital status and gender have no significant effect. Blacks and Hispanics are less likely to be informed. Those with higher 8 Mitchell (1988) uses a measure of firm size in her analysis of the determinants of pension information, and does not find it to have a significant effect. She had information, however, only on the average firm size at the industry level, while the HRS provides information on the actual firm size of the individual s employer. 21

23 average earnings are significantly more likely to be informed, while the effect of non-pension financial assets is insignificant. The second set of explanatory variables reflect information dissemination at the firm level. The indicator variables representing firm size and union status are significant, however, the negative coefficients on both firms larger than 500 employees and union status are unexpected. As expected, workers in firms with no usual age of retirement are more likely to be informed. The third set of explanatory variables reflect the benefits of acquiring information. The effects of poor health turn out to be insignificant. Those who report physical effort most of the time or stress on their jobs are significantly less likely to be informed, while those who report physical effort all the time are more likely to be informed. The availability of retiree health insurance does not significantly affect knowledge. Current pension wealth also does not explain knowledge. As expected, there is a negative and significant coefficient on years until the maximum pension wealth. This indicates that those with less benefit from being informed are less likely to be informed. Those with DB pension plans, which typically involve greater financial incentives for retiring at particular dates, are more likely to be informed. The full set of age dummies (not shown) are jointly significant and display the expected pattern: older individuals are more likely to be informed. It is worth noting that the acquisition of information seems to occur relatively late: the age dummies do not begin to increase until an individual s late 50s or early 60s and they are extremely large after age 65. Finally, the coefficient on employer reported pension gain is positive and signficant. Those facing stronger financial retirement incentives are more likely to be well-informed about those incentives. 9 This is similar in spirit to Luchak and Gunderson s (1999) finding that individuals at or beyond their wealth maximizing age are more knowledgeable about pension benefits. 22

24 Results from stage 2: the effects of information, controlling for selection To incorporate correlation between the determination of knowledge (equation [2]) and the retirement decision (equations [3]) we use the results from estimation of equation [2] to form selection correction terms. The results of estimating equation [3] are summarized in Table 5. Selection into the informed and uninformed regime is predicted using the specification summarized in the previous table. Unlike our previous results, retirement probabilities are estimated here using a linear probability model, to facilitate the inclusion of the selection terms. Focusing on the first few columns of Table 5, very few of the coefficients on control variables (other than pension gain) differ significantly across the informed and uninformed regimes. The selection terms (inverse mills ratios from the estimation shown in Table 5) are significant in both the informed and uninformed equations. For individuals with knowledge about retirement, the selection term is positive and significant, while for the uninformed group, the selection term is negative and significant. This is consistent with a positive correlation between being informed and retirement propensities. The effect of pension gain continues to vary dramatically between informed and uninformed individuals, however, with coefficients of and 0.001, respectively. In the linear probability model, this implies that a standard deviation increase in pension gain reduces retirement probabilities by approximately 3 percentage points among the well-informed. For comparison, column 2 shows that an otherwise identical specification that ignores selection gives a very similar coefficient on pension gain. Because of our concerns about correlation between INFORMED and non-linear patterns in pension gain, the next two columns of Table 5 repeat the selection model, but also including the dummy variables for having a non-positive pension gain (similar to the specification shown 23

25 in Table 3). In the informed regime, the selection term is positive and significant as before. The linear effect of pension gain is no longer statistically significant once we include the selection correction terms. The dummy variable for having a non-positive pension gain is positive and significant, suggesting that those who are well-informed and have a zero or negative pension gain are more likely to retire. These results stand in contrast to the linear probability model when the selection terms are not included: the pension gain coefficient for the informed is negative and significant, and the zero pension gain term is positive and significant. For uninformed individuals, including the selection correction term results in a positive and statistically significant coefficient on both pension gain and the non-linear term. While the positive coefficient on pension gain is surprising, this is consistent with informed and uninformed individuals behaving in quite different ways. While the variable for a non-positive pension gain is positive and significant for the uninformed group, it is roughly half the magnitude of the same variable among informed individuals. These results suggest that, while controlling for selection into the informed state makes a difference, there remains a larger responsiveness to pension gain among the informed (operating exclusively through the nonlinear term). Several caveats are required in interpreting the results from Table 5. First, we will only have adequately controlled for the endogeneity of information acquisition if our exclusion restrictions are valid, or if firm size, union status, having no usual age of retirement, and time before reaching the age of maximal pension wealth do not affect the probability of retirement. One concern is that the signs on firm size and union status are contrary to what theory and previous empirical work has suggested. We have experimented with using subsets of these exclusion restrictions, and get qualitatively the same results. Second, the differences in the 24

26 results depending upon whether the non-linear pension gain term is included require some additional investigation. We are not aware of any previous work in which this non-linearity has been reported. Future work will focus on better understanding the link between the process of information acquisition and the retirement decision. V. Conclusion Much previous work relating the financial incentives from private pensions to retirement behavior has been based on administrative pension data, the details of which seem beyond the information set of individual workers. In this paper, we ask whether there are systematic differences between informed and uninformed individuals in making retirement decisions. Without attributing a causal interpretation to the role of information, we first document large differences by information status in the relationship between pension incentives and retirement. Individuals who are uninformed about the incentive effects of their pensions show little or no responsiveness to these incentives. In contrast, the minority of individuals who do have (based on self-reported pension information) a reasonable idea of the nature of these incentives, show a substantial response to pensions in making their retirement decisions. We have also estimated a model that attempts to control for the endogenous selection of individuals into informed and uninformed states. While these preliminary results depend critically on our assumed exclusion restrictions, we do find evidence that information acquisition is endogenous. Selection correction terms are always statistically significant in retirement equations estimated separately by individuals information status. When we include both linear 25