Does Retiree Health Insurance Encourage Early Retirement?

Abstract

The strong link between health insurance and employment in the United States may cause workers to delay retirement until they become eligible for Medicare at age 65. However, some employers extend health insurance benefits to their retirees, and individuals who are eligible for such retiree health benefits need not wait until age 65 to retire with group health coverage. We investigate the impact of retiree health insurance on early retirement using employee-level data from 54 diverse firms that are clients of Towers Watson, a leading benefits consulting firm. We find that retiree health coverage has its strongest effects at ages 62 through 64. Coverage that includes an employer contribution is associated with a 6.3 percentage point (36.2 percent) increase in the probability of turnover at age 62, a 7.7 percentage point (48.8 percent) increase in the probability of turnover at age 63, and a 5.5 percentage point (38.0 percent) increase in the probability of turnover at age 64. Conditional on working at age 57, such coverage reduces the expected retirement age by almost three months and reduces the total number of person-years worked between ages 58 and 64 by 5.6 percent.

1. Introduction

In the United States, there is currently a strong link between health insurance and employment. Most individuals can only purchase health insurance at favorable group rates through their employer, and there are significant tax advantages to employer-based coverage. Employment-based health insurance can make it more difficult for individuals to retire before they become eligible for health insurance through Medicare at age 65. While some employers extend health insurance coverage to their pre-65 retirees, most do not. According to the Kaiser Family Foundation (2010), only 28 percent of large firms (with 200 or more employees) and 3 percent of small firms that offer employee health coverage also extend benefits to retirees. A worker whose employer does not offer retiree health coverage has limited options for obtaining health insurance if he or she retires before becoming eligible for Medicare. Buying an individual health insurance policy can be difficult, particularly for those with preexisting conditions. The Consolidated Omnibus Budget Reconciliation Act (COBRA) of 1985 allows workers who leave their jobs to continue to participate in their former employer’s health plan at group rates for up to 18 months. This law makes it possible for workers to retire at age 63½ without losing group coverage, although they would forego any employer contribution toward their premiums.

In this paper, through the lens of a new data source, we revisit the impact of the availability of group health insurance on the decision to retire. We have access to a unique and rich data source for examining this question. Our dataset consists of employee records from a large and diverse group of firms, drawn from among the clients of Towers Watson, a leading benefits consulting firm. These employee records are matched to detailed information about the firms’ benefit provisions. Some of Towers Watson’s clients offer health insurance to their retirees, while others do not. Moreover, the retiree health benefits that are offered vary in their generosity and eligibility criteria. Another advantage of our dataset is that we can control for a number of firm-level characteristics that influence retirement, including specific features of defined benefit and defined contribution pension plans. If access to health insurance does in fact influence retirement decisions, then we would expect to find a relationship between retiree health coverage and retirement for persons age 64 or younger.

This question is particularly important in light of the recently passed Patient Protection and Affordable Care Act (PPACA) of 2010, which will considerably weaken the link between employment and health insurance by making group coverage available to all individuals regardless of employment. Many individuals will also receive explicit subsidies to purchase group coverage, and older individuals will also likely receive substantial implicit subsidies through a legal limit on their premiums relative to those paid by younger individuals. One possible consequence of this reform is that it may encourage earlier retirements, as all older workers will be able to maintain group coverage – often with generous subsidies – even if they retire before Medicare eligibility. Studying the link between employer-provided health insurance and retirement can help us to understand the potential impact of PPACA on the labor market participation of older pre-Medicare workers.

To preview our results: We find that after controlling for individual and firm characteristics and pension plan features, being eligible for subsidized retiree health coverage (i.e., coverage in which the employer contributes towards the premium) raises the probability of turnover (leaving the firm) by 6.3 percentage points at age 62, representing a 36.2 percent increase relative to the turnover rate for individuals with no coverage. It raises the probability of turnover by 7.7 percentage points (48.8 percent) at age 63 and by 5.5 percentage points (38.0 percent) at age 64. We find no such effects for individuals who work for firms that offer retiree health insurance but do not meet the eligibility criteria for coverage. The effects of retiree health coverage are fairly consistent for men and women, and for high-salary and low-salary workers. We find little evidence that “access only” (i.e., coverage in which the retiree gets a group rate but the employer does not contribute towards the premium) influences retirement decisions in this age range. Subsidized retiree health coverage reduces the total years of employment between ages 58 and 64 by 5.6 percent. These results are consistent with the hypothesis that Medicare eligibility influences workers’ retirement decisions, specifically among individuals who are younger than age 65 and do not have access to subsidized retiree health coverage.

The remainder of the paper is organized as follows. Section 2 summarizes the previous literature on the relationship between health insurance and retirement, and describes the contribution of this paper. Section 3 describes our dataset. Section 4 presents our methodology, and Section 5 discusses our results. Section 6 concludes

2. Prior Research on Health Insurance and Retirement

Prior studies have used a variety of approaches to estimate the effect of health insurance on retirement. A number of studies use a reduced form approach to examine the retirement rates or labor force participation rates of those with and without retiree health coverage, controlling for other factors and, in some cases, for selection into retiree health coverage based on unobservable characteristics. In general, these studies find that retiree health coverage substantially increases the probability of early retirement among pre-Medicare eligible workers. Blau and Gilleskie (2001) estimate that subsidized retiree health coverage increases the rate of retirement (labor market exit) by about 2 percentage points per year among male workers aged 51–61, with an increase of 7.5 percentage points among 61-year-olds. They find that the effect on retirement is positive starting at age 54 and increases with age. Kapur and Rogowski (2011) estimate that retiree health insurance raises retirement rates by between 3 and 5 percentage points (depending on gender and marital status) for workers under the age of 65. Marton and Woodbury (2006) find effects of a similar magnitude, in the range of 3–4 percentage points for males aged 51–61. Karoly and Rogowski (1994) estimate that retiree health coverage roughly doubles (from 12 percent to 24 percent) the probability of retirement over a 2-year period for men aged 55–62. Robinson and Clark (2010) use a proportional hazard model to show that individuals aged 50–60 with retiree health insurance are 21.2 percent more likely to leave their job. Strumpf (2010) finds that retiree health coverage raises the probability of being retired by 8 percentage points, on average, for individuals under the age of 65. Madrian (1994) finds that retiree health coverage reduces the age of retirement by 5–16 months. Mulvey and Nyce (2004) use matched employer-employee administrative data to demonstrate that retiree health coverage is associated with a reduction in retirement ages of 9.3 months for men and 1.6 years for women. Marton and Woodbury (2012) find that retiree health coverage reduces retirement among workers in their early 50s (when they are generally ineligible for benefits) but raises among workers in their early 60s (when they are generally eligible).

An alternative approach followed by a number of authors is to estimate a structural model of retirement, and to use the estimated model to simulate the impact of retiree health coverage. These studies tend to find smaller effects than those that use the reduced form approach. Blau and Gilleskie (2008) estimate that retiree health coverage reduces the labor force participation rate of older men by 3.6 percentage points. Studying the behavior of married couples, Blau and Gilleskie (2006) predict an increase in retirement probability of less than half a percentage point for men and 1.6 percentage points for women. Gustman and Steinmeier (1994) find that retiree health coverage increases the probability of leaving full-time employment at age 62 by 2.1 percentage points, a 16 percent increase over the baseline exit rate. Lumsdaine, Stock, and Wise (1996) find that retiree health insurance raises retirement rates between ages 60 and 64 by about 2 percentage points per year. French and Jones (2011) estimate that retiree health coverage raises the retirement rate at age 62 by 8.5 percentage points.

A third approach is to estimate the impact of retiree health coverage using aggregate state-level data, and variation in state and federal policy. Gruber and Madrian (1995) examine the state and federal “continuation of coverage” requirements that were adopted during the 1970s and 1980s. They find that these mandates encouraged earlier retirement. In particular, they find that the availability of COBRA coverage reduced the labor force participation rate of 55–64 year-olds by 3.3 percentage points.

We have access to a unique dataset that contains much more detailed information about firm-level benefits than has previously been available to researchers. In particular, our dataset is derived from the employee records of 54 firms, matched to a firm-level survey of benefit provisions. These firms are quite diverse in terms of their industry, size, location, and other characteristics. Our paper contributes to the existing literature in several ways. First, compared with the HRS and other public data sources, we have more detailed information on the size of the employer contribution towards retiree health coverage. Thus, we are able to separate out the effects of subsidized coverage (i.e., retiree health coverage that includes an employer contribution) and access only coverage (i.e., coverage that does not include an employer contribution but provides the benefit of a group plan). Second, our data on retiree health provisions are subject to less measurement error than the self-reported information in the HRS. It is generally difficult to get access to such firm-level data. (Blau and Gilleskie’s (2008) study is one exception – they are able to supplement the HRS data with detailed employer-provided data on retiree health coverage and pensions.) Finally, we have detailed information on each firm’s age and service eligibility criteria for retiree health coverage. Hence, we are able to compare the behavior of eligible and ineligible individuals within the subset of individuals whose employers offer retiree health coverage. This allows us to control for the possibility that individuals may select into such jobs based on unobservable characteristics that also affect retirement behavior. Our dataset is similar to the one used by Mulvey and Nyce (2004), who also use administrative records from Watson Wyatt Worldwide (which merged with Towers Perrin in 2010 to form Towers Watson), although we control for additional individual and firm-specific characteristics relative to that study.

3. Data

a. Employee-level data

Our employee-level data come from the administrative data on the employees of a large number of Towers Watson’s clients. These records compose a panel spanning the years 2005–2009 and containing the employee-level actuarial information necessary to evaluate each client’s pension liabilities. The dataset includes each individual’s employer name, hire date, birth date, gender, salary, and employment status on January 1 of the relevant year. We select an initial sample of employees who are active with complete demographic and pay information in 2005; this restriction effectively excludes individuals hired after January 1, 2005. Following this initial selection, we use the 2006–2009 data on these employees in our analysis. Our dependent variable is an indicator for not being employed by the firm on January 1 of the current year, conditional on being employed in the previous year. We study the relationship between this turnover indicator and the retiree health benefits offered by an individual’s employer. Additional individual-level controls include salary in 2005, gender, and years of service. We use salary in 2005 rather than current year salary because current year salary tends to be very low for those who retired during the year. We also exclude individuals with a 2005 salary that is less than $10,000 (as these individuals are highly likely to have worked a partial year) and person-year observations with less than one year of service. Additional details on the steps taken to clean up the data are given in Appendix A.

Our main analysis is based on 172,343 person-year observations for employees aged 58 to 69, with complete information on individual and firm-level characteristics. Theory predicts that retiree health insurance should have an impact on retirement behavior for individuals under age 65, but not for those 65 and older. As an additional falsification check, we also perform some analysis using person-year observations for employees aged 40–49 (who should also not be affected by retiree health provisions). Table 1 presents the individual characteristics for all observations in our main sample of workers aged 58–69. These statistics are presented for the full sample, and broken down by retiree health coverage and eligibility. In general, individuals who are eligible for retiree health coverage tend to have higher salaries and longer service compared to those who have no coverage or are ineligible for coverage. They are also somewhat more likely to be male. These differences are not surprising as many firms have a minimum service criterion for retiree health eligibility. Moreover, salary is likely to be related to years of service, and women are more likely to have career interruptions that result in shorter service. In our regressions, we control explicitly for these observable differences.

Table 1.

Individual Characteristics

	Towers Watson Sample
	Full Sample	No Coverage	Access Only Ineligible	Subsidy<50% Ineligible	Subsidy≥50% Ineligible	Subsidy Unknown Ineligible
Male	0.51	0.53	0.30	0.48	0.47	0.47
Years of Service	19.95	17.62	7.80	6.75	10.83	17.46
Age	61.18	61.38	61.09	60.81	61.19	59.91
2005 Salary	56120	47062	50609	54080	41478	64448
Fraction of Obs.	1	0.44	0.03	0.03	0.06	0.02

	Access Only Eligible	Subsidy<50% Eligible	Subsidy≥50% Eligible	Unknown Eligible
Male	0.32	0.54	0.59	0.40
Years of Service	23.23	23.85	27.82	24.02
Age	61.11	61.13	60.79	61.96
2005 Salary	65201	73951	64400	69254
Fraction of Obs.	0.06	0.11	0.20	0.05

	HRS Sample
	Full Sample	No Coverage	Coverage	Coverage Unknown	Men	Women
Male	0.54	0.51	0.57	0.55	1.00	0.00
Years of Service	15.72	14.32	17.74	14.57	16.57	14.72
Age	61.14	60.73	60.50	62.78	61.22	61.06
Salary (in 2005 $)	53772	48328	59137	53175	62523	43386
Married	0.69	0.62	0.70	0.75	0.79	0.56
Spouse Agea	59.52	59.63	58.72	60.55	57.64	62.65
Spouse in LFa	0.66	0.61	0.67	0.69	0.65	0.67
Fraction of Obs.	1.00	0.36	0.39	0.25	0.54	0.46

Notes: Towers Watson sample includes 172,343 person-year observations in which the employee is between the ages of 58 and 69. HRS sample includes person-year observations from the 2004–2008 waves in which the respondant is between the ages of 58 and 69, works full time, is not self-employed, has a salary (in 2005 dollars) of $10,000 or more, and has one year or more of service. For HRS sample, salary includes wages from all jobs, not just the primary job. HRS respondant weights are used in computing means. Retiree health coverage includes coverage only through own employer.

^a- For married individuals only. In labor force includes working full time, working part time, “partially retired,” and unemployed. Not in labor force includes retired, disabled, and out of labor force.

For comparison, the bottom panel of Table 1 shows the individual characteristics of a comparable sample of HRS respondents. We select observations from the 2004–2008 waves of the HRS in which the respondent is between the ages of 58 and 69, works full time, is not self-employed, has a salary (in 2005 dollars) of $10,000 or more, and has one year or more of service.¹ On average, the individuals in our sample are somewhat more highly compensated and have more years of service than those in the HRS sample.

One potential shortcoming of our employee data is that we do not have data on employee marital status, and in particular, on whether an individual’s spouse has access to employee or retiree health insurance. Thus, we cannot control for the availability of retiree health insurance through a spouse. As the HRS data show, the majority of individuals in our sample are likely to be married. Lack of marital status information may cause us to underestimate the impact of retiree health insurance on retirement incentives. In particular, some of the individuals in our no coverage and ineligible groups are likely to be covered through a spouse, and may therefore behave more like individuals who are eligible for retiree health coverage. Thus our findings would likely be stronger if we were able to correct for this downward bias.

Another potential shortcoming of our employee data is that we only observe departure from the firm, rather than retirement. For example, retiree health insurance may facilitate a transition to self-employment. We believe that the impact of retiree health insurance on turnover – whether turnover represents full retirement, partial retirement through part-time work, or a transition to self-employment – is an interesting question in its own right. But if one wishes to distinguish between retirements and transitions to self-employment, the relevant question is whether retiree health insurance encourages transitions to self-employment. Bruce, Holtz-Eakin and Quinn (2000) do not find evidence that employment-based health insurance reduces the probability of a transition to self-employment among older workers. Boyle and Lahey (2007) find that expansions of non-employer based health coverage for veterans increased self-employment among more educated workers and reduced it among less educated workers. Overall, such coverage reduced labor supply among veterans. Based on these studies, we believe that the effect of retiree health coverage on turnover can be best understood as a reduction in labor supply (i.e., full or partial retirement) rather than a substitution of self-employment for current employment.

b. Firm-level Data

We merge our individual data with a firm-level survey of pension and retiree health benefit provisions. The firm-level survey was completed by actuaries at Towers Watson who are familiar with the specific firms’ provisions. The survey collects data on the provisions applying to three employee groups: (1) the typical full-career employee expected to retire in in 2010, (2) the typical full-career employee expected to retire in 2020, and (3) new hires. Table 2 summarizes our rules for assigning employees to these groups. For example, the top panel of the table shows that for retiree health provisions, individuals with less than 5 years of service are classified as new hires, while individuals with 5–9 years of service are classified as either new hires or expected 2010 retirees depending on their age. The rationale for these assignments is based on the fact that provisions vary across groups because employers made changes to their benefits – for example, a firm might have made its pension plan less generous or stopped offering retiree health coverage. When such changes occur, current employees are often grandfathered into the old provisions based on their age and years of service. We do not have information on when these benefit changes occurred or what the rules were for grandfathering existing employees.² However, we can approximate the provisions that apply to specific employees based on their age and years of service. Individuals with 10 or more years of service (as of 2005) are classified as expected 2010 retirees for all provisions. In addition, older employees with less service are classified as expected 2010 retirees for certain provisions. However, younger employees who were hired more recently are less likely to be grandfathered into older provisions, or are more likely to have arrived after any changes were made. This is true even if these individuals are old enough to retire in 2010, rather than 2020 or later. Such individuals are classified as either new hires or expected 2020 retirees depending on their age and service.

Table 2.

Assignment of Provisions to Individuals

Retiree Health Provisions
Service	Age 55–59	Age 60–64	Age 65–69
1–4 years	New hires	New hires	New hires
5–9 years	2020	2010	2010
10+ years	2010	2010	2010

Defined Benefit Provisions
1) Generosity, Formula, and Eligibility
Service	Age 55–59	Age 60–64	Age 65–69
1–4 years	New hires	New hires	New hires
5–9 years	2020	2020	2020
10+ years	2010	2010	2010
2) Plan Status
Hire Date >Close Date		→	No DB
Hire Date < Close Date		→	Covered
Year < Freeze Date		→	Open
Year > Freeze Date		→	Frozen

Defined Contribution Provisions
Service	Age 55–59	Age 60–64	Age 65–69
1–4 years	New hires	New hires	New hires
5–9 years	2020	2020	2010
10+ years	2010	2010	2010

While all firms in our sample offer employee health coverage, they vary considerably in terms of their retiree health offerings. For each of the employee groups, we have information on the existence and generosity of retiree health coverage, as well as the age and service requirements for eligibility. Firms may provide either pre-65 coverage or both pre-and post-65 (“Medigap”) coverage. In this paper, we focus on pre-65 coverage. If retiree health coverage is provided, the actuaries completing the survey are asked to indicate whether retirees have “access only” or subsidized coverage. Subsidized coverage means that the employer contributes towards the employees’ health insurance premiums, while access only coverage means that retirees can participate in a group health plan but without the benefit of an employer subsidy.³ For each individual, we construct a set of indicator variables for whether the individual has no coverage, is ineligible for coverage (but works for a firm that offers coverage), is eligible for access only, or is eligible for subsidized coverage.⁴ Among individuals whose employers offer retiree health coverage, a majority are eligible by age 55, and more than 75 percent are eligible by age 62.

We also have information on the DB pension provisions applying to each of the three employee groups. For each group, we know the plan formula (traditional or hybrid), the plan status (open, frozen, or closed), and the age and service eligibility criteria for an early (reduced) pension and a full pension. A traditional plan promises an annuity benefit based on a formula related to an employee’s earnings history. A hybrid plan operates more like a defined contribution plan. Plan contributions are credited to a notional account that earns interest or credits at a stipulated rate. A plan is open if new hires are enrolled in the plan and existing participants continue to accrue benefits. A plan is closed if new hires are not enrolled in the plan, but existing participants continue to accrue benefits. A plan is frozen if new hires are not enrolled in the plan and existing participants no longer accrue benefits. If a plan is closed or frozen we know, in most cases, the year in which the change occurred. Respondents are also asked to rate the generosity of the firm’s DB plan on a scale of 1 to 5, with 5 being the most generous. Respondents were provided guidelines for categorizing the generosity levels based on the percentage of a member’s salary that is notionally put aside, commonly called an “accrual rate” for traditional DB plans or a “pay credit” for hybrid pension plans. The guidelines suggested that traditional DB plans with accrual rates of 1 percent or lower are low generosity, plans with accrual rates of around 1.3 percent are average, and plans with accrual rates of 2 percent or higher are high generosity. Likewise, hybrid plans with 3 percent pay credits are low generosity, plans with 7 percent pay credits are average generosity, and plans with 10 percent pay credits are high generosity. We consolidate generosity into three categories – above average (4 or 5), average (3), and below average (1 or 2).

Individuals are matched to DB formula (hybrid or traditional) and generosity measures based on their classification into the three cohorts (new hires, retiring in 2020, or retiring in 2010). However, DB plan status (open, frozen, or no DB) is assigned based on freeze and close dates as summarized in Table 2. For each individual, we construct a set of indicator variables for every possible combination of generosity (above average, average, below average), formula (hybrid, traditional), status (open, frozen), and eligibility (ineligible, early, or full). One shortcoming of the DB data is that some firms offer multiple DB plans covering different groups of workers. In such cases, the plan provision survey contains information on the firm’s main DB plan. The firm’s main plan may not be the actual plan applying to a particular employee. However, the features of the main plan indicate the general direction of the firm’s DB policy. For example, a firm that freezes its main DB plan or makes its main DB plan less generous is likely to be moving in the same direction for its other plans.

Many DB plans provide strong incentives for early retirement by reducing the benefit accrual rate for workers starting in their late 50s. It is important to control for these incentives, as they may be correlated with the availability of retiree health coverage. Ideally, we would do this by controlling for the increase in DB pension wealth that an individual would earn from working for an additional year. However, we do not have sufficiently detailed information on DB plan design to construct such a measure. Instead, we approximate the early retirement incentives by including as a control the firm-level turnover rate at age 57. If a firm’s DB plan provides strong early retirement incentives, we would expect to observe a high turnover rate at age 57. By age 55, almost half of those covered by a DB plan are eligible for an early pension, and by age 62, more than a quarter are eligible for an unreduced pension.

Our firm-level dataset also includes information on the generosity of the defined contribution (DC) plan for each of the three cohorts. We capture this by including measures of both matching and non-matching contributions. We summarize DC matching contributions by calculating the total match amount offered, as a percentage of pay, if the employee contributes to the maximum pay threshold. This is commonly referred to as the effective or total match rate. Some firms also offer non-matching contributions. In other words, they contribute funds without requiring an employee contribution. These contributions are often discretionary based on company performance. Since these contributions can vary from one year to the next, we include the average over 2008 to 2010 as a proxy for the typical contribution, expressed as an average percentage of pay.

Both DB and DC provisions are correlated with retiree health coverage. Table 3 shows the joint distribution of DB plan characteristics and retiree health provisions. Firms that have frozen or terminated their DB plans are far less likely to offer retiree health coverage than firms with open DB plans. Moreover, firms that offer retiree health coverage have, on average, more generous matching and non-matching DC contributions. These relationships highlight the importance of controlling for pension incentives in determining the impact of retiree health coverage on early retirement.

Table 3.

Joint Distribution of Defined Benefit Plan Type and Retiree Health Coverage

DB Plan	No Coverage	Access Only	Subsidy	Total
Above Avg Traditional - Open	1.6	0.3	8.9	10.8
Above Avg Traditional - Frozen	0.5	0.0	0.0	0.5
Avg Traditional - Open	5.7	3.0	10.4	19.1
Avg Traditional - Frozen	14.6	0.0	1.4	16.0
Below Avg Traditonal - Open	0.4	0.0	1.0	1.5
Below Avg Traditional - Frozen	1.9	0.0	0.0	1.9
Above Avg Hybrid - Open	0.1	2.7	14.3	17.1
Avg Hybrid - Open	2.2	0.8	5.1	8.1
Avg Hybrid - Frozen	2.4	0.0	0.0	2.4
Below Avg Hybrid - Open	4.7	0.2	3.9	8.7
Below Avg Hybrid - Frozen	0.0	0.0	2.5	2.6
Unknown Formula	0.0	1.7	0.0	1.7
No DB	9.7	0.0	0.0	9.8
Total	43.7	8.7	47.6	100.0

Notes: Based on 172,343 person-year observations in which the employee is between the ages of 58 and 69. Number shown in cell is percent of total observations.

Additional firm-level variables that we use in our analysis include the employer’s size (total employment) and a set of dummies for the employer’s industry. These variables are obtained from a Towers Watson client database. Industry classifications are based on two-digit North American Industrial Classification System (NAICS) codes. Compared to a nationally representative sample of firms, our firms tend to be considerably larger. For example, according to 2006 data from the Small Business Administration (SBA), the vast majority of firms (more than 99 percent) have less than 1,000 employees, and such firms account for 55.4 percent of employment.⁵ None of the firms in our sample are this small. In fact, 44.4 percent of firms in our sample have more than 10,000 employees, compared to only 0.02 percent of firms in the SBA data. These large firms account for 84.2 percent of employment among firms in our sample, compared to 27.0 percent of employment in the SBA sample. In terms of industries, our employees are more likely to work in manufacturing and retail compared to a nationally representative sample.⁶ Within our sample, construction, retail, and professional services firms are less likely to offer retiree health coverage than firms in other sectors. Additional details of these comparisons are provided in Appendix B. The appendix also provides additional detail on the fraction of employees at each age who meet the eligibility criteria for DB pensions and retiree health insurance.

4. Methods

We estimate logit models in which the dependent variable is the indicator for turnover described in the previous section. That is, we model the probability of no longer being employed in the current year, conditional on being employed in the previous year. Our key independent variables are a set of dummies indicating whether an individual has no retiree health coverage, is ineligible for coverage (but works for a firm that offers coverage), is eligible for access only, or is eligible for subsidized coverage. We also include a set of age dummies and their interactions with the retiree health indicators. Thus, we allow the impact of retiree health insurance to vary with age. Other individual-level explanatory variables include gender, years of service, years of service squared, and a set of indicator variables indicating the individual’s decile in the overall age-specific salary distribution. We also include year dummies and controls for the pension provisions applying to the individual. Pension provision controls include the set of DB indicator variables for every possible combination of generosity (above average, average, below average), formula (hybrid, traditional), status (open, frozen), and eligibility (ineligible, early, or full) as well as DC total match rate and DC nonmatching percentage. To control for early retirement incentives in the firm’s DB plan, we include the firm’s turnover rate at age 57. Finally, we include the employer’s size (total employment) and a set of dummies for the employer’s industry.

Men and women may differ in their responsiveness to retiree health insurance. Some earlier studies have found that men and women are similarly responsive to retirement incentives that operate through health coverage and pensions (Coile 2004; Royalty and Abraham 2006). However, Kapur and Rogowski (2011) find that women are less responsive to their own retiree health coverage if their husbands are not yet retired. Thus, we might find that retiree health coverage has a smaller overall impact among women because our estimates combine the effects for women with working husbands and those for women with retired husbands (or single women). In addition, the effects of retiree health coverage may vary with income level. For example, low-income workers may tend to have more physically demanding jobs, and high-income workers may have more retirement resources. Both of these factors suggest that high-income workers may have more flexibility in choosing their retirement date than low income workers, and may therefore be more responsive to retiree health insurance. Thus, in addition to estimating a regression for the full sample of workers age 58–69, we also estimate separate regressions for men, women, and individuals in the top and bottom half of the age-specific salary distribution.

Our approach may be inappropriate for estimating the impact of retiree health coverage on retirement if individuals select into retiree health coverage based on unobservable characteristics that also affect turnover. Alternatively, despite our controls for pension incentives, there may be unobservable firm characteristics that influence retirement decisions and are also correlated with retiree health coverage. For example, French and Jones (2011) present evidence to suggest that employees of firms that offer retiree health coverage tend to have a stronger preference for leisure than employees of other firms; indeed, they find that differences in labor force participation between these groups persist beyond age 65. Blau and Gilleskie (2001) find that accounting for selection into retiree health coverage raises the estimated impact of such coverage; on the other hand, French and Jones (2011) find that it modestly lowers the estimated impact of coverage.

We have several ways of dealing with the possibility of such selection. First, we control for firm-level turnover at age 57, which should capture unobservable factors such as leisure preferences that tend to raise early retirement rates for particular firms. Second, by using eligibility information for retiree health coverage, we are able to compare individuals who are eligible for coverage with those who are ineligible but work for firms that offer coverage. If we find an effect for the eligible individuals, but not for the ineligible ones, selection is less likely to be a problem. Finally, we perform an additional falsification check by estimating our model on individuals aged 40–49, the vast majority of whom are ineligible for retiree health coverage.

5. Results

Table 5 reports the results from estimating our model on the full sample that includes all person-year observations aged 58–69. The model in top panel includes all controls described in the previous section, while the model in the bottom panel includes only the retiree health dummies, age dummies, and their interactions. We report only the average marginal effects of the retiree health coverage indicators, broken down by age. The omitted category is no retiree health coverage. Standard errors – clustered by firm – are in parentheses. Results for the full logit model – from which these marginal effects are computed – are available upon request.

Table 5.

Impact of Retiree Health Coverage on Probability of Turnover

VARIABLES	(1) Age 58	(2) Age 59	(3) Age 60	(4) Age 61	(5) Age 62	(6) Age 63	(7) Age 64	(8) Age 65	(9) Age 66	(10) Age 67	(11) Age 68	(12) Age 69
Ineligible	0.012 (0.011)	0.003 (0.011)	−0.015 (0.011)	−0.001 (0.013)	−0.008 (0.015)	0.024 (0.018)	0.011 (0.016)	−0.077** (0.038)	−0.088*** (0.022)	−0.070* (0.037)	−0.051**(0.023)	0.011 (0.021)
Access Only Eligible	−0.019* (0.010)	0.003 (0.013)	−0.016 (0.015)	0.004 (0.018)	0.029 (0.035)	0.029 (0.030)	0.021 (0.020)	−0.037 (0.059)	0.034* (0.020)	0.036 (0.040)	−0.047* (0.027)	0.024 (0.037)
Subsidy Eligible	−0.000 (0.010)	0.006 (0.010)	0.007 (0.013)	0.009 (0.012)	0.063*** (0.020)	0.077*** (0.020)	0.055*** (0.016)	−0.025 (0.051)	0.015 (0.028)	0.012 (0.036)	−0.015 (0.027)	0.069** (0.029)
No Coverage Turnover	0.093	0.097	0.125	0.121	0.174	0.158	0.145	0.326	0.315	0.253	0.222	0.192
Controls	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes

VARIABLES	(1) Age 58	(2) Age 59	(3) Age 60	(4) Age 61	(5) Age 62	(6) Age 63	(7) Age 64	(8) Age 65	(9) Age 66	(10) Age 67	(11) Age 68	(12) Age 69
Ineligible	0.018 (0.014)	0.005 (0.013)	−0.007 (0.015)	0.006 (0.011)	0.002 (0.018)	0.032** (0.016)	0.016 (0.014)	−0.075 (0.050)	−0.082*** (0.026)	−0.062* (0.036)	−0.039 (0.027)	0.028 (0.028)
Access Only Eligible	−0.012 (0.025)	0.005 (0.021)	−0.020 (0.026)	−0.001 (0.021)	0.016 (0.018)	0.015 (0.033)	0.007 (0.030)	−0.059 (0.053)	0.007 (0.036)	0.022 (0.032)	−0.055 (0.039)	0.010 (0.049)
Subsidy Eligible	0.013 (0.010)	0.018* (0.010)	0.013 (0.014)	0.012 (0.013)	0.066*** (0.019)	0.076*** (0.019)	0.053*** (0.016)	−0.030 (0.048)	0.009 (0.024)	0.007 (0.033)	−0.015 (0.024)	0.069*** (0.021)
No Coverage Turnover	0.093	0.097	0.125	0.121	0.174	0.158	0.145	0.326	0.315	0.253	0.222	0.192
Controls	No	No	No	No	No	No	No	No	No	No	No	No

^***p<0.01,

^**p<0.05,

^*p<0.1

Notes: Regressions based on 172,343 observations. Coefficients are average marginal effects from logit model. Standard errors (clustered by firm) in parentheses. All regressions include age dummies, retiree health dummies, and their interactions. Top panel also includes controls for gender, years of service, years of service squared, firm turnover at age 57, firm size, DB and DC pension provision dummies, age-specific 2005 salary decile dummies, year dummies, and firm sector dummies.

Looking at the top panel of the table, subsidized retiree health coverage has its strongest effects at ages 62–64, raising the probability of turnover (relative to no retiree health coverage) by 6.3 percentage points at age 62, by 7.7 percentage points at age 63, and by 5.5 percentage points at age 64. These effects are substantial. Relative to no coverage, subsidized coverage is associated with a 36.2 percent increase in the probability of turnover at age 62, a 48.8 percent increase in the probability of turnover at age 63, and a 38.0 percent increase in the probability of turnover at age 64.⁷ They are also not very different from the results given in the bottom panel of the table, suggesting that controlling for other factors does substantially change the relationship between retiree health coverage and turnover. Overall, these findings are similar in magnitude to those reported in prior studies. We find no evidence of an impact for individuals aged 65 and older. Additionally, we find no evidence of behavioral differences between individuals with no coverage and individuals who work for firms that offer coverage, but do not meet the eligibility criteria for coverage. The latter two results increase our confidence that the effects found at ages 62–64 are indeed causal effects.

Access only has a statistically insignificant effect for most age groups. One possible explanation for this finding is that COBRA effectively provides access only coverage. It allows employees to purchase health coverage at group rates, but without an employer contribution, for 18 months after leaving their job. Thus, access only coverage would not have any additional value to individuals above the age of 63½, and only a small value to individuals aged 62.

As mentioned previously, theory suggests that subsidized retiree health coverage should influence retirement behavior for workers under the age of 65, but not for those aged 65 and older.⁸ Indeed, if those with no coverage are waiting for Medicare to retire, we would expect to see a decrease in turnover for the eligible group relative to the no coverage group at age 65, as the no coverage group leaves in large numbers. Consistent with the theory, the point estimate for 65-year-olds who are eligible for subsidized coverage is negative; however, it is relatively small in magnitude and statistically insignificant. In addition, individuals 65 and older who do not meet the eligibility requirements for pre-65 retiree health coverage appear to have a lower turnover rate than their counterparts at firms that do not offer pre-65 retiree health coverage.

We suspect these results arise because the 65-and-older group is fundamentally different from the under-65 group. For example, individuals who have chosen to stay on the job until age 65 – whether they have retiree health coverage or not – may have different preferences for leisure compared to those who left before age 65. In addition, as eligibility for pre-65 retiree health coverage never requires an age that is greater than 65, it must be the case that the ineligible individuals in the 65-and-older group have inadequate service to obtain coverage. Indeed, the ineligible individuals in this age group have less than 9 years of service on average, compared to more than 16 years of service on average for the no coverage comparison group. This is a rather unusual group of workers, as it consists of those who changed jobs or entered the labor force later in life. Thus, the negative coefficients on the ineligible dummy for ages 65 and older may be picking up the effect of being in this group. For example, recent hires aged 65 and older might be inadequately prepared to retire, and may therefore be more likely to stay to qualify for pension benefits.

Tables 6 and 7 report the results from estimating our model for men and women separately. Both men and women appear to be responsive to subsidized retiree health coverage. While the coefficients for men are slightly larger than those for women, they are within two standard deviations for the women’s coefficients. In Tables 8 and 9, we report the results from estimating our model separately for the top half and bottom half of the overall age-specific salary distribution. Again, the effects for more highly compensated workers are slightly larger, though still within two standard deviations of those for less highly compensated workers. Thus, we conclude that the impact of retiree health is fairly consistent across gender and pay groups.

Table 6.

Impact of Retiree Health Coverage on Probability of Turnover (Males)

VARIABLES	(1) Age 58	(2) Age 59	(3) Age 60	(4) Age 61	(5) Age 62	(6) Age 63	(7) Age 64	(8) Age 65	(9) Age 66	(10) Age 67	(11) Age 68	(12) Age 69
Ineligible	0.013 (0.014)	0.013 (0.015)	−0.020 (0.015)	−0.005 (0.020)	−0.003 (0.022)	0.033 (0.023)	0.014 (0.024)	−0.071* (0.041)	−0.059 (0.036)	−0.062 (0.046)	−0.059* (0.033)	0.016 (0.033)
Access Only Eligible	−0.014 (0.012)	0.007 (0.019)	−0.029* (0.017)	−0.007 (0.030)	−0.013 (0.023)	0.007 (0.046)	−0.023 (0.030)	−0.013 (0.060)	−0.073* (0.040)	0.001 (0.070)	−0.028 (0.037)	0.083 (0.127)
Subsidy Eligible	−0.000 (0.012)	0.007 (0.013)	0.003 (0.018)	0.000 (0.015)	0.065** (0.027)	0.094*** (0.025)	0.063*** (0.020)	0.007 (0.056)	0.044 (0.034)	0.045 (0.040)	0.011 (0.035)	0.088** (0.039)
No Coverage Turnover	0.097	0.104	0.139	0.130	0.193	0.174	0.158	0.314	0.319	0.252	0.239	0.203

^***p<0.01,

^**p<0.05,

^*p<0.1

Notes: Regression based on 88,283 observations. Coefficients are average marginal effects from logit model. Standard errors (clustered by firm) in parentheses. All regressions include age dummies, retiree health dummies, and their interactions, as well as controls for gender, years of service, years of service squared, firm turnover at age 57, firm size, DB and DC pension provision dummies, age-specific 2005 salary decile dummies, year dummies, and firm sector dummies.

Table 7.

Impact of Retiree Health Coverage on Probability of Turnover (Females)

VARIABLES	(1) Age 58	(2) Age 59	(3) Age 60	(4) Age 61	(5) Age 62	(6) Age 63	(7) Age 64	(8) Age 65	(9) Age 66	(10) Age 67	(11) Age 68	(12) Age 69
Ineligible	0.011 (0.014)	−0.004 (0.012)	−0.007 (0.014)	0.004 (0.012)	−0.007 (0.015)	0.022 (0.021)	0.014 (0.017)	−0.084** (0.039)	−0.111*** (0.022)	−0.078** (0.037)	−0.040 (0.025)	0.010 (0.027)
Access Only Eligible	−0.025** (0.013)	−0.001 (0.013)	−0.008 (0.018)	0.007 (0.014)	0.047 (0.037)	0.036 (0.032)	0.038* (0.020)	−0.068 (0.064)	0.056* (0.029)	0.037 (0.031)	−0.049 (0.042)	0.007 (0.038)
Subsidy Eligible	0.001 (0.012)	0.007 (0.011)	0.013 (0.015)	0.020 (0.016)	0.063*** (0.018)	0.060*** (0.019)	0.049*** (0.016)	−0.056 (0.051)	−0.008 (0.029)	−0.017 (0.037)	−0.030 (0.034)	0.058* (0.034)
No Coverage Turnover	0.088	0.089	0.109	0.109	0.152	0.141	0.130	0.338	0.311	0.253	0.205	0.180

^***p<0.01,

^**p<0.05,

^*p<0.1

Notes: Regression based on 84,060 observations. Coefficients are average marginal effects from logit model. Standard errors (clustered by firm) in parentheses. All regressions include age dummies, retiree health dummies, and their interactions, as well as controls for gender, years of service, years of service squared, firm turnover at age 57, firm size, DB and DC pension provision dummies, age-specific 2005 salary decile dummies, year dummies, and firm sector dummies.

Table 8.

Impact of Retiree Health Coverage on Probability of Turnover (Top Half)

VARIABLES	(1) Age 58	(2) Age 59	(3) Age 60	(4) Age 61	(5) Age 62	(6) Age 63	(7) Age 64	(8) Age 65	(9) Age 66	(10) Age 67	(11) Age 68	(12) Age 69
Ineligible	0.016 (0.010)	0.006 (0.013)	0.006 (0.014)	0.014 (0.018)	−0.029* (0.016)	0.015 (0.019)	0.014 (0.019)	−0.010 (0.028)	−0.017 (0.030)	−0.012 (0.033)	0.029 (0.038)	0.060 (0.048)
Access Only Eligible	−0.021** (0.009)	0.002 (0.018)	−0.025 (0.019)	−0.009 (0.020)	0.028 (0.037)	0.020 (0.034)	0.007 (0.024)	0.065 (0.051)	0.054 (0.045)	0.100** (0.050)	−0.055 (0.034)	0.046 (0.054)
Subsidy Eligible	−0.003 (0.010)	0.008 (0.009)	0.017 (0.014)	0.015 (0.011)	0.066*** (0.019)	0.075*** (0.021)	0.041** (0.017)	0.048 (0.035)	0.044* (0.027)	0.068** (0.027)	0.019 (0.019)	0.129*** (0.028)
No Coverage Turnover	0.092	0.097	0.119	0.118	0.166	0.166	0.160	0.245	0.291	0.219	0.199	0.168

^***p<0.01,

^**p<0.05,

^*p<0.1

Notes: Regression based on 86,168 observations. Coefficients are average marginal effects from logit model. Standard errors (clustered by firm) in parentheses. All regressions include age dummies, retiree health dummies, and their interactions, as well as controls for gender, years of service, years of service squared, firm turnover at age 57, firm size, DB and DC pension provision dummies, age-specific 2005 salary decile dummies, year

Table 9.

Impact of Retiree Health Coverage on Probability of Turnover (Bottom Half)

VARIABLES	(1) Age 58	(2) Age 59	(3) Age 60	(4) Age 61	(5) Age 62	(6) Age 63	(7) Age 64	(8) Age 65	(9) Age 66	(10) Age 67	(11) Age 68	(12) Age 69
Ineligible	0.004 (0.015)	−0.004 (0.014)	−0.032** (0.014)	−0.015 (0.015)	−0.002 (0.021)	0.021 (0.024)	−0.000 (0.016)	−0.126*** (0.038)	−0.144*** (0.024)	−0.116*** (0.044)	−0.107*** (0.028)	−0.029 (0.028)
Access Only Eligible	−0.023 (0.016)	0.001 (0.015)	−0.007 (0.020)	0.016 (0.018)	0.030 (0.035)	0.027 (0.033)	0.024 (0.023)	−0.119** (0.059)	0.017 (0.041)	−0.038 (0.040)	0.017 (0.045)	0.029 (0.027)
Subsidy Eligible	−0.005 (0.013)	−0.005 (0.014)	−0.010 (0.018)	−0.005 (0.017)	0.053* (0.028)	0.053** (0.021)	0.048** (0.020)	−0.072 (0.051)	−0.027 (0.042)	−0.069 (0.043)	−0.059 (0.038)	−0.032 (0.046)
No Coverage Turnover	0.094	0.097	0.129	0.122	0.179	0.153	0.135	0.375	0.332	0.276	0.238	0.209

^***p<0.01,

^** p<0.05,

^*p<0.1

Notes: Regression based on 86,175 observations. Coefficients are average marginal effects from logit model. Standard errors (clustered by firm) in parentheses. All regressions include age dummies, retiree health dummies, and their interactions, as well as controls for gender, years of service, years of service squared, firm turnover at age 57, firm size, DB and DC pension provision dummies, age-specific 2005 salary decile dummies, year dummies, and firm sector dummies.

As our data span 2006–2009, it is possible that we may pick up some of the effects of the financial crisis and recession of 2007–09. Our regressions include year dummies, which can be used to check whether the recession had an impact on turnover overall. For example, we might expect the adverse shock to retirement wealth to reduce turnover, although this effect may be small for our largely DB-covered population. Alternatively, it is possible that some firms provided incentives for older workers to leave in 2009.⁹ While the coefficient on the 2009 dummy (reflecting turnover in 2008) is negative relative to the omitted category of 2006, it is statistically insignificant at conventional levels. Thus, our regressions do not seem to reflect a substantial impact from the recent recession.

Overall, our results suggest that of an initial group of 57-year-old workers with no retiree health coverage, 36.8 percent will remain by age 65. This result comes from iteratively applying the predicted turnover rates from the model in Table 5. In contrast, of an initial group of 57-year-olds with subsidized coverage, only 28.2 percent will remain by age 65. Thus, conditional on working at age 57, retiree health insurance reduces the probability of work at age 65 by 23.3 percent relative to no coverage.

One way to see the cumulative effects of retiree health coverage is to put these results in terms of the number of person-years worked over ages 58–64. That is, we can apply the predicted turnover rates from model in Table 5 and estimate the total number of person-years worked for each coverage level over ages 58–64. Summing over this age range, we find that subsidized retiree health coverage reduces the number of person-years worked by 5.6 percent relative to no coverage. For a workforce with 2,500 employees aged 57, this implies a loss of over 600 person-years. This result is shown in the top two rows of Table 10. Because retiree health has its strongest effects from ages 62–64, we also compute the reduction in person-years worked in that age range, conditional on working at age 61. As shown in the second two rows of Table 10, we find that subsidized retiree health coverage reduces the number of person-years worked over ages 62–64 by 14.2 percent.

Table 10.

Cumulative Effect of Retiree Health Insurance

	No Coverage	Subsidy
Person-Years Worked (58–64)	10,857.83	10,250.80
Difference from No Coverage (%)	-	−5.6%
Person-Years Worked (62–64)	5,280.39	4,528.45
Difference from No Coverage (%)	-	−14.2%
E(Retirement Age | Working at 57)	62.34	62.10
Difference from No Coverage (months)	-	−2.91

Notes: Based on model in Table 5. Expected retirement age calculation is conditional on working at age 57 and assumes all individuals retire by age 65. Person-years worked based on initial workforce of 2,500 57-year olds (top panel) or 61-year-olds (middle panel).

Another way to interpret these results is to compute the impact of retiree health coverage on the expected age of departure (or retirement, assuming most departures in our age range represent retirements). If we assume that all 64-year-olds who are working retire at age 65, we can use the predicted turnover rates from the model in Table 5 to compute the full probability distribution over retirement ages for a 57-year-old worker with different kinds of retiree health coverage. From this probability distribution, we can calculate the expected retirement ages. Our results are given in the bottom two rows of Table 10. We find that subsidized retiree health coverage reduces the expected age of retirement, conditional on working at age 57, by almost three months

As a final falsification check, we estimate our regression model for workers aged 40–49, almost all of whom are ineligible for retiree health coverage. Thus, we would not expect the existence of such coverage to influence their turnover rates. Because most of these workers are ineligible for coverage, we do not incorporate eligibility information into our retiree health coverage indicators. That is, we include a set of indicators for whether the individual’s employer provides no coverage, access only, or subsidized coverage. We find no statistically significant effects for this younger group. We report marginal effects by age for these regressions in Appendix B.

6. Conclusions

For most people under the age of 65, group health insurance coverage is only available through employment. In this paper, we have presented evidence that the link between health insurance and employment may cause individuals to delay retirement until they are eligible for Medicare. In particular, we have shown that after controlling for individual characteristics and pension incentives, employees under the age of 65 have substantially higher turnover rates at firms that offer subsidized retiree health coverage compared to their counterparts at firms that do not. Moreover, higher subsidy rates are associated with greater turnover than lower ones. Subsidized coverage has its largest effect at ages 62–64, raising the turnover rate at age 62 by 6.3 percentage points (36.2 percent), by 7.7 percentage point (48.8 percent) at age 63, and by 5.5 percentage points (38.0 percent) at age 64. Our model predicts that, conditional on working at age 57, only 28.2 percent of individuals with subsidized coverage remain at their firm at age 65, compared to 36.8 percent of individuals with no retiree health coverage.

The Patient Protection and Affordable Care Act (PPACA), which became law in 2010, will soon weaken the link between health insurance and employment by making it possible for all individuals to buy group coverage regardless of employment status. It could also provide a considerable number of individuals with subsidies towards their health insurance premiums, depending on their household income and employer-provided coverage. For those who qualify for a subsidy, the new law provides a tax credit such that the premium a person pays does not exceed 9.5 percent of household income. For those with income less than 400 percent of the Federal Poverty Level, the subsidies are even greater. Older Americans who do not qualify for explicit subsidies can still expect to receive substantial implicit subsidies. The new law prohibits insurers from charging older individuals – even those with pre-existing conditions – premiums that are more than three times the rates paid by younger individuals. As such, many older workers across all income groups will have new opportunities for affordable, non-employment based health care coverage that is comparable to today’s employer-provided subsidized pre-65 retiree medical coverage today.

Based on our results, we would expect these new alternatives to increase retirement rates among older workers who are below Medicare eligibility age. The primary effect will be on those with current employer-provided coverage who would not be able to obtain retiree coverage if they left their jobs; health care reform effectively provides these individuals with some level of subsidized retiree coverage.

We are hesitant to use our results to make a projection for the population as a whole given the complex structure of the subsidies and the fact that our sample consists of individuals whose employers have offered defined benefit (DB) pension plans in the past. However, this would be a valuable undertaking for future research. Despite these limitations, our results still suggest that the effects of affordable retiree medical coverage, of the sort that will be available to all Americans in 2014 under PPACA, could have a substantial impact on future retirement patterns.

The social welfare implications of this change in retirement incentives are not straightforward. Viewed by itself, the link between health insurance and employment (which results from the favorable tax treatment of employer-provided coverage) distorts retirement decisions, and breaking that link would increase efficiency. However, viewed in the context of other policies that affect retirement incentives, this may not be the case. For example, Social Security imposes high implicit tax rates on older workers and inefficiently encourages early retirement (see, e.g., Goda, Shoven, and Slavov 2009). For individuals below Medicare eligibility age, there is currently a countervailing effect: employment-based health insurance discourages early retirement and mitigates the distortion caused by Social Security. Breaking the link between employment and health insurance removes this countervailing effect, thereby amplifying the distortion caused by Social Security and potentially reducing efficiency. A valuable area for future research would be to examine the impact of PPACA in the context of other policies that affect the retirement incentives of older workers, including not only Social Security, but also other entitlement programs and private pensions.

Table 4.

Defined Contribution Paramters by Retiree Health Coverage

	Nonmatching % of Pay	Total Match Rate
No Coverage	0.41	1.60
Access Only	1.40	2.19
Subsidy	0.58	3.42

Notes: Based on 172,343 person-year observations in which the employee is between the ages of 58 and 69.

Table B5.

Impact of Retiree Health Coverage on Probability of Turnover (Three Subsidy Levels)

VARIABLES	(1) Age 58	(2) Age 59	(3) Age 60	(4) Age 61	(5) Age 62	(6) Age 63	(7) Age 64	(8) Age 65	(9) Age 66	(10) Age 67	(11) Age 68	(12) Age 69
Ineligible	0.013 (0.011)	0.004 (0.011)	−0.013 (0.011)	0.001 (0.013)	−0.005 (0.016)	0.026 (0.018)	0.013 (0.016)	−0.074* (0.038)	−0.085*** (0.022)	−0.068* (0.037)	−0.049** (0.023)	0.013 (0.021)
Access Only Eligible	−0.017 (0.011)	0.006 (0.014)	−0.014 (0.015)	0.007 (0.018)	0.033 (0.036)	0.032 (0.031)	0.024 (0.020)	−0.031 (0.059)	0.040* (0.020)	0.041 (0.040)	−0.043 (0.027)	0.028 (0.037)
Subsidy<50% Eligible	−0.008 (0.009)	−0.010 (0.010)	−0.011 (0.012)	−0.004 (0.012)	0.050** (0.021)	0.098*** (0.035)	0.065*** (0.023)	0.010 (0.052)	0.054 (0.034)	0.062* (0.038)	−0.004 (0.022)	0.116** (0.051)
Subsidy≥50% Eligible	0.006 (0.011)	0.015 (0.011)	0.017 (0.016)	0.022 (0.015)	0.082*** (0.026)	0.082*** (0.021)	0.053*** (0.018)	−0.038 (0.052)	−0.007 (0.030)	0.003 (0.039)	−0.014 (0.031)	0.058** (0.028)
Subsidy Unknown Eligible	−0.004 (0.014)	0.020 (0.025)	0.009 (0.017)	−0.011 (0.015)	0.040* (0.021)	0.040* (0.024)	0.057*** (0.017)	−0.029 (0.057)	0.028 (0.035)	−0.031 (0.039)	−0.021 (0.048)	0.050 (0.058)
No Coverage Turnover	0.093	0.097	0.125	0.121	0.174	0.158	0.145	0.326	0.315	0.253	0.222	0.192

^***p<0.01,

^**p<0.05,

^*p<0.1

Notes: Regressions based on 172,343 observations. Coefficients are average marginal effects from logit model. Standard errors (clustered by firm) in parentheses. All regressions include age dummies, retiree health dummies, and their interactions, as well as gender, years of service, years of service squared, firm turnover at age 57, firm size, DB and DC pension provision dummies, age-specific 2005 salary decile dummies, year dummies, and firm sector dummies.

Table B6.

Impact of Retiree Health Coverage on Probability of Turnover (Ages 40–49)

VARIABLES	(1) Age 40	(2) Age 41	(3) Age 42	(4) Age 43	(5) Age 44	(6) Age 45	(7) Age 46	(8) Age 47	(9) Age 48	(10) Age 49
Access Only	0.007 (0.011)	0.013* (0.008)	0.001 (0.009)	0.003 (0.007)	0.005 (0.006)	0.002 (0.007)	0.004 (0.008)	0.004 (0.008)	0.008 (0.010)	0.012 (0.009)
Subsidy	−0.006 (0.010)	−0.001 (0.010)	−0.009 (0.009)	−0.002 (0.008)	0.001 (0.008)	−0.005 (0.008)	−0.005 (0.007)	−0.004 (0.008)	0.000 (0.008)	−0.000 (0.007)
No Coverage Turnover	0.112	0.104	0.110	0.104	0.098	0.099	0.094	0.093	0.089	0.087

^***p<0.01,

^**p<0.05,

^*p<0.1

Notes: Regression based on 389,710 observations. Coefficients are average marginal effects from logit model. Standard errors (clustered by firm) in parentheses. All regressions include age dummies, retiree health dummies, and their interactions, as well as controls for gender, years of service, years of service squared, firm turnover at age 57, firm size, DB and DC pension provision dummies, age-specific 2005 salary decile dummies, year dummies, and firm sector dummies.

Highlights.

• We study the impact of retiree health insurance on the early retirement.

• We utilize a unique administrative dataset from a benefits consulting firm.

• Retiree health coverage increases turnover by 6.3 percentage points at age 62.

• Retiree health coverage increases turnover by 7.7 percentage points at age 63.

• Retiree health coverage reduces years worked between 58 and 64 by 5.6 percent.

Appendix A: Data

In this appendix, we detail the steps taken to clean up the employee-level administrative data, as well as the firm plan provisions data.

Employee Data

To begin, we drop individuals with missing birth dates and hire dates, individuals whose gender is coded inconsistently across years, and individuals with hire dates that imply they were hired before the age of 16. We convert all salaries to 2005 dollars using the CPI-U-RS. We exclude any person-year observations with less than one year of service or a salary of less than $10,000 in 2005. A low salary is likely to indicate a partial year and may result if, for example, an employee leaves or retires during the year.

A worker’s employment status is “active” if he or she is currently working and accruing benefits. We use employment status to create an indicator variable for employee turnover, which serves as the dependent variable in our analysis. For each employee, the turnover indicator for the current year takes on a value of zero if the employee is active as of January 1 of the current year, and was also active on January 1 of the previous year. The turnover indicator takes on a value of 1 if an employee who was active on January 1 of the previous year is inactive (retired, disabled, on leave, etc.) or missing from the dataset on January 1 of the current year. The turnover indicator is missing in all other cases (i.e., for employees who were not active in the previous year).

Our turnover indicator does not distinguish between retirements and other reasons for leaving the firm or dataset. For example, an employee may leave the firm to take a job at another firm. In addition, an employee can be dropped from the dataset for a number of reasons. In some cases, we do not have administrative data for a firm for a given year because of a merger or acquisition, divestiture, bankruptcy or severed client relationship with Towers Watson. In these cases, all employees of the firm are missing from the sample for a given year. Specific individuals can show up missing in a given year, yet be active in the previous year. This could reflect an employee retiring from the company and taking a lump sum benefit (rather than an annuity), which eliminates the pension liability from the employer’s books. In addition, a number of employee records are simply missing for some firms in particular years even if the employees have not departed or retired.

To deal with these issues, we impose several data restrictions. First, we drop all employees of a firm in a given year if the firm has a calculated turnover rate of more than 30 percent or less than 1 percent for that year. Second, we restrict the data to firms in which at least 60 percent of the active 2005 employees remain in the database in years 2006 to 2009 and do not drop out of the sample with an unknown status change. For the second criterion, note that individuals with a known status change (e.g., from active to retiree status) are retained in the sample for the entire sample period. Firms that fail to meet this criterion are excluded from the sample for that year. Finally, we drop all observations for a firm if we lose more than 40 percent of the firm’s observations due to invalid salary data (either missing our under $10,000).

In addition to these restrictions, we exclude one large firm that offers its employees a phased retirement option. Employees taking this option are classified as having left the firm, resulting in a very high measured turnover rate – in the 70-to-90 percent range – at age 65. Thus, measured turnover at this firm does not reflect actual turnover; for many workers, it reflects the start of a phased retirement. We also exclude all 2009 observations for one firm that we learned – through public sources – offered its older workers a buyout during 2008.

Provisions Data

Some firms have missing or inappropriate responses on benefit plan provisions or eligibility requirements for one or more of the three cohorts. We impute many of these responses by substituting values for another cohort.

In addition, we fill in some missing retiree health provision information from another firm-level survey conducted jointly by Towers Watson and the International Society of Certified Employee Benefits Specialists (ISCEBS). The Towers Watson/ISCEBS survey includes information for 2010 and 2020 retirees, and there is some overlap with the firms included in our provisions survey. If the alternative survey indicates coverage for 2010 retirees and no coverage for 2020 retirees, we assume no coverage for new hires as well.

Defined Benefit (DB) plans are classified as either hybrid or traditional. Any responses other than these are coded as either hybrid or traditional depending on the information provided. For example, “career average” is coded as traditional and “5 percent cash balance” is coded as hybrid.

Most closed or frozen DB plans provide a close or freeze year. For some firms, multiple close or freeze years are provided – for example, a response may indicate that benefit accruals for service and pay, or for different subsidiaries, were frozen in different years. In these cases, we used the earliest freeze year provided. Several firms indicate a freeze or close date of “Before 2003.” These are coded as 2003. If a DB plan close date is provided, but not a freeze date, and there is an indication that the plan is frozen (i.e., status is given as frozen for one or more of the employee groups), we assume the plan was frozen in the same year that it was closed to new entrants. If a freeze date is provided, but no close date, we assume the plan was closed to new entrants in the same year that it was frozen.

Some firms report eligibility requirements for early (reduced) retirement, but not for full retirement. For these firms, we assume the requirements for full retirement are an age of 65 and 5 years of service. We drop any remaining firms for which we lack early DB, full DB, or retiree health eligibility information. We also drop any remaining observations with missing provisions data.

Firm Characteristics Data

We obtain size and industry data on each firm from a Towers Watson client information database. For some firms, the size data appear to be for a division of the company, rather than the company as a whole. For example, for several firms, the number of employee records we have in our dataset exceeds the stated size. Thus, we define firm size as the maximum of the size given in the client information database and the number of 2005 employee records in our dataset. For one firm, we replace size with the firm’s stated size on its website. Industry is defined by two-digit North American Industrial Classification System (NAICS) code. For a handful of firms, we have invalid (and in one case, apparently incorrect) industrial classifications. As we know each firm’s name, we are able to update the NAICS codes in these instances.

Appendix B: Additional Tables

Table B1.

Firm Size Distribution

	Nationwide		Towers Watson Sample
Firm Size	Fraction of Firms	Fraction of Employment	Fraction of Firms	Fraction of Employment
0–999	99.8%	55.4%	0.0%	0.0%
1,000–1,499	0.05%	3.0%	1.9%	0.2%
1,500–2,499	0.04%	3.8%	5.6%	0.7%
2,500–4,999	0.03%	5.4%	25.9%	5.5%
5,000–9,999	0.02%	5.4%	22.2%	9.4%
10,000+	0.02%	27.0%	44.4%	84.2%

Notes: Nationwide distribution of firm sizes is based on 2006 data from the Small Business Administration, retrieved from http://archive.sba.gov/advo/research/us92_07ss.txt.

Table B2.

Employment by Sector

		Nationwide		Towers Watson Sample
NAICS code	Industry	Total	Percent	Total	Percent
11	Agriculture, forestry, fishing and hunting	153,829	0.1%	0	0.0%
21	Mining, quarrying, and oil and gas extraction	604,653	0.4%	2,200	0.2%
22	Utilities	641,552	0.5%	41,466	4.3%
23	Construction	5,967,128	4.4%	39,229	4.1%
31–33	Manufacturing	11,632,956	8.5%	229,041	23.7%
42	Wholesale trade	5,827,769	4.3%	46,000	4.8%
44–45	Retail trade	14,802,767	10.8%	330,645	34.2%
48–49	Transportation and warehousing	4,159,604	3.0%	25,900	2.7%
51	Information	3,288,109	2.4%	14,800	1.5%
52	Finance and insurance	6,171,240	4.5%	70,798	7.3%
53	Real estate and rental and leasing	2,036,590	1.5%	0	0.0%
54	Professional, scientific, and technical services	7,839,965	5.7%	11976	1.2%
55	Management of companies and enterprises	2,853,450	2.1%	4670	0.5%
56	Administrative and support and waste management and remediation services	9,060,987	6.6%	0	0.0%
61	Educational services	3,200,553	2.3%	0	0.0%
62	Health care and social assistance	17,531,142	12.8%	136,250	14.1%
71	Arts, entertainment, and recreation	2,010,339	1.5%	0	0.0%
72	Accommodation and food services	11,443,293	8.4%	0	0.0%
81	Other services (except public administration)	5,264,429	3.8%	0	0.0%
92	Public administration	22,479,000	16.4%	15000	1.5%
	Total	136,969,355	100.0%	967,975	100.0%

Notes: Nationwide data for all NAICS codes except for 92 (public administration) come from the U.S. Census Bureau’s 2009 Statisics of Business, retrieved from http://www.census.gov/econ/susb/. Data on government employment (for the public administration category) are for the end of 2009 and come from the Federal Reserve Bank of St. Louis, retrieved from http://research.stlouisfed.org/fred2/data/USGOVT.txt.

Table B3.

Joint Distribution of Retiree Health Coverage and Industry

NAICS code	Industry	No Coverage	Access Only	Subsidy	Total
21	Mining, quarrying, and oil and gas extraction	0.0	0.0	0.6	0.6
22	Utilities	0.0	0.8	3.9	4.7
23	Construction	4.5	0.0	0.0	4.5
31–33	Manufacturing	6.1	0.9	14.2	21.2
42	Wholesale trade	0.0	0.0	3.8	3.8
44–45	Retail trade	25.8	0.0	5.9	31.7
48–49	Transportation and warehousing	0.0	0.0	1.0	1.0
51	Information	0.1	0.0	1.4	1.5
52	Finance and insurance	0.5	2.5	6.2	9.2
54	Professional, scientific, and technical services	0.7	0.0	0.0	0.7
55	Management of companies and enterprises	0.0	0.0	0.1	0.1
62	Health care and social assistance	6.0	4.5	9.3	19.9
92	Public administration	0.0	0.0	1.1	1.1
	Total	43.7	8.7	47.6	100.0

Notes: Based on 172,343 person-year observations in which the employee is between the ages of 58 and 69. Number shown in cell is percent of total observations.

Table B4.

Eligibility for DB Pension and Retiree Health

	Early DB		Full DB		Retiree Health
Age	Ineligible	Eligible	Ineligible	Eligible	Ineligible	Eligible
40	100.0%	0.0%	100.0%	0.0%	100.0%	0.0%
41	100.0%	0.0%	100.0%	0.0%	100.0%	0.0%
42	100.0%	0.0%	100.0%	0.0%	100.0%	0.0%
43	99.8%	0.2%	99.8%	0.2%	100.0%	0.0%
44	99.7%	0.3%	99.7%	0.3%	100.0%	0.0%
45	99.5%	0.5%	99.5%	0.5%	100.0%	0.0%
46	99.3%	0.7%	99.3%	0.7%	100.0%	0.0%
47	99.0%	1.0%	99.0%	1.0%	99.7%	0.3%
48	98.6%	1.4%	98.6%	1.4%	99.5%	0.5%
49	98.3%	1.7%	98.4%	1.6%	99.3%	0.7%
50	97.3%	2.7%	98.3%	1.7%	86.1%	13.9%
51	97.0%	3.0%	98.1%	1.9%	85.9%	14.1%
52	96.7%	3.3%	98.1%	1.9%	84.8%	15.2%
53	96.4%	3.6%	98.0%	2.0%	84.5%	15.5%
54	96.5%	3.5%	98.1%	1.9%	84.6%	15.4%
55	50.7%	49.3%	96.5%	3.5%	31.5%	68.5%
56	51.1%	48.9%	96.8%	3.2%	31.1%	68.9%
57	52.1%	47.9%	97.1%	2.9%	29.5%	70.5%
58	50.0%	50.0%	96.7%	3.3%	29.6%	70.4%
59	45.8%	54.2%	96.8%	3.2%	29.6%	70.4%
60	40.4%	59.6%	97.0%	3.0%	22.2%	77.8%
61	38.4%	61.6%	97.4%	2.6%	22.1%	77.9%
62	38.2%	61.8%	74.7%	25.3%	21.9%	78.1%
63	37.9%	62.1%	74.9%	25.1%	23.7%	76.3%
64	38.4%	61.6%	75.1%	24.9%	23.8%	76.2%
65	38.4%	61.6%	40.8%	59.2%	24.0%	76.0%
66	39.4%	60.6%	42.4%	57.6%	24.9%	75.1%
67	40.9%	59.1%	44.7%	55.3%	27.2%	72.8%
68	42.6%	57.4%	47.8%	52.2%	28.2%	71.8%
69	42.5%	57.5%	48.7%	51.3%	26.8%	73.2%

Notes: DB Eligibility based on sample of individuals covered by a traditional DB plan. Retiree health eligibility based on sample of individuals whose employers offer pre-65 retiree health benefits. Retiree health eligibility for individuals 65+ based on criteria for individuals under 65.