Abstract
This article uses the longitudinal aspect of the Health and Retirement Study to explore the characteristics associated with reversals in retirement (referred to here as “unretirement”). Through the use of survival-time analysis with time-variant covariates, this article shows that health insurance status and its source are significant predictors of unretirement decisions. The relationship is important, as the potential impacts of the Affordable Care Act (ACA) are considered. The analysis finds that insurance is equally as predictive of retirement reversals as other financial explanations such as pensions and wealth at retirement. The analysis also shows that health insurance sources play a particularly predictive role for early retirees and those who were previously open to the idea of working in retirement. Rough estimates suggest that the ACA might reduce the number of reversals by between two and four percentage points, which would translate to 80,000 and 170,000 retirees annually.
Most research in the area of retirement has focused on an individual’s binary choice to retire or to continue working, with much of the debate among economists centered on what model best fits retirement decision behavior. This line of research often neglects the fact that an individual’s working career does not always end with retirement. 1 A nontrivial share of retirees choose to return to work either on a part-time or full-time basis after fully retiring, or return to full-time work after partially retiring (collectively referred to in this article as “unretirement”). Although generally ignored in retirement modeling in the existing literature, data from the Health and Retirement Study (HRS) show that between 25 percent to 35 percent of retirees later reversed their retirement decisions. This value is consistent with the rate of retirement reversals previous found by Ruhm (1990) using the 1970s Retirement History Longitudinal Survey. As the demographic makeup of the labor force in the United States changes, it is important to understand what factors may influence the movement of the increasing number of retirees in and out of the labor force. A better understanding of unretirement should help inform policy decision-making.
One important policy question that needs to be addressed is how the implementation of the Affordable Care Act (ACA) might impact the labor market, especially among the large group of workers from the baby boom generation who are approaching or have reached common retirement ages. The implementation of the law will have a potentially large impact on those who choose to retire before they reach Medicare’s minimum age requirement (currently sixty-five) because the ACA guarantees access to health insurance and limits the costs to consumers in multiple ways. Well over 50 percent of retirees are not eligible for Medicare at the time of their retirement, with the median retirement age at about sixty-four for men and sixty-two for women (Munnell 2015). Although previous research has identified a strong link between health insurance and the initial retirement decision, the scant research on retirement reversals has almost entirely ignored the potential relationship between health insurance and unretirement. 2 This article puts health insurance at the center of the examination and compares the results to other potential predictors of unretirement. Although a causal link cannot be proven with the available data, the following survival-time analysis using the longitudinal HRS with time-variant covariates finds a retiree’s source of health insurance to be a strong predictor of future retirement reversals.
The small, existing literature that examines retirement reversals has either excluded health insurance or come to contradictory findings. Ruhm (1990) found that those with pensions and higher levels of education were less likely to unretire than others when controlling for income, age, and gender but omitting health insurance and health care costs. Maestas (2010) identified preretirement expectations as a primary predictor of unretirement and found that health insurance provision (as measured simply by the loss of health insurance) and other financial changes at the time of retirement were not significant indicators. Maestas and Li (2007) employed a hazard model approach and found a statistically significant role for a retiree’s health insurance circumstances but did not examine this result in detail. 3
A separate and more in-depth examination of the relationship of health insurance and unretirement is important for a number of reasons. First and foremost, this study finds that the correlation between health insurance sources and unretirement is on par with financial concerns such as household wealth and pensions. The findings of this study suggest that postretirement behavior may change due to the ACA, and this should be accounted for when modeling labor force participation and other fiscal impacts of the law. This potential response may also need to be considered when predicting the impact from other potential changes to health insurance availability for older Americans, such as calls to increase the Medicare eligibility age to sixty-seven which was most recently proposed by House Speaker Paul Ryan (Beech and Cowan 2016). If the statistical relationship between health insurance and unretirement is interpreted as purely causal, the findings of the empirical analysis that follow would suggest that the ACA could decrease the overall unretirement rate by two to four percentage points (or by 8 percent to 16 percent), which would translate to roughly 80,000 to 170,000 fewer retirement reversals.
There are many good reasons to believe that the empirical relationship between health insurance sources and unretirement found in this analysis may be causal, at least in part. Specifically, the potentially high cost of health insurance and health care could be a surprise that jeopardizes an individual’s retirement when workers retire based on inaccurate or overly optimistic beliefs about their insurance options and future health. Older Americans have been shown in surveys and economic studies to be deficient in their financial retirement planning, which suggests they might also be making errors in their planning involving health care and health insurance needs. Potential retirees have been shown to have an inaccurate knowledge of their pension and Social Security income that influences their decisions (Gustman and Steinmeier 2005; Chan and Stevens 2008). Schur et al. (2004) found a similar disparity between near-retirement-aged individuals’ beliefs about their employers’ retiree health benefits and national averages. Planning for future health care expenses can be particularly difficult, given the potential postretirement changes to employer health benefits, the high variability in health care expenses, and consumers’ relatively low familiarity with and the complexity of the health care payment system, health insurance, and the nongroup health insurance market.
The second section describes the health insurance market in the United States, before and after the ACA, and describes American consumer’s knowledge and past decision-making regarding health insurance choices. The third section describes the HRS and the sample used. The fourth section defines retirement and unretirement. The fifth section discusses the general methodology used. The sixth section presents the main results, including a number of specification checks and detailed analysis of specific subgroups. The seventh section uses the estimates from the empirical analysis to conduct a naive simulation of the potential impact of the ACA on unretirement rates, given different potential changes to health insurance availability and costs under the law. The eighth section concludes.
Health Insurance and the ACA
Access to health insurance and its cost are important elements of retirement planning, especially since expected medical care increases with age and will often make up a large share of a retiree’s expenses. The federal Medicare program is the primary health insurance source for this group, but it is only available to those who are sixty-five years of age or older unless they have experienced a qualifying disability. For those who retired prior to reaching age sixty-five, a decreasing number are able to get continued coverage through their former employers (often referred to as retiree health insurance—RHI). Those retirees without offers of RHI or employer-provided health insurance (EPHI) from a spouse’s employer are faced with the choice of risking high medical costs by going uninsured or attempting to obtain private, nongroup insurance. The availability and the cost of the latter option have been dramatically changed by the ACA.
Prior to the ACA, private, nongroup health insurance plans were generally more expensive and less generous than EPHI with high variability in each policy’s terms. The Congressional Budget Office (2005) estimated that the average annual premium for a private, nongroup plan was a third higher than that of EPHI in 2002 despite the fact that this estimate does not control for the level of coverage (covered expenses and the level of copayments and deductibles). Furthermore, insurance companies in most states retained the option to limit benefits for preexisting conditions and to deny coverage to individuals whom they deem too risky or expensive, with estimates that companies rejected 10 percent to 14 percent of all applicants (Pauly and Nichols 2002; Merlis 2005) and up to 37 percent of those with preexisting conditions (Pollitz, Sorian, and Thomas 2001).
The major implications of the ACA for nongroup health insurance stem from the provisions that guarantees the issuance of standarized policies for anyone not covered by an employer or government health insurance program. Most significantly, insurance policies can no longer be refused, limited, or have different pricing based on preexisting medical conditions. For older Americans, this is particularly important, as they are more likely to have chronic medical issues than younger Americans. Additionally, the coverage and out-of-pocket maximums have to meet minimum standards, so retirees can be confident in the product they are purchasing. The ACA also limits the ratio between the premiums charged for older and younger Americans to three, which will likely result in some indirect subsidization of more costly, older customers by less costly, younger customers. Finally, the ACA provides government subsidies based on income with no consideration of accumulated wealth. Under the law, those with household incomes below 400 percent of the federal poverty level (US$46,680 for an individual in 2015) are eligible for subsidies that will make insurance more affordable. 4
There are many reasons related to health insurance for why retirees may make decisions regarding their retirement timing that are ex post suboptimal. First, the health benefits they expected to have in retirement may have changed. While some may have misunderstood their employer RHI benefits, others may have had their benefits change after retirement. Unlike pension benefits, RHI plans are subject to revision and rescission and are not insured by the federal government. A 2003 survey found that over 80 percent of large employers planned to increase cost sharing for retiree health benefits in the next year and 20 percent planned to terminate all subsidies for future retirees (Hewitt Associates and Henry J. Kaiser Family Foundation 2004). Some firms eliminate RHI all together. A Labor Department report stated that 2 percent of 1994 retirees lost their promised retirement health insurance benefits in the previous year (Government Accountability Office 1998) and a similar decline of 2 percent was found between 2013 and 2014 in a survey of employers by Towers Watson/National Business Group of Health (2014). This finding is consistent with the falling offer rates of EPHI and RHI. Surveys of employers have found that the percentage of employers offering EPHI fell from 68 percent in 2000 to 59 percent in 2007 and 55 percent in 2014, while the rate of RHI offers among large employers offering EPHI went down from 66 percent in 1988 to 40 percent in 1999 and 25 percent in 2014 (Kaiser Family Foundation and Health Research and Education Trust 2007, 2014).
A second possible explanation for why retirees may make suboptimal choices about the timing of their retirement relates to potential errors in forecasting medical costs and a lack of understanding of health insurance cost sharing. Forecasting errors could be due to unexpected health costs. Gruber and Madrian (1996) found that early retirement-aged individuals not only have higher average medical care expenditures than younger Americans but also much higher variability in those expenses. These high medical costs might be particularly unexpected for the large portion of retirees who are optimistic about their future health and coverage. A 2011 poll found that 87 percent of preretired individuals over fifty reported that their future health would be the same or better in retirement (Robert Wood Johnson Foundation 2011), while another poll found that 40 percent of the same population is not saving for retirement health care costs with almost 50 percent reporting that their costs will “be taken care of” and a third of all respondents reporting that they “won’t have health costs” (Skufca 2014). Although these findings may be true for some, it is also likely a sign of optimism regarding future health.
Even if a retiree is insured, he or she might find their coinsurance costs to be much higher than they expect. Surveys have shown that American consumers have generally poor “health insurance literacy.” A 2007 employee survey found that less than a third of workers felt they could explain common insurance terms such as “lifetime maximum” and “out-of-pocket medical expenses” (DiCenzo and Forstin 2008). A Kaiser Family Foundation survey found that only about half of adults could define the terms “deductibles” and “out-of-pocket maximums” and under 40 percent could calculate the cost of care with a formula using those concepts (Norton, Hamel, and Brodie 2014). For older Americans, studies of Medicare beneficiaries have found a similar lack of detailed knowledge about how the program works (Uhrig et al. 2006).
Finally, the complexity of health insurance offerings, especially in nonstandardized, nongroup, private health insurance markets, could have led consumers to make suboptimal decisions even if they have a good understanding of how health insurance works. A study by Sinaiko and Hirth (2011) found that a substantial number of employees choose an EPHI that was strictly dominated by another offered plan. Abaluck and Gruber (2011) and Kling et al. (2012) found evidence that older Americans made suboptimal decisions in their choices of Medicare drug plans despite the high availability of policy terms and the expected medication needed.
Data
The analysis that follows uses detailed longitudinal data on a nationally representative sample of American households from the HRS, produced by the Institute for Social Research at the University of Michigan. The empirical analysis will focus on retirees from the HRS prior to the 2008 recession and the passage of the ACA to limit confounding other factors during this traumatic economic period. The initial cohort of the HRS includes households where at least one member was between the ages of fifty-one and sixty-one in 1992. A newer cohort was added in 1998 and includes households in which one spouse was between the ages of fifty-one and fifty-six at that time. The HRS includes data from reinterviews of respondents that occur every two years, with the most recent interview “wave” analyzed here occurring in 2006. The HRS data used here are available through the Survey Research Center at the Institute for Social Research and the RAND Center for the Study of Aging (St. Clair 2008).
Although the HRS includes over 16,000 respondents, this study will only examine a restricted subset based on their observable characteristics and retirement timing. Specifically, those who were not age eligible for either cohort (e.g., respondents who are younger or older than their age-eligible spouses) were eliminated along with those who became deceased or otherwise attrited during the sample period. Because this study uses employment information from the wave prior to retirement, individuals are also required to be working in the first wave that their cohort was surveyed. Finally, respondents must be observed to retire and appear in at least one proceeding wave. With this restriction, each respondent has at least one opportunity to be observed to reverse his or her retirement decision. Upon unretirement, respondents are not reintroduced if they retire again. The final analysis sample includes 3,421 HRS respondents who are observed for an average of 2.7 waves after retirement before either reversing their retirement or the end of the observation period.
Definitions and Summary Statistics
In the HRS, individuals are given multiple opportunities to identify themselves as “retired.” As in much of the literature, full and partial retirement will be defined by both the amount worked by a respondent and their self-reported retirement status. Those who work full-time (defined as thirty-five hours or more per week and at least thirty-six weeks in the last year) are not considered retired regardless of their self-designation. Those working part-time are identified as partially retired if they self-report their retirement status as retired or partially retired. Finally, anyone not working and self-reporting being retired is considered “fully retired,” while those not working and not identifying themselves as retired (the unemployed, disabled, and those not in the labor force but not retired) will be excluded from the sample.
When defining retirement and unretirement, one must consider how to treat partial retirement. For the purposes of this study, the primary analysis will treat partial retirement as a form of retirement, based on the premise that the respondent has made a classification that any work he or she is doing is part of her retirement plan. The following section will also present results for alternative classifications of partial retirement, though the results are not substantively different. Since partial retirement is used to identify the onset of retirement, a “directional definition” of unretirement is used. Under this definition, a respondent would have to move from a higher to a lower state of retirement to be considered unretired. The highest state of retirement for these purposes would be full retirement, followed by partial retirement and not retired. For example, if a respondent was fully retired in the previous wave, he or she would be considered unretired if he or she moved to either a partially retired or not retired state. Respondents who are only partially retired will be considered to be unretired if they move to full-time employment or part-time work that they do not self-identify as partial retirement.
Table 1 presents the unretirement hazard rates following retirement. The hazard rate in this context is defined as the rate that unretirement occurs in a two-year period, given that it has not occurred prior to that period. The hazard rate is highest between the first observed retirement wave and the second wave (a period of two years) at 18.5 percent. The unretirement hazard falls consistently following the first wave after retirement, but over 5 percent of retirees unretire between six and eight years after they are first observed to retire. In total, over 31 percent of those in the sample were observed to unretire in some wave. Both values are similar to the rates found by Ruhm (1990; 25 percent to 35 percent) and Maestas (2010; 24 percent using the HRS but limiting the observation period to five years following retirement). The number of respondents that have “survived” to each wave following retirement are included in table 1, along with the exit rate due to the end of HRS waves studied.
Unretirement Hazard Rate by Wave after Retirement.
Note: All sample exits occurred after the 2006 wave of the Health and Retirement Study.
Although the key focus for this study is the relationship between health insurance sources and unretirement behavior, table 2 first looks at the role of demographic (age, gender, race, marital status, educational attainment, and Census region), health (own and spouse’s self-rated health), and financial (total wealth at retirement and pensions) factors. The table compares the mean values of these factors at the time of retirement for those who are not observed to reverse their retirement (permanent retirees) and future unretirees. At retirement, future unretirees are statistically significantly younger, male, more highly educated, and have better self-rated health. There is no statistical difference between the two groups based on marital status, but respondents are less likely to unretire in the future if their spouses were also retired. There is not a statistical difference between groups based on their spouses’ self-rated health, race, and region (with the exception of a respondent being statistically less likely to unretire if he or she lives in the north Census region). Although there is not a statistical difference in total household wealth at retirement between groups, permanent retirees are significantly more likely to report receiving a pension at their retirement wave than are future unretirees.
Comparison of Characteristics at Retirement.
Note: “Subsidized” and “unsubsidized” employer-provided health insurance (EPHI) refers to whether an employer contributes to the payment of the premiums of the EPHI.
Table 2 also presents the means for a number of health insurance sources at the time of retirement. Health insurance sources are divided into five categories based on expected cost and risk of large health-care expenditures: provision through a governmental program; provision through an employer or union subsidized program (“subsidized” EPHI tend to have the lowest consumer paid premiums among the private health insurance options); provision through an employer or union-based program in which the recipient pays the full cost (“unsubsidized” EPHI); provision through a private, nongroup health insurer (highest premiums, some exposure to large health care expenditures due to variations in the terms of policy, potential limits based on preexisting condition, and uncertainty of coverage following the current contract); and those without health insurance (high exposure to large health care expenditures). 5 Subsidized or unsubsidized EPHI could be provided by either the respondent’s current or former employer or by his or her spouse’s employer. Table 2 presents the share of respondents reporting each of these health insurance sources at the time they were first observed as retired. Future unretirees are significantly less likely to report subsidized EPHI or government-provided health insurance and more likely to report having no health insurance or private, nongroup health insurance than are permanent retirees. Statistically, there is not a significant difference in the rate that future unretirees and permanent retirees report having unsubsidized EPHI.
Empirical Methodology
To allow for changes in key variables across time, this study will primarily use a survival-time analysis to identify the statistical relationship between key static and dynamic characteristics and the unretirement hazard rate. The survival-time methodology accounts for censoring (in this case from the end of the survey) and for varying lengths of observation. It is also better equipped than static models to evaluate the effects of shocks and other changes to a respondent’s circumstances throughout retirement. This design element is one of the reasons survival-time models are commonly used in studies of unemployment and welfare spell duration, as many individuals do not become employed or do not exit welfare before the observation period ends. This study assumes a parametric hazard function and uses the Weibull distribution as the form of the baseline hazard, which will allow for negative duration dependence (Wooldridge 2002). 6 The proportional hazard model used can be expressed as
where t is measured in waves since retirement, λ0(t) captures duration dependence (Weibull distribution), and Xit represents the vector of covariates examined in the fourth section with additional interaction terms. Observations are weighted based on the HRS sampling weights.
As evidence of the importance of allowing key indicators to vary with time, table 3 summarizes the changes in health insurance sources between each respondent’s retirement wave and the period of observed, continued retirement. Of the 9 percent of respondents who report no health insurance at their retirement wave, only 36 percent report being uninsured in a subsequent wave. Among those who report private, nongroup health insurance at their retirement wave, 10 percent later report being uninsured. Similarly, 5 percent of those with unsubsidized EPHI at retirement later report being without health insurance and 13 percent later report having nongroup, private health insurance. Although 44 percent of respondents report employer-subsidized EPHI at retirement, 4 percent later report being uninsured, 7 percent report having private, nongroup insurance, and 11 percent report having unsubsidized EPHI during some future wave of their retirement. Not surprisingly, government programs tend to have the highest continuation rate, with no more than 3 percent reporting any other insurance source during their retirement.
Observed Changes to Health Insurance Sources after Retirement.
Note: “Subsidized” and “unsubsidized” EPHI refers to whether an employer contributes to the payment of the premiums for employer-provided health insurance (EPHI). Percentages only include waves after an individual retires but before any retirement reversals. Rows sum to more than 100 percent since each row represents health insurance sources in all waves after their retirement wave which may, therefore, include more than one source. HI = health insurance.
The survival-time analysis employed in this study examines the unretirement hazard between the current wave and the following wave, given that a respondent has not unretired prior to the current wave. Covariate values are based on the current wave, unless otherwise noted. For easier interpretation, the results are presented as mean marginal effects (MMEs), which represent the average change in the predicted unretirement hazard across the sample associated with a one-unit change in the covariate. 7 To be clear, the MMEs identify the statistical link between the covariate and the predicted hazard but cannot be interpreted as a causal relationship based on the available data and empirical methodology.
Estimation Results
Before examining the association of various health insurance sources with the unretirement hazard, table 4 presents the results of a survival-time analysis when only demographic, wealth, and health variables are included. Similar to the comparison of means described in the fourth section, the analysis finds that older retirees (over sixty-two), those in poor health, and those whose spouse was also retired have lower unretirement hazard rates while those with a college degree or who self-report their race as black have higher hazard rates. Although men in general do not have a statistically significant change in their hazard rate, men who were married or rated their health “poor” or “fair” have significantly higher hazard rates while the opposite is true for those with a spouse with low self-rated health. The results in table 4 also suggest that wealth at retirement is not a strong predictor of respondents’ unretirement hazard rates, but the receipt of a pension has a statistically significant and negative association with the likelihood of a retirement reversal.
Survival Analysis of Unretirement Including Demographic, Wealth, and Health Controls.
Note: Robust standard errors in brackets. Mean marginal effect represents the average impact of a one-unit change on the predicted hazard across the sample. Dummy variables for Northeast, Midwest, and West Census divisions are also included. Only the Northeast division is significantly different from the South Census division (negative relationship). Analysis weighted at the household level.
*Significant at 10 percent.
**Significant at 5 percent.
***Significant at 1 percent.
The results when health insurance sources variables are introduced to the survival-time analysis are presented in table 5. Although not reported in the table, the MMEs for demographic, health, and wealth factors do not change dramatically when health insurance sources are introduced. Column (1) presents the results using the primary definitions of retirement and unretirement identified in the fourth section. The MMEs for the four included health insurance categories show a clear pattern in relation to the omitted group, those reporting subsidized EPHI. There is a five percentage point increase in the hazard rate for those with nongroup, private health insurance and an eight percentage point increase for those with no health insurance, with the former statistically significant at the 10 percent level and the latter at the 1 percent level. In terms of magnitude, these MMEs are on par with that of retirement pensions for the retiree. On the other hand, retirees with the government insurance are over twenty percentage points less likely to reverse their retirement decision than those with subsidized EPHI. Those with an employer-provided plan but with unsubsidized EPHI have a three percentage point higher predicted rate of unretirement, though this is not statistically significant at traditional levels.
Survival Analysis of Unretirement Including Health Insurance Status and Alternative Categorizations of Retirement Status.
Note: Robust standard errors in brackets. Mean marginal effect represents the average impact of a one-unit change on the predicted hazard across the sample. “Subsidized” and “unsubsidized” EPHI refers to whether an employer contributes to the payment of the premiums for employer-provided health insurance (EPHI). Analysis weighted at the household level. PT = part-time work; PR = partially retired; FT = full-time work; FR = fully retired.
*Significant at 10 percent.
**Significant at 5 percent.
***Significant at 1 percent.
The results in columns (2) through (4) of table 5 suggest the correlations between retirement health insurance sources and unretirement rates are consistent across various definitions of retirement and unretirement. Column (2) presents the results when an individual is identified as retired only if they are fully retired and not working, while unretirement would then be defined as returning to any work. Column (3) examines only those who were working full-time when first entering the HRS, defining retirement as any exit from that full-time work and unretirement as a return to full-time work. Finally, column (4) represents a hybrid of these definitions by allowing retirement to begin when an individual is either partially or fully retired, but unretirement only occurs if the retiree moves to full-time work. Although the size of the MME might fluctuate based on these definitions, the direction and statistical significance of the MMEs associated with each health insurance category are consistent with the results from the primary definition in column (1).
To more closely examine the statistical relationship between health insurance and retirement reversals, table 6 presents the results for specific subgroups based on a retiree’s age, insurance sources prior to retirement, and a priori expectations of work in retirement. In column (1), the analysis is limited to only retirees that are under sixty-five years of age because this group has not yet qualified for Medicare and is therefore most likely to be impacted by health insurance. As expected, the MMEs are higher in magnitude for those without insurance or having nongroup insurance in this subsample, but the larger standard errors result in lower levels of statistical confidence. To identify the association for those transitioning from low-cost EPHI to either no insurance or higher cost, nongroup insurance, the second column of table 6 shows the results when the sample is limited to only those who report EPHI before retiring. The MME for those transitioning to no insurance is similar to that of the full sample analysis. Although the statistical significance falls below standard thresholds with a p value of .11, this finding still suggests that the results from the primary analysis are not simply reflecting the employment behaviors of those who are uninsured both before and after retirement. The MME for those with nongroup insurance is positive but only half as large as in the primary analysis, while the MMEs for unsubsidized and government insurance were relatively consistent with earlier results.
Survival Analysis of Unretirement in Subsamples.
Note: Robust standard errors in brackets. Mean marginal effect represents the average impact of a one-unit change on the predicted hazard across the sample. “Subsidized” and “unsubsidized” EPHI refers to whether an employer contributes to the payment of the premiums for employer-provided health insurance (EPHI). Planned work in retirement question was only asked in the initial, 1992, wave of the Health and Retirement Study. Analysis weighted at the household level.
*Significant at 10 percent.
**Significant at 5 percent.
***Significant at 1 percent.
The final two columns of table 6 examine the relationship between health insurance and retirement reversals when accounting for a retiree’s predisposition toward working in retirement. Using the first three waves of the HRS and a time-invariant model, Maestas (2010) found that a 1992 question asking respondents about their work expectations in retirement was the most important predictor of future unretirement while the loss of health insurance at retirement was not a significant predictor. Based on the same 1992 HRS question, column (3) presents the results of the primary analysis when the sample is limited to the 30 percent of the original HRS sample who reported they did not expect to work in retirement and column (4) presents the results for the other 70 percent. Although the two samples have very similar MMEs for the demographic variable explored in table 4, they have very different MMEs for their health insurance source variables. While nongovernmental health insurance sources do not appear to impact the unretirement hazard of those who did not plan to work in retirement, being uninsured or possessing nongroup health insurance leads to a significantly higher hazard rate for those who did plan to work in retirement. This result suggests that health insurance sources are strongly linked to the unretirement hazard rate for those who are considering future work as part of their retirement, but not those who were not planning to work. One possible reason is that those who are considering retirement reversals are more likely to respond to unexpected health care costs by returning to work than those who see retirement as a permanent decision. 8
Applying the Empirical Results to the ACA
To explore the potential impact of the ACA on retirement reversals, the empirical results above can be used to simulate the predicted retirement rates under different assumptions regarding the changing health insurance requirements and options under the ACA, including the requirement to purchase insurance, guaranteed access, cost controls, and subsidies. Specifically, the predicted unretirement hazard rates presented below will be based on the coefficients found using the hazard model described in the fifth section and whose MMEs were presented in column (1) of table 6. 9 Using the hazard model coefficients and each HRS respondent’s reported data within the sample, an individual’s predicted hazard is calculated and then aggregated to predict a full-sample unretirement hazard rate. The first two rows of table 7 compare the actual unretirement hazard to the predicted hazard values based on the above methodology. Compared to the actual values, the model predicts lower rates of unretirement in the initial wave and the final three waves of observed retirement, but overpredicts unretirement rates in the second and third waves. Overall, the empirical model underpredicts the percentage of the sample that will be observed to unretire by about five percentage points. Despite this difference, the predicted values are a fair approximation of the unretirement hazard rates under the health insurance landscape prior to the ACA.
Predicted Unretirement Hazard Rate Following the ACA.
Note: In the above predictions, all other variables are as reported by the respondent. “Subsidized” and “unsubsidized” EPHI refers to whether an employer contributes to the payment of the premiums for EPHI. ACA = Affordable Care Act; EPHI = employer-provided health insurance.
The bottom two rows of table 7 predict the unretirement hazard if the ACA is assumed to impact the respondents’ reported health insurance sources in particular ways, but a number of caveats are necessary before examining those results. The ACA is likely to have a very broad impact on the labor market. In addition to the aforementioned changes to the nongroup insurance market, the law requires large employers to offer full-time employees health insurance options or potentially pay a fine. Based on the prior research, an increase in the offer rate of EPHI could have an impact on the timing of retirement, which in turn could change the analysis of unretirement directly. Additionally, a higher offer rate of EPHI could coax more people out of retirement if they were more likely to receive health insurance with a full-time position. Finally, as mentioned earlier, the empirical analysis does not identify a causal link between health insurance and retirement.
In the third row of table 7, the ACA’s individual mandate to purchase health insurance and guaranteed access to that insurance with limited premium and coinsurance costs are examined. Specifically, the individual mandate should remove being uninsured as an option for retirees. In addition, the limitation on premiums and coinsurance with the minimum required benefits should bring the cost and coverage of health insurance more closely in line with those policies offered by employers prior to the ACA. This implies that those retirees who were previously uninsured or who purchased private, nongroup health insurance will instead have the opportunity to purchase insurance more like that of unsubsidized EPHI. To incorporate this change, predicted hazard rates are recalculated, using the same hazard model coefficients but with the underlying data modified so that anyone who reported no insurance or nongroup health insurance is assumed to have acquired unsubsidized EPHI instead. This method assumes that retirees do not qualify for federal subsidies to purchase health insurance (addressed in the final simulation). With this assumption, the predicted hazard rate drops by 1.6 percentage points in the first two years following retirement and 1.2 percentage points in the second two years. The overall unretirement rate is reduced by two percentage points, from 26.1 percent to 24.1 percent. Given that roughly 4 million Americans retire each year, this would project to a difference of about 80,000 unretirements. 10
The direct subsidies provided to low-income Americans under the ACA could lead to an even larger potential impact than those suggested above. Specifically, the federal government provides subsidies toward insurance premiums for those earning below 400 percent of the federal poverty line. The income limit for these subsidies is well above the average annual Social Security benefit of about US$15,220. If we assume most retirees are eligible for these subsidies, the out-of-pocket cost of insurance premiums and coinsurance for retirees will closely resemble the cost of subsidized EPHI prior to the law. The fourth row of table 7 presents the predicted unretirement hazard rates if everyone with no insurance, nongroup insurance, or unsubsidized EPHI were reassigned to have subsidized EPHI. With these assumptions, the predicted overall unretirement rate would be down 4.4 percentage points from the original baseline, which would translate to about 176,000 unretirements.
Conclusion
Despite the fact that unretirement is an important and relatively common phenomenon among retirement-aged Americans, it has been largely ignored in the existing literature. This article has set out to identify important factors related to the decision to reverse one’s retirement and specifically identify the potential channels in which the ACA’s reform of health insurance markets might influence these decisions. After controlling for a number of traits including gender, the influence of coordinated retirements, wealth at retirement, pensions, and health concerns, the analysis found that retirees’ lack of health insurance or reliance on private, nongroup health insurance is related to higher rates of unretirement, statistically similar to that of lacking a pension. The pattern of results was consistent across a variety of definitions for retirement and unretirement, while the relationship between health insurance and unretirement was especially acute for retirees under sixty-five years of age and those who envisioned working in retirement when asked in 1992. These findings suggest that the more accessible and lower cost insurance options provided under the ACA might lead to fewer retirement reversals than under the old regime, potentially on the magnitude of 80,000 to 175,000 per year.
The findings here suggest that previous retirement models should be modified to include the option of “unretirement” to truly capture retirement behavior. If future research can identify the reasons for the strong statistical link between RHI and retirement reversals, adapting existing models appropriately will allow policy makers to better evaluate the impact of changes to the health-care payment system in the United States. The implications of changes to the health care system have been studied in the context of the choice to retire, but not in the area of a retiree’s decision to return to work. The results suggest that the more accessible health insurance markets mandated under the ACA and the governmental subsidies for low-income individuals could decrease the likelihood of unretirement substantially.
Footnotes
Author’s Note
The findings and conclusions expressed are solely those of the author and do not represent the views of the Social Security Administration, any agency of the Federal government, or the Michigan Retirement Research Center
Acknowledgments
Special thanks to Charlie Brown, Jeff Smith, Tom Buchmueller, and David Albouy for their helpful guidance and support. Additional thanks to Robert Baumann, Melissa Boyle, Nzinga Broussard, Taryn Dinkelman, Ann Ferris, Samara Gunter, Jon Lanning and seminar participants at the University of Michigan and the College of the Holy Cross for their helpful input. All errors and omissions are my own.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was partially supported by a grant from the Social Security Administration through the Michigan Retirement Research Center (Grant # 10-P-98358-5).