Health Insurance and Retirement of Married Couples. David M. Blau and Donna B. Gilleskie. University of North Carolina at Chapel Hill

Transcription

1 Health Insurance and Retirement of Married Couples David M. Blau and Donna B. Gilleskie University of North Carolina at Chapel Hill December 2005 Abstract Most health insurance in the U.S. is provided by employers until eligibility for public health insurance (Medicare) begins at age 65. Retiring before 65 exposes workers who lack retiree health insurance coverage to the risk of catastrophic medical expenditure. We solve and estimate a dynamic model of the employment behavior of older married couples that includes risky medical expenditure and health insurance. Parameter estimates imply that the risk-reducing feature of health insurance can account for about half of the observed association between retiree health insurance and employment for married men, but can account for only one tenth of the much larger observed association for married women. Policy simulations imply very small effects on employment of changing the age of eligibility for Medicare from 65 to 67. JEL: J26 Keywords: health insurance, retirement, employment, Medicare Corresponding author: David Blau, Department of Economics, Gardner Hall, University of North Carolina, Chapel Hill, NC , david_blau@unc.edu, , Fax: Thanks to the National Institute on Aging grant R01 AG13406 for funding, and to John Bound, Charlie Brown, Eric French, Brigitte Madrian, John Rust, Jon Skinner, Ken Wolpin, seminar participants, and referees for helpful suggestions.

2 1. Introduction Most health insurance in the United States is provided by employers until eligibility for public health insurance for the elderly (Medicare) begins at age 65. Some employer health insurance plans provide coverage for retired workers, but others do not. The absence of retiree health insurance (RHI) coverage creates a link between employment decisions and health insurance coverage that may affect the incentive to retire before age 65. This link does not exist for workers who are eligible for RHI coverage from their employers, or who have coverage from other sources. Risk-averse workers who would like to retire before age 65 but who lack access to RHI may have an incentive to remain employed until age 65, in order to avoid exposure to the risk of catastrophic medical expenditure between the age of retirement and the age of Medicare eligibility. This is an important policy issue because reform proposals that would make health insurance coverage independent of employment status could increase the already-high rate of retirement before age 65, thus worsening the financial condition of Social Security and Medicare. Evidence from a variety of sources shows that there is a strong empirical association between retiree health insurance and employment of older individuals (e.g., Blau and Gilleskie, 2001; Madrian, 1994). An important study by Rust and Phelan (1997) showed that this association could be explained quantitatively by aversion to medical expenditure risk. They specified a forward-looking model of employment behavior in which RHI affects behavior through its impact on the risk of catastrophic medical expenditure. They estimated the distribution of medical expenditure for individuals with and without employer-provided health insurance and assumed that individuals behave as if they face these distributions. Their estimated model fits the data well: the difference between the medical expenditure 1

3 distributions for individuals with and without RHI together with the estimated risk aversion parameter can account for most of the large difference in the timing of exit from employment. Previous studies by Gustman and Steinmeier (1994) and Lumsdaine et al. (1994) estimated structural models that did not account for the risk-reducing feature of health insurance, and their estimates implied virtually no causal impact of health insurance on retirement. In this study, we revisit this important issue. We follow the approach of Rust and Phelan (1997) but extend their analysis in several ways. First, health insurance is often a family affair, so we model the joint employment decisions of married couples. Second, we include in our sample individuals who are covered by employer-provided pension plans. This makes our sample more representative than the sample analyzed by Rust and Phelan, who were forced to exclude men with private pensions as a result of lack of information on the provisions of the pension plans. Third, we model out-of-pocket medical expenditure as determined by health insurance plan rules, such as the premium, deductible, and coinsurance rate, applied to a random draw from a distribution of total medical expenditure. This approach provides a realistic and tractable link between employment decisions, health insurance, and medical expenditure risk. Finally, we analyze data from the 1990s, which provide a more up-to-date basis for policy analysis than the data from the 1970s used by Rust and Phelan. Data from the Health and retirement Study (HRS), reveal that after controlling for age, health, non-wage income, earnings, pension benefits, and Social Security benefits, older married men with employer-provided health insurance are about six percentage points less likely to be employed at a given point in time if they are covered by RHI than if they lack such coverage. Our estimated structural model is able to account for about half of this difference. A similar comparison for married women 2

4 shows that women with employer-provided health insurance are about 20 percentage points less likely to be employed if they have access to RHI than if they lack such access. Our estimated model cannot explain most of this large difference: it predicts a two percentage point difference. As in Rust and Phelan (1997), health insurance enters our model only through the budget constraint and expectations, so this suggests that the risk-reducing feature of health insurance is not the only factor behind the observed differences in employment choices by RHI coverage. Older individuals may value health insurance for other reasons not captured in our model. We use our results to simulate the impact of policy reforms that break the link between employment and health insurance The simulations indicate that even drastic reforms, such as universal health insurance independent of employment status, will have modest effects on employment behavior of married couples at older ages. A more realistic reform, increasing the age of eligibility for Medicare from 65 to 67, is predicted to have tiny effects on employment. These findings are rather surprising: they are more like those predicted by models that ignore risk aversion than like those of Rust and Phelan (1997). Yet they are generated by a model like that of Rust and Phelan (1997), in which risk aversion plays a central role. And the findings appear to be robust, as discussed in detail near the end of the paper. Two factors seem to account for our findings. First, the association between RHI and employment behavior is not as large as previously thought for men. As we show in the following section, over two thirds of the raw difference in employment behavior of men by RHI status is accounted for by other factors correlated with both RHI and employment, particularly Social Security and pension benefits. Previous studies have not been able to control for Social Security and pension benefits as carefully as we can, and this seems to matter a lot. Second, Rust and Phelan (1997) did not 3

5 actually use their model to simulate the impact of a reform such as changing the age of eligibility for Medicare, so we do not know what their estimates imply for a realistic reform of this type. A small effect of a two year change in Medicare eligibility may in fact be consistent with a fairly large impact of RHI such as found by Rust and Phelan (1997). Evidence in French and Jones (2004) supports this explanation: their structural model predicts a very small impact on employment of changing the age of Medicare eligibility from 65 to 67. In another paper in which we jointly model medical care and retirement decisions (Blau and Gilleskie, 2005a), we also find a very small predicted impact of this hypothetical policy change. 2. Descriptive Overview The HRS sampled men and women aged in 1992 and their spouses, and has interviewed them every two years since In this section, we describe the relationships among employment status, defined by whether an individual is employed at the survey date, employment transitions between survey dates, and health insurance coverage, using data from the first four waves of the HRS (1992, 1994, 1996, and 1998). Table I shows differences in employment status and employment transitions between individuals who are covered by employer-provided health insurance with and without retiree coverage at the first wave, estimated from logit models. Panel A shows that husbands with coverage from their own employer are 20.8 percentage points more likely to be employed if they lack RHI. This difference falls to 12.3 points with controls for age, earnings, health, and nonwage income; and falls further to 6.4 points with controls for current and future pension and Social Security benefits. 1 The raw 1 Social Security and pension benefits were computed using administrative Social Security and pension data described in more detail below. The estimates in column 3 are based on a model that 4

6 difference for wives in panel B is 34.6 percentage points. This falls to 29.5 points with the first set of controls, and 20.5 with the additional controls. The results are very similar if the sample is limited to individuals under the age of 63 or 65. These results suggest a large role for health insurance in early retirement decisions for women, and a sizeable role for men. Panels C and D show that the two-year exit rate from employment is higher for individuals with RHI than without it. The difference is about three percentage points for both men and women after controlling for other factors. These two-year transition rate differences are much smaller than those in the data used by Rust and Phelan (1997). Our extensive controls for earnings, health, and future Social Security and pension benefits may explain the smaller association. This highlights the importance of modeling the budget constraint as accurately as possible, a task to which we devote considerable attention below. 3. The Model and Estimation 3.1 Overview The key features of the model are that health insurance can help couples smooth the marginal utility of consumption across states of the world with different levels of medical expenditure, and for some couples health insurance is linked to employment. Because health insurance coverage is valuable to risk averse consumers, a worker whose own or spouse s insurance coverage is tied to his employment status may make different employment decisions than an otherwise similar worker whose controls for the Social Security benefit at every possible age of entitlement and the pension benefit at every possible age of exit from the pension-providing firm, assuming constant real earnings until the age of eligibility. This very flexible specification accounts for spikes in pension benefit accrual and in Social Security benefits. 5

7 insurance coverage is independent of his employment status. Three key assumptions are imposed for reasons of computational feasibility: there is no saving, health insurance coverage is not subject to choice (except via the choice of whether to continue in a job with employer-provided insurance without retiree coverage), and medical expenditure is not a choice variable. By forcing agents to satisfy a series of period-specific budget constraints, and by restricting choice over health insurance and medical expenditure, the model will provide an upper bound estimate of the risk-reducing effect of health insurance on employment decisions. If couples could self-insure by saving in anticipation of the possibility of large medical expenses, then health insurance would be less important and the incentive to be employed in order to retain health insurance would be weaker. If individuals could acquire health insurance coverage by changing jobs or purchasing private non-group insurance, then the employment decisions of individuals with and without retiree coverage might not be very different. And if agents could choose to forego medical treatment, then the absence of retiree coverage might not be a deterrent to retirement before 65. Thus by limiting other mechanisms for smoothing consumption, the model forces individuals who have a strong demand for insurance against medical expenditure risk and who lack retiree health benefits to remain employed until they become eligible for Medicare. If the estimates of this model imply little impact of health insurance on employment, then we would expect that relaxing these assumptions would also yield small impacts. 2 We specify a discrete time model of the employment choices of married couples. At the 2 Starr-McCluer (1996) presents evidence suggesting that precautionary saving by individuals without health insurance is rare. Blau and Gilleskie (2005a) model medical care decisions jointly with employment behavior of older men. French and Jones (2004) and Van der Klaauw and Wolpin (2005) model saving and employment jointly. 6

8 beginning of a period, a couple learns the realizations of all stochastic processes for the period, except the process governing medical expenditure. Given these realizations and the values of deterministic state variables such as work experience and job tenure, the husband and wife make their employment choices for the period and these choices remain fixed for the duration of the period. Total medical expenditure for the period is then realized, and together with the already-committed employment choices and the resulting health insurance coverage, this determines out-of-pocket medical expenditure and consumption net of out-of-pocket medical expenditure. State variables are updated at the end of the period, and the process repeats until the terminal date, T*. The length of a period is two years, corresponding to the average time between interviews in the HRS. The main features of the model are described in the following subsections, and additional details are provided in the unpublished Appendix (Blau and Gilleskie, 2005b) Employment and Health Insurance With probability * an individual who was employed at the end of period t-1 is laid off at the beginning of period t. With probability one, individuals receive a job offer from a new employer at the beginning of each period. Jobs are assumed to be identical except for the pension coverage they offer, as described below. Let j at represent the job choice of spouse a, a = m (male), f (female), in period t. If an individual was employed in period t-1 and not laid off at the beginning of period t, she chooses among non-employment (j at = 1), the new job offer (j at = 2), and the t-1 job (j at = 3). If the individual was laid off at the beginning of t or was not employed during t-1, she chooses between nonemployment and the new job offer. Until the age of Medicare eligibility, we assume that an individual can be covered by up to two 7

9 of the following types of health insurance during a given period: (0) none; (1) own-employer, with retiree coverage; (2) own-employer, without retiree coverage; (3) spouse-employer, with coverage available to the individual after the spouse retires; (4) spouse-employer, without retiree coverage; (5) private (non-employer); and (6) Medicare. Medicare is available to individuals under the age of 65 only if they receive Social Security Disability Insurance (SSDI). Upon reaching age 65, individuals are assumed to enroll in Medicare. We assume that if an individual is covered by his employer s health insurance plan without retiree coverage, he expects to become uninsured if he leaves employment before he turns He expects to remain uninsured until he becomes employed again or reaches age 65 and receives Medicare coverage. Health insurance coverage of individuals with retiree insurance or private plans is unaffected by employment decisions. The key implication of these assumptions is that health insurance coverage is fixed before age 65. Individuals who lack employer-provided health insurance cannot acquire such coverage by changing jobs, and individuals who have employer coverage without RHI cannot acquire RHI by changing jobs. Individuals who do not have private (non-employer) coverage cannot purchase such coverage. We assume that the only agents with access to private coverage are those who have such coverage in the first two waves of the HRS. Thus, the only agents whose health insurance coverage can 3 The Consolidated Omnibus Budget Reconciliation Act of 1986 (COBRA) requires firms with 20 or more employees to allow workers who leave the firm to continue their health insurance coverage for up to 18 months after the date of separation. Thus a worker who retires at age 63.5 can use COBRA coverage as a bridge to Medicare coverage at age 65. The firm can require the former worker to pay the full cost of the coverage. We ignore COBRA coverage because the time period in our model is two years, and because there is little information on COBRA in the HRS. See Gruber and Madrian (1995) for analysis of the effect of COBRA on retirement. 8

10 change are those who are covered by employer-provided insurance without RHI: they lose coverage if they leave employment, and regain coverage if they reenter employment. It is the behavior of this group relative to the other groups that drives the model Health and Medical Expenditure Health status of spouse a, h at, can take on values 0, 1, 2, where 0 denotes good health, 1 denotes bad health, and 2 denotes death. The probability of observing spouse a in health status hn during period t+1 given that he or she was in health status h during period t is B at (h, hn). The transition probabilities are specified as first-order Markov logit processes. Total medical expenditure for each spouse is modeled as a random draw from a known distribution that is realized after the individual has committed to an employment choice for the period. The individual s out-of-pocket expenditure is then computed by applying the rules of his or her health insurance plan(s). The rules include the premium, deductible, co-insurance rate, and maximum out-ofpocket expenditure. Out-of-pocket expenditure for spouse a is denoted m at *, and is given by m at * = F(m at, HI at ), where m at is total medical expenditure and HI at is an indicator of the type (or types) of health insurance plan. The function F characterizes the health insurance rules. We use a discrete approximation of the underlying continuous distribution of medical expenditure for each spouse. Let m ahtk represent the k th masspoint, k=1,...,k, of the distribution of total medical expenditure facing spouse a in health status h at age t, and let p ahtk be the probability of realizing this outcome Income and Consumption Earnings per period, W a, are fixed and non-stochastic. We make this assumption because Social Security benefits depend on average lifetime earnings, and allowing for earnings uncertainty 9

11 therefore results in a drastic increase in the size of the state space and computation time even if earnings shocks are transitory. We do allow for the most important form of earnings risk, namely layoffs. Given that earnings are non-stochastic, it follows that non-earned income b t is also non-stochastic and depends on the age, experience, tenure, earnings, and employment status of the husband and wife according to known rules summarized by the function b t = b(j mt, x 1mt, x 2mt, A mt, j ft, x 1ft, x 2ft, A ft, W m, W f, g t ) where x 1at and x 2at are work experience and job tenure, respectively, A is age, and g t is other nonwage income. This function is shorthand notation for the rules of the Social Security system and private pension plans. The employment choice (j at ) matters because of the Social Security earnings test and because pension benefits cannot be received while employed at the pension-providing firm. Earnings and work experience affect the average earnings measure on which Social Security benefits are based. Job tenure and earnings affect benefits in many pension plans. Age affects eligibility for Social Security and pension benefits and the level of benefits. An individual is assumed to become entitled to Social Security benefits in the first period in which he or she is not employed beginning at age 62. Thus, unlike Rust and Phelan (1997), we do not model the decision to apply for benefits. Asset accumulation is not modeled, so the asset income component of other nonwage income is treated as given. Given the absence of saving, consumption in period t equals total family income net of out-ofpocket medical expenses: C t = W t + b t - m t * - K(E jt, b t, m t *) (1) where W t = W m 1[j mt >1] + W f 1[j ft >1] is total earnings of the couple; 1[] is an indicator function equal to one if the statement inside the brackets is true, and equal to zero otherwise; m t * = m mt * + m ft * is out- 10

12 of-pocket medical expenditure for the couple in period t; and K( ) is a tax function, incorporating income and payroll taxes and accounting for the medical expense deduction Utility Function Utility of the couple in period t depends on each member s choice of employment, and on health, consumption, age, and a preference shock. Let e at =0 if spouse a chooses to be not employed in period t, and e at = 1 if he chooses employment, regardless of whether the job is new or old. Utility in period t if C t >0 is â t (C t, j mt, j ft, h mt, h ft ) +, t (j mt, j ft, h mt, h ft ) = " 0,hmt,hft,emt,eft + C 1-"1 t /(1-" 1 ) + " 2,hmt e mt-1 1[j mt =2] + " 3,hmt e mt-1 1[j mt =1] + " 4,hft e ft-1 1[j ft =2] + " 5,hft e ft-1 1[j ft =1] + " 6 (1-e mt-1 )1[j mt =1](1-e ft-1 )1[j ft =1] + " 7 e mt-1 1[j mt >1]e ft-1 1[j ft >1] + " 8,hmt,emt A mt + " 9,hft,eft A ft +, t (j mt, j ft, h mt, h ft ) The utility function allows employment preferences to depend on lagged employment and health, and to be non-separable across spouses. The utility of consumption is separable from employment and takes the isoelastic form with coefficient of relative risk aversion " 1. There are utility costs (or benefits) of changing jobs (" 2 and " 4 ) and exiting employment (" 3 and " 5 ). Some combinations of employment choices and realizations from the medical expenditure distributions would result in negative income net of out-of-pocket medical expenditure. Negative consumption is not possible, but carefully modeling the mechanisms that allow households to avoid such outcomes is too complex to attempt here. We simply assume that if C t #0, then â t (C t, j mt, j ft, h mt, h ft ) = " 10. The additive utility shocks (,) are assumed to be independently and identically Type I Extreme Value distributed within periods and over time. This rules out both serial correlation and contemporaneous correlation in the utility shocks, as well as 11

13 correlation with the medical expenditure and health shocks. This function characterizes the utility of a couple in which both members are alive. We do not model behavior following the end of a marriage, so we do not specify a utility function for individual members of the couple. 4 Given that the choices we model are discrete, the implications of the family utility function model specified here cannot be easily distinguished from those of a collective model in which each spouse has his or her own utility function and the couple reaches a Pareto-efficient bargain (Blundell et al., 2002). Thus, for simplicity, we use the family labor supply approach to modeling the joint employment decisions of husbands and wives (Blau, 1998; Gustman and Steinmeier, 2000; van Soest, 1995) Value Function and Solution The expected present discounted value (EPDV) of a couple s remaining lifetime utility in period t<t* resulting from given joint employment choices, conditional on the vector of state variables (s t ), and the vector of period t utility shocks (, t ), but not conditional on medical expenditure, is where V(s t,, t ; j mt, j ft, h mt, h ft ) = E t â t (C t, j mt, j ft, h mt, h ft ) +, t (j mt, j ft, h mt, h ft ) + $V t+1 (s t+1 ), 1 1 V t+1 (s t+1 ) = Pr(D t+1 =1)V(D t+1 =1) + (1-Pr(D t+1 =1))[33B mt (h mt, hn mt )B ft (h ft, hn ft )V(s t ; h mt, h ft ) hn m hn f 4 Divorce is modeled as an exogenous stochastic process. In order to avoid the complications of modeling behavior following divorce, we assign a terminal value to the event of the marriage ending as a result of divorce, and do not model behavior following divorce. The terminal value is denoted V(D t =1), and is normalized to We follow the same approach for widowhood: a terminal value is assigned and the behavior of the surviving spouse is not modeled. The terminal value function in the event of the death of spouse a is given by V(h at =2) and is also normalized to

14 1 1 + (3B mt (h mt, i)b ft (h ft, 2) + 3B mt (h mt, 2)B ft (h ft, i) + B mt (h mt, 2)B ft (h ft, 2)V(h t+1 =2)] i=0 i=0 V(h t+1 =2) is the value associated with death of one or both spouses, and $ is the discount factor. The value function accounts for the divorce probability Pr(D t =1), the probability of death of one or both spouses (B at (h at, 2), and the probability of arriving in period t+1 with the marriage intact and both spouses alive and in any combination of good and bad health. The â t term is unknown at the time the employment choice is made because medical expenditure is realized after the employment choice. E t â t (C t, j mt, j ft, h mt, h ft ) is the expected value of utility with respect to the distribution of medical expenditure: K K E t â t (C t, j mt, j ft, h mt, h ft ) = 33p mhtk p fhtr â t (C jtkr (m mhtk, m fhtl ), j mt, j ft, h mt, h ft ) k=1 R=1 where C jtkr (m mhtk, m fhtl ) denotes consumption as a function of medical expenditure, and is given by the budget constraint (eq. 1). This expression integrates (by summation) over the distribution of medical expenditure, and illustrates the fact that period-t medical expenditure has no future consequences. The maximal EPDV of remaining lifetime utility from being in health states h mt and h ft unconditional on period t choices is V(s t ; h mt, h ft ) = E t-1 maxv(s t,, t ; j mt, j ft, h mt, h ft ), where the expectation is taken with respect to, t, and the max is taken with respect to the set of feasible employment alternatives. The model is solved numerically by backward recursion with two approximations used in the solution. The model must be solved for every couple and every trial value of all parameters. The state space is very large, and the exact solution takes too long to compute. Therefore we truncate the decision period at T < T* and approximate the value function at T+1 as a nonlinear function of the state 13

15 space at T+1, with the parameters of the approximation estimated jointly with the other parameters of the model. The second approximation is to compute the value function for a randomly selected subset of points in the state space at each period. Following Keane and Wolpin (1994), the expected value of the value function is regressed on the state variables using the sample of randomly selected points. The estimated regression parameters are used to approximate the expected value of next period s value function. As part of the recursive solution, choice probabilities, which have the multinomial logit form, are computed for the observed choices for each couple Identification The two key sets of identifying assumptions are that (1) saving, health insurance, and medical expenditure are not subject to choice, and (2) there is no unobserved individual heterogeneity. Under the first set of assumptions, consumption equals current income net of medical expenses, and this is determined by employment choices and medical expense draws given an individual s health insurance coverage. With a normalization of one of the utility function intercepts (and of the terminal value functions in the event of divorce or death as described above), the remaining parameters are identified. The observed employment choices directly identify all of the utility function parameters except " 1, the risk-aversion coefficient, and " 10, the parameter determining the disutility of negative net income. " 1 is identified by the employment choices of couples whose health insurance depends on employment, relative to the choices of couples whose health insurance is independent of employment. 5 Given the medical expenditure risk they face, the employment choices of the former group relative to the latter 5 We fix the discount factor ($) and do not estimate it. This is not strictly necessary for identification, but it does help to empirically identify the risk aversion coefficient. 14

16 group identify willingness to bear medical expenditure risk. " 10 is identified by the employment choices couples make in order to avoid negative income. The terminal value function parameters are identified by the nonlinear form in which they enter the choice probabilities. The assumptions of no saving, no control over health insurance, and no control over medical expenditure are clearly quite strong. We argued above that our estimates are useful nevertheless because we can predict the direction of the bias caused by these assumptions. However, this line of reasoning does not account for unobserved heterogeneity in preferences. We do not allow for unobserved heterogeneity because it makes estimation too time-consuming, given that the model must be solved for every couple in the sample. There are a number of possible biases that could be induced by such heterogeneity, but we believe the most plausible story would result in an upward bias in the estimated impact of health insurance on retirement. If preferences for leisure vary in the population, then individuals who value leisure highly will expect to want to retire before age 65, giving them an incentive to seek employment at a firm that offers retiree health insurance (RHI) coverage. Individuals who place less value on leisure will be more likely to expect to remain employed until age 65, and therefore do not have an incentive to seek RHI. In this scenario, differences in RHI coverage do not cause differences in employment behavior - both differences result from unobserved heterogeneity in preferences. The observed association will yield an upward biased estimate of risk aversion and a spuriously large impact of health insurance on retirement Estimation Let t=1 denote the period in which we first observe a couple (the 1992 survey), and let t=0 denote the period prior to the first observation. We observe employment choices as well as health for 15