Table of Contents
Abstract
COBRA and Household Health Insurance Decisions We use the 1993 Survey of Income and Program Participation (SIPP) panel and the Health and Retirement Study (HRS) to model household health insurance coverage decisions over time. We estimate coverage decisions as functions of eligibility for various types of health insurance, including COBRA and coverage through current employers of household members, and other household characteristics. We then simulate changes in household health insurance coverage that result from COBRA qualifying events. In general, we find smaller estimated COBRA effects than obtained previously using cross-sectional data such as the Current Population Survey. One problem is that the use of the SIPP and HRS introduces potentially significant error in the measurement of COBRA qualifying events over time and the measurement of COBRA eligible health insurance plans. This measurement error could in part explain the small estimated COBRA effects found in this study. When health insurance is provided through the place of employment, interruptions in the employment relationship disrupt health insurance coverage. The Consolidated Omnibus Budget Reconciliation Act (COBRA), enacted in 1986, contains provisions designed to partly remedy this problem. Most employees are able to purchase health insurance from their former employer for up to 18 months after their employment ends, at a premium not to exceed 102% of the group rate. If workers are reluctant to change jobs because of health insurance considerations, COBRA can improve the efficiency of the labor market. Perhaps more important, COBRA potentially increases the health insurance coverage of the unemployed. While not universally elected, previous evidence suggests that COBRA increases the health insurance coverage of the unemployed. Previous work also suggests that health insurance coverage is a household decision. Since most health insurance plans allow a covered individual to include other family members as well, it is the household and not the individual that should be the unit of analysis. Coverage through a spouse's health insurance plan is an important option, especially for women. For that reason it is important to undertake a household analysis of health insurance coverage decisions. In this project, we examine health insurance decisions of families after the onset of unemployment, retirement, or a reduction in hours of work by either the husband or wife. This framework allows for a more unified treatment of health insurance coverage outcomes within households than has been possible in past work. We use the Health and Retirement Study (HRS), a panel survey of over 11,000 persons aged 51 to 61 in 1992, and the 1993 Survey of Income and Program Participation (SIPP), a panel survey of approximately 30,000 households. In each of these data sets, we observe the health insurance coverage that husbands and wives have at the beginning of the panel, observe whether or not a member of the household has experienced unemployment, a reduction in hours or retirement, and then observe the health insurance coverage of husbands and wives at a later date. Some individuals are eligible to take up COBRA after the onset of one of these events while others are not. We exploit this difference to identify the effect of COBRA on coverage rates. We also examine the impact of availability of coverage from other sources. The strengths of the HRS and SIPP are that they allow us to identify the effects of COBRA over time in a more sophisticated fashion than is possible using cross-section data such as the Current Population Survey (CPS). The major weakness of both data sets is that they introduce error in the measurment of COBRA eligibility and COBRA qualifying events, thus potentially biasing estimated COBRA effects toward zero. Overall, our results suggest that the estimated effect of COBRA on own or spouse's coverage is small. In fact, it appears that the use of panel data sets such as the HRS or SIPP produces a smaller estimated effect than that obtained with the April CPS cross section data. These smaller estimates could be the result of measurement error or could reflect the fact that the cross-section results overstate the true effect because cross-section models inadequately control for individual heterogeneity. Our estimated COBRA effect does appear somewhat larger for older workers than for younger workers. This result could also be due to measurement error: part of the effect attributed to COBRA may be due to early retirement buyouts and the health insurance that comes with them. We find a fairly large effect on coverage in the SIPP data that appears to come from eligibility for a spouse's employer plan. In addition, the estimates from both data sets indicate that the health insurance decisions of husbands and wives are strongly correlated, even after controlling for demographic factors, earnings, and health insurance status in earlier periods. These results reinforce the idea that it is important to consider health insurance options for the entire family when considering the effects of insurance continuation mandates such as COBRA. The results of this study may have implications for public policy. The effect of COBRA on health insurance coverage is relatively small, bounded by the estimates here and earlier cross-sectional estimates. This could suggest a greater role for the Health Insurance Portability and Accountability Act (HIPAA). However, because COBRA benefits must be exhausted first, many do not immediately qualify for HIPAA. The relatively strong effects of availability of spouse coverage found here and in other studies suggests that the key is for a family to have access to at least one group health insurance policy. There are massive differences in coverage rates among those with and without access to at least one group policy. Thus, public policy should be aimed at improving such access. Perhaps subsidies or tax credits for employers providing coverage to low-wage workers would help improve access for the working uninsured. If access while working could be improved, then if individuals lost or quit their jobs, they would then be eligible to continue coverage under HIPAA or COBRA. When health insurance is provided through the place of employment, interruptions in the employment relationship disrupt health insurance coverage. The Consolidated Omnibus Budget Reconciliation Act (COBRA), enacted in 1986, contains provisions designed to partly remedy this problem. Most employees are able to purchase health insurance from their former employer for up to 18 months after their employment ends, at a premium not to exceed 102% of the group rate. If workers are reluctant to change jobs because of health insurance considerations, COBRA can improve the efficiency of the labor market. Perhaps more important, COBRA potentially increases the health insurance coverage of the unemployed. More recently, the Health Insurance Portability and Accountability Act of 1996 (HIPAA) was passed to decrease the number of uninsured, including the unemployed. HIPAA waives preexisting conditions exclusions and guarantees individual health insurance for individuals who have been covered by a group plan for 18 or more months, who are not eligible for other coverage, and who have exhausted their COBRA continuation coverage. The General Accounting Office (1995) estimates that the HIPAA legislation could ensure continued health insurance coverage for up to 25 million individuals per year. These include workers who change jobs and their dependents, workers facing job lock, and those who are not eligible for COBRA. Because HIPAA only recently took effect, there is no direct evidence on its effectiveness. Given the relatively high cost of maintaining coverage under COBRA when workers change jobs, HIPAA may be a more effective way of eliminating job lock than COBRA.(1) For the unemployed, a worker must first exhaust COBRA benefits before turning to HIPAA to obtain coverage. Therefore it is important to first understand the effectiveness of COBRA legislation in increasing the health insurance coverage of the unemployed. While not universally elected, previous evidence suggests that COBRA increases the health insurance coverage of the unemployed. Using CobraServ data for 1990-91, Flynn (1992) reports that 21% of workers who qualified elected COBRA continuation coverage. Others attempt to provide a more precise estimate of the effect of COBRA on the health coverage of the unemployed by holding a number of demographic characteristics and other factors constant. Using data from the Survey of Income and Program Participation (SIPP) and holding constant age, education, and months since job loss, Klerman and Rahman (1992) find evidence of a positive effect of COBRA legislation on the health insurance coverage of the non-employed. In an important recent study, Gruber and Madrian (1995a) examine health insurance coverage among the non-employed, using longitudinal data from the SIPP for 1983 to 1989 for men aged 25-54. They find that the likelihood of having health insurance drops by approximately 20 percent after a worker is separated from his job. However, they find that state and federal health insurance continuation mandates such as COBRA increase the likelihood of coverage among the non-employed by 6.7 percent. They also find that the estimated effect of continuation mandates varies by the duration of the spell of unemployment. The effect of continuation mandates is insignificant for those with completed durations of one year or less. However, the effects are substantially larger for those with durations of more than one year, presumably the group with the greatest need. For instance, for those with unemployment durations of more than one year, a continuation mandate of one year increases the likelihood of insurance coverage by 9.4 percent (Gruber and Madrian, 1995a, p.23). In earlier work funded by the Pension and Welfare Benefits Administration, we used the 1993 April Current Population Survey and found that unemployed individuals who were eligible for both COBRA and spouse-provided health insurance were more likely to choose spouse provision, perhaps because it was cheaper or more convenient (Berger, Black, and Scott, 1996). In particular, among those eligible for coverage under both COBRA and their spouse's plan, nearly twice as many (52 percent vs. 27 percent) chose to be covered by their spouse's plan. After controlling for worker characteristics and eligibility for insurance from a spouse's employer, our binary logit models imply that COBRA eligibility increases the probability of health insurance coverage by .075. Our multinomial logit models provide a more detailed estimate of the COBRA eligibility effect: COBRA eligibility increases the probability of employer coverage by .154, reduces the probability of other coverage by .081, and reduces the probability of no coverage by .073. COBRA type legislation also appears to have the intended effect on labor market efficiency. Gruber and Madrian (1995a) find that health insurance continuation mandates increase turnover, and are associated with significant wage gains in subsequent jobs. Thus, these mandates appear to reduce job lock and to lead to more productive job search by individuals seeking new jobs. Finally, COBRA type mandates influence workers' decisions when to retire. Using SIPP and the March Current Population Survey (CPS) data, Gruber and Madrian (1995b) and Karoly and Rogowski (1994) find that health insurance continuation laws increase retirement probabilities among older workers. This previous work appears to show that, on average, COBRA legislation has the intended effects on health insurance coverage and labor market transitions. It is also clear from this work that health insurance coverage is a household decision. Since most health insurance plans allow a covered individual to include other family members as well, it is the household and not the individual that should be the unit of analysis. Coverage through a spouse's health insurance plan is an important option, especially for women. For that reason we need a household analysis of health insurance coverage decisions. More generally, employer-provided health insurance is the major form of coverage in the United States. Employer-provided coverage is not only the most important source of coverage for workers themselves, but also for other members of workers' families. In earlier work, we examined the health insurance decision using the March Current Population Surveys (Berger, Black, and Scott, 1994, 1995). In 1992, 69.8 million persons were covered by their own employer, but 63.1 million were covered by another family member's employer. Of these, 22.4 million were workers, 30.2 million were nonworking children, and 10.5 million were nonworking teenagers and adults. Clearly, coverage by employer-provided health insurance is a decision that involves all members of the household. Health insurance coverage from another family member's policy is important among the unemployed as well. Coverage through a spouse's plan can be particularly important for older individuals. While Gruber and Madrian (1995b) emphasize the importance of COBRA eligibility in explaining retirement decisions, spouse coverage can also be important for such households. Among unemployed individuals in the 1993 April CPS aged 50 and over and eligible for both COBRA and spouse coverage, 54 percent chose spouse coverage while only 29 percent chose COBRA. Seventy-four percent of individuals eligible only for spouse coverage chose to be covered (Berger, Black and Scott, 1996). In this project, we use the 1993 Survey of Income and Program Participation (SIPP) panel and the Health and Retirement Study (HRS) to analyze health insurance decisions made within families. Unlike the Current Population Survey, these data sets allow us to observe health insurance choices of individuals and households over time. We use these data sets to estimate models of household health insurance coverage that incorporate eligibility for various types of health insurance coverage and other characteristics of individuals within the household. The advantages of these data sets must be balanced against the disadvantage that they cannot be used to directly determine COBRA coverage. Klerman (1996) points out that individuals are asked about coverage from a previous employer, which may or may not be COBRA coverage. Especially in the case of the HRS, some of this coverage may be part of an early retirement package. Thus, there is error in the measurement of COBRA coverage, which must be kept in mind when interpreting our results. Klerman (1996) believes that analysis using the SIPP is still worthwhile, noting that that there is a strong connection between COBRA coverage and coverage from a previous employer. The remainder of this study is organized as follows. In the next section, we examine the determinants of COBRA eligibility based on the original legislation. We then discuss health insurance decisions in a household framework. In Section IV, we discuss the data used in this project, including the measurement of health insurance coverage, COBRA eligibility, and COBRA qualifying events. In Section V, we introduce the statistical framework, and in Section VI, we present the results of our analysis. We first present a set of preliminary results and then consider extensions of the basic model along with simulations under alternative assumptions. We provide an overall discussion of our results in Section VII along with conclusions. In 1987, the Consolidated Omnibus Budget Reconciliation Act (COBRA) went into effect, which contained provisions allowing certain former employees, spouses, and dependent children to buy temporary health insurance at group rates. "Employees" eligible for COBRA can be full-time or part-time workers, agents, independent contractors, directors, and certain self-employed individuals eligible to participate in a group plan. A qualified employee is anyone who was covered by a group health plan the day before a "qualifying event." Such events include voluntary or involuntary termination of employment for reasons other than gross misconduct, or a reduction in the number of hours worked. Spouses and dependent children qualify for coverage by demonstrating that either of these events were applicable to the covered employee (who was either their spouse or parent), because of death or divorce of the covered employee, or in the case of dependent children, if they lose dependent child status under the plan's rules. If a covered employee becomes eligible for Medicare benefits, their spouse and/or dependents qualify for COBRA coverage. After a qualifying event, beneficiaries have up to 60 days to elect COBRA coverage (U.S. Department of Labor, 1990, pp. 4-5, 9). Employees can receive coverage for up to 18 months at rates of up to 102 percent of the cost of the plan to similarly situated individuals who have not incurred a qualifying event. This coverage can be extended for up to 11 more months if a qualified beneficiary is determined under Title II or XVI of the Social Security Act to have been disabled at the time of termination or reduction in hours. The cost for the additional 11 months of coverage can be increased to 150 percent of the plan's cost. Spouses and dependent children are also eligible for 18 months of coverage if the covered employee terminates employment or suffers a reduction in hours. Spouses and dependent children can obtain up to 36 months of coverage if they become eligible through the death or divorce of the covered employee, or if the child loses his or her dependent status under the plan (U.S. Department of Labor, 1990, pp. 6-7,15). Certain employers are exempt from providing COBRA benefits. The law generally covers group health plans of employers with 20 or more employees during the previous year. The law covers plans provided in the private sector and by state and local governments. The law does not apply to Federally sponsored health plans or the plans of certain church-related organizations (U.S. Department of Labor, 1990, p. 2) III. Household Health Insurance Decisions In earlier work, we find that health insurance coverage of the spouse is important in household health insurance coverage decisions. It could be that the spouse coverage is cheaper or higher quality, or that it is just easier to obtain spouse coverage than to go through the trouble of signing up for COBRA. In any event, spouse coverage and intra-household decisions are important in explaining the health insurance coverage of the unemployed. While previous studies have not directly examined health insurance decisions within a household framework, there are a number of studies related to various aspects of individual health insurance choice. For example, the job lock literature has examined one of the potential side effects of employer-provided health insurance coverage (Gruber and Madrian, 1994; Holtz-Eakin, 1994; Madrian, 1994; Monheit and Cooper, 1994). Berger, Black, and Scott (1997) use Survey of Income and Program Participation (SIPP) data to estimate duration of employment spells as a function of employer-provided health insurance and individual health conditions. Scott, Berger, and Garen (1995) investigate employer-provided health insurance and firms' willingness to hire older workers. Another literature has focused on lengths of spells without health insurance. These papers include Swartz and McBride (1990), Nelson and Short (1990), Klerman (1992), and Swartz, Marcotte, and McBride (1993). While these studies have dealt with a variety of health insurance issues, none except for Berger, Black, and Scott (1996) have explicitly modeled spouse coverage, and their study used cross-sectional data from the 1993 April Current Population Survey. This study builds on previous work by modeling health insurance choice in a household framework over time. Our analysis is based on two different panel data sets: the 1993 Survey of Income and Program Participation (SIPP), which includes households in all age categories, and the Health and Retirement Study (HRS), which is restricted to older individuals, and therefore allows us to determine whether their household decisions differ from those in other age groups. This work has important implications for insurance continuation rules such as COBRA. Unlike previous research that has focused on individual choices, we emphasize household decisions regarding health-insurance coverage. For example, due to a change in employment circumstances members of a household may find themselves eligible for coverage under COBRA. However, we know little about the factors affecting households' decisions to elect COBRA coverage. The household approach helps us better understand gender differences in the decision to elect COBRA coverage in response to qualifying events. In addition, our research examines age differences in responses to qualifying events. In a young family in which the wife is laid off, the husband's employer-provided health insurance may be chosen for her coverage instead of electing COBRA. The same may not be true for older persons, where a couple may be more likely to elect COBRA to bridge the coverage gap between unemployment and the provision of Medicare benefits at the age of retirement. By following health insurance coverage patterns in households over time using the SIPP and HRS, we begin to shed light on such behavioral differences. (2)
A. Survey of Income and Program Participation We use the 1993 panel of the Survey of Income and Program Participation (SIPP) to estimate models of household health insurance coverage decisions. The SIPP is well suited for this project because it follows a sample of approximately 27,000 households over time and contains information on health insurance coverage, in addition to a rich set of demographic and financial variables. We observe individuals in the 1993 SIPP panel in the January, April, July, October 1993 and January 1994 waves. We observe their health insurance coverage at both time periods and can determine whether a COBRA qualifying event has taken place during the intervening year. We estimate models of health insurance coverage in later waves of the 1993 panel as functions of the type of coverage in earlier waves, whether there was a COBRA qualifying event, and other observable characteristics. Thus, we are able to estimate the effects of COBRA qualifying events on the coverage of different members of households over time. With the SIPP, we are able to study the effects of the onset of COBRA qualifying events and other changes on the health insurance coverage decisions of the household. This analysis is not possible with cross-section data sets such as the Current Population Survey. In the SIPP, we observe the health-insurance coverage decisions of the family at the start of the panel. Over the course of the survey, some households are affected by various COBRA qualifying events such as job layoffs or voluntary job quits. At a subsequent date, we observe the health insurance decisions that the household made in response to COBRA qualifying events and other factors. Through the use of discrete choice analysis, we estimate the extent to which these decisions are functions of a) COBRA eligibility, b) availability of other sources of coverage, and c) the demographic structure of the household. B. Health and Retirement Study We estimate similar models using the HRS. The HRS is a longitudinal study of persons born between the years of 1931 and 1941, and focuses on their economic status, health, and retirement. There were 12,652 respondents drawn from a total of 7,702 households in Wave 1, which was conducted in 1992. There were approximately 11,600 respondents in Wave 2, which was conducted in 1994. Again, consider the case of a husband-wife household, this time in the HRS. The household's health insurance coverage is estimated using a discrete choice framework as a function of eligibility for various health insurance plans, including employer coverage, retiree coverage from a former employer, COBRA, and other sources. Similar to SIPP, using the HRS we are able to observe the change in the health insurance coverage of the household in response to changing eligibility and other circumstances. The models also include demographic characteristics of the household, which are also available in the HRS. Our HRS estimates help to determine whether older households' health insurance choices are sensitive to a different set of factors than households in other age groups. C. Measurement of Health Insurance Coverage The dependent variable in our analysis is health insurance coverage in a particular wave of each of the surveys. Coverage statuses of husbands and wives in later waves are used as dependent variables in the analysis. Coverage statuses during earlier waves of each survey are divided into three categories and used as independent variables in the analysis. These categories are coverage provided by one's own employer, coverage provided through another source, including a spouse's employer, and no coverage. D. Measurement of COBRA Eligibility We measure COBRA eligibility using questions on employment, hours worked, health insurance coverage, industry, and firm size. We first measure whether an individual has had a COBRA qualifying event between earlier and later waves of the SIPP or HRS. We measure this using questions on employment status and hours worked. If an individual was working more than 25 hours in the earlier wave and less than 25 hours in the later wave, was working in the earlier wave and retired or unemployed in the later wave, or lost coverage and was working for a new employer, then we say that he or she had a COBRA event. According to the COBRA legislation, qualifying events for covered employees occur when there is a voluntary or involuntary termination or a reduction in hours worked. Our empirical definition of a qualifying event picks up terminations and hours reductions that occur between earlier and later waves of the surveys. We miss qualifying events that take place and result in new coverage prior to the second wave. Thus, we miss some short spells of unemployment. The amount by which this tends to understate the effect of COBRA eligibility depends on how much take-up rates differ with completed spell duration.(3) In order to qualify for COBRA, respondents must have had employer-provided coverage and been working for a COBRA eligible employer in the first wave. The COBRA legislation covers group health plans of employers with 20 or more employees in the private sector or state and local governments. Federal government and certain church-related plans are exempt from the legislation. In the SIPP, we measure COBRA eligible employers as those who have more than 25 workers, and are not federal government employers or church related employers.(4) In the HRS, there is no class of worker question to separate out government from private workers and there is not an industry category variable available at this time that separates out federal, state, and local government employees. Therefore, we classify all employers with more than 25 employees as COBRA eligible employers. Clearly the SIPP provides the superior breakdown of COBRA eligible and non-COBRA eligible employers at this time. V. Statistical Models of Household Health Insurance Decisions and COBRA Eligibility In this section, we construct statistical models to estimate the effects of COBRA eligibility on the health insurance coverage of households. Besides the health insurance coverage variables, we use a number of control variables. Variables that are likely to be exogenous to this choice and are available in the SIPP and the HRS include education, race, ethnic origin, gender, and age. The frequency and severity of health problems increases with age so we would expect health insurance coverage to increase as age increases. There may be higher frequency of medical care use among females, leading to a higher demand and a greater likelihood of coverage.(5) Thus, we estimate separate models for husbands and wives in our analysis. Those with more education are likely to be more knowledgeable about health and thus need fewer health services. On the other hand, expected earnings losses are higher for this group. In general, health outcomes tend to be worse among blacks than whites, suggesting a greater demand for services. However, because earnings are lower, earnings losses from poor health are lower among blacks. Also, information about health insurance coverage may differ between blacks and whites. In other words, it may be possible that the smaller earnings losses when an individual gets sick are canceled out by the greater demand and use of health services. Thus, the direction of the net effect of race on coverage, like many control variables, is uncertain a priori. We use binary probit models to estimate the effect of COBRA eligibility and other variables on the health insurance coverage of household members. These probits are estimated separately for husbands and wives, using data from the first through fourth waves of the 1993 SIPP panel, conducted in 1993 and 1994, and the first and second waves of the HRS, conducted in 1992 and 1994. More formally, the SIPP and HRS data allow us to estimate the probability of having health insurance, or E(Pt|X), where Pt is the probability of having insurance after being unemployed for a time t, and X is a vector of characteristics. We can estimate this conditional probability using probit analysis. This allows us to calculate the probability that an individual will have health insurance given a set of observed characteristics X. The probit model also allows us to determine the marginal effects of changes in the X variables on health insurance coverage. Included in X are our measure of COBRA eligibility, health insurance coverage status in an earlier wave of the HRS and SIPP, and demographic characteristics of the individual: age, education, race, and ethnic origin. We also include number of children in the SIPP models because the SIPP includes a number of younger parents in the sample. We estimate separate equations for husbands and wives. In our samples, we only consider households with both husbands and wives present in order to focus on household health insurance decisions. A. Samples used in the Analyses In the HRS, we include in our analysis any household in which both husband and wife are 62 years old or less at the first wave of the survey in 1992 and in which at least one spouse had a COBRA qualifying event between 1992 and 1994. Anyone who was 63 or more at the first wave of the survey would be eligible for Medicare when the second wave was conducted in 1994. In this sample of 1,150 husband-wife pairs, 504 husbands had a COBRA qualifying event between 1992 and 1994, and 768 wives had a COBRA qualifying event between 1992 and 1994. The SIPP sample is constructed in a similar fashion. All husband-wife pairs in which both spouses are aged 18-63 in 1993 and in which at least one spouse had a COBRA qualifying event are included in the full sample. Of this sample of 1,613 husband-wife pairs, 820 husbands had a COBRA qualifying event between 1993 and 1994, and 891 wives had a COBRA qualifying event between 1993 and 1994.(6) The sizes of the samples used in the analysis and the health insurance coverage rates of husbands and wives are shown in Table 1. The onset of a COBRA qualifying event for the husband lowers his coverage rate on average from 90.08% to 83.13%. On the other hand, the health insurance coverage rate of the wife actually increases between 1992 to 1994 when the husband has a COBRA qualifying event, from 88.10% to 91.02%. Similarly, when the wife has a COBRA qualifying event, her coverage rate falls from 93.62% to 90.49%. The husband's coverage rate stays nearly identical from 1992 to 1994 when the wife has a COBRA qualifying event (91.93% to 91.85%). Table 1 also shows the coverage rates of husbands and wives in 1993 and 1994 using the SIPP data, before and after COBRA qualifying events. Because the SIPP sample covers a wider age range of couples than the HRS data, we also include coverage rates disaggregated by age. Table 1 shows coverage rates where the husband is either under or over 40 in the case of a husband COBRA qualifying event, and in which the wife is either over or under 40 in the case of a wife COBRA qualifying event. In the SIPP data, as in the HRS data, the husband's coverage rate falls with the onset of his own COBRA qualifying event. His coverage rate remains nearly the same when his wife has a COBRA qualifying event. There are striking differences by age, however. Younger husbands are more likely to let their coverage lapse when they have a COBRA qualifying event than are older husbands. In both age groups, husband's coverage remains relatively steady when the wife has a COBRA qualifying event. In the case of wives, coverage drops by seven to eleven percentage points, whether it is the husband or the wife who has the COBRA qualifying event, and whether it is the younger or older sub-sample. Wives appear to be less age sensitive to disruptions in their health insurance coverage, and less sensitive to the source of the disruption. Table 2 shows 1994 health insurance coverage rates of husbands and wives as functions of the type of coverage they or their spouses had in 1992, prior to the onset of a COBRA qualifying event. In the HRS data (Table 2, Panel 2A), we first show the coverage rates for husbands and wives after the husband has a COBRA qualifying event. The coverage rate for husbands who have some form of coverage themselves in 1992, or whose wife has some form of coverage in 1992, hovers around 88%. The coverage rate for wives is somewhat higher, varying from 89.67% to 92.69%. This is not surprising in that it is the husbands that have undergone the COBRA event and have lost their coverage. Coverage rates are significantly lower in the cases in which the husband or the wife had no coverage in 1992. The husband's coverage rate in 1994 if he experienced a COBRA qualifying event but had no coverage in 1992 was 40%. For wives the coverage rate was 36%. Wives who had no coverage in 1992 increased it to 30% after the husband had a COBRA qualifying event. The coverage rate of husbands in this case was 41.67%. These cross-tabulations show that there are many changes in health insurance coverage status when the husband undergoes a COBRA qualifying event. These tabulations also show that the coverage rate of the wife is closely related to the coverage rate of the husband, giving additional motivation for our strategy of modeling the household health insurance coverage decision after the onset of a COBRA qualifying event. Coverage rates are generally higher when the wife experiences a COBRA qualifying event. In general, these events cause less disruption in the household's health insurance coverage than when the husband experiences a COBRA qualifying event. Thus, for example, when the wife experiences a reduction in her hours or changes employers, the household is less likely to experience a gap in their health insurance coverage. This could in part reflect the fact that families are more likely to rely on the health insurance plan of the husband, so changes in the employment status of the wife are less likely to disrupt health insurance coverage of family members. Table 2 also shows the coverage rates of husbands and wives after the onset of a COBRA qualifying event for the SIPP data (Panels B-D). In addition to providing cross-tabulations for the entire age range of the data (18-63), we provide separate tabulations for older (age>=40) and younger (age<40) sub-samples. In the full sample, when the husband has a COBRA qualifying event, coverage in 1994 varies widely depending on the coverage status of the husband and the wife in 1993. If the husband had employer-provided coverage in 1993, the coverage rate in 1994 for the husband is 70.44% and for the wife is 76.35%. These coverage rates are lower than if the wife had employer provided coverage in 1993 or the husband had other coverage in 1993, and are slightly higher than if the wife had other coverage in 1993. Coverage rates in 1994 are substantially lower if either the husband or the wife had no coverage in 1993. As can be seen in Table 2C, the coverage rates are somewhat higher for the older SIPP sub-sample, regardless of coverage status in 1993.(7) The differences are smaller when the wife has a COBRA qualifying event than when the husband has a COBRA qualifying event. Coverage rates are lower in the younger sub-sample as shown in Table 2D, especially when the husband has a COBRA qualifying event. These cross tabulations illustrate the difficulty of identifying the effect of COBRA and eligibility for other types of coverage on overall health insurance coverage when one spouse or the other undergoes a COBRA qualifying event. Consider the case of the HRS. We cannot simply compare the coverage rates in 1994 of persons who had employer-provided coverage in 1992 and persons who had no coverage in 1992. Such a comparison is tainted by at least two factors. First, only some of those who had employer-provided coverage in 1992 are COBRA eligible, namely those who worked for employers covered by the COBRA legislation. Second, those with no coverage in 1992 are a poor control group for those with employer-provided coverage in 1992; some are in worse jobs and others have chosen not to be covered. A cleaner identification strategy is the following: compare individuals who have had COBRA events and had employer-provided coverage in 1992, but recognize that some individuals work for COBRA eligible firms and others do not. Thus, we are identifying the COBRA effect from differences in the type of employer in 1992. The effect of eligibility for spouse-provided employer coverage is identified in a similar fashion. We hold constant type of coverage in 1992 and look at differences in coverage between those with spouses who were covered by their employer in 1992 and those with spouses who were not covered by their employer in 1992. Underlying this comparison is the assumption that if an individual's spouse had own-employer coverage in 1992, then that individual would be eligible for coverage under the spouse's employer plan in 1994. Our probit models specified below incorporate these identification assumptions. Table 3 shows the means of the data from the SIPP and HRS used in the analysis. The means are shown for the full samples of 1,613 families from the SIPP in which one or both family members have had a COBRA qualifying event, and the 1,150 families from the HRS in which one or both family members have had a COBRA qualifying event. There are some differences in the demographic characteristics of the two samples. The SIPP sample is younger because it covers a wider age range, while the HRS focuses on the experiences of older families. Education levels are somewhat higher in the SIPP, while the proportion of the sample that is black is higher in the HRS, with its planned oversampling of blacks. The health insurance coverage rates in 1994 are higher in the HRS than in the SIPP. The HRS sample is older, and thus has a higher demand for health insurance. In addition, many of the HRS COBRA events are planned retirements, in which workers are likely to maintain coverage, while the SIPP COBRA events are more likely to be unplanned layoffs. Families in this situation are likely to have lower income and less likely to maintain their health insurance coverage. We also show the means of the other variables used in the probit models, specifically the coverage statuses of husbands and wives in the first waves of the two surveys, conducted in 1992 for the HRS, and 1993 for the SIPP. The percentages of husbands and wives with no coverage is higher in the SIPP data than in the HRS data, perhaps reflecting the fact that the HRS is an older sample with a higher demand for health insurance. The coverage rates of spouses by their own employer is higher in the SIPP while the coverage rate from other sources is significantly higher in the HRS than in the SIPP. The difference in the rate of coverage from other sources is due to a couple of factors. The SIPP sample is younger, but more importantly, the set of health insurance questions are different in the SIPP than in the HRS. The lower rate on other coverage in the SIPP is in part due to the fact that it is more difficult to identify coverage from a spouse in the SIPP. While the HRS incorporates this source of coverage directly in the battery of health insurance questions, spouse coverage in the SIPP is identified only after the respondent says that he or she does not have coverage in his or her name and identifies the spouse as the source of the coverage. This is likely to miss some cases of spouse-provided coverage, if for no other reason than it misses all individuals who have coverage both in their own name and in their spouse's name. In the SIPP sample, there is approximately an equal breakdown between COBRA events experienced by the husband and experienced by the wife. In the HRS, a larger proportion of the events are experienced by the wife. COBRA eligibility is calculated as the product of a COBRA event and employment with a COBRA eligible employer. In the case of the HRS, a COBRA eligible employer is a small employer. In the case of the SIPP, both firm size and industry are used to construct the COBRA eligibility variable. In the SIPP, husbands are more likely to be COBRA eligible, while the reverse is true in the HRS. We report the binary probit estimates of health insurance coverage in 1994 in Tables 4 through 7. In Table 4 we report probit estimates from models using the HRS data for coverage of the husband and the wife when the husband has a COBRA qualifying event, and separate estimates when the wife has a COBRA qualifying event. Tables 5 through 7 are devoted to estimates using the SIPP data. Table 5 uses the full SIPP sample, Table 6 uses the samples of husbands and wives aged 40 and over, and Table 7 presents estimates for the samples of husbands and wives under age 40. The demographic controls in each model are age and age squared, and a series of dummy variables reflecting education completion, race, and ethnic origin. The education completion dummies are high school dropout, some college, bachelor's degree, and graduate or professional degree. The omitted category is high school graduate. The race variables are black and other race. The omitted category is white. The ethnic origin variable is Hispanic status. Also included besides the demographic variables are a series of variables indicating health insurance coverage status in 1992 and whether the individual or the spouse are eligible for COBRA and whether or not the spouse has had a COBRA qualifying event.(8) Consider the HRS models presented in Table 4. The health insurance coverage status variables are dummy variables indicating no coverage in 1992, whether the spouse had employer- provided coverage in 1992, whether the individual had some form of coverage other than own employer-provided coverage in 1992, whether the individual was eligible for COBRA, whether the spouse had a COBRA qualifying event between 1992 and 1994, and whether the spouse is eligible for COBRA. An individual is COBRA eligible if he or she had a COBRA qualifying event, was covered by employer-provided insurance in 1992, and worked for a COBRA eligible employer. The spouse COBRA eligible variable is defined in a similar fashion as the own COBRA eligible variable. The identification of the effect of COBRA eligibility is as follows. The omitted category includes individuals who had employer-provided coverage in 1992 and had a COBRA qualifying event, but did not work for a COBRA eligible employer. According to the COBRA legislation, these employers are those with 20 or fewer workers, the federal government, or churches. Unfortunately, the HRS does not yet have 3 digit industry breakdowns, so we were not able to identify church employees. More importantly, there is no class of worker question, and the available industry breakdown did not allow us to separate federal government workers from state and local government workers. Therefore, the omitted category in the HRS data consists only of workers at firms with 25 or less workers, which is the closest firm size breakdown to that under the COBRA legislation. Thus, we measure the own COBRA effect using the coefficient on the COBRA eligibility variable. This coefficient gives the difference in coverage, holding demographics constant, between individuals who have had a COBRA event, had employer-provided coverage in 1992, and who did and did not work for a COBRA eligible employer in 1992. Our previous work showed that eligibility for coverage under a spouse's employer plan is an important source of coverage for those experiencing a COBRA qualifying event. This effect can be observed directly from the coefficient on spouse's employer-provided coverage in 1992. An individual who experiences a COBRA qualifying event and has a spouse with coverage from his or her employer, can in most instances obtain coverage from that plan, and we measure that effect in our probit models. Finally, there are some families in our data in which both the husband and the wife have had COBRA qualifying events. Because families in this situation may have different coverage rates and make different health insurance decisions, we include two variables indicating whether the spouse has had a COBRA qualifying event and whether the spouse is eligible to elect COBRA coverage. The coefficient on the spouse COBRA eligible variable shows the effect of a spouse's eligibility for COBRA on one's own health insurance coverage. The first two columns of Table 4 show the coverage equations for husbands and wives when the husband has a COBRA qualifying event. In the husband's equation, none of the control variables have estimated effects that are significantly different from zero except for the two race variables. The other significant effect is that coverage rates are lower for those who started off with no coverage in 1992. The effects of COBRA eligibility are small and insignificant. In the wife's coverage equation after the husband has had a COBRA qualifying event, only the no coverage variable is significant at the five percent level, while black and husband's COBRA eligibility are significantly different from zero at the ten percent level. Surprisingly, a husband's eligibility for COBRA reduces the probability of his wife's coverage. This result is present in the raw data, in which 44 of 45 women whose husbands had COBRA qualifying events but worked for a non-COBRA eligible employer were covered in 1994, while 20 of 201 women whose husbands worked for a COBRA eligible employer were covered in 1994. The patterns in the raw data illustrate the difficulty of pinning down a COBRA effect in the HRS data. The third and fourth columns of Table 4 show the results for health insurance coverage in 1994 following a COBRA qualifying event experienced by the wife. In this case, COBRA eligibility significantly increases the probability of coverage, while spouse employer-provided coverage does so at the ten percent level. Blacks and those with no coverage in 1992 are significantly less likely to have coverage in 1994. Wives whose husbands also had a COBRA qualifying event between 1992 and 1994 are also significantly less likely to have coverage in 1994, as contrasted with the husbands whose wives have had a COBRA qualifying event, where there is no effect on coverage. The only very large and significant effect in the husbands' equation when the wife has a COBRA qualifying event is that those who had no coverage in 1992 are less likely to have coverage in 1994. Tables 5 through 7 report the binary probit estimates for the SIPP sample. Separate estimates are reported for the cases in which the husband has a COBRA qualifying event and in which the wife has a COBRA qualifying event. The full sample estimates are shown in Table 5. In general, there are significant age effects on coverage, unlike in the HRS data. This reflects the wider age range in the SIPP sample than in the HRS sample. When the SIPP sample is disaggregated into the 40 and over and under 40 groupings (Tables 6 and 7), the significant age effects disappear. As in the HRS data, there is some evidence of significant coverage differences by education, race, and ethnic groups. In general, lower educated individuals, blacks, and Hispanics are less likely to be covered. The insurance coverage variables in 1993 provide estimates of the effects on coverage after a COBRA event of COBRA and other health insurance coverage. Those with no coverage in 1993 are less likely to be covered in 1994, after the COBRA event. The estimated COBRA effect comes from the COBRA eligibility variables. In the husband COBRA event samples, the estimated marginal effect of husband COBRA eligibility on the husband's coverage is .02, and on the wife's coverage is .07. If the wife is COBRA eligible, in addition to the husband being COBRA eligible, coverage rates in the household are reduced, presumably due to the fact that both spouses may have experienced a disruption in their health insurance coverage. The effect of spouse's coverage is strongly positive and significant, as can be seen by the effect of wife's employer-provided coverage on the husband's coverage in column 1. In the second two columns of Table 5, the probit models for the wife COBRA event sample are estimated. In this case, the husband COBRA eligibility variables are strongly negative and significant. This again shows the effect of the husband being COBRA eligible (i.e. having a COBRA event and being COBRA eligible) on coverage. Thus when both the husband and the wife have had a COBRA event, household members' probability of being covered is reduced. The estimated COBRA effect is shown by the marginal effect of wife's COBRA eligibility variable. For husbands, the effect is near zero, and for wives, the effect is actually negative, although neither is statistically significant. However, the availability of coverage from the husband after the wife has had a COBRA qualifying event is large and significantly positive (column 4). In other words, for women, the effect of availability of spouse coverage is stronger than for men after the onset of a COBRA qualifying event. Tables 6 and 7 show the results for the SIPP sub-samples of those 40 and over and those under 40. While the age effects are insignificant, in general the effects of race, ethnic group, and education are similar to those found for the full SIPP sample. The SIPP regressions also include number of children in the household as a covariate because of the presence of younger families in the SIPP sample. However, the effect of number of children on health insurance coverage is insignificant in all of the SIPP estimated models. As before, those with no coverage in 1993 are less likely to be covered in 1994, and those with access to spouse's employer-provided health insurance are more likely to be covered in 1994. While smaller than the effect of access to spouse employer-provided health insurance, the COBRA effects are positive in all four cases considered in the older SIPP sub-sample. The estimated COBRA effects vary from almost zero in the case of husbands when the wife has a COBRA event to .09 for husband's coverage when the husband has a COBRA event. However, none of the estimated effects are statistically significant. In the younger SIPP sub-sample, the estimated COBRA effects are negative in three out of four cases but still statistically insignificant. Overall, the SIPP sub-samples suggest that the estimated COBRA effects increase with age, but are still much smaller than the estimated effects of spouse coverage after the onset of a COBRA event. In general, our estimated COBRA effects appear to be somewhat smaller than those we estimated in Berger, Black and Scott (1996) using April 1993 CPS data. While the SIPP and HRS data are useful in that they provide panel data and allow for identification of the COBRA effect over time, the April CPS has better questions regarding eligibility for COBRA, COBRA qualifying events, and for other health insurance coverage. Therefore, a tradeoff exists between identification using panel data and better survey questions. In this study, we show that when the COBRA effect is identified, its estimated magnitude is in general small and in many cases is not statistically significant. E. Extensions of the Binary Probit Models There are several extensions of the basic binary probit models reported in Tables 4 through 7 that can be considered. The models presented in Tables 4 through 7 used very parsimonious specifications of non-health insurance variables, including only measures of age, education, number of children, race, and ethnic origin. However, one could argue that other variables also affect health insurance coverage. For example, differences in income may affect coverage, and while age and education are correlated with income, it may be preferable to include income directly in the models. For example, there are earnings differences between men and women that may affect coverage rates. The earnings of an individual's spouse are also likely to be relevant. Those with lower earnings may be more likely to let their health insurance coverage lapse. In order to control for these differences, in the extensions to the basic model, we include the earnings of the individual and his or her spouse in the health insurance coverage models. Another variable that may be important in determining health insurance coverage and COBRA take up is self-reported health status. Those in poorer health may have a higher demand for health insurance coverage and may be more likely to take up COBRA. We include measures of disability status in our health insurance coverage models to control for differences in self-reported health in the HRS and SIPP. For the SIPP we constructed a measure of disability status by using the following question: "Has indicated having a physical, mental, or other health condition which limits the kind or amount of work ... can do during this panel?" In the HRS, there is no single measure of disability status. Therefore, we constructed our own measure by using the responses to several questions on the Health Status (Section B) module. In particular, we define a respondent as being disabled if he or she answered yes to any one of the following questions: that their health was in fair or poor condition, that their health was much worse than it was one year ago, that they were in poor emotional health, found it somewhat difficult or more difficult to walk a block, walk across a room, sit for two hours, get up from a chair after sitting for long periods, get out of bed without help, climb one flight of stairs without resting, lift a bag of groceries, stoop, kneel or crouch, pick up a dime from a table, bathe or shower without help, extend their arm above shoulder level, eat without help, or dress without help. Another potential problem is the length of time between the waves of the SIPP and HRS used in constructing our data sets. In the case of the HRS, there is a two-year period between the 1st and 2nd waves (1992 and 1994). In the SIPP analysis until now we have used waves that are one year apart (1993 and 1994). It is not possible to reduce the length of time between waves for the HRS. However, the SIPP is conducted every quarter for a given panel, so it is possible to reconfigure the data so that there is a three-month period between waves. This reduces the potential problem of missing unemployment spells and potential COBRA take-up decisions. In the SIPP analysis below, we use both the one-year and three-month data. Finally, the binary probit models shown in Tables 4 through 7 do not allow for an explicit connection between the health insurance decisions of the two spouses. In order to allow for such a relationship, we estimate bivariate probit models in the analysis presented in the next section. The bivariate probit models are similar to the binary probit models except that the disturbance terms in the underlying husband and wife coverage equations are allowed to be correlated. This means that unobservables affecting the probability that one spouse is covered by health insurance are correlated with unobservables affecting whether the other spouse is covered. This assumption allows the health insurance decisions of the husband and wife to be formally interrelated in the estimation process. These models potentially provide a more complete picture of family health insurance outcomes. The bivariate probit estimates are presented in Tables 8-11. We present two sets of estimates: the first which uses the same data and set of independent variables as did the single equation probit estimates presented in Tables 4-7, and the second which incorporates the extensions discussed in the previous section. For the HRS, this means that we include the earnings and disability variables. For the SIPP, we include the earnings and disability variables and use the quarterly data instead of the annual data. We use the full samples in the bivariate probit estimation; i.e. any family in which either the husband or the wife has had a COBRA qualifying event. In the HRS data, this amounts to 1,150 families. In the SIPP, there are 1,613 families in the annual sample and 3,871 families in the quarterly sample. The HRS bivariate probit estimates are reported in Table 8. The first two columns show the results for the basic specification. In general, demographics do not significantly explain variation in coverage rates in the bivariate probit models. The exception is that there is some evidence of racial and ethnic differences in coverage rates. Among husbands, Hispanics have lower coverage rates and among wives, blacks have significantly lower coverage rates. As before, those with no coverage in 1992 are significantly less likely to have coverage in 1994. A COBRA event for the husband is associated with lower coverage rates for both husbands and wives. The results are very similar in the expanded specification shown in the third and fourth columns. The earnings and disability variables have the expected effect on health insurance coverage, although only the variable measuring wife's earnings is statistically significant. The bivariate probit model also provides an estimate of the correlation between the error terms in the husband and wife health insurance coverage equations. In the case of the HRS data, the estimated correlation is .697 in the basic model and .644 in the expanded model, indicating a high degree of correlation between the disturbances in the husband and wife health insurance equations. In the case of the full SIPP samples, the estimated correlation is .876 in the basic model and .880 in the expanded model, indicating an even higher degree of correlation between the disturbances in the SIPP data than in the HRS data. The correlations for the old and young sub-samples are of a similar magnitude. Thus, it appears that the health insurance decisions of family members are highly interrelated after the onset of a COBRA event. Tables 9-11 show the estimated bivariate probit models for the full SIPP samples, and the older and younger sub-samples. The demographic results are broadly consistent with those we have seen in the HRS bivariate probits. One difference is that in the SIPP expanded models, the earnings of both the husband and the wife have statistically significant effects on health insurance coverage. The results for the insurance variables are qualitatively similar in the expanded and basic SIPP models, however the magnitudes of the estimates do change, sometimes substantially. The main differences in results between the SIPP and the HRS are in the effects of the insurance variables. In the SIPP models, those with no coverage in 1993 are less likely to be covered in 1994 and the presence of spouse employer-provided coverage in 1993 raises the probability of coverage in 1994, similar to the HRS results. However, the effects of COBRA eligibility are different than those estimated in the HRS. In the full SIPP samples, COBRA eligibility significantly lowers the probability of coverage in 1994. The same is true in the older and younger sub-samples, although the estimated effects are less negative in the older sub-sample. By contrast, COBRA eligibility in general has a positive effect on coverage in the HRS sample. This change in the sign of the estimate again points out the difficulty of identifying the COBRA effect using the HRS and SIPP data and the sensitivity of the estimate to differences in estimation method and identification assumptions. We use the bivariate probit estimates to generate predicted probabilities of coverage in 1994 for husbands and wives under different circumstances in the first year of the data (1992 for HRS and 1993 for SIPP). The six situations we consider after the onset of an own COBRA qualifying event are as follows: own employer-provided coverage by a COBRA eligible employer, spouse has coverage from employer; own employer-provided coverage by a COBRA ineligible employer, spouse has coverage from employer; own coverage from sources other than own employer, spouse has coverage from employer; own employer-provided coverage by a COBRA eligible employer, spouse has coverage from sources other than his or her employer; own employer-provided coverage by a COBRA ineligible employer, spouse has coverage from sources other than his or her own employer; no coverage for either spouse. These simulations allow us to see the effects of COBRA qualifying events on family health insurance coverage decisions under different circumstances. For example, comparing cases 1 and 2 or 4 and 5 provides an estimate of the effect of COBRA on coverage. Comparing 1 and 4 or 2 and 5 shows the effect of the independent source of coverage by the spouse when the other spouse has a COBRA qualifying event. Simulation 3 shows the effect of a COBRA qualifying event when the spouse already has coverage other than the own employer. Presumably, a COBRA qualifying event in this instance should have very little effect on coverage rates. Simulation 6 provides a kind of baseline in that it shows predicted coverage rates in 1994 when neither spouse had coverage in 1992 or 1993. We show simulations using the HRS data, for the full SIPP data, and for the older and younger SIPP sub-samples. For each data set, we perform simulations using both sets of bivariate probit parameter estimates reported in Tables 8-11. We first present the simulations using the parameter estimates from the basic specifications and then from the expanded specifications. The predicted probabilities are generated using the parameter estimates from Tables 8-11 and the same values for education, race, earnings, and disability status across the four samples. We use the means from the overall SIPP sample that are shown in Table 3.(9) We set age to be the same for the older SIPP sample and the HRS sample, using the mean ages for the HRS sample, and set it to lower values for the full SIPP sample and the younger sub-sample, using the SIPP average ages.(10) The insurance and COBRA event variables take on different values depending on the simulation being considered. Table 12 shows the results of simulations using the HRS data. In these simulations, a COBRA effect on coverage is visible. In comparing simulations 1 and 2, the coverage rates of husbands (column 1) and wives (column 4) after they have had a COBRA qualifying event is lower if they worked for a COBRA ineligible employer. The difference in coverage rates between rows 1 and 2 is thus an estimate of the COBRA effect on coverage. It should be noted, however, that this estimate is based on statistically insignificant parameter estimates in Table 8. None of the simulated events have much effect on the coverage rates of spouses, which remain almost constant and above .96 after the spouse has a COBRA event if the individuals were covered by their own employer in the first place. We find that both husband and wife coverage rates remain high after the onset of a wife's COBRA qualifying event. Also, the predicted coverage rates of those with no coverage in 1992, prior to the onset of the COBRA qualifying event, is very low. Table 13 shows the simulated coverage rates for the full SIPP sample. In the case of a husband COBRA qualifying event, the wife's rate stays above .95 if she has coverage from her own employer using the basic specification (Panel A) and above .86 using the expanded specification (Panel B). When the wife has a COBRA qualifying event, both the husband and wife coverage rates remain above .94, if the husband has coverage from his own employer. It appears that wives are more likely to switch to their husbands' plans than the reverse with the onset of a COBRA qualifying event. In comparing simulations 1 and 4 and 2 and 5, it appears that coverage rates fall if the spouse does not have employer-provided coverage when there is a COBRA qualifying event. This suggests that ultimately coverage depends in large part on whether the family has alternative sources of coverage when there is a COBRA qualifying event. The predicted coverage rates are substantially lower for husbands and wives with no coverage in 1993. This suggests that we would find much larger effects of COBRA and the presence of other insurance plans after the onset of a COBRA qualifying event if we were willing to use those with no coverage as a control group. However, these individuals are likely to be in unobservably worse jobs than individuals with employer-provided coverage in 1993, implying that using them as a control group would overstate the effects of COBRA and spouse coverage in 1993 on coverage after the event in 1994. Tables 14 and 15 provide the same simulations for the SIPP sub-samples in which the husband is age 40 and over, and under 40. While there are quantitative differences, reflecting differences in the parameter estimates in Tables 8-11, for the most part the results are qualitatively similar to those presented in Table 13. VII. Discussion and Conclusions We examine health insurance decisions of families after the onset of unemployment, retirement, or a reduction in hours of work by either the husband or wife. This framework allows for a more unified treatment of health insurance coverage outcomes within households than has been possible in past work. We use the Health and Retirement Study (HRS), a panel survey of over 11,000 persons aged 51 to 61 in 1992, and the 1993 Survey of Income and Program Participation (SIPP), a panel survey of approximately 30,000 households. In each of these data sets, we observe the health insurance coverage that husbands and wives have at the beginning of the panel, observe whether or not a member of the household has experienced unemployment, a reduction in hours or retirement, and then observe the health insurance coverage of husbands and wives at a later date. Some individuals are eligible to take up COBRA after the onset of one of these events while others are not. We exploit this difference to identify the effect of COBRA on coverage rates. We also examine the impact of availability of coverage from other sources. The strengths of the HRS and SIPP are that they allow us to identify the effects of COBRA over time in a more sophisticated fashion than is possible using cross-section data such as the Current Population Survey (CPS). The major weakness of both data sets is that they introduce error in the measurement of COBRA eligibility and COBRA qualifying events, thus potentially biasing estimated COBRA effects toward zero. We first present tabulations of coverage rates at the beginning of the panels, 1992 for the HRS and 1993 for the SIPP, and in 1994, after the onset of a COBRA event. We also look at tabulations of coverage in 1994 as a function of type of coverage at the beginning of the panels. We present binary probit estimates for samples in which the husband has had a COBRA event and separately for samples in which the wife has had a COBRA event. We then estimate bivariate probit models in which the health insurance coverage of the husband and wife are estimated jointly, using a sample of households for which either the husband or the wife has had a COBRA event. We present bivariate probit estimates using the same specification as the binary probit models and for expanded specifications using more control variables, and in the case of the SIPP, using a sample based on quarterly instead of annual data. Finally, we use the bivariate probit estimates to simulate coverage rates under different situations in the household. Our results suggest that the estimated effect of COBRA on own or spouse's coverage is small. In fact, it appears that the use of panel data sets such as the HRS or SIPP produces a smaller estimated effect than that obtained with the April CPS cross section data. These smaller estimates could be the result of measurement error or could reflect the fact that the cross-section results overstate the true effect because cross-section models inadequately control for individual heterogeneity. Our estimated COBRA effect does appear somewhat larger for older workers than for younger workers. This result could also be due to measurement error: part of the effect attributed to COBRA may be due to early retirement buyouts and the health insurance that comes with them. We also find a fairly large effect on coverage in the SIPP data that appears to come from eligibility for a spouse's employer plan. This suggests it is important to consider health insurance options for the entire family when considering the effects of insurance continuation mandates such as COBRA. The results of this study may have implications for public policy. The effect of COBRA on health insurance coverage is relatively small, bounded by the estimates here and earlier cross-sectional estimates. This could suggest a greater role for the Health Insurance Portability and Accountability Act (HIPAA). However, because COBRA benefits must be exhausted first, many do not immediately qualify for HIPAA. The relatively strong effects of availability of spouse coverage found here and in other studies suggests that the key is for a family to have access to at least one group health insurance policy. There are massive differences in coverage rates among those with and without access to at least one group policy. Thus, public policy should be aimed at improving such access. Perhaps subsidies or tax credits for employers providing coverage to low-wage workers would help improve access for the working uninsured. If access while working could be improved, then if individuals lost or quit their jobs, they would then be eligible to continue coverage under HIPAA or COBRA. Berger, Mark C., Black, Dan A., and Scott, Frank A. "Measuring the Uninsured by Firm Size and Employment Status, Part I," prepared for the U.S. Small Business Administration, Contract #SBA-7642-OA-92, June 1994.
Berger, Mark C., Black, Dan A., and Scott, Frank A. "Measuring the Uninsured by Firm Size and Employment Status, Part II," prepared for the U.S. Small Business Administration, Contract #SBA-7642-OA-92, May 1995.
Berger, Mark C., Black, Dan A., and Scott, Frank A. "Health Insurance Coverage of the Unemployed," final report prepared for the Pension and Welfare Benefits Administration, U.S. Department of Labor, Contract #41USC252C3, 1996.
Berger, Mark C., Black, Dan A., and Scott, Frank A. "Employer-Provided Health Insurance and Job Lock" final report prepared for the Agency for Health Care Policy and Research, Grant # R01 HS08188, 1997.
Flynn, Patrice. 1992. "Employment-Based Health Insurance: Coverage Under COBRA Continuation Rules." In Health Benefits and the Workforce, Pension and Welfare Benefits Administration, U.S. Department of Labor, Washington, DC, pp. 105-116.
Gruber, Jonathan and Madrian, Brigitte. "Limited Insurance Portability and Job Mobility: The Effects of Public Policy on Job Lock," Industrial and Labor Relations Review (48) October 1994, pp. 86-102.
Gruber, Jonathan and Brigitte C. Madrian. 1995a. "Non-Employment and Health Insurance Coverage." NBER Working Paper No. 5228, August.
Gruber, Jonathan and Brigitte C. Madrian. 1995b. "Health Insurance Availability and the Retirement Decision." American Economic Review 85:4 (September), pp. 938-948.
Holtz-Eakin, Douglas. "Health Insurance Provision and Labor Market Efficiency in the United States and Germany," in Rebecca Blank's Protection Versus Economic Flexibility: Is There a Tradeoff? Chicago, IL: University of Chicago Press, 1994.
Karoly, Lynn A. and Jeannette A. Rogowski. 1994. "The Effect of Access to Post-Retirement Health Insurance on the Decision to Retire Early." Industrial and Labor Relations Review 48:1 (October), pp. 103-123.
Klerman, Jacob Alex. "How Long is a Spell Without Health Insurance?" Unpublished Manuscript. Rand Corporation, October 1992.
Klerman, Jacob Alex. 1996. New Estimates of the Effect of Kassebaum-Kennedy's Group-to-Individual Conversion Provision on Premiums for Individual Health Insurance. Rand Corporation, Santa Monica, CA.
Klerman, Jacob Alex and Omar Rahman. 1992. "Employment Change and Continuation of Health Insurance Coverage." In Health Benefits and the Workforce, Pension and Welfare Benefits Administration, U.S. Department of Labor, Washington, DC, pp. 93-104.
Madrian, Brigitte. "The Effect of Health Insurance on Retirement," Brookings Papers on Economic Activity, 1994, pp. 183-200.
Monheit, Alan and Philip Cooper. "Health Insurance and Job Mobility: Theory and Evidence." Mimeo, Agency for Health Care Policy and Research, 1994.
Nelson, C. and K. Short. Health Insurance Coverage 1986-88, U.S. Bureau of the Census, Current Population Reports, March 1990, Series P-70, No. 17, pp.14-15.
Scott, Frank A., Berger, Mark C., and Garen, John. "Do Health Insurance Costs and Non-Discrimination Policies Reduce the Job Opportunities of Older Workers?" Industrial and Labor Relations Review 48, July 1995, pp. 775-791.
Sindelar, Jody L. 1982. "Differential Use of Medical Care by Sex." Journal of Political Economy 90:5 (October), pp. 1003-1019.
Swartz, Katherine and Timothy D. McBride. "Spells Without Health Insurance: Distributions of Durations and Their Link to Point in Time Estimates of the Uninsured," Inquiry 27 (Fall 1990): pp. 281-88.
Swartz, Katherine, John Marcotte, and Timothy D. McBride."Spells Without Health Insurance: The Distribution of Durations When Left-Censored Spells are Included." Inquiry 30 (Spring 1993), 77-83.
U.S. Bureau of the Census. 1995. Statistical Abstract of the United States, 115th edition. Washington, DC.
U.S. Department of Labor. 1990. Health Benefits Under the Consolidated Omnibus Budget Reconciliation Act (COBRA). Pension and Welfare Benefits Administration, Washington, DC.
U.S. General Accounting Office. 1995. Health Insurance Portability: Reform Could Ensure Continued Coverage for up to 25 Million Americans. GAO/HEHS-95-257, Washington, DC. Table 1
Table 2
Table 3
Table 4
Table 5
Table 6
Table 7
Table 8
Table 9
Table 10
Table 11
Table 12
Table 13
Table 14
1. While the Health Insurance Association of America estimates that premiums in the individual insurance market under HIPAA will increase as much as 20 percent, using Current Population Survey and Survey of Income and Program Participation data, Klerman (1996) estimates much more moderate increases, ranging from 1 to 5.7 percent. Return to Text 2. Of course, our inability to accurately characterize all coverage as COBRA coverage may adversely affect our ability to isolate the effect of COBRA on household health insurance decisions. Return to Text 3. This is less of a problem in our later analysis using the SIPP in which we define COBRA qualifying events over three month periods. This substantially reduces the likelihood that we miss short spells that end in new health insurance coverage. Return to Text 4. In both the SIPP and HRS data sets, the firm size of 25 or more employees is the closest we can get to the 20 or more employees designated in the COBRA legislation. Return to Text 5. See Sindelar (1982) and U.S. Census Bureau (1995, Tables 178, 179, 189, 190, 191, 192, 194) for evidence on greater frequency of use of health care services. Return to Text 6. In the HRS data, there were 122 families in which both the husband and the wife had a COBRA qualifying event between 1992 and 1994, and in the SIPP data, there were 98 families in which both the husband and the wife had a COBRA qualifying event between 1993 and 1994. Return to Text 7. The seemingly higher coverage rates may be an artifact of early-retirement offers being made to the older SIPP group. Early-retirement offers may also affect coverage rates in the HRS sample. Return to Text 8. A variable indicating whether or not the individual has had a COBRA qualifying event is not necessary because each sample includes only those who have had an event. Return to Text 9. In the SIPP data, three is a disability rate of 9.4 percent for men of all ages. For women, the disability rate is 7.4 percent across all ages. We use these means in our simulations of health insurance coverage. Correspondingly, in the HRS data, the disability rate was 16% for men and 10% for women. In the SIPP full-sample data earnings were estimated to be $2,085 per month for men, and $925 per month for women. In the HRS earnings are reported annually and we divided by 12 to obtain an estimate of monthly earnings. For men in the HRS average monthly earnings were $2,280, for women they were $996. The SIPP earnings means are used in the simulations. Return to Text 10. Thus, the SIPP full sample simulations used the SIPP average ages given in Table 3. The HRS and SIPP older sample simulations used the HRS average ages given in Table 3. The SIPP younger sample simulations used average ages of 31.9 years for husbands and 30.83 for wives. Return to Text
|
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||