How Do Households Choose Their Employer-Based Health Insurance?

Jean Marie Abraham William B. Vogt Martin S. Gaynor

How Do Households Choose Their EmployerBased Health Insurance?

Using the 1996 Medical Expenditure Panel Survey, this study estimates a model of household demand for employer-based health insurance to investigate the set of plan and household characteristics that influence coverage choices. Overall, we find that households are sensitive to price with respect to their coverage decisions, and that price sensitivity varies by marital status, wealth, and the number of offers of employer coverage available to the household. We also find that lower-income households are less likely to select an option that provides coverage for all household members. Using our model estimates, we simulate the effect of employers offering various levels of ‘‘opt-out’’ payments on changes in workers’ probabilities of not taking up coverage and on expected costs. One of the key institutional features of the employer-based health insurance system is that nearly all employers offering coverage as a fringe benefit offer family coverage. In 2003, approximately 64% of nonelderly Americans received their health insurance through an employer, either as active workers or as dependents (Carroll 2005). As a benefit that can be conferred upon all household members, health plan coverage choices are likely to depend not only on the attributes of the worker who is offered the fringe benefit, but also on the attributes of other family members. In this study, we estimate a model of household demand for employer-based health insurance to understand the extent to which cover-

age choices, price sensitivity, and the willingness to take up coverage from a particular employer source vary across different types of households. When the U.S. system of employer-based health insurance took root, most households had only one worker and one source of employerbased coverage. Over time, the distribution of household types has changed, as well as the employment decisions of adults within those households. In 1950, approximately 78% of the 43 million households in the United States included married adults, while the remaining 22% were equally split between single adults living alone and single adults with dependents. Today, this distribution has shifted with married adults

Jean Marie Abraham, Ph.D., is an assistant professor in the Division of Health Policy and Management, School of Public Health, University of Minnesota. William B. Vogt, Ph.D., is an associate professor at the H. John Heinz III School of Public Policy and Management, Carnegie Mellon University, and a faculty research fellow at the National Bureau of Economic Research. Martin S. Gaynor, Ph.D., is the E. J. Barone Professor of Economics and Health Policy at the Heinz School of Public Policy and Management, Carnegie Mellon University, and a research associate at the National Bureau of Economic Research. Address correspondence to Prof. Abraham at Division of Health Policy and Management, School of Public Health, University of Minnesota, Mayo Mail Code 510, 420 Delaware St., S.E., Minneapolis, MN 55455-0392. Email: [email protected]. Inquiry 43: 315–332 (Winter 2006/2007). Ó 2006 Excellus Health Plan, Inc. 0046-9580/06/4304–0315 www.inquiryjournal.org

Inquiry/Volume 43, Winter 2006/2007

now representing 50% of total households, while single adults and single adults with dependents comprise 33% and 17%, respectively (U.S. Census 2006). Growth in female labor force participation rates, particularly during the 1970s and 1980s, has been an important factor contributing to this shift. Additionally, this growth has led to a higher proportion of two-earner, married households. In fact, among nonelderly married households, approximately 60% now have both spouses working outside the home (U.S. Census 2005).1 Of course, not all households with two workers have two offers of employer-based coverage, since some individuals may be employed at firms that do not offer this fringe benefit. However, an estimated 40% of two-worker households do have two offers of employer coverage and have greater flexibility in their decision making, particularly if household members perceive offered plans to be substitutes.2 In a descriptive study, Abraham and Royalty (2005) found significant differences in the numbers and types of employer-based plans available to two-worker families as compared to other households, suggesting the importance of taking account of the full range of choices available when trying to understand how households make decisions about their health insurance. To date, only a few empirical studies have focused on households and the potential effects of having two offers of coverage on health plan decision making.3 Marquis and Kapur (2005) estimated a model of health insurance choice for families with two offers of coverage using the Current Population Survey and the 1997 Robert Wood Johnson Foundation Employer Health Insurance Survey. They found that price has a small but significant effect, and that the policyholder’s characteristics, including gender, wage income, and tenure also have an impact on family coverage decisions. Blumberg, Nichols, and Banthin (2001) used the 1996 Medical Expenditure Panel Survey (MEPS) to analyze a worker’s decision to take up any coverage. They found that workers are significantly less likely to take up coverage through their place of employment if their spouses also are offered insurance and, more generally, that workers are not very price sensitive. Using earlier data from the 1987 National Medical Expenditure Survey, Monheit, Schone, and Taylor (1999) investigated a household’s decision to take up two health insurance policies that 316

together provide ‘‘double coverage,’’ whereby at least one household member is covered under both policies. They found the probability of a household having double coverage positively related to having at least one policy with no employee contribution. Almost certainly, the primary reason there is not more research using the household as the unit of analysis is the lack of data. There are very few data sources that track detailed information about insurance offerings and allow for the matching of household members. In this paper, we estimate a model of demand using the 1996 MEPS, a rich data source containing information on insurance offerings by employers and individuals’ choices. With these data, it is possible to specify the set of plans available to all eligible workers in a household, as well as to incorporate into the analysis household attributes that may influence preferences regarding plan and coverage type. Understanding households’ coverage decisions also may have important implications for how employers design their health benefits. During this most recent period of premium inflation, many employers have attempted to contain shortrun cost growth by increasing employee contributions and/or decreasing benefit generosity. Some employers also have created policies to reward employees who take up coverage from another employer source when it is available (Costello 2003). In a 1999 Kaiser Family Foundation/Health Research and Educational Trust (HRET) Employer Health Benefit Survey, approximately 13.6% of firms reported offering these financial incentives, with larger firms and those in the finance and technology sectors being more likely to do so (Gabel et al. 2001). In a related literature stream, three studies examined the relationship between fringe benefit design and workforce composition. Dranove, Spier, and Baker (2000) developed a theoretical model of ‘‘employer competition’’ to explain why employee contributions toward health insurance premiums have risen over time. The authors contend that subsidizing family coverage is costly to employers, who may choose to decrease benefit generosity or increase employee contributions as a way to encourage employees to take up coverage through a spouse’s employer when available. Using establishment-level data from the 1993 Robert Wood Johnson Employer Health Insurance Survey, they found

How Households Choose

that employee contributions are higher in firms with more female workers and more part-time workers—groups shown to be more likely to have an alternative source of health insurance available to them in their households. Gruber and McKnight (2003) also estimated a model to examine the rise in employee contributions during the time period of 1982–1996. Using the Current Population Survey, they found some empirical evidence to support the hypothesis that as an employee’s outside options increase, including having coverage through a spouse, required contributions rise. Finally, Vistnes, Morrisey, and Jensen (2006) explored the effects of two-earner household prevalence, as well as other firm and labor market characteristics, on the way employers set the marginal employee premium contribution for family coverage. Using establishment-level data from the 1997–2001 MEPS Insurance Component, they found a positive relationship between the proportion of two-earner spouses in the local labor market and higher single and family premium contributions. A key contribution of this study is that we use household information and our plan choice model estimates to simulate the potential impact on enrollment and employers’ costs of providing a cash payment to workers who do not take up coverage. For an employee with an alternative source of coverage, choosing the ‘‘opt-out’’ payment and becoming a dependent on a spouse’s policy may be welfare-improving if the incentive exceeds the additional costs associated with switching. Predicting how likely employees are to switch from taking up coverage to not taking up coverage depends on both individuals’ sensitivity to price and the size of the opt-out payment. If an employer is seeking to reduce expenditures by offering this provision, then the savings generated from not having to pay the employer portion of costs for those who opt out should be greater than the new costs associated with payments made to those workers who choose not to take up coverage. Offering this type of incentive also may serve as an effective tool for discouraging double coverage by employees, which potentially can increase moral hazard and employers’ costs.4 However, this type of provision also may raise concerns about selection, stability of the risk pool, and plan sponsorship. For example, if

young workers at a firm have lower demand for medical care and choose to opt out, then an employer’s costs could rise if these individuals contribute more to premiums than they use in services or if the opt-out payment exceeds what the employer would have paid on their behalf if they were covered. Furthermore, insurers in some states may require small employer groups to enroll a minimum percentage of eligible employees as a prerequisite for plan sponsorship. In these situations, employers would want to encourage, rather than discourage, enrollment in the company’s plan. Overall, our results reveal that households are sensitive to price with respect to their plan choices and that price sensitivity varies by marital status, wealth, and the number of offers of employer coverage available to the household. We also find that lower-income households are less likely to select an option that provides coverage for all household members. In what follows, we outline the conceptual framework and provide a description of the data and measures. The fourth section discusses our econometric approach, and the fifth section reports our results. In the sixth section we present the ‘‘opt-out’’ simulation and end with concluding remarks.

Conceptual Framework In the standard model of health insurance demand, money and health are the key determinants of a household’s choice of health insurance plan, including the choice of not taking up any employer-based coverage (Cutler and Zeckhauser 2000). Money consists of household income net of health care expenses, which may include an employee contribution toward the total premium, as well as out-of-pocket spending for which copayments and deductibles can serve as proxies. Health state depends on the medical care consumption of household members, demographic characteristics, and a random shock parameter that captures some loss of health by one or more household members during the coverage period that can be restored at least partially with medical care. In this model we assume, as have other studies of health insurance demand, that the employment status of workers and health care coverage options offered to them are exogenous. Of course, if workers sort into jobs or if there is sorting 317


Table 1. Household type distribution One-offer Two-offer households households Single Single with children Married without children Married with children Total number of households

567 196 270 421 1,454

... ... 102 130 232

within households based on particular preferences regarding the trade-off between wage and nonwage benefits, this may bias the parameter estimates. However, the sign of the bias is indeterminate if individuals choose their employers and employment status based on their health insurance options and/or employers respond to employee preferences when establishing their health insurance benefits programs and employee contributions.

Empirical Specification Data Description We use two components of the 1996 MEPS to estimate the health plan choice model (U.S. DHHS 2001). The first is the Household Component (HC), which is a nationally representative survey of the civilian noninstitutionalized population of the United States, containing individual-level data on demographic characteristics, employment status, health status, medical expenditures, and health insurance coverage for 22,601 individuals in approximately 11,000 households. Our definition of a household is based on the constructed health insurance eligibility unit (HIEU) identifier contained in the data file. Specifically, an HIEU is a subfamily relationship unit constructed to include adults plus those family members who typically would be eligible for coverage under private family plans. These family members include spouses, unmarried natural or adopted children who are age 18 or under, and children under age 24 who are full-time students. Household survey respondents who indicated that they were employed were asked for contact information regarding their place of employment, as well as permission to contact their employer. Employers of these household respondents then were surveyed and the results were compiled into the MEPS Insurance Component (IC) database. For the MEPS-IC, information was collected on 318

up to four comprehensive health plans for employees of private establishments and all plans offered by public employers.5 Information was collected on the total premium and employee contribution for both single and family coverage; plan type category (e.g., exclusive provider organization, mixed provider organization, or any choice of provider organization); cost-sharing provisions including coinsurance, copayments, and deductibles; and covered benefits (e.g., prescription drug coverage, well-child care, immunizations) for plans that were both chosen and not chosen by an employee. Employers also were asked to verify employee eligibility and to confirm the plan and coverage type held by the employee. Our study population includes households in which one or more members are between the ages of 19 and 64, employed, and offered employerbased coverage through their current main job.6 By imposing these criteria, our sample of eligible households falls to 4,913. We further classify these households by whether they have one offer of coverage or two, since this designation is important for constructing households’ choice sets. Households with one offer of coverage include single adults, single adults with dependents, married adult households in which only one spouse is employed and eligible for coverage, and married households in which there are two workers, but only one is offered insurance. The other category includes households with two workers who both are offered coverage.7 While the data have substantial advantages over other sources, they suffer from one limitation. The sample of workers with complete information contained on the IC is not nationally representative due to a high rate of nonresponse on the combined surveys. Some workers refused to grant survey takers permission to contact their employer, while others provided incomplete or inaccurate contact information. Furthermore, some employers chose not to respond to certain items on the IC, which are necessary for estimating models of health plan choice (e.g., confirming which plan was held). After merging the data files and checking for missing information, the final sample includes 1,686 households. Table 1 reports the distribution of household types represented in the final sample. A set of descriptive statistics was tabulated to examine potential differences between those


households included in the final sample and those that were excluded due to incomplete information. For the households with one offer of coverage, those included in the final sample have higher incomes, have larger numbers of serious medical conditions, and are more likely to have a federal government worker in the household. No statistically significant differences were identified for the households with two offers of coverage.

Table 2. Number of offered plans available to households One-offer households

Two-offer households

Available plans Number Percentage Number Percentage 1 2 3 4 5 6 or more

660 236 130 139 21 268

45.4 16.25 8.94 9.55 1.44 18.43

0 56 27 28 24 97

0 24.14 11.64 12.07 10.34 41.8

Choice Sets Table 2 provides the distribution of the number of health plans available to households in the final sample. Of the 1,454 households with one offer of coverage, approximately 55% have more than one plan from which to choose, and by definition, all households with two offers have plan choice. Furthermore, we consider the choice between single and family coverage since it not only affects which household members are covered, but typically the required employee contribution too. For households with one offer of coverage, we define the set of plan options to consist of all plan-coverage type combinations offered to the employed member, as well as the option of not taking up any employer-based coverage. Not taking up coverage encompasses three possibilities: having public insurance, having nongroup coverage, and choosing to be uninsured.8 To illustrate, if a worker is offered two plans, Plan A and Plan B, then he has five plan options consisting of Plan A-single coverage, Plan A-family coverage, Plan B-single coverage, Plan B-family coverage, and no take-up. For households with two offers of coverage, the choice set is more complex since it consists of all possible combinations of plan coverage type options belonging to each worker in the household, plus the decision to not take up any employer coverage. These households may choose to take up coverage from neither, one, or both employer sources. Moreover, they may choose to take up plans with different combinations of coverage types, such as two single policies, one single and one family policy, or two family policies. To illustrate the choice set construction for households with two offers of coverage, suppose a husband and wife each are offered one plan

through their respective employers (Plan H for the husband and Plan W for the wife). If we define all plan coverage type combinations belonging to each worker, then the household’s choice set consists of nine options including: Plan Hfamily coverage, no take-up-W; Plan H-single coverage, no take-up-W; Plan H-family coverage, Plan W-single coverage; Plan H-single coverage, Plan W-family coverage; no take-up-H, Plan W-family coverage; no take-up-H, Plan Wsingle coverage; Plan H-family coverage, Plan W-family coverage; Plan H-single coverage, Plan W-single coverage; no take-up-H, no take-up-W. Explanatory Variables We hypothesize that several plan attributes affect demand. The relevant price for analyzing household behavior is the annual employee contribution toward the total premium for the particular coverage type (single or family).9 For households with two offers of coverage, some of their options include one plan from each employer. For the options with two plans, we use the sum of the contributions as our measure of price. We also control for whether a plan option covers all household members (‘‘all covered’’). For households with one offer of coverage, we simply use information on whether the option’s coverage type is family coverage. For households with two offers of coverage, we use information on the actual number of household members, as well as the coverage type associated with the plan or plans that comprise the option, to construct a variable for whether that option provides coverage for all household members or not. In the MEPS-IC, plan types are defined by provider organizational structures including exclusive provider organizations (e.g., health 319


maintenance organizations [HMOs]), any provider organizations (e.g., fee-for-service [FFS] plans), and organizations that are a mixture of exclusive and any providers (e.g., preferred provider organizations [PPOs] or point-of-service [POS] plans). In our multivariate analyses, we combine the last two organizational structures into a single category to represent an option with freedom of choice (FOC) of provider. Often, health plans vary in terms of out-ofpocket spending. In the specification, we include measures of the coinsurance rate and annual deductible for outpatient care. For plans that reported copayments rather than coinsurance, we converted these values by dividing the copayment by the average price of an office visit in 1996 and multiplying by 100.10 In doing so, we recognize that the constructed measure is not a true coinsurance rate, since this rate may not apply to all costs for services provided on an outpatient basis. For two-offer household options that include two plans, we use the average deductible and coinsurance rate as our measures. Our rationale for using the average rather than the minimum stems from the fact that when households take up two plans, it is not always the case that all household members have coverage under both plans, which would be necessary for them to be able to use the one with the lowest cost-sharing.11 Covered benefits represent a final set of variables that we consider for the model. For each plan in the linked MEPS data, information is collected on the following services: routine physicals, Pap smears, mammograms, well-child care, prenatal care, immunizations, chiropractic care, alcohol and substance abuse treatment, and prescription drug coverage. We chose to not include these measures in the specification since descriptive analyses revealed that there was too little variation within households’ choice sets to identify any impact on plan choice. Several household characteristics also are included in the model. The first set of measures control for differences with respect to household composition.12 In particular, we include measures for the total number of household members, and the number of children age 18 and younger, an indicator variable for whether the household includes two married adults, and an indicator variable for whether the household has one offer of employer coverage (two offers is the excluded category). 320

Second, given the interrelated nature of medical care consumption and health insurance demand, we include a measure of health status, defined as the number of serious medical conditions per capita in the household.13 These medical conditions include stroke, cancer, heart disease, gall bladder disease, high blood pressure, arteriosclerosis, rheumatism, emphysema, arthritis, and diabetes. We hypothesize that households in which one or more members have a serious medical condition are more likely to take up coverage and also more likely to select a plan option with a freedom of provider choice. Third, purchasing insurance is consistent with being risk averse, which implies diminishing marginal utility of income. Following Holmer (1984), we include the natural log of household wage income in the specification. While the MEPS does not collect actual wealth or asset information, we proxy for it with an indicator variable based on the type of income tax form filed by the household. We assume that households that file a 1040 form, rather than a 1040A or 1040EZ, are more likely to have investment income that requires them to do so. Finally, we include an indicator variable to control for whether at least one individual in the household is employed by the federal government, since information on federal employees’ health plan options and employee contributions were coded directly by survey administrators, rather than through the standard data collection methods. Tables 3 and 4 provide summary statistics for the household and plan characteristics.

Econometric Specification For most discrete choice models, recovering parameter estimates for attributes of the choosers that do not vary across alternatives (e.g., household characteristics) is achieved by interacting those personal attributes with a set of indicator variables corresponding to the specific alternatives in the choice set. Unfortunately, we cannot employ this strategy since each household’s set of plan options is unique, resulting in too many parameters to estimate. Instead, we adopt a more general econometric approach in which we interact our set of household characteristics with two plan attributes including the employee contribution and an indicator variable for whether the option provides coverage for all household


Table 3. Household attributes (means)

Number in household Number of children under age 18 Ln(household wage income) Household files 1040 tax form Serious medical conditions per capita Household has a federal government worker

One-offer households

Two-offer households

2.28 .770 10.33 .474 .425 .082

3.09 1.02 10.94 .659 .376 .082

(1.37) (1.07) (.70) (.50) (.75) (.27)

(1.18) (1.14) (.47) (.48) (.524) (.27)

Note: Standard deviations are in parentheses.

members. Using this specification allows us to test for whether households of different types (e.g., single vs. married, low-income vs. high-income) vary with respect to their price sensitivity and coverage type preferences. In the empirical literature on choice, it is commonly assumed that the errors in a random utility model are additive and independent and indentically distributed (i.i.d.) Weibull, leading to a conditional logit specification. Although simple to implement, this model makes the restrictive assumption that the relative probabilities for any two alternatives depend only on the attributes of these alternatives. This is called the independence of irrelevant alternatives (IIA) assumption. For example, suppose that the probability of choosing FFS vs. an HMO is .5. Now add an HMO and suppose that consumers choose between the two HMOs with equal probability. IIA implies that the probability of choosing the FFS plan must now drop to .33, which is clearly implausible (Maddala 1983). The validity of this assumption has been questioned in the context of health plan choice (Feldman et al. 1989; Royalty and Solomon 1999). Following Feldman et al.

(1989) and Royalty and Solomon (1999), we adopt a nested logit model specification to address this concern. While sharing similar computational properties as the model described previously, the nested logit allows for a general pattern of dependence among the alternatives. In particular, this specification permits some plan options to be more similar than others by placing them into the same ‘‘nest.’’ Although the IIA assumption should still hold among plan options in the same nest, it does not apply to plan options that are in different nests. Figure 1 provides a graphical representation of the particular nesting structure that we adopt. We define the nesting structure based on dimensions for which we believe the plan options are dissimilar in ways not completely observable by the researcher. The first of these, as represented by the ith level in Figure 1, reflects the fact that the option of not taking up any employer coverage may be distinct from options which include actual plan coverage. The second dimension (jth level of Figure 1) captures whether a plan option provides household members with freedom of choice of provider or constrains them to seek care

Table 4. Plan attributes by provider type (means) NFOC plans Exclusive provider organizations (HMOs) Total annual premium, single coverage ($) Annual employee contribution, single coverage ($) Total annual premium, family coverage ($) Annual employee contribution, family coverage ($) Annual deductible ($) Outpatient coinsurance rate (%)

1,967 263 5,033 1,289 8.2 7.7

(460) (313) (929) (1,043) (70.31) (4.38)

FOC plans Any provider organizations (FFS plans) 2,151 302 5,218 1,256 330 17.3

(914) (433) (1,778) (1,331) (600) (7.46)

Mixed provider organizations (PPOs, POS plans) 2,341 529 5,300 1,604 144 9.4

(838) (584) (1,460) (1,372) (249) (6.38)

Note: Standard deviations are in parentheses. FOC ¼ freedom of choice of provider; NFOC ¼ plan without freedom of choice of provider; HMO ¼ health maintenance organization; FFS ¼ fee for service; PPO ¼ preferred provider organization; POS ¼ point of service.

321


Figure 1. Nesting structure (FOC ¼ freedom of choice of provider; NFOC ¼ plan without freedom of choice of provider) from an exclusive network of providers (e.g., non-FOC plans or NFOC). In the MEPS-IC, plans that are consistent with the FOC designation include FFS plans, PPOs, and POS plans, while the NFOC designation represents HMOs.14 The bottom level (k) is the actual choice of plan option, conditional on being in either the FOC or NFOC nest and taking up coverage. We estimate the nested logit in three steps, starting from the bottom level and working upward. As the first step, for the FOC nest and NFOC nest, we estimate separate plan choice equations as conditional logit models. Then, us^ and values of ing the parameter estimates (bs) the explanatory variables (X), we construct an inclusive value term (I1), defined as the following: " # J1 X ^ FOC Þ expðX b I1 ¼ log k¼1

"

log

J2 X

# ^ NFOC Þ expðX b

k¼1

with robust standard errors for a household’s decision to choose an FOC plan option or not.16 This decision is specified to depend on household attributes (H) entered by themselves and the inclusive value term (I1).17 In order to estimate the household’s decision to take up any employer-based coverage or not, we construct a second inclusive value term (I2), which encompasses information from the previous steps. It is defined as the following: 8 log½expðH^ g þ ^tIFOC Þ þ expð^t INFOC Þ > > > if household has both NFOC and > > > > > FOC in its choice set > > > > g þ ^tIFOC Þ if household does < log½expðH^ I2 ¼ not have a NFOC option in its > > choice set > > > > > log½expðH^ g þ ^tINFOC Þ if household > > > > does not have a FOC option in its > : choice set: The take-up regression is also estimated as a binary logit with robust standard errors.

I1 ¼ IFOC INFOC : Intuitively, I1 captures the difference in ‘‘quality’’ between the set of FOC plan options and the set of NFOC plan options available to a household and it is included in the next step of estimation.15 For the second step, we estimate a binary logit 322

Results Descriptive Statistics Figures 2 and 3 show the coverage and plan types held by the one-offer and two-offer households,


Figure 2. Number of plans, types of plans, and coverage types held by one-offer households (FOC ¼ freedom of choice of provider; NFOC ¼ plan without freedom of choice of provider; EBHI ¼ employer-based health insurance) respectively. For households with one offer, 92% take up coverage, with approximately 47% choosing single coverage and the remaining selecting family coverage. Of those holding coverage, 62% choose FOC plans, while the remainder select NFOC plans. For households with two offers of coverage, approximately 97% take up at least one employer-based plan, and 56% select an option with two plans. In fact, nearly 38% of households with two offers take up a combination of plans that provides ‘‘double coverage’’ for at least one member of the family.18 Nested Logit Results We use the full sample of households to estimate the model and report the results in Tables 5 and 6.19 All of the regressions are statistically significant and have pseudo-R2 statistics ranging from .082 to .32. To evaluate how well the model predicts households’ plan choices, we computed a statistic for the proportion of households in the sample for which the predicted probability of the plan option actually chosen by the house-

hold exceeds the probability of choosing that plan option at random from the household’s choice set. For the reported specification, this statistic is .82. Since all of the regressions are estimated as logit specifications, only the sign and the significance of the parameter estimates are directly interpretable; the magnitudes are not. Panel 5A reports the results for a household’s decision to take up any employer-based coverage. Not surprisingly, household wage income is positively related to the probability of taking up any employer-based coverage, ceteris paribus. To illustrate the magnitude of this effect, we compute the change in probability of take-up by a ‘‘representative’’ household for an increase in income from one standard deviation below the mean (9.709 log units or $16,465) to the sample mean (10.414 log units or $33,322).20 For this change, the household’s probability of take-up increases from .866 to .923. In addition to finding a positive effect of income, household wealth, as measured by our 1040 tax form proxy, also is positively associated with the decision to take up coverage. 323


Figure 3. Number of plans, types of plans, and coverage types held by two-offer households (FOC ¼ freedom of choice of provider; NFOC ¼ plan without freedom of choice of provider; EBHI ¼ employer-based health insurance) A final note regarding this first regression is that the parameter estimate on the inclusive value term (I2) equals .574 and is statistically significant (p , .01). Intuitively, this result suggests that as the overall ‘‘quality’’ of a household’s set of options increases, it is more likely to take up coverage, all else equal. With this result, we also can reject the null hypothesis that the parameter equals one, a finding which suggests violation of the IIA assumption and that use of the nested logit model is appropriate. In Panel 5B, we report results for our equation for whether a household chooses an FOC plan option or not, conditional on taking up coverage. Like the take-up decision, we find some empirical support that higher income households exhibit stronger preferences for greater provider choice, a result that is consistent with findings from earlier studies such as Cutler and Reber (1998), 324

Royalty and Solomon (1999), and Parente, Christianson, and Feldman (2004). The health status of household members appears to influence decision making as well. Specifically, households with more serious medical conditions per capita are more likely to select a plan option that gives them the ability to go outside of an exclusive provider network when seeking medical care. Results for the final two regressions are reported in Table 6 and correspond to the specific plan options chosen, conditional on being in either the FOC nest or the NFOC nest. Regardless of whether the household chooses an FOC or NFOC plan, overall, it is more likely to choose an option that covers all members, relative to one that provides only partial coverage. This inference is based on the positive coefficient estimates of 2.69 and 2.83 on the ‘‘all covered’’ variable in the FOC and NFOC regressions,


Table 5. Nested logit: take-up of employer-based coverage and selection of FOC plan (5A) Dependent variable: take-up any EBHI Variable name Number of household members Married household Number of children Ln (household wage income) File 1040 Medical conditions per capita Federal government One-offer household Inclusive value (I2) Inclusive value (I1) Constant Number of households Pseudo R2

(5B) Dependent variable: choose an option with FOC plan

Parameter estimate

Robust standard error

Parameter estimate

Robust standard error

.084 2.048 2.118 .872*** .518** .313* 2.508 2.038 .574*** . . .. 2.72

.459 .497 .472 .151 .217 .172 .417 .486 .134 ... .124

2.243 .190 .062 .248* .228 .233** 2.239 2.004 ... .481*** 2.574

.257 .336 .261 .138 .171 .116 .226 .231 ... .081 .091

1,686 .136

734 .082

Note: Model was estimated using explanatory variables that were de-meaned, which in turn affects the value of the constant. EBHI ¼ employer-based health insurance; FOC ¼ freedom of choice of provider. * p , .10. ** p , .05. *** p , .01.

respectively. By examining the interactions of this variable with household characteristics, we can evaluate whether different types of households vary in terms of their coverage preferences. For example, based on the positive and significant estimates on the interaction with income (.535 in the FOC nest and .643 in the NFOC nest), we infer that the preference for complete household coverage is even stronger as household wage income increases, or alternatively, that lower-income households are less likely to choose an option that provides complete household coverage. Our results are mixed in terms of households’ preferences on the dimensions of coinsurance and deductibles. Within the NFOC nest, our estimates are negative, but imprecisely estimated. However, in the FOC nest, the parameter estimate on the coinsurance rate is positive and significant, which is opposite in sign to what would be predicted. One possible explanation is that there are unobservable plan characteristics (e.g., a larger or higher quality provider network) that are positively correlated with cost-sharing, and therefore individuals might be willing to choose plans with higher cost-sharing if it means having better provider access.21 The model reveals strong support for the effect of price on health plan choice – households are

less likely to choose plan options that have higher employee contributions, holding all else constant. This conclusion is based on the negative and statistically significant parameter estimates on the contribution variable in both equations (.081 in the FOC nest and .091 in the NFOC nest). Once again, we can assess whether households of different types are more or less price sensitive by looking at whether significant coefficients on the interaction terms are negative or positive, respectively. In the FOC nest, price sensitivity appears to vary on two household dimensions. First, households that reported filing a 1040 tax form are less price sensitive—a finding consistent with expectations. In contrast, our results show that households that have only one offer of employer coverage are more price sensitive, as indicated by the .027 parameter estimate in the FOC nest and the .068 estimate in the NFOC nest. One possible explanation for this result is that it is capturing an unobserved income effect. In the NFOC nest, we also find that married households, as compared with singles and singles with dependents, are less price sensitive, while federal government households exhibit stronger price sensitivity. To obtain a better sense of how price changes affect decision making for different types of 325


Table 6. Nested logit: choice of plan options Choice of plan options FOC Variable name Contribution ($100s) Number of household members*Contribution Married*Contribution Ln (household wage income)*Contribution File 1040*Contribution Medical conditions per capita*Contribution Federal government*Contribution One offer*Contribution Number of children*Contribution All covered Number of household members*All covered Married*All covered Number of children*All covered Ln (household wage income)*All covered File 1040*All covered Medical conditions per capita*All covered Federal government*All covered One offer*All covered FFS plan type Deductible ($) Coinsurance rate (%) Pseudo R2

NFOC

Parameter estimate

Standard error

Parameter estimate

Standard error

2.081*** 2.008 .038 .005 .041*** 2.007 2.014 2.027* .018 2.69*** .827 2.972 2.868 .535** 2.077 2.183 2.426 .626 2.847*** 2.0001 .040***

.008 .025 .029 .011 .014 .009 .022 .015 .026 .158 .609 .671 .641 .230 .308 .176 .449 .533 .187 .0001 .014

2.091*** .008 .088** 2.008 .006 2.013 2.102** 2.068** 2.005 2.83*** .079 2.668 2.077 .643** .618 .056 .271 1.66* ... 2.0008 2.014

.012 .035 .043 .014 .022 .022 .045 .028 .036 .214 .712 .855 .722 .322 .410 .273 .592 .880 ... .0007 .022

.32

.25

Note: FOC ¼ freedom of choice of provider; NFOC ¼ plan without freedom of choice of provider. * p , .10 ** p , .05 *** p , .01.

households, we compute price elasticities for six ‘‘representative’’ households, including a single person, a single person with one child, a married couple with no children (one offer), a married couple with children (one offer), a married couple with no children (two offers), and a married couple with children (two offers). For our calculations, we use average values of the type-specific household characteristics and assign each household two plans (one FOC plan and one NFOC plan) that have characteristics with mean or modal values of available plans in the data. With plan options varying by coverage type, plan type, and even the number of plans, we modify conventional interpretation methods. More specifically, our elasticities are defined in terms of the effect of a one-percentage increase in the out-of-pocket contribution on the percentage change in ‘‘expected covered lives.’’ For each plan option in a household’s choice set, we calculate covered lives by allocating the number of household members to each category based on 326

information about plan type(s) and coverage type(s). For example, if a household has three members and the option is an FOC plan with single coverage, then there are zero NFOC covered lives, one FOC covered life, and two ‘‘no take-up’’ lives.22 Expected covered lives are calculated for each plan option by multiplying the predicted, unconditional probability of an option by the covered lives for each category (FOC, NFOC, and no take-up), and then summing over all of the options in the household’s choice set.23 Table 7 reports these ‘‘own-price’’ elasticity estimates. Overall, the results suggest that households are sensitive to changes in the price of coverage, and that households tend to be somewhat more sensitive to changes in the price of FOC plan options relative to NFOC options. Across all household types, the magnitudes we compute are similar to those reported in other studies that use national samples, but they are smaller than estimates found in recent, single-firm studies of health plan choice such as Royalty and


Solomon (1999), Cutler and Reber (1998), and Buchmueller and Feldstein (1997). Greater product differentiation and/or unobserved plan characteristics within choice sets that are correlated with the contribution may partially explain our smaller price elasticities (Berry, Levinsohn, and Pakes 1995).

Robustness Checks In our analysis, we perform several sensitivity checks to evaluate the degree to which our results change when we modify the nested logit specification or estimated separate models for different household types.24 One reason that households may not take up employer-based coverage is because they are eligible for public insurance. To address this concern, we include an indicator variable for whether a household’s income is below the federal poverty line. This variable and its interactions with plan characteristics are statistically insignificant across all three levels of the nested logit. As a second approach, we re-estimate the model excluding those households that reported having other health insurance (e.g., nongroup and public insurance coverage). The results reveal that households are more likely to select an option that covers all members, that they are less price sensitive, and that individuals in worse health status are more likely to take up coverage. Given the large proportion of two-offer households that select an option with double coverage, we estimate a model with an indicator variable for whether a plan option in a household’s choice set includes either the combination of single and family coverage or two family coverage policies. Multicollinearity between this variable and the ‘‘all covered’’ variable leads us to substitute one for the other in the specification. We find no significant interactions between choosing double coverage and household characteristics, and this model has a poorer fit relative to the reported specification. Another important determinant of a household’s choice of plan may be its anticipated out-of-pocket spending.25 We estimate an ordinary least squares (OLS) regression of each household member’s total medical expenditure as a function of age, sex, race, chronic medical conditions, and health insurance status.26 Then, we predict each individual’s expenditure assuming private, employer-group coverage and aggregate up to the household level. We then include this measure as an explanatory

Table 7. Price elasticities of demand FOC price elasticity

NFOC price elasticity

Workers in one-offer households Single 2.19 Single with children 2.79 Married no children 2.78 Married with children 2.30

2.09 2.48 2.74 2.23

Workers in two-offer households Married no children 2.32 Married with children 2.17

2.09 2.04

Note: Elasticities are defined as the percentage change in expected covered lives given a one-percentage increase in price. FOC ¼ freedom of choice of providers; NFOC ¼ plan without freedom of choice of provider.

variable in the specification. We find no significant relationship between a household’s predicted medical expenditures and its decision either to take up coverage or to choose a plan with freedom of choice of provider. Our final robustness check includes partitioning the sample and estimating separate models for single-member households, multiple-member households with one offer of coverage, and multiple-member households with two offers of coverage. For singles, we still find an inverse relationship between the contribution and choice of plan option, but do not find that it varies by other household attributes such as income. Singles with serious medical conditions also are more likely to take up coverage. For the multiplemember households with one offer of coverage (e.g., singles with dependents, married couples with and without children), we find an inverse relationship between the contribution and plan choice, although the magnitude of this effect is smaller. Other results are qualitatively similar to those in the reported specification. Finally, because only seven ‘‘two-offer’’ households do not take up coverage, it is not possible to model the take-up decision. However, for the plan choice equation we still find significant price and coverage type effects. Our results also reveal that wealthier households, as measured by our 1040 tax form proxy, are more likely to choose a plan with FOC of provider.

Opt-Out Payment Simulation One hypothesis that has gained traction in the past few years is that given the rapid growth of 327


Table 8. Average probability of employee take-up of coverage from a particular employer source (means) Baseline

$400 incentive

$600 incentive

$800 incentive

$1,000 incentive

Workers across all household types Workers in one-offer households Single Single with children Married no children Married with children

.900 (.003)

.888 (.003)

.882 (.003)

.875 (.004)

.868 (.004)

.920 .931 .873 .935 .917

.911 .912 .852 .935 .923

.907 .907 .846 .933 .921

.904 .901 .839 .931 .919

.900 .896 .832 .928 .917

Workers in two-offer households Married no children Married with children

.776 (.012) .809 (.017) .752 (.016)

(.002) (.003) (.008) (.005) (.004)

(.002) (.003) (.009) (.005) (.004)

.741 (.013) .774 (.018) .717 (.017)

(.003) (.003) (.009) (.006) (.004)

.720 (.013) .753 (.019) .695 (.018)

(.003) (.004) (.009) (.006) (.004)

.697 (.014) .730 (.020) .672 (.018)

(.003) (.004) (.009) (.006) (.004)

.673 (.014) .706 (.021) .648 (.019)

Note: Standard errors are in parentheses.

health care costs as well as the prevalence of twoworker households, employers may design their health benefits to compete to be the ‘‘employer not chosen’’ to subsidize coverage for workers with families. While decreasing benefit generosity and increasing contribution requirements represent two competitive strategies, a third such alternative for an employer is to provide a financial incentive to its workers to not take up its offered coverage. The potential cost savings to the employer of offering this incentive depends on the size of the payment, the switching behavior of workers and their characteristics, and the employer portion of health insurance costs. Only when the savings generated from not paying the employer portion of health insurance for those opting out exceeds the new costs associated with the optout payments is the provision cost-saving to the employer. To help understand which types of employees may be more likely to accept opt-out payments, we simulate a situation with an opt-out selection in the mix of choices. First, we estimate the likelihood that an employee switches from taking up coverage to not taking up coverage through his own employer source when an opt-out payment becomes available. To do this, we use our model estimates to calculate the baseline probability that a worker chooses a plan option from his employer that includes actual coverage. Then, assuming a particular opt-out payment amount, we re-calculate this probability of taking up coverage.27 The difference between the two proba328

bilities represents the magnitude of switching. We calculate this for all offered workers in our households using four different opt-out payment levels ($400, $600, $800, and $1,000). Table 8 reports the average probability of take-up by the number of offers available to the worker’s household, as well as by marital status and presence of children. On average, we find that the baseline predicted take-up rate for workers in households with one offer of coverage is quite high at 92%, while the percentage of workers from two-offer households taking up coverage from a particular employer is only 77.6%. Furthermore, providing a $1,000 payment to workers in two-offer households is associated with a reduction of .103 or 13.3% in the average probability of taking up coverage. In contrast, we find a much smaller response by workers in households with one offer of coverage, whereby the resulting change is .021, or a 2.3% decrease in take-up. The size of workers’ responses to the incentive also varies by the marital status and presence of dependents in the household. Among one-offer households, single adults with children exhibit the largest response to the incentive, which may be related to the fact that among all household types, this group has the lowest average income. Table 9 reports a set of descriptive statistics identifying associations between households’ likelihood of opting out of coverage (reported as quartile categories) and household factors including income, the total number of serious medical conditions reported by members, and predicted


Table 9. Average household characteristics by opt-out probability quartile

Variable

Quartile 1 (Lowest opt-out probability)

Quartile 2

Quartile 3

Quartile 4 (Highest opt-out probability)

.834 (.06)

1.269 (.09)

.720 (.06)

.602 (.05)

52,216 (1,660) 5,256 (120)

46,729 (1,261) 3,594 (120)

32,114 (1,014) 3,164 (122)

35,303 (1,473) 3,410 (109)

Total number of serious medical conditions* Household wage income ($)* Predicted household medical expenditures ($)*

Note: Standard errors are in parentheses. * Indicates a statistically significant difference of means by opt-out probability quartile with p , .05.

household medical expenditures. From this pattern of results, we see that households with higher incomes and higher predicted medical expenditures are the least likely to opt out of coverage from a particular employer source. For the final part of the simulation, we calculate the potential cost savings to the employer associated with offering an opt-out payment. This is accomplished in several steps. First, we calculate the baseline probability of choosing each plan option in a worker’s choice set. Second, we multiply the employer contribution associated with each particular option by this baseline option probability, and then sum over the set of options. This provides a baseline average cost of providing coverage for the worker. Third, we calculate four new sets of plan option probabilities, one corresponding to each opt-out payment level. Additionally, we modify employers’ costs associated with the ‘‘no take-up’’ option from a value of zero to the dollar amount associated with the opt-out payment. Using the new probabilities

Table 10.

and employer costs, we produce a set of estimates to show the extent to which costs change under the various opt-out payment scenarios. Table 10 reports the average employer cost of coverage for the four levels of opt-out payments. In this table, we additionally break down the costs by whether the household has one or two offers of coverage, as well as by the marital status and presence of children. Across all categories, there is a small reduction in the employer portion of costs resulting from offering an opt-out provision, with a somewhat larger effect for the employers of workers in households with two offers of coverage. For example, given a $1,000 opt-out payment, we find that for employers of workers in one-offer households, the average cost per employee decreases $40, from $2,505 to $2,465, whereas for employers of workers in two-offer households the average cost falls $267. Also, changes in cost vary by the marital and dependent status of workers’ households. For single workers with children, costs fall by

Average employer cost of coverage Baseline ($)

$400 incentive ($)

$600 incentive ($)

$800 incentive ($)

$1,000 incentive ($)

Workers across all household types Workers in one-offer households Single Single with children Married no children Married with children

2,465 (35)

2,440 (35)

2,426 (35)

2,410 (35)

2,394 (35)

2,505 (38) 1,668 (29) 2,826 (103) 3,149 (93) 3,068 (77)

2,491 1,643 2,771 3,146 3,080

2,482 1,635 2,752 3,139 3,074

2,474 1,627 2,733 3,133 3,069

2,465 1,619 2,713 3,126 3,063

Workers in two-offer households Married no children Married with children

2,215 (81) 2,037 (117) 2,359 (112)

2,125 (80) 1,958 (116) 2,256 (110)

(38) (29) (102) (93) (78)

(38) (29) (103) (93) (77)

2,069 (80) 1,910 (115) 2,193 (109)

(38) (29) (102) (93) (77)

2,010 (79) 1,859 (115) 2,128 (109)

(38) (29) (102) (93) (77)

1,948 (79) 1,805 (114) 2,060 (108)

Note: Standard errors are in parentheses.

329


$113 on average given the $1,000 incentive. However, for workers who are married with children, this change is only $5. Relating these results back to the model proposed in Dranove, Spier, and Baker (2000), our findings provide more direct evidence that optout provisions influence behavior of workers with respect to taking up coverage and, indeed, may serve to provide another method by which employers compete to not be chosen when workers have an alternative source of coverage. Two potential limitations are worth noting. First, we do not account for the differential tax treatment of an opt-out payment relative to health insurance premium contributions. Second, in our calculation of employers’ costs, we do not explicitly consider any new costs that may be incurred to administer an opt-out payment provision.

Concluding Remarks Using the 1996 Medical Expenditure Panel Survey, we have estimated a nested logit model of household demand for employer-based health insurance to understand the extent to which coverage choices, price sensitivity, and the willingness to take up cov-

erage from a particular employer source vary across different types of households. In contrast to other studies in the literature, we are able to take account of the full range of plan choices available to household members including the option of not taking up any employer-based coverage. Additionally, we are able to include detailed information about the characteristics of workers’ households to better understand how those factors influence health plan decision making. As employers face significant growth in health insurance premiums, many strategies are being proposed to contain costs, including changes in plan offerings, increasing employee contribution requirements, and using financial incentives to influence workers’ demand for health insurance. Our results show that different types of households respond differently to these benefit design strategies. In particular, our opt-out simulation reveals that households with two offers of coverage are able to take advantage of their increased flexibility and are more likely to switch from taking up to not taking up coverage from a particular employer source when they are offered alternative compensation.

Notes Special thanks to Jessica Vistness, John Sommers, and Ray Kuntz at AHRQ, Anne Beeson Royalty, Roger Feldman, Jon Christianson, and seminar participants at the University of Minnesota for valuable advice. The usual caveat applies.

5 6

1 We recognize that household members’ employment decisions may be related to their health insurance options and that empirical analyses are needed to better understand this relationship. Here, we take as given the employment status and health insurance eligibility of household members. 2 This statistic was computed using the 2001 MEPS Household Component database. 3 Scanlon, Chernew, and Lave (1997) and Morrisey (1992) provide excellent reviews of the empirical literature on health plan choice. Other more recent studies focus on health plan choice and switching behavior by university employees. These include Royalty and Solomon (1999), Cutler and Reber (1998), and Buchmueller and Feldstein (1997). All three studies found that employees, on average, were price-elastic with respect to their demand for coverage. 4 This can occur when cost-sharing (e.g., coinsurance and deductibles) is reduced as a result of having primary and secondary payers. However, this 330

7

8

9

effect may be mitigated somewhat by coordination of benefits policies. For this analysis, we do not consider supplemental plans such as dental, vision, or separate prescription drug coverage. We also excluded a small number of households in which members were enrolled in Champus/VA coverage. We restrict our two-offer household designation to only those households in which both workers are able to cover each other under their respective policies, ruling out such arrangements as a parent and an adult child over the age of 24. By our definition, the parent and adult child would be treated as two separate households. Unfortunately, the choices of public insurance and nongroup coverage could not be modeled separately due to a lack of specific information on plan characteristics and small sample sizes. In our final sample, of the 122 households that did not take up employer-based coverage, approximately 12% were enrolled in public insurance programs, 29% had nongroup coverage, and 59% were uninsured. The MEPS IC asks employers to report the contribution for an ‘‘average full-time employee,’’ which may or may not accurately reflect what the particular household member would pay. However, through discussions with MEPS surveyors, we have been


10

11

12

13

14

15

16

17

informed that among establishments offering coverage to all workers, only a small percentage vary contribution requirements between full-time and part-time workers who are offered insurance. The average price of an office visit was computed from the MEPS Household Component file. We also investigated the relationship between our constructed coinsurance values and plan type. For exclusive provider organizations, the mean coinsurance rate was 7.7% with a standard deviation of 4.38. For mixed provider organizations, the mean was 9.4% with a standard deviation of 6.38%. Finally, any provider organizations had a mean coinsurance rate of 17.4% and a standard deviation of 7.46%. While the mean and standard deviation are smaller for plans without freedom of choice of provider, there is still quite a bit of variation both within and across plan types. As a robustness check, we estimated the model using the minimum coinsurance and deductible for options that include two plans. Our results are qualitatively similar. An alternative strategy is to split the sample on one or more of these dimensions and estimate separate models. We pursued this strategy in addition to the one we report in the paper. Comments on our findings are discussed in the results section. We tried several measures of health status, including whether anyone in the household had a serious condition and the total number of conditions. The results did not change. Due to the small number of households having a choice of FFS plans, we are unable to separately nest FFS and PPO/POS plans. However, recognizing that these plan types are different, we include a dummy variable for whether the plan option includes a FFS plan in the choice equation within the FOC nest. We assign a value of zero to IFOC if the household does not have FOC plan options. Similarly, we assign a value of 0 to INFOC if the household does not have any NFOC plan options. Hardin (2002) shows that the sandwich estimate of variance (computed using the robust option in Stata 7.0) is asymptotically equivalent to the MurphyTopel estimator of variance that is often used to correct the standard errors obtained by estimating the model in steps. By testing the null hypothesis that the parameter on the inclusive value equals one, we can evaluate the validity of the IIA assumption. Rejecting the hypothesis implies a rejection of the IIA assumption (Maddala, 1983).

18 Using the 1987 National Medical Expenditure Survey, Monheit et al. (1999) found that approximately 68% of households in which both workers are offered health insurance had double coverage. 19 Different subsets of households are used to estimate each of the regressions. For the decision to take up any coverage, we use the entire sample. For the decision of whether a household chooses an FOC plan option, only those households that actually have both an FOC and NFOC option are used. Finally, for the plan choice decisions conditional on being in either the NFOC or FOC nest, only those households that hold that particular plan type and have choice of that plan type are included. 20 The ‘‘representative’’ household is married with one child, does not file a 1040, has .333 serious medical conditions per capita, does not work for the federal government, has one offer of coverage, and has a set of plan options that yield an inclusive value term equal to the sample mean. 21 We also investigated whether this predicted effect of coinsurance differed by plan type by including interactions of the cost-sharing measures with the fee-for-service plan type indicator. We did not find evidence to support this hypothesis. 22 For two-offer household options including two plans, this is more complicated. For example, two-offer households whose options include two family coverage plans of different types had their members split equally. So, for a household with four members, if the option contains an FOC plan with family coverage and an NFOC plan with family coverage, then there are two FOC covered lives, two NFOC covered lives, and zero no take-up covered lives. 23 Marginal effects were computed discretely to account for the interaction terms in the model. Elasticities were then calculated using these marginal effects and average employee contributions by plan type. 24 Model results for these specifications are available from the corresponding author by request. 25 We thank an anonymous reviewer for suggesting this analysis. 26 To account for intra-household correlation with respect to medical expenditures, we used the cluster option in Stata. 27 To simulate the availability of the opt-out payment, we increase the contribution of all plan options that include coverage by the amount of the incentive. This is a statistically equivalent way to compute workers’ behavioral responses, given the empirical framework.

References Abraham, J.M., and A. Beeson Royalty. 2005. Does Having Two Earners in the Household Matter for Understanding How Well Employer-Based Health Insurance Works? Medical Care Research and Review 62(2): 167–186.

Berry, S., J. Levinsohn, and A. Pakes. 1995. Automobile Prices in Market Equilibrium. Econometrica 63(4): 841–890. Blumberg, L.J., L.M. Nichols, and J.S. Banthin. 2001. Worker Decisions to Purchase Health Insurance. 331


International Journal of Health Finance and Economics 1(3/4): 305–325. Buchmueller, T.C., and P.J. Feldstein. 1997. The Effect of Price on Switching among Health Plans. Journal of Health Economics 16: 231–247. Carroll, W. 2005. The Health Insurance Status of U.S. Workers, 2003: Estimates for Civilian Noninstitutionalized Workers Ages 16–64. MEPS Statistical Brief #71, Rockville, Md.: Agency for Healthcare Research and Quality. Available at: http://www.meps.ahrq.gov/papers/st71/stat71.pdf. Accessed October 2006. Costello, D. 2003. Firms Cut Back Medical Coverage; Faced with Soaring Costs, Many Employers are Discouraging Workers from Adding Spouses or Children to their Insurance Plans. Los Angeles Times Oct. 6: F1. Cutler, D., and R. Zeckhauser. 2000. The Anatomy of Health Insurance. In Handbook of Health Economics, Vol. 1A, A.J. Culyer and J.P. Newhouse, eds. Elsevier: New York. Cutler, D., and S. Reber. 1998. Paying for Health Insurance: The Tradeoff Between Competition and Adverse Selection. Quarterly Journal of Economics 113: 433–466. Dranove, D., K.E. Spier, and L. Baker. 2000. Competition among Employers Offering Health Insurance. Journal of Health Economics 19: 121–140. Feldman, R., M. Finch, B. Dowd, and S. Cassou. 1989. The Demand for Employment-Based Health Insurance Plans. Journal of Human Resources 24(1): 115–142. Gabel, J., L. Levitt, E. Holve, et al. 2002. Job-Based Health Benefits in 2002: Some Important Trends. Health Affairs 21(5): 143–151. Gabel, J., J. Pickreign, H. Whitmore, and C. Schoen. 2001. Embraceable You: How Employers Influence Health Plan Enrollment. Health Affairs 20(4): 196–208. Gruber, J., and R. McKnight. 2003. Why Did Employee Health Insurance Contributions Rise? Journal of Health Economics 22: 1085–1104. Hardin, J.W. 2002. The Robust Variance Estimator for Two-Stage Models. Stata Journal 2(3): 253– 266. Holmer, M. 1984. Tax Policy and the Demand for Health Insurance. Journal of Health Economics 3: 203–222.

332

Maddala, G.S. 1983. Limited Dependent and Qualitative Variables in Econometrics. New York: Cambridge University Press. Marquis, M.S., and K. Kapur. 2005. Family Decision Making When Two Workers are Offered Group Coverage. Working paper. Arlington, Va.; RAND. McFadden, D. 1974. The Measurement of Urban Travel Demand. Journal of Public Economics 3: 303–328. Monheit, A., B. Steinberg Schone, and A. Taylor. 1999. Health Insurance Choices in Two-Worker Households: Determinants of Double Coverage. Inquiry 36(1): 12–29. Morrisey, M. 1992. Price Sensitivity in Health Care: Implications for Health Care Policy. Chicago: American Hospital Association. Parente, S.T., J. Christianson, and R. Feldman. 2004. Employee Choice of Consumer-Driven Health Insurance in a Multiplan, Multiproduct Setting. Health Services Research 39(4): 1091–1111. Royalty, A. Beeson, and N. Solomon. 1999. Health Plan Choice Price Elasticities in a Managed Competition Setting. Journal of Human Resources 34: 1–41. Scanlon, D., M. Chernew, and J. Lave. 1997. Consumer Health Plan Choice: Current Knowledge and Future Directions. Annual Review of Public Health 18: 507–528. StataCorp. 2001. Stata Statistical Software: Release 7.0. College Station, Texas: Stata Corporation. U.S. Census Bureau. 2006. U.S. Households and Families, Table S1101. Available at: http://factfinder. census.gov. Accessed October 2006. U.S. Census Bureau, Population Division, Fertility and Family Statistics Branch. 2005. America’s Families and Living Arrangements: 2005. Available at: http://census.gov/population/www/socdemo/ hh-fam/cps2005.html. Accessed October 2006. U.S. Department of Health and Human Services (DHHS), Agency for Healthcare Research and Quality. 2001. Medical Expenditure Panel Survey: Household Component. Available at: http:// www.meps.ahrq.gov/data_pub/hc_toc.htm. Accessed October 2006. Vistnes, J., M. Morrisey, and G.A. Jensen. 2006. Employer Choices of Family Premium Sharing. International Journal of Health Care Finance and Economics 6(1): 25–47.