The key to simulating changes in the health care system is to develop a baseline database that depicts the U.S. health care system in detail. Our HBSM baseline data is based upon the 1999 through 2001 Medical Expenditures Panel Survey (MEPS) data, which provide information on sources of coverage and health expenditures for a representative sample of the population. These data are adjusted to reflect the population and coverage levels reported in the 2006 Current Population Survey (CPS) data (with adjustments for under-reporting discussed below). We also statistically match workers in these data to the Kaiser/HRET survey of employers which provides additional detail on coverage provided through work. The methods used to develop baseline data for households and employers is presented in the following sections.
The Household Database. The HBSM baseline data is derived from a sample of households that is representative of the economic, demographic and health sector characteristics of the population. HBSM uses the 1999 through 2001 MEPS data to provide the underlying distribution of health care utilization and expenditures across individuals by age, sex, income, source of coverage and employment status. The use of data for three years substantially increases sample size, thus permitting us to develop more stable estimates of narrowly defined policy options.
We re-weighted the MEPS household data to reflect population control totals reported in the 2006 March CPS data. The March CPS is used for the annual Census Bureau estimates of the number of uninsured in the US and each state. While the CPS provides the most current data on insurance coverage, it under-reports the number of people covered under the Medicaid program, which causes these data to over-estimate the number of uninsured. Consequently, we corrected the CPS data for under-reporting of Medicaid coverage to provide a more accurate count of the number of people without coverage.
We corrected the CPS for under-reporting of Medicaid using the HBSM. The model first allocates earnings over the number of weeks each individual worked during the prior year and creates information on income for each month of the year. The model then simulates eligibility for Medicaid and SCHIP using these monthly income data to identify people who appear to be eligible for these programs based upon the income eligibility levels actually used in these programs for various categories of eligibility (e.g. children, parents, etc…). The model does this in a way that accounts for changes in eligibility over the year as people move into and out of employment from month-to-month. We then select a portion of the people who appear to be eligible for Medicaid or SCHIP and assign them to enrolled status so that these data report the correct number of people participating in these programs.
Another issue to deal with is that the CPS reports the number of people who were without coverage from any source during all 12 months of the prior year. However, this definition omits those who were uninsured for only a portion of the year. This not only understates the number of uninsured, it would also lead us to under-estimate the cost of covering these people under various proposals to expand insurance coverage. Thus, the most appropriate measure of the uninsured for policy purposes is the average monthly number of uninsured.
In order to estimate average monthly figures, we allocate reported coverage from each source over the 12 months of the year based upon employment and duration of enrollment data reported in the CPS. We allocate employer wages and employer health insurance coverage over the periods of work reported in the CPS. We also allocated Medicaid and SCHIP coverage over the number of months they report (or are assigned) being enrolled for months where these individuals appear to be income eligible. We assume that people reporting coverage from Medicare, TRICARE or non-group coverage are insured by these sources all year. This enables us to estimate the number of people without insurance coverage in each month.
After adjusting the CPS data, we are able to develop the weights to adjust the MEPS household data to reflect the population control totals reported in the CPS data. These weight adjustments were performed with an iterative proportional-fitting model, which adjusts the data to match approximately 250 separate classifications of individuals by socioeconomic status, sources of coverage and job characteristics in the CPS. Iterative proportional fitting is a process where the sample weights for each individual in the sample are repeatedly adjusted in a stepwise fashion until the database simultaneously replicates the distribution of people across each of these variables in the state.1 This approach permits us to simultaneously replicate the distribution of people across a large number of variables while preserving the underlying distribution of people by level of healthcare utilization and expenditures as reported in MEPS.
The health spending data are adjusted to reflect projections of the health spending by type of service and source of payment in the base year (i.e., 2009). These data are used to estimate health insurance premium costs for people with private health insurance. These spending estimates are based upon health spending data provided by the Centers for Medicare & Medicaid Services and detailed projections of expenditures for people in Medicare and Medicaid spending across various eligibility groups.2 The result is a database that is representative of the base year population by economic and demographic group, which also provides extensive information on the joint distribution of health expenditures and utilization across population groups.
The Employer Database. HBSM includes a database of employers for use in simulating policies that affect employer decisions to offer health insurance. We used the survey of employers conducted by the Kaiser Family Foundation and the Health Research and Educational Trust (Kaiser/HRET). These data include about 2,000 randomly selected public and private employers with 3 or more workers, which provide information on whether they sponsor coverage and the premiums and coverage characteristics of the plans that insuring employers offer.
We statistically match each MEPS worker with one of the firms in the Kaiser/HRET data. Experience has shown that it is important that the individuals assigned to each firm be consistent with the employer’s workforce characteristics. The Kaiser/HRET data provide information on the distribution of workers by wage level. However, additional information such as age of worker and marital status for insured people are not included in the database. Thus, in order to use these data in our analysis, we statistically matched the Kaiser/HRET data with employers surveyed in the 1991 Health Insurance Association of America (HIAA) employer survey data, which provides detailed information on the characteristics of each employer’s workforce including number of workers by:3
- Full-time/part-time status;
- Coverage status (eligible enrolled, eligible not enrolled and ineligible);
- Policy type for covered people (i.e., single/family); and
- Wage level.
The employer health plan eligibility data in the database is important to simulations of policies affecting employers. One important consideration is that many of those who do not have employer coverage work for a firm that offers coverage to at least some of their workers. About 81.5 percent of all workers are employed by a firm that covers at least some of their workers (Figure 2). However, only about 75 percent of these people are eligible and enrolled. About 10.2 percent are ineligible and about 14.3 percent are eligible but have declined coverage.4 Some of these workers have declined coverage because they receive insurance elsewhere (e.g., through their spouse’s employer or the individual market).
Figure 2 Workers by Employer Insurance Status (in millions)
Source: Lewin Group Estimates using the Health Benefits Simulation Model (HBSM).
The model controls for the workforce characteristics for each firm in matching individuals to firms. While the firm data provide information on the number of people in the firm with these characteristics, they do not provide the “joint distribution” across these groups (e.g., by age, sex, income etc.). We estimate the joint distribution for each firm using an iterative proportional fitting process. In this approach, we begin with the joint distribution of workers across these variables as reported nationally in the CPS, and scale them in an iterative process so that in the aggregate they replicate the aggregate number of workers in the firm for each worker characteristic. Each non-zero cell of the joint distribution matrix for each firm is treated as an individual worker, who is matched to MEPS individuals based upon these individual characteristics.
Thus, if a firm reports that it employs mostly low-wage female workers, the firm tended to be matched to low-wage female workers in the MEPS data. This approach helps assure that Kaiser/HRET firms are matched to workers with health expenditure patterns that are generally consistent with the premiums reported by the firm. This feature is crucial to simulating the effects of employer coverage decisions that impact the health spending profiles of workers going into various insurance pools. Controlling for the joint distribution of workers within firms is crucial to simulations of program impacts because premiums and behavioral responses vary widely by age, wage level, part time/full-time status, and the number of workers with dependents.