Few quantitative studies have analyzed the relationship between demographic characteristics and purchase of non-group health insurance. In a 1996 study, the U.S. General Accounting Office (GAO), now the Government Accountability Office, analyzed data from the Bureau of the Census. The authors found that the likelihood of having non-group insurance varied widely by race and ethnicity. Whites were twice as likely to have non-group insurance as blacks or Hispanics (other variables were not controlled for). The GAO hypothesized that this difference was due to the higher median income of whites.

Saver and Doescher (2000) also examined demographic factors associated with purchasing non-group insurance. In contrast to the GAO study their analysis controlled for income, wealth, health status, and other factors using multivariate regression methods. Using data from the 1987 National Medical Expenditure Survey, and a logistic model, they found that Hispanics, Non-Hispanic blacks, and other minorities had less than half the odds of non-Hispanic whites of purchasing non-group insurance. They also found that individuals with less than a high school education had less then half the odds of college graduates of purchasing non-group insurance. These results persisted even after responses to questions on the value of health insurance and healthcare were added to the analysis.

In another study Saver and colleagues (2003) analyzed the roles of community-level and family-level factors in the purchase of non-group health insurance using data from the 1996-1997 Community Tracking Survey.3 The community-level factors were the role of the local social safety net (measured by factors such as the percentage of medically-underserved persons served
3The survey was of 60,446 individuals in sixty communities from across the country, 51 urban and 9 rural, using stratified sampling. With federal and state funds), social measures of income equality and housing segregation, and the percentage of residents of the community who were uninsured. Family-level factors included health and insurance status of all family members. Their analysis, conducted at the family level, confirmed the earlier results that being a minority, or having a lower education level, makes one less likely to purchase non-group health insurance. They also found that among Hispanic respondents those who were interviewed in Spanish were less likely to purchase insurance compared to those who were interviewed in English. Safety-net factors were not found to be significant, with one exception: the annual per capita number of teaching and public hospital outpatient department visits in the community, which was associated with somewhat lower odds of purchasing non-group insurance.

Many studies have examined the impact of price and health status on the decision to purchase insurance in the non-group market. Susan Marquis and Stephen Long (1995) used national data from the Current Population Survey and the Survey of Income and Program Participation to estimate the price elasticity of demand for non-group insurance among workers who do not have employer-sponsored insurance. They found that a 10 percent decrease in price would lead to an increase of about 3 to 4 percent in the number of individuals who would purchase non-group insurance. They estimated that a subsidy that reduced the price of insurance by 40% would increase the number of families who purchase non-group insurance by 8 percentage points. Pauly and Nichols (2002) used multiple regression analysis to look at the probability of having non-group insurance for people in poor health using data from the Community Tracking Study household survey conducted in 2000-2001. Their somewhat contradictory results were that being in a household with a member who reports his or her health status as poor decreases the likelihood of having coverage, but that being in a household with a member with a chronic condition increases the likelihood of buying insurance controlling for age, gender, race, income, education, marital status, and other factors. The authors concluded that those in families with a member with a chronic condition are willing to pay for insurance even at a high price, while those who have a family member in overall poor health are less likely to be willing to pay the price they are asked to pay.

Another study which used data from the 1991 and 2001 Community Tracking Survey found that the probability of buying non-group health insurance decreases significantly with declining health status controlling for education, race, ethnicity, income, and other factors (Hadley and Reschovsky, 2003). The predicted probability of a person in fair or poor health being a policyholder was half that of a similar person in excellent health with no chronic conditions. In addition, the study found that the probability of purchasing non-group insurance decreases somewhat when other family members are reported to be in fair or poor health in comparison to being in excellent health. The study also found that racial and ethnic minorities are less likely to purchase non-group insurance, and that increasing education levels make one more likely to purchase non-group insurance.

DATA
This project uses the 2005 full-year consolidated data file of the household component of the Medical Expenditure Panel Survey (MEPS) produced by the Agency for Health Care Research and Quality (AHRQ). MEPS, initiated in 1996, is the third in a serious of national surveys conducted by AHRQ and its predecessors on healthcare financing and utilization (AHRQ, 2008). Data collected, at the person level, include information on demographic characteristics, health status, health insurance status on a monthly basis, healthcare utilization, and healthcare expenditures. The MEPS sample is representative of the non-institutional civilian U.S. population.
MEPS is designed as a panel survey where each panel is a sub-sample of household respondents to the previous year’s National Health Interview Survey, a survey sponsored by the National Center for Health Statistics. MEPS is a set of surveys including a household survey of the civilian non-institutionalized population, a survey of medical providers, and an independent survey of employers and unions. Each panel of the MEPS household survey is administered through five in-person interviews, called rounds, over a two and a half year period.

This analysis uses the 2005 consolidated person-level data file for the household component. The 2005 file combines data from rounds 3, 4, and 5 of panel 9 and rounds 1, 2, and 3 of panel 10. The overall combined response rate was 61.3% (AHRQ, 2007). The combined data in the 2005 file has a sample size of 12,810 families and 32,320 individuals (AHRQ, 2008). The sample design is based on a complex stratified multi-stage probability design. The survey design includes over sampling of low-income individuals, African Americans, and people with Asian and Hispanic backgrounds. Sample weights are provided to enable the production of nationally-representative estimates.

ANALYSIS PLAN Study Population
Due to the complexity of studying purchase decisions across multiple types of insurance options, e.g. employer sponsored, public programs, non-group market, this study aims to evaluate the decision to purchase non-group insurance when that is the only insurance option available. Information available in the data was used to approximate the population of adult non-elderly individuals who faced the option of purchasing non-group insurance or being uninsured.

This project uses a sub-sample of individuals from the MEPS Household Component consisting of non-elderly adults ages eighteen to sixty-four, who were uninsured all year or who exclusively held non-group health insurance for two or more months. Monthly insurance status variables available in the data were used to identify those who held only non-group insurance (that is, no other private or public insurance alongside non-group) on a monthly basis. The 12 monthly non-group insurance indicators were summed to create an index of the number of months of non-group insurance for each person.
The study population was then restricted to those who either were: uninsured all year, held non-group insurance all year, or purchased non-group during the year and held it for at least two months. This last group was constructed by including people with at least two months of non-group insurance who did not hold the insurance in January.

People who were ever offered employer-sponsored insurance during the year were excluded from the study population. Specifically, respondents were asked at three points during the year whether they had the option of obtaining insurance through their employer. Information on whether individuals could obtain insurance through their spouse’s employer was not available.

In addition, information on whether people were eligible for publicly-sponsored programs but did not enroll was not available. The study population consists of 4,066 individuals. Regression Model This study uses the following logistic regression model:
Log (odds of non-group insurance purchase) = β1Race + β2Ethnicity + β3Education Level + β4English Language Ability + β5 Other Demographic Characteristics + β6Health Status Characteristics + β7Attitudional Characteristics Dependent Variable
Covered by Non-group Insurance: This is a binary variable indicating that the individual had non- group insurance for the entire year or newly-acquired it during the year (as defined in the description of the study population above). In the study population, anyone who does not have non-group insurance is uninsured.
Key Demographic Variables English Language Difficulty: This is a binary variable indicating difficulty with the English language. Language difficulty may hinder one’s ability to navigate the choices in the non-group market, leading the person to not purchase insurance. Hispanic ethnicity: The model includes a binary variable indicating Hispanic ethnicity, which is expected to be associated with a lower probability of purchasing non-group insurance compared to non-Hispanics. Education: This set of binary variables measure the highest degree an individual obtained (reference group no degree): general equivalency diploma, high school diploma, bachelor’s degree, and advanced degree. Higher education levels are expected to be associated with a higher probability of purchasing insurance since education increases an individual’s ability to understand the value of insurance and to navigate the market to purchase insurance. Race: This is a set of binary variables with the categories of white and other and the reference group of black. Because of small sample sizes, American Indian/Alaska Native, Asian, Native Hawaiian/Pacific Islander, and multiple races reported were placed in the “other” category.

Being a minority, black or other, is expected to be associated with a lower likelihood of purchasing non-group insurance in comparison to whites. Other Demographic Variables Age: Age is included in the model as a set of categories, 18-24, 25-44, 45-64, with 18-24 serving as the reference group. This variable is based on a continuous variable edited by MEPS administrators that provides the individual’s age on the last day of the data year (that is, 12/31/05). Age is an important factor in the model since prior research shows that age influences insurance purchasing. On a bivariate level younger individuals are less likely to purchase insurance, in spite of the fact that it is typically cheaper for them (Fronstin, 2008).

Employment Status: This is a binary variable indicating whether an individual is working or not working. This variable was created from a variable with the following responses: currently employed, has a job to return to, employed during the reference period, and not employed with no job to return to. Those who were currently employed or had a job to return to were considered working, and everyone else was considered not working. Employment status was based on the employment variable from the mid-year round (round 4/2).4
4 For those missing information in round 4/2, information from round 3/1 was used to assign a status (13 people); for those who were also missing information in round 3/1 information from round 5/3 was used (103 people). Seven people whose status remained missing after this process were assigned to not working.

Gender: The model includes the binary variable female to indicate respondent’s gender. Past studies have shown that women are more likely to utilize medical care than men, which would make them more likely to buy health insurance (Bertakis et al, 1999)

Poverty Category: Poverty category is included in the model as a series of binary variables (with poor as the reference category): near poor (100% to 125%), low income (125% to less than 200%), middle income (200% to less than 400%), and high income (greater than or equal to 400%). These variables are based on a variable constructed by MEPS administrators that combines income for all members of an identified family, and then places families into poverty categories. The categories are based on the family size and income in relation to the poverty line.

Marital Status: This is a binary variable with categories of married or not married. This is marital status as of 12/31/05. Single people and married couples may have different opinions about insurance which influence purchasing. For example, married individuals may be more likely to purchase insurance since they have responsibilities to someone other than themselves.
Health Status/Health Insurance Status Variables

Health Condition Indicator Variables: Binary variables are included for the following health conditions (numbers in parenthesis indicate the number of individuals who were missing data and who were reassigned to never diagnosed): arthritis (45), asthma (29), diabetes (27), emphysema (28), high blood pressure (30), high cholesterol (83), heart disease (31), and joint pain (44). All condition variables indicate having ever been diagnosed with the condition, except for joint pain which indicates experiencing joint pain in the last 12 months.

Perceived Health Status: Perceived health status is included in the model as a series of binary variables (with excellent as the reference category): very good, good, fair, and poor. Individuals were asked what they perceive their health status to be in all three rounds. This variable is primarily based on round 4/2 information.5 An individual’s perception of their health status likely impacts their decision to purchase insurance, with those in poor health more likely to purchase insurance since they feel they need it.

Smoker: This is a binary variable indicating if an individual currently smokes. Smokers may be less likely to purchase insurance since insurance companies may deem them high risk, and therefore charge them more or deny them coverage. In addition, those who smoke may be more willing to take risks, making them less likely to buy insurance. Individuals who were missing data on this variable were reassigned to non-smoker (455).
Attitudinal Characteristics

Don’t Value Health Insurance: How strongly individuals agreed or disagreed with the statement “I don’t value health insurance” is included in the model as a series of binary variables (with strongly agree as the reference category): agree somewhat, uncertain, disagree somewhat, disagree strongly. This variable indicates attitudes that likely have a strong impact on buying insurance. Individuals who were missing data on this variable were reassigned to uncertain (482).

Willingness to take risks: How strongly individuals agreed or disagreed with the statement “I am more likely to take risks” is included in the model as a series of binary variables with the reference category of strongly agree. It is measured on the same scale as “don’t value health insurance.” Individuals willing to take risks would be expected to be less likely to purchase insurance, as they would value it less. Individuals who were missing data on this variable were reassigned to uncertain (487).
5 For those missing information in round 4/2 values were assigned based on values from round 3/1 (1 value), and round 5/3 (115 values).

Both of these questions were asked on a self-administered questionnaire which may explain the high level of missing information.
Factors included in the conceptual framework that are not available in the data include availability of charity care, eligibility for public programs, generosity of benefits, contact with insurance agents, intelligence level, state regulations or presence of high risk pool, and price.