Data for this analysis came from the Household Component (HC) of the 2017 Medical Expenditure Panel Survey (MEPS), a nationally representative survey of the civilian noninstitutionalized population of the United States, that was publicly available for download (https://www.meps.ahrq.gov/mepsweb/data_stats/download_data_files_detail.jsp?cboPufNumber=HC-201). The analysis included all the participants who had a positive sampling weight (PERWT17F). For descriptive statistics, we further restricted the sample to those under 65 years old at the time of being surveyed since those over 65 are eligible for government Medicare insurance.
The primary outcome was insurance coverage status, a categorical variable with five levels: (1) publicly insured: having public insurance for every of the 12 months of 2017; (2) privately insured: having private insurance for every of the 12 months of 2017; (3) dually insured: having public or private insurance for every of the 12 months of 2017 AND for at least one of these months the participant had both public and private insurance coverage, including the scenario that for one month the participant had private insurance and for another month the participant had public insurance; (4) partially insured: having either public or private insurance for at least one month but less than 12 months of 2017; and (5) uninsured: having no insurance coverage for every of the 12 months of 2017.
We used three sets of variables to construct the insurance coverage status variable – any insurance in a month (INSJA17X – INSDE17X), any public insurance in a month (PUBJA17X – PUBDE17X), and private insurance in a month (PRIJA17 – PRIDE17). Public insurance referred to TRICARE, Medicare, Medicaid or State Children’s Health Insurance Program (SCHIP), or other public hospital/physician programs. Private insurance included union or employer group insurance, non-group insurance, other group insurance, and private insurance through federally facilitated, state-based or state partnership marketplace or exchange. We considered a subject as having any insurance if covered by any of these insurance sources above.
Operationalization of the exposures used for this study was based on the vulnerability framework promulgated by Shi and Stevens (11). According to Shi and Stevens, vulnerability refers to the likelihood of experiencing poor health and is determined by a convergence of predisposing, enabling, and need characteristics. In their access to care framework (12-13), Aday and Andersen described predisposing characteristics as those that describe the propensity of individuals to use health services including basic demographic characteristics (e.g., age, sex, and family size), social structure variables (e.g., race and ethnicity, education, employment status, and occupation), and beliefs (e.g., general beliefs and attitudes about the value of health services); enabling characteristics as the means individuals have available to them for the use of services including resources specific to individuals and families (e.g., income and insurance coverage) and attributes of the community or region in which an individual lives; and need characteristics as health status or illness, which is the most important cause of health services use. Thus, individuals are most vulnerable if they experience a convergence of predisposing, enabling, and need attributes of risk.
Based on this framework, we identified measures within MEPS that denote predisposing, enabling, and need attributes of risk. We combined three of these variables into a new vulnerability measure that reflects the convergence of predisposing, enabling, and need attributes of risk. These were race (predisposing dimension), income (enabling dimension), and self-perceived health status (need dimension), and are among the most, although not the only, significant indicators of vulnerability. It is possible to create a measure incorporating other vulnerable attributes (e.g., the more objective chronic illness measure for health status, the behavioral risks such as smoking, alcohol, and drug abuse for predisposing factor). However, the trade-off is that the resulting sample size for some subgroups would be too small for comparative analysis (e.g., chronic illness is likely to be more concentrated among the elderly population). To avoid small subgroup sample size, we further re-coded these variables into dichotomous categories so that our final vulnerability measure was limited to eight categories: (1) the minority-low-income-bad health group (the most vulnerable group with vulnerable attributes in all three dimensions), (2) the minority-low-income-good health group, (3) the white-low-income-bad health group (3) the minority high-income-good health group, (4) the white-low-income-good health group, (5) the minority-high-income-bad health group, (6) the minority-high-income-good health group, (7) the white-high-income-bad health group, and (8) the white-high-income-good health group (this is the least vulnerable group with none of the three vulnerable attributes measured).
Minority included all non-white racial and ethnic groups including blacks, Hispanics, Asians, American Indians, and others. Measure of income was based on the variable family income as a percentage of poverty within MEPS (POVCAT). Family income comprised annual earnings from wages, salaries, bonuses, tips, commissions; business and farm gains and losses; unemployment and workman’s compensation; interest and dividends; alimony, child support, and other private cash transfers; private pensions, IRA withdrawals, social security, and veterans payments; supplemental security income and cash welfare payments from public assistance, aid to families with dependent children, and aid to dependent children; gains and losses from estates, trusts, partnerships, S corporations, rent, and royalties; and a small amount of ‘‘other’’ income. Family income excluded tax refunds and capital gains. Person-level income totals were then summed over family members to yield the family-level total. POVCAT was constructed by dividing family income by the applicable poverty line (based on family size and composition), with the resulting percentages grouped into five categories: negative or poor, near poor, low income, middle income, and high income. For the purpose of this study, we grouped ‘negative or poor, near poor, low income’ as ‘low income’ and ‘’middle income, high income’ as ‘high income.’ For health, ‘good health’ included excellent, very good, and good health, whereas ‘bad health’ included fair and poor health). Self-rated health has strong predictive validity for mortality, morbidity, and mental health, independent of other physiological, behavioral, and psychosocial risk factors (14-17).
We selected additional measures of vulnerability as the other covariates, including age (≤ 17 years; 18-64 years; 65-85 years), sex (male; female), education (college; General Educational Development [GED] or high school; none), employment status (employed; unemployed), census region (northeast; Midwest; west; south), perceived mental health status (excellent; very good; good; fair; poor), and need Activities of Daily Living (ADL) help (yes; no).
We summarized the insurance coverage status by vulnerability groups as percentages (95% confidence intervals) (PROC SURVEYFREQ). We constructed three logistic regression models (PROC SURVEYLOGISTIC) to estimate the associations between insurance coverage status and vulnerability measures, comparing insured with uninsured or partially insured, partially insured only, and uninsured only, respectively.
We excluded observations with missing outcome (insurance coverage status) or exposures (race, family income, self-perceived health status). We treated missing values for the other covariates as a separate category for the multivariable logistic regressions. We performed data management and analyses using SAS 9.4 (SAS Institute, Cary, NC, USA), accounting for complex survey design, i.e., strata, cluster, and sampling weights. The designated statistical significance level was two-sided p-value < 0.05.