Evaluation of Changes in US Health Insurance Coverage for Individuals With Criminal Legal Involvement in Medicaid Expansion and Nonexpansion States, 2010 to 2017

This cross-sectional study compares changes in health insurance coverage from 2010 to 2017 for low-income US adults with criminal legal involvement in states that did and did not adopt the Medicaid expansion provision of the Affordable Care Act.


eMethods. Expanded Discussion of Methods
We used restricted data from the National Survey on Drug Use and Health (NSDUH) from 2010-2017. Sponsored by the Substance Abuse and Mental Health Services Administration (SAMHSA), the NSDUH is a cross-sectional, nationally representative, annual survey of non-institutionalized residents of the United States age 12 years and older. The NSDUH interviews around 65,000 individuals a year using a combination of telephone, in-person, and computer-assisted survey techniques. Across all years included in this study, NSDUH weighted screening response rates are reported as greater than 75% and weighted interview response rates greater than 67%. For this study we used restricted-use NSDUH data to utilize state-level variables that are not available in the public-use data file. SAMHSA provides access to a restricted-use data file to approved researchers through the US Census' Federal Statistical Research Data Center system.

Study Sample
To target our analysis on the ACA's Medicaid expansion, we limited our sample to adults aged 18-64, with household income ≤138% the federal poverty level. We then further limited our cohort to individuals with past-year criminal legal involvement. We defined presence of past year criminal legal involvement based on responses to questions regarding past year arrest (excluding minor traffic violations) and being on community correctional supervision (parole or probation) in the past year.

Primary exposure of interest: Medicaid Expansion
We defined states that expanded Medicaid between 2010-2017 using data from the Kaiser Family Foundation. To account for differences in when states implemented Medicaid expansion, we defined exposure to Medicaid expansion by state of residence and quarter-year of survey administration. Residence in a state after Medicaid expansion was defined as living in a state for any quarter year after Medicaid expansion was implemented. See supplementary eTable 1 for state-by-state quarter year Medicaid expansion exposure definition used in our analysis.

Primary outcomes of interest: Insurance coverage
We categorized respondents as insured if, at the time of the survey, they were enrolled in a private or public health insurance plan. We subsequently categorized their insurance coverage as being Medicaid coverage, private insurance, or other (e.g.,. Tricare, Veterans Health Administration, or Medicare).

Covariates
We adjusted for respondent age, gender, race and ethnicity, marital status, and employment status in our models. Age was included as a continuous variable. Gender categories were based on selfreport, and individuals were categorized as either male or female. Self-reported race and ethnicity categories included non-Hispanic Black, Hispanic, non-Hispanic White, or Other. The Other category includes respondents categorized as non-Hispanic Native American/Alaskan Native, non-Hispanic Native Hawaiian/Pacific Islander, and non-Hispanic Asian, as well as respondents who reported more than one race. Sample size precluded us from further refining "Other" race and ethnicity. We defined employment status as unemployed, employed part-time, employed full-time, or not in labor force.

Analysis
We estimated weighted proportions of sociodemographic characteristics in those who resided in Medicaid expansion and non-expansion states. Next, to estimate the impact of Medicaid expansion in individuals residing in a state that expanded Medicaid, we used difference-in-differences (DiD) methods frequently used to analyze the Affordable Care Act and its Medicaid expansion provision. DiD is a quasiexperimental approach to estimate the impact of an intervention by comparing outcomes before and after the intervention between groups that are exposed and unexposed to an intervention. This method can account for secular trends, as well as unobservable differences between the exposed and unexposed groups. This method relies on the assumption that absent the intervention, the groups exposed and unexposed to the intervention would have parallel trends in the outcome. We tested this assumption by visual confirmation and by testing trends in periods pre-ACA for rates of insurance coverage between expansion and non-expansion states.
To perform the difference-in-differences estimation, we used multivariable linear regression models which included an interaction term between a variable for quarter-years before and after Medicaid expansion and a variable for Medicaid expansion status. We performed both an unadjusted analysis. Our unadjusted models included variables indicating pre/post-Medicaid expansion and whether a state expanded Medicaid and an interaction of these variables. The adjusted analyses utilized the interaction term, but also controlled for state and year fixed effects and covariates for age, gender, race/ethnicity, marital status, and employment status. Standard errors were clustered at the state level. We used predictive margins to generate adjusted estimates of insurance coverage and a DiD estimate.
All analyses used sample weights provided by NSDUH to account for the survey's complex sample design and were performed in Stata 15 (Stata Corp). We conducted all our analyses of restricted NSDUH data in a Federal Statistical Restricted Data Center managed by the U.S. Census Bureau. All results were cleared for disclosure by SAMHSA, but the agency did not have a role in study design, analysis, or interpretation of results. Our analysis did not require institutional review board approval because it falls under the Yale University policy for research using de-identified, publicly available data sets. A STROBE checklist for the reporting of cross-sectional studies was completed and is included in supplementary materials.

Expanded discussion of limitations
All outcomes were measured by self-report and cannot be verified through claims or administrative data. They are therefore subject to recall or social desirability bias although self-reported data on criminal justice involvement has been shown to be a valid measure in previous studies. The NSDUH does not survey adults who are institutionalized or unhoused, which may underestimate the proportion of the population with past-year criminal legal involvement. Response rates of the NSDUH as reported above may lead to selection bias, although its use for national estimates, despite this limitation, is widely accepted. DiD methodology cannot account for the possible effect of differences between expansion and non-expansion states that would have coincided temporally with implementation of the ACA and affected our outcome of interest. Despite this limitation DiD methods have been used widely to account for the impact of Medicaid expansion on a variety of outcomes.