Data

This study used publicly available data that was waived from review by the Institutional Review Board at Tulane University. Data analysis was based on repeated cross-sectional national population level data from the 2012–2018 7-year American Community Survey (ACS) Public Use Microdata Sample (PUMS) files. ACS is a very large survey, in terms of number of respondents who complete the survey, making it possible to obtain estimations for narrowly defined subpopulations by FPL [47]. PUMS are a sample of actual responses from the ACS and include most population and housing characteristics. This data source, with the flexibility to prepare customized tabulations, can be used for detailed research and analysis. In the PUMS, files have been de-identified to protect the confidentiality of all individuals and households. Further, this survey includes information about socio-economic, demographics, and other characteristics. The PUMS file contains over 3.5 million respondents each year with an average response rate of over 97 %. The differences in this cross-sectional survey were constant over time, which makes the estimation of trends in insurance coverage comparable [48].

Analytic Strategy

This study employed differences-in-differences-in-differences (triple-D) design by comparing the outcome (health insurance coverage), before and after the intervention (implementation of Medicaid expansion) for the treatment group (individuals under 138 % FPL) and control group. This quasi-experimental design has been widely used to estimate the effects of a specific intervention or treatment by comparing the changes in outcomes over time between a population that is enrolled in a policy/program (i.e. the intervention group) and a population that is not (i.e. the control group) [49, 50]. The approach removes biases in post-policy period comparisons between the treatment and control group that could be the result from permanent differences between those groups, as well as biases from comparisons over time in the treatment group that could be the result of trends due to other causes of the outcome.

In this study, we did not include data prior to 2012 and excluded individuals less than 26 years old, as a provision of the ACA enables young adults to remain on their parents’ insurance plans until 26 years old. Without capturing effects of that provision, which have been well studied in previous research, our results excluded this potential confounding impact [51,52,53]. Additionally, a 0 to 1 intensity was used to demonstrate the effect of the dynamic enrollment of adoption of the policy for each state in order to generate more precise estimates.

The first regression was performed with the entire sample, with the treatment group being those under 138 % FPL and the control group being those with FPL above 138 %. The second regression included those under 138 % FPL as the treatment group and those above 400 % FPL as the control group. Those with an FPL range from 138 to 400 % were excluded in order to eliminate the ACA market exchange subsidy influence on the insurance coverage rate. The equation is as follows:

Variables and Measures

Uninsured_isy was the study outcome. It was derived by the survey question “whether the individual has any insurance coverage” and was transformed into a dichotomous variable where 1 equaled no health insurance and 0 equaled having any insurance.

Expanded_isy, variable of interest, was the interaction term that measured the individual level of Medicaid expansion status. This variable was calculated by Medicaid_Expansion_sy times Poverty_isy. In the first regression, the coefficient of this variable reflected the impact Medicaid expansion had on insurance coverage between persons below 138 % FPL (treatment group) and above 138 % FPL (control group) in year \(i\)and state \(s\). In the second regression, the control group were the individuals above 400 % FPL.

Medicaid_Expansion_sy was the Medicaid expansion policy indicator variable, defined as a varying intensity variable from 0 to 1. When a state expanded Medicaid, this variable equaled 1, otherwise, it equaled 0. For states that adopted the policy at the beginning of the year, time before this certain year was the pre-policy period and from this year to 2018 was the post-policy period. Coverage under Medicaid expansion became effective January 1, 2014 in all states that adopted the policy except the following: Michigan (4/1/2014), New Hampshire (8/15/2014), Pennsylvania (1/1/2015), Indiana (2/1/2015), Alaska (9/1/2015), Montana (1/1/2016), Louisiana (7/1/2016) [54]. For the above states, the indicator variable was coded as the portion of the year and then equaled to 1 in the following post-policy period. Specifically, the treatment for Michigan in 2014 was 275/365; the treatment for New Hampshire in 2014 was 138/365; the treatment for Indiana in 2015 was 334/365; the treatment for Alaska in 2015 was 122/365; and the treatment for Louisiana in 2016 was 184/365. Substantive Medicaid expansion policies prior to the state’s official implementation date of Medicaid expansion were not considered in the study. However, the prior adoption of Medicaid expansions in some states (IN, ME, TN, and WI) were quite limited with capped or closed enrollment. In some states a mild form of Medicaid expansion was adopted prior to the enactment of the ACA for both parents and childless adults (DE, DC, MA, NY, VT), which turned out to be an equivalent of the ACA expansion.

Poverty_isy, was defined as a categorical variable using FPL threshold. FLP was calculated using poverty guidelines (one of the federal poverty measures) and the number of persons living in a household [55]. The poverty guidelines are updated each year by the US Department of Health and Human Services, for use for administrative purposes (e.g., determining financial eligibility for federal Medicaid programs). In the first regression, 138 % FPL was used as the cut-off point, where 1 was the population below 138 % FPL and 0 represented the population above 138 % FPL (control group). In the second regression, 1 represented those below 138 % FPL and 0 represented those above 400 % FPL (control group).

\({\gamma }_{sy}\) was the interaction term for year and state.\({\delta }_{is}\) was the interaction term for whether the individual was under 138 % FPL in a specific state. \({\eta }_{iy}\) was the interaction term for whether the individual was under 138 % FPL in a specific year (2012–2018).\({State}_{iy}\) indicated the residence of state for the individuals in a given year. \({Year}_{is}\) represented the calendar year for the individuals in a state.

\({X}_{isy}\)was a series of individual demographic and socioeconomic covariates based on the Andersen model. Age, sex, race/ethnicity, marital status, education attainment, and employment status were adjusted in each model and these variables have been widely used in the literature to examine health care access and utilization [4, 5, 56, 57]. Age was measured as a categorical variable: 26–34, 35–44, 45–54 and 55–64. Sex was also a categorical variable, where 1 was male and 0 was female. Survey participants were asked to self-report their race (White, Black/African American, American Indian/Alaska Native, Asian, Native Hawaiian/other Pacific islander, or multiple race) and ethnicity (Hispanic/Latino: yes/no) separately. A 4-category race/ethnicity variable was coded as follows: participants reporting Hispanic/Latino ethnicity were considered Hispanic, regardless of race; all non-Hispanics were categorized as White, Black, or other races. Thus, Whites were coded as 0, Blacks, Hispanics, and Others were coded as 1, 2 and 3, respectively. Marital status was self-reported and categorized as married, widowed, divorced, and separated, and never married. For the model, we coded married as 0 and combined widowed, divorced, and separated into one group, and coded this group as (1) Self-identified education level was grouped by less than high school coded as 0, high school graduate, GED, or alternative coded as 1, and bachelor’s degree or higher coded as (2) Current employment status was categorized as unemployed coded as 0, employed coded as 1, and not in the labor force coded as 2.

Robustness checks were performed to test whether Medicaid expansion had an effect [58, 59]. To test this, we verified whether the control group (individuals with < 138 % FPL in non-expansion states) served as a good comparison for the treatment group (individuals with < 138 % FPL in states that expanded Medicaid). Therefore, we tested if the trends in the two groups were parallel before the policy was implemented. This assumption was tested indirectly by employing an event study model that interacted the treatment variables with the full set of fixed year effects. The regression took the form as Eq. 2 below. To satisfy the parallel trends assumption, no statistical significance on the first coefficient was expected, which suggested no change associated with Medicaid expansion between the period at least 2 years and 1 year prior to expansion. Statistical significance in later years’ coefficients indicated the expansion effected the treatment group with the given poverty level. In addition, we also examined the association of Medicaid expansion with a placebo outcome unrelated to insurance status that would not be affected by the policy. The placebo outcome was set as the probability of getting higher education. The placebo outcome check was performed in the same format as Eq. 1.

Similar to the main estimating equation (Eq. 1), only the Expanded_isy treatment variable was changed from the main model into a set of time dummies for individuals given the year and state (\(Pre2, Same0, Post1,Post2\), time \({Poverty}_{isy}\), respectively) in Eq. 2. The \(Pre2\) term was an indicator variable set as 1 and represented at least 2 years before a state expanded Medicaid and 0 otherwise. The second \(Same0\)was 1 when in the year a state expanded Medicaid and 0 otherwise. Same algorithm followed for the \(Post1\) and \(Post2.\) The omitted category \(Pre1\) was the year immediately before expansion. Thus, the coefficients on each of these variables gave the change for 2-year before expansion and 2-year post expansion. Other variables in Eq. 2 represented the same state/year fixed effects and control vectors as Eq. 1. Standard errors were clustered at the state level to account for state-level differences and serial autocorrelation. The 7-year data was a multi-year combination of the 1-year PUMS file with appropriate adjustments to the weights and inflation adjustment factors. The significance level for tests was set as 0.05. The data was weighted by the ACS survey weight and analyses were performed using SAS 9.4 and Stata 12.0.