Association of Medicaid Expansion With Medicaid Enrollment and Health Care Use Among Older Adults With Low Income and Chronic Condition Limitations

This cross-sectional study uses data from the National Health Interview Survey to assess the association of Medicaid expansion to working-age adults with Medicaid enrollment and health care use among older adults with low income and chronic condition limitations.


III. Identifying Individuals Likely Eligible for Medicaid
We restricted the NHIS to adults ages 65 and older who together with their spouse had income less than or equal to 100% federal poverty level (FPL). Adults ages 65 and older with income at or below 100% FPL and assets below $7,730 (single) and $11,600 (couple) are eligible for partial Medicaid through the Qualified Medicare Beneficiary (QMB) Medicare Savings Program, which pays Medicare premiums and cost sharing for Parts A and B. Our inclusion criteria also identified adults eligible for full Medicaid through the Aged, Blind, Disabled (ABD) pathway which also covers services that Medicare does not. In 2015, the ABD income eligibility limit ranged from 53% FPL for single adults in Connecticut to 100% FPL in 17 states; in most states, the ABD asset eligibility limit was $2000 for singles ($3000 for couples). ii There are two main barriers to identifying these individuals in the NHIS data. First, in the NHIS only earned income is reported at the individual level while unearned income is reported at the family level. Since some seniors lived with family members aside from their spouse, we could not attribute all familial unearned income to the couple and instead must do an accounting exercise. Second, for 39% of the seniors in the years of our sample, exact income was not reported. Instead, NHIS provided five different imputed income values. We address our approach to these challenges below.

a. Accounting for family unearned income when computing total couple income
The NHIS reports total individual earned income (ERNYR) from the past year and total earned and unearned combined family income (FAMINCI). To calculate total earned and unearned spousal income combined, we followed the following procedure. For all individuals (i) in family (F), we summed the total individual earned income within a family to get gross family earned income. We then subtracted gross family earned income from total family income. The residual amount was assumed to be unearned income. We divided this amount among all individuals in the family who reported having some form of income. If there was a positive estimate of unearned income, but no individuals report having income, we divided the unearned income among all adults. If unearned income was estimated to be less than zero, we assumed there had been some error in reporting and set unearned income equal to zero so that individuals were evaluated only based on their self-reported earned employment income. This reduced our sample relative to what we would have had if we incorporated negative unearned income. We summed the individual reported employment earnings with each individual's share of the family-level unearned income to get total individual income.
If an individual is single, we used this total individual income as the evaluation income to compare to the federal poverty level. If individuals were married, the evaluation income was the sum of both spouses. In the 2010-17 NHIS, 46% of adults ages 65 and older resided in a family that had more than just the respondent and his or her spouse.

b. Multiple Imputation: Addressing Income Nonresponse
The other major challenge is the high rate of nonresponse to income questions among NHIS respondents. For the years 2010-2017, exact income was not reported for 39 percent of adults aged 65 and older. For these adults, the National Center for Health Statistics (NCHS) used imputation methods to generate five different imputed values for each variable. iii We followed NCHS guidelines to use these five different imputed values to conduct our analyses. In the exhibits, we reported the sample size for one of the five imputations. The sample size differs from imputation to imputation because respondents are only included in the sample when imputed income is at or below 100% FPL (i.e., they are excluded when imputed income is above 100% FPL). We report sample size from the first imputation. The multiple imputation procedure is outlined below.
We first used each different imputed value to separately calculate the evaluation income as defined in the previous section. For each of these five values, we then created a binary variable for being above or below the income threshold for sample inclusion. Each individual will then have five separate variables indicating sample inclusion which may not agree. We then used the multiple imputation package MI to run our analyses. We used the command MI SVYSET to register the complex structure of our data. For estimating sample means, we then used the command MI ESTIMATE, ESAMPVARYOK: SVY MEAN. To test the significance of differences in sample means, we used MI ESTIMATE TTESTTRANSFORM. To derive regression estimates, we used the MI ESTIMATE: REGRESS command along with NHIS survey sampling weights and estimated heteroskedasticity-robust standard errors clustered by state using the CLUSTER() option. The package runs a separate regression over each of the five imputation wave samples then combines point estimates and standard errors.

IV. Alternative Specifications, Sample Definitions, and Methods to Conduct Inference
In eFigure 1 we present a flow diagram of the sample construction.
In eFigures 2-9, we present results of an event study specification which allows us to both formally test the parallel trends assumption underlying the difference-in-differences analysis and understand how the effects evolve over time. In this model, we replace the term "POST x EXPANSION" in Equation (1) with a full set of indicator variables representing each year, relative to the expansion year, interacted with the EXPANSION indicator variable; non-expansion states are coded as zero in all years. We omit the indicator for the year prior to expansion in the state.
The Supplementary Figures (eFigures 2-9) plot the coefficients of these interaction terms and their confidence intervals and show the absence of differential pre-expansion trends between expansion and non-expansion state respondents for all outcomes. We formally tested this with an F-test of joint significance for the pre-period coefficients. In every case, the p-value for this F-test exceeded 0.05. These event studies show that effects generally arose in the third year following expansion.
In eTable 1, we present the results of a balance test showing that the likelihood that a respondent reports a limitation from a chronic condition did not vary with ACA Medicaid expansion.
In eTable 2, we examine the robustness of the results to including potentially endogenous covariates and different definitions of the included sample. That is, we included controls for income, employment status, number of chronic conditions for which the respondent is experiencing limitations, number of limitations from activities of daily living (ADLs), family size, marital status, and the state unemployment rate. Estimates of the association of the Medicaid expansion with insurance coverage and utilization were robust to the inclusion of these variables.
Since Medicaid eligibility is determined after several adjustments are made to income, we first expanded the sample to include respondents with total income at or below 150% FPL, who were likely to be income-eligible for either QMB or full Medicaid. With the restricted-use NHIS data, we were able to examine effects for those who are income-eligible for full or ABD Medicaid in their state. Finally, it is possible that the welcome mat affected new enrollment in both Medicare and Medicaid, so we showed that our DD estimates are robust to including respondents not enrolled in Medicare. Our results are largely robust to these different sample definitions (eTable 2). The signs and magnitudes are similar to our baseline models, but the estimates on measures of office visits are less precise.
In eTable 3, we tested the robustness of the estimates to changes in the states included in the sample. In column (1) (2017). vii The signs and magnitudes of the coefficient estimates are robust to these exclusions, though with a smaller sample size, the effect on Medicaid enrollment was estimated less precisely in the smallest sample and the effects on office visits were also estimated less precisely.
In eTable 4, we examined the effect of the Medicaid expansion on two other types of outpatient care: receiving advice or test results by phone in the past two weeks and home health care visits from health professionals in the past two weeks. Among respondents with a limitation from a chronic condition, there was no increase in the likelihood of receiving either type of care (-

eTable 1. Changes in Likelihood Respondent Has a Limitation from a Chronic Condition Among Older Adults with Low Income in States with Medicaid Expansion (Linear Probability Model), 2010-2017 NHIS Coefficient Estimate (95% CI) N
Has a limitation from a chronic condition 1.90 (-1.38 -5.17) 21,859 Note: Coefficients were estimated using multiple imputation over five different samples that were defined based on five different imputed incomes.