Comparison of Utilization, Costs, and Quality of Medicaid vs Subsidized Private Health Insurance for Low-Income Adults

Key Points Question How do utilization, cost, and quality compare between public (Medicaid) and private (Marketplace) health insurance? Findings This cross-sectional study of 8182 participants used a propensity score–matched sample narrowed to 5 percentage points above and below the federal poverty level threshold that separates Medicaid and Marketplace eligibility (138%). Marketplace coverage was associated with fewer emergency department visits and more office visits than Medicaid, total costs were 83% higher in Marketplace coverage owing to much higher prices, and out-of-pocket spending was 10 times higher in Marketplace coverage; results for quality of care were mixed. Meaning This study found that Medicaid and Marketplace coverage differ in important ways: more emergency department visits in Medicaid may reflect impaired access to outpatient care or lower copayments; Marketplace coverage was more costly owing to higher prices and also had higher cost sharing for consumers.

The CO APCD includes comprehensive data on enrollment; utilization and payments for outpatient, inpatient, and prescription drug claims; and some information on beneficiary demographics. We also obtained time-stamped family income data, as a percentage of FPL. The Medicaid sample was limited to childless adults and parents ages 19-64 who qualified for Medicaid under the ACA expansion. Our sample excluded women whose diagnosis codes indicates a pregnancy during the enrollment year. We also omitted from our sample those with incomes less than 75% of FPL, since the state's disability-related pathway to Medicaid extends through 74% of FPL. Above that income threshold, adults in the state (unless they are pregnant) are generally only eligible for subsidized coverage via the ACA's provisionsi.e. Medicaid expansion or Marketplace coverage. Medicaid data were directly matched to the CO APCD using a common identification number, whereas Marketplace data required a probabilistic match.
We tested two approaches, either limiting the sample only to exact Marketplace matches by name and birthdate, or using a fuzzy match algorithm developed by the state Exchange in which we kept all observations with a cosine similarity of 0.6 or greater, which roughly corresponds to exact birthdate match and similar names, or matching names with a birthdate off by a single © 2021 Allen H et al. JAMA Network Open.
digit. Our primary model included these inexact but probabilistic matches, but results were quite similar when we limited the sample to exact matches.
Income was based on the first reported value within each calendar year. While many individuals only reported income at the time of their initial enrollment for coverage, some individuals in Medicaid reported monthly changes in income during the course of the year. For Marketplace enrollees, there is only a single FPL measure per year; we attributed this measure to their first month of Marketplace enrollment, consistent with an intent-to-treat analysis of eligibility. We added 5 percentage-points of FPL to income for all Medicaid enrollees in our sample, given the ACA's statutory income disregard equal to this amount. This placed the key eligibility transition point at 138% of FPL.

Propensity Score Matching
Our propensity score model was based on a logistic regression of the following form, Equation (1) The predicted value for this model was then used as the propensity score for having an income between 139-143% of FPL, and we implemented a 1:1 nearest neighbor match with the 134-138% FPL sample. We used a caliper of 0. from the sample, and produced very similar overall results. Descriptive statistics and standardized mean differences were then calculated for each covariate, with a threshold ≤ 0.1 used to indicate adequate balance.

Main Regression Models
Our regression analysis used Generalized Linear Models, using the distributions and link functions described in the main methods of the paper. The equation for coverage outcomes was as follows: Equation (2) where MonthsCoverageit is the months of coverage in the prior year with either Medicaid or Marketplace insurance, i indexes the individual and t the year. Xit is a vector of demographics (sex and age group), Elixhauser is the Elixhauser comorbidity score, and ChronicConditions is a vector of indicator variables for the five most chronic conditions in our sample. µ is a vector of year fixed effects (for 2014 vs. 2015), and Ω is a vector of area fixed effects at the level of the 3digit zip code. We created a residual category for the 2.84% of our sample that did not have a three-digit zip code or resided in a zip-code with fewer than 1,000 people.
The coefficient of interest is 1, which measures the outcome difference associated with having an initial income above 138% of FPL, making that person eligible for Marketplace coverage, rather than Medicaid.
The regression used Huber-White robust standard errors clustered at the level of the individual, to account for repeated measures for those appearing separately in both years of the dataset.
The equation for utilization, cost, and quality outcomes was as follows: © 2021 Allen H et al. JAMA Network Open.
The only differences between Equations 2 and 3 were the outcome variable, and the addition of Months of Coverage as an independent variable.
For secondary quality outcomes, several of which had much smaller condition-specific samples, we used a more parsimonious version of Equation 3 to reduce the risk of overfitting and/or dropping observations due to perfectly predicting outcomes. This equation replaced the three digit zip code fixed effects with an indicator for urban vs. rural residence, and only used the overall Elixhauser score but not specific condition indicators: All coefficients were then converted into adjusted outcome estimates using the "margins" command in Stata for Marketplace-Eligible, which provides separate sample-wide marginal outcomes for the Marketplace-Eligible population and the Medicaid-Eligible population, using each observations' actual covariates and the coefficients from the relevant regression model described above, except for the coverage measures, which used the margins "at means" option (using the means of the covariates)see Table 2 footnotes for details.

Medicaid Price-Normalized Cost
We analyzed Medicaid price-normalized costs to facilitate comparisons between the overall health care utilization across coverage types, after removing the impact of differential prices for health care services in the difference insurance plans. To create this outcome, we calculated the mean costs of CPT procedure codes found in claims in the Medicaid database, and © 2021 Allen H et al. JAMA Network Open.
then applied the mean Medicaid cost per procedure to all claims (whether they were Medicaid or Marketplace) to derive the price-normalized cost for each individual in our sample.
The initial average costs of CPT procedure codes were calculated from claims with only a single unique CPT code, which accounted for 8,157 CPT codes. For the remaining codes found only in multiple-CPT claims, we iteratively calculated the average costs of the unknown CPT codes by subtracting total claim costs by the costs calculated from known codes. For the claims with "j" total CPT codes with "j-1" CPT codes with a calculated cost, where "j" is an integer, the remainder provides an estimate of the unknown CPT code's cost. These averages were then stored and used in the next iteration to calculate remaining prices. This process had reached completion with a total of six iterations. This process allowed us to calculate prices for 10,051 CPT codes, leaving 922 unmatched codes, which represented only 0.07% of claims, for which we set the effective price to $0.
These mean Medicaid prices were then assigned to all Medicaid ad Marketplace claims via CPT codes in order to calculate the price-normalized costs per enrollee. Negative yearly enrollee costs (which resulted from the above iteration for multiple CPT codes) affected 0.05% of the sample and were truncated to zero in our final analysis.

Clinical Quality Measures
We created binary indicators for the following measures of high-value care: • Influenza Vaccination -The denominator included all adults aged 19-64, based on current National Quality Forum guidelines, 2 and the numerator included anyone with a claim for influenza immunization during the calendar year.
• Chlamydia Screening -The denominator included all non-pregnant women aged 19-24, based on current HEDIS guidelines (note that the claims data do not allow us to distinguish between women who are and are not sexually active), 3 and the numerator included anyone with a claim for a chlamydia test during the calendar year.

• Beta Blockers in Patients with Coronary Artery Disease -
The denominator included all adults with coronary artery disease, and the numerator included anyone with a prescription claim for beta blockers during the calendar year. 4 While the benefit of indefinite therapy with beta blockers in uncomplicated CAD remains unclear and an area of clinical investigation, this remains a commonly-used quality measure that can be evaluated with claims data. • Annual Eye Exam in Patients with Diabetes -The denominator included all adults with diabetes, and the numerator included anyone with a claim for eye exam during the measurement year. 7 We also created binary indicators for the following measures of low-value care, identified in prior research (Barnett, et al. 2017) 8 : • Advanced Imaging for Uncomplicated Back Pain < 6 Weeks Duration -The denominator included all individuals with a claim including a diagnosis code for back pain, excluding those with fever, weight loss, neurologic symptoms, cancer, fracture, myelopathy, prior back surgery, or a prior visit at least two months earlier for back pain. 8 The numerator included a claim for CT or MRI of the back within the month or month following the initial visit.
• Advanced Imaging for Uncomplicated Headache -The denominator included all individuals with a claim including a diagnosis code for headache, excluding those with HIV, pregnancy, neurologic symptoms, cancer, fracture, or epilepsy. 8 The numerator included a claim for CT or MRI of the head within the month or month following the initial visit.
• Narcotic Prescription for Headache -The denominator included all individuals with a claim including a diagnosis code for headache, excluding those with HIV, pregnancy, neurologic symptoms, cancer, fracture, or epilepsy. 8 The numerator included a prescription claim for any narcotics for the same month of the initial visit. respiratory illness, cancer, HIV, or sexually transmitted infections. 8 The numerator included a prescription claim for antibiotics for the same month of the intial visit.

Modified NYU Emergency Department Visit Algorithm
We have classified emergency department visits according to the updated Emergency • Non-emergent -"Immediate medical care was not required within 12 hours." • Emergent but primary care treatable -"Treatment was required within 12 hours, but care could have been provided effectively and safely in a primary care setting. The complaint did not require continuous observation, and no procedures were performed or resources used that are not available in a primary care setting." • Emergent but preventable -"Emergency department care was required based on the complaint or procedures performed/resources used, but the emergent nature of the condition was potentially preventable/avoidable if timely and effective ambulatory care had been received during the episode of illness." • Emergent and not preventable -"Emergency department care was required and ambulatory care treatment could not have prevented the condition."   (overall score and top five  conditions), year, and three-digit zip code; utilization, cost, and quality outcomes also adjusted for total months of Medicaid or Marketplace coverage. Coverage, utilization, and quality outcomes were analyzed using a generalized linear model (GLM) with a negative binomial distribution. Costs outcomes were analyzed using a GLM with a gamma distribution and log link, with outcomes in 2015-inflation adjusted terms. All regression results were converted to adjusted means based on the observed distribution of covariates using the margins command in Stata, other than for coverage outcomes. Coverage outcomes were assessed using margins at covariate means, due to totaling errors with the margins command at the observed distribution (i.e. total months of coverage < months Medicaid

Notes:
Sample contains propensity-score matched adults ages 19-64, with incomes between 134% and 143% of FPL, and excludes all individuals whose first recorded claim was an ED or hospital inpatient claim during their first month of coverage in the calendar year (N=8,000). Data are from the Colorado All Payer Claims Database, linked to income data from Medicaid and Marketplace eligibility files. Models adjusted for age, sex, Elixhauser comorbidity index (overall score and top five conditions), year, three-digit zip code, and total months of Medicaid or Marketplace coverage. Utilization and quality outcomes were analyzed using a generalized linear model (GLM) with a negative binomial distribution. Costs outcomes were analyzed using a GLM with a gamma distribution and log link, with outcomes in 2015-inflation adjusted terms. All regression results were converted to adjusted means based on the observed distribution of covariates using the margins command in Stata. 95% CI = "95% Confidence Interval" § Out-of-pocket costs are the charged amount; the dataset does not indicate whether patients paid the required amount. † This outcome was calculated using mean Medicaid price per service provided, to provide an aggregate measure of health care resources consumed but using the same price regardless of the person's type of health insurance. See Appendix methods for further details.

Notes:
Data are from the Colorado All Payer Claims Database, linked to income data from Medicaid and Marketplace eligibility files. Sample contains propensity-score matched adults ages 19-64, with incomes between 129% and 148% of FPL (N=17,282). Models adjusted for age, sex, Elixhauser comorbidity index, urban vs. rural residence, and total months of Medicaid or Marketplace coverage. All outcomes were analyzed using a generalized linear model (GLM) with a Binomial distribution and logistic link. All regression results were converted to adjusted means based on the observed distribution of covariates using the margins command in Stata. "95% CI = "95% Confidence Interval * These p-values were adjusted according to the family-wise error rate, using the Westfall and Young (1993) free step-down resampling approach, to account for multiple outcomes within each category.