Disparities in Health Care Spending and Utilization Among Black and White Medicaid Enrollees

Key Points Question Are there racial differences in health care spending and utilization for low-income individuals in the US who are covered by the Medicaid program? Findings In this cross-sectional study of 1 966 689 Black and White Medicaid enrollees in 3 states, Black enrollees used fewer services, including primary care, and generated lower spending than White enrollees, but were more likely to utilize the emergency department for avoidable reasons. Differences persisted among enrollees residing in the same zip codes who were treated by the same health care professionals. Meaning The results of this study suggest that stark differences in spending and primary care use exist between Black and White Medicaid enrollees, and additional steps to ensure equity are needed.


B. Definition of Racial and Ethnic Collective Terms
In each of the three states used for the analysis racial and ethnic data are self-reported. However, because the exact terms differ across states we define 7 mutually exclusive and completely exhaustive collective terms so that we can pool the analyses across states using common terminology. The 7 collective terms we use are Asian, Black, Hispanic, Native American or Alaska Native, Native Hawaiian or Pacific Islander, White, and Other. In eTable 1, we present the exact descriptions of the self-reported race and ethnicity data as well as the collective term we used for the analyses.

C. HHS-HCC Model
The HHS-HCC model combines demographic and diagnostic information to construct an enrollee level risk score which measures how costly that enrollee is expected to be (in terms of health care expenditures). Instead of constructing a risk score, we use the 141 conditions to generate enrollee-level controls that we add to our regressions. These conditions are hierarchical so that enrollees are ultimately assigned to only the most severe manifestation of their disease as opposed to all manifestations (e.g., an enrollee with diagnoses qualifying for both chronic and acute pancreatitis would be hierarchically categorized with just chronic pancreatitis).

D. Usual source of care attribution
Using all paid and denied medical claims in 2016 we attribute each claim to a single usual source of carethe billing provider identified by their National Provider Identifier (NPI). The billing provider field is always present for all claims across all states. For each enrollee we determine the billing provider they received the plurality of their care from based on the largest count of unique claims. In the case of ties, an enrollee was attributed to the health care professional or medical institution that they had a claim with earlier in the year. Enrollees with 0 medical claims could not be attributed to a usual source of care and were instead attributed to a "Missing" usual source of care (unique for each state).

E. Model for estimated risk
Using the full population of enrollees, we generate a concurrent estimation of total health care spending in 2016 using enrollee level characteristics and the 141 HHS-HCC indicators. Total annual health care spending Y for each enrollee i is modeled as: where are 1-year indicators for an enrollee's age, are indicators for an enrollee's gender, are indicators for an enrollee's eligibility category, are the 141 HHS-HCC condition indicators for that enrollee, and is a noise term. We also include controls for an enrollee's state and ZIP code of residence to deal with any differences that might arise due to differences in the Medicaid program across states or regions. Enrollee race is not included in this model.
Using the coefficients obtained from the above model, we create our measure of enrollee-level estimated risk ̂ from all of the predictors, except state and ZIP code.

F. Construction of enrollee categories for drug conditioning
To identify enrollees with different health conditions, we use the following logic: These health conditions are then used to condition our drug utilization measures and Medication Possession Ratios in Table 2 and eTable 6.

G. Multiple Inference Correction
To adjust for multiple inference, we use the Benjamini-Hochberg procedure to account for testing of multiple outcomes in Tables 2, 3, and 4. To adjust for multiple outcomes we examined within each group of outcomes within each table (and in Table 4, separately within the kids and adults samples), we use the Benjamini-Hochberg procedure to control the false discovery rate at the 5% significance level. We report adjusted P values in braces below the unadjusted P values in Tables 2, 3, and 4. Additional information on the Benjamini-Hochberg procedure is available elsewhere. 1,2 eFigure 3. Annual health care spending by quantile of estimated risk and race without truncation Legend: Total annual health care spending by race as a function of estimated risk. The estimation is a full-population, concurrent estimation based on an enrollee's age in years, Medicaid eligibility category, gender, and the 141 HHS-HCC indicators. + markers show the 50-quantiles, dots indicate deciles. Percentiles are based on the estimated risk in the entire population pooling across race.

eFigure 4. Annual health care spending by quantile of estimated risk and race, stratified by adults and children
Legend: Total annual health care spending by race as a function of estimated risk. The estimation is a full-population, concurrent estimation based on an enrollee's age in years, Medicaid eligibility category, gender, and the 141 HHS-HCC indicators. + markers show the 50-quantiles, dots indicate deciles. Percentiles are based on the estimated risk in the entire population pooling across race, and then the sample is divided by age where adults are 19 and older. .03] † Drug groupings are defined using different levels of the Anatomical Therapeutic Chemical Classification (ATC) system. Anti-hypertensives are defined by ATC level 2 C02, C03, C07, C08 and C09 and exclude ATC C02KX01, C03BA08, C03CA01, C07AA07 and C07AA12; asthma medications are defined by ATC level 2 R03; diabetes medications are defined by ATC level 2 A10; and statins are defined by ATC level 4 C10AA. Measures are assessed for enrollees with related diagnosed conditions (see eMethods 1 in the Supplement). For measures based on a subset of the population, sample sizes are presented under unadjusted means.