We conducted a retrospective cohort study using pharmacy data from CVS Health and Walgreens linked to Medicare claims. Medicare is the U.S. federal health insurance program that provides coverage to people aged 65 years or older and select people younger than age 65 who have certain disabilities or end-stage renal disease (16). An orientation on the use of Medicare administrative data for observational research are available (17). The database, named COVVAXAGE, includes all CVS Health and Walgreens pharmacy customers at any time between 2018 and 2021 who were 65 years or older and had a pharmacy record paid by Medicare. Pharmacy customers who were Medicare beneficiaries were matched deterministically to the 100% Medicare Enrollment File based on name, address, and date of birth. COVVAXAGE has been used in prior studies of COVID-19 vaccine safety and effectiveness in older adults (18). Approximately 95% of eligible customers were matched, creating a cohort of over 38 million individuals overall and more than 27 million individuals aged 65 years or older as of January 1, 2021 (representing 70% of Medicare beneficiaries 65+) (16). COVVAXAGE links complete Medicare enrollment information, inpatient claims, outpatient and provider (i.e., carrier) claims, and CVS Health and Walgreens pharmacy records.

Our study population included community-dwelling Medicare beneficiaries aged ≥66 years with 2 doses of an mRNA vaccine (BNT162b2 or mRNA-1273) documented in Medicare carrier and outpatient files or pharmacy vaccination records between January 1 and August 1, 2021. We excluded individuals who died or were not enrolled in Medicare as of or before the start of follow-up (August 1, 2021). We then excluded those who were in institutional long-term or post-acute care or did not have 12 months of continuous enrollment in Medicare FFS or MA as of the start of follow-up. Finally, we excluded individuals who had a history of three or more mRNA vaccine doses or a Johnson and Johnson vaccine, as the likelihood for mRNA booster vaccination was potentially different for these individuals during the study follow-up period versus those with 2 doses of the primary series of the mRNA vaccines. We restricted the study population to individuals ≥66 to ensure a one-year lookback in Medicare data for covariates, as age-entitled Medicare enrollment begins at age 65.

Race and ethnicity was the primary independent variable of interest, defined using the Research Triangle Institute (RTI) race code that provided the following mutually exclusive categories: non-Hispanic White (henceforth “White”), Black, Asian/Pacific Islander (henceforth “Asian”), Hispanic, American Indian/Alaska Native (henceforth Native American), Other Race/Ethnicity, and Missing Race/Ethnicity (19, 20). As the RTI measure poorly identifies Native American ethnicity and “other” races and ethnicities (21), we focused our analysis on comparing vaccine uptake between White (reference group), Black, Asian, and Hispanic individuals. However, individuals belonging to other racial and ethnic groups were included in all models, with their estimates of booster vaccine uptake reported in the supplement.

The outcome was receipt of a third dose of a COVID-19 vaccine (henceforth “booster vaccine”), which was defined using Common Procedural Terminology (CPT) codes or relevant pharmacy records (definitions available in Supplementary Table S1). Individuals were followed from August 1, 2021 until receipt of a booster vaccine, or they were censored upon death, disenrollment from Medicare, or end of follow-up (December 31, 2021). We did not censor on entry into long-term or post-acute care as we anticipated that the rate of new admission during the five-months of follow-up would be low and non-differential between race/ethnicity groups. The follow-up period for the primary analysis ended on December 31, 2021 because of an administrative change in COVID-19 vaccine billing to Medicare for MA beneficiaries that began on January 1, 2022 (22). However, as a sensitivity analysis, we restricted the population to FFS beneficiaries and extended follow-up for booster vaccine uptake to May 15, 2022 (latest available data at the time of analysis). As some individuals likely received a booster vaccine prior to their official eligibility based on the time since their second dose, we did not apply any time-based restrictions on when booster vaccines could be ascertained based on CDC recommendations during the study period. Finally, we followed individuals on a calendar time scale (i.e., starting on August 1, 2021) rather than a vaccine exposure-anchored time scale (i.e., from the day after completion of their primary vaccine series) to estimate rates of vaccine uptake that only included follow-up time in which booster vaccines were available to patients (in or after August 2021).

To describe the population, we collected demographic (e.g., age, sex, and zip code of residence) and clinical information inclusive of and before the start of follow-up. Geographic region of residence was defined using the US Department of Health and Human Services (HHS) Regions, which are categorized by the major US city in that particular region (e.g., Boston) (23). An area-level measure of social deprivation was obtained from the American Community Survey (2015–2019) (24). We used the National Center for Health Statistics urban–rural classification scheme to categorize counties as large central metropolitan counties (i.e., inner cities), large fringe metropolitan (i.e., suburbs), medium metropolitan, small metropolitan, micropolitan, and non-core. Metropolitan areas are considered urban whereas micropolitan and non-core areas are considered rural. Medicare enrollment type was defined as a categorical variable based on the combination of an individual’s Medicare insurance (FFS, MA, or a mix of FFS/MS) and Medicaid (supplemental insurance for low-income individuals) dual eligibility status in the 12 months prior to follow-up (e.g., dual Medicaid eligible and FFS; not dual Medicaid eligible and FFS).

Finally, using all available FFS Medicare claims (inpatient, outpatient, carrier, etc.) in the 12 months prior to the start of follow up, we measured the following covariates: comorbidities were defined using Hierarchical Condition Categories (HCC) codes (25); frailty was measured using the Kim cumulative deficit claims-based frailty index (CFI) (26); and comorbidity burden was measured using the Combined Comorbidity Index (27, 28). Clinical information was only captured among FFS beneficiaries because MA encounter data (used to derive clinical covariates) was unavailable for the study period at the time of analysis. This clinical information was used only for descriptive purposes and not included in regression models used for the final analysis.

We plotted the cumulative incidence of booster vaccination with 95% confidence intervals (CIs) for each group using the Kaplan–Meier estimator. The cumulative incidence at the end of follow-up corresponded to the incidence of booster vaccination over the average follow-up time of 15.5 weeks. Generalized linear models with a Poisson distribution and log link function estimated rate ratios (RR) and 95% CIs for booster vaccination across racial and ethnic groups. Models included a log-transformed offset for follow-up time and covariates for age, sex, and geographic region. To test the robustness of results from the primary analyses to alternate model specifications in a sensitivity analysis, we included a 4-level covariate combining type of Medicare insurance and dual Medicaid eligibility (i.e., FFS or MA with or without dual Medicaid enrollment) (29).

To describe potential variation in racial and ethnic differences in booster vaccination by age, sex, geographic region, degree of urbanicity, and Medicare plan/Medicaid eligibility, we first plotted cumulative incidence curves for each group. As a formal assessment of effect measure modification on the multiplicative scale and to obtain subgroup-specific estimates of booster vaccine uptake between race/ethnicity groups, we fit a series of adjusted Poisson models with statistical interaction product terms between race and ethnicity group and each sociodemographic characteristic of interest (e.g., race/ethnicity*region) – specifying separate models for each interaction being assessed. Two-sided p-values for homogeneity were calculated, and values <0.05 were considered indicative of heterogeneity in differences. Interacted models were adjusted for the other sociodemographic characteristics as in the primary analyses (age, sex, geographic region). All analyses were conducted in SAS v9.4 (SAS Institute, Cary NC) and Stata version 17.0 (StataCorp LLC, College Station, TX).

Brown University’s Institutional Review Board approved the study with a waiver of informed consent (Protocol # 2103002950). We established data use agreements with CVS Health, Walgreens, and the Medicare & Medicaid Resource Information Center (MedRIC) for Centers for Medicare and Medicaid (CMS) data files.

We observed significant differences in the cumulative incidence of booster vaccination by race and ethnicity by December 31, 2021 (Kaplan Meier cumulative incidence estimates: White = 68.5%; Asian = 67.0%; Black = 57.0%; Hispanic = 53.3%, Figure 1) by the end of follow up. After controlling for age, sex, and region in regression models, Black individuals (RR = 0.78 [95%CI = 0.78–0.78]) and Hispanic individuals (RR = 0.72 [95% = CI 0.72–0.72]) had a significantly lower rate of booster vaccination compared to White individuals. Asian individuals had a slightly lower rate of booster vaccination compared to White individuals (RR = 0.97 [95%CI = 0.97–0.98]). Unadjusted, partially-adjusted, and fully-adjusted model results are available in Supplementary Table S4. Results from the sensitivity analysis including Medicare/Medicaid enrollment type were concordant with the primary results (Supplementary Table S5), except that after adjusting for enrollment, Asian beneficiaries were more likely to receive a booster vaccination (RR = 1.08 [95% CI (1.07–1.08)]). Finally, when we restricted the study population to FFS beneficiaries and extended follow-up to May 15, 2022, the overall proportion of those with a booster vaccination increased (Kaplan Meier cumulative incidence for follow-up to December 2021 vs. May 2022 for FFS beneficiaries: 69% vs. 76%; Supplementary Figure S3). Relative rates of booster vaccination between race and ethnicity groups were similar with this extended follow-up in the FFS population, though disparities were modestly lessened (e.g., Black versus White beneficiaries: RR = 0.78 changed to RR = 0.82; model results available in Supplementary Table S6).

Little variation in the degree of racial and ethnic differences was observed between sex and age subgroups; however, estimates varied widely between regions, degrees of urbanicity, and types of Medicare enrollment and Medicaid eligibility (Figure 2; Supplementary Figures S4, S5; Supplementary Table S7). For example, in the Boston region, Black individuals showed 34% lower relative booster uptake compared to White individuals (RR = 0.66; 95%CI = 0.65–0.68); however, in the Dallas region, Black individuals showed a 9% lower relative booster vaccination rate than White individuals (RR = 0.91; 95%CI = 0.90–0.92; homogeneity p < 0.001) (Figure 2A). Similar patterns were observed across levels of urbanicity, with larger racial and ethnic differences in booster vaccination in the most urban counties than in the most rural counties (e.g., Black vs. White race among those living in large central metropolitan counties = 0.70 [95%CI = 0.69–0.70]; Black vs. White race among those living in non-core counties = 0.91 [95%CI = 0.90–0.93]) (Supplementary Figure S5).

The magnitude of observed relative rates in booster vaccination between racial and ethnic groups when stratified by geographic location was largely the result of changes in the absolute vaccine uptake in White individuals; the degree of absolute uptake in Black and Hispanic individuals varied to a lesser degree. For example, in the Boston region, 73% of White individuals received a booster vaccine compared to 55% of Black individuals; however, in the Dallas region 61% of White individuals received a booster vaccine compared to 58% of Black individuals (Supplementary Table S7). The range of absolute booster uptake across regions was greatest for White individuals (maximum = 73% [Boston region]; minimum = 61% [Dallas]) and smallest for Black individuals (maximum = 58% [San Francisco]; minimum = 51% [Denver]).

Wide variation in booster vaccination across race and ethnicity was observed by type of Medicare enrollment and Medicaid eligibility (Figure 2B; Supplementary Table S7). For example, among those who switched between FFS and MA in the year prior to the start of follow-up, after adjustment, Black individuals had 0.78 times the rate of booster vaccination compared to White individuals (95%CI = 0.77–0.79). However, racial and ethnic differences were generally smaller among those with dual Medicaid eligibility (e.g., Hispanic vs. White among those with FFS and dual Medicaid eligibility = 0.88 [95%CI = 0.87–0.89]; Hispanic vs. White among those with FFS and no dual Medicaid eligibility = 0.82 [95%CI = 0.82–0.83]). Uptake of booster vaccines was much lower among those with versus without dual eligibility overall, and similar to the geography-based analyses, the degree of racial and ethnicity differences was largely a result of lower vaccine uptake in White individuals (Supplementary Table S7).