Association of Medicare Mandatory Bundled Payment System for Hip and Knee Joint Replacement With Racial/Ethnic Difference in Joint Replacement Care

This cohort study analyzes Medicare claims data to assess whether changes in joint replacement care are associated with Medicare’s Comprehensive Care for Joint Replacement (CJR) model among White, Black, and Hispanic patients.


eFigure. Sample selection
Notes: We identified hospitals participating in BPCI model 1, phase 2 of model 2 or 4 and post-acute care providers participating in BPCI model 3 for joint replacements based on BPCI participant list on the CMS website. We excluded hospitals and post-acute providers from our sample if they ever participated in BPCI models I, phase 2 of model II, model III, and phase 2 of model IV at least once during our study period.

eTable 1. Definition of outcomes and explanatory variables Definition
Data source Primary outcomes in the main analysis (Table 2-4) Total spending Standardized, inflation-adjusted Medicare payments that occurred during episode of care (i.e. index LEJR inpatient hospitalization and 90 day post-discharge period) except payments to durable equipment and hospice care Medicare 100% claims Discharge to institutional post-acute care A binary variable that indicates whether a patient was discharged to institutional post-acute settings (inpatient rehabilitation facility, skilled nursing facility, swing bed, and long-term care hospital) Home Health Agency, Inpatient, and Skilled Nursing Facility Claims

Relevant readmission
A binary variable that indicates whether a patient was readmitted to a hospital within 90 day postdischarge period. We excluded irrelevant readmissions based on CMS definition. Inpatient Claims and list of MS-DRG and ICD diagnosis codes from CMS 1

Standardized, inflation-adjusted spending
We reported all Medicare-allowed payment amounts in 2016 dollars. We also standardized payment amounts by removing Medicare payment variation caused by differences in wage-index, special payments related to medical education, or disproportionate share hospital status. 2

Index hospitalization
Medicare payments for index inpatient hospitalization. Notes: a Outcome violates parallel pre-trend assumption and should be interpreted with caution. Column (1)- (6) show unadjusted value and Column (7) adjusted value. All analyses used linear regression models at the episode level, and adjusted for the interaction between a treatment MSA measure (i.e., whether a joint replacement occurred in treatment MSAs) and a post-CJR measure (i.e., whether a joint replacement occurred during the post-CJR period), and three-way interactions between a treatment MSA measure, a post-CJR measure, and race/ethnicity measures (black and Hispanic measures with white as the reference group). Models also included interactions between race/ethnicity and post-CJR measure, interactions between race/ethnicity and treatment MSA measure, race/ethnicity measures, binary measures of each hospital to account for time-invariant hospital characteristics, and binary measures of each year and quarter. Models also adjusted for patient age, gender, and surgery type. More specifically, we estimated the following care-episode-level regression: where ℎ is an outcome variable for joint replacement that occurred in hospital ℎ located in MSA in year . takes a value of 1 if a joint replacement occurred in the treatment MSAs and 0 otherwise. takes a value of 1 if a joint replacement occurred after April 2016 and 0 otherwise. takes a value of 1 for black patients and takes a value of 1 for Hispanic patients. includes patient age groups, gender, and surgery type (elective knee, elective hip, and hip fracture surgery). ℎ , , and represent binary measures of each hospital, year, and quarter. 1 and 2 measures changes in black-white and Hispanic-white differences under CJR while 3 measured the changes associated with CJR for white patients.
We obtained changes associated with CJR for black patients by adding 1 and 3 (using a post-estimation command that computes linear combinations of coefficient estimates) and changes associated with CJR for Hispanic patients by adding eMethods 3. Propensity score calculation and weight application As a sensitivity analysis, we used propensity score weighting in an attempt to account for differences in hospital and patient composition between treatment and control MSAs. Propensity score weighting, also known as inverse probability of treatment weighting, is a method designed to make observations in the treatment and control groups more comparable by weighing observations based on their probability of being in the treatment group. 8 Rather than running this method on the entire population, we stratified our study population based on year of joint replacement and race/ethnicity (e.g. white patients with a surgery in 2013, black patients with a surgery in 2017, etc.). We chose to run this process separately for each racial/ethnic group to avoid adjusting out meaningful differences between the racial/ethnic groups. The stratified process instead attempts to make observations in treatment and control MSAs comparable within each racial/ethnic group. We chose to additionally stratify the process by year, as the differences between treatment and control populations may change over time, especially if the CJR program lead to changes in patient composition.
Within each strata, we used patient demographics, hospital characteristics and patient baseline health status to predict whether the surgery occurred within a treatment or control MSA. More specifically, our models contained the following covariates: CMS MSA sampling cluster, hospital volume of joint replacements, hospital bed count, hospital ownership type (public, non-profit, or for-profit), hospital major teaching status, hospital disproportionate patient percentage (DPP), patient gender, patient age, patient Medicaid eligibility, type of joint replacements (elective hip, elective knee, and hip fracture surgery), and Elixhauser mortality and readmission risk scores.
We then used the resulting predictive probability of each surgery occurring in a treatment MSA to calculate Average Treatment Effect (ATE) weights (equation 1).
where ( = 1| ) is the predicted probability that surgery i occurred in a treatment MSA, conditional on covariates . Resulting propensity score weights were multiplied by the CJR sampling weights to create final weights that were applied to all models. We used kernel density plots to confirm that treatment and control groups met the sufficient overlap or common support required for propensity score weighting.