Association of Physician Group Practice Participation in Bundled Payments With Patient Selection, Costs, and Outcomes for Joint Replacement

This cross-sectional study assesses whether physician group participation in Medicare bundled payments is associated with changes in costs or patient outcomes.

However, when we linked these 154 TINs back to the claims data, we discovered that many of the PGPs that signed up for the MJRLE group in BPCI were large multispecialty or hospitalist practices that signed up for many conditions (e.g. heart failure, pneumonia), and did not actually care for MJRLE patients under the same TIN -only 96 of these practices ever billed for even a single MJRLE on an inpatient basis as the operating or attending surgeon over the study period, although they had many other attributed hospitalizations. We were therefore concerned that these multispecialty practices might have complex billing arrangements under multiple TINs that were impacting our ability to correctly identify them. However, of the 96 practices with one or more MRJLE episodes, 93 were orthopedic surgery practices. We therefore limited the analytic sample, both for BPCI participants and potential controls, to orthopedic surgery practices to enable more appropriate comparisons and because we were confident that the data reflected actual participation in the BPCI model.
We did additional manual verification of the orthopedic practices, in part because many were quite large and we wanted to be sure that the data were accurate. For example, our largest participating practice, located in the Northwest, had 360 NPIs affiliated with its identified TIN per year on average in the MD-PPAS data during our study period, of which 199 were surgeons. We confirmed that as of early 2021, this group practice has over a hundred clinic locations and over 400 clinicians in its online directory, including over 200 surgeons. Our second largest participating practice, located in the Southeast, had 346 NPIs affiliated with its identified TIN per year on average in the MD-PPAS data during our study period, of which 146 were surgeons. Its online directory included over 30 locations and over 400 clinicians that would likely have unique NPIs, including over 140 physicians, as well as physician assistants, physical therapists, hand therapists, occupational therapists, and trainers. Other handchecked practices across a range of practice sizes were similarly comparable between the MD-PPAS and online-searched data.

B. Propensity Matching
Using propensity scores based on PGP and market characteristics, each BPCI PGP was matched without replacement with up to 3 control orthopedic PGPs within the same region and the same baseline volume tertile (0-430, 431-811, and 812 or more admissions for MJRLE in 2013). Automated matching was restricted to PGPs with a log odds propensity score absolute difference below 0.5. We then handmatched 20 practices (14 large, 4 medium, and 2 small) by removing the within-region match requirement and selecting the remaining potential control within the same volume group with the closest number of surgeons in the practice (the dominant factor in the propensity model). Any practice or market characteristic with an SMD of 0.2 or higher after matching was included in the regression models described below as a covariate.

C. Analyses
The regression model described in the methods section was implemented using a marginal, generalized equation approach (the GENMOD procedure in the SAS statistical package). The GEE approach is robust in that a particular distribution for the outcome variable is not required to be specified, only that the distribution be within the exponential family, which includes many common distributions including normal, gamma, and logistic. And the approach does not require specification of the correlation structure since the correlation is estimated empirically from the model residuals. We did however specify an independent working correlation structure so that each patient would count equally in the effect estimates and so that imbalances in samples sizes between practices would not create a bias. Regardless of the specified working correlation, the effect estimates will be consistent as long as there are a sufficient number of practices. We also assumed a linear model with an identity link for the mean function, so that covariates would have additive rather than multiplicative effects on the outcome. Absolute rather than relative changes in costs, as well as in outcome rates, are simpler to interpret and are the conventional way to present the results of policy interventions. To reduce the impact of outliers, costs were Winsorized, in concordance with CMS conventions.
In addition to adjusting for correlation within practices, the model included fixed effects for match groups so that the effect of the intervention was estimated solely by comparing each BPCI practice to their matched controls. The primary predictors in the model were an indicator for time period (preversus post-intervention), intervention group (BPCI versus matched control) and the interaction between these two indicators. The interaction term determined whether the change in outcome was greater for BPCI practices than for their matched controls. The model also adjusted for temporal trends by including a linear term for calendar quarter, and for differences between practices by including fixed effects for practice characteristics which were not balanced by the matching algorithm: number of patients, number of surgeons, percentage Medicare Advantage penetration, and the number of rehabilitation hospitals at the county level. Differences in patients seen at different practices were accounted for by fixed effects for DRG, patient age, patient gender, and the 27 Chronic Condition Warehouse (CCW) co-morbidities.

The primary model equation is:
Expected Total payments = Intercept + TimePeriod + BPCI + TimePeriod*BPCI + Calendar-Quarter + Match-Group(1-91) + Number-of-Surgeons + Number of Patients + Medicare-Advantage-Penetration + Rehabilitation-Hospitals + Patient-Age + Patient-Gender + DRG (1-2) + CCW (1-27) Where TimePeriod includes four group-specific time frames: pre-intervention, burn-in, burn-out, and post-intervention. For each PGP, patients with procedures between 1/1/2013 and 9/30/2013 constitute the reference pre-intervention period and patients seen from 10/1/2013 until the participation date of the BPCI-A PGP constitute the burn-in period. Patients from the participation date until three months later constitute the burn-out period, and those seen from three months post-joining until 9/30/2017 constitute the post-intervention period. Therefore, the post-intervention period is specific to each match group, as are burn-in and burn-out periods surrounding the participation date.
In a sensitivity analysis, calendar quarter was added to the model as a categorical predictor to account for changes over time. Because of the redundancy between this time variable and the study TimePeriods, changes between TimePeriods could not be estimated in the control practices. However, the interaction between TimePeriod and BPCI (i.e., the diff-in-diffs effect) could still be estimated and tested.

D. Model coefficients
Model coefficients for the primary outcome (Medicare payments per episode) are shown below.