Does care-coordination work in a fee-for-service setting? An Analysis of Colorado’s Accountable Care Collaborative

In July 2011, Colorado Medicaid began the Accountable Care Collaborative (ACC) which included elements of Accountable Care but retained fee-for-service reimbursement. Regional Coordinated Care Organizations (RCCOs) were formed to coordinate the care of ACC enrollees through primary care medical providers (PCMPs) who agreed to join the collaborative. RCCOs and PCMPs both received a per-member-per-month (PMPM) payment to coordinate care and bonuses based on RCCO performance on risk-adjusted measures of utilization of low-value care (ED visits, hospital readmissions, and high cost imaging) and well-child visits.  The state also with a data analytics provider to support the RCCOs care coordination efforts and to attribute patient to primary care. The approach taken by Colorado is unique in that the changes were layered onto the existing FFS system and there was no risk shifted to providers making the incentives to reduce utilization relatively weak compared to capitation and other mechanisms designed to influence providers and patients. We estimated a difference-in-differences model that compared utilization of Medicaid enrollees attributed to PCMPs who joined the ACC to those who did not. We used PCMP fixed effects because the PCMP’s decision to join the ACC was not random but due to fixed unmeasured characteristics such as practice style. Qualitative evidence supports this assumption because PCMPs who joined the ACC did so because the goals of the ACC matched their own goals for their practice. Medicaid enrollees entered the ACC by default when their attributed PCMP joined the collaborative. We used inverse probability weighting (IPW) to insure comparability of the ACC and the FFS-only comparison group. The propensity score was estimated using enrollee characteristics (CDPS risk adjusters, time within Medicaid), provider type and specialty because these may be correlated with the PCMP’s decision to join the ACC and utilization. The propensity specification is supported by balancing and overlap tests. The IPW specification with PCMP fixed effects satisfied tests of pre-period parallel trends. We found that ACC-related utilization declined by $14-$16 PMPM relative to the control group (p<0.01). This translates to about a 3% decline in expenditures. The estimate when we do not control for nonrandom provider and patient selection are twice as large, implying that there was favorable selection of PCMPs (and their patients) into the ACC. Our findings suggest that the ACC enrollees experienced a small but significant decline in utilization. A failure to control for nonrandom selection into the program would lead to significantly larger estimates. The results show that relatively small incentives can be used to influence utilization in a Medicaid fee-for-service population. This may be due to the fact that Medicaid reimbursement is relatively low and our results do not generalize to privately-insured or Medicare populations.