Population Health in Primary Care: Cost, Quality, and Experience Impact
Tracy L. Johnson, PhD, MA; Mary van der Heijde, FSA, MAAA; Stoddard Davenport, BA; Carlos Irwin Oronce, MD, MPH; Daniel Brewer, BA; Rachel Everhart, PhD, MS; Patricia Gabow, MD; Simon J. Hambidge, MD; Adam Atherly, PhD; and Holly Batal, MD, MBA
Constraining healthcare cost growth has been a focus of health policy since the enactment of Medicaid and Medicare in the 1960s.1 Despite substantial efforts, high medical inflation persists, governmental health spending continues to crowd out other priorities, and health outcomes remain poor compared with those of international peers.2-4 Post Affordable Care Act, efforts to reverse these trends have focused on patient-centered medical homes (PCMHs), complex care management, transitions of care, and accountable care organizations (ACOs). Results to date have been inconsistent, with a limited number of multifaceted programs reducing healthcare expenditures.5-8
Programs that have lowered costs and improved quality and patient experience (the Triple Aim) often share features of Wagner’s Chronic Care Model.9 Most have implemented comprehensive team-based care, enabling proactive provider-patient interactions and tailoring care models to better support the needs of high-risk patients.10-15 Clinical information systems to facilitate population health management have also been foundational, including high-risk patient identification, decision support, and clinical performance feedback.16,17
Most cost-saving initiatives have targeted Medicare or commercial populations and have achieved savings by reducing hospitalization.10,12,13,18-22 Among the few well-designed Medicaid programs, evaluations assessing costs or hospitalizations have shown mixed findings.23-25 This suggests the need to adapt successful interventions to Medicaid populations, which will likely require attending to social determinants of health.22,26,27
Finding effective Medicaid approaches takes on increased salience as Medicaid/Children’s Health Insurance Program (CHIP) enrollment has outpaced that of Medicare.28,29 To this end, the Center for Medicare & Medicaid Innovation (CMMI) funded Health Care Innovation Challenge Awards (HCIAs) to accelerate at-scale, delivery system reform.30 We describe the reach and Triple Aim outcomes of an HCIA-funded primary care practice transformation at a large urban safety-net institution. This intervention employed predictive risk modeling to segment patients according to clinical and financial criteria, and sought to target higher-risk patients for more, and more intensive, services. This evaluation tested whether a risk-stratified, enhanced primary care delivery model demonstrated savings through reduced inpatient spending, while improving quality and patient experience for Medicaid and Medicare adults.
Denver Health (DH) is an integrated, academic safety-net delivery system and the largest provider of Medicaid and uninsured services in Colorado, serving approximately 200,000 patients annually. DH provides comprehensive outpatient and inpatient services and operates a managed care plan. This integration enables data capture across the care continuum.
Although the program targeted both adults and children in need of primary care services, this evaluation focused on Medicaid and Medicare beneficiaries older than 19 years, specifically: DH primary care users (≥1 primary care visits in the prior 18 months); DH managed care members; or frequent users of DH’s emergency department, urgent care, or hospital services (≥3 services in a year). The target population was dynamically redefined monthly, according to a validated population attribution and risk-stratification algorithm, described briefly below and detailed elsewhere.31 Excluded were individuals who could not be risk-stratified due to no claims history or short enrollment periods.
DH adapted commercial predictive risk modeling software to assign patients to 1 of 4 tiers of care needs. The algorithm relied primarily on age, gender, diagnosis, clinical procedure, and medication history. Through a multidisciplinary process, DH-defined rules were developed and integrated into the algorithm. These rules considered additional diagnostic information, clinical registries, and utilization patterns signaling unmet social or behavioral health needs. Algorithm development for high-risk tiers focused on defining patient groups at risk of hospitalization with distinctive care support needs, not solely on high costs. As noted, the algorithm was rerun monthly to capture new patients and changes in health status.
Intervention: Tiered Population Health Approach to Primary Care
This population health intervention built on DH’s National Committee for Quality Assurance–certified PCMHs and used Wagner Chronic Care Model principles to implement team-based care for complex populations.32-36 Consistent with Wagner’s vision of health information technology–enabled, proactive, prepared teams, DH provided real-time patient risk-tier information for care planning at the point of care. We specified a graduated set of enhanced clinical and electronic services appropriate to each risk tier, with more higher-intensity services targeted to higher-tier patients. Standard work that considers both tier and individual needs guided care teams’ service provision.
All patients (tiers 1-4) were provided “usual care” medical home services, complemented by new, optional electronic messaging reminder services. DH also expanded primary care staffing to include new team members to provide disease management, care transition, and patient navigation services (tiers 2-4). This enhanced-care team included nurse care coordinators, clinical pharmacists, behavioral health consultants, and patient navigators. More comprehensive multidisciplinary care management support was available to complex patients (tiers 3-4) during and between visits.
For tier 4, separate high-intensity clinics were established for targeted subpopulations: children with special healthcare needs, medically complex adults with multiple admissions, and adults with severe mental health diagnoses. All high-risk teams sought to empower patients through multidisciplinary care planning, goal setting, and problem-solving approaches. Patients also received referrals for specialty care, substance abuse treatment, housing, and other community resources (Figure).
Process and Outcome Measures
Primary outcomes were program reach, patient experience, quality, and total cost of care. Patient reach was quantified as the number of patients receiving any face-to-face or phone-based patient interaction by the HCIA-funded enhanced-care or high-intensity teams. Usual care clinical contacts were not included, nor were low-touch services, such as phone messages, letters, and text messages. These exclusions sought to avoid overstating program reach by focusing on new, intensive HCIA services.
To assess overall quality, DH uses an internal composite quality metric comprising individual indicators largely adapted from the Health Resources and Services Administration’s Health Center Program Uniform Data System. DH used the Consumer Assessment of Healthcare Providers and Systems’ PCMH-aligned items to assess patient experience.
Assessed from the payer perspective, the total cost of care measure was calculated by aggregating medical expenditures across all inpatient, outpatient, professional, and other services, and was expressed on a per-member per-month (PMPM) basis to account for differing lengths of enrollment. DH had nearly complete data capture for its managed care members because the DH health plan pays members’ medical claims. For those enrolled in fee-for-service (FFS) programs, spending estimates for non-DH services were derived from public sources (listed in the eAppendix [eAppendices available at ajmc.com]). Savings net of program expenses were also calculated.
Program expenses included salary and benefits of HCIA-funded staff, including the enhanced-care team, information technology, and evaluation personnel. Clinical staffing associated with reimbursable clinical care for program participants was not counted, nor were program development costs. This represented DH’s annual, unreimbursed costs to continue the program post award (eAppendix).
This quasi-experimental analysis compared trends in payer spending between a baseline performance period and 2 intervention performance periods. Because DH enrolls a majority of the Denver Medicaid market and the program was implemented at scale, no concurrent DH or non-DH control population exists. Therefore, it was necessary to develop an appropriate historical comparative population to estimate what the baseline costs would have been without the intervention. Observed costs were then compared with this historical cost benchmark. CMS uses a similar methodology to assess the financial performance of Medicare ACOs.37 Consistent with CMS’ approach, population-level cost outcomes were assessed whether or not targeted individuals were reached by the intervention. In the actuarial literature, this is approach is known as quasi-experimental analysis using an adjusted historical control.38
To construct performance populations for analysis, DH used its patient attribution and tiering algorithm to identify the target population during each month of baseline and intervention performance periods, resulting in 38 cross-sections, each with different member month counts and tier distributions. These 38 cross-sections were subsequently aggregated into 3 performance periods: a 12-month baseline period, a 14-month early program period, and a 12-month mature program period. The decision to create 2 intervention periods balanced several considerations: distinguishing early and mature program effects, minimizing seasonal effects by selecting measurement periods of approximately equal length, and isolating the potentially confounding effect of the Medicaid expansion in 2014 into a separate performance period.
Changes in medical costs were calculated by subtracting observed costs during the 2 intervention periods from the historical cost benchmark. In contrast to a pre-post analysis that focuses on a single population cross-section that is reached by the intervention, we refresh the population (38 cross-sections). This addresses regression by accounting for population dynamics such as changes in health status. To the extent that high-risk patients became lower-risk, died, or left the population during the intervention periods, this phenomenon also occurred during the baseline period, enabling detection of performance improvement net of these effects.
During the intervention periods, observed member months (which differ by time period) were the denominator used to calculate PMPM costs, but were not the multiplier to calculate total costs. To ensure the intervention period costs were comparable to those of the baseline period, the PMPMs were applied to the baseline population’s tier distribution. This “tier mix” adjustment ensures that patient health status was comparable across performance periods, effectively holding tier mix constant (see eAppendix for details).
To establish the historical cost benchmark, against which the intervention period costs can be compared, we quantified payer spending for the baseline population, adjusted by a medical inflation factor to account for secular cost trends. This adjustment is necessary to enable direct comparison of costs during the baseline and intervention periods. The Medicaid trend factor was 3.7%, consistent with the trend developed during Colorado state Medicaid agency annual capitation rate setting for DH’s health plan. The 2.3% Medicare trend factor was derived from the National Health Expenditures Projections report. Baseline costs for the Medicaid expansion population were estimated according to insurance rate-setting methods for a new population (eAppendix). We stratified cost outcomes by payer to account for payer-specific reimbursement levels.
Table 1 compares baseline and intervention populations according to tier, gender, and age. Member months are post adjustment, reflecting the above-described tier-mix adjustment.
Program Reach Results
As intended, a greater proportion of higher-tier patients were reached by DH interventions than were lower-tier patients. More than half of adult tier 4 patients were reached by either primary care–based enhanced-care team members or visited a high-intensity clinic during the ramp-up (51.0%) and mature program periods (56.4%). Program reach was lower for lower-tier patients: for the ramp-up and mature program periods, respectively, the percentages were 24.6% and 25.6% for tier 3, 15.7% and 12.9% for tier 2, and 1.5% and 1.4% for tier 1.
Quality and Patient Experience Results
The DH composite quality metric measured 77% at baseline. It increased to 82% during the intervention ramp-up, declined to 72% in 2014, and rebounded to 81% by mid-2015. Four of 6 patient experience metrics improved or remained constant during intervention ramp-up, and improved relative to baseline in 2014 and 2015 (Table 2).
Total Cost of Care Results
Reductions in total costs were observed for the majority of periods and payer populations, except for Medicaid managed care during the ramp-up period and Medicare managed care in the mature program period (Table 3). DH achieved a cumulative $10.9 million reduction in the total cost of care for its Medicaid and Medicare FFS populations across the 2 program implementation periods. These savings largely accrued to the state and federal governments. The largest share of this FFS cost avoidance ($8.2 million) was attributable to Medicare FFS. An additional $5.0 million reduction in the total cost of care was estimated for DH’s capitated managed care during this same timeframe. The annualized personnel cost of the adult program totaled $1.8 million.
During both the ramp-up and the mature program periods, reductions in the total cost of care for tier 4 adults were observed across all managed care and FFS payers, ranging from –$40.88 PMPM to –$737.20 PMPM. Reduced tier 4 spending was concentrated in inpatient reductions. For lower-risk populations (tiers 1-3), changes in claims costs varied by payer and by year. In aggregate, the analysis shows small cost reductions for tiers 1 and 3 (totaling $1 million across years and payers).
Tier 2 costs consistently exceeded the benchmark, totaling $4.8 million for Medicaid and $2.6 million for Medicare. For both Medicare and Medicaid, this higher-than-expected tier 2 spending offset, but did not eliminate, overall cost reductions.
We conducted a subanalysis on DH’s Medicaid and Medicare managed care populations to further explore the changes in tier 4 inpatient costs. We chose these populations due to the more complete data capture. Table 4 reveals that reductions in inpatient costs are not consistently observed in lower-risk tiers, suggesting that tier 4 inpatient cost reductions are not due to a broader secular trend affecting all DH populations. Consistent with prior research, Table 4 reveals that tier 4 inpatient savings are offset by increased spending in other areas, possibly reflecting program-driven referrals that resulted in additional service provisions.12,24,20
This study provides an important contribution to the literature as one of the first cost analyses of a major CMMI/HCIA initiative. Compared with an inflation-adjusted baseline period, net reductions in PMPM spending were observed in 5 out of 6 payers during 2 subsequent intervention periods. Patient satisfaction measures also improved during this same timeframe, suggesting that cost reductions were not achieved at the expense of patient experience. Overall quality improved in 2013 and 2015 over baseline performance. Reduced performance during 2014 may be partially explained by pent-up demand during the Medicaid expansion.
As hypothesized, reduced inpatient spending among high-risk adults drove the overall reduction in the total cost of care. High-risk (tier 4) adults with multiple chronic conditions and repeated hospitalizations were targeted for multiple interventions, and more than half were reached. The program reached fewer lower-tier patients and concomitant changes in associated payer spending were less pronounced, and even increased in some cases. This latter finding suggests that team-based care for lower-risk patients may need to be more targeted or evaluated over a longer time horizon.
Although we observed year-to-year fluctuations in performance, average cost reductions for Medicare exceeded Medicaid performance and FFS reductions were generally larger than those for managed care members. These results are consistent with literature that finds that short-term return on investment requires carefully targeting patients at risk of hospitalization, reducing Medicare inpatient costs may be easier than for Medicaid inpatient costs, and opportunities may be greater among unmanaged FFS populations. Differences in cost-avoidance by payer and by year underscore the importance of all-payer approaches to overall financial sustainability.
Gross cost reductions of nearly $16 million over the entire 26-month intervention period are substantially larger than the approximately $3.9 million in staffing expenses, supporting the self-sustaining potential of risk-stratified primary care delivery models. However, neither FFS nor experience-based capitation permits ongoing reinvestment of savings into program costs. Cost reductions accrue to at least 3 separate payers: the federal and state government for FFS Medicare and Medicaid patients, and the DH health plan for its managed care members. Since DH owns and operates its own health plan—an unusual arrangement among safety net institutions—there is a direct means to “capture” a portion of the payer savings. Combined Medicaid and Medicare managed care cost avoidance was estimated at $2.3 million (annualized) compared with the $1.8 million in annual program costs. However, because capitation rates are based on historical claims costs, funds for program reinvestment will decline over time as capitation payments are rebased to reflect lower levels of medical expenditures.
Thus, current payment models—even capitated managed care—do not align incentives. Although both Medicare and Medicaid have implemented new FFS care coordination, care transitions, and integrated care reimbursement opportunities, requirements are often highly prescriptive, process-oriented, and not consistent across payers. As a result, staffing models and clinical work flows that meet Medicare rules do not necessarily satisfy those of Medicaid, and vice versa. However, multipayer approaches are necessary to ensure sufficient funds and to smooth out year-to-year variations, especially for high-volume Medicaid providers. This underlines the need to accelerate implementation of advance payment models that better align financial incentives.
Recognizing that current payment models do not adequately incentivize population health approaches, CMS has set a goal to have 90% of Medicare non-FFS arrangements by 2018.39 This evaluation offers both programmatic and financing insights relevant to alternative payment model development, such as those being developed under the Medicare Access and CHIP Replacement Act.
First, although we applied a tier-mix adjustment and stratified by payer to ensure equivalence between the baseline and intervention populations, some differences in population characteristics remain that could partially account for the findings. Second, incomplete data capture is a concern. Although we estimated service use at non-DH facilities among FFS populations, we did not attempt to estimate cross-sector impacts to the criminal justice or social service systems. Third, there is no widely accepted best practice for trend assumptions to establish the cost benchmark. Because DH accounts for approximately 57% of the Denver Medicaid primary care market, we cannot use actual trend. We therefore selected trend assumptions that were based either on state rate-setting practices (3.7% for Medicaid) or national health spending trends (2.3% for Medicare). These trends are equivalent or lower than a recent CMS actuaries’ analysis that found that US health spending growth was “historically low” between 2009 and 2013, averaging 3.7%.3 Our trends are also lower than a touted Oregon Medicaid program analysis, assuming a 5.4% trend on a 2011 baseline.31,40 Although our approach does not exclude the possibility that savings result from a broader, secular phenomenon of lower inpatient spending, reductions were concentrated in tier 4 where program reach was greatest. Generalized inpatient reductions across all tiers were not observed. Finally, the program was implemented at a single institution; findings may not be generalizable.
Recent leveling-off of costs is often attributed to the aggregate effect of readmission reduction programs nationally, most of which have been launched without formal research designs. Although at-scale implementation complicates efforts to identify unexposed, concurrent comparison groups, from an institutional perspective, improvements over past performance are relevant even while the broader sector also improves.34 This analysis demonstrates that a large, multifaceted program may be evaluated by benchmarking against historical costs, using well-accepted actuarial methods to address regression to the mean.41 We conclude that risk-stratified, enhanced primary care delivery models hold Triple Aim promise, assuming supportive payment models.
The authors would like to acknowledge the project team responsible for designing and implementing the intervention, including the core management team, clinical teams, information technology team, and evaluation team, as well as Ambulatory Care Services and Executive Leadership (past and present).
This study was conducted for quality improvement/quality assurance purposes. Colorado Multiple Institutional Review Board (COMIRB) determined this project to be “not human subjects” research. Results are not generalizable.