The study protocol (Text S1
) and CONSORT checklist (Text S2
) are provided as supporting information. Though nested within the larger MCCD, this study's design and execution were undertaken by the authors independently of CMS or MPR. The national and HQP program-specific evaluation plans designed by MPR for the MCCD have previously been reported 
All participants randomized into the HQP program from the start of the MCCD in April 2002 through March 2008 are included in this study. Traditional, fee for service Medicare beneficiaries with Parts A (hospitals, skilled nursing facility, hospice, home health care) and B (physician services, outpatient care, home health services) insurance coverage, residing in eastern Pennsylvania, 65 y of age and older, with heart failure, coronary heart disease, asthma, diabetes, hypertension, or hyperlipidemia, and receiving care at a primary care practice agreeing to work with the HQP program, were eligible to participate in this study. No minimum prior health care utilization or hospitalization was required for eligibility. Exclusion criteria included dementia, end-stage renal disease, schizophrenia, active cancer (except skin) in the prior 5 y, life expectancy less than 6 mo, and current or imminent residence in a long-term care facility. Individuals at very low risk for future health complications based on a pre-enrollment assessment were also excluded from the study. In September 2006, a protocol change made a pre-enrollment assessment of low risk an additional exclusion criterion, because interim evaluations indicated that control group participants in this stratum were not utilizing enough health care services to allow for a sufficient realization of savings in the intervention group to offset program costs.
In the US, Medicare is provided in two basic forms: (1) fee for service coverage (traditional Medicare), funded and administered by the federal government, and (2) managed care coverage (Medicare Advantage), funded by the federal government, but sold and administered by private health plans. In areas of the country where insurers offer Medicare Advantage plans, Medicare beneficiaries can choose between these two types of coverage. Medicare Advantage plan members often have financial incentives to use providers within networks recognized by the health plan, and such plans may provide various forms of care coordination or chronic disease management services. By contrast, traditional Medicare beneficiaries can choose to receive their care from any participating Medicare provider and can switch providers at any time without financial penalties. Traditional Medicare, to date, lacks significant care coordination or chronic disease management benefits. All the participants in the current study were beneficiaries receiving traditional US Medicare.
Potential study subjects were referred to the study from participating primary care practices. Practices were assisted to utilize administrative billing data to identify Medicare beneficiaries that might be eligible for the study based on ICD9 diagnosis codes, age, and insurance information. Primary care providers reviewed the patient list generated from billing data queries and selected patients to refer to the study. Outreach to potentially eligible patients was undertaken by HQP by way of a mailed letter of introduction and follow-up phone calls inviting referred patients to learn more about the study.
Developing a Network of Participating Primary Care Practices
A network of primary care practices was developed by meeting with and describing the HQP program and the MCCD to hospitals, physician-hospital organizations, independent physicians associations, and individual practices. The basic requirements of practices agreeing to participate include: (1) responding to communications about their patients initiated by the nurse care managers on an as needed basis, (2) making the office medical records available to the nurse care managers and chart auditors, and (3) assisting in case-finding potentially eligible individuals on their patient panels, using billing system reports or extracts, or other mutually agreed to processes. The program was designed and promoted as easy to use and free of burdens related to: paperwork, recurring authorizations or pre-certifications, routine case reviews, or administrative tasks.
Practices were encouraged to “test drive” the program by initially referring a small number or select set of patients meeting eligibility criteria. Offices were not required to sign a contract or commit to a minimum length of participation and there were no financial transactions involved. It was explained to offices that by virtue of the randomization process roughly half of their referred and randomized patients would be assigned to the control (usual care) group and half to the intervention group; underscoring that half of all patients from their practice that participated would not receive any extra services. Business Associate agreements committing HQP to safeguard the privacy and confidentiality of the personal health information provided by the practices were executed.
During the time period of this study, 93 primary care practices in and around the 4,662-km2, four-county service area of eastern Pennsylvania (Bucks, Montgomery, Lehigh, and Northampton) agreed to participate. Patients of these practices received most of their acute care services from seven hospitals owned by six different health systems. Most practices solicited (greater than 80%) agreed to participate except for those affiliated with two hospital-owned, multi-practice networks (one operating as a Preferred Provider Organization) that declined to participate, citing their desire to: (1) implement and manage their own care coordination programs to enhance their ability to negotiate with health plans, and (2) maintain more direct control over such programs.
Participating practices varied widely in terms of size (most had four or fewer primary care providers), use of electronic records, and organizational affiliation (most were independent). In the past few years, an increasing number of practices have implemented some form of the patient-centered medical home (PCMH); designed to support primary care physicians to improve the proactive coordination and tracking of patient care, typically involving the use of information systems, disease registries, and care team models. There have been no observed barriers, operational difficulties, or decreased interest in collaborating with the HQP program as the result of offices adopting the PCMH.
CMS administered the overall conduct of the MCCD. As previously reported, “The Secretary of Health and Human Services, acting through the CMS, determined that the overall demonstration and evaluation met all criteria in both the Common Rule and National Institutes of Health's Exemption Number 5 for exemption from institutional review board review for research and demonstration projects on public benefit and service programs.” 
(page 604). All participants provided written informed consent prior to study enrollment. HQP separately sought and received approval of the Institutional Review Board of Doylestown Hospital (Doylestown, Pennsylvania, US) for the present study.
Classification Prior to Randomization
After providing consent, but prior to study randomization, each participant was classified using two different schema: primary diagnosis and risk stratum. The nurse care management supervisor made the determination of the primary diagnosis. For participants with only one of the chronic health conditions required for study eligibility, that condition was considered the primary diagnosis. For participants having more than one qualifying diagnosis, the condition judged most likely to precipitate a future hospitalization, on the basis of the participant's clinical measures, self-management skills, disease-specific symptoms, and hospital utilization in the prior 6 mo, was chosen as the primary diagnosis.
Eligible participants were also classified into discrete categorical risk strata 
(page 13). The first step in the algorithm HQP used to determine risk strata, is an assessment of geriatric-related risks using the Sutter Health Questionnaire (used with permission, Cheryl Phillips) 
. A number of domains are covered in this questionnaire including: self-rated health, number of medications taken, change in weight, falls, health care utilization in prior 6 mo, living arrangement, care giver status, activities of daily living, instrumental activities of daily living, ancillary health care services used, physical activity level/mobility, chronic illnesses, depression, and tobacco use. Individuals scoring at or above a level 3 on the Sutter instrument were defined as the high-risk stratum for this study. Individuals scoring below this breakpoint on the Sutter tool received a second, disease-specific risk assessment developed by HQP, which was used to classify participants into one of three additional risk strata: moderate, low, and very low. Individuals in the very low risk stratum were excluded from study participation from the outset, and those in the low-risk stratum were also excluded beginning in September 2006.
In the course of administering the pre-randomization Sutter Health Questionnaire a numeric risk score (total score) was also calculated. This score was used to augment the outcomes analysis in this study by creating risk subgroups according to total score tertiles: lower, middle, and upper, defined by total scores of <15, 15–35, and >35, respectively.
Participants randomized into the control group received the usual care afforded to traditional Medicare beneficiaries and following notification of their study group assignment, had no further contact with HQP. Participants randomized into the intervention group were provided the HQP model of community-based nurse care management. This model was previously described in a report by the MCCD contracted evaluators 
. The HQP program was developed over several years in multiple care delivery settings and incorporated a broad portfolio of evidence-based preventive and care management interventions delivered longitudinally by nurse care managers in collaboration with local health care and social service providers. A detailed listing of the elements of this intervention is provided as a supplemental table (Text S3
). Nurse care managers used a database developed by HQP to track their activities and participant contacts as well as key assessments and clinical data on participants. Additional paper-based documentation and assessment tools were organized and maintained in participant chart records. All intervention group participants received additional assessments to identify their physical, functional, cognitive, psychological, behavioral, social, and environmental needs. Participants determined to be in the high-risk stratum, on pre-randomization assessment, received a comprehensive, in-home geriatric assessment involving 15 specified elements, including: physical assessment (HQP), Index of Independence in Activities of Daily Living (Katz), Mini-Mental State Exam (Folstein), Clock Drawing Test (Heinik et al.), Geriatric Depression Screen-Short Form (Sheikh and Yesavage), Nutritional Risk Assessment – Nutrition Screening Initiative (NSI), violence screening (HQP), alcohol abuse screening using the CAGE Questionnaire (Ewing), behavioral and caregiver assessment, home environment safety checklist, Numeric Pain Scale (Jacox), sleep, incontinence, assessment of immunizations and preventive screenings, and psychosocial support needs (HQP).
Regardless of enrollment risk strata assignment, however, the nurse care manager developed an individualized plan for each participant. Three factors were used to establish priorities for this plan: (1) the participant's self-articulated primary concerns and unmet needs, (2) findings from risk assessments and evaluations (initial and repeated), and (3) the participant's motivational readiness. Though a structured instrument was not used to assess an individual's motivational readiness, care managers were trained to recognize stages of readiness for change and to apply interventions appropriate to each stage using the Transtheoretical Model of Behavior Change (Prochaska and DiClemente).
Interventions typically incorporated into an individualized plan included: education, symptom monitoring, medication reconciliation and counseling for adherence, and help identifying, arranging, and monitoring community health and social service referrals. Group interventions such as curriculum-based education; structured lifestyle and behavior change programs for weight loss; weight loss maintenance; exercise classes for improving strength and increasing physical activity; and a balance and mobility program for fall prevention were also provided directly to participants by the nurse care managers. Nurses collaborated with the participants' primary care physicians and specialists on an as needed basis to help participants achieve target clinical goals and receive appropriate and timely preventive care according to guidelines. Collaboration also allowed early identification of new or worsening conditions or symptoms, and facilitation of timely medical interventions in an effort to prevent disease exacerbation, hospital admissions, and unnecessary use of the emergency department.
The nurse care managers were community based and, depending on the size of a practice's patient panel, served patients from multiple primary care practices. Participant encounters consisted of in-person visits, group sessions, and telephone contacts. In-person encounters occurred in the participants' homes, physicians' offices, and other accessible community settings, such as HQP's offices, hospitals, community centers, libraries, and faith-based organizations. Contact frequency was determined by participant need with a minimum standard of a monthly contact. On average, participants received 17.4 total contacts per year during the period included in the current study. More than half of all contacts were made in-person either as one-to-one meetings or as group classes. Individualized intervention plans were continuously updated to match the dynamic needs of participants and their caregivers. Once enrolled into the program, intervention participants received services until they died, moved out of the area, requested disenrollment, had a change in insurance coverage making them ineligible for the demonstration, or were placed in a care environment in which the nurse care manager felt they were unable to significantly add to the effectiveness of care (e.g., hospice placement). Once fully trained, each care manager served 85 to 110 participants depending on caseload complexity, geographic distribution, experience, and phase of study recruitment.
In 2007, a protocol of intensified follow-up was added for participants transitioning home or to another level of care upon discharge from hospital. The protocol established guidelines by which nurses provided timely coordination and communication with hospital and post-hospital care providers. The goals were to ensure well informed, safe, and expeditious discharge plans, perform timely patient follow-through on discharge instructions, reconcile medications, and identify and address any errors, omissions, or contraindications in order to prevent readmissions and other serious adverse events.
Program implementation and reliability were supported by careful nurse selection and recruitment practices, pre-service training, ongoing coaching and supervision, structured protocols, explicit operating procedures, clearly articulated performance standards, and a system of data management and statistical process control analysis and reporting to support organizational decision making. A further description of the management elements of this model is provided as a supplemental table (Text S4
). This set of management practices has been described as “core implementation components” 
. Program improvement efforts were ongoing and continuous and resulted in numerous refinements to the program over the course of its implementation within the MCCD.
The main objective of this study was to determine whether HQP's model of community-based care management, as implemented in the MCCD, is associated with a reduction in all-cause mortality overall and within subgroups of risk strata and primary diagnoses. Another objective was to determine whether there was an intervention-associated reduction in all-cause mortality within subgroups defined by tertiles of a numeric risk score obtained on intake assessment using the Sutter Health Questionnaire. The main reasons to explore treatment effect within these subgroups included: (1) refine future program eligibility criteria to direct resources to those that benefit most from the intervention, and (2) permit comparison of impacts on health outcomes to financial outcomes using similar or identical subgroups used by MPR and CMS in their separate and independent financial analyses. It was hypothesized that participants classified as belonging to one or more high-risk subgroups were more likely to demonstrate an intervention-associated reduction in mortality over the follow-up period of this study. The pre-specified and post hoc analyses of the study are summarized in .
Outcomes and subgroup analyses specified in the study protocol.
The primary outcome of this study was the risk of death from any cause among intervention participants compared to control participants overall and within subgroups defined by risk strata and primary diagnosis. Vital status as of March 31, 2009 was assessed for all participants. The data source used to establish death was the online Social Security Death Master File (SSDMF) (http://www.ssdmf.com
). Social security numbers obtained from participants following informed consent and prior to randomization were used to check vital status in the SSDMF.
Outcomes: Specified Post Hoc
Analyzing deaths within subgroups defined by tertiles based on the numeric risk score obtained from the Sutter Health Questionnaire was not pre-specified in the study plan. After the study began, but before analysis commenced, this outcome was added. On the basis of random samples, we estimated an overall error rate of 3%–5% in the assignment of participant risk stratum. This rate was due to mistaken Sutter level determinations resulting from the manual tallying of risk scores and variation in the optional use of “flags” (specific question responses defined in the Sutter Questionnaire), which can, if four or more are present, result in increasing the Sutter level by one level. The numeric risk score of the Sutter Health Questionnaire when calculated retrospectively by computer using questionnaire data fields was more reliable.
The risk score derived from the Sutter Questionnaire is obtained in the first step of a multi-step process required for final risk stratum assignment. The risk score is a numeric variable (range in our data: 1–136, mean 29). The use of a computer calculated risk score alone, if predictive of outcomes, could offer a more streamlined, reliable, and efficient method of risk classification, potentially improving future program operations.
The original minimum enrollment recommendation for MCCD study sites (686 in total; 343 participants each for treatment and control groups) made by MPR as part of their sample size estimation was based on the expected impact of effective interventions on hospitalization as described in MPR's study plan for the MCCD 
. These original sample size calculations were not based on estimated impacts on mortality. Given the actual number of overall participants in this study (1,736), the observed probability of death in the control group (0.129), and the observed unadjusted hazard ratio (0.75), with alpha set at 0.05, this study is estimated to have a power of 58% for analysis of overall mortality risk using the Cox proportional hazard method. Similarly calculated power estimates for subgroups were lower, with the exception of the upper risk tertile (power
67%) and coronary heart disease (power
Randomization and Blinding
The study was conducted as a parallel group, randomized, controlled trial. Randomization took place at the individual participant level within each of the risk strata determined by HQP prior to enrollment (high, moderate, and low) using a secure website managed by MPR. Participants were randomized on a 1
1 (intervention: control) basis. All randomization was done offsite by MPR per a protocol established by them and approved by CMS using randomly generated, concealed 4-digit “strings” of treatment-control assignments. By excluding strings of all treatment or all control assignments runs of more than six consecutive assignments to any group were prevented. The random assignment result was available to the program site via the website almost immediately. For practical reasons, study group assignment was not blinded.
All participants randomized into the trial from its start in April 2002 through March 2008 are included in the outcome analysis according to their original study group assignment. The primary outcome (vital status) on all randomized participants (regardless of early program discontinuation) was collected and analyzed through March 31, 2009. The observation period available for each individual participant ran from his or her date of randomization through March 31, 2009 or the completion of a full 5 y of observation (whichever occurred first). Discontinuation from study participation occurring before observation endpoints were reached, for any reason including lost to follow-up, was not a reason for exclusion from the outcome analysis.
Mortality over time was plotted using the Kaplan-Meier method with p-values calculated using the log-rank test. The Cox proportional hazard method was used to calculate hazard ratios. Covariates selected for inclusion in Cox regression models had a significant association with the risk of death in univariate analysis and a recognized association with mortality (sex, age group, primary diagnosis, perceived health rating, number of medications taken, hospital stays in the past 6 mo) or failed to reach significance in univariate analysis, but are widely acknowledged to have a strong association with death (tobacco use). The proportional hazard assumptions for Cox regression models were tested using Schoenfeld residuals and no violations were identified (all p-values ≥0.05). Subgroup analyses include significance testing of interaction effects using likelihood ratio testing to compare proportional hazards models with a subgroup-treatment interaction term to one without.
Comparison of categorical data was performed using Fisher exact test. Comparison of continuous data was performed using the Student's t
test when data was normally distributed or Wilcoxon's rank sum method when data significantly departed from a normal distribution. All values for p
were calculated using two-sided tests. Statistical tests were performed using Stata/MP 10.1 for Macintosh (http://www.stata.com