Study Design and Data Sources
We had the opportunity to study naturally occurring differences in the rate of implementation of patient engagement initiatives in two large ACOs—Advocate Health Care (AHC) in Chicago, IL, and DaVita HealthCare Partners (DHCP) in Los Angeles, CA. Key background characteristics of the two ACOs are shown in Table 1. Each is a large and long-established healthcare organization participating in the Medicare Shared Savings Program and in other risk-bearing contracts that create incentives to increase patient involvement in their care in order to achieve better outcomes and to reduce the costs associated with emergency department visits and preventable hospital admissions and re-admissions.
Based on the existing literature, we developed a baseline survey of 39 patient activation and engagement activities (available online, Appendix 1). The survey was completed by a clinical or operational leader from each of the 44 AHC and 27 DHCP practices. Practices were scored based on their responses to the 39-item instrument indicating the extent of implementation of each activity within their practice. From each ACO, we randomly sampled four practices from the top quartile and four from the bottom quartile of the score distribution in order to maximize the baseline variance in patient activation and engagement activities of the study sites. This yielded a total of 16 practices for analysis—eight scoring in the highest quartile and eight in the lowest quartile of implementation of patient engagement activities (see Table 2 for descriptive characteristics of the 16 practice sites). The eight from the highest quartile scored an average of 79.1 (range 71.8–100) of the 100 possible points for each of the questions on the baseline survey, and those in the lowest quartile scored an average of 30.6 (range 5.1–42.0) out of the 100 possible points for each question.
We based sample size power calculations on the detection of clinically meaningful changes in PROMIS physical function scores17 and on blood pressure, assuming a 50% rate of response to the patient survey. This resulted in sampling 273 adult patients with CVD and/or diabetes who had at least one visit in 2014 from each practice using the inclusion criteria shown in Table 3. From the electronic health record we collected data on blood pressure, glycated hemoglobin (HbA1c) levels, low-density lipoprotein (LDL) cholesterol levels, comorbid conditions, sociodemographic characteristics, and insurance status on each patient. We collected data on patient-reported outcomes of care (PROs), patient assessment of the chronic illness care that they received (PACIC), and patient-reported activation and engagement (PAM) from a mailed survey administered in both English and Spanish, as needed, with telephone follow-up, obtaining a 51% completion rate (n = 2176). We also collected survey data from adult primary care team members at each of the 16 practices regarding the extent to which the practice exhibited a patient-centered culture and the degree of relational coordination existing among the people occupying different roles on the team, including primary care physicians, nurses, medical assistants, diabetic nurse educators, nutritionists, and receptionists.18
19 We obtained an overall 86% response rate from the team members (n = 411). The study was approved by the institutional review board (IRB) of the University of California, Berkeley, prior to data collection.
PROs included the validated 12-item Patient-Reported Outcomes Measurement Information System (PROMIS) Physical Function form (Short Form 12a), the validated eight-item PROMIS Social Function form (Short Form 8a), and the validated four-item Patient Health Questionnaire for Depression and Anxiety (PHQ-4) emotional health screening tools17
21 (available online, Appendix 2). The Cronbach alpha internal consistency reliability coefficients were 0.92, 0.95, and 0.87, respectively.22
Practice-Level Independent Variables
Practice-reported patient engagement in decision-making activities was measured by a seven-item subscale (α = 0.89) based on factor analysis of the 39-item baseline survey. The items included in the subscale were as follows: “clinicians encourage patients to discuss their work, home life, and social situation”; “staff note patient preferences for treatment in the patient’s record”; “clinicians consistently involve patients in developing treatment goals”; “physicians have follow-up discussions with patients regarding their treatment options and preferences”; “ clinicians discuss the importance of patient advance directives”; “clinicians discuss the hospice care options with patients”; and “clinicians discuss the availability of hospital-based and community-based palliative care”. The response categories included the following: “Yes, fully implemented”; “Yes, partially implemented”; “Yes, but not regularly”; and “No”. This measure was calculated as a continuous score from 0 to 7. We hypothesized that the efforts of practices to better engage patients would be positively associated with better patient-reported outcomes of care through the increased motivation and participation of patients in achieving treatment goals.6
Patient centeredness was measured by a five-item scale (α = 0.92) used in previous research,23 composed of each team member’s degree of agreement with the following statements: 1) the practice does a good job of assessing current patient needs and expectations; 2) staff promptly resolve patient complaints; 3) patients’ complaints are studied to identify patterns and to prevent the same problems from recurring; 4) the organization uses data from patients to improve services; and 5) the organization uses data on customer expectations and/or satisfaction/experiences when designing new services (available online, Appendix 4). We hypothesized that these types of patient-responsive practice behaviors would be associated with better patient-reported outcomes of care given that existing research has found them to be positively associated with greater use of evidence-based care management processes.23
The seven-item between-role relational coordination measure α = 0.87 used in previous research18
25 asked each team member to assess the frequency, timeliness, accuracy, and problem-solving focus of communication with each other team member with whom they interacted, in addition to the degree of shared goals, shared knowledge, and mutual respect among team members (available online, Appendix 3). For example, each team member was asked, “How frequently do people in each of these groups communicate with you about patients with diabetes and/or cardiovascular disease?” With regard to shared knowledge, each team member was asked, “Do people in each of these groups know about the work you do with patients with diabetes and/or cardiovascular disease?” For purposes of these analyses, we restricted the role relationships to those involving the patient’s primary care physician, the nurse, and the medical assistant, as they most frequently interacted with each other and with the patient. Based on existing research, we hypothesized that greater within-group relational coordination would be positively associated with better patient-reported outcomes.17
Patient-Reported Independent Variables
Patients’ perceptions of their chronic illness care were assessed by the 11-item Patient Assessment of Chronic Illness Care (PACIC) scale used in previous research26
28 (α = 0.93). Sample items included the following: “Over the past six months, when I received care for my chronic condition, how often was I: given choices about treatments to think about; helped to set specific goals to improve my eating or exercise; and helped to plan ahead so I could take care of my condition even in hard times”. Response categories included “never”, “sometimes”, “usually”, and “always”. We hypothesized that patients who reported that they were more satisfied with their chronic illness care would report better patient-reported outcomes.27
We measured patient activation using the 13-item Patient Activation Measure (PAM) developed by Hibbard et al.29 (α = 0.90). Sample questions completed by all surveyed patients included “When all is said and done, I am the person who is responsible for managing my health condition;” “I am confident that I can take actions that will help prevent or minimize some symptoms or problems associated with my health condition” and “I am confident that I can follow through on medical treatments I need to do at home”. Response categories were “strongly disagree”, “disagree”, “agree”, and “strongly agree”. The PAM has been associated with positive health behaviors such as aerobic exercise and receiving preventive cancer screenings as well as more favorable emotional health and lower costs of care.6
30 In addition to a continuous measure, the response scores were summarized into four quartiles representing different categories of activation levels. An individual in the lowest level is a passive participant in healthcare decisions. An individual in the second level of activation has the knowledge and confidence to take a more active role in their healthcare, but has not yet done so. In the third level of activation, the patient plays an active role in making healthcare decisions with their providers. In the highest level of activation, the patient has the knowledge and confidence to take action concerning their own healthcare, even during times of stress.30 Extending previous research noted above, we hypothesized that PAM would be positively associated with PROs of physical function, social function, and depression symptoms, and would largely mediate the positive association of better patient experiences of chronic care and better patient-reported outcomes of care.
From the electronic health record, we collected data on the presence or absence of up to 13 co-morbid medical conditions, including other forms of heart failure, atherosclerosis, aortic aneurysm, aortocoronary bypass, hypertension, asthma, emphysema, chronic obstructive pulmonary disease (COPD), mood disorders, other nonorganic psychoses, anxiety, adjustment reaction, and depression.
We controlled for the patient’s latest reported blood pressure <140/90 mmHg; LDL-C ≤ 100 mg/dl, and HbA1c ≤8.0%. We also adjusted for patient age, sex, education, insurance status, and limited English language proficiency.
To account for the clustering of patients within the 16 practices, we used hierarchical linear models (HLM), with patients as the first-level analysis and practices as the second level31 to estimate the association between predictors and each PRO. Given the relative skewness of the PROs toward more positive outcomes, and for ease of interpretation, we report logistic regression results dichotomizing patients’ scores above or below the median on each of the PHQ-4 depression symptom, physical, and social outcome measures. For these analyses, the results are reported as odds ratios (ORs). We also estimated multilevel linear regression models using continuous PRO measures, and obtained nearly identical results (data not shown). We tested for the mediating effect of PAM on each of the patient-reported PHQ-4 depression, physical, and social outcomes by running multilevel mediation tests32 to estimate the direct and indirect effect of PAM on PROs.
We conducted a number of sensitivity analyses involving different measures of disease burden, including 2+ comorbid conditions, 3+ comorbid conditions, and whether the patient had one or more mental health conditions or at least one physical plus at least one mental health condition.33
36 We also examined the effect of including a broader number of patient care team roles in the measure of relational coordination, including diabetes nurse educators, social workers, and receptionists. We tested for a number of potential moderating interaction effects involving disease burden and the PAM to see whether the effects might be greatest for the sickest patients. We also tested for an interaction effect of patient-reported shared decision-making and PAM to see whether shared decision-making mattered most for patients who were very highly activated or were very disengaged. The intraclass correlation coefficient in all models supported model assumptions (p < 0.001). We analyzed the data using Stata language 14.0 (StataCorp LP, College Station, TX) and considered the regression coefficients significant at a level of ≤ 0.05.