Understanding What Inuences Community-based Care Coordination Improvement: Interpreting Variation in a Medicare Hospital Readmissions Reduction Program Using the Consolidated Framework for Implementation Research (CFIR)

Background Quality Improvement Networks Quality Improvement Organizations (QIN-QIOs) developed community coalitions to align care coordination efforts for Medicare beneciaries in order to reduce readmission rates within geographically dened communities. This CMS (Centers for Medicare & Medicaid Services) funded national quality improvement program worked with 380 coalitions from 2014-2019, facilitating a variety of interventions within each community. Baseline readmission rates among communities, calculated from claims data, varied from 17.7 to 112 readmissions/1000 beneciaries. Program results ranged from +40.7% (high performance) to -35.8% (low performance) relative improvement. We applied an implementation framework (CFIR) to the QIN-QIO efforts to dene common characteristics of interventions, implementation strategies, and contexts in which improvement efforts took place. We identify features associated with successful and unsuccessful intervention implementation, and with changes in readmission rates. We selected 22 communities representing a range of relative improvement, geographic characteristics and baseline readmissions rates. We measured the QIN-QIO’s perception of inuence of individual CFIR constructs on community readmission rates over time using a written assessment and elicited details and mechanisms through structured interviews. Two independent reviewers qualitatively coded transcribed interviews. Final ratings for the inuence of each CFIR construct on community performance were assigned by consensus, ranging from -2 (strong negative inuence) to +2 (strong positive inuence). Some adaptation of the CFIR, such including codes in a coalition domain, and adding constructs to the outer setting domain, such as healthcare market characteristics, helped t the framework to the QIN-QIO work. The characteristics of individuals domain was less applicable to this study. Several constructs were found to be associated with improvement, or lack of, in readmission rates in communities. the negative among high RIR communities compared to at or low RIR communities, indicating interventions well aligned with the values and workows of providers implementing those interventions are more likely to result in lower readmissions. Similarly, constructs of readiness for implementation were more positive among high RIR communities relative to low and at RIR communities, indicating that the capability of implementers to inuence readiness is an important component for success. The subconstruct of leadership engagement showed the strongest relationship within this construct, indicating that visible commitment and involvement of leadership was a distinguishing element of successful implementation within this project.

QIN-QIOs worked with more than 380 community coalitions between 2014 and 2019; these coalitions performed root cause analyses of hospital readmissions and facilitated implementation of interventions within each community. Most QIN-QIOs also supplied direct technical assistance, such as data collection, analysis, or training in intervention or quality improvement strategies, to individual providers or organizations as possible given the scale of the project. QIN-QIOs and provider participants tracked and reported process and intermediate outcome measures; by design, these measures were not standardized to encourage customized and adaptable approaches as needed. QIN-QIOs submitted quarterly narrative progress reports describing successes, challenges, and harm mitigation strategies.
Program progress was tracked by quarterly readmission rates for the Fee-for-Service (FFS) Medicare bene ciaries in each community (readmissions/1000 bene ciaries), calculated from claims data. Readmissions results varied widely among the participating communities, from a 35.0% reduction from baseline to a 44.2% increase over the course of the SOW.
Despite robust data and reporting requirements, a large number of participating communities, and variability in outcomes performance it is unclear what interventions and implementation strategies were effective, and in what context. The QIN-QIO program allowed communities maximum exibility, resulting in wide variation in multiple aspects of individual projects. These included the characteristics of communities and providers engaged in the work, interventions selected, measures collected, populations targeted by efforts, and state and local policy infrastructure. This variation has made it di cult to distill a set of common lessons that might guide future work in care coordination.
We sought to de ne the implementation characteristics of individual community efforts, and aggregate insights, by applying the Consolidated Framework for Implementation Research (CFIR) (16) (16). The CFIR consolidates a menu of constructs, or characteristics that in uence an implementation effort, from multiple theoretical implementation frameworks, organized within ve domains. Domains include the organizational setting in which implementation efforts occur (Inner Setting), the external environment beyond the in uence of the people engaged in implementation (Outer Setting), features of the intervention itself (Intervention Characteristics), the individuals working to implement an intervention, including the participants in the intervention (Characteristics of Individuals), and the processes used to put an intervention into practice and adapt as necessary (Process). We selected the CFIR because it accounts for Inner and Outer Setting contexts, has readily available support tools (17), and has been used to evaluate complex, large-scale, publicly funded improvement efforts across a variety of settings (18) (19) (20). The purpose of this project was to determine whether the CFIR could serve as an appropriate taxonomy for describing QIN-QIO care coordination efforts, identify CFIR constructs associated with intervention implementation, and associate community characteristics with changes in 30-day hospital readmissions.

Methods
Theoretical Qualitative Approach: We chose a realist evaluation approach to understand factors in uencing intervention implementation, and by extension, success or lack of success in achieving desired outcomes. Realist evaluation acknowledges that programs are self-transformational, embedded within dynamic social, political, and environmental systems (21), and best understood through accounting for contextual in uence. We used pre-determined constructs from the CFIR supplemented by inductive coding to create a menu of constructs that could incorporate multiple theories of behavioral change. Inductive determination of constructs allows theory application efforts to be tailored to the speci cs of the situation being evaluated, and better re ects the realist philosophies than one individual theory of change (22).

Sample Selection:
We sorted communities by relative improvement rate (RIR) in readmissions per 1,000 Medicare FFS bene ciaries between two annual timeframes, from baseline (7/1/2013-6/30/2014) to remeasurement (7/1/2016-6/30/2017). We identi ed communities by RIR in the top 10%, the bottom 10%, and the 10% centered on 'no improvement'. We de ned 'no improvement' as 0.00% RIR. The lowest RIR communities showed increases in readmissions per 1,000, or negative RIR. We sorted each set of communities by readmission rate at baseline to identify candidate communities within each stratum that started with high vs. low readmission rates. We selected 22 communities aiming for a group that represented geographic distribution, differently sized Medicare FFS populations, and balance between rural and urban settings (Table 1 located in the appendix).

Study Participants:
We invited QIN-QIO staff who led work for the care coordination task from each of the 22 representative communities we identi ed (Table 1 located in the appendix) and 100% agreed to complete an assessment (Additional le 1) and participate in a semi-structured interview. Participants were invited to an orientation webinar to explain the goals of the project and to provide an introduction to the CFIR, including links to online resources and a sample paper demonstrating use of the CFIR to evaluate a complex initiative (18).

Data Collection:
QIN-QIOs were asked to select an intervention for the sample community that they believed either had the greatest impact on readmission rates or impacted the greatest number of resident bene ciaries. They were sent a written assessment that asked for a description of that intervention, and a rating of each CFIR construct according to its impact on implementation of the intervention, and on the community's RIR. In addition, they were asked to indicate the degree of con dence in their ratings, as it was unlikely that QIN-QIO staff would always have the information needed to assess each construct. After each assessment, the research team wrote a brief narrative summary of the community and the intervention and implementation strategies, including identi ed facilitators and barriers. We then performed 90 minute follow up interviews with each QIN-QIO using a standard guide customized according to the results of the written assessment (Additional le 2). We used open ended questions to clarify our initial understanding, and probes to elicit details related to each CFIR construct.
Interviews captured background information about the community, selection of the intervention, and how various CFIR constructs impacted the team's ability to implement the intervention as planned and reduce community readmissions incidence. Interviewees were also asked to re ect on elements that in uenced the community's readmission rate regardless of implementation success. Interviews were audio-recorded and transcribed.
Coding and Analysis: The rst transcript was coded by all ve coders through open discussion to establish common interpretation of constructs and identify needs for additional or adapted construct de nitions. We created three categories for the construct of evidence strength and quality to re ect the methods used in the task: evidence based on traditional experimental research and reported in peer-reviewed literature; degree to which interventions aligned with root cause analyses (RCA) of readmissions for Medicare bene ciaries in the community; and evidence of success using the chosen approach in a different QIN-QIO-facilitated community. Formal experimental research testing intervention techniques for driving readmissions reduction through coalition-facilitated multi-stakeholder approach is sparse (24), and most QIN-QIO experience, although extensive, is unpublished.
Transcripts were then independently coded using Atlas.ti by two of the ve team members, using our codebook (Additional le 3) modi ed from the publicly available CFIR codebook. Coding pairs varied for each community to increase consistency in interpretation of codes. We used a consensual qualitative approach which allowed inductive coding as needed; new suggested codes were discussed with the larger team and added to our codebook through agreement. The team met regularly during the coding process to discuss challenges and coding decisions, and to formalize code de nitions to improve interrater reliability. Key decisions were recorded in an audit document for reference.
After coding the rst few interviews, all coding teams noted two signi cant challenges related to accurate assessment of setting. First, assigning structures, processes and implementation elements associated with the coalition to either the inner or outer setting resulted in frequent disagreement and inconsistency, as the coalition was neither the primary setting where most implementation activity occurred (inner setting) nor a setting characterized by elements outside of the control of participants engaged in implementation activities (outer setting). Thus, the team unanimously agreed to establish a coalition domain and allowed any relevant construct to be applied to this domain. (Additional le 3) Second, given the nature of the work as a community-wide initiative, interviewees often commented on the in uence of community-speci c elements that did not t neatly into any of the four outer setting constructs. This was especially apparent in re ections on community readmission rate in uences, such as statements noting antagonism between competing hospital systems resulting in insu cient spread of interventions successfully implemented within one system. Coders attributed poorly tted elements to a broad construct of community characteristics. After coding all transcripts, community characteristics was the most frequently coded construct, often double coded to other outer setting constructs. We aggregated all quotes attributed to the constructs of needs and resources of those served by the organization, external policy and incentives, and community characteristics, recoded them inductively, then established an expanded set of outer setting constructs through consensus. Examples of quotes recoded to new outer setting constructs are included in table 2 in the appendix. These new constructs were double coded with the original constructs, and not treated as subconstructs or replacements of the original constructs.

Data Analysis:
Coder pairs, after agreeing on codes, used a process of data transformation to assign ratings for the in uence of each construct on intervention implementation through discussion guided by the CFIR rating rules. All text assigned to a particular code was reviewed by each member of the pair to assign preliminary ratings. The pair then met to reconcile any differences and agree on nal ratings. Ratings ranged from -2 (strong negative in uence) to +2 (strong positive in uence), and were not assigned by comparing communities, but by considering the relative in uence of each construct compared to other constructs scored for that community. Constructs were rated 'mixed' if the construct had both positive and negative in uences in relative balance; '0' if the construct was present but did not in uence implementation; or 'not rated' if the QIN indicated that it was not present or not su ciently known to determine in uence on implementation. We excluded two communities from further analysis because interviews did not elicit ratable descriptors of interventions, nor allow reliable assessment of coalition functions separable from provider efforts. Ratings were documented for each remaining community (n=20), along with a brief rationale for each rating. Once the rating process was complete, the larger team reviewed all ratings and rationales for the entire group of communities in open discussion to ensure consistency and agreement. Rationales were revised as needed for accuracy and clarity.
We used a convergent mixed method design by summarizing our transformed qualitative data with the RIR quantitative data in a matrix (table 3 located in the appendix and additional le 4). The matrix included all constructs, ratings and rationales for each community, with color coded ratings sorted by RIR performance for visual assessment of associations between RIR and the number of positive and/or negative in uencing CFIR constructs. To create a parsimonious set of constructs most relevant to the work, we elected to remove constructs that were rated for less than 40% of communities (8/20). We assigned the most extreme rating for any subconstruct to the main construct and used expert consensus to balance mixed ratings for multiple subconstructs for communities where this occurred. We retained any subconstructs, as well as added constructs, regardless of why or when they were added that were rated for at least 40% of communities (table 4 located in the appendix).
To further explore the relationship between implementation constructs and RIR performance, we tabulated codes most frequently co-occurring with "explanation of RIR performance." Results CFIR as a taxonomy for describing QIN-QIO efforts to coordinate care We found that the CFIR provided an appropriate taxonomy for characterizing the QIN-QIO care coordination efforts and their associated contexts. We note some modi cations and additions made to the CFIR to adapt to some of the unique aspects of the QIN-QIO role and the community-based nature of the work.
Of the 41 CFIR constructs and subconstructs de ned in the online codebook template (25), all were scored for at least one community, and 30 (73%) were coded and rated in at least eight (40%) communities. The CFIR framework allows for adaptation in selecting the most applicable domains or constructs for a particular inquiry, or in adding and adjusting constructs as appropriate. After adding coalition constructs, three evidence strength & quality subconstructs, and seven outer setting codes, our codebook had 70 possible constructs. Of those, 47 were scored in at least eight (40%) of the 20 communities (table 4 located in the appendix).
The domains that QIN-QIOs found most relevant to their coalition's efforts were outer setting, where all constructs (100%) were consistently scored, and process, for which nine out of ten (90%) constructs and subconstructs were scored. Characteristics of individuals was the least applicable domain, with only one construct consistently detected and rated.
The most frequently applied outer setting constructs were external policy and incentives, needs and resources of population served by the organization, and community characteristics. After inductive coding of text originally coded to community characteristics, we found the newly added sub-construct healthcare market characteristics to be applicable to all 20 communities.
Among constructs in the added coalition domain, coalition structure and coalition implementation climate were scored for all communities. QIN-QIOs described many features of coalition structure, such as composition, meeting frequency, maturity, and use of subgroups like workgroups or committees.
Finally, the constructs related to characterizing individuals in the characteristics of individuals domain and the functional roles in the process engaging subconstructs proved challenging for us. Because individuals were engaged to serve roles within a coalition as well as within their inner settings, we were often unclear on how to assign the engagement roles as suggested by the CFIR process subconstructs (e.g., champion versus key stakeholder). In interviews, we heard many compelling stories about a speci c person in uencing implementation efforts, but there were no consistent elements in the roles those people served within the coalition, their organizations, or the broader community, nor in the activities that they performed. Additionally, QIN-QIOs were seldom able to provide the depth of personal information about these individuals that would be necessary to reliably ascertain characteristics such as stage of change or self-e cacy. We therefore did not further assess constructs within the characteristics of individuals domain and recommend future work to better de ne important roles for engagement within a multi-provider coalition-facilitated program.

Constructs associated with care coordination intervention implementation
We found apparent relationships between the in uence of certain constructs and performance on RIR. High RIR communities had more constructs rated as positive in uences (138/182), as compared to low RIR communities (117/216) and low RIR communities had more constructs rated as negative in uences (66/216) compared to high RIR communities (14/182) (table 3 located in the appendix and additional le 4). Of the 70 constructs assessed, 14 distinguished high RIR from low RIR communities.
Coalition Domain: The quality of networks and communication was more often negative in low performing communities compared to at or high performers. One of the most valued aspects of coalitions generally was that they provided a rare opportunity for individuals in similar roles from different organizations (e.g., case managers) to interact, even in communities in which coalition participation did not seemingly facilitate implementation of interventions. Among lower performing communities, there were several examples of coalitions in which attendees remained grouped within their own organizations and interacted very little or shared little information.
The aggregated implementation climate construct and the sub-constructs of tension for change and the perceived relative priority of interventions within the coalition environment were also more likely to be negative in low performers. During interviews, we heard that many stakeholders, despite valuing coalition participation, did not necessarily have a sense of urgency to address readmissions or implement an intervention in their organizations. Reasons cited included low baseline readmissions rates, and/or other population factors being perceived as a higher priority for the community.
The coalition structure construct, despite being one of the most commonly scored constructs, did not distinguish performance. This construct incorporates a variety of characteristics, and it is likely that more detailed subconstructs should be established to tease out in uential differences.

Intervention characteristics:
The constructs within the Intervention Characteristics domain are more likely to be negative for low performing communities compared to others, although this relationship is less evident for the individual constructs.
Inner setting: The compatibility construct was less negative among high RIR communities compared to at or low RIR communities, indicating interventions well aligned with the values and work ows of providers implementing those interventions are more likely to result in lower readmissions.
Similarly, constructs of readiness for implementation were more positive among high RIR communities relative to low and at RIR communities, indicating that the capability of implementers to in uence readiness is an important component for success. The subconstruct of leadership engagement showed the strongest relationship within this construct, indicating that visible commitment and involvement of leadership was a distinguishing element of successful implementation within this project.
The engaging construct (including all subconstructs) had more positive ratings among high RIR communities compared to all others. Despite our challenges in assigning roles to those engaged, this provides evidence that an important characteristic of high performers is the capability to get the appropriate people involved in facilitating implementation. Conversely, being unable to engage the appropriate people is a signi cant barrier to implementation within a setting.
Outer setting: The outer setting domain constructs were also rated as negative in uences more frequently among low RIR communities, and positive in uences among high RIR communities. The construct community characteristics showed the strongest relationship. This code had been added inductively to capture notable features in a community as described in interviews and was rated in all but one of the communities. No low performing community received a positive rating for this construct. The similarly created new construct healthcare market characteristics also showed a strong relationship with RIR performance. Only one low RIR community had a positive rating for this construct, and two high RIR communities rated negatively for this construct. This may indicate that while healthcare market characteristics can be a signi cant barrier to implementation, it is not necessarily an insurmountable one.
Community characteristics associated with changes in 30-day hospital readmissions We tabulated codes that co-occurred with explanation of RIR performance, a code we added to capture response to a speci c interview question to elicit perceptions of what most in uenced successful reduction of readmissions in each community (table 5 located in the appendix).
Healthcare market characteristics was by far the most frequently co-occurring construct (14/20), and was more often cited by low and stable RIR communities (9 of 13), than high RIR communities (3 of 7). Other outer setting constructs, including population characteristics and physical features of the community also frequently co-occurred with RIR performance, though neither of these showed clear association with performance.
The implementation climate construct within the coalition domain also frequently co-occurred (7) with performance, with more high RIR communities (4) noting this than low performers (2). Inner setting implementation climate was also noted (5), though not as often, and not as strongly associated with community RIR. Example quotes associated with the most frequent co-occurring codes with RIR performance are shown in Table 6 located in the appendix.

Discussion
The CFIR provided a useful taxonomy for describing and capturing meaningful aspects of implementation in this complex, coalition-facilitated, multistakeholder initiative to reduce readmissions incidence. The conceptual domains and constructs of the CFIR were readily accepted as natural categories, and generally intuitive, to QIN-QIO staff who provided technical assistance through stakeholder coalitions, and domains often overlapped with elements already being reported by QIN-QIOs. Our interviews, guided by these domains and constructs, were frequently perceived as helpful to interviewees in organizing program lessons, and cataloguing important aspects of community efforts to implement interventions: "And it's really made me think about the consistency, continuity of interventions, and how we easily were pulled off into these different directions. It would be interesting to see how this particular intervention and this particular community t in the framework that you're researching for this project. It's really helped us to think about things…" The elds of quality improvement and implementation science are conceptually distinct but with intersecting goals. Quality improvement has been characterized as beginning with an identi ed performance gap in a speci c healthcare system, then using small tests of change to drive development of intervention and implementation strategies simultaneously. Implementation science typically begins with an underused evidence-based practice, and studies and remedies organizational, contextual and other factors that impede the practice's implementation (26). In reality, the distinction is blurry as many quality improvement initiatives promote established interventions, and most implementation efforts involve signi cant, ongoing adaptation (27). Greater integration of formal implementation science tools into national quality improvement initiatives might advance both elds: quality improvement through standardizing implementation elements to support rigorous monitoring and evaluation designs; and implementation science through dramatically increasing the number of implementation examples available to study. Hundreds of implementation tests generated through national quality improvement initiatives, coded to an implementation taxonomy, could be used to test hypotheses (28) and generate distinct explanatory and predictive models.
Indeed, in 2012, a conference on Advancing Methods for Healthcare Quality Improvement Research (29) was convened to synthesize quality improvement research methods and generate ways to build on its strengths in the future. It was noted that for quality improvement efforts to make signi cant contributions to the health care services research literature, it would need to reach a "higher level of validity and value." In other words, methods need the rigor of current medical and economic research efforts. However, quality improvement efforts are not well suited to controlled trials; they take place in uncontrolled settings and contexts and are met with different levels of investment and commitment from organizational leadership and front line staff. As noted by Toulany and colleagues, what is needed are research methods that allow for and understand the impact of this variation, with more de ned and detailed descriptions of local contexts, to learn lessons from both high and low performing teams rather than controlling variable elements (30). The CFIR provides a useful and applicable framework by which to identify contextual features of a given setting and understand which ones matter and how.
If applied to national quality improvement campaigns it would result in a large database of standardized elements and outcomes for testing complex intervention and implementation relationships, even with a high degree of variation among individual projects.
Other quality initiatives have found the CFIR useful for informing rapid-cycle approaches to improving implementation (31). The CFIR aggregates common elements from many theories and frameworks and is therefore consistent with most conceptual models in dissemination and implementation research (16). It has su cient detail to capture a comprehensive spectrum of pertinent information and can be easily tailored to relevant aspects of a given implementation effort. We added three new codes for evidence strength & quality, seven to the Outer Setting domain, and 17 existing CFIR constructs to a new Coalition Setting domain to capture elements important to this community-based initiative. We believe our set of domains and constructs aligned to the CFIR is a promising start to a standardized tool suitable for guiding and assessing future coalition-facilitated, community-based quality improvement strategies. In addition, other quality improvement organizations have found that using CFIR may be useful to inform rapid-cycle approaches to improving implementation.
As this initiative aimed to foster change that would impact the whole Medicare population of the community, QIN-QIO staff had detailed and nuanced insights into many aspects of the community environment. As a result, they provided much information about outer setting contextual features, and we re ned these codes to ensure key details were accounted for. This was particularly evident with the new code, healthcare market characteristics, which was rated as in uential to implementation in every community interviewed and demonstrated a clear relationship between the nature of that in uence (positive or negative) and performance on RIR. Interviewees frequently referenced the presence of other initiatives (eg. an Accountable Care Organization) or other provider arrangements within the community as being in uential on how providers interacted with each other, and their subsequent capacity to prioritize interdependent improvement activity or resources for care coordination interventions. For example, many answers about what in uenced readmission rates over the course of the project contained descriptive information such as "there are three separate hospital systems in this community" or "all our hospitals are for pro t". One helpful addition to the outer setting domain would be a distinction between local outer setting in uences from non-local outer setting characteristics, as suggested by the Promoting Action on Research Implementation (PARiHS) framework (32). For example, general awareness of readmissions payment penalties is probably less in uential to a community coalition-led effort than having one or more coalition participants either receiving or being at imminent risk of receiving a penalty. Future efforts should seek to catalogue the types of healthcare market characteristics that seem important in greater detail and test their associations with both successful implementation and change in readmission rates. Constructs shown to in uence implementation and/or outcomes could then be organized into existing outer setting constructs or appended to a framework for describing community-based implementation efforts.
Our decision to create a new coalition domain acknowledges the interface where the QIN-QIOs, who were our data source, were most knowledgeable. In a single setting implementation, the inner setting is where the actors make change, and the coalition is the inner setting where QIN-QIOs were making change.
Results of coding existing CFIR constructs to the coalition domain re ects this, as the majority of constructs we coded as coalition constructs are de ned by the CFIR as inner setting constructs. Because many QIN-QIOs also had meaningful insights into provider settings where patients were directly impacted by implementation efforts, we retained the inner setting domain as a separate contextual level.
The CFIR is an evaluation framework designed to comprehensively assess factors relevant to successful program implementation in a speci c inner setting, but it was not designed to explain nor evaluate a dynamic program that incorporates many settings, participants, and interventions. Nevertheless, we found consistent practices that suggest future directions should focus on standardizing implementation strategies, data collection, and evaluation approaches for similarly complex initiatives. We found that all interviewed QIN-QIOs formed and supported coalitions; all supplied data and analyzed root causes of readmissions; and most communities had providers that successfully implemented interventions. Our results suggest that coalitions may be helpful for driving change among a group of providers by offering a networking and communication platform, and by fostering an implementation climate that enhances a sense of priority and urgency for solving the problem of readmissions. The subconstructs of readiness for implementation, leadership engagement and available resources are likely to matter most at the implementation point of action, that is, within the provider organizations carrying forward those activities.
Similarly, communities that more effectively reduced readmissions may have been better at engaging the right people in the right roles to succeed in implementing appropriate interventions. These communities also seem to have had more robust planning and execution of planned interventions.

Limitations
Our data is qualitative and obtained from interviews with QIN-QIO staff re ecting a single perspective within the communities from which we sought to learn about parameters of complicated interventions, relationships, and contexts. We did not interview provider staff. The information obtained and constructs identi ed were therefore biased towards what an external technical assistance agency could reasonably know about providers, their activities, and the attitudes and actions that they brought to the work. QIN-QIO interviewees may have wanted to present their communities and their work positively, leading to overestimation or underestimation of the in uence of any given construct. Quarterly trend data of readmission rates and RIR was openly shared by QIN-QIOs throughout the program, so both interviewers and interviewees knew the community's performance during interviews, transcription, coding and rating. Our ratings, rationales, and community summaries were not veri ed for accuracy by interviewees. This project aimed to create assessment tools for future similar efforts, therefore, our results were obtained after the project had ended. Additionally, contracts stipulated that each QIN-QIO work with enough communities to impact 60% of the Medicare FFS bene ciaries in their states. Inevitably we received more information about low performing communities because QIN-QIO staff were likely to have been more engaged in those communities, and furthermore required by the funding agency (CMS) to provide more detailed and frequent assessments of these communities. Our methods may have therefore systematically underestimated either positive or negative in uences among high performing communities. Although these limitations may have signi cantly biased inferences around the strength of association between any construct and either implementation or readmissions outcomes, they are unlikely to have in uenced the degree to which the CFIR served as a taxonomy.

Conclusions
The CFIR could form the backbone of a common taxonomy for complex, multi-stakeholder community-based care coordination efforts, as shown by our data and others' experiences (31). Establishing that a taxonomy would have the potential to accelerate detection and replication of best practices by allowing aggregation of insights from communities with a wide range of context and performance, as well as variations in improvement strategies. Future efforts should aim to apply our modi ed constructs to a larger group of community-based improvement efforts and thereby to validate or further adapt them. Once a taxonomy is established, formal implementation research techniques should test the impact of identi ed constructs and the interaction of constructs in explanatory and predictive implementation models.
"The CEOs and the corporations that own the hospitals, they're heads in the bed. You know the type that still have that old school culture in their walls and this comes from the top down, that comes from the CEOs that comes from the corporations. I think that we are late adopters in this area, and we realized that because we see the program, they're across the country and the one key feature for success in some of these programs is that it comes from CEOs that make this a priority. It comes from C-suite to make readmissions a priority and we just don't see that here." Community 7 "And also, too, what we were nding out was that the primary care physicians, some of the major ones were not accepting Medicare patients, new Medicare patients. And then, one of the problems in the coalition that we really had faced and work on was that of the physicians that were taking Medicare patients, none of them were taking call, so it ended up, I mean it was just…and then with the economic downturn they lost their mental health stuff, all their mental health." Community 16 "Many of the hospitals with this zip code fall into what they call safety-net hospitals. And I believe this was the time period where there was so many issues with funding and whether or not hospitals were going to be shut down. One hospital in particular, that is a part of the community, had intense turnover. So, I think that those represent some of the barriers." Community 5 Population Characteristics "This is a very rural community -It has struggled with many things in terms of food insecurity, housing, income, opioids. And in probably the rst year of our contract, this was the community that actually led the way to be an accountable community for healthcare. So, their community team sort of changed their focus from admissions and re-admissions to addressing what they felt might be preventable issues in terms of keeping people from the ED, but addressing population health." Community 3 "…particularly around behavioral health, it was sometimes hard to nd the right person to get to, and nding availability, especially when some of these patients were homeless, trying to coordinate how to get the patient connected to those resources." Community 10 "I know they approach their county leaders to address, also their city council to gure out a way of how they can expand some of the transportation. So that would be to look at the bus systems, also any type of taxi services. So there was some limitations that existed. But they were moving forward on trying to gure out how can we expand some of the bus systems for the county area?" Community 12 "…and so with this healthcare system, they now have in place a community health program. I might be wording it wrong, but within this community health program, they're now partnering with some of the food storage places, also with transportation folks. And then also with other programs that are available for patients, so that way when a patient gets referred to that community health program, they can work with the case managers and the patients to nd out how they can meet those needs, within a certain geographical area." Community 1