Clinical Performance Feedback Intervention Theory (CP-FIT): a new theory for designing, implementing, and evaluating feedback in health care based on a systematic review and meta-synthesis of qualitative research

1Providing health professionals with quantitative summaries of their clinical performance when treating specific groups of patients has been used for decades as a quality improvement strategy (Table 1). Such approaches may be called “audit and feedback”, “clinical performance feedback”, “performance measurement”, “quality measurement”, “key performance indicators”, “quality indicators”, “quality dashboards”, “scorecards”, “report cards”, or “population health analytics”. In this paper, we use the term “feedback” intervention to encompass all these approaches and to refer to the entire process of selecting a clinical topic on which to improve, collecting and analysing population-level data, producing and delivering a quantitative summary of clinical performance, and making subsequent changes to clinical practice.

2Feedback has been extensively researched in numerous quantitative and qualitative studies. However, despite its popularity, the mechanisms by which it operates are poorly understood. In this paper, we define mechanisms as underlying explanations of how and why an intervention works. Three consecutive Cochrane reviews have found feedback produces “small but potentially important improvements in professional practice” with wide variations in its impact: the most recent demonstrated a median clinical practice improvement of 4.3%, ranging from a 9% decrease to a 70% increase. When feedback interventions target suboptimally performed high-volume and clinically impactful practices, such as hypertension management or antimicrobial stewardship, this variation can translate to thousands of quality-adjusted life years.

3Policymakers and practitioners only have a tentative set of best practices regarding how feedback could be optimally conducted; thus there is a need to better understand how and why feedback works in order to maximise their impact. One approach is to consider the underlying theory of feedback, which has often been over-looked. In this paper, we define theory as a “coherent description of a process that is arrived at by inference, provides an explanation for observed phenomena, and generates predictions”. In the 140 randomised controlled trials in the most recent Cochrane review, 18 different theories were used in only 20 (14%) of the studies, suggesting a lack of consensus as to which is most appropriate for feedback. More recently, three theories have gained popularity in the feedback literature: Control Theory, Goal Setting Theory, and Feedback Intervention Theory. However, these theories address only part of the feedback process, and even if used in conjunction may still miss potentially important factors specific to health care (Table 2).

4Qualitative evaluations of quality improvement interventions can generate hypotheses regarding their effect modifiers (i.e. variables that influence success) and mechanisms of action. For example, by helping explain why a particular intervention was ineffective (e.g.), or developing a logic model for success (e.g.). Synthesising findings from different qualitative studies can help build theories of how interventions may be optimally designed and implemented. Such approaches have been used to improve interventions in tuberculosis therapy, smoking cessation, skin cancer prevention, and telephone counselling. A similar approach may therefore be useful for feedback and, to the best of our knowledge, has not yet been attempted.

5We aimed to synthesise findings from qualitative research on feedback interventions to inform the development of a comprehensive new health care-specific feedback theory. Informed by our definition of theory, our objectives were to (1) describe the processes by which feedback interventions effect change in clinical practice, (2) identify variables that may predict the success of these processes, (3) formulate explanatory mechanisms of how these variables may operate, and (4) distil these findings into parsimonious propositions.

6We published our protocol on the International Prospective Register of Systematic Reviews (PROSPERO; registration number CRD42015017541).

7We replicated the latest Cochrane review’s search strategy, adding qualitative research filters (Additional file 1). MEDLINE (Ovid), EMBASE (Ovid), and CINAHL (Ebsco) were searched without time limits on 25 March 2015. Citation, related article, and reference list searches were undertaken up to 31 December 2016 for all included studies, relevant reviews, and essays (e.g.). Further studies were found through international experts and Google Scholar alerts.

8Table 3 describes our inclusion criteria. Two reviewers independently screened titles and abstracts. Full manuscripts of potentially relevant citations were obtained and the criteria re-applied. Data from included articles were extracted independently by BB and WG regarding the study and feedback intervention details (Additional file 2; e.g. study setting, who provided the feedback, and what information the feedback contained). Critical appraisal was conducted concurrently using 12 technical and theoretical criteria including the appropriateness of data collection and analysis methods, adequacy of context description, and transferability of findings. Any disagreements were resolved through discussion, with the wider team consulted as necessary.

9Study findings were extracted as direct quotations from participants and author interpretations found in the abstract, results, and discussion sections. Data were synthesised in five stages (Fig. 1; please see Additional file 3 for details): coding excerpts from individual papers in batches using framework analysis and realistic evaluation, generalising findings across papers and building on pre-existing theories to formulate hypotheses, iteratively testing and improving these hypotheses on new batches of papers using Analytic Induction, assessing confidence in our hypotheses using the GRADE-CERQual method, and summarising high-confidence hypotheses into a core set of propositions.

10We screened 15,413 papers (Fig. 2). Sixty-five were ultimately included, reporting 61 studies of 73 feedback interventions involving 1791 participants, from which we synthesised 1369 excerpts. Table 4 summarises their main characteristics, full details of which are in Additional file 4.

11From our synthesis, we developed Clinical Performance Feedback Intervention Theory (CP-FIT). CP-FIT argues that effective feedback works in a cycle, the success of progressing round which is influenced by variables operating through a set of common explanatory mechanisms related to the feedback itself, the recipient, and wider context (Fig. 3). How these variables and mechanisms influence the feedback cycle is illustrated by 42 high-confidence hypotheses (Table 5), which are in turn summarised by three propositions (Table 6). CP-FIT draws on concepts from 30 pre-existing behaviour change theories (Table 7) and has over 200 lower confidence hypotheses (Additional file 5).

12We describe CP-FIT in detail below. To maintain readability, we focus on its high-confidence hypotheses and provide only key example references to supporting studies and theories. CP-FIT’s constructs are in italics. Table 8 provides example illustrative quotes and Additional file 5 contains the full descriptions of constructs, with references to supporting papers and theories. Additional file 6 provides case studies demonstrating how CP-FIT can explain the success of different feedback interventions included in the synthesis.

13Similar to existing feedback, goal setting, and information value theories, we found that successful feedback exerts its effects through a series of sequential processes, each of which required a non-trivial work commitment from health professionals (Fig. 3). This started with choosing standards of clinical performance against which care would be measured (Goal setting), followed by collection and analysis of clinical performance data (Data collection and analysis); communication of the measured clinical performance to health professionals (Feedback); reception, comprehension, and acceptance of this by the recipient (Interaction, Perception, and Acceptance respectively); a planned behavioural response based on the feedback (Intention and Behaviour); and ultimately positive changes to patient care (Clinical performance improvement). A further step of Verification could occur between Perception and Acceptance where recipients interrogated the data underlying their feedback (e.g.). The cycle then repeated, usually starting with further Data collection and analysis. Feedback interventions became less effective if any of the above processes failed, halting progress round the cycle. For example, if Data collection was not conducted (e.g.), or a recipient did not Accept the feedback they were given (e.g.; Table 8, quote 1).

14In addition to potentially improving clinical performance, we found both positive and negative unintended outcomes associated with feedback. Health care organisations often noted improved record-keeping (e.g.), and recipient knowledge and awareness of the feedback topic (e.g.). However, it could also result in: Gaming, where health professionals manipulated clinical data or changed the patient population being measured scrutiny to artificially improve their performance (e.g.), or Tunnel vision, where health professionals excessively focused on the feedback topic at the detriment of other clinical areas (Table 8, quote 2).

15We found four feedback variables that influenced progress round the feedback cycle: (1) the goal, (2) data collection and analysis methods, (3) feedback display, and (4) feedback delivery.

16This variable refers to the clinical topic and the associated clinical behaviours or patient outcomes measured by the feedback intervention. For example, the proportion of diabetic patients with controlled cholesterol in primary care, or whether nutritional advice is provided to nursing home residents. Similar to feedback-specific and general behaviour change theories, we found Acceptance and Intention more likely when feedback measured aspects of care recipients thought were clinically meaningful (Importance; Table 8, quote 3). Acceptance and Intention were also more likely when feedback targeted goals within the control of recipients (Controllability e.g.) and that were relevant to their job (Relevance e.g.).

17When undertaken by feedback recipients themselves (Conducted by recipients e.g.) or performed manually (Automation e.g.), we found the Data collection and analysis process was inhibited, often due to a lack of time or skills. In extreme cases, the Goal setting process was re-visited in order to find more suitable methods (e.g.).

18We found Acceptance was more likely when recipients believed the data collection and analysis process produced a true representation of their clinical performance (Accuracy), which often related to the positive predictive value of the feedback (i.e. its ability to correctly highlight areas of clinical performance requiring improvement). If perceived Accuracy was low, recipients were also more likely to undertake Verification (Table 8, quote 4).

19Likewise, Acceptance was facilitated when feedback recipients could exception report patients they felt were inappropriate to include in feedback (Exclusions e.g.). Potential reasons for exception reporting are discussed in the “Patient population” section.

20We found Intention and Behaviour were more likely when feedback communicated recipients’ performance level had room for improvement (Performance level). This violated their perception they delivered high-quality care, thus providing motivation and opportunity to change (e.g.). It also encouraged Verification as recipients often wanted to clarify this alternative view of their achievements themselves (e.g.). We found some support for theories that suggested the feedback process could be inhibited if performance was so extreme that improvement was unlikely: for example, non-Acceptance if current performance was too low (e.g.), or Goal setting re-visited if performance too high (e.g.); though these findings were inconsistent.

21Feedback that detailed the patients included in the clinical performance calculation (Patient lists) facilitated Verification, Perception, Intention, and Behaviour by enabling recipients to understand how suboptimal care may have occurred, helping them take corrective action (where possible) for those patients and learn lessons for the future (e.g.). It also facilitated Acceptance by increasing transparency and trustworthiness of the feedback methodology (Table 8, quote 5).

22Feedback focusing on the performance of individual health professionals rather than their wider team or organisation increased Acceptance, Intention, and Behaviour because, similar to Controllability and Relevance (see “Goal” section), it was more likely to highlight situations for which they had responsibility (Specificity e.g.). Using recent data to calculate recipients’ current performance (Timeliness) had a similar effect because it was based on what recipients could change currently, rather than events that had long passed (e.g.).

23Feedback often compared recipients’ current performance to other scores, such as their past performance (Trend e.g.), others’ performance (Benchmarking e.g.), or an expected standard (usually determined by experts; Target e.g.). We found that Trend facilitated Perception by helping recipients interpret their current performance in a historical context. Benchmarking worked in a similar fashion by helping recipients understand how they performed relative to other health professionals or organisations, stimulating Intention and Behaviour because they wanted to do better than their colleagues and neighbours. Benchmarking also worked by motivating recipients to maintain their social status when they saw others in their peer group behaving differently. These findings contradicted Feedback Intervention Theory, which predicts that drawing attention to other’s performance reduces the impact of feedback. It was unclear whether Benchmarking was more effective when the identities of the health professionals were visible to each other, or to which health professionals’ performance should be compared. We found only minimal evidence that Targets influenced feedback effectiveness despite their prominence in existing feedback theories.

24Feedback was more effective when it communicated the relative importance of its contents (Prioritisation) and employed user-friendly designs (Usability), because it reduced cognitive load by helping recipients decide what aspects of their performance required attention (e.g.). Studies provided little detail on how this could be practically achieved, though strategies may include limiting the number of clinical topics in the feedback (Number of metrics e.g.) or using charts (Graphical elements e.g.). We found insufficient evidence that feedback’s effectiveness was influenced by whether it was presented positively or negatively (Framing).

25Recipients often rejected feedback whose purpose they believed was to punish rather than support positive change because it did not align with their inherent motivation to improve care (Function e.g.). Similarly, when feedback was reported to external organisations or the public, it often drew negative reactions with little evidence of impact on clinical performance (External reporting e.g. anxiety and anger; Table 8, quote 6).

26Acceptance was also less likely when delivered by a person or organisation perceived to have an inappropriate level of knowledge or skill (Source knowledge and skill). This could relate to the clinical topic on which feedback was provided (e.g.) or quality improvement methodology (e.g.). We found inconsistent evidence that the location of feedback delivery, for example whether internal or external to the recipients’ organisation, influenced effectiveness (Source location).

27Feedback that was “pushed” to recipients facilitated Interaction more than those requiring them to “pull” it (Active delivery). For example, feedback sent by email (e.g.) was received more frequently than when published in a document that was not distributed (e.g.). An exception was feedback solely delivered in face-to-face meetings, as the significant time commitments often meant health professionals could not attend (e.g.).

28Feedback delivered to groups of health professionals improved Teamwork (see “Organisation or team characteristics” section) by promoting engagement and facilitating discussion (Delivery to a group e.g.). There was inconsistent evidence on the effects of how often feedback was delivered (Frequency), and little insight into whether it was best delivered electronically or on paper (Medium).

29We found two recipient variables that influenced progress round the feedback cycle: (1) health professional characteristics and (2) their behavioural response.

30Often health professionals did not possess the knowledge and skills to effectively engage with and respond to feedback. This included technical quality improvement skills such as interpreting data or formulating and implementing action plans, rather than the clinical topic in question (e.g.). We found interventions targeting those with greater capability (both technical and clinical) were more effective because recipients were more likely to successfully proceed through the Perception, Intention, and Behaviour feedback processes (Knowledge and skills in quality improvement and the clinical topic, respectively; e.g.). This seemed to undermine the rationale of interventions predicated on addressing health professionals’ presumed lack of clinical knowledge (e.g.; Table 8, quote 7).

31Understandably, health professionals with positive views on the potential benefits of feedback were more likely to engage with it (Feedback attitude e.g.). And although health professionals often had profound emotional reactions to feedback, both positive and negative (e.g.), we found no reliable evidence that these directly influenced the feedback cycle.

32We found two main types of action taken by recipients (if any) in response to feedback: those relating to the care of individual patients one-at-a-time (Patient-level) or those aimed at the wider health care system (Organisation-level). Patient-level behaviours included retrospectively “correcting” suboptimal care given in the past, or prospectively providing “better” care to patients in the future. For example, resolving medication safety errors by withdrawing previously prescribed medications versus optimising treatment when a patient with uncontrolled diabetes is next encountered. In contrast, Organisation-level behaviours focused on changing care delivery systems. For example, changing how medications are stored in hospital, or introducing computerised decision support software to support clinician-patient interactions. We found Organisation-level behaviours often led to greater Clinical performance improvement because they enabled multiple Patient-level behaviours by augmenting the clinical environment in which they occurred. For example, changing how medications were stored reduced the likelihood of delayed administration to all patients, and decision support software could remind clinicians how to optimally treat diabetic patients. Conversely, by definition, Patient-level behaviours only ever affected one patient (Table 8, quote 8). We found no clear evidence that feedback success was affected if it required an increase, decrease, change, or maintenance of recipients’ current clinical behaviours to improve their performance (Direction).

33We found three context variables that influenced progress round the feedback cycle: (1) organisation or team characteristics, (2) patient population, (3) co-interventions, and (4) implementation process.

34We found all organisations and teams had a finite supply of staff, time, finances, and equipment (e.g.), stretched by the complexity of modern health care, such as serving increasing numbers of elderly multimorbid patients and dealing with wider organisational activities such as existing quality improvement initiatives and re-structures (e.g.). Consequently, if an organisation had less capacity (Resource) or significant other responsibilities (Competing priorities), they were less able to interact with and respond to feedback (e.g.). However, if senior managers advocated for the feedback intervention or individuals were present who were dedicated to ensuring it was a success, they often influenced others and provided additional resource to enable more meaningful engagement with feedback (Leadership support e.g. and Champions e.g., respectively; Table 8, quote 9).

35Increased Resource and Leadership support also increased the likelihood that Organisation-level behaviours were undertaken (see “Behavioural response” section), because they often required process redesign and change management (e.g.). In turn, Organisation-level behaviours also had the potential to further increase Resource, for example by recruiting new staff (e.g.) or purchasing new equipment (e.g.), which in turn further increased their capacity to engage with and respond to feedback.

36Feedback was more successful when members of organisations and teams worked effectively towards a common goal (Teamwork; e.g.), had strong internal communication channels (Intra-organisational networks; e.g.), and actively communicated with other organisations and teams (Extra-organisational networks; e.g.). These characteristics often co-existed and provided practical support for feedback recipients during Interaction, Perception, Intention, and Behaviour (Table 8, quote 10).

37Organisations and teams also commonly had long-established systems and processes that were often difficult to change, such as methods of care delivery and technical infrastructure (e.g.). Therefore, if the feedback intervention fitted alongside their existing ways of working (Workflow fit), it required less effort to implement (e.g.).

38Health professionals felt it was inappropriate to include certain patients in the their clinical performance calculation. For example, patients that refused the aspects of care measured by the feedback (Choice alignment; e.g.), or those who already received maximal therapy or had relevant clinical contraindications (e.g. medication allergies; Clinical appropriateness; Table 8, quote 11). Including such patients in their clinical performance calculation inhibited Acceptance and Intention, with some evidence it may have also led to Gaming ( “The feedback cycle (research objective 1)” section e.g.).

39Synthesised papers used eight different quality improvement interventions alongside feedback (Table 4). However, only four appeared to impact feedback success because they addressed specific barriers. The provision of support for health professionals to discuss their feedback with peers (Peer discussion) and to identify reasons for and develop solutions to suboptimal performance (Problem solving and Action planning) facilitated Perception, Intention, and Behaviour. These co-interventions addressed shortcomings in health professionals’ quality improvement skills (see “Health professional characteristics” section). Peer discussion had the added benefit of improving Teamwork (see “Organisation or team characteristics” section). Such approaches often co-existed, and could be delivered in different ways, for example as didactic workshops (e.g.) or led by recipients themselves (e.g.), though it was unclear which was most effective (Table 8, quote 12).

40Co-interventions that provided additional staff to explicitly support the implementation of feedback helped overcome time and staffing issues (see “Organisation or team characteristics” section; External change agents). These personnel could either be directly involved in feedback processes (e.g. carrying out improvement actions), or indirectly support recipients (e.g. facilitating Perception and Intention).

41We found little support for education (Clinical education) or financial incentives (Financial rewards). There was some evidence that Financial rewards could negatively impact feedback success by conflicting with recipients’ motivation and sense of professionalism (e.g.).

42How feedback was introduced into clinical practice impacted all feedback cycle processes. Feedback tailored to the specific requirements of the health care organisation and its staff appeared more successful because it aligned with their needs and improved Workflow fit (see “Organisation or team characteristics” section; Adaptability). For example, if quality indicator definitions could be amended to fit existing data sources or focus on local clinical problems.

43When training and support were provided on how to use an intervention (not the clinical topic under scrutiny; Training and support), it improved recipients’ Knowledge and skills in quality improvement (see “Health professional characteristics” section; e.g.). Further, if the training demonstrated the intervention’s potential benefits (Observability), recipients were also more likely to engage with it. These benefits could be to recipients themselves, such as improved feedback user-friendliness (Usability, the “Feedback display” section; e.g.), or to patient care (e.g.). Trend (see “Feedback display” section) could also increase Observability if its trajectory was positive (Table 8, quote 13).

44Interventions considered “expensive” to deploy, in terms of time, human, or financial resources, were generally less effective because they required more resource or effort (Cost). Examples of expensive interventions included when data collection was Conducted by recipients (see “Data collection and analysis method” section; e.g.) or when feedback was delivered solely face-to-face (see “Feedback delivery” section; e.g.).

45We found more consistent evidence to support interventions that made recipients feel like they “owned” the feedback intervention rather than those imposed via external policies or directives because they harnessed their autonomy and internal motivation to improve patient care (Ownership). Despite this, we found little support for seeking input from recipients into the design and implementation of feedback (Linkage at the development stage).

46We found seven explanatory mechanisms through which the above variables operated. Many mirrored constructs from existing theories of context and implementation, and variables often effected change through multiple mechanisms (Table 5).

47Complexity is how straightforward it was to undertake each feedback cycle process. This could refer to the number of steps required or how difficult they were to complete. Simple feedback facilitated all feedback cycle processes.

48Relative advantage refers to whether health professionals believed the feedback had a perceived advantage over alternative ways of working, including other feedback interventions. Understandably, variables operating via this mechanism depended on the specific circumstances into which they were implemented. Relative advantage facilitated all feedback cycle processes.

49Resource match details whether health professionals, organisations, and teams had adequate resources to engage with and respond to those required by the feedback intervention. It included time, staff capacity and skills, equipment, physical space, and finances. When Resource match was achieved, all feedback cycle processes were facilitated.

50Compatibility characterises the degree to which the feedback interventions aligned with the beliefs, values, needs, systems, and processes of the health care organisations and their staff. Compatibility facilitated all feedback cycle processes.

51Credibility was how health professionals perceived the trustworthiness and reliability of the feedback. Recipients were more likely to believe and engage with credible feedback, which facilitated Interaction, Verification, Acceptance, Intention, and Behaviour.

52Social influence specifies how much the feedback harnessed the social dynamics of health care organisations and teams. Exploiting Social influence could facilitate all feedback cycle processes.

53Actionability describes how easily health professionals could take action in response to feedback and in turn how directly that action influenced patient care. Actionability facilitated Intention, Behaviour, and Clinical performance improvement.

54We distilled the above hypotheses of how context and intervention variables influenced feedback cycle processes (Table 5) into three propositions that govern the effects of feedback interventions (Table 6). Each proposition summarised multiple variable hypotheses, though only related to a mutually exclusive set of explanatory mechanisms.

55CP-FIT describes causal pathways of feedback effectiveness synthesised from 65 qualitative studies of 73 interventions (Table 4), and 30 pre-existing theories (Table 7). It states that effective feedback is a cyclical process of Goal setting, Data collection and analysis, Feedback, recipient Interaction, Perception, and Acceptance of the feedback, followed by Intention, Behaviour, and Clinical performance improvement (the feedback cycle; Fig. 3). Feedback becomes less effective if any individual process fails causing progress round the cycle to stop and is influenced by variables relating to the feedback itself (its Goal, Data collection and analysis methods, Feedback display, and Feedback delivery), the recipient (Health professional characteristics, and Behavioural response), and context (Organisation or team characteristics, Patient population, Co-interventions and Implementation process). These variables exert their effects via explanatory mechanisms of Complexity, Relative advantage, Resource match, Compatibility, Credibility, Social influence, and Actionability (Table 5) and are summarised by three propositions (Table 6).

56Each of Table 5’s 42 high-confidence hypotheses can be viewed as specific design recommendations to increase feedback effectiveness. For example, hypothesis 12 (Trend) recommends feedback should display recipients’ current performance in relation to their past performance; hypothesis 17 (Source knowledge and skill) recommends feedback should be delivered by a person or organisation perceived as having an appropriate level of knowledge or skill by recipients; and hypothesis 26 (Leadership support) recommends that feedback interventions should seek the support of senior managers in health care organisations when implemented. For practitioners and policy-makers, CP-FIT therefore provides guidance they should consider when developing and deploying feedback interventions. This includes national clinical audits (e.g.), pay-for-performance programmes (e.g.), and learning health systems (where routinely collected health care data is analysed to drive continuous improvement)—such programmes are large-scale, address impactful clinical problems (e.g. cardiovascular mortality or antimicrobial resistance), and require substantial expenditure to develop and maintain (e.g. data collection and analysis infrastructure). Using CP-FIT thus has the potential to improve care for large numbers of patients, in addition to reducing the opportunity cost from unsuccessful feedback initiatives and improving returns on health care systems’ financial investments.

57Table 5’s hypotheses can also be translated into explanations why feedback may or may not have been effective. Additional file 6 provides examples of how to do this by presenting three case studies of different feedback interventions included in our meta-synthesis, and using CP-FIT to explain their successes and failures. CP-FIT can therefore help researchers and feedback evaluators assess and explain feedback’s observed or predicted effects. Specifically for qualitative methodologists, Additional file 5 provides a comprehensive codebook that can be used to analyse data and discover causal pathways. For quantitative investigators, both Table 5 and Additional file 5 provide over 200 potentially falsifiable hypotheses to test. As illustrated in Additional file 6, CP-FIT may be particularly useful in process evaluations to identify weak points in a feedback interventions’ logic model (i.e. the feedback cycle; Fig. 3) and barriers and facilitators to its use (i.e. its variables).

58Although developed specifically for feedback, CP-FIT may also have relevance to other quality improvement strategies that analyse patient data and communicate those analyses to health professionals in order to effect change. Examples include computerised clinical decision support and educational outreach, where CP-FIT concepts such as Accuracy (see “Data collection and analysis method” section), Timeliness (see “Feedback display” section), Credibility (see “Credibility” section), and Actionability (see “Actionability” section) may all be important. CP-FIT concepts related to population-level feedback (e.g. Benchmarking and Trend; the “Feedback display” section) may be less relevant when the focus of the intervention is on individual patient-level care, such as in clinical decision support.

59Table 9 shows how CP-FIT may explain reasons for feedback effectiveness variation found in the latest Cochrane review. CP-FIT suggests further sources of variation not identified that could be operationalised in a future update of the review: for example, if feedback allows Exclusions or provides Patient lists (see “Data collection and analysis method” and “Feedback display” sections, respectively).

60CP-FIT aligns well with tentative best practices for effective feedback interventions and provides potential evidence-based explanations as to why they may work (Table 10). It also provides additional potential recommendations such as automating data collection and analysis (Automation; see “Data collection and analysis method” section) and gaining leadership support (Leadership support; see “Organisation or team characteristics” section). An advantage of CP-FIT over these existing best practice recommendations is that it provides parsimonious generalisable principles (in the form of its explanatory mechanisms and propositions; see “Mechanisms (research objective 3)” and “Propositions (research objective 4)” sections, respectively). Consequently, CP-FIT’s hypotheses can be extended beyond those in Table 5 if they conform to these constructs. For example, the Complexity (see “Complexity” section) of a feedback interventions’ targeted clinical behaviour (Goal; see “Goal” section) may be reasonably expected to influence its effectiveness, despite not being a consistent finding in our synthesis.

61Table 7 demonstrates how pre-existing theories contribute to, and overlap with, CP-FIT. In comparison to other theories used to model clinical performance feedback, CP-FIT adds value for health care settings by specifying potential unintended consequences (see “The feedback cycle (research objective 1)” section); detailing new context-related constructs, for example in relation to the organisation or team (see “Organisation or team characteristics” section); and elaborating on specific aspects of the feedback process, for example data collection and analysis (see “Data collection and analysis method” section). This wider and more detailed view may explain why CP-FIT occasionally provides different predictions: for example, Feedback Intervention Theory predicts the presentation of others’ performance (Normative information) decreases effectiveness by diverting attention away from the task at hand, whereas CP-FIT states it does the opposite by harnessing the social dynamics between recipients (Benchmarking; see “Data collection and analysis method” section).

62To our knowledge, a systematic search and synthesis of qualitative evaluations of feedback interventions has not been previously undertaken. However, two reviews exploring the use of patient-reported outcome measure (PROM) feedback in improving patient care have been recently published. Although neither explicitly attempted to develop theory, their main findings can be mapped to CP-FIT constructs. Boyce et al. found there were practical difficulties in collecting and managing PROMs data related to an organisation’s resources and compatibility with existing workflows (cf. CP-FIT Propositions 1 and 2, respectively; Table 6); whereas Greenhalgh et al. note “actionability” as a key characteristic in the effective use of PROM data (cf. CP-FIT Proposition 3; Table 6). Both noted the “credibility” of data and source from which it was fed back were essential to securing health professional’s acceptance (cf. CP-FIT’s Credibility; see “Credibility” section).

63Colquhoun et al. generated 313 theory-informed hypotheses about feedback interventions by interviewing subject experts. Many of the hypotheses appear in CP-FIT (e.g. feedback will be more effective if patient-specific information is provided cf. CP-FIT’s Patient lists; see “Data collection and analysis method” section), though some are contradictory (e.g. feedback will be less effective when presented to those with greater expertise cf. CP-FIT’s Knowledge and skills in quality improvement and clinical topic; see “Health professional characteristics” section). A possible explanation is that Colquhoun et al.’s hypotheses have been informed by disparate research paradigms (including those outside health care) rather than attempting to develop a unifying theory based on empirical evaluations of feedback interventions like CP-FIT. Work is ongoing to prioritise these hypotheses for empirical testing, which will also help further validate CP-FIT.

64Like all literature syntheses, our findings reflect only what has been reported by its constituent studies. Consequently, CP-FIT may not include features of feedback interventions or contexts associated with effectiveness that have been under-reported. This may manifest by such findings being absent, having “low” or “moderate” GRADE-CERQual ratings (Additional file 5) or unclear effects (e.g. Frequency; see “Feedback delivery” section or Direction, see “Behavioural response” section). For similar reasons CP-FIT’s current form may also lack detail regarding certain construct definitions (e.g. how is good Usability [see “Feedback display” section] or effective Action planning [see “Co-interventions” section] best achieved?), how particular variables may be manipulated in practice (e.g. how can we persuade health professionals of a feedback topic’s Importance [see “Goal” section] or to undertake Organisation-level as well as Patient-level behaviour [see “Behavioural response” section]?), and inherent tensions within the theory (e.g. how do we ensure Compatibility [see “Compatibility” section] whilst also attempting to change health professional behaviour?). Future research should address these evidence gaps by evaluating innovative new feedback designs delivered in different contexts, employing both robust qualitative and quantitative approaches, using CP-FIT as a framework.

65Finally, CP-FIT does not currently quantify the relative effect sizes of its variables and mechanisms. It is possible that variables appearing to influence feedback effectiveness with “high” GRADE-CERQual confidence may in fact have negligible effects on patient care. Consequently, future work should aim to quantitatively test CP-FIT’s hypotheses and refine its assumptions.

66The advent of electronic health records and web-based technologies has resulted in widespread use and expenditure on feedback interventions. Whilst there is pressure to provide higher quality with lower costs, the messy reality of health care means feedback initiatives have varying success. This results in missed opportunities to improve care for large populations of patients (e.g.) and see returns on financial investments. Feedback interventions are often as complex as the health care environments in which they operate, with multiple opportunities and reasons for failure (Fig. 3 and Table 5). To address these challenges, we have presented the first reported qualitative meta-synthesis of real-world feedback interventions and used the results to develop the first comprehensive theory of feedback designed specifically for health care. CP-FIT contributes new knowledge on how feedback works in practice (research objective 1) and factors that influence its effects (research objectives 2 and 3, respectively), in a parsimonious and usable way (research objective 4). CP-FIT meets the definition of a theory provided in the “Background” section because it (1) coherently describes the processes of clinical performance feedback (see “The feedback cycle (research objective 1)” section and Fig. 3), (2) was arrived at by inferring causal pathways of effectiveness and ineffectiveness from 65 studies of 73 feedback interventions (as detailed in Additional file 3), (3) can provide explanations as to why feedback interventions were effective or ineffective (as demonstrated by the case studies in Additional file 6), and (4) generates predictions about what factors make feedback interventions more or less effective (see hypotheses in Table 4 and Additional file 5). We hope our findings can help feedback designers and practitioners build more effective interventions, in addition to supporting evaluators discern why a particular initiative may (not) have been successful. We strongly encourage further research to test CP-FIT whilst exploring its applicability to other quality improvement strategies, refining and extending it where appropriate.

67Providing quantitative summaries of clinical performance when treating specific groups of patients (“feedback”) is a widely used quality improvement strategy, yet it has varying success.Theory could help explain what factors influence feedback success; however, existing theories lack detail and specificity to health care.This is the first systematic review and meta-synthesis of qualitative evaluations of feedback interventions and presents the first comprehensive health care-specific feedback theory that can be used to design, implement, and evaluate feedback (Clinical Performance Feedback Intervention Theory; CP-FIT).Using CP-FIT could help improve care for large numbers of patients, reduce opportunity costs from unsuccessful interventions, and improve returns on feedback infrastructure investment.

68Providing quantitative summaries of clinical performance when treating specific groups of patients (“feedback”) is a widely used quality improvement strategy, yet it has varying success.

69Theory could help explain what factors influence feedback success; however, existing theories lack detail and specificity to health care.

70This is the first systematic review and meta-synthesis of qualitative evaluations of feedback interventions and presents the first comprehensive health care-specific feedback theory that can be used to design, implement, and evaluate feedback (Clinical Performance Feedback Intervention Theory; CP-FIT).

71Using CP-FIT could help improve care for large numbers of patients, reduce opportunity costs from unsuccessful interventions, and improve returns on feedback infrastructure investment.