Using Normalization Process Theory in feasibility studies and process evaluations of complex healthcare interventions: a systematic review

NPT is a theory of implementation that focuses on what people—both individuals and groups—do rather than what they believe or intend, and it has been built up from studies of practice in many different healthcare systems. This means that it focuses attention on aspects of individual and collective behaviour shown to be important in empirical studies of implementation processes. The development of NPT first involved the iterative development of a robust generic theory of implementation [5‐9, 11, 12]. From this, tools were developed to assist implementation practitioners and researchers [13‐16] in thinking through and measuring important elements of implementation processes. In its most recent iteration, we have shown how the basic mechanisms characterised in NPT function as self-organising mechanisms in complex adaptive social systems [10]. Theory development in NPT has been iterative, with three phases of development around practical questions.

Underpinning these practical questions is one that is fundamental to the social and behavioural sciences—and especially to behavioural economics, sociology and social psychology—which is how can we best understand the dynamics of human agency under conditions of constraint [10]? The important implication of this question is that well-designed, theoretically informed studies in implementation research actually offer opportunities for basic investigations in the social sciences.

A review by McEvoy et al. [17], published in 2014, provided a qualitative synthesis of 29 early and heterogeneous studies in which NPT was used. It drew attention to a positive response from healthcare researchers to the theory, but it also made three important critical points about the emerging NPT literature. McEvoy et al. [17] pointed to the ways that early studies using NPT did little work to justify the choice of theory, called for the prospective application of NPT to data analysis and collection and stressed the importance of moving beyond single stakeholder perspectives.

In the period since McEvoy et al.’s review [17], studies using NPT have proliferated. There are now a large number of protocols, empirical studies and reviews in which NPT plays a role. Importantly, a large number of NPT studies have now been completed by groups who are independent of the theory’s architects. It is therefore an opportune time to undertake a qualitative systematic review that will (i) identify and characterise the uses and limits of NPT in research on the implementation and integration of healthcare interventions and (ii) explore NPT’s contribution to understanding the dynamics of these processes.

As the aim of this qualitative systematic review was to identify the uses of NPT in research on the implementation and integration of health care interventions since the publication of the first iteration of the theory in 2006, our search strategy was focused on citations. Following Kirk et al.’s review of reports of the Consolidated Framework for Implementation Research [18], we searched two bibliographic two databases (Scopus and Web of Science), and a search engine (Google Scholar), to search for citations of key papers that developed or expounded the main constructs of NPT [5‐9, 11, 12], papers that developed NPT related methods or tools [13‐15] and citations of the NPT web-enabled on-line toolkit (www.​normalizationpro​cess.​org) [16]. Searches were conducted by AC, MG, CRM, MM and TLF. The sensitivity of the search strategy was tested against a database of studies using NPT that had been collected by three of the co-authors (CRM, TR, TF). All studies already known to use NPT at December 2015 were identified by the first round of systematic searches. Searches were initially undertaken in June 2015 and were updated in December 2015, August 2016 and March 2017. A final search was undertaken in December 2017.

We included the following: peer-reviewed English language journal articles reporting empirical research on the implementation of healthcare interventions, in which NPT was the primary analytic framework (applied either prospectively in study design and data collection, or retrospectively in the interpretation of already collected data) and which were undertaken in any healthcare setting. We define an empirical paper as one that contains evidence of data collection and analysis. We included studies that used any method of empirical investigation (qualitative, quantitative, and mixed methods).

We excluded the following: papers in which NPT was used as a framework for systematic reviews or meta-syntheses; papers solely on patient and caregiver experiences; papers in which NPT was not the primary analytic theory; editorials, theory and methods discussion papers; papers containing passing references to NPT; study protocols; papers describing work undertaken in settings other than healthcare; and papers published in languages other than English. We also excluded theses or dissertations, books and book chapters, conference proceedings and abstracts. We did not exclude papers on the grounds of methodological quality. We already knew that the literature ranged from student projects through to process evaluations in large and well-designed clinical trials in which NPT informed all activities from design through process evaluation and follow-up, to interpretation of trial outcomes. All studies were equally interesting to us, because we were searching for information about the way in which the theory was used rather than the summative results of NPT analyses.

Screening started with an assessment of citations and abstracts’ relevance by reviewers who had not been involved in the development of NPT (AC and MM). Reports that met eligibility criteria were obtained in full text. Full-text papers were screened by pairs of reviewers (AC with MM or CRM; MB with CRM; or CRM and TF) working independently of each other. Full-text screening consisted of identifying papers where NPT was clearly the analytic framework for an empirical study. Because no ‘one best way’ to operationalise NPT and its constructs has been prescribed, we did not apply judgments about this to screened papers. This meant that screening involved a simple Yes/No question, and references were sorted within Endnote Libraries accordingly.

We developed an extraction instrument, (see Additional file 1: Appendix 1). Data were extracted by all authors except CMM, FSM and EM. To avoid conflicts of interest, authors or co-authors of included papers were not involved in extracting data from those papers. Data were extracted on authors, year of publication, health care problem addressed, study type and methods, data collection procedures, how NPT was used in the study and whether this had been pre-specified in the study protocol. We looked for data on whether and how NPT had contributed to understanding the dynamics of the processes of implementation and integration, and for authors’ views about the limitations of NPT in terms of both its scope (what the theory explains) and application (what happens when researchers use the theory). As this was a qualitative review, we included data from both the results and discussion sections of included papers.

Coding and initial interpretation work was undertaken using the extraction instrument. To ensure consistency, CRM and TLF jointly checked coding on 75/130 of included papers, and CRM and CMM jointly checked categorisation of all included papers. The analysis aimed (i) to identify and characterise the uses and limitations of NPT in research on the implementation and integration of healthcare interventions and (ii) to explore NPT’s contribution to understanding the dynamics of these processes. Hence, we started by describing how NPT had been used and subsequently analysed the data to explore the ways that mechanisms defined by NPT have been revealed to operate. We sought to understand the relative importance of specific NPT constructs across different settings (core processes and mechanisms) and differences that seemed to apply in relation to different intervention types and healthcare systems (contingent processes and mechanisms).

PROSPERO deemed this review ineligible for public registration on the grounds that NPT was not a healthcare intervention.

Searches revealed 3322 citations. In Fig. 1, we show that after eliminating 2337 duplicates and broken or junk URLs, 985 were screened as titles and abstracts. Of these, 101 were excluded because they did not fit the inclusion criteria for the review. This left 884 articles for full-text screening. Of these, 754 did not fit the inclusion criteria for the review. This left 130 papers presenting results from 108 identifiable studies to be included in the review.

In this review, 130 papers reported the application of NPT in 108 identifiable studies. Included articles presented both controlled (n = 26) and uncontrolled (n = 82) studies.

In Table 1, we show that NPT was employed in 26 controlled studies—mainly complex intervention trials—and these generated 40/130 (30.8%) articles [19‐58]. These included an intervention design study (n = 1), feasibility studies (n = 5), process evaluations (n = 19) and retrospective documentary analyses (n = 1), embedded in complex intervention trials. Three of these studies used mixed methods, and one [55] was a survey. The remainder (n = 22) all used qualitative methods.

In Table 2, we show that NPT was employed in in 82 uncontrolled studies, and these generated 90/130 (69.2%) articles [59‐148]. These included feasibility studies (n = 20) and process evaluations (n = 54), and seven were what we have called ‘field studies’ which focused on general conditions in which interventions might take place, rather than the progress of specific interventions. One study was an ethnography of a set of socio-technical practices [103]. Qualitative methods were used in 72 studies. Of the remainder, seven were mixed methods studies, two were surveys, and one was a prospective cohort study.

Studies included in this review fell into seven categories. The most numerous group of studies were those concerned with service organisation and delivery (n = 29, 26.9% [23, 27, 32‐35, 43‐46, 58, 76, 79, 82, 84, 86, 89, 91, 92, 99, 105‐107, 110, 115, 116, 119, 122, 127, 133‐136, 140, 146, 148]). For example, in the UK, Grant et al. [34, 35] evaluated a complex intervention aimed at reducing risk in prescribing in primary care. They used NPT in ‘identifying and describing the components and sub-components of the intervention’ to understand ‘the nuances associated with collective implementation’. The next most numerous group of studies focused on the implementation of diagnostic and therapeutic interventions (n = 28, 25.9% [19, 24‐26, 28‐31, 36, 37, 47, 48, 55‐57, 67, 69, 78, 80, 90, 95, 97, 98, 103, 104, 111, 117, 123, 126, 137, 142, 143]). For example, in the USA, Hoberg et al. [98] examined the implementation of a new form of group therapy for people with mental health problems, while Leon et al. [48] showed how provider initiated testing and counselling for HIV was successfully normalised in a South African setting. Studies of implementation of E-Health and telemedicine—including telephone advice—were also numerous (n = 21, 19.4% [21, 22, 49, 59‐62, 66, 71, 73‐75, 83, 85, 87, 88, 93, 112, 114, 118, 125, 128, 130, 145]). Here, a Norwegian team led by Wilhelmsen et al. [145] showed how problems of participation and action—and especially the interactional workability—of a service providing internet-based cognitive behavioural therapy led to ambivalence on the part of general practitioners about its use, to low levels of follow-up and to doctors reverting to ‘standard treatment’ [145]. Less numerous (n = 11, 10.1%, were studies of the implementation of screening and surveillance tools [38‐42, 53, 64, 65, 72, 77, 113, 129, 132, 139, 144]). In a feasibility study, Ahmed et al. [64] showed that integrating a family history questionnaire about common genetic diseases into the workflow of primary care was unlikely without significant changes to the pattern of GP-patient interactions, and these were unlikely to be supported by clinicians. Such professional factors also affected the outcome of studies of decision support and shared decision­making (n = 8, 7.4% [81, 96, 100, 108, 109, 120, 121, 124, 131, 147]). In this category, in the USA, Scalia et al. [124] compared the implementation and integration of decision support tools between two major healthcare systems. This study raised important questions about how the interactions between clinicians’ (micro-level) experiences of the workability of complex interventions and meso-level organisational processes through which reflexive monitoring mechanisms play out their effects. Some studies were also explicitly concerned with implementing change in professional roles (n = 7, 6.5% [20, 54, 63, 70, 94, 101, 138]). For example, Thomas et al. [56, 57] showed how changes in roles and workload interacted to promote the routine embedding of an intervention intended to manage incontinence in stroke patients. Finally, a small group of studies were concerned with guideline implementation (n = 4, 3.7% [50‐52, 68, 102, 141]). Here, Vest et al. [141] described a study in the USA of the implementation of guidelines for the management of chronic kidney disease in primary care. They asserted that NPT could not only identify key barriers to practice but could also guide intervention choice.

Studies included in this review were mainly focused on reporting the implementation of complex healthcare interventions. Most of these studies had a translational component and made some claim about the evidence underpinning interventions. This evidence was heterogeneous and included qualitative studies [120, 121], implementation appraisals [133], meta-ethnographies [137] and previous trial results [38‐41]. However, the most common appeal to an evidence base in studies included in this review was through references to systematic reviews and rigorously developed clinical guidelines. Across the studies included in the review, 64/108 (59.2%) were linked to such support by their authors [149‐201]. As Tables 1 and 2 show, systematic reviews and rigorous guidelines were cited in support of 17/26 (65.4%) controlled studies and 47/82 (57.3%) uncontrolled studies.

As Tables 1 and 2 show, in 54/108 (50%) of the studies included in this review, the use of NPT appeared to have been planned in advance, and this was included in the study protocol. Amongst controlled studies, 19/26 (73%) of studies made explicit reference to including NPT in study protocols, while only 35/82 (42.7%) of uncontrolled studies did so. Not all papers offered a justification for using NPT. For the most part, authors characterised NPT as a conceptual framework that explains implementation processes and thus structures study design and data analysis. For example, Brooks et al. [23] justify it thus:

‘Normalisation Process Theory (NPT) has been used to consider complex interventions prior to the development of a randomised control trial to test their effectiveness (…). It has also been used in the context of mental health to explore the impact of new forms of collaborative care on the way in which professionals carry out their routines of work in primary care (…). The four constructs (coherence, cognitive participation, collective action and reflexive monitoring) permit a means of appraising factors that might ‘promote and inhibit the routine incorporation of complex interventions into everyday life’ (…). It focuses on the work that people need to do to ensure interventions become ‘normalised’. As a heuristic framework it can support the optimisation of a trial intervention at three points:

• supporting intervention design
• describing the context of a trial
• supporting the interpretation of a trial’s results’ [23].

Other papers reflected in more general terms on NPT’s empirical grounding (e.g. [28, 50, 52, 67, 73‐75, 87, 114, 122]) and its usefulness in thinking about implementation design (e.g. [27, 33, 67, 106, 147, 202, 203]).

In all but one study in the review [84], there was evidence that implementation outcomes could be explained by reference to the mechanisms specified by NPT. For example, Scalia et al. [124], state that their study

suggests that patient decision aids that are specifically designed for use in clinical encounters can be embedded in clinical settings, provided there is agreement about the need to use them, that the team members are willing to work together to make sure that such tools can be integrated in existing work patterns, and understood as making a positive overall contribution to the work that has to be performed. These considerations match the mechanisms of the NPT, which provides an explanatory framework for understanding the sustained use of these tools by the two systems examined. The motivation for the use of the Option Grid at CapitalCare was their wish to achieve success in an external quality improvement initiative. At HealthPartners, implementation efforts were motivated by a ‘champion’ physician. The nursing staff also played a pivotal role by systematically identifying eligible patients and providing those patients with the relevant encounter tool. These organizations, in different ways and to different degrees, exhibited coherence, collective action and cognitive participation that supported the sustained use of the tools. The organizational appraisal, in other words, their reflexive monitoring, was positive overall, despite concerns about readability and time pressures.

(Part omitted)

Implementing patient decision aids into clinical settings is a difficult process (…) In the UK, an implementation program known as MAking Good Decisions In Collaboration (MAGIC) highlighted the need for an organizational coherence, i.e. a widely held and agreed understanding of SDM principles in order to facilitate the implementation of patient decision aids (…). Commitment at multiple organizational levels has been recognized as an important precondition for implementation (…). This lack of commitment was noticeable at the CapitalCare sites that did not use patient encounter tools [124].

Differences between participant groups were characterised in 69/108 (64%) studies and between settings in 36/108 (33%) studies. For example, Clarke et al. [26] placed this in the wider context of levels of analysis.

‘This paper briefly considers implementation theories in respect of complex interventions and provides an overview of process evaluations to set the context for the study. We draw on Normalisation Process Theory (NPT) (…) as a conceptual lens through which to explore those features of the implementation process that were intended to secure practice change and to engage caregivers in the program. We also consider the interaction between influential macro and micro contextual factors that affected delivery by multi-disciplinary stroke unit staff and suggest that prior focus on generative mechanisms identified within NPT can be used to inform implementation processes within complex healthcare settings’ [26].

NPT thus characterises core elements of implementation processes and the factors that shape them, and using NPT enabled researchers to explain the ‘work’ that is involved in implementation. Implementation involves interactions between mechanisms and contexts that are highly complex and emergent. Dynamic elements of context can exercise powerful constraints on action. The sources of these constraints included system-level processes that structured behaviour (e.g. the role of fee for service payments in undermining the implementation of self-care programmes [43]) and micro-level conflicts within contexts (e.g. disagreements over participation and intervention legitimacy [20, 25]).

Implementation processes in NPT are explained by the operation of social mechanisms that motivate and shape collective action. Researchers using the theory employed its constructs in four distinctive ways, irrespective of the iteration of the theory that they used. We show examples of these diagrammatically below. First of all, some researchers clearly found it helpful to see the theory as describing a linear process in time [22, 63], in which the operation of mechanisms followed sequentially from each other (Fig. 2). In these studies, sense-making was seen as a necessary precursor to participation, and a degree of cognitive participation was required before collective action—in the form of an actual implementation process—could take place. Reflexive monitoring was seen as the final stage in the implementation process. However, research reported in this review often focused on feasibility studies or on the early stages of implementation life cycles in process evaluation. This skews their analyses towards the implementation phase of studies rather than their embedding and integration in everyday practice.

As Fig. 3 shows, the focus on the ‘front end’ of studies leads to an emphasis on ‘coherence’ and sense-making work as an obligatory point of departure for implementation processes (e.g. [27, 31, 50, 52, 57, 74, 85, 108, 113, 123, 131, 203]), sometimes at the expense of other activities. Figure 4 suggests a novel analysis of the relationship between mechanisms. Holtrop et al. examined the operation of components of collective action in the restructuring of provider reimbursement. Their analysis emphasised the role of relational integration as a precondition for normalisation. In this context, the operation of one mechanism might be an obligatory point of passage for the others. Holtrop et al. [99] state that

‘We found that effective care management normalization required relationship development between practice providers and staff and the care manager. Since identification and referral of patients needing care management was key to care management happening at all, the practice personnel understanding and appreciating the care manager role through a relationship with the care manager was critical. This was captured well through the NPT collective action component of relational integration. We interpreted relational integration to be the professional relationship development that occurred when care manager, providers and practice staff work together and understand and appreciate each other’s roles and contribution to patient care. Although it is its own component in NPT, we found it to be more of an outcome that occurred when the other components worked well (contextual integration, skill set workability and interactional workability). (…) We found that when any of the other components were not in place, there was also a lack of development of trust around shared patient care. Since care management is a relationship rich endeavor, the lack of this relationship is a key factor in care management’s disuse’ [99].

Finally, as Fig. 5 shows, NPT assumes that its constituent mechanisms can operate simultaneously—but unevenly—rather than sequentially. Few studies in this review tracked the implementation of a complex intervention over its whole life. When they did, they tended to present summative rather than a formative accounts. An interesting example of a longitudinal study may be found in work by Tazzyman et al. [148] that depicted NPT in precisely these terms. They state that the mechanisms specified by NPT are

non-linear and interact dynamically to provide a comprehensive explanation of the implementation processes. NPT was designed to be applied flexibly, can be used at one or more points in a qualitative study, has been successfully used beyond its original field and provides a robust theoretical framework to understand the dynamics of implementation [148].

Tazzyman et al. [148] explored the processes that underpinned revalidation of medical practitioners in a qualitative study of senior decision-makers undertaken at three time points (2011, 2013 and 2015). They characterised respondents in their study in NPT terms as ‘sense-makers’ and then explored the process of implementing and embedding of revalidation as a broad policy initiative. They state that their contribution

has been to extend the use of NPT to explore the implementation of a broad and complex policy, with wide ranging implications for an entire profession, and the wider healthcare system. Much previous work using NPT in healthcare has addressed the implementation of micro level interventions. This expanded application of NPT has highlighted a number of factors which seem to have affected the implementation of revalidation. The four dimensions of the framework (see Table 3) had an intuitive relevance and provided a useful explanatory framework for understanding the implementation of revalidation. There is scope to apply NPT more widely to complex social interventions and policy initiatives at the organisational and system level in future [148].

More usually, longitudinal studies using NPT were process evaluations embedded in large complex intervention trials. As we have noted above and elsewhere [10], these permitted a more structured analysis of implementation processes and their motivating mechanisms over time [29‐31, 38‐41, 43‐46, 50‐52].

Researchers used two main strategies to translate the constructs of NPT into practically useful analytic tools. Some used deductive strategies that relied on framework or directed content [204, 205] analyses and in which interpretation of data was structured prospectively by the theory. These approaches often took the form of relating data to matrices of varying degrees of complexity. In Table 3, we show how Alverbratt et al. [65] have created a detailed matrix in which they reinterpret and operationalise all constructs and sub-constructs of NPT. This partly replicates the way that these were originally characterised in May and Finch’s account of NPT [7]. This approach defines and sets out all of the constituent elements of the work that drives implementation processes and permits data collection and coding using framework, or directed, content analysis [205]. The approach taken by Alverbratt et al. focuses on translating the content of the theory into practical research questions in a very precise way. Others focused on the main constructs of the theory prospectively, but within a more flexible framework. In Table 4, we show how Røsstad et al. [122] set out a matrix that links theory constructs to a description of data collected and in Table 5, we show how Nordmark et al. offer an even simpler data matrix, in which core constructs are linked to data collection opportunities [116]. Tazzyman et al. [148] used an analytic approach included both deductive and inductive elements.

A coding framework was developed using the four domains and sub-domains of NPT by using an adapted version of the NoMAD instrument (part omitted), which was developed to assess implementation processes (Normalization Measure Development is an instrument designed for assessing the implementation of complex interventions). The adapted NoMaD instrument was applied to the transcripts by coding evidence of the sub-domains in Dedoose [206]. Following coding, two members of the research team (AT and JF) analysed the data across the three interview stages, using the constant comparative method, in order to understand changes and continuities over time. The inductive method of constant comparison analysis involved searching within individual transcripts, making comparison between transcripts within the same cohort, and comparing transcripts from different cohorts for conceptual similarities and differences. This method was combined with the deductive approach of using the four domains on NPT as a framework for the analysis.

Tazzyman et al.’s hybrid approach enabled them to develop a theory-led analysis, without needing to force data into a rigid theoretical framework. However, many studies took a more straightforward inductive approach to data collection and analysis. When studies collected and analysed qualitative data inductively—in the light of NPT—rather than deductively using framework approaches, there was less pressure on them to interpret their qualitative data within an inflexible coding framework. For example, in Table 6, we show how Bamford et al. [20] described the ways that their inductively generated data categories mapped on to NPT constructs. This group of papers includes a group of highly illuminating studies across the life course of complex intervention trials. Bamford et al.’s [20] process evaluation of the CAREDEM trial, and Kennedy et al.’s [43‐46] account of the WISE trial explain how structural factors militated against processes of cognitive participation. In their longitudinal accounts of the MOVE [38‐41] and STEPPING-UP [29‐31] Trials, Hooker et al. and Furler et al. show how mechanisms of coherence, cognitive participation and collective action interact to support the embedding of complex interventions in practice. Importantly, these studies also showed that the intervention remained in play once the trials themselves had concluded.

Critique of NPT as a theory was rare amongst the papers included in this review. However, it was not absent. For example, Clarke et al. [26] criticised an over-emphasis on agency at the expense of implementation contexts in NPT.

‘While May et al (…) acknowledge that the NPT generative mechanisms are in dynamic interaction with local contexts and external drivers, the framework primarily addresses the mechanisms. Indeed, the theory tends to place undue emphasis on individual and collective agency without explicitly locating this within, and as shaped by, the organisational and relational context in which implementation occurs’ [26].

Segrott et al. [126] take this further. They point to what they perceive as a focus on the agency of those involved in implementation, as opposed to those who experience the effects of that agency.

‘ENPT places considerable emphasis on the notion of implementation as an expression of agency. However, the agents in question appear to be mainly conceptualised as professional practitioners (e.g. nurses), rather than the participants who receive interventions. There is scope to consider further how the key constructs of ENPT can be applied to understand how participant (and non-participant) agency may shape whether interventions become integrated and embedded within delivery systems’ [126]

Beyond this, Alharbi et al. [63] criticised NPT for presenting a normative model of implementation that paid insufficient attention to idealised temporal aspects of implementation, a point echoed by Alverbratt et al. [65]. Critique was more often about the interaction between theory and method. Some articles (9/108) observed that NPT constructs overlapped, that the technical vocabulary of the theory was difficult and that as a result coding qualitative data was difficult [39, 44, 48, 59, 64, 67, 82, 99, 207]. Problems of this nature seemed less evident when researchers used a more inductive approach to qualitative data analysis (e.g. [25, 26, 38‐41]) than they did when authors employed a framework approach (e.g. [39, 99]).

In this review, we identified 108 discrete studies of complex healthcare interventions and related implementation processes. These studies were reported in 130 journal articles published after 2008. In papers included in this review, researchers collected and analysed their data in ways that effectively provided a basis (i) for intervention design and implementation planning and (ii) for understanding the dynamics of implementation, embedding and integration. Three key results of the review are as follows:

The use of NPT has coalesced around two main types of study: feasibility studies and process evaluations. However, unlike McEvoy et al.’s [17] review of NPT studies, we found that authors were justifying their choice of theory, and NPT was more frequently embedded in study protocols and thus being operationalised prospectively. However, concerns raised by McEvoy et al. about the lack of prospective application do not just apply to NPT. For example, Kirk et al. [18] point to the problem of low levels of prospective use of the CFIR [208] and PARIHS [209] frameworks. They point to the additional problem of lack of integration of theory into implementation research. Against this background, our review suggests that—although some authors have experienced difficulty with NPT’s technical vocabulary—users of NPT appear to be able to operationalise its concepts in consistent, stable ways to inform their work, and we can see evidence of theoretical integration in four kinds of studies.

These different approaches to mobilising theory suggest that NPT’s users have developed flexible explanatory strategies, and we have pointed to some of these in Figs. 2, 3, 4 and 5. In earlier papers [5‐9, 11, 12], we have argued that theories are conceptual toolkits that can be used flexibly to deal with practical problems. This means that there is no definitive ‘right way’ to employ NPT. It can be used on its own or in combination with other theories in ways that are locally defined to solve problems in intervention design and evaluation.

This review contributes to the literature on the incorporation of theory in implementation research, the benefits of this incorporation and the problems that can arise as a result. There are, of course, limitations to the review. Searches were undertaken in two databases, so it is possible that some studies were missed. It is questionable whether this would have altered the main findings and conclusions. Because Google Scholar is a search engine, and not a database, results of searches using it were not stable. Searches on Google Scholar also identified multiple versions of the same reference (e.g. versions of the published paper on publisher’s websites, records on institutional repositories, versions on personal websites and on academic social media sites such as Academia.edu and ResearchGate.Com). Sorting these involved significant additional work. It did however identify about 20 papers that would not otherwise have been included in the review. One paper was brought to our attention before appearing in any databases; however, subsequent updated searches did identify this paper. Equally, we excluded studies published in languages other than English, although there is some evidence that this is unlikely to be a major limitation [211].

We deliberately followed a two-step approach to data analysis, first identifying and characterising the use of NPT in implementation research, and then exploring the contribution made by NPT to understanding the dynamics of the processes of implementation and integration, and the limitations of its use. The characterisation is likely to be replicable by another team, but it is possible that a different group of researchers, with different backgrounds and different prior experience of NPT, would reach different conclusions. We have maximised the robustness of our findings by following a transparent process for analysis, including NPT-naïve researchers in the team, and holding frequent discussions amongst the team during the analysis. Finally, we made a deliberate decision to focus solely on the health care literature, and in light of this decision, our findings only apply to research on implementation in health care.

Most papers in the review used the elaboration of NPT published by May et al. [7, 8] in 2009. More recent iterations of the theory have focused on (a) the important role that social structural and social cognitive features of context play in mobilisation for implementation [9] and (b) the ways in which implementation processes demand that their participants negotiate with other actors and elements in the context in which they are set [10]. In these papers, we have already gone some way to answering the critique of Clarke et al. [26] on the relationship between agency and context. The critique offered by Segrott et al. [126], however, focused on the experiences of different groups of actors in implementation processes. They saw NPT as primarily being about the agency of professionals, rather than the experiences of patients and other participants in implementation processes. NPT both can be, and is, applied to those groups. We have developed theory in this area to explore the relationship between the implementation of complex interventions and burden of treatment (e.g. [212‐215]), and there is now a discrete body of primary research literature (e.g. [216‐220]) and systematic reviews (e.g. [221‐223]) that utilises these theoretical perspectives to understand patient and caregiver experience.

NPT has developed iteratively. Future work to develop it will explore variations in the ways that NPT mechanisms motivate and shape implementation processes across and between settings, and between micro, meso and macro levels of activity. This will engender a comprehensive ‘whole system’ approach to understanding implementation processes. Future empirical research will also help us to explore and test the hypothesis that collective action mechanisms operate cumulatively and that some mechanisms are more significant than others in determining implementation process outcomes. Rigorous quantitative research will assist in this, but until recently, there has been no robust instrument through which quantitative investigations of NPT mechanisms could be done. However, the NoMAD instrument is now available to perform this task [224]. This will make possible both large-scale and comparative quantitative and mixed methods studies that will provide important insights into the role of NPT mechanisms and the form and direction of implementation processes. This should lead to rigorous statistical models of NPT mechanisms at work and so to new insights about implementation processes. Finally, despite attempts to make NPT more user friendly through the development of explanatory toolkits, some users have difficulty with its technical vocabulary. NPT training packages are now coming on stream that will help to solve this problem [225].