Select and specify the target behaviour/s

The first step is to identify the behaviour(s) that need to be changed to address the implementation problem. In the contexts in which they are performed, the key behaviours are often interdependent with other behaviours within the individual and with behaviours of others. Other attributes include the inherent complexity of behaviour, including whether it is performed by individual healthcare professionals or by healthcare teams, and the frequency of opportunities for performing the behaviour.

The next step is to specify these behaviours in terms of who needs to do what differently, when, where, how and with whom? If there are multiple behaviours, it is helpful to start with one or, possibly, two behaviours to target in the first instance. The criteria to consider when prioritising behaviours include (i) how modifiable it is likely to be and (ii) how central it is in bringing about the desired change in clinical practice; (iii) the ‘spillover’ effect, i.e. the positive or negative effect on other related behaviours if change occurred (known in the literature as conflicting and facilitating behaviours) [45]; and (iv) the amenability to measurement. Selection is usually influenced by a thorough assessment based on a range of sources of information about the problem and careful examination of evidence-based recommendations and empirical research, both published and local. There are inevitable trade-offs in prioritising behaviours for investigation.

The more precisely the behaviour is specified, the greater the specificity of the facilitators and barriers identified. There are three aspects of this process: (1) decide the appropriate level of behavioural specificity; (2) identify who performs the behaviour, when, where and how; and (3) consider the attributes of the target behaviour such as complexity, action sequences and interdependence of team-level behaviours. We explain each of these aspects below, using examples from the implementation research literature to illustrate each point.

Behavioural specificity

There is a balance between being highly behaviourally specific (to maximise the likelihood of identifying barriers to and facilitators of that behaviour) and being general enough to be relevant to a range of contexts. For example, to investigate the management of diabetes in primary care, a more specific description is ‘general practitioners measure the blood glucose levels of their patients with diabetes every 6 months’ whereas a less specific description is ‘general practitioners managing their patients with diabetes according to guidelines’. The more specific description is more likely to identify the sources of implementation problems that need to be changed because it is clear what the behaviour is, who needs to perform the behaviour and how often it is performed. Thus, study findings are more likely to be interpretable if the behaviour targeted for change is defined carefully in terms of who needs to perform the behaviour, what they need to do, when they need to do it, where they need to do it, how often they need to do it and with whom will they need to do it [43, 46, 47]. Furthermore, it is important that the behaviour be specified in terms of target behaviour, e.g. GPs to advise patients with sore throats to take painkillers and drink plenty of cool or warm fluids, rather than the problem behaviour, e.g. GPs prescribing antibiotics for sore throats. There are cases where it is not possible to isolate and target one behaviour for change, for example if designing an intervention to help GPs improve diabetes control; there are more than 10 interdependent behaviours that could be targeted for change. One way of addressing this challenge is to prioritise two or three key behaviours. The example in Table 3 illustrates the specification of a professional behaviour according to the principle of behavioural specificity. It may be that a goal has been set, e.g. reducing infections in a particular setting, but the behaviours required to achieve that goal are not immediately obvious. In these cases, analysis of audit data and discussion with stakeholders can support the identification of relevant behaviours and agreement on target behaviours.

Select the study design

As with all research, the appropriate study design depends on the research question and the state of knowledge in the given field of research. For example, qualitative interviews may be more useful when little is known about an implementation problem and the study design allows researchers to probe in greater detail providing richer data which can be helpful when developing theory-informed interventions in that they may provide better insight into needed content of interventions. They are also likely to be useful to understand the mechanism of action in interventions. Survey studies may be more relevant when more is known about the problem and potentially relevant influencing factors, but the aim is to identify those factors/domains predictive of behaviour change in a more representative sample, or to explore mechanism of action of interventions (mediation analyses). Structured observation and approaches such as documentary analysis may be useful to supplement interview/survey studies (data can be triangulated), but they are unlikely to be sufficiently comprehensive for all domains (for example cognitions are not observable or documented). As the TDF has largely been applied at exploratory and formative stages of a research programme to inform problem analysis and intervention development, most reported work have used qualitative interviews (one-on-one or focus groups) to elicit health professionals’ perceptions of TDF-related barriers and facilitators. However, the TDF is potentially applicable to other research designs for which methods can be further developed, e.g. structured observations, documentary analysis, case study designs.

The TDF has been used in questionnaire studies (Table 4). There are three published validated questionnaire measures of the TDF to identify influences on the following behaviours: health care professionals’ patient safety behaviours [48], physical activity in the general population [20] and generic health professional behaviours [49].

The TDF also has the potential to inform systematic reviews by synthesising influences on behaviours across studies according to theoretical domains (Table 5) and understanding effect size (Table 6).

Decide the sampling strategy

The target population needs to include the target adopters of the behaviours and/or other relevant stakeholders. These could be individuals (e.g. clinicians, patients, students or members of the public), dyads (e.g. clinicians and patients; teachers and students) or teams (e.g. teams of clinicians and managers or business workgroups). The organisational level at which change is proposed to occur could be at different levels, e.g. individual, team, organisation or population levels. Change may need to be coordinated across different organisational levels [50], with different types of behaviours being enacted by a range of individuals or groups.

There are several challenges to collecting data for implementation research which need to be considered when deciding sampling strategy. First, studies have largely relied on self-report data and individuals may be biassed in their views about the problem and attribute failures to external (environment or other people) rather than internal (ability, effort) factors [51, 52]. Therefore, it is important to include multiple perspectives (e.g. from users, managers, commissioners as well as providers of health care) and, where possible, to use multiple sources of data (e.g. clinician self-report of influences on behaviour via interviews and/or surveys, practice and policy documents and direct observation of behaviour) [53, 54]. In this way, the validity of findings are likely to be improved through integration or ‘triangulation’. Triangulation is the ‘process of studying a problem using different methods to gain a more complete picture’ [55]. A number of triangulation techniques are available to researchers, and integration can be carried out at the analysis and/or interpretation stages (for an overview of methods, see O’Cathain et al. [55]).

Although sample size for interview and focus group studies can be determined by the sampling procedure (such as purposive sampling for maximum variation) and the implementation problem under investigation, specifying a minimum sample size a priori is recommended. Francis et al. recommend that a minimum of 10 interviews be conducted for initial data analysis, followed by three additional interviews until no new theme emerges (stopping criterion) [56]. Sample size will also depend on whether the study involves different groups of health care professionals and whether they are being analysed together or as separate groups. If multiple groups are involved and the plan is to analyse them separately to get varying group perspectives, 10 interviews plus three per group are advised as a minimum (Francis et al. [56]). Table 7 illustrates the selection of professional stakeholder groups to achieve maximum variation in the sample of a TDF-based qualitative interview study. Focus groups involving all stakeholders have the potential to provide multiple perspectives and potentially reduce the tendency to focus only on external influences on behaviour. Our recommendations would be a minimum of 3 groups if the focus is on a specific care setting.

Develop interview schedule

Interview schedule

As in all interview studies, a key step in a TDF-based interview study is the development of an interview schedule. We advise using language relevant to the target population and piloting schedules to check comprehension. The schedule typically consists of an open question for each theoretical domain to elicit the first response, followed by a series of follow-up prompts to probe more deeply. Each question focuses explicitly on the target behaviour. The TDF was developed to promote a comprehensive consideration of possible influences on a given behaviour so there is no specific order in which the questions should be asked. We recommend flexibility in the order in which domains are covered to harness the natural flow of the conversation if a respondent volunteers’ information relating to a domain not yet covered. The number of domains covered and number of questions within each domain depend on the target behaviour and existing evidence. For example, where previous research has established a domain is not relevant to a target behaviour, researchers may consider omitting questions relating to that domain and focusing more on exploring domains considered more relevant to the target behaviour. However, as with all qualitative research, coding using the TDF can only code the text in the interview transcript. If questions are not asked, the text cannot be coded. Researchers will have to determine the value of including all domains to ensure coverage whilst balancing the evidence surrounding the target behaviour. An in-depth understanding of the theoretical content of domains and context of the implementation problem will help ensure interview schedules elicit a maximum amount of useful information. Published TDF questionnaires may be helpful in developing interview schedules [20, 48, 49]. Example questions to explore domains in implementation research taken from Huijg et al. [49] are provided in Additional file 2.

Collect the data

Data can be collected using the TDF by one or more of the following: interviews, surveys, observation and documentary analysis. For example, in a study to inform implementation of a hospital care pathway to reduce sepsis mortality, the TDF was used to gather data by interviews with the care team, observation of ward staff and analysis of hospital protocols. Behaviour change techniques identified through observation and interview were then mapped to TDF to identify mechanisms of action [54]. Interviews can be conducted in groups or individually either face-to-face or by telephone. Our experience has been that interviews typically last on average between 25–45 min for one-on-one interviews and 50–90 min for focus group interview but is of course dependent on the number of behaviours being investigated.

As with all interviewing, follow-up questions are the key to eliciting a good understanding of the ways in which the domains contribute to the target problem or could be used to bring about change. For example, a question such as ‘how confident are you in doing x?’ with a follow-up probing question ‘what has made you confident?’ allows for the participant to be specific as they reflect on their confidence but also gives them an opportunity to reflect on situations that limit their confidence. Anchoring discussion to the target behaviour(s) can help to keep discussion focused and avoid a drift into general issues. As with developing study materials, interviewers with an underlying understanding of the theoretical constructs underpinning domains will promote appropriate probing during interview.

Analyse the data

The TDF is intended for use by researchers and practitioners from many disciplines. Whilst users do not necessarily need expertise in using particular theories, a good understanding of the domains and the theoretical constructs each represents are recommended to aid interpretation of data. Data can be analysed deductively, using the TDF to generate the framework for a content analysis and, inductively, generating themes that can then be considered in relation to domains. Some research teams have used this approach as the basis for designing predictive questionnaires to collect quantitative data to test out hypotheses generated by the qualitative analysis (Fig. 1). Intervention designers can select behaviour change techniques either directly from identified relevant theoretical domains using validated linkages [7, 42] or by linking to the Behaviour Change Wheel to guide the selection of intervention functions, policy categories and behaviour change techniques [43].

Develop a coding guideline

A coding guideline is a set of explicit statements of how the TDF is to be applied to a specific data set. Statements provide guidance on strength of confidence that a piece of text indicates a domain where change is likely to be helpful in changing behaviour.

The coding guideline should be developed at the same time as the interview schedule and updated iteratively during data collection. Independent coding by two people allows discrepancies to be discussed and coding guidelines refined until acceptable reliability is achieved between coders. Coding difficulties may arise because the interviewer has not sufficiently probed to clarify how responses relate to the domain under investigation. Reviewing transcripts during the data collection period rather than at the end will allow interviewers to refine the interview schedule as more understanding of the problem being studied is gained. Many uncertainties by the coding team can be eliminated if the interviewer is familiar with the implementation problem under investigation, the TDF and published studies using it.

Deductive analyses

Coding interview transcripts into theoretical domains: Coding begins by reading participants’ responses in the transcript, considering their relevance to the definitions of the domains and/or the constructs within the domains and then attributing them to one or more domains. This directed content analysis technique is guided by theory and/or relevant research findings to interpret meaning from the content of qualitative data for initial codes [57]. New users are advised to ensure that all coded texts relate to the target behaviour and not other behaviours interviewees or focus group participants may discuss but which are not relevant. Whilst the domains are purposively design to be broad groupings of the possible factors to influence behaviour, the intent is to explore the important domains in further detail. Some text may seem to fit in multiple domains. For example, everything a HCP does can be dependent on context so everything can be coded in environmental context and resources. However, this is a somewhat simplistic assumption and other domains permit the division of the contextual factors influencing behaviour and identify those that are amenable to change (i.e social context reflected in social influencing, reinforcement likely delivered by the context, beliefs about capabilities which are typically situation-specific and organisations which can make certain actions easier or more difficult). Text should be coded into the domains that best reflect the key theme, despite the inclination to code everything into one domain. Users with no or limited experience with the TDF should initially meet frequently to discuss coding and challenges to address concerns early in the process. Once the users are comfortable with the TDF, it is recommended that two researchers code data independently into theoretical domains following a mutually agreed coding guideline to increase the reliability of coding.

Disagreements in coding are not uncommon, and discussions to resolve them can be informative both about the process and the substantive questions addressed by the study. In order to facilitate consensus among coders, we recommend that the coders articulate their understanding of the coded text (i.e. key meaning) and justify their rationale for selecting the domain. Justification for why the text should not be coded in the alternate domain should also be discussed, with each coder given the opportunity to discuss the other’s point. When consensus cannot be reached, discussion with an expert in the area being studied and an experienced researcher who is familiar with the TDF and the theories from which it draws can help guide coders to interpret the text in relation to the domains and the theoretical constructs within the domains. When agreement on assigning text to a single domain cannot be reached, consensus can be achieved by assigning the text to all the domains identified by both the coders. Coders should document which text is attributed to which theoretical domain/s. Table 8 illustrates this step.

Data saturation: Data saturation is reached when the data collected do not contribute any new information about barriers and facilitators influencing the implementation problem. If data are collected progressively, concurrent with analysis, this will inform final sample size [58]. A number of factors can influence saturation including the design or scope of the study [59], the heterogeneity of the population [60] and the nature of the implementation problem [61].

Reliability: Reliability between the two coders may be assessed by an inter-rater reliability coefficient, for example through assessing a kappa score across all domains [28]. Calculating simple percentage agreement can also be used to establish agreement among coders [62]. Reliability between two coders is acceptable if kappa score > 0.6 or percentage agreement > 60% is achieved [63]. PABAK kappa which corrects for negative agreement when Cohen’s kappa is marginal can also be used.

Statistical software: NVivo software, or other qualitative analysis package, may be used for data analysis. It can enhance analysts’ efficiency at data storage, retrieval, coding, editing and revising coding, organising data and sharing files across researchers and can be used to assess reliability.

Inductive Analyses

Generating themes and/or belief statements: For thematic analyses, researchers are encouraged to follow established methods [64, 65] and report explicitly the processes used to analyse the data.

After coding data into theoretical domains, some researchers have used the following methods to further analyse the data within domains: generate overarching themes for a number of responses with similar underlying ideas and/or generate statements of specific underlying beliefs for each response [26]. The overarching themes represent the factors which are perceived to influence performance of the target behaviour. A belief statement is a collection of responses with a similar underlying belief that suggest a problem and/or influence of the beliefs on the target implementation problem [25]. For example, these responses, ‘guidelines are just guidelines’, ‘guidelines are not gospel’ and ‘there are no rules about going outside guideline’, were grouped under the belief statement ‘I can make my decision outside the guidelines’ [26]. For efficient use of time, one coder can generate belief statements and the other coder can interrogate and confirm those. In the example given above reported in Francis et al. [15], this step resulted in a list of belief statements supported by responses made in the interviews within each theoretical domain. Each belief statement was counted once within each interview to generate a frequency count across all interviews. This step in the analysis results in a list of belief statements with frequency counts for each of the belief statements and/or overarching themes within each theoretical domain. Frequency count of belief statements is not warranted in the case of focus group interviews as nonverbal behaviour such as nodding in agreement with a belief statement from another participant would not always be captured. Also, social influence effects may elicit more agreement with such a statement than would be identified in one-to-one interviews.

Challenges in coding data

Sometimes there can be uncertainties when coding data into theoretical domains, a feature not uncommon in qualitative coding. This should not impede progress of the study. Coding into theoretical domains may require a certain amount of interpretation of theoretical constructs by researchers.

Coding interview transcripts into some theoretical domains can be more challenging compared to other domains depending upon the implementation problems under investigation. A single response may involve more than one theoretical domain, and often, several domains are addressed in a single response to a particular interview question (see Additional file 3 for examples). An attempt to tease out different domains for the purpose of coding into appropriate theoretical domain may result in losing the context of the response. To avoid such occurrences, we recommend that the entire response be coded in all identified domains.

Sometimes responses may not clearly fit any theoretical domain despite guidance on interpretation by trained psychologists or those with a good understanding of the TDF from the study team. It is important to note such occurrences as these effectively test the capacity of the TDF to account for all the interview data. The most common reason for utterances not fitting into a domain is that it is not about the target behaviour. Coding difficulties then become an opportunity to check the adequacy of the behavioural specification. Hence, we recommend coding wider contextual information into a separate code for ease of retrieval and completion of descriptive summaries of participants and their practice characteristics.

Identification of relevant theoretical domains: In this step, relevant theoretical construct domains (i.e. domains that should be targeted in an intervention) are identified by judging the importance of specific beliefs or themes. The following three criteria have been applied in published studies: (1) relatively high frequency of specific beliefs and/or themes (not relevant to focus group interviews); (2) presence of conflicting beliefs; and (3) evidence of strong beliefs that may affect the target behaviour [28]. At the completion of this step, the researchers will have a list of relevant theoretical domains that are most likely to influence the target implementation problem and associated behaviours. Table 9 illustrates this point.