Implementation research is the scientific study of methods to promote the systematic uptake of clinical research findings into routine clinical practice, and hence to reduce inappropriate care. It includes the study of influences on healthcare professionals' behaviour, and methods to enable them to use research findings more effectively. Over the last decade a considerable body of such research has been both published and reviewed 12. This research demonstrates that several (variously complex) interventions can be effective in changing health care professionals' behaviour. However, these studies have substantial heterogeneity within interventions, targeted behaviours, and study settings that render generalisability problematic 3.

In order to optimise the number of costly and time consuming future trials that need to be conducted – and to enhance their generalisability, it is necessary to identify the 'active ingredients' in professional behaviour change strategies. Interventions could be effective at changing behaviour for two reasons: they may contain components that are always effective in changing any behaviour, or they may contain components that overcome specific barriers encountered in relation to a particular behaviour. Hence, two approaches are necessary to identify the 'active ingredients' in the complex interventions of implementation trials: 1) Develop an understanding of the factors underlying professional behaviour in order to identify what sorts of processes should be targeted by interventions (process modelling) 4, and 2) Develop an understanding of how the interventions work and can be optimised (intervention modelling) 5.

In intervention modelling, key elements of an intervention are manipulated in a manner that simulates a real situation as much as possible, and interim endpoints are measured rather than changes in professional behaviour or healthcare outcome. A typical interim endpoint is a stated intention to behave in a particular way. Intention has been defined as "indications of how hard people are willing to try, of how much effort they are planning to exert, in order to perform a behaviour" 6 (see page 181). Compared to large scale trials, modelling experiments have two potential advantages – smaller size and shorter timescales – whilst still offering experimental control. Although proof of concept work has demonstrated the feasibility of such studies 5, a prerequisite of this method is a predictable link between the interim endpoints and changes in actual professional behaviour. For the method to be valid, interim endpoints (e.g. measures of intention) must be predictive of real world outcomes. This is the case for behavioural intention in non-clinical populations as demonstrated by reviews of both observational and experimental studies. Godin and Kok 7 reported averaged correlations between intention and different health-related behaviours ranging from 0.35 to 0.56 (i.e. intention was accounting for between 12% and 31% of the variance in behaviour). Armitage and Connor 8, using 63 independent studies reporting prospectively measured behavioural data, reported that the Theory of Planned Behaviour (TPB) variables that directly influence behaviour (intention and perceived behavioural control) accounted for a similar proportion of the variance in behaviour. When behavioural measures were self-reported, the TPB accounted for more of the variance in behaviour than when behaviour measures were objective or observed. A meta-analysis of 10 meta-analyses by Sheeran 9 reported that intention accounted for almost one-third of the variance in behaviour. Finally, Webb and Sheeran 10 reviewed experimental studies to relate change in intention to change in behaviour. From a meta-analysis of 47 experimental tests of the intention-behaviour relationship, they concluded that a "medium-to-large" change in intention leads to a "small-to-medium" change in behaviour. None of these reviews 78910 specifically identified studies of healthcare professionals and clinical behaviours.

These reviews demonstrate that there is a reliable, but not a perfect, relationship between stated intention and behaviour. Considerable research efforts have been directed to addressing the 'intention-behaviour gap' and two approaches have been proposed. One addresses the variability of the link by focusing on moderators of the intention-behaviour relationship, such as intention certainty and attitudinal versus normative control 11. According to this approach, it is possible to predict which individuals will enact their intentions (e.g. those whose intentions are attitudinally controlled). A second approach focuses on mediators of the intention-behaviour relationship, or processes that might be regarded as 'post-intentional,' such as implementation intentions 12, action plans and coping plans 13. This approach identifies processes that assist individuals to enact their intentions, thereby minimising the size of the intention-behaviour gap.

It has been argued 14 that the intentions and behaviour of clinicians are influenced by measurable psychological variables (e.g. attitudes) in the same way as the intentions and behaviour of any individual. However, most psychological models of behaviour are predicated on the basis of perceived consequences of a behaviour being experienced by the actor (e.g. If I give up smoking, I will improve my physical fitness). For clinicians, the perceived (or actual) consequences of their clinical behaviours are often (though not always) experienced by another person (e.g. If I lower my patient's blood pressure, she will have a reduced risk of premature mortality). Furthermore, clinical behaviour is constrained by a number of factors, such as imperatives dictated by a professional role, legal responsibilities, and principles of clinical governance. Although in a theoretical context these may be seen as control factors that influence behaviour together with intentions, some of these factors may not be articulated by clinicians. As they always form part of the context of any clinical behaviour, they may be implicit but powerful influences on intentions and behaviour, or the relationship between them. Therefore, it is important to explore the intention-behaviour relationship relating to clinical behaviour.

This review addresses the following two questions: 1) What is the nature of the relationship between measures of intention and clinical behaviours in clinicians? and 2) How does this compare to the intention-behaviour relationship in studies of non-clinicians?

This review found 10 studies examining the relationship between intention and clinical behaviours in 1623 health professionals. The proportion of variance in behaviour explained by intention was of a similar magnitude to that found in the literature relating to non-health professionals and corresponds to a medium to large effect 29. This was more consistently the case for studies in which intention-behaviour correspondence was good and behaviour was self-reported for the studies. However, it did also apply to studies with poor correspondence between intention and the measure of behaviour.

The literature assessing the correlation of intention and behaviour in health professionals is small compared to that available for non-health professionals. Sheeran 9, for example, presents data for 82,107 subjects from studies of non-health professionals. The considerably smaller literature for behaviour in health professionals makes it hard to draw firm conclusions based solely on these studies. However, despite the potential constraints on health professional behaviours, this review, viewed in the context of the larger populations of studies, provides encouragement for the contention that there is a predictable relationship between the intentions of a health professional and their subsequent behaviour.

Our review highlights a number of methodological issues within the available literature: the lack of experiments, the methods of measuring behaviour, and the reporting of studies. All of the studies that we found were observational. Unlike Webb 10 we found no reports of experiments that examined the relationship between changes in intention and changes in behaviour. Whilst it is not possible to comment on this relationship in health professionals, it would seem prudent to assume that the same pattern of results would apply and that although we have grounds to support there being medium to large effects in correlational studies, these may well be smaller when the process of change is evaluated within experimental designs.

Our results highlight the major challenge in the practicalities of measuring healthcare professional behaviour. Using self-reported measures of behaviour (compared to measuring actual behaviour) is relatively quick, cheap and easy. Moreover, it allows us to avoid having to deal with the logistics of gathering observed or recorded behaviour data and makes it straightforward to ensure good correspondence between measures of intention and behaviour (a central facet of operationalising theories).

However, healthcare systems are not interested in changing health professionals self-reported behaviour; they want to change actual behaviour in the expectation that this should lead to improved patient outcomes. A major advantage of using theory to design interventions to change the behaviour of health professionals is that it offers a generalisable framework with which to work 3031. When building theory-based interventions, outcomes, like intention or self-reported behaviours, can be useful proxies for actual behaviour 532. However, self-reported behaviour has the drawback of not accounting for the range of external factors (e.g. organisational or patients factors) that could be important effect modifiers of a heath professional's behaviour, and any estimate of the potential effect of an intervention based on self-reported behaviour is likely to overestimate its impact 8.

Whilst it may be relatively straight forward to define the behaviours of interest and be able to word a measure of intention, the choice of measures of behaviour can be more problematic. If direct observation is close to a (usually expensive) gold standard then all other methods of measuring behaviour usually involve some degree of compromise, either in terms of the correspondence between the measured behaviour and the behaviour of interest – or in terms of the feasibility of data collection. Three studies in this review that did not find a significant correlation between intention and behaviour did not attempt to measure behaviour through direct observation, but used an alternative measure – number of prescriptions 21, chart review 25, and patient report 22. These have advantages; for example, two of them are routinely available and therefore relatively cheap to obtain. However, it is more difficult to guarantee the one-to-one correspondence necessary between a measure of intention and a measure of behaviour when the latter is drawn from a potentially unreliable source. Under such circumstances it is not clear whether the absence of a correlation between intention and behaviour is a true lack of correlation or is due to measurement error. For instance, the study using chart review as a data source for teaching behaviour 25 makes the implicit assumption that all teaching activity is recorded, but this may not have been the case.

There was considerable variation and discrepancy in how the studies reported their findings. From the perspective of the systematic reviewer, if no other, it would be helpful if there was an agreed format of presenting the results of such studies. This could relate to issues around the theories used (i.e., clear statements of how each construct was operationalised), study conduct (i.e., clear statement of the timelines relating to administration of measures of intention and behaviour), and analysis (i.e., reporting of Pearson correlations between all variables that are entered into multiple regression analyses).

The findings from this review of health care professionals are broadly consistent with those found in the non-health professional literature. Intention appears to be a valid proxy measure for behaviour for use in the development of implementation interventions. However, there remain significant methodological challenges.

The author(s) declare that they have no competing interests.

All authors contributed to the conception, design and analysis of the study and approved the submitted draft. ME, JF, EK and SH reviewed the articles and abstracted the data.