Over the last two decades, the application of randomised controlled trials and systematic reviews has extended to the evaluation of ever more complex interventions. A range of facets of the complexity of these interventions has been identified. While most health care interventions have some degree of complexity, interventions that include a number of components, which may be independent or inter-dependent, are at the more complex end of the spectrum . Examples include case management and discharge planning, both of which aim to minimise the fragmentation of health care ,. More recently, complex interventions of a more conceptual nature have been systematically reviewed, such as continuity of care or “trust” between doctors and patients ,. These present the additional challenge of defining concepts that are often poorly developed.
Compared with single evaluations of complex interventions ,, less attention has been paid to the methodological issues arising from the synthesis of data from complex interventions. Important limitations include difficulties in (1) defining the intervention within the review; (2) searching for and locating relevant evidence; (3) standardising the selection of studies for a review; and (4) synthesising data. In this Viewpoint, we describe the implications of these limitations and suggest some approaches to help systematic reviewers reflect on the conceptual and analytical challenges posed by these types of review.
1. Defining the Complex Intervention
To varying degrees, complex interventions can be standardised and defined in individual prospective studies. However, the lack of an agreed definition of complex interventions that have the same aim but are described differently, or inadequately, across studies poses inherent difficulties for systematic reviewers and users of these reviews . Case management illustrates this well. Despite there being no agreed typology, case management has become a generic concept across different specialities to improve inter-professional collaboration and co-ordination of care for individual patients. Case management comes in many forms, including a brokerage model, an integrated care pathway, a liaison service, and self-managed care . All these variations can occur alone or in combination.
Defining the intervention is made difficult when the core purpose of an intervention, such as behavioural counselling , varies according to the characteristics of the participants or the trial setting, or if the intervention aims to promote an abstract concept, such as promoting trust or continuity of care ,. In the field of service delivery, interventions become complex and difficult to define if, as is often the case, (1) they are delivered across the primary-secondary care interface; (2) they are delivered in new settings; or (3) there is an added behavioural dimension and staff perform new behaviours or current behaviours in a new context .
Solutions to improve the description and conceptual understanding of the content of a complex intervention include (1) typologies to guide the classification of interventions and (2) supplementary evidence, such as qualitative or descriptive data , (see Table 1).
Typologies can guide the classification of common elements of interventions into homogeneous groups. The Cochrane Effective Practice and Organisation of Care Review Group (EPOC) has developed a typology for interventions aimed at professional practice, organisational, financial, and regulatory systems (see http://www.epoc.cochrane.org/). The Cochrane Consumers and Communication Review Group has developed a typology for consumers' interactions with health care professionals, services, and researchers (see http://www.latrobe.edu.au/chcp/assets/downloads/TopicList.pdf). As an example of how such typologies are used, the EPOC typology was used to classify quality improvement strategies designed for the care of people with type 2 diabetes . Subsequent correspondence highlighted the risk of misclassifying interventions due to inadequate detail . In another example, a typology for heart failure disease management guided the grouping of clinical service interventions (multi-disciplinary, case management, and clinic models) ,.
An alternative is to develop a typology of interventions by consensus. For example, in a systematic review of occupational therapy interventions for rheumatoid arthritis, four occupational therapists identified seven types of intervention (comprehensive therapy, training of motor function, skills training, joint protection, advice on assistive devices, counselling, and provision of splints) .
B. Supplementary evidence: (i) Trial-related data
Contacting trialists to obtain the protocols they followed, and identifying supplementary research related to the trial, may help define interventions. A qualitative study, conducted alongside a trial of intensive case management for people with severe mental illness, investigated the active ingredients of the intervention in terms of staff roles and organisational features . Team management, comprehensive assessment, and needs-led service were regarded as the key mechanisms of this intervention. In another example, trialists contributing to a systematic review of stroke units were surveyed to build a description of the active components of stroke unit care. These included comprehensive assessment, active physiological management, early mobilisation, skilled nursing care, early rehabilitation, and discharge planning involving carers .
B. Supplementary evidence: (ii) Qualitative data, descriptive data, and policy documents
When trial-related evidence is inadequate , other sources of information may be relevant . A systematic review of barriers and facilitators to healthy eating among children used qualitative evidence, unrelated to the trial data, to gain a better understanding of children's perspectives . The qualitative synthesis guided the categorisation of interventions, according to the degree to which they combined health advice with the promotion of eating fruit and vegetables. A synthesis of qualitative studies aided a fuller understanding of the interventions included in a systematic review of directly observed therapy (DOT) for tuberculosis, by identifying factors that improved adherence . Factors included flexible delivery systems, involving patients in decisions, and social and family support systems . Policy documents can be particularly informative in understanding the development of service interventions across settings, for example, interventions designed to reduce reliance on hospital beds and interventions involving school feeding programmes ,.
B. Supplementary evidence: (iii) Theory
Theory may help to explain how an intervention is related to similar interventions in a particular field . However, many reviews fail to locate interventions within a theoretical model. Realist synthesis, which attempts to provide an explanatory analysis of how and why complex social interventions work (or not) in particular contexts, can aid this process. For example, a realist review of school feeding programmes identified the theory and processes that promoted the success of these interventions . Theory can also guide the classification of interventions; behavioural and contingency theory successfully guided the classification of interventions designed to implement change in practice . The theory of planned behaviour was used to address an individual's motivation, attitude, and perceived behavioural control; whereas contingency theory was used to take into account the fit between clinical practice and environmental constraints. However, theory can only improve our understanding of how an intervention works if it is part of an integrated body of knowledge that differentiates the explanatory role of one theory from another and provides robust predictions of causal pathways. An attempt to categorise studies in a Cochrane review of tobacco cessation for young people failed due to the complexity of the interventions and the simultaneous use of several psychosocial theories .
2. Searching For and Identifying All Relevant Data
The lack of consistent terminology and the inconsistent use of existing terminology to describe complex interventions means that identifying potentially eligible studies can be difficult. Search strategies may be incomplete and risk introducing bias if they identify only a proportion of all possible configurations of a complex intervention. For example, “continuity of care”, a concept that is considered to contribute to high-quality care, can be delivered through numerous mechanisms (shared care, telephone follow-up, patient-held records, and case management, to name a few) .
Solutions include characterising the elements of an intervention through an iterative scoping exercise and searching outside the traditional health care domains to include engineering, social sciences, and management journals. Data may be unpublished and only accessed through policy documents, conference proceedings, or book chapters, but need to be obtained to minimise the effects of publication bias ,. Contacting those working in the field, retrieving references of references, and tracking citations will also increase the efficiency of finding relevant evidence.
3. Selecting Studies for Inclusion in a Review
A major threat to validity from an imprecise definition of an intervention is the non-standardised and potentially non-reproducible selection of studies for inclusion in a review. Based on the available information, considerable judgement may be required when assessing how similar any given intervention is to the intervention of interest, particularly for multi-faceted interventions and those at the boundary of the content area. For example, reaching a common understanding of patient-centred interventions was not easily achieved in a review of interventions intended to promote patient-centred care .
Solutions to this problem of definition include: (1) refining the definition of an intervention through an iterative process to accommodate previously unseen configurations; (2) contacting study authors for further information; (3) recording the components of an intervention during data extraction; and (4) being explicit in the review about where disagreement occurred.
4. Synthesis of Data
Complex interventions with a large number of ill-defined elements may result in a high degree of heterogeneity. Conversely, applying a narrow definition limits generalisability by losing the potential relevance gained from examining an intervention being implemented across a range of settings. A meta-analysis of DOT for tuberculosis provided in clinics, by lay health workers, or in the home provides an example of the usefulness of exploring sources of heterogeneity . The authors found no important difference between DOT and self-administered treatment (risk ratio 1.02, 95% confidence interval 0.86 to 1.21; I2 64%). However, when the trials were grouped by the location of DOT there was a small beneficial effect for delivering DOT in a home setting compared with self administration (risk ratio 1.10, 95% confidence interval 1.02 to 1.18; I2 53%). This beneficial effect allows for several possible interpretations; for example, the burden upon patients of travelling to a clinic five days a week is minimised by having their therapy supervised at home by a lay health worker or a community or family member.
Solutions to the problem of data synthesis include categorising interventions by key variables and retaining these in the analysis. For example, a meta-analysis of discharge planning and post-discharge support categorised interventions by intensity, which varied from a single home visit, to increased clinic follow-up with telephone contact through to extended home care services . If meta-analyses cannot be performed, similar processes can be conducted whilst performing narrative synthesis. The quality of narrative analysis and applicability of review findings have recently received more attention  (see Box 1).
Box 1. Presentation of Review Findings: Information To Support Assessment of the Applicability of Evidence of Effectiveness
Intervention Content 
- Describe the content (the active ingredients) of the interventions
- Describe any interventions received by the control group, including the content of “usual care”
- Describe how the interventions were delivered and any differences in delivery across the included trials
- Describe the contextual similarities and differences between the trials
- Include details describing whether the interventions included in a review do what is intended or if they deviate from the intended shape or form during the course of implementation
- Include an assessment of whether an intervention failed because it was poorly implemented or it was not effective
- Include details on the sustainability of intervention effects over time
Roll Out/Scaling Up of the Intervention 
- Report data on accessibility, the risk of adverse events, cost-effectiveness, or budget impact of interventions
- Address the following questions regarding the applicability of the evidence to individual patients (where applicable) :
- Have biological results (e.g., sex, co-morbidities, age) that might modify the treatment response been excluded?
- Can consumers comply with the treatment requirements?
- Can health care providers comply with the treatment requirements?
- Are the likely benefits worth the potential risks and costs?
- Address the following questions regarding the applicability of the evidence in other health systems (where applicable) :
- Are there important differences or similarities in the structural elements of health systems or of health services between where the research was done and where it will be applied that might mean that an intervention could not work in the same way?
- Are there important differences in on-the-ground realities and constraints (i.e., governance, financial, and delivery arrangements) between where the research was done and where it could be applied that might substantially alter the potential benefits of the intervention?
- Are there likely to be important differences in the baseline conditions between where the research was done and other settings? If so, would this mean that the intervention could have different absolute effects, even if the relative effectiveness was the same?
- Are there important differences in perspectives and influences of health system stakeholders between where the research was done and where it could be applied that might mean an intervention will not be accepted or taken up in the same way?
Despite the range of supplementary methods available to improve the synthesis of complex interventions, most of these methods are infrequently used. There are several reasons for this lack of use. In some cases the theory underpinning a specific complex intervention has not been assembled. However, there are usually few data reporting the characteristics of complex interventions, and what data there are tend to be of poor quality. Although simple in concept, providing an adequate description of complex interventions can be technically difficult. The need to address this is becoming urgent as interventions with multiple components evolve in response to the complex health problems faced by health services. Current criteria to improve the reporting of health research are primarily concerned with the internal validity of studies , and while these include criteria related to the intervention, these guidelines can still be followed without providing adequate details of the intervention (see Table 2). It is essential that methods to improve the descriptions of complex interventions are further developed and tested with the expectation that they will complement existing systematic review methodology.