A systematic review of empirical studies examining mechanisms of implementation in health

1This is the first systematic review of implementation mechanisms across health that assesses the quality of studies and the extent to which they offer evidence in support of establishing mechanisms of implementation.We summarize nine examples of models for evaluating mechanisms.We offer conceptual, theoretical, and methodological guidance for the field to contribute to the study of implementation mechanisms.

2This is the first systematic review of implementation mechanisms across health that assesses the quality of studies and the extent to which they offer evidence in support of establishing mechanisms of implementation.

3We summarize nine examples of models for evaluating mechanisms.

4We offer conceptual, theoretical, and methodological guidance for the field to contribute to the study of implementation mechanisms.

5Implementation research is the scientific evaluation of strategies or methods used to support the integration of evidence-based practices or programs (EBPs) into healthcare settings to enhance the quality and effectiveness of services. There is mounting evidence that multi-faceted or blended implementation strategies are necessary (i.e., a discrete strategy is insufficient), but we have a poor understanding of how and why these strategies work. Mechanistic research in implementation science is in an early phase of development. As of 2016, there were only nine studies included in one systematic review of implementation mediators1 specific to the field of mental health. Mediators are an intervening variable that may statistically account for the relation between an implementation strategy and outcome. We define the term mechanism as a process or event through which an implementation strategy operates to affect one or more implementation outcomes (see Table 1 for key terms and definitions used throughout this manuscript). Mechanisms offer causal pathways explaining how strategies operate to achieve desired outcomes, like changes in care delivery. Some researchers conflate moderators, mediators, and mechanisms, using the terms interchangeably. Mediators and moderators can point toward mechanisms, but they are not all mechanisms as they typically are insufficient to explain exactly how change came about.

6In addition to these linguistic inconsistencies and lack of conceptual clarity, there is little attention paid to the criteria for establishing a mechanistic relation. Originally, Bradford-Hill, and more recently Kazdin offers at least seven criteria for establishing mechanisms of psychosocial treatments that are equally relevant to implementation strategies: strong association, specificity, consistency, experimental manipulation, timeline, gradient, plausibility, or coherence (see Table 2 for definitions). Taken together, these criteria can guide study designs for building the case for mechanisms over time. In lieu of such criteria, disparate models and approaches for investigating mechanisms are likely to exist that make synthesizing findings across studies quite challenging. Consequently, the assumption that more strategies will achieve better results is likely to remain, driving costly and imprecise approaches to implementation.

7Understanding the mechanisms of implementation strategies, defined as the processes by which strategies produce desired effects, is important for both research and practice. For research, it is important to specify and examine mechanisms of implementation strategies, especially in the case of null studies, in order to understand why a strategy did or did not achieve its intended effect. For practice, it is crucial to understand mechanisms so that strategies are designed and selected to directly target implementation determinants or barriers. In the absence of this kind of intentional, a priori matching (i.e., strategy targets determinant), it is possible that the “wrong” (or perhaps less potent) strategy will be deployed. This phenomenon of mismatched strategies and determinants was quite prevalent among the 22 tailored improvement intervention studies included in Bosch et al.’s multiple case study analysis. Upon examining the timing of determinant identification and the degree to which included studies informed tailoring of the type versus the content of the strategies using determinant information, they discovered frequent determinant-strategy mismatch across levels of analysis (e.g., clinician-level strategies were used to address barriers that were at the organizational level). Perhaps what is missing is a clear articulation of implementation mechanisms to inform determinant-strategy matching. We argue that, ultimately, knowledge of mechanisms would help to create a more rational, efficient bundle of implementation strategies that fit specific contextual challenges.

8Via a systematic review, we sought to understand how mechanisms are conceptualized and measured, how they are studied (by characterizing the wide array of models and designs used to evaluate mechanisms) and evaluated (by applying Kazdin’s seven criteria), and how much evidence exists for specific mechanisms. In doing so, we offer a rich characterization of the current state of the evidence. In reflecting on this evidence, we provide recommendations for future research to optimize their contributions to mechanistic implementation science.

9The databases, PubMed and CINAHL Plus, were chosen because of their extensive collection of over 32 million combined citations of medical, nursing and allied health, and life science journals, as well as inclusiveness of international publications. We searched both databases in August 2018 for empirical studies published between January 1990 and August 2018 testing candidate mechanisms of implementation strategies. This starting date was selected given that the concept of evidence-based practice/evidence-based treatment/evidence-based medicine first gained prominence in the 1990’s with the field of implementation science following in response to a growing consciousness of the research to practice gap. The search terms were based on input from all authors who represent a variety of methodological and content expertise related to implementation science and reviewed by a librarian; see Table 3 for all search terms. The search string consisted of three levels with terms reflecting (1) implementation science, (2) evidence-based practice (EBP), and (3) mechanism. We adopted Kazdin’s definition of mechanisms, which he indicates are the basis of an effect. Due to the diversity of definitions that exist in the literature, the term “mechanism” was supplemented with the terms “mediator” and “moderator” to ensure all relevant studies were collected.

10Studies were included if they were considered an empirical implementation study (i.e., original data collection) and statistically tested or qualitatively explored mechanisms, mediators, or moderators. We did not include dissemination studies given the likely substantive differences between strategies, mechanisms, and outcomes. Specifically, we align with the distinction made between dissemination and implementation put forth by the National Institutes of Health program announcement for Dissemination and Implementation Research in Health that describes dissemination as involving distribution of evidence to a target audience (i.e., communication of evidence) and implementation as involving use of strategies to integrate evidence into target settings (i.e., use of evidence in practice). However, the word “dissemination” was included in our search terms because of the tendency of some researchers to use “implementation” and “dissemination” interchangeably. Studies were excluded if they were not an implementation study, used the terms “mediator,” “moderator,” or “mechanism” in a different context (i.e., conflict mediator), did not involve the implementation of an EBP, or were a review, concept paper, or opinion piece rather than original research. All study designs were considered. Only studies in English were assessed. See Additional File 1 for exclusion criteria and definitions. We strategically cast a wide net and limited our exclusions so as to characterize the broad range of empirical studies of implementation mechanisms.

11Citations generated from the search of PubMed and CINAHL were loaded into EPPI Reviewer 4, an online software program used for conducting literature reviews. Duplicate citations were identified for removal via the duplicate checking function in EPPI and via manual searching. Two independent reviewers (MRB, CWB) screened the first ten citations on title and abstract for inclusion. They then met to clarify inclusion and exclusion criteria with the authorship team, as well as add additional criteria if necessary, and clarify nuances of the inclusion/exclusion coding system (see Additional File 1 for exclusion criteria and definitions). The reviewers met once a week to compare codes and resolve discrepancies through discussion. If discrepancies could not be easily resolved through discussion among the two reviewers, the first author (CCL) made a final determination. During full text review, additional exclusion coding was applied for criteria that could not be discerned from the abstract; articles were excluded at this phase if they only mentioned the study of mechanisms in the discussion or future directions. Seven studies from the previous systematic review of implementation mechanisms were added to our study for data extraction; these studies likely did not appear in our review due to differences in the search strategy in that the review undertaken by Williams hand searched published reviews of implementation strategies in mental health.

12The methodological quality of included studies was assessed using the Mixed Methods Appraisal Tool (MMAT-version 2018). This tool has been utilized in over three dozen systematic reviews in the health sciences. The MMAT includes two initial screening criteria that assess for the articulation of a clear research question/objective and for the appropriateness of the data collected to address the research question. Studies must receive a “yes” in order to be included. The tool contains a subset of questions to assess for quality for each study type—qualitative, quantitative, and mixed methods. Table 4 summarizes the questions by which studies were evaluated, such as participant recruitment and relevance and quality of measures. Per the established approach to MMAT application, a series of four questions specific to each study design type are assigned a dichotomous “yes” or “no” answer. Studies receive 25 percentage points for each “yes” response. Higher percentages reflect higher quality, with 100% indicating all quality criteria were met. The MMAT was applied by the third author (CWB). The first author (CCL) checked the first 15% of included studies and, based on reaching 100% agreement on the application of the rating criteria, the primary reviewer then applied the tool independently to the remaining studies.

13Data extraction focused on several categories: study information/ background (i.e., country, setting, and sample), methods (i.e., theories that informed study, measures used, study design, analyses used, proposed mediation model), results (i.e., statistical relations between proposed variables of the mediation model tested), and criteria for establishing mechanisms (based on the seven listed in Table 2;). All authors contributed to the development of data extraction categories that were applied to the full text of included studies. One reviewer (MRB) independently extracted relevant data and the other reviewer (CWB) checked the results for accuracy, with the first author (CCL) addressing any discrepancies or questions, consistent with the approach of other systematic reviews. Extracted text demonstrating evidence of study meeting (or not meeting) each criterion for establishing a mechanism was further independently coded as “1” reflecting “criterion met” or “0” reflecting “criterion not met” by MRB and checked by CWB. Again, discrepancies and questions were resolved by the first author (CCL). Technically, mechanisms were considered “established” if all criteria were met. See Additional File 2 for PRISMA checklist for this study.

14The search of PubMed and CINAHL Plus yielded 2277 studies for title and abstract screening, of which 447 were duplicates, and 183 moved on to full-text review for eligibility. Excluded studies were most frequently eliminated due to the use of mechanism in a different context (i.e., to refer to a process, technique, or system for achieving results of something other than implementation strategies). After full article review, 39 studies were deemed suitable for inclusion in this review. Two of the included studies appeared in the only other systematic review of implementation mechanisms in mental health settings. For consistency and comprehensiveness, the remaining seven studies from the previously published review were added to the current systematic review for a total of 46 studies.2 See Fig. 1 for a PRISMA Flowchart of the screening process and results.

15Table 5 illustrates the characteristics of the 46 included studies. Twenty-five studies (54.3%) were completed in the USA, while 21 studies were conducted in other countries (e.g., Australia, Canada, Netherlands, UK). Settings were widely variable; studies occurred in behavioral health (e.g., community mental health, residential facilities) or substance abuse facilities most frequently (21.7%), followed by hospitals (15.2%), multiple sites across a health care system (15.2%), schools (15.2%), primary care clinics (10.9%), and Veteran’s Affairs facilities (8.7%). Sampling occurred at multiple ecological levels, including patients (17.4%), providers (65.2%), and organizations (43.5%). Seventeen (40.0%) studies examined the implementation of a complex psychosocial intervention (e.g., Cognitive behavioral therapy;, multisystemic therapy).

16Our review included six qualitative (10.9%), seven mixed methods (15.2%), and 34 quantitative studies (73.9%). The most common study design was quantitative non-randomized/observational (21 studies; 45.7%), of which 11 were cross-sectional. There were 13 (28.3%) randomized studies included in this review. Twenty-nine studies (63.0%) were longitudinal (i.e., included more than one data collection time point for the sample).

17Table 4 shows the results of the MMAT quality assessment. Scores for the included studies ranged from 25 to 100%. Six studies (13.0%) received a 25% rating based on the MMAT criteria, 17 studies (40.0%) received 50%, 21 studies (45.7%) received 75%, and only three studies (6.5%) scored 100%. The most frequent weaknesses were the lack of discussion on researcher influence in qualitative and mixed methods studies, lack of clear description of randomization approach utilized in the randomized quantitative studies, and subthreshold rates for acceptable response or follow-up in non-randomized quantitative studies.

18Twenty-seven (58.7%) of the studies articulated their plan to evaluate mechanisms, mediators, or moderators in their research aims or hypotheses; the remaining studies included this as a secondary analysis. Thirty-five studies (76.1%) cited a theory, framework, or model as the basis or rationale for their evaluation. The diffusion of innovations theory was most frequently cited, appearing in nine studies (19.6%), followed by the theory of planned behavior, appearing in seven studies (15.2%). The most commonly cited frameworks were the theoretical domains framework (five studies; 10.9%) and Promoting Action on Research in Health Services (PARiHS) (three studies; 6.5%).

19Four studies (8.7%) incorporated theories or frameworks that focused exclusively on a single ecological level; two focusing on leadership, one at the organizational level, and one at the systems level. There was some discordance between the theories that purportedly informed studies and the potential mechanisms of interest, as 67.4% of candidate mechanisms or mediators were at the intrapersonal level, while 30.4% were at the interpersonal level, and 21.7% at the organizational level. There were no proposed mechanisms at the systems or policy level. Although 12 studies (26.1%) examined mechanisms or mediators across multiple ecological levels, few explicitly examined multilevel relationships (e.g., multiple single-level mediation models were tested in one study).

20The vast majority of studies (38, 82.6%) utilized self-report measures as the primary means of assessing the mechanism, and 13 of these studies (28.3%) utilized focus groups and/or interviews as a primary measure, often in combination with other self-report measures such as surveys. Multiple regression constituted the most common analytic approach for assessing mediators or moderators, utilized by 25 studies (54.3%), albeit this was applied in a variety of ways. Twelve studies (26.1%) utilized hierarchical linear modeling (HLM) and six studies (13.0%) utilized structural equation modeling (SEM); see Table 6 for a complete breakdown. Studies that explicitly tested mediators employed diverse approaches including Baron and Kenny’s (N = 8, 17.4 causal steps approach, Preacher and Hayes’ (N = 3, 6.5%) approach to conducting bias-corrected bootstrapping to estimate the significance of a mediated effect (i.e., computing significance for the product of coefficients), and Sobel’s (N = 4, 8.9%) approach to estimating standard error for the product of coefficients often using structural equation modeling. Only one study tested a potential moderator, citing Raudenbush’s. Two other studies included a potential moderator in their conceptual frameworks, but did not explicitly test moderation.

21There was substantial variation in the models that emerged from the studies included in this review. Table 7 represents variables considered in mediating or moderating models across studies (or identified as candidate mediators, moderators, or mechanisms in the case of qualitative studies). Additional file 3 depicts the unique versions of models tested and their associated studies. We attempted to categorize variables as either (a) an independent variable (X) impacting a dependent variable; (b) a dependent variable (Y), typically the outcome of interest for a study; or (c) an intervening variable (M), a putative mediator in most cases, though three studies tested potential moderators. We further specified variables as representing a strategy, determinant, and outcome; see Table 1 for definitions.3

22The most common model type (29; 63.0%) was one in which X was a determinant, M was also a determinant, and Y was an implementation outcome variable (determinant ➔ determinant ➔ implementation outcome). For example, Beenstock et al. tested a model in which propensity to act (determinant) was evaluated as a mediator explaining the relation between main place of work (determinant) and referral to smoking cessation services (outcome). Just less than half the studies (22; 47.8%) included an implementation strategy in their model, of which 16 (34.8%) evaluated a mediation model in which an implementation strategy was X, a determinant was the candidate M, and an implementation outcome was Y (strategy ➔ determinant ➔ implementation outcome); ten of these studies experimentally manipulated the relation between the implementation strategy and determinant. An example of this more traditional mediation model is a study by Atkins and colleagues which evaluated key opinion leader support and mental health practitioner support (determinants) as potential mediators of the relation between training and consultation (strategy) and adoption of the EBP (implementation outcome). Five studies included a mediation model in which X was an implementation strategy, Y was a clinical outcome, and M was an implementation outcome (strategy ➔ implementation outcome ➔ clinical outcome).

23While the majority of quantitative studies tested a three-variable model, there were some notable exceptions. Several studies tested multiple three variable models that held the independent variable and mediator constant but tested the relation among several dependent variables. Several studies tested multiple three variable models that held the independent variable and dependent variables constant but tested several mediators.

24Five studies included in this review utilized qualitative methods to explore potential mechanisms or mediators of change, though only one explicitly stated this goal in their aims. Three studies utilized a comparative case study design incorporating a combination of interviews, focus groups, observation, and document review, whereas two studies employed a cross-sectional descriptive design. Although three of the five studies reported their analytic design was informed by a theory or previously established model, only one study included an interview guide in which items were explicitly linked to theory. All qualitative studies explored relations between multiple ecological levels, drawing connections between intra and interpersonal behavioral constructs and organization or system level change.

25Finally, with respect to the seven criteria for establishing mechanisms of change, the plausibility/coherence (i.e., a logical explanation of how the mechanism operates that incorporates relevant research findings) was the most frequently fulfilled requirement, met by 42 studies (91.3%). Although 20 studies (43.5%), of which 18 were quantitative, provided statistical evidence of a strong association between the dependent and independent variables, only 13 (28.2%) studies experimentally manipulated an implementation strategy or the proposed mediator or mechanism. Further, there was only one study that attempted to demonstrate a dose-response relation between mediators and outcomes. Most included studies (39; 84.8%) fulfilled three or fewer criteria, and only one study fulfilled six of the seven requirements for demonstrating a mechanism of change; see Table 8.

26Mechanism-focused implementation research is in an early phase of development, with only 46 studies identified in our systematic review across health disciplines broadly. Consistent with the field of implementation science, no single discipline is driving the conduct of mechanistic research, and a diverse array of methods (quantitative, qualitative, mixed methods) and designs (e.g., cross-sectional survey, longitudinal non-randomized, longitudinal randomized, etc.) have been used to examine mechanisms. Just over one-third of studies (N = 16; 34.8%) evaluated a mediation model with the implementation strategy as the independent variable, determinant as a putative mediator, and implementation outcome as the dependent variable. Although this was the most commonly reported model, we would expect a much higher proportion of studies testing mechanisms of implementation strategies given the ultimate goal of precise selection of strategies targeting key mechanisms of change. Studies sometimes evaluated models in which the determinant was the independent variable, another determinant was the putative mediator, and an implementation outcome was the dependent variable (N = 11; 23.9%). These models suggest an interest in understanding the cascading effect of changes in context on key outcomes, but without manipulating or evaluating an implementation strategy as the driver of observed change. Less common (only 5, 10.9%) were more complex models in which multiple mediators and outcomes and different levels of analyses were tested (e.g.,), despite that this level of complexity is likely to characterize the reality of typical implementation contexts. Although there were several quantitative studies that did observe significant relations pointing toward a mediator, none met all criteria for establishing a mechanism.

27Less than one-third of the studies experimentally manipulated the strategy-mechanism linkage. As the field progresses, we anticipate many more tests of this nature, which will allow us to discern how strategies exert their effect on outcomes of interest. However, implementation science will continue to be challenged by the costly nature of the type of experimental studies that would be needed to establish this type of evidence. Fortunately, methodological innovations that capitalize on recently funded implementation trials to engage in multilevel mediation modeling hold promise for the next iteration of mechanistic implementation research As this work unfolds, a number of scenarios are possible. For example, it is likely the case that multiple strategies can target the same mechanism; that a single strategy can target multiple mechanisms; and that mechanisms across multiple levels of analysis must be engaged for a given strategy to influence an outcome of interest. Accordingly, we expect great variability in model testing will continue and that more narrowly focused efforts will remain important contributions so long as shared conceptualization of mechanisms and related variables is embraced, articulated, and rigorously tested. As with other fields, we observed great variability in the degree to which mechanisms (and related variables of interest) were appropriately specified, operationalized, and measured. This misspecification coupled with the overall lack of high-quality studies (only three met 100% of the quality criteria), and the diversity in study methods, strategies tested, and mediating or moderating variables under consideration, we were unable to synthesize the findings across studies to point toward promising mechanisms.

28Despite the important advances that the studies included in this review represent, there are clear conceptual and methodological issues that need to be addressed to allow future research to more systematically establish mechanisms. Table 1 offers a list of key terms and definitions for the field to consider. We suggest the term “mechanism” be used to reflect a process or event through which an implementation strategy operates to affect desired implementation outcomes. Consistent with existing criteria, mechanisms can only be confidently established via carefully designed (i.e., longitudinal; experimentally manipulated) empirical studies demonstrating a strong association, and ideally a dose-response relation, between an intervening variable and outcome (e.g., via qualitative data or mediation or moderator analyses) that are supported by very specific theoretical propositions observed consistently across multiple studies. We found the term “mediator” to be most frequently used in this systematic review, which can point toward a mechanism, but without consideration of these full criteria, detection of a mediator reflects a missed opportunity to contribute more meaningfully to the mechanisms literature.

29Interestingly, the nearly half of studies (43.5%) treated a variable that many would conceptualize as a “determinant” as the independent variable in at least one proposed or tested mediation pathway. Presumably, if researchers are exploring the impact of a determinant on another determinant and then on an outcome, there must be a strategy (or action) that caused the change in the initial determinant. Or, it is possible that researchers are simply interested in the natural associations among these determinants to identify promising points of leverage. This is a prime example where the variable or overlapping use of concepts (i.e., calling all factors of interest “determinants”) becomes particularly problematic and undermines the capacity of the field to accumulate knowledge across studies in the service of establishing mechanisms. We contend that it is important to differentiate among concepts to use more meaningful terms like preconditions, putative mechanisms, proximal and distal outcomes, all of which were under-specified in the majority of the included studies. Several authors from our team have articulated an approach to building causal pathway diagrams that clarifies that preconditions are necessary factors for a mechanism to be activated and proximal outcomes are the immediate result of a strategy that is realized only because the specific mechanism was activated. We conceptualize distal outcomes as the eight implementation outcomes articulated by Proctor and colleagues. Disentangling these concepts can help characterize why strategies fail to exert an impact on an outcome of interest. Examples of each follow in the section below.

30With greater precision among these concepts, the field can also generate and test more specific hypotheses about how and why key variables are related. This begins with laying out mechanistic research questions (e.g., How does a network intervention, like a learning collaborative, influence provider attitudes?) and generating theory-driven hypotheses. For instance, a testable hypothesis may be that learning collaboratives [strategy] operate through sharing [mechanism] of positive experiences with a new practice to influence provider attitudes [outcome]. As another example, clinical decision support [strategy] may act through helping the provider to remember [mechanism] to administer a screener [proximal outcome] and flagging this practice before an encounter may not allow the mechanism to be activated [precondition]. Finally, organizational strategy development [strategy] may have an effect because it means prioritizing competing demands [mechanism] to generate a positive implementation climate [proximal outcome]. Research questions that allow for specific mechanism-focused hypotheses have the potential to expedite the rate at which effective implementation strategies are identified.

31Ultimately, theory is necessary to drive hypotheses, explain implementation processes, and effectively inform implementation practice by providing guidance about when and in what contexts specific implementation strategies should or should not be used. Implementation theories can offer mechanisms that extend across levels of analysis (e.g., intrapersonal, interpersonal, organizational, community, macro policy). However, there is a preponderance of frameworks and process models, with few theories in existence. Given that implementation is a process of behavior change at its core, in lieu of implementation-specific theories, many researchers draw upon classic theories from psychology, decision science, and organizational literatures, for instance. Because of this, the majority of the identified studies explored intrapersonal-level mechanisms, driven by their testing of social psychological theories such as the theory of planned behavior and social cognitive theory. Nine studies cited the diffusion of innovations as a theory guiding their mechanism investigation, which does extend beyond intrapersonal to emphasize interpersonal, and to some degree community level mechanisms, although we did not see this materialize in the included study analyses. Moving forward, developing and testing theory is critical for advancing the study of implementation mechanisms because theories (implicitly or explicitly) tend to identify putative mechanisms instead of immutable determinants.

32Inadequate measurement has the potential to undermine our ability to advance this area of research. Our coding indicated that mechanisms were assessed almost exclusively via self-report (questionnaire, interview, focus group) suggesting that researchers conceptualize the diverse array of mechanisms to be latent constructs and not directly observable. This may indeed be appropriate, given that mechanisms are typically processes like learning and reflecting that occur within an individual and it is their proximal outcomes that are directly observable (e.g., knowledge acquisition, confidence, perceived control). However, conceptual, theoretical, and empirical work is needed to (a) articulate the theorized mechanisms for the 70+ strategies and proximal outcomes, (b) identify measures of implementation mechanisms and evaluate their psychometric evidence base and pragmatic qualities, and (c) attempt to identify and rate or develop objective measures of proximal outcomes for use in real-time experimental manipulations of mechanism-outcome pairings.

33The multilevel interrelations of factors implicated in an implementation process also call for sophisticated quantitative and qualitative methods to uncover mechanisms. With respect to quantitative methods, it was surprising that the Baron and Kenny approach to mediation testing remains most prevalent despite that most studies are statistically underpowered to use this approach, and the other most common approach (i.e., the Sobel test) relies on an assumption that the sampling distribution of the mediation effect is normal, neither of which were reported on in any of the 12 included studies that used these methods. Williams suggests the product of coefficients approach is more appropriate for mediation analysis because it is a highly general approach to both single and multi-level mediation models that minimizes type I error rates, maximizes statistical power, and enhances accuracy of confidence intervals. The application of moderated mediation models and mediated moderator models will allow for a nuanced understanding of the complex interrelations among factors implicated in an implementation process.

34Because this was the first review of implementation mechanisms across health disciplines, we believed it was important to be inclusive with respect to methods employed. Qualitative studies are important to advancing research on implementation mechanisms in part because they offer a data collection method in lieu of having an established measure to assess mechanisms quantitatively. Qualitative research is important for informing measure development work, but also for theory development given the richness of the data that can be gleaned. Qualitative inquiry can be more directive by developing hypotheses and generating interview guides to directly test mechanisms. Diagramming and tracing causal linkages can be informed by qualitative inquiry in a structured way that is explicit with regard to how the data informs our understanding of mechanisms. This kind of directed qualitative research is called for in the United Kingdom’s MRC Guidance for Process Evaluation. We encourage researchers internationally to adopt this approach as it would importantly advance us beyond the descriptive studies that currently dominate the field.

35There are several limitations to this study. First, we took an efficient approach to coding for study quality when applying the MMAT. Although it was a strength that we evaluated study quality, the majority of studies were assessed only by one research specialist. Second, we may have overlooked relevant process evaluations conducted in the UK where MRC Guidance stipulates inclusion of mechanisms that may have been described using terms not included in our search string. Third, although we identified several realist reviews, we did not include them in our systematic review because they conceptualize mechanisms differently than how they are treated in this review. That is, realist synthesis posits that interventions are theories and that they imply specific mechanisms of action instead of separating mechanisms from the implementation strategies/interventions themselves. Thus, including the realist operationalization would have further confused an already disharmonized literature with respect to mechanisms terminology but ultimately synthesizing findings from realist reviews with standard implementation mechanism evaluations will be important. Fourth, our characterization of the models tested in the identified studies may not reflect those intended by researchers given our attempt to offer conceptual consistency across studies, although we did reach out to corresponding authors for whom we wished to seek clarification on their study. Finally, because of the diversity of study designs and methods, and the inconsistent use of relevant terms, we are unable to synthesize across the studies and report on any robustly established mechanisms.

36This study represents the first systematic review of implementation mechanisms in health. Our inclusive approach yielded 46 qualitative, quantitative, and mixed methods studies, none of which met all seven criteria (i.e., strong association, specificity, consistency, experimental manipulation, timeline, gradient, plausibility or coherence) that are deemed critical for empirically establishing mechanisms. We found nine unique versions of models that attempted to uncover mechanisms, with only six exploring mediators of implementation strategies. The results of this review indicated inconsistent use of relevant terms (e.g., mechanisms, determinants) for which we offer guidance to achieve precision and encourage greater specificity in articulating research questions and hypotheses that allow for careful testing of causal relations among variables of interest. Implementation science will benefit from both quantitative and qualitative research that is more explicit in their attempt to uncover mechanisms. In doing so, our research will allow us to test the idea that more is better and move toward parsimony both for standardized and tailored approaches to implementation.