Many health problems are complex, requiring systems thinking tools to visualize causal networks and simulate interventions. Causal loop diagrams (CLDs) are valuable for this purpose, but their construction often relies on single sources of evidence, leading to biases and limitations. This study addresses this by proposing a triangulation approach that combines three distinct methods: group model building (GMB), literature review, and causal discovery. GMB leverages the collective knowledge of domain experts, literature review synthesizes existing research, and causal discovery uses data-driven methods to identify causal relationships. The integration of these methods aims to overcome the limitations of each individual approach, resulting in more robust and reliable CLDs. The researchers hypothesize that this triangulation will lead to more comprehensive, accurate, and transparent CLDs, ultimately improving our understanding of complex health issues. The study's importance lies in its potential to significantly improve the reliability and validity of CLDs, a critical tool for understanding and addressing complex health problems where multiple factors interact in intricate ways. The use of a case study allows for a detailed examination of the triangulation process and its impact on the resulting CLD. The chosen case study focuses on the trajectory of depressive symptoms in response to stressors in healthy adults, providing a relevant and well-defined context for evaluating the proposed methodology.

The paper reviews existing literature on CLD construction methods, highlighting the limitations of relying solely on expert knowledge or literature reviews. It emphasizes the increasing adoption of causal diagrams in epidemiology for building computational models and simulating interventions. The authors discuss the benefits of participatory methods like GMB in eliciting expert insights, particularly for complex problems with limited data. However, they also acknowledge the potential for bias and lack of reproducibility when relying on a single source of evidence. The literature highlights the increasing trend of combining GMB with literature reviews, but this combination still faces subjectivity issues. Therefore, the authors advocate for a mixed-methods approach integrating qualitative and quantitative data to develop CLDs that accurately represent real-world systems. The review also discusses the potential of causal discovery as a quantitative approach for identifying causal diagrams from observational data, noting both its advantages (data-driven perspective, assessment of all possible links) and challenges (sensitivity to unmet assumptions, potential deviation from theory-driven models). The need for validating causal discovery results with domain knowledge is emphasized.

The study employed a mixed-methods triangulation approach to develop a CLD for the trajectory of depressive symptoms in response to stressors in healthy adults. The methodology involved four key steps: 1. **Group Model Building (GMB):** A group of four domain experts (two each from biological, psychological, behavioral, and social domains) participated in GMB sessions to identify variables and map causal links between them. The Healthy Brain Study (HBS), a longitudinal cohort study, provided the context and variables for the model. A consensus-based approach was adopted, ensuring agreement on all links. 2. **Literature Review:** After each GMB session, a literature review was conducted for each newly added link to assess existing evidence. The review process examined longitudinal associations, robustness across studies, and plausible mechanisms. Links lacking supporting evidence were either removed or flagged as 'hypothetical'. 3. **Causal Discovery:** Causal discovery analysis was performed using the J-PCMCI+ algorithm on longitudinal data from the HBS. This algorithm identifies causal links from observational data, considering various factors including time lags and conditional independencies. Sensitivity analyses were conducted to assess the robustness of the results. Education and sex were included as context variables in the algorithm. 4. **Triangulation:** The findings from GMB, literature review, and causal discovery were integrated in a final GMB session. Experts evaluated inconsistencies and made decisions on which links to include in the final CLD. Links were categorized based on agreement or disagreement between GMB and causal discovery, with literature review used to resolve discrepancies. The final CLD included links supported by both GMB and causal discovery, links found only by causal discovery but supported by literature, and links found only by GMB but deemed plausible by experts. The J-PCMCI+ algorithm, a constraint-based method, was used due to its flexibility in handling mixed data types and the ability to identify time-dependent feedback loops. The algorithm's assumptions, including the causal Markov and faithfulness conditions, and sensitivity to missing data were considered. The specific steps involved data preprocessing, running the algorithm, and sensitivity analysis regarding significance level, independence tests, and the impact of participants with missing data. Linear models were used to quantify the strength of identified links.

The study revealed a number of key findings, illustrating the impact of triangulation on CLD development. The GMB process initially identified 33 links between 14 variables representing biological, psychological, behavioral, and social domains. The literature review resulted in the removal of two implausible links and the modification of another based on expert discussion. The causal discovery analysis using J-PCMCI+ identified 12 links, with sensitivity analysis showing some sensitivity to the inclusion of participants with complete data across all assessments. The triangulation process, combining these three sources of evidence, resulted in a final CLD with 36 causal links. This final CLD demonstrated three improvements: 1. **Increased Comprehensiveness:** The integration of literature review and causal discovery significantly expanded the CLD's scope, including links from multiple research domains that were not initially identified by GMB. 2. **Modified Feedback Structure:** The CLD's feedback structure changed across the different triangulation steps. For instance, some reinforcing feedback loops were removed or modified, while new loops were introduced based on causal discovery findings and expert consensus. 3. **Increased Transparency:** The triangulation process provided greater transparency regarding the uncertainty associated with specific links, showing which findings were supported by all three sources (high confidence) and which were supported by only one or two (lower confidence). The final CLD visually represented this uncertainty using different line styles and thicknesses. The resulting CLD provided preliminary insights into the interplay of factors influencing depressive symptoms, revealing reinforcing feedback loops involving depressive symptoms, proinflammatory processes, perceived stress, loneliness, and sleep disturbances. The CLD also showed cross-scale loops involving multiple research domains and highlighted the significant role of prosocial behavior as a potential intervention target. The specific links identified by J-PCMCI+, their polarities, and whether or not they were included in the final CLD are detailed in Table 2.

The findings demonstrate that the triangulation approach successfully addressed the limitations of relying on a single source of evidence for CLD construction. The increased comprehensiveness, modified feedback structure, and enhanced transparency of the final CLD highlight the significant benefits of combining expert knowledge, literature review, and data-driven causal discovery. This approach addresses the potential biases and incompleteness inherent in each individual method. The case study's findings suggest that integrating causal discovery with GMB, a participatory modeling method, can be particularly beneficial for identifying links missed by experts or for revealing potential research gaps. The improved feedback structure reveals a more nuanced understanding of the dynamics underlying depressive symptoms. The enhanced transparency regarding uncertainty enables a more informed assessment of the CLD's reliability and facilitates future research by identifying areas requiring further investigation. The results support the recommendation of triangulation for future studies aiming to model complex health problems.

This study successfully demonstrated a novel mixed-methods triangulation approach for developing CLDs. The integration of GMB, literature review, and causal discovery produced a more comprehensive, accurate, and transparent CLD for the case study on depressive symptoms. The findings highlight the benefits of this approach in overcoming the limitations of individual methods, improving the robustness of CLDs, and identifying areas for future research. Future work could explore alternative triangulation approaches, incorporate data-driven variable selection, use artificial intelligence for automated literature reviews, and extend the CLD to include additional variables and timeframes. This triangulation approach has the potential to significantly advance our understanding of complex health problems and improve the development of effective intervention strategies.

The study's limitations primarily stem from those inherent in each of the individual methods. The causal discovery analysis was sensitive to the removal of participants with incomplete data, highlighting the potential impact of missing values on the results. The literature review, while thorough, was not fully systematic, which could impact the comprehensiveness of evidence considered for each link. The Healthy Brain Study, while rich in data, had limitations in the number of assessments (three over one year), restricting the ability to correct for complex confounding. Additionally, the time-consuming nature of the triangulation process, involving multiple methods and expert discussions, should be considered when planning similar studies.