Systematic approaches to identify participants in a participatory modeling study include the application of selection criteria, nomination by a committee, and referral by the participants (i.e., ‘snowball sampling’) (Hedelin et al. 2021). We use these three strategies as follows. Selection criteria for the SMEs included having an established track record in research and public health prevention of suicide and/or ACEs and/or clinical expertise in the therapy and treatment of youth presenting suicidal behavior. The inclusion of therapists, clinical practitioners, and public health experts is essential to obtain a comprehensive coverage of risks and protective factors for suicide, which include but are not limited to ACEs. Our participants were experts within (n = 10) and outside the CDC (n = 5). A track record consists of graduate-level training (e.g., MS, MPH, PhD) and a minimum of 6 years of experience either in research or practice. Based on these criteria, a committee created a purposeful sample of SMEs to ensure we would engage with participants who could communicate experiences and expertise related to suicide in a comprehensive and reflective manner, and who were available and willing to participate within the study time frame. The committee included CDC experts from the National Center for Injury Prevention and Control’s (NCIPC) Division of Injury Prevention Suicide Prevention Team and Division of Violence Prevention Child Abuse Neglect and Adversity Team. Given the cross-sectoral nature of ACEs and suicide risks or preventative factors, the committee assembled an interdisciplinary team of SMEs including behavioral scientists, health scientists and epidemiologists, and physicians. Referrals from participating SMEs served to recruit additional colleagues across other divisions within NCIPC and outside the CDC (i.e., SMEs from other federal agencies, non-profit organizations, private practice, and universities). After each SME was identified, an email invitation was sent with proposed calendar dates within the next two weeks for scheduling.

We note that participatory modeling studies do not have a ‘typical’ or one-size-fits-all number of participants. On the one hand, several studies using mapping and quantification methods have been conducted with fewer than 10 participants (Giabbanelli et al. 2012; van Vliet et al. 2010) or at most 15 (Penn et al. 2013; Henly-Shepard et al. 2015; Isaac et al. 2009) participants. On the other hand, a previous study included 264 participants (Lavin et al. 2018). The number of participants partly depends on the task (Giabbanelli and Tawfik 2020) (e.g., if we want to compare the perspectives of participants and group them, then we may need a larger sample) and the complexity of the problem (e.g., saturation of ideas is obtained faster in problems with a smaller domain space and well-accepted evidence). In line with our previous study on a similar task (developing a map of a system) and cross-disciplinary problem (Firmansyah et al. 2019), we use validation methods (including structural network analysis and assessment of saturation) in each step to ensure that the participatory modeling process is sound.

A common approach to create a system map with SMEs is to perform semi-structured interviews. Examples include the study of Owen et al. (2018) on obesity with 16 SMEs or our previous study on smart cities (Firmansyah et al. 2019). Interviews can either be performed with a group, which directly leads into a group-level model, or with individuals, whose models are later combined into a single group-level model. Many variables drive the choice of one approach or the other: manpower and associated timeline (the additional work of combining individual models takes longer), the availability of the SMEs (a later combination of individual models requires the SME to provide feedback on the resulting group-level model), the possibility that SMEs have different viewpoints which would need to be preserved (thus avoiding the standardization or power imbalances that occur in a group), or a preference for transparency by tracking how the knowledge of each SME exactly impacts the model (Voinov et al. 2018; Jordan et al. 2018; Andersen et al. 1997). We chose the individual approach as the interdisciplinary nature of suicide research involves SMEs from different fields, thus raising the possibility of individual differences and the need to reflect them into a group-level model. The approach was feasible as SMEs were available for a follow-up and as we included two team members to cope with the additional work of model combination.

Before conducting an individual interview, a mapping approach requires an identification of the problem of interest and a definition of the system’s boundaries (Allender et al. 2019; De Pinho 2017). The problem must be sufficiently isolated to avoid ambiguities, which can lead to seemingly contradictory answers from the SMEs and raise issues in data analysis. Asking “what causes suicide?” would potentially be ambiguous as SMEs may think of different stages, from suicidal thoughts to suicide attempt and completion. In line with other models (Page et al. 2017), we thus decomposed suicide into four problems for the suicide SMEs: suicide ideation, suicide planning, suicide attempt, and suicide. In other words, the causes and consequences of each of the four interrelated problems was discussed during the semi-structured interviews with the suicide SMEs. With experts on ACEs, we set ACEs as the problem of interest. Given different definitions of ACEs in the literature, we did not decompose ACEs into mandatory categories; in other words, each participant was allowed to share their operational definition of ACEs without being restricted to one by the research team.

Complex systems are often highly interactive and cannot be considered fully in isolation. For instance, suicide is partly shaped by the family context, the community, and society. As an analogy, it is difficult to model the behavior of the respiratory system by ignoring the rest of the body, yet the complete body should not be included if the model’s focus is only on breathing (Murray-Smith 2012, p. 10). In our case, we set the boundaries on factors that shape, or are affected by, ACEs and suicide for youth in the U.S. Situations that would not be representative of the target population (e.g., trauma of famine or civil war) were thus considered beyond the scope of this project, similarly to situations encountered in adulthood (e.g., health consequences of ACEs). While the situation of parents was relevant to ACEs and suicide (e.g., mental health in the family, adversity), our model only captures it as it pertains to understanding dynamics in the child. For instance, parental separation or divorce would be relevant for the child, but the complex reasons underlying a divorce would be out of scope for our child focused model.

Each interview is conducted and recorded using WebEx video conferencing after receiving informed consent from the SME. The interview starts with a description of our project, which also serves as an opportunity to reiterate the system boundaries. SMEs are asked if they have any questions, which are then answered either by the modeler (to explain how the interview proceeds and the data are used) or by the project manager (to explain how the interview fits into the portfolio of CDC projects). Then, the interview proceeds with eliciting the map. Our process is briefly described below and shown in Fig. 2. For additional information on the flow of such interviews, we refer the reader to Reddy et al. (2019); for the cognitive underpinnings of the process, we refer to the overview in McNeese and Ayoub (2011).

We begin with the problem of interest and systematically ask the participant for causes and consequences, while clarifying with the participant whether the causal link they mentioned leads to an increase (e.g., internalized trauma increases the risk of suicide ideation) or a decrease (e.g., connectedness lowers the risk of suicide ideation). We do not ask confirmatory questions (e.g., “do you think homelessness increases suicide ideation?”), as they may bias an SME in endorsing an idea that would not have otherwise been present in their perspective. The list of factors is frequently read back to the SME such that they can identify any additional ones (e.g., “You mentioned alcohol abuse and mental health in the family—are there any other risk factors in the family?”). Once the proximal causes and consequences are identified, we shift into a second phase to identify mediators, more distal factors, or interrelationships. Questions include identifying how causal links work, particularly where there appears to be a logical gap; forming connections between existing factors; and identifying additional causes. The order in which these questions are asked depends on the conversation, as we avoid ‘jumping’ across parts of the map through successive questions (which can come across as disordered) and instead fully characterize one area before moving onto a closely related one. To establish a positive rapport with each SME, we avoid interrupting them even when they cross the system’s preferred boundaries, and instead bring them back within the boundaries through the next question. One hour was allocated for each interview such that SMEs do not need to be rushed or interrupted by lack of time.

There are several related approaches that elicit the perception of SMEs through an interview and later organize it as a network (Voinov et al. 2018). In a causal map or Causal Loop Diagram (CLD), each causal link has a type (it either increases or decreases its target) and any two factors can be connected by a link. In contrast, mind maps do not have types and only extend radially, while perception graphs also have types but cannot represent loops (Düspohl and Döll 2016). In this project, we use causal maps as this type of causation helps us to better characterize the system and because causal loops are present in complex problems such as suicide (Naumann et al. 2019; Chung 2016). As a causal map summarizes the logical impact of each factor onto others, every relationship has a cause and a consequent. To avoid confusion across disciplines, note that the terminology may differ in studies on suicide, in which a causation is the direct reason for death (e.g., gunshot wounds) while concepts such as ‘parental divorce’ would be categorized as risk factors rather than causes. In addition, the notion of ‘causation’ depends on fields and the context: here, we study the logical connections perceived by subject matter experts (e.g., “in my view, an increase in A will create a change in B”); this participatory approach to determining causation differs from demonstrating causations in health through the use of randomized controlled trials.

As case studies have abundantly demonstrated that cognitive limitations prevent stakeholders from identifying loops (Axelrod 2015), the facilitator actively monitors the structure of the network as the interview unfolds. The network is approximately structured by the modeler during the discussion using MentalModeler (Gray et al. 2013) to track what has been said and prepare the next questions.

We note that an alternative approach is to identify variables and directly connect them, for instance, by drawing them on a circle and asking participants to draw connections as well as specify their type (increasing or decreasing). This approach is feasible and promotes transparency when the causal maps are small, as exemplified by the work of Waqa et al. (2017) in which maps have about 20 connections. When the maps are larger (as in our study where each map averages 83.6 connections), visualizations can be challenging to use for participants; thus, we used the approach of interviewing and then transcribing into maps. That is, only the facilitator sees the ‘sketch’ of the map used to keep track of the conversation; a participant does not see this network during the interview and hence remains focused on the conversation rather than potentially distracted by a visual artifact.

We use the systematic method from Kim and Andersen (2012) to structure a recorded interview into a causal map. We identify the concepts in the discussion (e.g., homelessness, suicide ideation, untreated mental illness in the family) and track their causal connections (e.g., homelessness increases the risk of suicide ideation). In a causal map, concepts are represented as nodes while connections are represented as typed directed edges. In other words, each part of the interview is transformed into a schematic description of a cause variable, an effect variable, and a relationship polarity. Three sample transformations are shown in Fig. 3. In Fig. 3a, the SME provides a list of causes for separation. Rather than simplifying, we capture all such causes, as they may later connect to parts mentioned by other SMEs (e.g. on racial differences in parental incarceration). In Fig. 3b, we note that a causal mechanism may have both a positive and negative impact, which is a common sign that the mechanism depends on additional constructs. Once these constructs are brought into the conversation, we see indirect and direct pathways, which capture the difference between the positive and negative impacts. In Fig. 3c, we note that parts of a conversation may consist of examples or reflections rather than new causal mechanisms, thus not every sentence results in a node or edge. We also strictly use the vocabulary of the SME in a transcription; thus, we would not include the construct of ‘hopelessness’ based on Fig. 3c (“there is no end to it”) until and unless the SME states it. Occasionally, a logical connection is made such as how the ability or barriers to perform a task will impact this task (e.g., the ability to connect impacts connectedness).

After each interview has been transformed into a map, we examine the structure of the map for validation. Although many studies proceed without structural validation (Owen et al. 2018), causal maps can be assessed with respect to their structure or ‘fundamental causal patterns’ (Gray et al. 2019; Levy et al. 2018) to validate the interview process (Firmansyah et al. 2019). Note that ‘validation’ should be interpreted here as the presence of a quality control to ensure that the trained facilitator followed the process in Sect. 2.2. We analyze the structure of each map in terms of:

Types of nodes. As shown in Fig. 4 (colored circles), causal impacts can end at a node (receiver; typically used as output), start at a node (source; typically used as parameter), or pass through a node (transmitter). We generally observe more sources than receivers, as models tend to have few outputs of interest but many parameters driving the model’s behavior (Vasslides and Jensen 2016). Most nodes are neither sources nor receivers, as they constitute the core of the model logic that transforms parameters into outputs. When an SME categorizes a variable as a source, “this means that he perceives the relative causal relationship as a ‘forcing’ function, which cannot be controlled by any other variables. In contrast, a receiver variable is seen as not affecting any of the other variables in the system” (Papageorgiou and Areti 2012). Although an early model may have a lot of parameters, we expect this to be reduced in more elaborate models as SMEs start to capture that such parameters are in fact partly driven by other concepts.

We note that there is a plethora of measures for the structure of a causal map, as evidenced by a summary of 20 years of research in the area (Yoon and Jetter 2016). Additional measures would include whether SMEs see many alternatives (Fig. 4), or assessments of centrality, which can be done in dozens of ways. We thus focused on measures that are commonly adopted and can be interpreted in the context of structural validation (Giabbanelli and Tawfik 2021). Such measures can reveal problematic causal structures indicative of issues in the facilitation or expertise of the participants (Fig. 5).

Combining all individual maps into one is a challenge due to variations in language across the SMEs (Fig. 6). Some studies prevent this problem from appearing by limiting the SMEs to only designate concepts from a pre-established list (Gray et al. 2015), possibly including ‘distractors’ or ‘misleading’ concepts to ensure careful selection (Ruiz-Primo 2000). However, such limitations may force SMEs to artificially think alike and prevent the emergence of concepts that were not previously considered (Lavin et al. 2018). After adding all causal links and concepts across individual maps into a single one, we manually identify all equivalent forms of a concept and resolve them into a single term. Figure 6 exemplifies how variations (nodes of the same color) are resolved. Although some of the resolutions are straightforward (e.g., “mental illness of the parents” vs “parents’ mental illness”), others require an expert understanding of the context (e.g., terminology for mental health disorders). Note that some of the changes may require inverting the causal direction of a link (e.g., a ‘lack of connectedness’ is a risk factor but, ‘connectedness’ is a protective factor), as shown in Fig. 6 for hopelessness.

Five of the authors (including a suicide SME and an ACE SME) thus identify and reduce variations in language iteratively, until each concept appears through a unique term. Variations are tracked and documented in a ‘thesaurus,’ available on our repository (https://osf.io/7nxp4/). The thesaurus can be reused by other researchers needing to reduce variability in similar qualitative inquiries featuring open-ended coding schemes (Naumann et al. 2019; Chung 2016; Giabbanelli et al. 2019).

Based on previous studies, the first aggregate map of a complex system can be ‘large and unwieldy’ (Owen et al. 2018); thus, it needs to undergo a simplification process. We simplify the model based on consultations with SMEs, guided by an analysis of the network to identify possible candidates for simplification: (i) source concepts, characterized as having no incoming link, whose removal would reduce the number of model parameters; (ii) receive concepts, defined as having no outgoing link, whose removal would reduce the number of model outputs; and (iii) intermediate concepts on a chain of causation, defined as having exactly one incoming and one outgoing link, which can be simplified by rewiring their endpoints. As an unfiltered aggregate map can be challenging to navigate, network analysis served to automatize the identification of candidates, using the scripts openly available at https://osf.io/7nxp4/.

The main drawback of this approach is that it is very time-consuming, as an SME on ACEs or suicide needs to be involved to validate each of the decision to remove, and several iterations are necessary. The key advantages are that such changes do not affect the existence of feedback loops (i.e., any loop present before simplification will be preserved afterward) (Owen et al. 2018) and the model remains entirely connected unlike in some extensive pruning methods (Hayward et al. 2020).

By the end of the previous section, we obtain one map. The map is directed as it represents causality, and its concept nodes have unique labels identified by resolving linguistic variations. However, every link that potentially matters is put together without any simple way to filter or prioritize. In other words, every link mentioned was included regardless of whether it was stated by a single SME or multiple ones. Indeed, the aggregation process includes a link in the final map if it was present in any of the individual maps, without counting the number of maps that included it. This treats SMEs equally, but a systematic filtering is necessary to analyze and navigate the system. This is akin to having a road map that lists every path between all locations without differentiating whether they are highways or trails, hence preventing effective navigation. Two broad approaches can be then considered. One approach is to filter the map by including links between nodes that either occur at least at a certain user-defined frequency (e.g., only retain connections made by at least two experts) or at a frequency that is statistically significant (e.g., the use of z-scores by Bakeman and Gottman (1997), which also requires at least five experts). Another approach is to extend the map by equipping each edge with information that characterizes its nature, function, or level of endorsement (Lavin et al. 2018). For instance, they can be characterized in terms of causal intensity (when connectedness increase, how strongly does it impact suicide ideation?), timing (is the impact immediate and sustained through time or is there a lag and gradual decrease?), and previous history. The use of a follow-up survey to quantify the links thus makes up for a deficiency at the aggregation stage (not counting the endorsement of each link in individual models) and potentially goes further depending on the items on the survey.

Modeling approaches such as System Dynamics (SD) can represent all these characteristics, but the issue is then to obtain data: there are neither sufficient data to populate all aspects of the model (as evidenced by our analysis), nor the possibility to ask subject matter experts to precisely state the duration of every time lag. In contrast, Fuzzy Cognitive Mapping (FCM) can represent knowledge domains with high complexity and without a single accepted theory. In this situation, an FCM combines the qualitative information provided by different sources, thus aggregating their partial knowledge to achieve an overall understanding of the system (Hester and Adams 2017). We thus adopt the FCM methodology to characterize the causal intensity/strength of the links in the map. For a broader discussion on FCM in comparison with other network-based methods, we refer the reader to (Hester and Adams 2017).

The process is depicted in Fig. 7 and proceeds in three steps. First, we decompose the set of causal links in the map into three groups matching the expertise of the SMEs in the problem domain (ACEs, suicide, therapy/clinical practice). This avoids the creation of an overwhelmingly long questionnaire where most questions may be skipped. Second, we create an online questionnaire for each group. One question is asked for each causal edge, focusing on its causal strength. The answers are qualitative and allow for a skip pattern as well as to report that the SME disagrees with the existence of the proposed causal mechanism. The third step uses fuzzy logic to transform the qualitative constructs (e.g., ‘very low’ or ‘high’ strength) into quantitative weights.³ Fuzzy Logic was designed to deal with imprecise concepts in real-world problems, thus providing a mathematical tool that also deals with the degree or partiality of truth. Concepts such as ‘very low’ or ‘high’ are defined mathematically through fuzzy membership functions (triangular functions in Fig. 7) which partly overlap, thus indicating that an SME’s perception of a ‘medium’ strength can partly resemble what another may label a ‘high’ strength. After SMEs have completed a questionnaire, we count the proportion of respondents for each of the options (Fig. 7, Step 3a). Then, we apply fuzzy logic, which projects the ratio of each response onto its fuzzy membership function (Fig. 7, Step 3b). These responses are aggregated by ‘clipping’ the prorated parts of each function and then a consensus response is established by identifying the center of gravity (Step 3c) which results in a number in the range [−1, 1] where −1 indicates the most negative causal strength and 1 the highest one.⁴

All data reported in this section are openly accessible without registration on the third-party Open Science Framework repository (OSF) at https://osf.io/7nxp4/. The repository is organized in numbered subfolders following the progress of the model building process, from (1) individual maps and the subsequent completion of (2) individual questionnaires, to the creation of (3) the complete map. As a large map is best explored through interactive visualizations, we provide the map together with (4) an open-source visualization software and a brief tutorial. The key application demonstrating the usefulness of the map consists of (5) examining the data landscape. The remainder of this section reports key results on the model building process, complete map, and application.

A total of 17 invitations were emailed, with 15 affirmative responses, and 2 non-responses. Their professional characteristics are summarized in Table 1, and their complete list together with professional affiliations is provided in the Acknowledgements section. The fifteen SMEs accepting to participate were interviewed for the study between June 29, 2020, and July 21, 2020. Interviews lasted 55.1 ± 7.4 min. Characteristics of the individual maps (transcribed from the interview recordings) are shown in Fig. 8. As the importance and definition of metrics was provided under section ‘Transcribing an interview into an individual map,’ we now briefly analyze these results as they pertain to the validation of the model building process. Participants aptly saw the system as being composed of loops, rather than as a list of determinants (Fig. 5b). Rather than focused on direct causes or consequences, participants addressed distal factors as evidence by an average diameter of 8. On average, 15 out of the 85 causal links are concentrated around a single factor (i.e., max degree), which indicates a small level of centralization and is far away from the problematic case of the ‘star graph’ (Fig. 5a).

In the process of combining the fifteen individual maps into a single one, we found that 131 terms appeared in several maps under different names. These differences were resolved by adopting a single name, such as “racism” instead of “structural racism,” “racism and discrimination,” “racial injustice,” or “racism and forms of systemic oppression.” Each of the 131 terms had an average of 4 linguistic variations. Eight concepts had over ten linguistic variations, which indicates their importance across interviews: mental health disorders (e.g., mental health issues, mental illness, child with mental illnesses, emerging psycho-pathologies), ‘connectedness’ (e.g., social support, friends, social connections), coping skills (e.g., healthy coping strategies, problem-solving skills), family financial stress (e.g., financial distress, economic stressors on families), economic policies for ACEs, parents’ substance use, protective factors, and child abuse or neglect.

When simplifying the combined map, we emphasize that network science only serves to identify potential areas for simplifications, but decisions are ultimately made by SMEs. For example, the vast majority (62 out of 71) of concepts serving as intermediate on a chain were preserved by SMEs. Such concepts are considered as important to understand ACEs and suicide, or as potential targets for prevention. For example, we have a chain stipulating that ACEs arise when a child engaging in conflicts is harshly disciplined. Although ‘harsh discipline’ only plays the role of an intermediate from a network perspective, it is an important target in suicide prevention. Similarly, we see a chain noting that exposure to violence is followed by normalization of violence, ultimately contributing to suicide fatality. The concept of normalization was preserved, as a noteworthy social construct. Among the nine concepts that were simplified, some were merged (e.g., ‘ruminative thoughts about dying’ subsumed ‘repeated suicide ideation,’ ‘access to immediate effective clinical cares’ subsumed ‘requires medical intervention’ in the context of saving life by treatment following an attempt), and others were rewired such as directly connecting in-patient treatment to mental health treatment, without noting the ‘short stay’ intermediate (as going into a hospital triage model is out of scope).

Characteristics of the map after simplification are presented in Table 2. We note its total of 946 edges and 361 concepts, which makes it the largest map on ACEs and suicide (Brenas et al. 2019). There are only two receiver nodes, which stand for consequences that go ‘beyond’ the life of a child or adolescent. One is the consequence of a suicide, which results in exposing the community to suicide. The other is an increased propensity to be involved in violence in adulthood. The most important factor in the map in terms of degree are Adverse Childhood Experiences (ACEs). As the concept encompasses a large number of forms of abuse, it has a total of 92 antecedents and consequents. The existence of a few such ‘hub’ concepts with many connections and a majority of concepts with few connections evokes a power-law degree distribution, which is the hallmark of scale-free network. We confirmed this hypothesis by fitting a power-law degree distribution to the map (Fig. 9) with a p-value of 0.008, which is statistically significant (< 0.05). Although we see that the power-law distribution does not perfectly match our empirical distribution, such deviations from a perfect fit are commonplace in empirical data (Broido and Clauset 2019).

Previous Fuzzy Cognitive Maps related to suicide or Adverse Childhood Experiences were able to quantify all causal connections either because they were much smaller, such as 12 concepts in Merlin et al. (2020), or because they used fully automated techniques rather than a participatory approach, such as White and Mazlack (2011) who detected suicide notes based on connections between words. In contrast, our participatory modeling approach resulted in 361 nodes, connected by 946 edges. We thus selected small subsets for further quantification with the SMEs through three questionnaires: an ACE questionnaire (116 links; n = 6 respondents), a suicide questionnaire (114 links; n = 5 respondents), and a therapy questionnaire (108 links; n = 2 respondents). Twelve questions were asked to two groups of SMEs, as the questions were related to the expertise of both groups. When a large difference was observed (≥ 0.10), the group on the receiving end of the causal link always saw it as larger (Table 3). In other words, when a causal link ‘left’ the domain of expertise of a group, they tended to see it as less important.

Although the use of a follow-up questionnaire to quantify weights with the participants is a common practice (Firmansyah et al. 2019; Giabbanelli et al. 2012; Mkhitaryan et al. 2020), a recurring concern is whether participants will see ‘everything as important’ and hence score every link as Very High instead of providing nuances. We do not see this concern in the data as the (defuzzified) scores had a large range: from 0.34 to 0.86 (average 0.63 ± 0.09) for ACEs SMEs, from 0.29 to 0.86 (average 0.61 ± 0.13) for suicide SMEs, and 0.37 to 0.79 (average 0.67 ± 0.10) for therapy SMEs. This indicates that SMEs carefully reflected on the nuances of causal strengths.

The option of labeling a causal edge as inexistent (i.e., ‘no causality’) was extremely rare. Across all three questionnaires, no edge was labeled as inexistent by more than one participant. In addition, only one respondent within each of the three groups ever categorized edges as inexistent, with four edges receiving this label for suicide, 5 for ACEs, and 6 for therapy. There is thus a large consensus that most edges in the map do play a role in the ACEs and suicide system.

Given its large size (Table 2), the final map is not meant to be visualized all at once. Rather, it provides a decision support tool that can be utilized to address asking specific questions such as “in which ways does factor A eventually impact factor B,” or “what are the direct and indirect consequences of an intervention on factor A.” In this section, we selected some common queries in consultation with SMEs. In Fig. 10, we examined the many ways in which ACEs of parents eventually cause suicide fatality of the child, thus highlighting intergenerational effects. In Fig. 11, we selected a sample of six feedback loops covering different domains, from the benefits of therapy and changes in parental practices (Fig. 11a or c) to the many loops involving parental frustration and coping via substance abuse or perpetration of ACEs (Fig. 11b or d). Thanks to its broad coverage of proximal and distal factors, the loops also include high-level societal constructs such as poverty, thus demonstrating the perpetuation of social issues (Fig. 11e). Finally, we examined opportunities for interventions via structural changes focused on one factor, and the rippling effects through the broader map. Economic policies (Fig. 12a) can directly improve financial stability and lower stress, which ultimately reduces ACEs. Social support for parents (Fig. 12b) also reduces ACEs, in addition to preventing suicide ideation. Additional analyses provided in S1 Table list all of the quantified links contributing to suicide ideation, attempt, and fatality. When a direct link is not present, we measured indirect effects through paths and whether such paths all acted via factors such as ACEs, or included ACEs among other mediating variables.