From childhood trauma to complex PTSD: a network analysis

ABSTRACT

Background: Network analysis enables the understanding of complex posttraumatic stress disorder (CPTSD) as interconnected symptoms, with Bayesian networks allowing causal inference. However, existing network studies have included populations with varied trauma types and durations, yielding mixed results, and have rarely incorporated trauma histories into the analyses, despite their etiological significance.

Objective: This study examines the structural relationships and potential pathways between childhood trauma types and CPTSD symptoms in individuals with interpersonal childhood trauma, aiming to identify possible key variables that could guide intervention targets.

Method: The analysis included 249 adults with histories of interpersonal childhood trauma. Childhood trauma histories were assessed through the Childhood Trauma Questionnaire (CTQ), and CPTSD symptoms were measured by the International Trauma Questionnaire (ITQ). Network analysis comprised complementary methodologies: a regularized partial correlation network using graphical LASSO with centrality and bridge centrality, and a directed acyclic graph (DAG) using a Bayesian hill-climbing algorithm. Additional methodologies included exploratory graph analysis, unique variable analysis, predictability, and bootstrapped stability metrics to enhance methodological rigour.

Results: The analysis suggested a potential pathway from emotional neglect to emotional abuse, with negative self-concept bridging trauma experiences and subsequent symptoms; however, supplementary analyses warrant caution in interpretation. Accordingly, rather than pinpointing individual symptoms, the findings highlight upstream variables – particularly negative self-concept and disturbances in relationships – as potential intervention targets. Notably, incorporating trauma experiences into the DAG analysis revealed directional patterns that contrasted with symptom-only networks, suggesting that trauma experiences serve as crucial reference points.

Conclusions: This exploratory study demonstrates the value of integrating trauma variables and complementary network methodologies in understanding the CPTSD network. The findings suggest the potential significance of emotional trauma and upstream variables in symptom development, while highlighting how trauma context can anchor symptom relationships in the Bayesian network.

HIGHLIGHTS

• Network analysis suggested emotional trauma as the primary pathway through which childhood trauma influences CPTSD symptom development, with negative self-concept serving as a bridge.

• Upstream variables – negative self-concept and disturbances in relationships – emerged as potential intervention targets, while downstream symptoms – affect dysregulation, sense of threat, re-experiencing, and avoidance – appeared to form a feedback loop maintaining the network.

• Integrating childhood trauma experiences into network analysis revealed directional patterns that contrasted with symptom-only networks, suggesting that trauma experiences serve as crucial reference points.

1. Introduction

The 11th revision of the International Classification of Diseases (ICD-11) has formalized complex posttraumatic stress disorder (CPTSD) as a distinct diagnosis, characterized by specific symptom constellations following prolonged and repeated interpersonal trauma (Herman, 1992; Hyland et al., 2017). This diagnosis encompasses posttraumatic stress disorder (PTSD) symptoms of re-experiencing, avoidance, and sense of threat, alongside disturbances in self-organization (DSO), which comprises affect dysregulation, negative self-concept, and disturbances in relationships.

Network analysis has emerged as a particularly apposite methodology for examining the interplay of co-occurring symptoms in CPTSD, conceptualizing the disorder as a complex system of interconnected symptoms rather than manifestations of a latent construct. While network analysis provides insights into the structure and centrality of variables, causal inference remains limited in cross-sectional data (Borsboom et al., 2021; Briganti et al., 2023; Epskamp, Waldorp et al., 2018). Specifically, partial correlation networks cannot determine whether a symptom causes or is caused by other symptoms, even when identified as the most central in the network.

To address the limitation, Bayesian network analysis has emerged in the field of psychopathology (McNally, Heeren, et al., 2017; McNally, Mair, et al., 2017). Bayesian networks can infer causal relationships through a directed acyclic graph (DAG) and joint probability distributions by modelling directional relationships under the assumptions of no unmeasured variables or cyclic relationships (Briganti et al., 2023; McNally, Mair, et al., 2017). DAGs enable directional prediction through edges with arrowheads, where each directed edge represents the potential causal flow between variables.

Both partial correlation and Bayesian networks have been employed to investigate CPTSD symptoms. However, they yielded varying results, where partial correlation networks identified negative self-concept, sense of threat, avoidance, and affect dysregulation as the most central symptoms across studies (Karatzias et al., 2020; Knefel et al., 2016, 2019; Levin et al., 2021; McElroy et al., 2019), and a Bayesian network identified hypervigilance as a central symptom in the upstream network, suggesting it as a key target for intervention (Yang et al., 2023).

Two explanations may account for the varying results. One is considerable heterogeneity in trauma experiences across studies, ranging from the general population to those with specific trauma exposures, including short-term war trauma, lifetime trauma, combined direct and indirect trauma, and childhood trauma (Karatzias et al., 2020; Knefel et al., 2016, 2019; Levin et al., 2021; McElroy et al., 2019; Yang et al., 2023). While CPTSD can develop following any type and duration of trauma at any stage of life (Hyland et al., 2017; Kairyte et al., 2022; Levin et al., 2021; Maercker et al., 2022; Yang et al., 2023), Karatzias et al. (2020) reported that different trauma types were associated with different central symptoms in CPTSD networks; specifically, sense of threat and disturbances in relationships for illness-related trauma, avoidance for accidents or assaults, and negative self-concept for poly-trauma.

Moreover, clinical observations and subsequent empirical evidence have established CPTSD as primarily developing from repeated interpersonal trauma, particularly involving coercive control, during childhood (Cloitre et al., 2019; Herman, 1992; Hyland et al., 2017; Karatzias et al., 2017; Maercker et al., 2022). Therefore, this study focuses specifically on individuals with interpersonal childhood trauma to generate findings most relevant to the prototypical presentation of CPTSD.

Another possible explanation for the varying results is the exclusion of trauma variables in the network analysis. Existing CPTSD network studies have rarely incorporated trauma histories into their analyses, despite their etiological significance in symptom development. Because network estimation is sensitive to missing variables, leaving out key variables can significantly affect the results (Borsboom et al., 2021; Bringmann et al., 2019). The present study integrates five distinct categories of trauma (physical abuse, physical neglect, emotional abuse, emotional neglect, and sexual abuse) from the Childhood Trauma Questionnaire (CTQ; Bernstein et al., 2003) and examines how childhood trauma influences CPTSD symptom development and maintenance using both partial correlation (including centrality and bridge centrality) and Bayesian networks. To further strengthen our analytical approach, we conducted additional analyses: exploratory graph analysis, unique variable analysis, predictability, and bootstrapped stability metrics.

As an exploratory investigation, this study aims to: (1) examine the structural relationships and potential directional pathways from childhood trauma types to CPTSD symptoms in individuals with interpersonal childhood trauma, and (2) identify possible key variables that may activate the symptom network, guiding potential intervention targets.

2. Method

2.1. Participants and procedure

A total of 249 adults with at least one interpersonal trauma before the age of 18 participated in the study. All participants were recruited through online portals and bulletin boards at counselling centres in Korea and were able to complete questionnaires in Korean. They completed an online survey related to childhood trauma and trauma symptoms between November 2023 and February 2024, for monetary compensation. The inclusion criterion was having experienced at least one interpersonal childhood trauma, screened by a self-report questionnaire. Psychiatric comorbidities were assessed through questions regarding formal mental health diagnoses and specific diagnosed conditions.

2.2. Measures

2.2.1. Childhood Trauma Questionnaire

Childhood trauma history was measured using the Childhood Trauma Questionnaire (CTQ; Bernstein et al., 1994, 2003). The translation and psychometric properties were established with Cronbach's α of .79 (Yu et al., 2009). The scale consists of 28 items (25 clinical items and 3 validity items) assessing five types of childhood trauma: (1) physical abuse, (2) physical neglect, (3) emotional abuse, (4) emotional neglect, and (5) sexual abuse. All items are rated on a 5-point Likert scale ranging from 1 (Never True) to 5 (Very Often True) to indicate the frequency of abuse experiences.

2.2.2. International Trauma Questionnaire

ICD-11 PTSD and CPTSD symptoms were measured using the International Trauma Questionnaire (ITQ; Cloitre et al., 2018). The translation and psychometric properties were established with Cronbach's α of .92 for PTSD and .91 for DSO items (Choi et al., 2021). This 18-item measure consists of 9 items for PTSD (6 symptom items, with 2 items each assessing re-experiencing, avoidance, and sense of threat; 3 functional impairment items) and 9 items for DSO (6 symptom items, with 2 items each measuring affect dysregulation, negative self-concept, and disturbances in relationships; 3 functional impairment items). All items are rated on a 5-point Likert scale ranging from 0 (Not at all) to 4 (Extremely) to indicate symptom severity.

2.3. Data Analysis

All analyses were conducted in R (version 4.4.2). Individual questionnaire items were aggregated into a total of 11 subscale scores: five from CTQ and six from ITQ (three PTSD and three DSO subscales). There were no missing data. In our trauma-exposed group, Mardia’s test indicated violations of multivariate normality. Therefore, the nonparanormal transformation was applied using the huge package (version 1.3.5; Liu et al., 2012). This transformation preserves the underlying relationship structure while accommodating non-normal distributions.

Prior to network estimation, we employed Unique Variable Analysis (UVA) to detect and handle potentially redundant nodes in our network, using the EGAnet package (version 2.2.0; Christensen et al., 2023). UVA identifies redundant variable pairs through weighted topological overlap (wTO) metrics. While we initially tested two recommended strategies to handle the redundancy (Christensen et al., 2020): (1) creating composite variables for redundant pairs, and (2) removing redundant nodes, neither method proved suitable. Therefore, we proceeded with the original variable set for our final network analysis. To examine the community structure, particularly whether PTSD and DSO symptoms form distinct clusters, Exploratory Graph Analysis (EGA) was conducted using the Walktrap algorithm for community detection, implemented in the EGAnet package (Golino & Demetriou, 2017; Golino & Epskamp, 2017).

2.3.1. Partial correlation network (Graphical LASSO)

Network estimation was conducted using Gaussian Graphical Models (GGM; Epskamp, Waldorp et al., 2018) with the graphical least absolute shrinkage and selection operator (Graphical LASSO; Friedman et al., 2008), implemented in the bootnet package (version 1.6; Epskamp, Borsboom et al., 2018). LASSO employs a regularization penalty that shrinks edge estimates to zero, where weaker edges are dropped from the model, leading to a sparse network (Tibshirani, 1996). The Extended Bayesian Information Criterion (EBIC; Chen & Chen, 2008) with a tuning parameter of γ = .5 was used to optimize the level of regularization.

The partial correlation network was undirected, with edges representing bidirectional associations rather than causal relationships, and weighted, with edge thickness indicating the strength of relationships between nodes. Blue and red edges represented positive and negative associations, respectively. The network was visualized using the qgraph package (version 1.9.8; Epskamp et al., 2012), which implements the Fruchterman-Reingold algorithm. This algorithm optimizes node placement by positioning strongly interconnected nodes toward the network's centre, thereby facilitating the interpretation of structural patterns (Fortunato, 2010).

2.3.2. Predictability

Predictability, the proportion of variance (R²) explained by all neighbouring nodes in the network, was calculated using the mgm package (version 1.2.15; Haslbeck & Fried, 2017; Haslbeck & Waldorp, 2018). Unlike centrality indices, which provide relative importance measures, predictability offers absolute values ranging from 0 to 1, serving as a complementary metric to centrality by quantifying each node's influence in absolute terms (McNally, 2021).

2.3.3. Centrality and bridge centrality estimations

To assess the importance of each node in the network, four centrality indices were calculated using the qgraph package: strength, expected influence, betweenness, and closeness (Epskamp et al., 2012; Opsahl et al., 2010; Robinaugh et al., 2016). Strength centrality measures a node's overall connectivity by summing the absolute values of its connected edge weights. Expected influence (EI) works similarly but retains the original signs of edge weights, allowing positive and negative connections to offset each other – making it particularly useful for networks with both positive and negative relationships (Robinaugh et al., 2016). Betweenness centrality quantifies how frequently a node lies on the shortest path between any node pair in the network. Closeness centrality measures the average distance between a node and all other nodes in the network.

To identify nodes that act as ‘bridges’ between different clusters, bridge centrality was conducted using the networktools package (version 1.6.0; Jones et al., 2019). Bridge measures included bridge strength, expected influence (1-step and 2-step), betweenness, and closeness. Bridge strength quantifies a node’s connections to other nodes outside its community by summing the absolute values of edge weights. Similarly, bridge expected influence measures connections to other clusters but preserves the signs of edge weights, quantifying the overall positive influence across clusters. Bridge expected influence includes one-step expected influence (EI1), which measures a node’s direct influence on immediate neighbours in other clusters, and two-step expected influence (EI2), which extends this measurement to include nodes that are two steps away in other clusters to account for direct and indirect effects. Bridge betweenness measures how often a node appears on the shortest path between any pair of nodes from distinct communities. Lastly, bridge closeness measures how close a node is to all other nodes in other communities by calculating average distances.

2.3.4. Stability and accuracy analysis

Three stability metrics were calculated using the bootnet package (Epskamp, Borsboom et al., 2018). First, the edge weight accuracy was estimated using bootstrapped 95% confidence intervals (BCIs) with 1000 iterations around the edges (Hastie et al., 2015). Second, the stability of edges, centrality, and bridge centrality was examined via the correlation stability (CS) coefficient. The CS-coefficient quantifies what proportion of cases can be maximally dropped while maintaining original estimates with a correlation of at least .7 within a 95% confidence interval; CS-coefficients of at least .25 are recommended, and .5 are preferred for strong robustness (Epskamp & Fried, 2018; Epskamp, Borsboom et al., 2018). Third, bootstrap difference tests using 1000 iterations were conducted to examine pairwise comparisons in centrality indices (Epskamp, Borsboom et al., 2018).

2.3.5. Bayesian network (DAG)

A Bayesian network was computed with the hill-climbing algorithm and visualized as a DAG, using the bnlearn package (version 5.0.2; Briganti et al., 2023; McNally, Heeren, et al., 2017; McNally, Mair, et al., 2017). The hill-climbing algorithm modifies every possible edge by adding, removing, or reversing until the best fit is obtained (i.e. Bayesian Information Criterion [BIC]). For the study, edges from any PTSD or DSO symptoms to trauma types were excluded (‘blacklisting’) as they are improbable directions in clinical or real-life settings (i.e. current symptoms cannot cause past childhood trauma experiences). To establish a stable DAG, we conducted bootstrapping with 1000 samples and averaged them to obtain the resultant network. We employed Scutari and Nagarajan (2013) method, which demonstrates superior sensitivity in edge detection. Since relationships between childhood trauma and symptoms can be relatively weaker compared to symptom-symptom relationships, Scutari and Nagarajan (2013) approach was particularly valuable for our analysis, as it better captures complex relationships and retains more edges. Following this approach, a DAG with an optimized threshold of .48 was presented.

3. Results

3.1. Demographic information

The sample comprised 249 participants (196 women, 53 men; mean age = 28.8 years, SD = 6.64), predominantly Korean (n = 247, 99.2%) and unmarried (83.5%). Psychiatric comorbidity was present in 26.5% (n = 66) of participants, with 68.2% reporting a single diagnosis, 25.8% two diagnoses, and 6% three or more diagnoses. The most prevalent conditions were depressive disorders (19.3%), anxiety disorders (6.8%), bipolar disorders (2.0%), attention-deficit/hyperactivity disorder (2.0%), and posttraumatic stress disorder (1.6%). Other conditions (3.2%) included sleep disorders, social communication disorder, obsessive-compulsive disorder, and alcohol use disorder.

Participants reported varying levels of childhood trauma exposure. Emotional neglect was the most prevalent type (72.3%) and had the highest mean score (M = 13.4, SD = 5.9), followed by emotional and physical abuse (61.9%, M = 11.8, SD = 5.5, and 61.9%, M = 10.8, SD = 5.5, respectively). Physical neglect (51.0%; M = 8.7, SD = 3.6) and sexual abuse (39.7%; M = 6.8, SD = 3.5) were reported less frequently and had lower mean scores.

3.2. Unique variable analysis

Initial UVA revealed several redundant pairs, despite their theoretical distinctness (wTO > .30; Christensen et al., 2023): emotional abuse-emotional neglect (EA-EN, wTO = .37), emotional abuse-physical abuse (EA-PA, wTO = .32), physical neglect-emotional neglect (PN-EN, wTO = .34), negative self-concept-disturbances in relationships (NSC-DR, wTO = .37), and re-experiencing-avoidance (RE-AV, wTO = .33). In the two recommended approaches to handle redundancies, composition and removal (Christensen et al., 2020), new redundant pairs emerged after each intervention, leading to excessive data reduction. The composition method ultimately merged all trauma types except sexual abuse into one cluster and all CPTSD symptoms into another. The removal method retained only physical neglect, sexual abuse, and re-experiencing.

We considered limiting redundancy handling to the first iteration. However, item-level analysis showed that while initial between-variable redundancies were not present (wTO < .20), within-category redundancies emerged. These within-category redundancies would have produced our current aggregated variables through composition, creating uncertainty about where to stop the process. Given that these variables represent theoretically distinct constructs and that both redundancy handling methods would oversimplify the complex relationships in our data, we retained all original nodes in our network analysis.

3.3. Exploratory graph analysis

EGA identified three distinct communities within the symptom network (Figure 1). The first community comprised five childhood trauma types from CTQ. The second community included PTSD symptoms: re-experiencing, avoidance, and sense of threat, from ITQ. The third community comprised DSO symptoms: negative self-concept, disturbances in relationships, and affect dysregulation, from ITQ.

Figure 1.

Network constructed via the graphical LASSO, depicting regularized partial correlations between trauma types from CTQ and CPTSD symptoms from ITQ. Ring segments around nodes represent predictability (shared variance) of each node.

Notes: PA = physical abuse; PN = physical neglect; EA = emotional abuse; EN = emotional neglect; SA = sexual abuse; RE = re-experiencing; TH = sense of threat; AV = avoidance; NSC = negative self-concept; DR = disturbances in relationships; AD = affect dysregulation.

3.4. Regularized partial correlation network (Graphical LASSO)

The regularized partial correlation network is depicted in Figure 1. With regularization of the EBIC glasso, 24 edges were retained from 55 possible edges (43.6%), with edge weights ranging from – .03 to .47. All retained edges were positively associated, except for a negative partial correlation between emotional neglect and re-experiencing (edge weight – .03). The strongest connections among trauma types emerged between emotional neglect and emotional abuse (edge weight .44), emotional neglect and physical neglect (edge weight .38), and emotional abuse and physical abuse (edge weight .37). For CPTSD symptoms, the strongest connections were between negative self-concept and disturbances in relationships (edge weight .47) and between re-experiencing and avoidance (edge weight .38). The CS-coefficient of the edge weights was .75, indicating highly robust network stability (Epskamp & Fried, 2018; see Supplementary Figure S1).

3.5. Predictability

Predictability was calculated and visualized as ring segments around each node in the regularized partial correlation network (Figure 1). Consistent with both centrality and bridge centrality findings, emotional neglect (67.9%) and emotional abuse (66.3%) showed the highest shared variance with other nodes in the network; physical abuse (49.3%) and physical neglect (35.5%) showed moderate shared variance; and sexual abuse (5.6%) showed markedly lower shared variance. Among DSO symptoms, negative self-concept (53.1%) and disturbances in relationships (52.0%) showed the highest shared variance, while affect dysregulation showed comparatively lower shared variance (40.7%). PTSD symptoms demonstrated moderate shared variance, with re-experiencing (50.7%), avoidance (48.7%), and sense of threat (47.4%).

3.6. Node and bridge centrality

In the centrality analysis, the stability coefficient revealed robust stability for strength and expected influence (EI) (CS-coefficient .75) and moderate stability for closeness (CS-coefficient .29), while betweenness demonstrated poor stability (CS-coefficient .05) and was therefore excluded from further interpretation (see Supplementary Figure S2; Epskamp & Fried, 2018). As depicted in Figure 2, strength and EI showed highly similar patterns, where emotional neglect emerged as the most central node, followed by emotional abuse.

Figure 2.

Centrality plot of node strength, expected influence, and closeness.

Notes: PA = physical abuse; PN = physical neglect; EA = emotional abuse; EN = emotional neglect; SA = sexual abuse; RE = re-experiencing; TH = sense of threat; AV = avoidance; NSC = negative self-concept; DR = disturbances in relationships; AD = affect dysregulation.

Among symptoms, negative self-concept emerged as the most central node in strength and EI, followed by disturbances in relationships and sense of threat, while negative self-concept also showed the highest centrality in closeness. However, the bootstrapped difference test across the indices revealed that many of these relative positions were not statistically significant. Specifically, negative self-concept was not significantly different from emotional abuse, re-experiencing, disturbances in relationships, sense of threat, affect dysregulation, and avoidance in strength; emotional neglect, emotional abuse, re-experiencing, disturbances in relationships, sense of threat, and affect dysregulation in EI; and all except sexual abuse in closeness centrality (see Supplementary Figure S3).

In the bridge centrality analysis, the stability coefficient demonstrated moderate stability for bridge strength and expected influence (1-step and 2-step) (CS-coefficient .59) and bridge closeness (CS-coefficient .36), while bridge betweenness showed poor stability (CS-coefficient .05) and was therefore excluded from interpretation (see Supplementary Figure S4; Epskamp & Fried, 2018).

As presented in Figure 3, among the five trauma types, emotional abuse consistently emerged as the most influential bridge, followed by emotional neglect across the indices. In symptoms, the bridge centrality identified three key bridging variables: affect dysregulation, sense of threat, and negative self-concept. In both bridge strength and one-step expected influence (EI1), affect dysregulation emerged as the strongest bridge, followed by sense of threat and negative self-concept. Interestingly, while affect dysregulation maintained its position as the strongest bridge in two-step expected influence (EI2), negative self-concept rose to the second rank, surpassing sense of threat. Negative self-concept also emerged as the highest node in bridge closeness.

Figure 3.

Bridge centrality plot of bridge strength, expected influence (1-step and 2-step), and closeness, in order of value.

Notes: PA = physical abuse; PN = physical neglect; EA = emotional abuse; EN = emotional neglect; SA = sexual abuse; RE = re-experiencing; TH = sense of threat; AV = avoidance; NSC = negative self-concept; DR = disturbances in relationships; AD = affect dysregulation.

3.7. Bayesian network

Figure 4 presents a directed acyclic graph (DAG) averaging 1000 bootstrapped networks (optimized threshold = .48; Scutari & Nagarajan, 2013) with edge thickness representing bootstrap strength. Several key findings emerged. First, among the five trauma types, emotional trauma (both neglect and abuse) showed the strongest edge direction towards CPTSD symptoms, with the DAG depicting a path from emotional neglect to emotional abuse and into the symptoms. Second, negative self-concept bridges trauma experiences and CPTSD symptoms, potentially triggering cascading effects on the symptom network. Third, affect dysregulation appears to be a converging variable for multiple symptom pathways.

Figure 4.

Directed acyclic graph (DAG) depicting trauma types from CTQ and CPTSD symptoms from ITQ with Scutari and Nagarajan (2013) method.

Notes: PA = physical abuse; PN = physical neglect; EA = emotional abuse; EN = emotional neglect; SA = sexual abuse; RE = re-experiencing; TH = sense of threat; AV = avoidance; NSC = negative self-concept; DR = disturbances in relationships; AD = affect dysregulation.

The acyclic constraint of a DAG prevented the representation of several important relationships. McNally, Mair et al. (2017) suggest directional probabilities close to 50% may indicate bidirectional relationships. Edges with potential bidirectional relationships were found between sense of threat to affect dysregulation and avoidance (directions 51% and 52%, respectively). Furthermore, despite strong connection metrics between re-experiencing and sense of threat (strength 97%; direction 58%), this relationship was not represented due to the acyclic nature of the model. These ‘hidden loops,’ while not visible in the DAG, were considered when interpreting the results.

4. Discussion

This study examined the network structure of childhood trauma types and ICD-11 CPTSD symptoms in individuals with interpersonal childhood trauma. We employed partial correlation and Bayesian networks to generate preliminary insights into structural relationships, potential pathways, and key variables influencing the symptom network. By incorporating prototypical trauma associated with CPTSD into our analysis, we attempted to develop more comprehensive CPTSD models while addressing both the sensitivity of network estimation and the omission of key etiological components (Borsboom et al., 2021).

In examining the overall symptom structure, exploratory graph analysis (EGA) demonstrated that PTSD and DSO symptoms formed distinct clusters, supporting the ICD-11's conceptual framework. This finding is consistent with both network analysis (Levin et al., 2021; Liu et al., 2024; McElroy et al., 2019) and traditional factor analysis studies examining CPTSD (Litvin et al., 2017; Rácz et al., 2022; Tian et al., 2021).

In our analysis of trauma types, emotional abuse and neglect emerged as the most influential nodes in the centrality and bridge centrality measures. Predictability analysis similarly revealed their highest shared variance with other nodes in the network. The DAG further depicts edge direction from emotional neglect to emotional abuse, which then bridges trauma types and CPTSD symptoms. These findings align with network studies that consistently identify emotional abuse as having the strongest centrality among childhood trauma types (Güreşen & Dereboy, 2024; Huang et al., 2024; Volgenau et al., 2023; Zhang et al., 2024), along with factor analysis and latent class analysis identifying emotional abuse as the most significant predictor of CPTSD (Haselgruber et al., 2020; Lortye et al., 2024; Tian et al., 2021, 2024).

Among symptoms, negative self-concept and affect dysregulation stood out across metrics. Negative self-concept emerged as the most central symptom in all centrality indices and showed the highest bridge closeness and increasing bridge expected influence, with 53.1% predictability. This aligns with other network studies identifying its central role in CPTSD (Knefel et al., 2019, 2020; Levin et al., 2021; Liang & Yang, 2023). Conversely, affect dysregulation showed the strongest bridge strength and bridge expected influence, but displayed mid to lower influence across centrality indices, with comparatively lower predictability (40.7%).

The DAG provided complementary directional and positional insights. Within the DAG, negative self-concept bridged trauma and subsequent symptoms. This positioning, combined with its high centrality and increasing importance in bridge centrality, suggests it may act as a key receiver from the trauma network and an activator of the overall symptom network. This is consistent with temporal network findings demonstrating negative self-concept with high outward influence (‘provider’) (Liu et al., 2024). In contrast, affect dysregulation appears as a convergence point for multiple symptom pathways. As a DSO symptom, its strong connection to downstream PTSD variables may explain its high bridge centrality but lower centrality importance.

Given our findings and their consistency with previous literature, negative self-concept seems to emerge as a potential key target for intervention to prevent symptom cascades, rather than affect dysregulation. However, several findings warrant caution in this interpretation. First, bootstrapped difference tests in centrality revealed that these key variables did not significantly differ from many other nodes. Second, initial UVA results showed redundancies among the key variables. While such overlap is theoretically sensible (e.g. it is difficult to imagine emotional abuse occurring without emotional neglect), these findings highlight the close interconnection between variables, precluding definitive conclusions about specific targets.

Nevertheless, our preliminary findings suggest a potentially useful direction for intervention development: upstream variables, especially negative self-concept and disturbances in relationships, may be more strategic intervention targets; while downstream variables: affect dysregulation, sense of threat, re-experiencing, and avoidance, could serve as maintaining factors by forming a feedback loop within the symptom network.

Notably, Yang et al. (2023), who conducted an item-level DAG analysis of CPTSD symptoms, reported a relatively reversed directional pattern compared with our results. In their analysis, avoidance and sense of threat were positioned upstream, while negative self-concept and disturbances in relationships appeared downstream. They suggested that the two most central symptoms served different positional roles: sense of threat (‘hypervigilance’) as an activator of other symptoms and negative self-concept (‘feelings of failure’) as a symptom activated by other symptoms in the network.

To investigate this pattern, we conducted a supplementary DAG analysis at the item-level, excluding trauma variables (see Supplementary Figure S5). Our result converged with the pathways reported by Yang et al. (2023), except for disturbances in relationships positioned upstream rather than downstream. This discrepancy may reflect differences in trauma exposure, as our sample comprised only individuals with interpersonal childhood trauma, whereas Yang et al. (2023) included individuals with both interpersonal and non-interpersonal trauma. The consistency between symptom-only analyses despite different trauma backgrounds, which contrasts with our earlier analysis including trauma context, suggests that childhood trauma experiences may function as an anchor, a reference point, revealing trauma-contingent symptom relationships. This highlights the importance of considering trauma context when examining CPTSD symptoms in future Bayesian network analyses.

4.1. Limitations and future directions

Although we employed complementary analyses to enhance methodological rigour, the cross-sectional design and the assumptions underlying centrality and DAGs warrant caution in interpreting the findings (Briganti et al., 2024; Epskamp, Waldorp et al., 2018). The moderate sample size (N = 249) and reliance on self-report further suggest that results should be viewed as hypothesis-generating rather than confirmatory.

Questions of ergodicity – the validity of generalizing between-person analyses to within-person dynamics – also warrant attention (Fisher et al., 2018; Spiller et al., 2020). While perspectives vary on whether strict ergodicity is necessary (Adolf & Fried, 2019; van Borkulo et al., 2016) or exists on a continuum (Brose et al., 2015; Voelkle et al., 2014), emerging idiographic network approaches offer promising avenues for examining individual-level processes and developing personalized interventions (Hoekstra et al., 2024; van der Tuin et al., 2023). Additionally, while maintaining original nodes allowed their unique contributions to be examined, topological overlap remains a concern (Fried & Cramer, 2017; McNally, 2021).

DAG models also rest on strong assumptions of the absence of cycles and unmeasured confounders (Briganti et al., 2023; McNally, 2021; Ryan et al., 2022). Although we attempted to address bidirectionality by calculating strength and direction rather than relying solely on a DAG presentation, self-feedback and longer loops remain difficult to capture. Moreover, despite incorporating trauma experiences, the complexity of psychopathology suggests that some influential variables likely remained unmeasured. Lastly, while we used an optimized threshold, Briganti et al. (2023) recommend adopting the .85 threshold of Sachs et al. (2005) to improve specificity and edge retention. Given these limitations, future studies should consider employing longitudinal data to better understand the causal and dynamic nature of CPTSD networks or implementing the recommended threshold to enhance precision for Bayesian networks with cross-sectional data.

5. Conclusion

Despite its exploratory nature and limitations, the present study offers several key contributions. First, while preliminary, our findings regarding the importance of upstream variables suggest directions for future research and potential intervention targets. Second, by incorporating trauma types into our analysis, we demonstrate how trauma context serves as a critical anchor, revealing trauma-contingent symptom relationships. Third, the use of complementary methodologies strengthened inference by integrating converging evidence.

Funding Statement

This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) under a grant funded by the Korean government (MSIT) [No. RS-2023-00224386, Psychological risk factor (stress) mitigation content authoring tool technology, 50%]; and National Research Foundation (NRF), Korea, under project BK21 FOUR.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used Claude-3.5-Sonnet in order to enhance language clarity and assist with proofreading. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Ethical approval

The study obtained ethical approval from the Korea University Institutional Review Board (protocol code: IRB No. 2023-0338-03 and date of approval: 19 December 2023). All participants provided online written informed consent for their participation in the study.

Data availability statement

The raw data cannot be made publicly available due to participant privacy protections. For transparency and reproducibility, we have developed a covariance matrix. The matrix, together with the R script, is openly available on OSF at https://osf.io/p8bvc/?view_only = f21cfe0b378b496e94287b4a8bc8f11e

Supplemental Material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/20008066.2025.2571365.