Introduction
Conflict-affected populations are disproportionately burdened by mental health disorders, with depression and post-traumatic stress disorder (PTSD) among the most prevalent and debilitating conditions (Steel et al., Reference Steel, Chey, Silove, Marnane, Bryant and Van Ommeren2009; Charlson et al., Reference Charlson, Flaxman, Ferrari, Vos, Steel and Whiteford2016, Reference Charlson, van Ommeren, Flaxman, Cornett, Whiteford and Saxena2019). Forced displacement, exposure to violence and chronic post-displacement stressors contribute to a complex array of psychological symptoms among refugees (Miller and Rasmussen, Reference Miller and Rasmussen2017; Tinghög et al., Reference Tinghög, Malm, Arwidson, Sigvardsdotter, Lundin and Saboonchi2017). Contemporary refugee mental health models emphasize that psychological distress is shaped not only by past traumatic experiences but also by ongoing adversities such as economic insecurity, social disruption, restricted mobility, uncertainty regarding legal status, and limited access to opportunities and services (Miller and Rasmussen, Reference Miller and Rasmussen2017; Silove et al., Reference Silove, Ventevogel and Rees2017). This perspective is particularly relevant for Rohingya refugees living in protracted displacement settings, where daily stressors may persist long after the initial traumatic events. Understanding the potential interactions between symptoms of depression and PTSD is critical for developing targeted interventions aimed at mitigating mental health consequences in conflict-affected settings.
Depression and PTSD frequently co-occur in populations exposed to traumatic events, with high comorbidity rates reported across refugee and conflict-affected populations (Hoppen et al., Reference Hoppen, Priebe, Vetter and Morina2021; Nickerson et al., Reference Nickerson, Schick, Schnyder, Bryant and Morina2017; Tinghög et al., Reference Tinghög, Malm, Arwidson, Sigvardsdotter, Lundin and Saboonchi2017). Meta-analytic evidence indicates that depression and PTSD are among the most prevalent mental health conditions in conflict-affected populations globally (Hoppen et al., Reference Hoppen, Priebe, Vetter and Morina2021). The substantial overlap between trauma-related and depressive symptomatology suggests that these disorders may share common psychological processes and mutually reinforcing symptom patterns (Nickerson et al., Reference Nickerson, Schick, Schnyder, Bryant and Morina2017). Both disorders share overlapping symptoms, including negative mood, concentration difficulties and sleep disturbances, complicating the diagnosis and treatment (Choi et al., Reference Choi, Batchelder, Ehlinger, Safren and O’Cleirigh2017). Traditional latent variable models conceptualize these disorders as distinct entities causing observed symptoms. However, this perspective may overlook the dynamic symptom interactions that sustain and exacerbate these psychopathologies (Borsboom, Reference Borsboom2017).
Network analysis offers a promising framework to examine the complex, reciprocal relationships among individual symptoms of mental disorders (Epskamp et al., Reference Epskamp, Borsboom and Fried2018). Unlike conventional approaches, network models conceptualize symptoms as mutually interacting elements rather than passive indicators of an underlying disorder (Fried et al., Reference Fried, van Borkulo, Cramer, Boschloo, Schoevers and Borsboom2017). This approach allows identification of central symptoms that may serve as intervention targets and bridge symptoms that link comorbid conditions (Jones et al., Reference Jones, Ma and McNally2021). Recent network analyses in trauma-exposed populations have revealed distinct symptom clusters and highlighted key symptoms bridging PTSD with depression, anxiety and other disorders (Armour et al., Reference Armour, Fried, Deserno, Tsai and Pietrzak2017; McNally et al., Reference McNally, Heeren and Robinaugh2017a; Malgaroli et al., Reference Malgaroli, Maccallum and Bonanno2018; Price et al., Reference Price, Legrand, Brier and Hébert-Dufresne2019; Lazarov et al., Reference Lazarov, Suarez-Jimenez, Levi, Coppersmith, Lubin, Pine, Bar-Haim, Abend and Neria2020).
For instance, a study among treatment-seeking veterans identified feeling distant or cut off from others as the most central symptom, followed by feeling very upset when reminded of the event and repeated disturbing memories or thoughts of the event. Concentration difficulties and anhedonia were among the five most central symptoms in the combined network (Duek et al., Reference Duek, Spiller, Pietrzak, Fried and Harpaz-Rotem2021). Another study found role confusion, meaninglessness and loneliness as the most central symptoms of persistent complex bereavement disorder among conjugally bereaved individuals (Malgaroli et al., Reference Malgaroli, Maccallum and Bonanno2018). Longitudinal analyses following strict COVID-19 lockdowns reported irritability/anger as the most influential bridge symptom (Chen et al., Reference Chen, Bi, Lyu, Sun and Bonanno2022). Among sexual minorities with trauma histories, sleep and concentration problems emerged as the strongest connecting symptoms between depression and PTSD (Choi et al., Reference Choi, Batchelder, Ehlinger, Safren and O’Cleirigh2017). Similarly, Australian adults reported sleep problems, irritability, concentration problems and loss of interest as key overlapping symptoms between depression and PTSD (Afzali et al., Reference Afzali, Sunderland, Teesson, Carragher, Mills and Slade2017). These findings reinforce the importance of exploring the co-occurrence of depression and PTSD symptoms.
Complementing network analysis, directed acyclic graphs (DAGs) provide an advanced method for modelling directional dependencies among symptoms by estimating directed edges that imply probabilistic conditional relationships among variables (McNally et al., Reference McNally, Heeren and Robinaugh2017a, Reference McNally, Mair, Mugno and Riemann2017b; Chen et al., Reference Chen, Bi, Lyu, Sun and Bonanno2022). DAGs enable researchers to explore probabilistic directional dependencies among symptoms, moving beyond undirected associations while remaining hypothesis-generating in cross-sectional data (Pearl et al., Reference Pearl, Glymour and Jewell2016). This is particularly valuable in understanding psychopathology in trauma-exposed populations, where symptom dynamics may unfold over time or via causal cascades.
Despite growing interest, there remains a paucity of network analytic studies, including DAG applications, focusing on populations affected by protracted conflicts in low-resource settings, such as forcibly displaced Rohingya refugees and host communities in Bangladesh. The Cox’s Bazar district, home to over one million forcibly displaced persons, represents one of the largest humanitarian crises globally (UNHCR, 2025). The Rohingya constitute one of the world’s largest stateless populations and have experienced decades of discrimination, structural violence, religious persecution and systematic restrictions on fundamental rights in Myanmar (Tay et al., Reference Tay, Islam, Riley, Welton-Mitchell, Duchesne, Waters, Varner, Silove and Ventevogel2018). Many Rohingya refugees have been exposed to repeated episodes of displacement, restrictions on citizenship rights, limited access to education and healthcare, and widespread human rights violations. Following displacement to Bangladesh, most Rohingya refugees reside in highly crowded camps where movement outside camp boundaries is restricted, and opportunities for employment and social participation remain limited. These ongoing adversities may shape both the expression and organization of psychological distress and contribute to symptom patterns that differ from those observed in non-refugee populations (Tay et al., Reference Tay, Riley, Islam, Welton-Mitchell, Duchesne, Waters, Varner, Moussa, Mahmudul Alam, Elshazly, Silove and Ventevogel2019b). Emerging evidence also suggests that patterns of distress may differ across Rohingya populations residing in Bangladesh and other host countries such as Malaysia (Tay et al., Reference Tay, Rees, Miah, Khan, Badrudduza, Morgan, Fadil Azim, Balasundaram and Silove2019a; Khan and Haque, Reference Khan and Haque2021).
Depression and PTSD symptoms have been documented among Rohingya refugees, yet symptom-level investigations remain limited (O’Connor and Seager, Reference O’Connor and Seager2021; Ritsema and Armstrong-Hough, Reference Ritsema and Armstrong-Hough2023; Ullah et al., Reference Ullah, Pratley, Shariful Islam, Islam and Roy2023). Studies conducted in Bangladesh have shown that systematic human rights violations, trauma exposure, food insecurity, restricted freedom of movement, inadequate humanitarian assistance and limited employment opportunities are associated with poorer mental health outcomes (Riley et al., Reference Riley, Varner, Ventevogel, Taimur Hasan and Welton-Mitchell2017, Reference Riley, Akther, Noor, Ali and Welton-Mitchell2020; Hossain et al., Reference Hossain, Baten, Sultana, Rahman, Adnan, Hossain, Khan and Uddin2021). Similarly, research among Rohingya refugees in Malaysia highlights the cumulative effects of pre-migration trauma, post-migration living difficulties, and statelessness on mental health and functioning (Tay et al., Reference Tay, Rees, Miah, Khan, Badrudduza, Morgan, Fadil Azim, Balasundaram and Silove2019a). Collectively, these findings suggest that psychological distress among Rohingya refugees is shaped by the combined effects of trauma exposure and chronic displacement-related stressors. Understanding symptom networks and potential directional dependencies in this context is therefore important for developing culturally sensitive and contextually relevant mental health strategies.
The current study aimed to fill this gap by applying both psychological network analysis and DAG modelling to investigate the structure, connectivity and exploratory directional dependencies among depression and PTSD symptoms among participants residing in conflict-affected Cox’s Bazar and neighbouring Bandarban. Specifically, we aimed to identify central symptoms within and across disorders, bridge symptoms linking depression and PTSD, and evaluate the stability, accuracy and exploratory edge orientations of the symptom network. Findings are expected to enhance understanding of the complex interplay of symptoms in conflict-affected populations and inform more targeted and effective mental health interventions.
Methods
Study design and setting
This study used baseline data from the Cox’s Bazar Panel Survey (CBPS), a multi-year panel survey conducted from March 2019 to August 2019 in Cox’s Bazar and Bandarban districts of Bangladesh (Baird et al., Reference Baird, Davis, Genoni, Lopez-Pena, Mobarak, Palaniswamy, Seager and Vishwanath2021). The CBPS was implemented as a collaboration between Yale University, Innovations for Poverty Action Bangladesh, the World Bank and the Gender & Adolescence: Global Evidence program. It aims to assess the well-being, socio-economic conditions and mental health of forcibly displaced refugee populations alongside host communities in a conflict-affected region. The survey includes modules on trauma exposure and mental health symptoms, including depression and PTSD.
Participants and sampling
The CBPS baseline sample included 5,020 households drawn from 2 main populations: refugees residing in 27 officially recognized camps in Cox’s Bazar, comprising newly arrived refugees (since August 2017) and earlier arrivals, and host community households from six upazilas in Cox’s Bazar and one upazila in Bandarban. Sampling was designed to represent both groups equally, with approximately 2,600 households from camps and 2,535 from host communities.
Households within camps were selected using multistage probability sampling based on majhee blocks, which were randomly selected proportional to size. Host community households were sampled using administrative units called mauzas, segmented using geographic and census data, with random selection of households from each segment.
From each household, two participants aged 15 years or older were randomly selected to participate in adult interviews. The mental health and trauma modules were completed by 9,885 adults. After excluding participants with incomplete symptom data, 9,397 participants constituted the analytic sample for this study.
Data collection procedure
Before the main survey, a detailed household listing was conducted in all selected clusters between January and February 2019 to create sampling frames and collect locator information. The survey consisted of a household questionnaire administered to one knowledgeable adult household member, and an adult questionnaire generally administered to two randomly selected household members. The adult questionnaire included assessments of trauma exposure and mental health symptoms, specifically depression, measured by the Patient Health Questionnaire-9 (PHQ-9), and PTSD symptoms, assessed using the 16-item Harvard Trauma Questionnaire (HTQ-16).
Data collectors completed multiple rounds of comprehensive training covering ethical guidelines, trauma module administration and the use of computer-assisted personal interviewing devices. The training included classroom instruction, field practice and refresher sessions.
The main data collection was conducted by 16 teams of enumerators and supervisors between March 2019 and July 2019, with a 1-month pause during Ramadan. Quality control measures included daily tracking of survey completion, high-frequency data quality checks to monitor enumerator performance, backchecks to verify responses and spot checks by supervisors. Participants received a modest cash incentive upon completing the interview to encourage participation.
Measures
Depression
Depression was measured using the PHQ-9 over the past 2 weeks (Kroenke et al., Reference Kroenke, Spitzer and Williams2001). Each item represents a feature or symptom of depression: anhedonia, sad mood, sleep, energy, appetite, guilt, concentration, motor and suicidal ideation. The responses were recorded on a four-point Likert scale ranging from 0 (not at all) to 3 (nearly every day). The total score ranges from 0 to 27, where higher scores indicate higher depressive symptom severity.
Trauma exposure
Lifetime exposure to traumatic events was assessed via 12 items adapted from conflict-affected population surveys, capturing experiences relevant to the study context. The trauma exposure items included: imprisonment, injury, combat experience, rape, social isolation, witnessing the death of a close person, family separation, murder of a family member, unnatural death, murder by a stranger, being lost or displaced, and torture. Each trauma item was assessed with four response categories to capture varying levels of exposure – experienced: direct personal experience of the event; witnessed: observed the event happening to others; heard about it: learned about the event occurring to others; and no exposure: no experience or knowledge of the event.
Post-traumatic stress disorder
Post-traumatic stress symptoms were assessed using the HTQ-16 (Mollica et al., Reference Mollica, Caspi-Yavin, Bollini, Truong, Tor and Lavelle1992). The HTQ-16 measures symptoms of PTSD experienced over the past week, including re-experiencing, avoidance, numbing and hyperarousal. Each item describes a specific PTSD symptom, such as intrusive thoughts, nightmares, emotional numbness and heightened startle response. Responses were recorded on a four-point Likert scale ranging from 1 (not at all) to 4 (extremely), reflecting the degree to which each symptom bothered the respondent. The total score ranges from 16 to 64, with higher scores indicating greater symptom severity.
Ethical considerations
This study is a secondary analysis of publicly available, de-identified data from the CBPS. The original CBPS received ethical approval from the Innovations for Poverty Action Institutional Review Board (Protocol No. 14742) and the George Washington University Institutional Review Board (Protocol No. 071721). Verbal informed consent was obtained from all adult participants prior to data collection. The present study analysed anonymized publicly available data; therefore, additional ethical approval was not required.
Statistical analysis
All statistical analyses were conducted in R (version 4.4.1). Descriptive analysis was performed to measure the frequency and percentages of exposure to traumatic events among participants. Network analysis was employed to explore the complex relationships between depression and PTSD symptoms, with a focus on identifying central or bridge symptoms.
Data preparation
Stata data files were imported using the haven package. Twenty-five symptom variables, comprising nine depression symptoms (D1–D9) and 16 PTSD symptoms (PTSD1–PTSD16), were selected for analysis. Participants with missing data on one or more symptom variables were excluded using complete-case analysis, resulting in a final analytic sample of 9,397 participants (95.1% of the original sample). To evaluate the potential impact of missing data, participants included in the final analytic sample were compared with those excluded due to missing symptom data. Age differences were assessed using Welch’s t-test, whereas sex and settlement status (host community vs refugee camp) were compared using chi-square tests. The results of these comparisons are presented in Supplementary Table S1. Consistent with standard psychopathology network methodology, the network analyses focused on conditional dependencies among symptom variables and did not adjust for demographic covariates within the network estimation process. The primary objective was to characterize symptom-to-symptom relationships rather than model demographic or contextual determinants of symptom occurrence. Therefore, age, sex and settlement status were examined descriptively and through subsequent sensitivity analyses using the network comparison test (NCT), rather than included as nodes in the primary symptom network.
Network estimation
The network structure was estimated using the Extended Bayesian Information Criterion Graphical Lasso (EBICglasso) algorithm through the estimateNetwork() function in the bootnet package (Epskamp et al., Reference Epskamp, Borsboom and Fried2018). Given the ordinal nature of the PHQ-9 and HTQ symptom items, polychoric correlations were estimated using the cor_auto() function from the qgraph package before network estimation. The EBICglasso approach estimates a regularized partial correlation network and applies Least Absolute Shrinkage and Selection Operator (LASSO) regularization to reduce spurious associations while retaining the most robust symptom connections. The EBIC tuning parameter was set to 0.5, a commonly recommended value that balances sensitivity and specificity in psychological network estimation. The network was visualized using the qgraph package with the Fruchterman–Reingold layout. Nodes represented individual symptoms, whereas edges represented regularized partial correlations after adjusting for all remaining nodes. Edge thickness and colour saturation reflected the magnitude of the regularized partial correlations. To enhance interpretation, depression symptoms (D1–D9) were displayed in blue and PTSD symptoms (PTSD1–PTSD16) in coral. Node labels were customized using full symptom descriptions.
Centrality, predictability and bridge centrality
Centrality indices including strength, closeness and betweenness were computed using the centrality() function in the qgraph package and visualized with centralityPlot() (Epskamp et al., Reference Epskamp, Cramer, Waldorp, Schmittmann and Borsboom2012). These indices quantified how influential each symptom was within the network. Predictability was estimated using the mgm package and quantified as node-wise classification accuracy (CC), representing the extent to which each symptom could be predicted from its neighbouring nodes within the network (Haslbeck and Waldorp, Reference Haslbeck and Waldorp2020). To examine cross-domain connectivity, bridge centrality metrics were computed with the networktools package (Jones et al., Reference Jones, Ma and McNally2021). The bridge() function was used to compute bridge strength, bridge closeness, bridge betweenness and bridge expected influence (BEI). These metrics allowed the identification of key bridge symptoms linking depression and PTSD symptoms.
Network stability and accuracy
Network robustness was assessed via bootstrapping procedures using the bootnet package. Centrality stability was evaluated through a case-dropping procedure with 1,000 iterations. The correlation stability (CS) coefficient was computed to indicate how much of the sample could be dropped while maintaining a correlation ≥0.7 with original centrality values. A CS coefficient >0.50 is generally considered acceptable, whereas values above 0.70 indicate excellent stability. Edge-weight accuracy was assessed using nonparametric bootstrapping (1,000 samples), and 95% confidence intervals for edge weights were plotted. Narrower confidence intervals indicated more precise edge estimates, while wider confidence intervals, particularly those including zero, indicated greater uncertainty (Epskamp et al., Reference Epskamp, Borsboom and Fried2018).
Bayesian network (DAG) estimation
To examine exploratory directional dependencies between depression and PTSD symptoms, we estimated a Bayesian network structure, represented as a DAG, using the bnlearn R package (Scutari, Reference Scutari2010). All symptom variables were treated as categorical and converted to factors, ensuring compatibility with the BIC scoring function for discrete data. We employed the hill-climbing (HC) algorithm for structure learning, optimizing the Bayesian information criterion (BIC) to identify the most probable network configuration. To ensure the stability and robustness of the estimated network structure, we first conducted bootstrapping with 1,000 resamples for preliminary exploration. For the final analysis, we increased the number of bootstrap resamples to 10,000 (McNally et al., Reference McNally, Heeren and Robinaugh2017a). In each bootstrap iteration, the HC algorithm searched for the optimal network structure by adding, removing or reversing edges to maximize model fit. The final consensus DAG was computed by averaging across all bootstrap-derived networks. Following the procedure recommended by Scutari and Nagarajan (Reference Scutari and Nagarajan2013), we retained only those edges present in at least 50% of bootstrap samples (i.e. threshold = 0.5), balancing sensitivity and specificity (Scutari and Nagarajan, Reference Scutari and Nagarajan2013). Edge strength was defined as the proportion of bootstrap samples in which a given edge appeared, and direction probability was quantified as the probability that a directed edge consistently pointed from one node to another across bootstraps. The resultant DAG was visualized using the Rgraphviz package, displaying all consensus edges. For enhanced clarity, we also highlighted the strongest edges (edge strength > 0.85) in an additional network visualization (McNally et al., Reference McNally, Heeren and Robinaugh2017a). A complete summary of edge frequencies and direction probabilities is provided in the Supplementary Materials.
Sensitivity analyses
Sensitivity analyses were conducted using the NCT to evaluate whether the estimated symptom network differed according to settlement status (refugee camp vs host community) and sex (male vs female). Separate NCTs examined network structure invariance, global strength invariance and node strength centrality invariance using 1,000 permutations. These analyses were performed to assess the potential influence of key demographic characteristics on the observed symptom relationships.
Results
Missing data analysis
Of the 9,885 participants initially available for analysis, 9,397 (95.1%) had complete data on all depression and PTSD symptoms and were included in the network analyses, whereas 488 participants (4.9%) were excluded because of missing symptom data. Comparisons between included and excluded participants are presented in Supplementary Table S1. No significant differences were observed in age (p = 0.288) or settlement status (p = 0.267). A small difference was observed for sex (p = 0.036), with females comprising a slightly higher proportion of excluded participants than included participants. Given the low proportion of missing data and the minimal differences between groups, the complete-case sample was considered broadly representative of the original study population.
Network structure
The network structure illustrating the conditional dependencies among 25 symptoms of depression (PHQ-9, nodes labelled D1–D9) and PTSD (HTQ-16, nodes labelled PTSD1–PTSD16) is presented in Figure 1 (see Supplementary Table S2 for the strongest regularized partial correlations). Nodes representing depression symptoms are shown in blue, while PTSD symptoms are shown in orange. Edges indicate regularized partial correlations between symptoms, with edge thickness and colour intensity reflecting the strength of associations.
Network structure of depression and PTSD symptoms among conflict-affected adults in Cox’s Bazar, Bangladesh.
Figure 1 Long description
Network diagram with circular nodes labeled D1, D2, D3, D4, D5, D6, D7, D8 and D9 and circular nodes labeled PTSD1 through PTSD16. The nodes form a single network with D and PTSD labels intermingled across the layout. Many curved and straight connecting lines run between pairs of nodes, including connections among D labeled nodes, among PTSD labeled nodes and between D and PTSD labeled nodes. A text list on the right maps each code to a symptom name: D1, Anhedonia; D2, Sad mood; D3, Sleep; D4, Fatigue; D5, Appetite; D6, Guilty; D7, Concentration; D8, Motor; D9, Suicide ideation. PTSD1, Recurrent traumatic thoughts; PTSD2, Nightmares; PTSD3, Emotional detachment; PTSD4, Emotional numbing; PTSD5, Irritability; PTSD6, Avoiding social interaction; PTSD7, Hopelessness about future; PTSD8, Difficulty in new situation; PTSD9, Physical health problems; PTSD10, Inability to plan; PTSD11, Feeling misunderstood; PTSD12, Perceiving hostility; PTSD13, No support; PTSD14, Misuse; PTSD15, Powerlessness to help; PTSD16, Rumination.
Within the depression subnetwork, strong associations were observed among affective, cognitive and somatic symptoms. The strongest edge connected ‘Anhedonia’ (D1) and ‘Fatigue’ (D4), followed by a robust association between ‘Anhedonia’ (D1) and ‘Sad mood’ (D2), highlighting the central role of core depressive symptoms. Strong connections were also observed between ‘Sleep’ (D3) and ‘Appetite’ (D5), as well as between ‘Guilt’ (D6) and ‘Suicidal ideation’ (D9), suggesting close interrelationships among somatic symptoms and negative self-evaluative cognitions. Additionally, a strong association was observed between ‘Concentration’ (D7) and ‘Motor difficulties’ (D8).
The PTSD subnetwork also revealed several tightly connected symptom clusters. The strongest association was observed between ‘Lack of support’ (PTSD13) and ‘Mistrust’ (PTSD14), followed by the connection between ‘Recurrent traumatic thoughts’ (PTSD1) and ‘Rumination’ (PTSD16). Furthermore, ‘Recurrent traumatic thoughts’ (PTSD1) were strongly linked with ‘Nightmares’ (PTSD2), reflecting the clustering of re-experiencing symptoms. Strong associations were also observed between ‘Feeling misunderstood’ (PTSD11) and ‘Perceiving hostility’ (PTSD12), as well as between ‘Emotional detachment’ (PTSD3) and ‘Avoiding social interaction’ (PTSD6), highlighting interpersonal and social withdrawal dimensions of PTSD.
Cross-domain edges between depression and PTSD symptoms highlighted potential bridge symptoms linking the two disorders. Notable cross-domain associations included links between ‘Fatigue’ (D4) and ‘Physical health problems’ (PTSD9), ‘Guilt’ (D6) and ‘Rumination’ (PTSD16), ‘Guilt’ (D6) and ‘Recurrent traumatic thoughts’ (PTSD1), and ‘Anhedonia’ (D1) and ‘Physical health problems’ (PTSD9). These connections suggest that somatic complaints, trauma-related intrusions and maladaptive self-referential cognitions may contribute to symptom overlap between depression and PTSD.
Overall, the network revealed distinct yet interconnected clusters of depression and PTSD symptoms, with several cognitive, interpersonal and somatic symptoms serving as key points of connection between the two symptom domains.
Network centrality and predictability
Centrality indices were examined to identify the most influential symptoms within the combined depression and PTSD symptom network (Figure 2). Three centrality metrics were computed: strength, closeness and betweenness. However, interpretation focused primarily on strength centrality because it is generally considered the most stable and interpretable measure in psychological networks.
Network centrality for depression and PTSD symptoms.
Strength centrality quantifies the sum of absolute edge weights connected to a node and reflects the overall connectedness of a symptom within the network. The highest strength centrality was observed for ‘Guilt’ (D6; strength = 1.873), indicating that it was the most strongly connected symptom within the network. Other highly central symptoms included ‘Anhedonia’ (D1; strength = 1.539), ‘Recurrent traumatic thoughts’ (PTSD1; strength = 1.476), ‘Emotional detachment’ (PTSD3; strength = 1.454), ‘Hopelessness about future’ (PTSD7; strength = 1.435) and ‘Rumination’ (PTSD16; strength = 1.412). These findings suggest that these symptoms occupy influential positions within the symptom network (Supplementary Table S3). Closeness and betweenness centrality indices are presented in Figure 2 for completeness but were not interpreted substantively because these metrics are often less stable and less reliable in psychopathology networks.
Node-wise predictability values, estimated as CC using the mgm package, are presented in Supplementary Table S4. Predictability reflects the proportion of correctly classified responses for each symptom when predicted from the remaining symptoms in the model.
Overall, predictability values ranged from 0.542 to 0.791, indicating moderate to high levels of predictability across both depression and PTSD symptoms. Among the depression symptoms, ‘Motor difficulties’ (D8) exhibited the highest predictability (0.769), suggesting that this symptom was most accurately predicted by the remaining symptoms in the network. Other highly predictable depression symptoms included ‘Concentration difficulties’ (D7; 0.704) and ‘Anhedonia’ (D1; 0.691).
Within the PTSD domain, the highest predictability was observed for ‘Avoiding social interaction’ (PTSD6; 0.791) and ‘Emotional detachment’ (PTSD3; 0.701), indicating that these symptoms were strongly associated with other symptoms in the network. In contrast, ‘Powerlessness to help’ (PTSD15; 0.542) and ‘Irritability’ (PTSD5; 0.562) exhibited the lowest predictability values.
Bridge centrality
Bridge centrality analysis was conducted to identify symptoms that connect the depression and PTSD symptom communities (Figure 3). Although several bridge centrality metrics were calculated, interpretation focused primarily on bridge strength and one-step BEI, which are generally considered the most stable and clinically interpretable bridge indices in psychopathology networks (Supplementary Table S5).
Bridge centrality for depression and PTSD symptom communities.
Figure 3 Long description
A) Line graph with title text, Bridge Betweenness. Horizontal axis label not legible, tick labels 0, 10, 20, 30, 40. Vertical axis label not legible, category labels D1, D2, D3, D4, D5, D6, D7, D8, D9, PTSD1, PTSD2, PTSD3, PTSD4, PTSD5, PTSD6, PTSD7, PTSD8, PTSD9, PTSD10, PTSD11, PTSD12, PTSD13, PTSD14, PTSD15, PTSD16. A single polyline with point markers crosses the categories. B) Line graph with title text, Bridge Closeness. Horizontal axis label not legible, tick labels 0.05, 0.06, 0.07, 0.08, 0.09. Vertical axis label not legible, category labels D1, D2, D3, D4, D5, D6, D7, D8, D9, PTSD1, PTSD2, PTSD3, PTSD4, PTSD5, PTSD6, PTSD7, PTSD8, PTSD9, PTSD10, PTSD11, PTSD12, PTSD13, PTSD14, PTSD15, PTSD16. A single polyline with point markers crosses the categories. C) Line graph with title text, Bridge Expected Influence 1 step. Horizontal axis label not legible, tick labels 0.0, 0.2, 0.4, 0.6. Vertical axis label not legible, category labels D1, D2, D3, D4, D5, D6, D7, D8, D9, PTSD1, PTSD2, PTSD3, PTSD4, PTSD5, PTSD6, PTSD7, PTSD8, PTSD9, PTSD10, PTSD11, PTSD12, PTSD13, PTSD14, PTSD15, PTSD16. A single polyline with point markers crosses the categories. D) Line graph with title text, Bridge Expected Influence 2 step. Horizontal axis label not legible, tick labels 0.0, 0.25, 0.50, 0.75, 1.0. Vertical axis label not legible, category labels D1, D2, D3, D4, D5, D6, D7, D8, D9, PTSD1, PTSD2, PTSD3, PTSD4, PTSD5, PTSD6, PTSD7, PTSD8, PTSD9, PTSD10, PTSD11, PTSD12, PTSD13, PTSD14, PTSD15, PTSD16. A single polyline with point markers crosses the categories. E) Line graph with title text, Bridge Strength. Horizontal axis label not legible, tick labels 0.25, 0.50, 0.75. Vertical axis label not legible, category labels D1, D2, D3, D4, D5, D6, D7, D8, D9, PTSD1, PTSD2, PTSD3, PTSD4, PTSD5, PTSD6, PTSD7, PTSD8, PTSD9, PTSD10, PTSD11, PTSD12, PTSD13, PTSD14, PTSD15, PTSD16. A single polyline with point markers crosses the categories.
Among depression symptoms, ‘Guilt’ (D6) had the highest bridge strength (0.896), and the BEI was 0.274. This finding suggests that guilt may serve as a prominent factor linking depressive and PTSD symptom clusters. Other depressive symptoms with high bridge strength included ‘Motor difficulties’ (D8; 0.600), ‘Suicidal ideation’ (D9; 0.574), ‘Anhedonia’ (D1; 0.572) and ‘Fatigue’ (D4; 0.545).
Within the PTSD symptom community, ‘Physical health problems’ (PTSD9) demonstrated the highest bridge strength (0.580) and the highest BEI (0.580), indicating a particularly important role in connecting PTSD symptoms with depressive symptoms. Other notable PTSD bridge symptoms included ‘Recurrent traumatic thoughts’ (PTSD1; bridge strength = 0.449), ‘Rumination’ (PTSD16; bridge strength = 0.408) and ‘Feeling misunderstood’ (PTSD11; bridge strength = 0.377).
Network accuracy and stability
The accuracy and stability of the estimated network were evaluated using nonparametric and case-dropping bootstrap procedures. Edge-weight accuracy was assessed using a nonparametric bootstrap with 1,000 iterations. As shown in Supplementary Figure S1, the bootstrap mean estimates closely corresponded to the original sample estimates, indicating limited estimation bias. Confidence intervals were generally narrower for stronger edges and wider for weaker edges, suggesting greater precision in the estimation of the strongest symptom associations and greater uncertainty surrounding weaker connections.
The stability of the network was assessed using a case-dropping bootstrap procedure. As illustrated in Supplementary Figure S2, the CS coefficient was 0.750 for both edge weights and strength centrality, indicating excellent stability. These values exceed the recommended threshold of 0.50 and suggest that the estimated network structure and strength centrality indices were highly robust. Furthermore, the average correlation between the strength centrality estimates obtained from subset samples and those from the original sample remained high across all levels of case dropping, indicating that the relative ordering of symptom importance was largely preserved even when substantial proportions of the sample were removed.
Overall, these findings support the reliability and robustness of the estimated network structure and the identified central symptoms.
Bayesian network estimation
A Bayesian network was estimated using a bootstrapped HC algorithm to examine exploratory directional dependencies among depression and PTSD symptoms (Figure 4; Supplementary Table S6). In the resulting DAG, edges represent conditional dependencies between symptoms, whereas edge strength indicates the proportion of bootstrap samples in which a given edge was identified. Direction probabilities indicate the relative consistency of the estimated edge orientation across bootstrap samples and should be interpreted as exploratory rather than causal.
Averaged Bayesian network of depression and PTSD symptoms estimated using a bootstrapped HC algorithm.
Figure 4 Long description
Directed network diagram with labeled boxes connected by one way arrows. Layout: PTSD16 is near the top center. D1 is below PTSD16. D6 is to the right of D1. D7 is left of D1. D2 is below D6. D4 is below D1 and left of D2. D3 is below D2. D5 is below D4. PTSD1 is right of D3. D9 is right of PTSD1. PTSD9 is left of D4. D8 is left of PTSD4. PTSD11 and PTSD10 are at the far left. PTSD12 is at the lower left. PTSD4 is below D8. PTSD3 is below PTSD4. PTSD8 is left of PTSD6. PTSD6 is below PTSD3. PTSD5 is right of PTSD6. PTSD2 is right of PTSD5. PTSD14 is near the bottom center. PTSD13 is right of PTSD2. PTSD7 is at the right edge. PTSD15 is at the lower right. Arrow connections as drawn: PTSD16 points to PTSD1. PTSD16 points to D6. PTSD16 points to PTSD11. PTSD16 points to PTSD7. D1 points to D7. D1 points to D4. D1 points to D2. D6 points to D2. D6 points to D9. D2 points to D4. D2 points to D3. D4 points to D3. D4 points to D5. D7 points to D8. PTSD9 points to PTSD10. PTSD11 points to PTSD10. PTSD11 points to PTSD12. PTSD10 points to PTSD4. PTSD4 points to PTSD3. PTSD3 points to PTSD8. PTSD4 points to PTSD6. PTSD3 points to PTSD6. PTSD3 points to PTSD2. PTSD8 points to PTSD14. PTSD13 points to PTSD14. PTSD7 points to PTSD15. PTSD14 points to PTSD15. PTSD1 points to D9.
Within the depression symptom cluster, several highly stable edges were identified. Notably, anhedonia (D1) showed strong exploratory directional dependencies with sad mood (D2), fatigue (D4) and concentration difficulties (D7), with edge strengths ranging from 0.94 to 1.00. Similarly, fatigue (D4) was connected to sleep problems (D3) and appetite disturbances (D5), whereas guilt (D6) showed strong exploratory directional dependencies with sad mood (D2) and suicidal ideation (D9). These findings suggest that core affective and somatic symptoms may occupy relatively prominent positions within the depressive symptom structure.
Within the PTSD symptom cluster, several exploratory directional dependencies were observed. Recurrent traumatic thoughts (PTSD1) were strongly associated with nightmares (PTSD2), emotional detachment (PTSD3) showed an exploratory directional dependency with avoiding social interaction (PTSD6) and difficulty in new situations (PTSD8), and hopelessness about the future (PTSD7) demonstrated strong exploratory directional dependencies with lack of support (PTSD13) and powerlessness to help (PTSD15). In addition, feeling misunderstood (PTSD11) showed an exploratory directional dependency with perceiving hostility (PTSD12) and other interpersonal PTSD symptoms.
Several exploratory directional dependencies were observed between depression and PTSD symptom communities. Rumination (PTSD16) showed a strong exploratory directional dependency with guilt (D6), whereas physical health problems (PTSD9) were linked to fatigue (D4). Feeling misunderstood (PTSD11) also showed an exploratory directional dependency with motor difficulties (D8). These cross-community exploratory directional dependencies were broadly consistent with the bridge symptoms identified in the undirected network analysis and may represent potentially important patterns of symptom organization linking depression and PTSD symptom communities.
Most identified edges demonstrated high bootstrap support (edge strength > 0.85), indicating that these relationships were consistently recovered across resampled datasets (Figure 5). However, several edge orientations showed exploratory direction probabilities close to 0.50, indicating substantial uncertainty regarding the direction of these relationships. Consequently, the observed edge directions should be interpreted as exploratory directional dependencies rather than evidence of temporal or causal processes. Overall, the Bayesian network complements the undirected network findings by highlighting exploratory directional dependencies that generate hypotheses regarding symptom organization for future longitudinal studies.
Averaged Bayesian network with highly stable edges highlighted.
Figure 5 Long description
Directed network diagram with rectangular nodes labeled PTSD1, PTSD2, PTSD3, PTSD4, PTSD5, PTSD6, PTSD7, PTSD8, PTSD9, PTSD10, PTSD11, PTSD12, PTSD13, PTSD14, PTSD15, PTSD16 and D1, D2, D3, D4, D5, D6, D7, D8, D9. Arrows connect many node pairs and two line weights are visible, thinner arrows and thicker arrows. PTSD16 is near the top center. A thick curved arrow runs across the top from PTSD16 toward the left side and ends at PTSD11. PTSD16 has a thick downward arrow to D6. PTSD16 also has thick arrows to to PTSD1. Upper and central nodes include D1, D7, D6, D2, D4, D3 and D5. D1 has a thick arrow to D7 and a thick arrow to D4. D6 has a thick arrow to D2. D4 has a thick arrow to D3 and a thick arrow to D5. D3 has a thick arrow to D5. Left side nodes include PTSD11, PTSD10, PTSD9 and D8. PTSD11 has a thick arrow to PTSD10. PTSD11 also has a thick arrow to PTSD12. PTSD10 has a thick arrow to PTSD4. . Lower left and lower center nodes include PTSD12, PTSD8, PTSD4, PTSD3, PTSD6, PTSD5, PTSD2 and PTSD14. PTSD4 has a thin downward arrow to PTSD3. PTSD3 has a thick downward arrow to PTSD6. PTSD3 has a thin arrow to PTSD5. PTSD3 has a thin arrow to PTSD2. PTSD8 has a thine arrow to PTSD14. PTSD12 has a thick arrow to PTSD14. Right side nodes include PTSD7, PTSD13 and PTSD15. PTSD7 and PTSD13 are connected by a thick arrow. A thick curved arrow runs along the bottom edge from the left side toward the right and ends at PTSD15. A thick curved arrow runs along the left edge from the lower left toward the upper left and ends at PTSD11.
Sensitivity analyses
To evaluate the potential influence of demographic characteristics on the estimated symptom network, NCTs were conducted across settlement status and sex. The comparison between refugee camp participants and host-community participants revealed a significant difference in overall network structure (M = 0.137, p = 0.001), indicating that the pattern of symptom associations varied across settlement settings (Supplementary Table S7). However, no significant difference was observed in global network strength (S = 0.235, p = 0.297), suggesting comparable overall levels of symptom connectivity. Several symptoms exhibited significant differences in strength centrality, including guilt (D6), concentration difficulties (D7), suicidal ideation (D9), nightmares (PTSD2), difficulty adapting to new situations (PTSD8), perceiving hostility (PTSD12) and rumination (PTSD16) (Supplementary Table S8).
In contrast, the comparison between males and females showed no significant differences in network structure (M = 0.071, p = 0.144) or global network strength (S = 0.118, p = 0.646) (Supplementary Table S9). Only one symptom, avoiding social interaction (PTSD6), demonstrated a significant difference in strength centrality between sexes (Supplementary Table S10). Overall, these findings suggest that settlement status may influence the organization of symptom relationships, whereas the overall network structure appears largely invariant across sexes.
These findings indicate that the primary network results were generally robust across demographic subgroups, although some variation in symptom organization was observed between refugee camp and host-community participants.
Discussion
This study employed comprehensive psychological network analysis alongside Bayesian network modelling to elucidate the complex interrelations and exploratory directional dependencies linking depression and PTSD symptoms in a large sample of participants affected by a protracted conflict and living in Cox’s Bazar, Bangladesh. The estimated symptom network revealed distinct yet interconnected clusters of depression and PTSD symptoms, with specific symptoms such as guilt, rumination, recurrent traumatic thoughts, emotional detachment and anhedonia emerging as highly central and influential within the combined network. Bridge centrality analyses identified guilt and physical health problems as key symptoms linking depression and PTSD symptom communities. Furthermore, network stability and accuracy assessments supported the sampling stability of the estimated associations within the analysed sample. The Bayesian network (DAG) analysis provided additional insights by identifying exploratory directional dependencies involving anhedonia, guilt, rumination, recurrent traumatic thoughts and emotional detachment. Collectively, these findings enhance our understanding of the complex symptom interrelationships and directional dependencies underlying depression and PTSD in conflict-affected populations and highlight symptoms that may warrant further investigation in future intervention research.
The psychological network analysis revealed a coherent structure with distinct symptom clusters reflecting the underlying depressive and PTSD symptom domains. Within depression, somatic symptoms such as sleep problems, fatigue and appetite disturbances formed tightly connected subnetworks, consistent with prior network studies highlighting the salience of somatic manifestations in depression (Bringmann et al., Reference Bringmann, Lemmens, Huibers, Borsboom and Tuerlinckx2015; Jurado-González et al., Reference Jurado-González, García-Torres, Contreras, Muñoz-Navarro, González-Blanch, Medrano, Ruiz-Rodríguez, Cano-Vindel and Moriana2024). Cognitive-affective symptoms like anhedonia and sad mood were strongly interconnected, aligning with established clinical conceptualizations of core depressive features (Jurado-González et al., Reference Jurado-González, García-Torres, Contreras, Muñoz-Navarro, González-Blanch, Medrano, Ruiz-Rodríguez, Cano-Vindel and Moriana2024). Similarly, PTSD symptoms clustered around classic re-experiencing and avoidance domains, with symptoms such as recurrent traumatic thoughts, nightmares and emotional detachment forming a central core consistent with previous trauma-focused network analyses (Armour et al., Reference Armour, Fried, Deserno, Tsai and Pietrzak2017; McNally et al., Reference McNally, Heeren and Robinaugh2017a).
Centrality analyses further emphasized the influential role of specific symptoms within the network. Guilt had the highest estimated strength centrality. This is consistent with existing literature documenting guilt’s prominence in trauma-related psychopathology and its association with PTSD symptom severity (Kubany and Manke, Reference Kubany and Manke1995; Kip et al., Reference Kip, Diele, Holling and Morina2022). Other highly central symptoms included rumination, recurrent traumatic thoughts, emotional detachment, anhedonia and hopelessness about the future, all of which occupied highly central positions within the network. These symptoms collectively reflect cognitive, affective and trauma-related processes that may represent important features of psychological distress in conflict-affected populations. The prominence of these symptoms is consistent with cognitive models of depression and PTSD that emphasize maladaptive appraisals, repetitive negative thinking, emotional disengagement and diminished reward processing as important processes proposed to underlie psychopathology (Ehlers and Clark, Reference Ehlers and Clark2000; Beck, Reference Beck2002).
The prominence of guilt, rumination, recurrent traumatic thoughts, emotional detachment, anhedonia and hopelessness about the future may be particularly understandable in the context of prolonged displacement, uncertainty regarding legal status, restricted opportunities and ongoing exposure to daily stressors. Previous research among Rohingya refugees has shown that both trauma exposure and post-displacement stressors, including restrictions on movement, economic hardship, inadequate humanitarian assistance and concerns regarding safety and the future, contribute substantially to psychological distress (Riley et al., Reference Riley, Varner, Ventevogel, Taimur Hasan and Welton-Mitchell2017, Reference Riley, Akther, Noor, Ali and Welton-Mitchell2020; Hossain et al., Reference Hossain, Baten, Sultana, Rahman, Adnan, Hossain, Khan and Uddin2021). These findings are consistent with broader frameworks suggesting that refugee mental health is shaped by the interaction of traumatic experiences and chronic post-displacement adversity (Miller and Rasmussen, Reference Miller and Rasmussen2017; Silove et al., Reference Silove, Ventevogel and Rees2017).
The predictability analyses demonstrated that motor difficulties and avoidance of social interaction had the highest predictability values. Conversely, symptoms with lower predictability may represent unique clinical features influenced by factors external to the measured network or necessitate tailored interventions (Haslbeck and Waldorp, Reference Haslbeck and Waldorp2020).
An important contribution of this study lies in the identification of bridge symptoms that link depression and PTSD symptom clusters. Bridge centrality analyses identified guilt as the most prominent bridge symptom within the depression community and physical health problems as the most prominent bridge symptom within the PTSD community. Additional bridge symptoms included psychomotor difficulties, suicidal ideation, recurrent traumatic thoughts, rumination and feelings of being misunderstood. Within the PTSD community, physical health problems emerged as the strongest bridge symptom, followed by recurrent traumatic thoughts, rumination and interpersonal symptoms. These findings corroborate prior research emphasizing the pivotal role of bridge symptoms in sustaining comorbidity and symptom interaction between disorders (McNally et al., Reference McNally, Mair, Mugno and Riemann2017b; Jones et al., Reference Jones, Ma and McNally2021). Within the network framework, bridge symptoms are hypothesized to facilitate interactions between symptom communities (Robinaugh et al., Reference Robinaugh, Hoekstra, Toner and Borsboom2020). Notably, guilt’s dual role as a central and bridge symptom underscores its significance in the complex interplay between depressive and trauma-related symptoms, consistent with the clinical relevance of guilt in trauma-exposed populations (Kubany and Manke, Reference Kubany and Manke1995). Physical health problems’ role as a bridge symptom may reflect the somatic burden of PTSD and its capacity to exacerbate depressive symptomatology, highlighting the need for integrated somatic and psychological interventions (Afari et al., Reference Afari, Ahumada, Wright, Mostoufi, Golnari, Reis and Cuneo2014). These findings suggest that bridge symptoms may represent promising candidate intervention targets for future research. However, longitudinal and intervention studies are required to determine whether targeting these symptoms results in broader changes across the symptom network.
The Bayesian network analysis provided valuable insights into exploratory directional dependencies among symptoms, complementing the undirected network findings. The DAG identified several highly stable exploratory directional dependencies involving anhedonia, guilt, rumination, recurrent traumatic thoughts and emotional detachment. For instance, anhedonia showed exploratory directional dependencies with sad mood, fatigue and concentration difficulties, suggesting that it may occupy a relatively prominent position within the depressive symptom network. Similarly, stable exploratory directional dependencies were observed between rumination and guilt, and between guilt, sad mood and suicidal ideation, suggesting close associations between cognitive and affective symptoms. Within the PTSD symptom cluster, recurrent traumatic thoughts were strongly connected with nightmares and rumination, and emotional detachment showed exploratory directional dependencies with nightmares, irritability and avoidance-related symptoms, consistent with previously observed symptom associations (McNally et al., Reference McNally, Heeren and Robinaugh2017a; Chen et al., Reference Chen, Bi, Lyu, Sun and Bonanno2022). Cross-disorder exploratory directional dependencies, such as those linking rumination and guilt, suggest the close interconnections between PTSD and depressive symptoms at the symptom level. The high bootstrap edge strengths indicate that these exploratory directional dependencies were consistently identified across bootstrap samples. Overall, the Bayesian network analysis enriches understanding by identifying exploratory directional dependencies that generate hypotheses regarding symptom organization for future longitudinal research (McNally et al., Reference McNally, Heeren and Robinaugh2017a; Robinaugh et al., Reference Robinaugh, Hoekstra, Toner and Borsboom2020).
Sensitivity analyses indicated that symptom network structure differed between refugee camp and host-community participants, although overall network connectivity remained comparable. These findings suggest that contextual factors associated with displacement, living conditions, trauma exposure and access to resources may influence symptom organization while preserving overall network connectivity. The observed difference in network structure, despite similar global strength, suggests that displacement-related environments may influence how individual symptoms interact with one another rather than the overall intensity of symptom connectivity. This finding underscores the importance of considering contextual variation when interpreting symptom networks and developing mental health interventions in humanitarian settings. Several symptoms, including guilt, nightmares, rumination and concentration difficulties, differed in centrality across settlement groups, indicating potential heterogeneity in symptom importance across humanitarian contexts. These findings support the symptom-level focus of the primary network analyses while highlighting that settlement-related context may shape how symptoms are organized rather than the overall degree of symptom connectivity. In contrast, no significant differences in overall network structure or global connectivity were observed between males and females, suggesting that the overall organization of depression and PTSD symptoms was largely consistent across sexes.
Practical implications
Although the present findings derive from cross-sectional data and thus cannot establish causal relationships, they offer valuable practical guidance for mental health practitioners working in conflict-affected and resource-constrained settings. The identification of central symptoms such as guilt, rumination, recurrent traumatic thoughts, emotional detachment, anhedonia and hopelessness about the future suggests that these symptoms occupy highly influential positions within the broader depression–PTSD symptom network. Clinicians may therefore consider paying particular attention to these symptoms during assessment and intervention planning. However, whether interventions targeting these symptoms lead to broader changes across the symptom network remains to be established in future longitudinal and intervention studies (Kubany and Manke, Reference Kubany and Manke1995; Fried et al., Reference Fried, van Borkulo, Cramer, Boschloo, Schoevers and Borsboom2017). The identification of bridge symptoms highlights the importance of integrated approaches to the assessment and treatment of depression and PTSD. In particular, guilt and physical health problems emerged as the strongest bridge symptoms connecting the two symptom communities, suggesting that these symptoms may play important roles in the co-occurrence of depressive and trauma-related distress. Additional bridge symptoms, including psychomotor difficulties, suicidal ideation, recurrent traumatic thoughts and rumination, may also contribute to cross-disorder symptom interactions. Interventions addressing both psychological and somatic manifestations of distress may warrant further evaluation in humanitarian settings where comorbidity is common, and resources are limited (Afari et al., Reference Afari, Ahumada, Wright, Mostoufi, Golnari, Reis and Cuneo2014).
The present findings also support transdiagnostic intervention approaches that focus on interconnected symptoms rather than diagnostic categories alone. The identification of central and bridge symptoms spanning both depression and PTSD domains suggests that these symptoms may serve as candidate markers for future intervention development, and that interventions targeting shared symptom processes warrant further empirical evaluation. This perspective is consistent with Integrative Adapt Therapy, a culturally adapted intervention that has demonstrated promising outcomes among Rohingya refugees in Bangladesh and Malaysia (Mahmuda et al., Reference Mahmuda, Miah, Elshazly, Khan, Tay and Ventevogel2019; Tay et al., Reference Tay, Mung, Miah, Balasundaram, Ventevogel, Badrudduza, Khan, Morgan, Rees, Mohsin and Silove2020, Reference Tay, Miah, Khan, Mohsin, Alam, Ozen, Mahmuda, Ahmed, Silove and Ventevogel2021). By targeting common mechanisms underlying multiple symptoms, such approaches may be especially well suited to humanitarian settings characterized by high levels of comorbidity and limited mental health resources.
Although causal inferences cannot be drawn, the Bayesian network analysis generated hypotheses regarding exploratory directional dependencies among symptoms. Several highly stable exploratory directional dependencies involved anhedonia, guilt, rumination, recurrent traumatic thoughts and emotional detachment, suggesting that these findings may help identify candidate symptoms for future longitudinal and intervention studies. Therapeutic approaches incorporating behavioural activation, cognitive restructuring and trauma-focused strategies targeting these symptoms may represent promising avenues for future research (Robinaugh et al., Reference Robinaugh, Hoekstra, Toner and Borsboom2020).
Moreover, symptoms demonstrating high predictability and strong integration within the network, particularly psychomotor difficulties (D8) and avoidance of social interaction (PTSD6), may serve as useful indicators for treatment monitoring and clinical evaluation. Monitoring changes in these highly predictable symptoms could help clinicians assess treatment progress and optimize resource allocation in routine practice.
Finally, sensitivity analyses demonstrated that symptom network structure differed across settlement settings while remaining largely invariant across sexes. These findings underscore the importance of considering contextual factors associated with displacement and living conditions when designing and implementing mental health interventions for conflict-affected populations.
Despite the inherent limitations of cross-sectional designs, the present symptom-level insights provide potentially useful information for clinicians and program planners aiming to optimize mental healthcare in humanitarian contexts. Collectively, the results highlight several candidate symptom targets that may help inform more focused, integrated and culturally responsive intervention strategies.
Strength and limitations
This study has several notable strengths that enhance its contribution to understanding depression and PTSD symptomatology in conflict-affected populations. First, it utilized a large, representative sample of adults drawn from both forcibly displaced Rohingya refugees and host communities, supporting robust estimation of symptom networks within a protracted humanitarian setting. Second, the combined use of psychological network analysis and Bayesian network modelling provided complementary perspectives on symptom interrelationships, enabling the identification of central symptoms, bridge symptoms and exploratory directional dependencies beyond traditional latent variable approaches. Third, rigorous assessments of network accuracy, stability and predictability supported the robustness and reliability of the reported findings.
Despite these strengths, several limitations warrant consideration. The cross-sectional design precludes conclusions regarding temporal ordering or causality among symptoms. Consequently, the exploratory directional dependencies identified in the Bayesian network should be interpreted as exploratory and hypothesis-generating rather than causal. Furthermore, although the proportion of missing symptom data was low (4.9%) and comparisons between included and excluded participants revealed only minimal demographic differences, the possibility of residual bias cannot be entirely excluded. Measurement limitations include reliance on self-reported symptom scales, which may be subject to reporting biases and cultural variation in symptom expression. In addition, symptoms and contextual factors not included in the present analyses may contribute to depression and PTSD symptom dynamics. The study’s focus on a specific humanitarian setting may also limit the generalizability of findings to other conflict-affected or displaced populations. Although sensitivity analyses were conducted across settlement status and sex, the primary network models were estimated using symptom variables only and did not explicitly adjust for demographic characteristics. Consequently, some symptom associations may partly reflect shared relationships with demographic or contextual factors. Future network studies incorporating trauma exposure, displacement-related stressors and other contextual variables as additional nodes or covariates may further improve understanding of symptom organization in conflict-affected populations. Finally, some symptoms demonstrated lower predictability, suggesting that unmeasured variables or external factors may influence symptomatology beyond the modelled network.
An additional limitation relates to cultural and contextual specificity. The expression, interpretation and communication of psychological distress among Rohingya refugees may differ from those observed in Western populations and may be influenced by local cultural idioms, collective experiences of persecution and the social realities of displacement (Tay et al., Reference Tay, Riley, Islam, Welton-Mitchell, Duchesne, Waters, Varner, Moussa, Mahmudul Alam, Elshazly, Silove and Ventevogel2019b). Consequently, the identified symptom network may not generalize to other refugee populations, conflict-affected settings or non-refugee populations that differ in cultural, historical and sociopolitical context.
Future research should prioritize longitudinal designs to examine the temporal evolution of symptom networks and to evaluate the exploratory directional dependencies identified by Bayesian network modelling. Such studies may help clarify symptom progression patterns, mechanisms underlying depression–PTSD comorbidity and evaluate whether the candidate intervention targets identified in the present network are associated with improved clinical outcomes. Replication across diverse conflict-affected populations would help determine the generalizability of the observed network structure and bridge symptoms. Furthermore, integrating biological, social and contextual variables into future network models may improve understanding of external influences on symptom dynamics and enhance ecological validity. Finally, intervention studies informed by network analytic findings are needed to evaluate whether targeting central and bridge symptoms yields improved clinical outcomes, particularly in resource-limited humanitarian settings.
Conclusion
This study advances the understanding of depression and PTSD symptomatology in conflict-affected populations by applying complementary network analytic approaches to identify influential symptom relationships and exploratory directional dependencies. The findings highlight guilt, rumination, recurrent traumatic thoughts, emotional detachment, anhedonia, hopelessness about the future and physical health problems as particularly important symptoms within the depression–PTSD network. In particular, guilt and physical health problems emerged as prominent bridge symptoms linking depressive and trauma-related symptom communities, underscoring their potential relevance for future integrated intervention development.
Although the cross-sectional nature of the data precludes causal inference, the identified symptom network demonstrated excellent stability and moderate-to-high predictability, supporting the robustness of the findings. The results provide clinically informative insights for assessment, intervention planning and program development in humanitarian settings. Future longitudinal and intervention studies are needed to evaluate the temporal relationships suggested by the network models and to determine whether interventions focusing on these candidate central and bridge symptoms improve mental health outcomes among forcibly displaced and vulnerable populations. Ultimately, a symptom-focused approach may contribute to more precise, contextually informed and culturally responsive mental healthcare in complex conflict settings.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S2045796026101152.
Availability of data and materials
The dataset associated with the manuscript is available from the following website: https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi%3A10.7910/DVN/V614VB.
Acknowledgements
The authors thank all study participants for their valuable contributions to this research. Generative AI (OpenAI ChatGPT) was used for language refinement and editing of the manuscript. All AI-generated suggestions were reviewed and approved by the authors, who take full responsibility for the content and accuracy of the manuscript.
Financial support
Dr ALmerab is currently receiving funding support from the Princess Nourah bint Abdulrahman University Researchers Supporting Project (No. PNURSP2026R563), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.
Competing interests
None.
Ethical standards
This study is a secondary analysis of publicly available, de-identified data from the CBPS. The original CBPS received ethical approval from the Innovations for Poverty Action Institutional Review Board (Protocol No. 14742) and the George Washington University Institutional Review Board (Protocol No. 071721). Verbal informed consent was obtained from all adult participants prior to data collection. The present study analysed anonymized, publicly available data; therefore, additional ethical approval was not required.
Open access
