Participants in this study consisted of students attending universities in Beijing, China. Sixty-seven universities in Beijing were divided into 13 types according to their disciplines. Taking the running levels of the universities into account as well, in other words, trying to cover not only key universities but also ordinary universities, we included 31 universities in this study: comprehensive (5), science (5), engineering (5), agriculture (2), normal (2), finance and economics (3), forestry (1), politics and law (1), medicine (1), language (3), nationality (1), art (1) and sports (1). Random stratified sampling strata were made on the universities, majors (liberal arts or sciences) and grades. In this way, we ensured the diversity and representativeness of the participants. Participants first read the instructions for the present study. Those who agreed to participate provided informed consent. Then, they were given an online questionnaire and completed the assessment. The distribution and collection of questionnaires were conducted by teachers in the universities. We continuously recruited students until the number in each stratum reached the number we formulated in advance. Overall, 2048 participants from 29 universities completed the survey. We first screened to obtain valid data, and 221 participants were excluded due to careless answers (e.g., failure to pass the attention check items or answering the same answer to each item). Then, 1827 (89.2%) valid data points were further screened according to the inclusion criteria in this study: (a) had direct or indirect trauma histories, which were determined by the score of the Life Events Checklist for DSM-5; and (b) were aged 18–25 years when the data were collected. Finally, 1368 met the inclusion criteria.

Approval for this study was granted by the ethics committee on human research protection of East China Normal University. All participants gave informed consent. They had a mean age of 20.36 ± 1.45 years, and there were more women (65.4%) than men. Other demographics were the prevalence of PTSD and CPTSD, the participants’ majors and the education levels of their parents (see Table 1).

To classify childhood traumatic events that may contribute to CPTSD symptoms, we used a revised version of the Life Events Checklist for DSM-5 (LEC-5), which has 17 items such as natural disasters, physical or sexual assault, serious injury, and violent death (homicide or suicide) [19]. Considering that the target population were young Chinese adults, four events that would hardly happen to them were deleted: exposure to war, captivity, serious accident at work and severe human suffering. For each event, participants were asked to recall and indicate the type of exposure (e.g., whether they directly experienced or witnessed the event and whether it was related to occupational activities) before they were 18 years old. Each item was scored on a six-point Likert scale, ranging from 0 (does not apply) to 5 (happened to me). Only those young adults who reported having witnessed or experienced at least one event were considered to have childhood traumatic experience and were identified as having a history of childhood trauma.

The International Trauma Questionnaire (ITQ) was adopted to measure ICD-11 PTSD and CPTSD [20]. The Chinese version of ITQ was utilized to assess CPTSD symptoms in this study [21]. The ITQ consists of 18 items, 12 of which correspond to 12 symptoms of CPTSD and 6 that measure functional impairment. PTSD symptoms (re-experiencing, avoidance, and sense of current threat) were assessed by six items, with each symptom measured by two items. There were three items assessing functional impairment associated with PTSD symptoms. Similarly, DSO symptoms (negative self-concept, affective dysregulation, and disturbances in relationship) were assessed by six symptom-related items and three function-related items. All items were rated on a 5-point Likert scale ranging from 0 (not at all) to 4 (extremely). The diagnosis of PTSD or DSO requires all three PTSD or DSO symptoms to be present (scored 2 or greater), while functional impairment was also observed (at least one of the three function-related items scored 2 or greater). CPTSD was diagnosed when both PTSD and DSO met the criteria. In other words, participants who only meet the diagnostic criteria for PTSD are diagnosed with PTSD, while participants who meet the diagnostic criteria for both PTSD and DSO are diagnosed with CPTSD. In this study, Cronbach’s alpha of the scale was 0.88.

We relied on IBM SPSS Statistics 23.0 to evaluate the prevalence of the reported childhood traumatic events and descriptive statistics of CPTSD symptoms. There was no missing data.

We first tested the factor structure of ITQ in our sample. The correlated six-factor first-order model (model 1) and the two-factor second-order model (model 2; see Fig. 1) were tested using confirmatory factor analysis (CFA). The CFA analyses were performed in Mplus 8.3 [22]. We evaluated the model fit using the following fit indices: chi-squared test, the comparative fit index (CFI) [23], the Tucker‒Lewis index (TLI) [24], and the root mean square error of approximation (RMSEA) [25]. CFI and TLI values ≥ 0.95 reflect excellent model fit; RMSEA values ≤ 0.08 and ≤ 0.06 reflect acceptable and excellent model fit, respectively. The change in the RMSEA value (ΔRMSEA) was used to compare the two CFA models, and a ΔRMSEA value of ≥ 0.015 suggests a meaningful difference in model fit [26].

Correlated six-factor first-order model (a) and two-factor second-order model (b). Notes: Re = re-experiencing; Av = avoidance; Th = sense of threat; AD = affective dysregulation; NSC = negative self-concept; DR = disturbed Relationships; Re1: nightmares; Re2: flashbacks; Av1: internal avoidance; Av2: external avoidance; Th1: hypervigilance; Th2: exaggerated startle response; AD1: long-term upset; AD2: emotional numbing; NSC1: feelings of failure; NSC2: feelings of worthlessness; DR1: feeling distant or cut off from others; DR2: difficulties feeling close to others

A statistical procedure described by Epskamp and Fried was conducted to identify the overall network of ICD-11 CPTSD symptoms [27]. All analyses were performed using R 4.1.2 and visualized with the R package qgraph [28]. Because previous studies found that CPTSD is more likely associated with repeated trauma and poly-traumatized exposure [13], we first performed network analysis in all samples with trauma history and then in people who experienced 2 or more trauma types. Finally, we compared the results of network analysis in two samples.

The partial correlation network was used to prescribe the association parameters between all nodes according to Gaussian graphical models (GGMs). Sixty-six pairwise associated parameters between a total of 12 symptom nodes were estimated using the least absolute shrinkage and selection operator (LASSO) [29].

Centrality estimation was made for every symptom in the network, consisting of two categories of indices: strength centrality and bridge strength. Strength centrality, the most common and stable centrality metric [30], refers to the weighted sum of all edges connected to a particular node [31]. It was analyzed to predict the most connected node in a network. Bridge strength indicates a node’s total connectivity with other disorders or other clusters in the same disorder [12]. It was obtained for the two distinct subgroups of CPTSD (PTSD and DSO).

Robustness analyses were performed by the R package bootnet [32]. To account for the edge weight accuracy, we used the R package bootnet to bootstrap the 95% confidence intervals (CIs) around the edge weights (bootstrapped samples = 1000). Fewer overlaps among those CIs indicate higher accuracy. Centrality stability was estimated by case-dropping bootstraps, which extracted subsets from the original data, calculated node centrality based on the subsets, and correlated the ranking results of subset centrality with that of the total sample. The correlation-stability coefficients (CS coefficients) were used as an outcome measure. When it is above 0.50, it indicates that the stability is strong [33]. The edge weight difference test and centrality difference test were also estimated.

A Bayesian network was explored, with accessible causal interpretations of relationships between nodes [11]. We used the hill-climbing algorithm [34] provided in the R package bnlearn [35] to evaluate the directed edges (i.e., arrows) among symptoms, with all variables placed in a putative causal cascade, where upstream variables constitute the cause of downstream variables. The modeling process randomly added, subtracted and reversed the direction of edges while gradually optimizing the Bayesian information criterion (BIC) at the same time. For the stability of the Bayesian network, multiple bootstrapping samples were drawn, and their averaged results were used as the final network [36]. The Bayesian network was visualized in the form of a directed acyclic graph (DAG), and the source nodes or the upstream nodes revealed the most noteworthy symptoms in the Bayesian network.