Psychotic-like experiences as precursors to emotional symptoms in adolescents: a longitudinal network analysis

Abstract

Background

Psychotic-like experiences (PLEs), anxiety, and depressive symptoms frequently co-occur in adolescents. While traditional research has relied on cross-sectional data and overall symptom levels, this approach fails to clarify their dynamic interactions over time. Network analysis offers a novel methodological framework to address this critical limitation.

Methods

In this longitudinal study, 2,101 adolescent participants from Hunan Province, China, completed a battery of questionnaires at two assessment time points (T1: March 2024; T2: March 2025). The questionnaires measured psychotic-like experiences (PLEs) via the 15-item Community Assessment of Psychic Experiences (CAPE-P15), anxiety symptoms via the 7-item Generalized Anxiety Disorder Scale (GAD-7), and depressive symptoms via the 9-item Patient Health Questionnaire (PHQ-9), with sociodemographic and clinical characteristics also collected via this survey. Cross-lagged panel network (CLPN) analysis and contemporaneous network analysis were conducted.

Results

The contemporaneous network revealed extensive anxiety-depression comorbidity. The strongest connections were between “Worrying too much” (GAD3) and “Feeling bad about yourself” (PHQ6), and between “Trouble relaxing” (GAD4) and “Sleep problems” (PHQ3). “Perceptual abnormalities” (PA) and “Suicidal ideation” (PHQ9) showed the strongest cross‑domain connections between PLEs and depression. In the CLPN analysis, T1 “Persecutory ideation” (PI) predicted T2 “Irritability” (GAD6) and “Feeling bad about yourself” (PHQ6), while T1 “Bizarre experiences” (BE) predicted T2 “Fear of something terrible happening” (GAD7). Centrality analysis indicated that PI and BE had the highest out‑degree influence, whereas T2 anxiety symptoms “Feeling nervous” (GAD1), “Inability to stop or control worries” (GAD2), and “Worrying too much” (GAD3) had the highest in‑degree centrality, reflecting their susceptibility to influence from other symptoms.

Conclusions

This study pinpoints specific symptomatic pathways, particularly certain PLEs’ predictive role for emotional symptoms. This provides a solid foundation for precise early intervention and prevention strategies for at-risk adolescents.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13034-026-01063-y.

Introduction

Psychotic-like experiences (PLEs) refer to subtle, subclinical hallucinations and delusions that are common in the general population [1, 2]. Notably, PLEs are relatively prevalent among children and adolescents. Specifically, the prevalence of PLEs reaches 17% in childhood, while it decreases to 7.5% during adolescence, showing a distinct age-related downward developmental trend [3]. Existing studies have confirmed that if a small subset of PLEs that onset during adolescence do not remit but persist instead, they are highly likely to further progress to clinical psychotic disorders [4, 5]. The adolescent stage is a critical period when PLEs may persist and advance to clinical psychotic disorders. Furthermore, previous research has indicated that adolescence is also the peak period for the first onset of affective disorders, such as anxiety and depression [6, 7]. Therefore, adolescence constitutes a crucial window in the emergence and development of mental disorders, and it is also of great importance to explore the interactive relationship between PLEs and other emotional symptoms during this stage.

Adolescent PLEs are often accompanied by greater clinical severity, including a higher burden of emotional problems [8]. Current research consistently confirms that exposure to PLEs increases the risk of depression and anxiety in adolescents [9–11]. As common mental disorders in adolescents, depression and anxiety have relatively high incidence during this developmental stage and have become key challenges affecting adolescent mental health [12]. Cross-sectional studies have shown that adolescents with a history of PLEs have a 4- to 6-fold higher risk of developing depressive and anxiety symptoms than those without PLE experiences [13–15]. Longitudinal studies have further expanded these findings. When adolescents experience persistent PLEs, they show a significant worsening of depressive and anxiety symptoms during a 1–2-year follow-up period [16–19]. Among the common subtypes of PLEs in children and adolescents, auditory hallucinations have been linked to depressive disorders both 2 years later [16] and 8 years later [9], according to population-based sample studies.

However, most prior studies investigating the association between PLEs, anxiety and depressive symptoms have relied on latent variable models [20–22]. These models aggregate symptoms into latent constructs, for instance a PLEs factor or a depression factor. This structure makes it impossible to explore dynamic relationships between specific PLE subtypes and individual anxiety or depressive symptoms at the item level, and this limitation hinders our understanding of the precise mechanisms that underpin their association. In contrast, network analysis is an innovative methodological framework that conceptualizes mental health issues as a network of individual symptoms (nodes) and their associations (edges) [23, 24]. This approach enables the investigation of item-level dynamic relationships, thereby offering a complementary perspective to latent variable modeling and addressing the aforementioned limitation [25]. Unlike traditional approaches that assume symptom independence, network theory proposes that mental health symptoms mutually reinforce each other [23, 26]. Recent empirical studies have supported this proposition. Specifically, positive correlations have been observed between different PLE symptoms and between different anxiety and depression symptoms [27, 28]. These findings indicate that the symptom interaction hypothesis inherent in network analysis is more consistent with the dynamic characteristics of PLEs, anxiety, and depression.

Recent progress in network analysis has led to the development of the Cross-Lagged Panel Network (CLPN) model [29, 30]. By examining how variables measured at one timepoint predict their own values and the values of other variables at subsequent timepoints, CLPN models clarify the temporal order of processes across different domains of psychopathology. Similar to cross-sectional network models, CLPN models also evaluate variable centrality [31], which is defined as the overall connection strength between a specific variable and all other variables in the network. Notably, CLPN models further specify the directionality of centrality through two key metrics: output centrality (representing the strength of outgoing connections from a variable) and input centrality (representing the strength of incoming connections to a variable). From a clinical perspective, variables with the highest output centrality are suitable as intervention targets, while variables with the highest input centrality may serve as key indicators of intervention outcomes.

Consistent empirical evidence supports associations among PLEs, anxiety, and depressive symptoms, yet uncertainties remain regarding the specific details of these relationships. Prior research has primarily used latent variable models to explore these associations [20–22]. In the present study, we adopt network analysis, which offers a complementary perspective to latent variable modeling. Notably, this analytical approach enables the investigation of dynamic relationships between variables at the item level, thereby addressing the aforementioned limitation of latent variable models [25]. Against this methodological and empirical backdrop, the present study aims to address two interrelated objectives using network analysis. The first objective is to identify the longitudinal influence relationships between PLEs, anxiety, and depressive symptoms. The second is to pinpoint key nodes within the dynamic influence network of PLEs, anxiety, and depressive symptoms among adolescents. The identification of these key nodes is intended to provide effective targets for the intervention of adolescent psychological symptoms. To achieve the aforementioned research objectives, CLPN models were constructed and analyzed in the current study.

Methods

Participants and procedure

Participants were adolescents from three middle schools in Hunan Province, China, who completed two rounds of questionnaire surveys in March 2024 (T1) and March 2025 (T2), respectively. A total of 2,416 adolescents were initially recruited at T1. Of these, 2,101 (87.0%) completed the assessment at both T1 and T2 and were included in the final longitudinal analysis. Psychiatric diagnosis history was screened via a validated self-report item adapted from the Mini International Neuropsychiatric Interview for Children and Adolescents (MINI-KID) [32]: “Have you ever been formally diagnosed with a mental disorder (e.g., depression, anxiety disorder, schizophrenia) by a psychiatrist or clinical psychologist?” Participants who responded “Yes” were excluded to focus on subclinical symptom networks in a community-dwelling adolescent sample, avoiding confounding with chronic psychiatric conditions. Sociodemographic and clinical data were also collected using a dedicated questionnaire, which captured basic information including age, gender, residence, only-child status and economic status. Economic status was assessed with the MacArthur Scale [33], a 10-rung ladder that measures perceived socioeconomic status (SES). The top rung of this ladder represents the highest SES, while the bottom represents the lowest. This scale has been validated in Chinese adults [34].

The ethical review of this study was approved by the Ethics Committee of the First Affiliated Hospital of Hunan University of Chinese Medicine. All participants and their guardians completed the informed consent forms as required, and the questionnaires were administered via paper-and-pencil format under the guidance of school teachers.

Among the 2101 participants, the mean age was 15.83 years (SD = 0.89). There were 1111 females (52.88%) and 990 males (47.12%). Regarding only-child status: 1830 were non-only children (87.10%), and 271 were only children (12.90%). For residential background: 1134 were from urban areas (53.97%), and 967 were from rural areas (46.03%). The participants’ mean score on subjective economic status was 5.47, with a standard deviation of 1.56.

Measures

Anxiety symptoms were assessed using the Generalized Anxiety Scale (GAD-7), a commonly used tool for evaluating generalized anxiety symptoms [35]. The GAD-7 comprises seven items, covering dimensions such as “Nervous”, “Uncontrollable worrying”, “Worry too much”, “Trouble relaxing”, “Restlessness”, “Irritability”, and “Feeling afraid”. Items are rated on a 4-point scale (from 0 = not at all to 3 = nearly every day), with total scores reflecting the severity of anxiety symptoms—higher scores indicate more severe anxiety. The Chinese version of the GAD-7 has demonstrated good reliability and validity in the Chinese adolescents population [36]. The Cronbach’s alpha coefficients of this scale were adequate at T1 (α = 0.912) and T2 (α = 0.921).

Depressive symptoms were assessed using the Patient Health Questionnaire-9 (PHQ-9), a widely utilized tool for screening and assessing the severity of depression [37]. The PHQ-9 consists of nine items rated on a scale of 0 to 3, with total scores ranging from 0 to 27. Higher scores are indicative of a greater presence of depressive symptoms. The Chinese version of the PHQ-9 has shown reliable psychometric properties [38], with a Cronbach’s α of 0.891 in the current study. The Cronbach’s alpha coefficients of this scale were adequate at T1 (α = 0.889) and T2 (α = 0.911).

PLEs were assessed using the Community Assessment of Psychic Experiences (CAPE-P15) [39]. The CAPE-P15 consists of two subscales that assess the frequency and distress of PLEs across 15 items. These 15 items are grouped into three factors: persecutory ideation (PI), bizarre experiences (BE), and perceptual abnormalities (PA). Each item was rated on a four-point Likert scale, from 1 (never), 2 (sometimes), 3 (often), to 4(nearly always). The Chinese version of the CAPE-P15 has demonstrated satisfactory psychometric properties among adolescents [40]. The Cronbach’s alpha coefficients of this scale were adequate at T1 (α = 0.817) and T2 (α = 0.881).

Data analysis

First, SPSS 18.0 was used for descriptive statistics and paired-sample t-tests. Second, R 4.5.1 was used to construct a CLPN model to analyze the network relationship between psychotic-like experiences, anxiety and depressive symptoms at T1 and T2 [41]. In the CLPN model, gender, age and economic status were included as covariates.

Contemporaneous network analysis

Contemporaneous networks for both waves were estimated and visualized using the Bootnet and qgraph packages in R [42]. Given the non-normal distribution of the data, Spearman’s rank correlation matrices were used as input. The EBICglasso method was applied to estimate Gaussian Graphical Models (GGM), with the extended Bayesian information criterion guiding model selection. The resulting networks consist of nodes (symptoms) and edges (regularized partial correlations). Expected Influence (EI) centrality was employed to measure the total expected impact a node has on the network upon activation, calculated by summing the weights (both positive and negative) of all its connected edges [43]. For Contemporaneous Networks, covariates (age, gender, economic status) were included as predictors in a linear regression model prior to EBICglasso analysis. Residuals from these regressions were used to construct the contemporaneous networks, ensuring that covariate effects were partialled out [25].

Cross-lagged panel network analysis

The CLPN was estimated using R packages including glmnet, qgraph, and bootnet. The analysis incorporated 19 symptoms (3 PLEs, 7 GAD, and 9 PHQ items) that were assessed at both T1 and T2, along with gender, age, and economic status as covariates. The model was fitted using regularized regression with LASSO penalty, implemented via penalized maximum likelihood estimation [44]. This approach reduces overfitting and shrinks trivial regression coefficients to zero, improving interpretability and generalizability [45]. For each T2 symptom, a 10-fold cross-validated LASSO regression was fitted with all T1 symptoms and covariates as predictors [46]. Lambda (regularization parameter) for CLPN was selected via 10-fold cross-validation using the cv.glmnet function in the R package glmnet. The lambda value that minimized the mean squared error (lambda.min) was chosen to balance model fit and parsimony, with a penalty factor of 1 assigned to all edges to ensure consistent regularization across the network [47]. This produced a 19 × 19 adjacency matrix in which diagonal elements represent autoregressive effects (within-symptom prediction) and off-diagonal elements represent cross-lagged effects (between-symptom prediction), adjusted for all other T1 symptoms and covariates.

The network was visualized using the qgraph package [41]. Nodes were grouped by symptom domain and arranged using a spring layout. To aid interpretability, autoregressive edges were omitted from the visualization, and only edges with weights ≥ 0.05 were retained.

Two centrality indices (out-degree and in-degree) were derived. Out-degree (equivalent to out-expected influence) reflects the total strength with which a T1 symptom predicts other T2 symptoms. In-degree (equivalent to in-expected influence) quantifies the total influence of other T1 symptoms on a given T2 symptom. Both indices were Z-score standardized and visualized using ggplot2.

Network stability and replication

The stability and accuracy of the network were evaluated using the bootnet package [41]. The case-drop bootstrapping method was used to calculate correlation stability (CS) coefficients for centrality indices (a CS > 0.25 was considered acceptable, and > 0.5 was deemed excellent) [25]. Bootstrap-based difference tests were also conducted for edge weights and centrality indices.

Subgroup analysis

Subgroup sensitivity analyses were performed to assess network heterogeneity and robustness, stratifying the total sample (N = 2,101) by gender (male: n = 990; female: n = 1,111) and residential status (urban: n = 1,135; rural: n = 966). Consistent with full-sample procedures, LASSO regularization regression was used to construct T1→T2 cross-lagged networks for each subgroup, with standardized out-degree (influence) and in-degree (susceptibility) centrality calculated (Z-score normalized).

Statistical inferences (e.g., bootstrap significance testing, centrality calculations) were based on the complete estimated network containing all edges. Pearson correlation analysis quantified the stability of centrality indices across five groups (including the full sample) to verify the consistency of core findings. Nodes ranked in the top 5 for centrality within each subgroup were defined as “key nodes,” and their frequency across subgroups was counted to identify cross-subgroup robust core nodes. Additionally, the top 3 cross-community cross-lagged edges (by strength) were extracted from each subgroup network to compare the cross-population stability of core associations.

Results

Descriptive analyses

Descriptive statistical analyses were performed on data for each variable at time points T1 and T2, results are presented in the Table 1. The variables encompass PI, BE, and PA within psychotic-like experiences, along with items related to generalized anxiety (GAD1-GAD7) and depression (PHQ1-PHQ9).

Table 1.

Item labels and descriptive statistics of anxiety symptoms, depressive symptoms and psychotic-like experiences at two-time points (N = 2101)

Item	Label	T1 (M ± SD)	T2 (M ± SD)	t	p	Cohen’s d
Generalized anxiety symptoms
Feeling nervous, anxious, or on edge	GAD1	0.96 ± 0.81	0.82 ± 0.79	8.43	< 0.001	0.18
Inability to stop or control worries	GAD2	0.72 ± 0.81	0.64 ± 0.83	4.38	< 0.001	0.10
Worrying too much about different things	GAD3	0.89 ± 0.87	0.75 ± 0.86	7.91	< 0.001	0.17
Trouble relaxing	GAD4	0.66 ± 0.81	0.55 ± 0.79	6.29	< 0.001	0.14
Restlessness	GAD5	0.55 ± 0.76	0.48 ± 0.77	4.12	< 0.001	0.09
Irritability	GAD6	0.85 ± 0.88	0.73 ± 0.87	6.86	< 0.001	0.15
Fear of something terrible happening	GAD7	0.79 ± 0.88	0.61 ± 0.84	8.75	< 0.001	0.19
Depressive symptoms
Feeling little interest or pleasure	PHQ1	0.99 ± 0.83	0.78 ± 0.80	10.62	< 0.001	0.23
Feeling down or hopeless	PHQ2	0.77 ± 0.75	0.65 ± 0.79	6.54	< 0.001	0.14
Problems sleeping	PHQ3	0.66 ± 0.82	0.62 ± 0.85	2.17	0.030	0.05
Being tired or having little energy	PHQ4	0.87 ± 0.80	0.75 ± 0.88	6.78	< 0.001	0.15
Poor appetite or overeating	PHQ5	0.65 ± 0.83	0.62 ± 0.85	1.74	0.082	0.04
Feeling bad about yourself	PHQ6	0.86 ± 0.89	0.71 ± 0.88	8.03	< 0.001	0.17
Trouble concentrating	PHQ7	0.59 ± 0.81	0.57 ± 0.83	1.15	0.250	0.02
Speaking/moving slowly or restless	PHQ8	0.43 ± 0.74	0.38 ± 0.70	2.89	0.004	0.07
Suicidal ideation	PHQ9	0.43 ± 0.70	0.37 ± 0.69	3.26	0.001	0.08
Psychotic-like experiences
Persecutory ideation	PI	1.44 ± 0.44	1.41 ± 0.51	2.45	0.014	0.06
Bizarre experiences	BE	1.35 ± 0.44	1.35 ± 0.51	0.00	1.000	0.00
Perceptual abnormalities	PA	1.23 ± 0.42	1.23 ± 0.47	0.00	1.000	0.00

For reference, a post-hoc Bonferroni correction was implemented with an adjusted alpha level of α = 0.05/19 = 0.0026 (reflecting the 19 total variables included in the analysis). Under this corrected threshold (p < 0.0026), statistically significant results were observed for all generalized anxiety symptoms (GAD1 to GAD7), as well as depressive symptoms PHQ1, PHQ2, PHQ4, PHQ6, and PHQ9; non-significant results were noted for depressive symptoms PHQ3, PHQ5, PHQ7, PHQ8, and all psychotic-like experiences (PI, BE, PA)

Paired-samples t-tests were performed to examine the temporal changes in all variables from T1 to T2. The results indicated that all generalized anxiety symptoms showed significant decreases from T1 to T2 (all p < 0.05). Specifically, the reductions were significant for GAD1 (t = 8.43, Cohen’s d = 0.18), GAD2 (t = 4.38, Cohen’s d = 0.10), GAD3 (t = 7.91, Cohen’s d = 0.17), GAD4 (t = 6.29, Cohen’s d = 0.14), GAD5 (t = 4.12, Cohen’s d = 0.09), GAD6 (t = 6.86, Cohen’s d = 0.15), and GAD7 (t = 8.75, Cohen’s d = 0.19), with small effect sizes observed for all these items.

For depressive symptoms, six out of nine items exhibited significant reductions over time (all p < 0.05). The significant decreases were found in PHQ1 (t = 10.62, Cohen’s d = 0.23), PHQ2 (t = 6.54, Cohen’s d = 0.14), PHQ4 (t = 6.78, Cohen’s d = 0.15), PHQ6 (t = 8.03, Cohen’s d = 0.17), PHQ8 (t = 2.89, Cohen’s d = 0.07), and PHQ9 (t = 3.26, Cohen’s d = 0.08), with effect sizes ranging from trivial to small. In contrast, no significant temporal changes were detected for PHQ3, PHQ5, and PHQ7 (all p > 0.05).

Regarding psychotic-like experiences, only PI showed a significant reduction from T1 to T2 (t = 2.45, p < 0.05, Cohen’s d = 0.06), with a small effect size indicating a mild magnitude of change. No significant differences were observed between T1 and T2 for BE and PA, suggesting these two PLE dimensions remained stable over time.

For reference, a post-hoc Bonferroni correction was implemented with an adjusted alpha level of α = 0.05/19 = 0.0026 (accounting for 19 total variables). Under this stricter threshold, significant reductions were retained for all generalized anxiety symptoms (GAD1-GAD7) and the five aforementioned depressive symptoms (PHQ1, PHQ2, PHQ4, PHQ6, PHQ9), while PHQ8 and PI no longer reached statistical significance.

Network analyses

Contemporaneous networks

The contemporaneous networks for the two waves, controlling for gender, age, and economic status, are presented in Fig. 1, with their corresponding weighted adjacency matrices detailed in Supplementary Tables S1 and S2. Analysis of these networks indicated that PA and PHQ9 shared the strongest cross-domain connection (mean weight = 0.105), forming the primary link between the psychotic-like experiences (PLEs) and depression domains. Connections within the anxiety and depression domains were notably robust; for instance, GAD3 and PHQ6 were connected with a mean weight of 0.093, and GAD4 and PHQ3 with a mean weight of 0.092. Among the top ten strongest cross-domain connections, nine involved pairs of anxiety and depression measures. In contrast, connections between the PLEs and anxiety domains were substantially weaker, with the strongest link observed between BE and GAD2 (mean weight = 0.042) (Fig. 1a, b).

Fig. 1.

Structure of cross-sectional networks (T1: Panel a, T2: Panel b) and expected influence centrality indices of all items (T1: Panel c, T2: Panel d). Arrow thickness illustrates the strength of observed associations

Regarding symptom centrality, GAD2 and GAD3 exhibited the highest average Expected Influence (EI) (Fig. 1c, d), indicating that anxiety symptoms may exert the strongest average predictive effects on other variables in the network.

The stability of centrality indices was assessed by examining changes in the correlation between centrality and original centrality as sample size decreased (Supplementary Figure S1-S2). The CS-coefficients for EI were 0.439 at T1 and 0.443 at T2. According to the criterion proposed by Epskamp and Fried (2018) (requiring values > 0.25) [25], the node centrality in the contemporaneous networks can be considered stable.

Temporal networks

The CLPN is visualized as a directed network in Fig. 2. Edges with weights ≥ 0.05 were retained in the network visualization to enhance clarity. This threshold was chosen because all estimated negative edges had absolute weights below 0.05, and a weight of 0.05 represents a conservative cutoff above negligible effect sizes [48]. For all statistical interpretation (e.g., bootstrap significance testing, centrality calculations), the complete estimated network including all edges was used. A network plot containing all edges is provided in Supplementary Material (Figure S3). This network incorporates cross-lagged edges, representing temporal associations between node pairs after controlling for gender, age, and economic status at T1. To reduce visual complexity, autoregressive edges are excluded from Fig. 2 but are fully presented in Fig. 3. The network contained 154 non-zero cross-lagged edges (45.03% of possible edges), with the adjacency matrix provided in Supplementary Table S3. Within this longitudinal network, the three strongest cross-lagged edges connecting distinct symptom communities were: from PI to GAD6 (weight = 0.16), from PI to PHQ6 (weight = 0.15), and from BE to GAD7 (weight = 0.15). The most substantial autoregressive effects were observed for PI, BE, and PHQ3.

Fig. 2.

The network analyzed in the present study. Arrows show significant longitudinal associations. Arrow thickness illustrates the strength of observed associations. Auto-regressive effects and the effects of covariates are not shown to ease visual interpretation. (edges weight ≥ 0.05)

Fig. 3.

Autoregressive effects in the T1−> T2 network

Centrality estimates for the temporal network are shown in Fig. 4. The nodes with the highest out-degree centrality were PI (out-degree = 2.97), BE (out-degree = 2.18), and PHQ4 (out-degree = 0.45), suggesting they exert strong predictive influences on other symptoms. Conversely, the nodes with the highest in-degree centrality were GAD3 (in-degree = 1.37), GAD2 (in-degree = 1.13), and GAD1 (in-degree = 1.09), indicating they are more susceptible to being influenced by other symptoms. Notably, PI demonstrated both the highest out-degree and the most negative in-degree centrality (-1.43), positioning it as a potential core driver in the network, initiating cascading effects on other variables while being minimally influenced by them. In contrast, the anxiety symptom cluster (GAD1–GAD3) appears to function predominantly as receivers of influence, potentially representing endpoints in the symptom interaction pathways.

Fig. 4.

Network centrality estimates in the T1 → T2 CLPN

The stability of the CLPN centrality was evaluated, with CS-coefficients for in-EI and out-EI being 0.672 and 0.283, respectively (Supplementary Figure S4). Based on the same stability criterion [25], the node centrality in the temporal network was stable. Furthermore, bootstrapped confidence intervals for edge weights are presented in Supplementary Figure S5; the narrow intervals suggest good accuracy of the estimated edge weights. Additional edge weight difference tests and centrality difference tests are provided in Supplementary Figures S6, S7, and S8.

Subgroup analysis results

Subgroup analysis of the cross-lagged network revealed moderate stability in the centrality patterns of the network structure across demographic subgroups (Supplementary Table S4-5). In-degree centrality showed high consistency among subgroups (mean r = 0.695, range 0.447–0.809), while out-degree centrality had relatively lower consistency (mean r = 0.577, range 0.002–0.908).

Two symptoms exhibited stably consistent centrality and consistently ranked among the top five nodes across all five subgroups (full sample, males, females, urban, rural) (Supplementary Figure S9): PI emerged as a top-five high-impact node (high out-degree) across all subgroups, acting as a core driver of network dynamics; GAD3 was identified as a top-five highly susceptible node (high in-degree) across all subgroups, thereby demonstrating heightened susceptibility to the influence of other symptoms in the network.

The three strongest cross-lagged edges differed distinctly across subgroups (Supplementary Figure S10-11): in the male network, all three strongest edges were from BE to GAD symptoms; in the female network, the three strongest edges were dominated by PI-driven effects; in urban samples, the three strongest edges were all PI-driven; in rural samples, the three strongest edges were mainly BE-driven, with this subgroup showing the only strong cross-symptom-group cross-lagged edge (PHQ4→GAD1) in the study. The core driving pathways (PI/BE→GAD/PHQ symptoms) were stable across all subgroups, further verifying the robustness of the study results.

Discussion

The current study is the first to explore the longitudinal influential relationships between PLEs, anxiety symptoms, and depressive symptoms in adolescents from a network perspective. Contemporaneous networks revealed extensive comorbidity between anxiety symptoms and depressive symptoms. The strongest connections between anxiety and depressive symptoms were observed between GAD3 (Worrying too much about different things) and PHQ6 (Feeling bad about yourself), and between GAD4 (Trouble relaxing) and PHQ3 (Problems sleeping). In addition, Perceptual abnormalities (PA) and PHQ9 (Suicidal ideation) formed the strongest cross-domain link among contemporaneous networks. CLPN analysis suggested temporal predictive relationships, indicating that specific PLE nodes at T1 were associated with subsequent anxiety and depressive symptoms at T2: Persecutory ideation (PI) and Bizarre experiences (BE) at T1 were associated with subsequent emotional symptoms. Specifically, PI at T1 was linked to GAD6 (Irritability) and PHQ6 (Feeling bad about yourself) at T2, while BE at T1 was linked to GAD7 (Fear of something terrible happening) at T2. Both PI and BE at T1 emerged as nodes with high predictive influence (out-degree) in the entire network. Autoregressive analyses further confirmed that PI and BE exhibited significant temporal stability, remaining persistent and resistant to change in affected individuals. It is noteworthy that although most symptoms showed statistically significant decreases from T1 to T2, the effect sizes were small (Cohen’s d < 0.20). The small reduction in symptom severity is consistent with natural fluctuations in adolescent emotional symptoms over a one-year period, as reported in previous longitudinal studies [49]. This may reflect transient stressors (e.g., academic pressure) that abated between assessments.

Contemporaneous network results refine our understanding of the mechanisms underlying comorbidity between anxiety and depressive symptoms. Epidemiological studies have long documented high comorbidity between anxiety and depression in adolescents [50, 51], but prior research often focused on domain-level associations rather than symptom-specific interactions. Our identification of key symptom pairs fills this gap: the link between GAD3 (Worrying too much about different things) and PHQ6 (Feeling bad about yourself) could be interpreted as a “cognitive worry-emotional devaluation” pathway that may underpin core comorbid processes [52, 53], while the connection between GAD4 (Trouble relaxing) and PHQ3 (Problems sleeping) points to a potential “somatic tension-sleep disturbance” physiological symptom chain. This granularity aligns with network theory’s emphasis on symptom interactions as drivers of psychopathology [26], providing empirical support for heterogeneous comorbidity pathways. The robust contemporaneous link between perceptual abnormalities (PA) and suicidal ideation is a critical finding with notable theoretical and clinical value. Disturbed reality testing associated with PA can trigger overwhelming fear and hopelessness, driving a desire to escape distressing subjective experiences, a pathway linked to adolescent suicide risk [54]. The association may also stem from shared neurocognitive deficits (e.g., impaired source monitoring), which are transdiagnostic correlates of both PLEs and suicidal thoughts [55]. Additionally, PA can serve as a marker of severe psychological distress that directly amplifies suicidal ideation, consistent with evidence that PA confers unique suicidality risk in young populations [54]. Clinically, this finding underscores that assessing PLEs, particularly perceptual disturbances, should be part of routine suicide risk assessment for adolescents, even in the absence of a formal psychotic disorder diagnosis [56]. This finding addresses a limitation of prior latent variable models, which treated PLEs as a unified construct and failed to identify subtypes specifically tied to suicide risk [57, 58].

While the contemporaneous network illuminated the structure of symptom co-occurrence, the longitudinal CLPN model was crucial for uncovering temporal dynamics. CLPN findings suggest that PI (Persecutory ideation) and BE (Bizarre experiences) may function as potential precursors or influential nodes in symptom development. The observed associations of PI at T1 with GAD6 (Irritability) and PHQ6 (Feeling bad about yourself) at T2 are consistent with two distinct pathways: interpersonal distrust from persecutory beliefs may translate to irritability, while victimization experiences may erode self-worth. This aligns with Freeman et al.’s cognitive model, which posits that persecutory delusions activate negative emotional cycles by distorting threat perception and self-cognition [59]. Similarly, the association between BE at T1 and GAD7 (Fear of something terrible happening) at T2 lends support to the theoretical notion that bizarre experiences may heighten threat vigilance, thereby reinforcing anxiety. Our findings also align with the stress-vulnerability model [60], which posits that preexisting vulnerabilities (e.g., PLEs) interact with environmental stressors to predict emotional symptoms. Additionally, the transdiagnostic vulnerability theory [61] suggests that shared mechanisms (e.g., negative affectivity) underlie both PLEs and emotional symptoms, which may explain the cross-domain edges observed in our network. These theories collectively support the notion that PLEs are not isolated phenomena but part of a broader transdiagnostic symptom cluster in adolescence.

Notably, both PI and BE at T1 exhibited the highest out-degree centrality and significant temporal stability, which was consistent with prior research indicating that a subset of adolescents experience persistent PLE symptoms [5]. Their stability, evident in strong autoregressive effects, suggests that once present, these specific PLE subtypes may persist and exert a sustained influence on the long-term progression of emotional symptoms. This highlights PI and BE as potential markers for identifying high-risk adolescents. Prior research has linked PLE persistence to poor outcomes [62], but our study uniquely identifies PI and BE as the most predictive and central subtype within the PLE spectrum, suggesting they could be a precise target for early intervention.

These findings have direct clinical applications. First, the specific associations and pathways identified highlight potential targets for intervention for both PI and BE: given their strong predictive power and stability, cognitive-behavioral strategies should be tailored to not only address the core irrationality of persecutory beliefs and bizarre experiences but also to preemptively mitigate their downstream effects on irritability (GAD6) and negative self-concept (PHQ6) [63]. Second, monitoring high in-degree anxiety symptoms (GAD1, GAD2, and GAD3), which were previously shown to be susceptible to external influences, through techniques such as relaxation training or worry management can prevent them from becoming “endpoints” of symptom cascades. Third, the association between PA (Perceptual abnormalities) and PHQ9 (Suicidal ideation) mandates concurrent screening for perceptual abnormalities and suicidal ideation to mitigate overlooked risk.

However, this study has several limitations. First, the data were collected via self-report questionnaires which may be subject to response bias. Future research could integrate multiple assessment methods such as clinical interviews and objective physiological measures to enhance the reliability of findings. Second, our sample was recruited from urban and rural areas of Hunan province, China, which limits generalizability to other cultural contexts. Chinese adolescents face unique stressors (e.g., intense academic pressure, collectivistic family values emphasizing filial piety and achievement) that may shape symptom networks differently than in Western populations [64]. For example, the strong PA-suicidal ideation link may be amplified by stigma surrounding mental health in China, leading to underreporting of emotional symptoms and overreliance on somatic/perceptual symptoms. Future studies should replicate these findings in diverse cultural and geographic samples to test cross-cultural invariance of the symptom network. Third, the two-wave design, while sufficient for initial CLPN modeling, prevents the examination of feedback loops, equilibrium states, or more nuanced developmental courses that require three or more time points. Future multi-wave studies are needed to validate and extend the temporal dynamics observed here. Finally, while we controlled for key demographic factors, other important confounders such as childhood trauma, substance use, or family psychiatric history were not measured. These factors represent shared vulnerabilities for both PLEs and emotional symptoms, and their absence means the identified predictive pathways might be partially explained by these unmeasured third variables.

In conclusion, this network analysis clarifies both comorbid patterns and longitudinal dynamics between adolescent PLEs, anxiety, and depression. Our results highlight PI and BE as potentially influential nodes, PA as having a strong positive correlation with suicidal ideation, and specific anxiety symptoms as vulnerable targets. These insights could inform precision prevention and intervention strategies. Future research should explore mediating mechanisms (e.g., cognitive attribution, peer relationships) and test targeted interventions specifically designed around these symptom-to-symptom pathways to validate their clinical utility.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Research question conceptualization and study design (all authors). Data acquisition (C.X., S.W.), statistical analysis, and responsibility for integrity and accuracy of data analysis (X.X., C.X.). Interpretation of data (all authors). Drafting of the manuscript (C.X.). Critical revision of the manuscript for important intellectual content (all authors).

Funding

This study was supported by the National Natural Science Foundation of China (No: 62402170), the National Social Science Fund of China (No: 25CSH033) and the Natural Science Foundation of Hunan Province, China (No: 2023JJ30446).

Data availability

The datasets generated and analyzed during the current study are not publicly available but are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The ethical review of this study was approved by the Ethics Committee of the First Affiliated Hospital of Hunan University of Chinese Medicine (Approval Number: HN-LL-GZR-2024-045). Written informed consent was obtained from all participants and their legal guardians in accordance with the relevant requirements.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.