Network structure and temporal stability of depressive symptoms after a natural disaster among children and adolescents

ABSTRACT

Background: Previous studies have found that the postdisaster developmental course of depression is more stable than that of other mental disorders among children and adolescents. However, the network structure and temporal stability of depressive symptoms after natural disasters among children and adolescents remain unknown.

Objective: This study aims to understand the depressive symptom network and evaluate its temporal stability among children and adolescents after natural disasters.

Methods: Three-wave measurements were conducted among 1,466 children and adolescents at 3, 15, and 27 months following the Zhouqu debris flow. Depressive symptoms were evaluated by the Child Depression Inventory (CDI), which was dichotomised to signify the presence or absence of depressive symptoms. Depression networks were estimated with the Ising model, and expected influence was used to assess node centrality. A network comparison test was used to test the differences in the depression networks among the three temporal points.

Results: Overall, the depressive symptom network was temporally stable regarding symptom centrality and global connectivity over the two-year study period. Self-hate, loneliness, and sleep disturbance were central symptoms and had low variability in the depressive networks at the three temporal points. Crying and self-deprecation had large temporal variability in centrality.

Conclusion: The present study provides the first evidence for the temporal stability of the youth depressive symptom network postdisaster. The similar central symptoms and connectivity of depression symptoms at different temporal points after natural disasters may partially explain the stable prevalence and developmental trajectory of depression. Self-hate, loneliness, and sleep disturbance could be central characteristics, and sleep disturbance and reduced appetite, sadness and crying, and misbehaviour and disobedience could be key associations in the endurance of depression among children and adolescents after experiencing a natural disaster.

Highlights

• This study provides the first evidence for the temporal stability of the youth depressive symptom network.

• The depressive symptom network had temporal stability.

• Self-hate, loneliness, and sleep disturbance were the central symptoms among youths after a natural disaster.

1. Introduction

One of the most common mental health outcomes of a natural disaster is depression, symptoms of which may continue for many years (Beaudoin, 2007; Liang et al., 2019). Depressive symptoms have a range of short- and long-term negative consequences on children and adolescents, such as decreased quality of life (Bertha & Balázs, 2013) and increased suicidal thoughts (Gijzen et al., 2021). Moreover, high levels of childhood emotional problems are related to mental disorder diagnoses in adulthood (Hofstra et al., 2002). Therefore, it is worth paying attention to the diagnosis and development of postdisaster depression among children and adolescents.

Because of the destructive power of natural disasters, which threaten people’s lives (McGuire et al., 2018) and expose them to massive negative consequences (e.g. others’ injuries/trapping, home damage, and the injury or death of a loved one) (Ying et al., 2014), prevalent depressive symptoms are often found in postdisaster samples (Li et al., 2020; Liang et al., 2019; Zhou et al., 2016). Research has shown that natural disasters often cause common depressive symptoms among children and adolescents (Guo et al., 2017; Li et al., 2020), and 30.5% of child survivors had depression 4 years after the Wenchuan earthquake (Liang et al., 2021a). A large proportion of survivors who were initially symptomatic continued to experience depressive symptoms several years after an earthquake (Guo et al., 2017). Our previous longitudinal study among children exposed to the Wenchuan earthquake found that the postdisaster developmental course of depression was more stable than that of other mental disorders, and postdisaster stressful life events might have led to the maintenance of depressive symptoms among children from the chronic depression trajectory (Liang et al., 2021a). However, no research has explored whether the structure and key symptoms of depression networks remain stable after a disaster.

Recent research suggests that the sum score of depression obfuscates the depressive symptom effects, and the analysis of specific depressive symptoms is essential (Fried & Nesse, 2014, 2015). The field that emerged from the network perspective on psychology is different from the traditional latent variable model, which suggests that all disorder symptoms are caused by an underlying common cause (McNally, 2016). The traditional view implies that each symptom is independent (Borsboom & Cramer, 2013). However, in the actual developmental process of mental disorders, symptom interaction is common (Borsboom, 2017). The network approach defines mental disorders as a series of interacting symptoms (Borsboom & Cramer, 2013); nodes represent a specific symptom, and edges represent relationships between symptoms in a network. The triggering of one symptom may lead to the activation of other symptoms. In this way, the dynamic causality between symptoms constitutes the essence of mental disorders (Bryant et al., 2017; McNally, 2016). The network approach addresses the deficiency of the latent variable model to a certain extent. In psychopathology research, the network approach can reveal central symptoms, which are more closely connected with other symptoms and may activate other symptoms (McNally, 2016). Thus, central symptoms are valued in network analysis studies (Borsboom & Cramer, 2013).

Existing network analyses of adolescents’ depressive symptoms are based on high school students and have identified loneliness, sadness, self-hatred, self-deprecation, fatigue, pessimism, and crying as central symptoms of depression (Gijzen et al., 2021; Mullarkey et al., 2019). Network analyses have identified loss of pleasure (Bringmann et al., 2015), fatigue or loss of energy, feelings of guilt (van Borkulo et al., 2015) and sadness (Fried et al., 2016; Santos et al., 2017) as central symptoms of depression in adult samples. Therefore, network structure and influential symptoms of depression are different among children and adolescents compared to adults. These differences in depressive symptoms are mainly attributed to the fact that children and adolescents are in different developmental periods compared to adults in which they have different biological and psychosocial characteristics (Crone & Dahl, 2012; Mezulis et al., 2014). Compared to adults, children and adolescents are likely to experience loneliness because from the late childhood period to puberty, individuals become more sensitive to social relations, and their requirements for relationships rise from quantity to quality (Crone & Dahl, 2012). Moreover, individuals have become more reliant on themselves and peers rather than parents since this period (Mezulis et al., 2014). In addition, individuals are also more likely to hate themselves in this period due to dissatisfaction with body image and susceptibility evaluation of others (Ge & Natsuaki, 2009).

However, the results of the depression network differ between clinical and general population samples. Low severity levels of depressive symptoms and prevalence rates of depression in community samples cannot reflect the complexity and severity of depression in clinical samples (Armour et al., 2017). Thus, previous conclusions (Gijzen et al., 2021; Mullarkey et al., 2019) may be difficult to generalise to children and adolescents with high severity levels of depressive symptoms after natural disasters, which often cause common depressive symptoms among children and adolescents (Li et al., 2020).

Moreover, the stability of the network analysis results is an issue of great concern in the field of psychopathology (Robinaugh et al., 2020). A longitudinal study constructed a regularised partial correlation network on 7 occasions between ages 5 and 14 years and found stable central symptoms and increasing connectivity in depression and anxiety networks as the children aged (McElroy et al., 2018). A strongly connected network depends on the connectivity degree of symptoms. If a network has high symptom connectivity, the activation of only one symptom can quickly trigger other symptoms, leading to a state of mental disorder (Cramer et al., 2010; Robinaugh et al., 2020). Previous research found that strongly connected networks featured stronger feedback among their symptoms and thus might increase the vulnerability to depression and lead to fewer positive prospects for recovery from depression (van Borkulo et al., 2015). Therefore, we hypothesise that highly stable levels of depression after natural disasters could be associated with the temporal stability of certain central symptoms and the increase in network connectivity over time.

In the current study, a large-scale longitudinal sample of children and adolescents who experienced the Zhouqu debris flow was selected to understand the depressive symptom network and evaluate its temporal stability among children and adolescents. The Zhouqu debris flow was a major natural disaster that occurred in Gansu Province, China. At 11:00 pm on August 7, 2010, a sudden downpour in the mountainous area in northeastern Zhouqu County triggered landslides. The mudslides passed through the densely populated areas of the county and blocked the Bailong River. More than half of the Zhouqu County urban area was flooded, causing significant loss of personnel and property. In the disaster, 1,557 people were killed, 284 people were missing, and 2,315 outpatients were treated. Additionally, 4,496 households and 20,227 people were affected by the disaster, 1,417 acres of farmland were destroyed by water, and 5,508 houses were destroyed by the disaster (Buzohre et al., 2018).

2. Methods

2.1. Participants and procedure

The study samples were collected after the Zhouqu debris flow. Data collection was conducted at 3 months (T1: early November 2010), 15 months (T2: early November 2011), and 27 months (T3: end of October 2011) following the disaster. At T1, 3957 participants were recruited in grades four to nine from 2 primary and 2 secondary schools located in Zhouqu County. These 4 schools were those most affected by the disaster and covered the vast majority of children in this age range in Zhouqu County. The exclusion criteria were a prior diagnosis of mental illness or brain injury. At T2, 5344 participants were surveyed in grades five to ten from the same 4 schools and a high school, including most of the high school students in this county. At T3, 3724 participants were surveyed in grades six to eleven from the aforementioned 5 schools. In each survey, the participants were measured collectively through questionnaires in a group format during class. The questionnaires were distributed and administered by two or three volunteers who were uniformly trained by the same standardised instructions, and the class teachers were present in the classrooms. The ethics review committee of the Institute of Psychology, Chinese Academy of Sciences approved the study design and procedures. Detailed information on the study design and data collection process is presented elsewhere (Liang et al., 2021b). The data that support the findings of this study are available from the corresponding author upon reasonable request.

The data analysis included 1600 participants who participated in all three investigations. Transferring to other schools or entering secondary or high school were the main reasons for sample loss. After screening out careless answers and missing questionnaires, 1,466 participants were included in the final analysis. The screening criterion depended on the CDI response rate at each temporal point. The participants who answered no less than 22 of 27 items were included in the final analysis. The retention rates at each temporal point were 93.88%, 98.25% and 98.88%. The overall retention rate was 91.63%. In the final sample, 678 participants were male, and 788 were female. In the first survey, 719 students were from grades four to six, 747 students were from grades seven to nine, and the mean age of the participants was 12.71 (SD = 2.30). Detailed demographic and trauma exposure information for the participants is shown in Table 1.

Table 1.

Demographic and trauma exposure information (n = 1,466).

	Category	N (%)
Gender	Male	678 (46.25)
	Female	788 (53.75)
Grade	4–6	719 (49.05)
	7–9	747 (50.95)
Father’s education	Never went to school	176 (12.01)
	Primary school	296 (20.19)
	Junior high school	506 (34.52)
	High school	277 (18.89)
	College and above	150 (10.23)
Mother’s education	Never went to school	608 (41.47)
	Primary school	355 (24.22)
	Junior high school	253 (17.26)
	High school	132 (9.00)
	College and above	48 (3.27)
Parental marital status	First marriage	1386 (94.54)
	Divorced	21 (1.43)
	Remarried	22 (1.50)
Prior trauma exposure	Yes	541 (36.90)
	No	873 (59.55)
Trauma exposure	Was trapped	212 (14.46)
	Was injured	63 (4.30)
	Witnessed a building collapse	411 (28.04)
	Witnessed a dead body	454 (30.97)
	Lost one (or more) relative(s)	586 (39.97)
	Home damage	106 (7.23)

Note: Some variable totals did not reach 1,466 because of missing data.

2.2. Measures

Depressive symptoms and demographic information were assessed in the classrooms. The Child Depression Inventory (CDI; Kovacs, 1992) was administered to evaluate depressive symptoms. The CDI includes 27 items divided into 5 subscales—anhedonia, negative mood, negative self-esteem, ineffectiveness and interpersonal problems—to assess depressive symptoms in children. All the items are assessed on a 3-point scale ranging from 0 to 2, and the children rated each item according their situation during the preceding two weeks. The CDI is widely used in clinical practice and research among children and adolescents, and a cutoff for a total score equal to 19 or greater is adequate for the general screening of depression (Timbremont et al., 2004). In this study, the scale exhibited good internal consistency at each temporal point (Cronbach’s αs of 0.84, 0.81, and 0.85).

2.3. Data analysis

All descriptive statistical analyses were performed with SPSS (Version 22.0 for Windows), and network analysis was implemented in R software (Version 4.1.2) according to the standard guidelines (Epskamp et al., 2018). In the final sample, the depression item-level data were missing for 1.0%, 0.5% and 0.3% at each temporal point. The missing data were handled using the expectation maximisation (EM) algorithm. The EM algorithm is a viable option for dealing with missing data based on a maximum-likelihood estimation of parameters (Chen et al., 2020). When the parameters are estimated (e.g. means, variances/covariances, etc.), expected values for the missing data can be derived from the EM algorithm.

The presence of the CDI items at each temporal point is presented in Table 2. Means, standard deviations, and the differences between the CDI item scores at the temporal points are shown in the Supplemental Materials (Table S4). Since the optimal way to model a network analysis with trichotomous items remains debatable (Fried & Nesse, 2015), each item in the CDI was dichotomised to signify whether depressive symptoms were absent (0) or present (1). Item values of 0 were recoded as absence (0), while item values of 1 or 2 were scored as presence (1), which is consistent with previous network analysis studies using the CDI (Gijzen et al., 2021; Mullarkey et al., 2019).

Table 2.

Prevalence of the CDI symptoms at 3 temporal points (n = 1,466).

Symptoms	T1		T2		T3
	% Absence (‘0’)	% Presence (‘1’)	% Absence (‘0’)	% Presence (‘1’)	% Absence (‘0’)	% Presence (‘1’)
AN1	27.35	72.65	36.63	63.37	36.77	63.23
AN2	50.34	49.66	60.71	39.29	52.52	47.48
AN3	70.60	29.40	73.19	26.81	66.98	33.02
AN4	43.52	56.48	53.62	46.38	47.82	52.18
AN5	57.84	42.16	64.67	35.33	62.96	37.04
AN6	52.05	47.95	54.77	45.23	53.34	46.66
AN7	50.68	49.32	10.16	89.84	9.75	90.25
AN8	33.22	66.78	38.95	61.05	41.00	59.00
NM1	68.83	31.17	77.90	22.10	78.24	21.76
NM2	44.95	55.05	52.59	47.41	49.59	50.41
NM3	41.75	58.25	38.68	61.32	41.13	58.87
NM4	81.58	18.42	87.38	12.62	85.88	14.12
NM5	57.57	42.43	56.89	43.11	60.16	39.84
NM6	26.19	73.81	27.22	72.78	27.29	72.71
NS1	31.11	68.89	28.99	71.01	26.88	73.12
NS2	75.92	24.08	76.74	23.26	75.38	24.62
NS3	66.51	33.49	67.39	32.61	62.76	37.24
NS4	38.68	61.32	41.00	59.00	45.91	54.09
NS5	40.79	59.21	36.02	63.98	32.74	67.26
IF1	63.85	36.15	64.26	35.74	61.60	38.40
IF2	47.68	52.32	41.06	58.94	38.95	61.05
IF3	24.90	75.10	23.12	76.88	21.08	78.92
IF4	40.18	59.82	33.70	66.30	31.17	68.83
IP1	77.01	23.99	78.04	21.96	76.67	23.33
IP2	70.19	29.81	70.12	29.88	70.12	29.88
IP3	73.06	26.94	76.81	23.19	74.42	25.58
IP4	88.47	11.53	90.79	9.21	89.02	10.98

Note: AN1 = anhedonia, AN2 = sleep disturbance, AN3 = fatigue, AN4 = reduced appetite, AN5 = somatic concerns, AN6 = loneliness, AN7 = school dislike, AN8 = lack of friendship, NM1 = sadness, NM2 = pessimistic worrying, NM3 = self-blame, NM4 = crying, NM5 = irritability, NM6 = indecisiveness, NS1 = pessimism, NS2 = self-hate, NS3 = suicidal ideation, NS4 = negative body image, NS5 = feeling unloved, IF1 = self-deprecation, IF2 = school work difficulty, IF3 = school performance decrement, IF4 = low self-esteem, IP1 = misbehaviour, IP2 = social withdrawal, IP3 = disobedience, IP4 = fighting.

2.3.1. Network estimate

The network structure of depressive symptoms at 3 temporal points (3, 15 and 27 months following the Zhouqu debris flow) was estimated by the Ising model, which is used for estimations of binary data (van Borkulo et al., 2014). The Ising model can be regarded as a series of pairwise associations among binary variables after controlling for all the other associations. The Ising model also integrates eLasso based on the extended Bayesian information criterion, which shrinks small edge coefficients to zero in the underlying network structure accurately (Ravikumar et al., 2010). To avoid potential node overlap, we adopted the ‘Goldbricker’ function from the networktools package to examine the data (Jones, 2018). The threshold was set as 0.25 (Levinson et al., 2018).

The R package qgraph was used to plot networks, twenty-seven items in the CDI were depicted as nodes, and the associations between the depressive symptoms were depicted as edges in the network. The green (red) edges indicate positive (negative) associations between two nodes. To facilitate the visual network comparison at different temporal points, the averageLayout function in the qgraph package was used, which presents an accordant layout of the nodes according to the average position across several networks.

2.3.2. Centrality estimation

Strength, expected influence, betweenness and closeness are four common graph theoretical centrality measures (Burger et al., 2022). Strength is the absolute value of the weights of all edges connected to a node, which is the most popular and robust centrality index in network research (Birkeland et al., 2020). Expected influence is used to solve the problem that networks contain negative edges: it considers both negative and positive edges and outperforms strength centrality in networks containing negative edges. Betweenness is the number of times that a node lies on the shortest path connecting two other symptoms, and a symptom with a high betweenness centrality may serve as a bridge connecting other symptoms. Closeness is the average distance from a given node to all other nodes in the network and is calculated by taking the inverse of all the shortest paths’ distances from a node to all other nodes. Because closeness makes more sense in epidemiological studies than it does in psychological studies (McNally, 2016), we did not estimate it in this study. In addition, previous studies have demonstrated that betweenness and closeness are often not reliably estimated (Birkeland et al., 2020). Therefore, we mainly report the results of the expected influence in the main text, and the results of strength and betweenness are shown in the Supplemental Materials. Consistent with previous studies, we considered nodes with EI greater than one standard deviation as having high centrality (Birkeland et al., 2020).

2.3.3. Stability estimation

To examine the edge weight accuracy and centrality stability of the three estimated depression networks, two robustness analyses were analysed using the R package bootnet (Epskamp et al., 2018). The edge accuracy was estimated by bootstrapping the 95% confidence intervals (CIs) of the edge weights (bootstrapped samples = 1,000), and fewer overlaps among the CIs indicated a higher accuracy. Node centrality stability was estimated by case-dropping bootstraps. The correlation stability (CS) coefficients represent the maximum proportion of cases that can be excluded from the subsets while retaining a high correlation (0.7 or higher) with the original centrality indices. CS coefficients greater than 0.25 (0.5) indicate moderate (strong) stability (Epskamp et al., 2018).

2.3.4. Network comparison

A network comparison test (NCT) was used to compare the depression networks among the three temporal points, including tests of network invariance, global strength invariance, edge invariance and centrality invariance (Fried et al., 2017). Network and global strength invariance were used to test for differences in the entire network structure and global connectivity, which weighted the absolute sum of all the edges in a network (Opsahl et al., 2010). Edge invariance was used to test for differences in edge weights, and centrality invariance was used to test the invariance of the expected influence centrality between the networks. All the NCTs used the R package NetworkComparisonTest (van Borkulo et al., 2022).

3. Results

3.1. Descriptive statistics

The average depression scores among the participants at the three temporal points (3, 15, 27 months) were 15.56 (SD = 7.57), 15.20 (SD = 6.78), and 15.60 (SD = 7.34), respectively. The prevalence rates of probable depression (≥ 19) at each temporal point were 34.99%, 29.88%, and 33.63%.

3.2. Networks and centrality estimation

The ‘Goldbricker’ function was used to check the data for potential overlap and did not yield any redundant nodes at any of the temporal points. Figure 1 shows the depressive symptom networks for children and adolescents at 3 (T1, Panel A), 15 (T2, Panel B), and 27 months (T3, Panel C). Many similar patterns of symptom connections appeared in the depression networks across the temporal points. The strong connections at each temporal point included sleep disturbance and reduced appetite (AN2:AN4), sadness and crying (NM1:NM4), and misbehaviour and disobedience (IP1:IP3).

Figure 1.

Depressive symptom networks at 3 (a), 15 (b) and 27 months (c). Note: Node colors refer to the symptoms on the five subscales. The edges are colored green (red) for positive (negative) correlations. The thickness of an edge indicates the association strength.

The expected influence centrality of the depressive symptoms at each temporal point is shown in Figure 2. Self-hate (NS2) and loneliness (AN6) were among the most central nodes in all networks, and sleep disturbance (AN2) was among the more central nodes in all the networks. Overall, the expected influence centrality of the depressive symptoms was similar at each temporal point, except for crying (NM4) and self-deprecation (IF1). The expected influence of crying was high at T1 but decreased to a relatively high level at T2 and T3, while the expected influence of self-deprecation (IF1) was low at T1 but increased to a high level at T3. The results of the node strength centrality and betweenness for the networks at the three temporal points are shown in Supplementary Materials, Figure S1.

Figure 2.

Standardized estimates of the expected node influence for the networks at different temporal points. Note: AN1 = anhedonia, AN2 = sleep disturbance, AN3 = fatigue, AN4 = reduced appetite, AN5 = somatic concerns, AN6 = loneliness, AN7 = school dislike, AN8 = lack of friendship, NM1 = sadness, NM2 = pessimistic worrying, NM3 = self-blame, NM4 = crying, NM5 = irritability, NM6 = indecisiveness, NS1 = pessimism, NS2 = self-hate, NS3 = suicidal ideation, NS4 = negative body image, NS5 = feeling unloved, IF1= self-deprecation, IF2 = school work difficulty, IF3 = school performance decrement, IF4 = low self-esteem, IP1 = misbehaviour, IP2 = social withdrawal, IP3 = disobedience, IP4 = fighting.

3.3. Network accuracy and stability

The results of edge weight bootstrapping revealed that the depressive networks at each temporal point were moderately accurately estimated (see Supplementary Materials, Figure S2). In the networks at all temporal points, the 95% CIs of the edge weights exhibited considerable overlap; nonoverlapping CIs also existed. Moreover, most of the strongest edges were significantly different from most of the other edges in the network (see Supplementary Materials, Figure S3).

The subset bootstrap showed stable and interpretable estimates of the order of the node expected influence centrality in the networks at the 3 temporal points (see Figure S4). The CS coefficients for the expected influence were 0.52, 0.52 and 0.59 at T1, T2 and T3, respectively. The results of subset bootstrapping for the node strength centrality and betweenness are shown in Supplementary Materials, Figure S5. Moreover, the centrality difference tests favoured the node centrality order in each network. The node with the greatest centrality was statistically stronger than most of the other nodes in the networks at each point (see Supplementary Materials, Figure S6).

3.4. Network comparison

The NCT results showed that global connectivity in the 3 networks did not differ significantly (all p > .05; global strength: T1 = 38.61, T2 = 42.96, T3 = 44.11). Structurally, the network at T1 was significantly different from that at T2 and T3 (all p < .02), but the networks at T2 and T3 were not significantly different (p = .937). The NCT also showed that 11.11% (T1 compared to T2), 6.91% (T2 compared to T3) and 12.31% (T1 compared to T3) edges in the networks were significantly different at the three temporal points. For instance, the association between sleep disturbance (AN2) and fatigue (AN3) was stronger over time, while pessimism (NS1) and self-deprecation (IF1) had a strong association at T3 but not at T1 and T2. In addition, the edge between sleep disturbance (AN2) and irritability (NM5) was observed only at T1. Moreover, the expected influence of school dislike (AN7), lack of friendship (AN8) and pessimism (NS1) were significantly different between T1 and T2; the expected influence of self-deprecation (IF1) was significantly different between T2 and T3; and loneliness (AN6), school dislike (AN7), lack of friendship (AN8), sadness (NM1), pessimism (NS1) and self-deprecation (IF1) were significantly different between T1 and T3. Overall, the depressive network at T1 was slightly different from the networks at T2 and T3 in terms of network structure, edge and node expected influence, while the depressive networks at T2 and T3 were very similar (see Table 3).

Table 3.

Comparison of the node expected influence among the temporal points (n = 1,466).

Symptoms	p (T1 vs. T2）	p (T2 vs. T3）	p (T1 vs. T3）
AN1	.130	.876	.086
AN2	.172	.871	.284
AN3	.946	.587	.541
AN4	.530	.578	.695
AN5	.160	.355	.412
AN6	.545	.214	.041
AN7	<.001	.479	<.001
AN8	.005	.650	<.001
NM1	.187	.573	.048
NM2	.133	.436	.280
NM3	.614	.403	.170
NM4	.306	.899	.358
NM5	.417	.892	.530
NM6	.582	.932	.537
NS1	.011	.248	<.001
NS2	.114	.620	.234
NS3	.624	.835	.515
NS4	.507	.353	.962
NS5	.866	.550	.512
IF1	.404	.037	.006
IF2	.568	.747	.746
IF3	.425	.659	.817
IF4	.142	.156	.912
IP1	.925	.328	.383
IP2	.962	.882	.974
IP3	.483	.665	.784
IP4	.892	.845	.933

Note: AN1 = anhedonia, AN2 = sleep disturbance, AN3 = fatigue, AN4 = reduced appetite, AN5 = somatic concerns, AN6 = loneliness, AN7 = school dislike, AN8 = lack of friendship, NM1 = sadness, NM2 = pessimistic worrying, NM3 = self-blame, NM4 = crying, NM5 = irritability, NM6 = indecisiveness, NS1 = pessimism, NS2 = self-hate, NS3 = suicidal ideation, NS4 = negative body image, NS5 = feeling unloved, IF1 = self-deprecation, IF2 = school work difficulty, IF3 = school performance decrement, IF4 = low self-esteem, IP1 = misbehaviour, IP2 = social withdrawal, IP3 = disobedience, IP4 = fighting.

Discussion

To our knowledge, the present study is the first to evaluate the temporal stability of depressive symptom networks among children and adolescents. Beginning three months after the Zhouqu debris flow, depressive symptom networks were found to be similar regarding symptom centrality and global connectivity over the two-year study period. At the three temporal points, self-hate, loneliness, and sleep disturbance were the central symptoms in the networks, and the associations between sleep disturbance and reduced appetite, sadness and crying, and misbehaviour and disobedience were strong in the depression network.

Self-hate and loneliness were the symptoms with the greatest centrality in the depressive symptom networks across time among children and adolescents, which is similar to previous findings (Gijzen et al., 2021; Mullarkey et al., 2019). Unlike in adults, in youth, self-hate is an important symptom of depression (Bringmann et al., 2015; Fried et al., 2016). Individuals entering puberty experience identity development (Dahl, 2004). Meanwhile, information-processing biases may be likely to emerge during adolescence (Jacobs et al., 2008). These processes may lead to increased negative self-referent thinking (Connolly et al., 2016). In the face of the destructive power of debris flows, youth might feel powerless and unable to give aid to family, friends or other lives, which gave rise to a harmful self-narrative. In addition, we found that the connection between negative body image and self-hate was strong in the depression network over time, which is consistent with previous research (Mullarkey et al., 2019). This link may be important for depression in children and adolescents. In early puberty, individuals face body image changes, and dissatisfaction with body image is a common stressor in youth, which may lead to self-hate, especially in females (Mezulis et al., 2014; Mullarkey et al., 2019). Additionally, interventions targeting dissatisfaction with the body can decrease depressive symptoms (Bearman et al., 2003).

Loneliness is another important depressive symptom in youth that differs from that in adults. Several network analyses identified loneliness as a high-central symptom of depression among children and adolescents (Gijzen et al., 2021; Mullarkey et al., 2019) but not among adults (Bringmann et al., 2015; Fried et al., 2016). The main reason for this difference may be that children and adolescents are more likely to experience loneliness than adults. From late childhood to puberty, individuals become more sensitive to social relations, and their relationship requirements change from quantity to quality (Crone & Dahl, 2012). Moreover, individuals in this period become more reliant on themselves and their peers than on their parents (Mezulis et al., 2014). Following the disaster, a family might invest more time in returning to normalcy and temporarily pay less attention to children, which can lead to loneliness.

Sleep disturbance was among the more central nodes in depression networks, which has not been found in previous studies of both children and adults (Bringmann et al., 2015; Fried et al., 2016; Gijzen et al., 2021; Mullarkey et al., 2019). A possible reason is that sleep disturbance is a characteristic symptom in children after natural disasters. Traumatic events such as natural disasters can severely disrupt an individual’s sleep integrity and continuity and have been shown to be one of the most important triggers of sleep disturbances (Charney, 2003; Geng et al., 2013). Traumatic events may produce a sustained neurobiological response, thus destroying the normal sleep arousal regulation mechanism and resulting in central and physiological excitement (Sinha, 2016). Posttraumatic sleep problems are also common among children and adolescents, and the prevalence of sleep problems in adolescents after the Wenchuan earthquake was 28.8%−30.2% (Geng et al., 2013). In addition, the association between sleep disturbance and reduced appetite was strong in the depression network, which may be related to the disturbance of the 24-hour sleep–wake cycle and the underlying circadian system in adolescence (Crouse et al., 2021).

The expected influence of crying was at a high level at T1 but decreased at T2 and T3, while self-deprecation (IF1) was at a low level at T1 but increased at T3. Crying, as an intense emotional expression in the face of natural disasters and their dire consequences, decreases as life returns to normal. Regarding self-deprecation, children and adolescents may form a feeling of smallness and fragility in the face of nature and its destructive power, which challenges their belief stability and leads to an increase in self-deprecation (Janoff-Bulman, 2010; Zhou et al., 2019). These findings indicated that self-deprecation might develop into an important depression symptom over time after natural disasters.

Sleep disturbance and reduced appetite, sadness and crying, and misbehaviour and disobedience all showed strong connections at each temporal point. The results illustrated the stability of these edges in the postdisaster depression network among children and adolescents and suggested that they could contribute to the maintenance of the depression network. Uncontrollable catastrophic natural disasters are bound to trigger negative physical and mental outcomes among children and adolescents (Tchounwou, 2004). They might become preoccupied with thoughts and information about the disaster, which can affect their sleep quality and lead to loss of appetite. Meanwhile, exposure to the painful consequences of the disaster may lead to sadness, and crying may become a means of venting. Additionally, a seemingly uncontrollable world may give rise to challenging rules and order, namely, engaging in disobedience or a spontaneous rise in misbehaviour. These behaviours have been proven to have a significant effect on the development of depression (Burke & Loeber, 2010).

Despite the high proportion of edge similarity, changed edges are also worth attention. Connectivity between sleep disturbance and fatigue became stronger over time. A possible reason could be self-reinforcing loops of psychopathological symptom activation after the waning of triggering, and the results reflected the phenomenon of hysteresis of the symptom network (Borsboom, 2017). Due to the prolonged effect of debris flow and reconstruction in the local area, the normal lives of children and adolescents were disrupted; they were likely to then become trapped in bad sleep patterns and thus become fatigued. Such a vicious cycle can lead to a tighter connection of symptoms. In addition, the connection between sleep disturbance and irritability was observed only at T1. Sleep disturbance and irritability are hyperarousal symptoms and are usually interrelated in the aftermath of a disaster (Sinha, 2016). As time passed, individuals’ moods gradually eased; thus, the connection between irritability and sleep disturbance subsided.

Overall, the temporal stability of youth’s depressive symptom networks was high across the 2 years after the Zhouqu debris flow. The longitudinal stability observed in the present study lends confidence to the precision and generalizability of parameter estimates of psychopathology networks. Persistent high-central depressive symptoms may play key roles in the onset and development of depression among children and adolescents (Borsboom & Cramer, 2013; Mullarkey et al., 2019).

The present study has several strengths and limitations. We collected a large sample of children and adolescents who encountered the Zhouqu debris flow. The 3-wave data enable us to examine the temporal stability of the depressive symptom network and to identify high central symptoms among youth’s depressive symptoms. These findings provided novel evidence to understand why the prevalence of postdisaster depression was high and stable from the network perspective. Several limitations should also be acknowledged. First, we did not measure children’s depression before the Zhouqu debris flow, so we cannot compare the network structure of depressive symptoms before and after the disaster. Second, the measure of depressive symptoms relied on a self-report questionnaire (CDI) rather than clinical interviews. Despite the good diagnostic utility of CDI (Timbremont et al., 2004), more studies based on structured clinical interviews are still needed. Additionally, items were dichotomised to present the presence of symptoms that could not identify the severity of symptoms across different time points. Third, although approximately 30% of our sample screened positive for depression in each wave, our findings should be cautiously generalised to clinical populations of children and adolescents with depression. Fourth, the present study used a cross-sectional approach that does not allow the edge directionality to be determined. Additionally, cross-sectional findings derived from undirected networks can be understood only at a between-subjects level, rather than at a within level (Fisher et al., 2018). Future studies should discuss the node activation order by using longitudinal network methods such as a cross-lagged panel network (Epskamp et al., 2018). Fifth, our study focused on depressive symptoms in children and adolescents after a natural disaster, and the results may not be generalisable to other disaster types, such as anthropological disasters.

Despite these limitations, our study provides the first evidence for the temporal stability of youth depressive symptom networks. The similarity of networks at different temporal points after natural disasters may aid in understanding the high prevalence and maintenance of postdisaster depression (Guo et al., 2017; Liang et al., 2021c). These findings also lend confidence to the repeatability and generalizability of the analyses of psychopathology networks. Overall, self-hate, loneliness, and sleep disturbances were the central network symptoms across the two years after the Zhouqu debris flow. These symptoms might play important roles in the maintenance of depression among children and adolescents after natural disasters. According to the network perspective, intervening with symptoms with high centrality may speed up the recovery of mental disorders by partially breaking down the interactive network (Ross et al., 2018). After natural disasters, group counselling is the most commonly chosen treatment since numerous local residents suffered from the disaster. Sleep problems, self-hate and loneliness can be given more attention in the diagnosis and group intervention of youths’ depression after disasters. Additionally, an increasing number of researchers have highlighted that symptoms with high centrality cannot be automatically serviced as viable intervention targets (Stelzer et al., 2020). Moreover, it is difficult to target a single symptom without affecting other symptoms simultaneously in clinical work (McNally, 2021). Therefore, the efficacy of treatments related to high central symptoms needs to be further verified in intervention studies.

Funding Statement

This work was supported by National Planning Office of Philosophy and Social Sciences of China [Grant Number No. 22CSH092].

Disclosure statement

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