Core features of positive mental health in adolescents and their protective role against psychopathology

Participants

This study was part of a larger longitudinal project focused on positive mental health among adolescents. Participants completed this battery of measures four times over two months (every two weeks) in the larger project. For this study, we used data from the first timepoint. Participants were aged between 16 and 19 years old (M = 17.53 years, SD = 1.02 years) and were residing in Singapore (n = 1909 after removing 82 participants who had missing data for all variables, 78.9% females). There were 79.94% of Chinese ethnicity, 6.5% Malays, 8.12% Indians, 0.84% Eurasians, and 4.61% of other ethnicities. They were invited to participate through email blasts through their schools and advertisements posted on Instagram.

Procedure

Interested participants responded to the invitation via a link in their email or Instagram advertisements. They were first presented with the information sheet detailing the study and their rights as research participants. After indicating their informed consent to participate, participants were directed to fill out some demographic information, followed by a battery of questionnaires. The A*STAR Institutional Review Board approved this project and the study was conducted in accordance to approved protocol.

Materials

Emotional quotient inventory

The Bar-On Emotional Quotient Inventory Youth, short version³², had 30 items that measured four aspects of emotional intelligence in adolescents: Adaptability (α = 0.85; e.g., “I can easily use different ways of solving problems”), Interpersonal Intelligence (α = 0.74; e.g., “I care what happens to other people”), Intrapersonal Intelligence (α = 0.27; e.g., “It is easy to tell people how I feel”), and Poor Stress Management (α = 0.84; e.g., “I get too upset about things”). Participants rated how much they agreed with each statement on a four-point Likert scale from “not true of me, never, or seldom” to “very much true of me, or very often”. The scale showed reasonable fit on the confirmatory factor analysis (χ²[224] = 2238.35, p < .001, CFI = 0.88, TLI = 0.86, RMSEA = 0.069, 90% CI[0.066, 0.071], SRMR = 0.067).

EPOCH adolescent well-being

The EPOCH Measure of Adolescent Well-Being³³ is a 20-item with five dimensions of positive characteristics that together support higher levels of well-being: Engagement (α = 0.85; e.g., “I get so involved in activities that I forget about everything else”), Perseverance (α = 0.81; e.g., “I finish whatever I begin”), Optimism (α = 0.82; e.g., “In uncertain times, I expect the best”), Connectedness (α = 0.83; e.g., “When something good happens to me, I have people who I like to share the good news with”), and Happiness (α = 0.88; e.g., “I feel happy”). Participants rated how often each statement described them on a five-point Likert scale from 1 (almost never) to 5 (almost always). The confirmatory factor analysis showed acceptable model fit (χ²[160] = 1389.75, p < .001, CFI = 0.94, TLI = 0.92, RMSEA = 0.065, 90% CI[ 0.062, 0.069], SRMR = 0.04).

Flourishing

The Flourishing scale consisted of eight items that described the essential factors of human functioning (α = 0.87), such as positive relationships, feelings of competence, and purpose in life³⁴. Statements included “I lead a purposeful and meaningful life” and “People respect me”. Participants rated how much they agreed with each statement on a six-point Likert scale from one (strongly disagree) to six (strongly agree). The scale showed reasonable fit on the confirmatory factor analysis (χ²[20] = 292.73, p < .001, CFI = 0.95, TLI = 0.93, RMSEA = 0.087, CI [0.078, 0.095], SRMR = 0.034).

Growth mindset

The Growth Mindset questionnaire³⁵ measured how much individuals believe that they can get smarter if they work on it (α = 0.91; e.g., “You have a certain amount of intelligence, and you can’t really do much to change it”, reverse coded). On a six-point Likert scale from one (strongly disagree) to six (strongly agree), participants rated how much they agreed with each item. Confirmatory factor analysis could not be done with a three-item model.

Inventory of parent-peer attachment

The revised version of the Inventory of Parent-Peer Attachment is a 28-item measure of attachment towards parents and peers³⁶. Within the parents or peers domain, there were three factors: Communication (parents α = 0.90, e.g., “I like to get my parents’ view on things I’m worried about”; peers α = 0.86, e.g., “I like to get my friends’ opinions on things I’m worried about”), Trust (parents α = 0.92, e.g., “I can depend on my parents to help me solve a problem”; peers α = 0.83, e.g., “I can count on my friends to listen when something is bothering me”), and alienation (parents α = 0.77, e.g., “I feel silly or ashamed when I talk about my problems with my parents”; peers α = 0.71, e.g., “I feel silly or ashamed when I talk about my problems with my friends”). On a five-point Likert scale of 1 (almost never or never true) to 5 (almost always or always true), participants rated how true each statement is to them regarding their relationship with peers or parents respectively. The confirmatory factor analysis showed reasonable model fit (χ²[242] = 2545.95, p < .001, CFI = 0.91, TLI = 0.89, RMSEA = 0.076, 90% CI [0.074, 0.079], SRMR = 0.05).

Rosenberg self-esteem

The Rosenberg Self-Esteem scale is a ten-item questionnaire on global self-worth³⁷. The participants rated the items on a six-point Likert scale from one (strongly disagree) to six (strongly agree). As the single-factor structure with reverse coded items has been criticized in the literature³⁸, we limited the scoring to the five positively-worded items (e.g., “I feel that I am a person of worth, at least on an equal basis with others”) as measure of self-esteem (α = 0.89). The confirmatory factor analysis of the five-item scale showed reasonable fit (χ²[5] = 418.30, p < .001, CFI = 0.91, TLI = 0.83, RMSEA = 0.21, 90% CI [0.197, 0.232], SRMR = 0.04).

Self-concept clarity

The twelve-item Self-Concept Clarity scale (α = 0.87) measures the extent to which one’s self-beliefs are clearly defined, internally consistent, and stable³⁹. Some sample questions included “My beliefs about myself often conflict with one another” (reverse-coded) and “In general, I have a clear sense of who I am and what I am”. Participants indicated how much they agreed with each statement on a six-point Likert scale from one (strongly disagree) to six (strongly agree). The confirmatory factor analysis showed reasonable model fit (χ²[56] = 1033.30, p < .001, CFI = 0.87, TLI = 0.84, RMSEA = 0.10, 90% CI [0.097, 0.11], SRMR = 0.05).

Subjective well-being

Subjective Well-Being is a five-item scale (α = 0.90) that measures global life satisfaction²⁰. On a Likert scale from one (strongly disagree) to six (strongly agree), participants rated how much they agreed with each statement. Some sample items included “The things in my life are excellent” and “I am happy with my life”. The confirmatory factor analysis showed good model fit (χ²[5] = 23.22, p < .001, CFI = 0.997, TLI = 0.99, RMSEA = 0.05, 90% CI[0.03, 0.07], SRMR = 0.01).

Singapore youth resilience scale

The Singapore Youth Resilience Scale⁴⁰ has 50-item that measured ten domains of resilience: Emotional Regulation (α = 0.87; e.g., “I stay calm in difficult circumstances”), Personal Control (α = 0.42; e.g., “Failure does not easily discourage me”), Personal Confidence (α = 0.70; e.g., “I am confident that I can solve problems in life”), Flexibility (α = 0.65; “I can accept it when things are unclear and uncertain”), Committed (α = 0.81; e.g., “When I start doing something I try to finish it”), Positive Self-image (α = 0.87; “I accept myself”), Positive Coping (α = 0.72; e.g., “I know which situations I can handle and which I cannot”), Humor (α = 0.84; e.g., “I can see the funny side of things”), Social Support (α = 0.75; e.g., “I have good friends that I can trust”), and Spirituality (α = 0.85; e.g., “I believe my life has meaning and purpose”). Participants rated how much they agreed with each statement on a six-point Likert scale from one (strongly disagree) to six (strongly agree). “Personal control” was removed from the confirmatory factor analysis due to very low internal consistency. The remaining subscales showed reasonable model fit (χ²[953] = 8426.89, p < .001, CFI = 0.81, TLI = 0.79, RMSEA = 0.068, 90% CI[0.067, 0.069], SRMR = 0.067).

Tromsø social intelligence scale

The Tromsø Social Intelligence scale⁴¹ measures three aspects of social intelligence: Social Information Processing (α = 0.85; e.g., “I can predict other peoples’ behavior”), Social Skills (α = 0.80; e.g., “I fit in easily in social situations”), and Social Awareness (α = 0.80; e.g., “I often feel that it is difficult to understand others’ choices”, reverse-coded). Participants rated how much each statement described them on a seven-point Likert scale from one (describes me extremely poorly) to seven (describes me extremely well). The scale showed reasonable model fit (χ²[132] = 1423.59, p < .001, CFI = 0.89, TLI = 0.87, RMSEA = 0.073, 90% CI[0.069, 0.076], SRMR = 0.066).

Positive-negative affect schedule

The Positive-Negative Affect Schedule⁴² measures one’s experience of either positive (α = 0.88; e.g., “determined”) or negative (α = 0.88; e.g., “distressed) affect over a week. Participants rated how frequently they experienced each of the twenty emotions over the past week on a five-point Likert scale from one (very slightly or not at all) to five (extremely). The scale showed reasonable model fit (χ²[169] = 2334.15, p < .001, CFI = 0.83, TLI = 0.81, RMSEA = 0.09, 90% CI[0.089, 0.096], SRMR = 0.07).

Perceived stress scale

The four-item Perceived Stress Scale⁴³ measures the degree to which the individual appraised situations in life as stressful (α = 0.70). Participants indicated their response to each question according to their experience in the last week on a five-point Likert scale, from zero (never) to four (very often). Sample statements included “How often have you felt that things were going your way?” and “How often have you felt that you were on top of things?” The scale showed good model fit (χ²[2] = 1.70, p = .43, CFI = 1.00, TLI = 1.00, RMSEA < 0.001, 90% CI[0, 0.049], SRMR = 0.006).

PHQ-9

The Patient Health Questionnaire-9 (α = 0.90) is a nine-item screening tool for depressive symptoms⁴⁴. On a four-point Likert scale ranging from zero (not at all) to three (nearly every day), participants rated how often they felt bothered by each statement over the past two weeks. The items included “little interest or pleasure in doing things” and “poor appetite or overeating”. The scale showed good model fit (χ²[27] = 444.02, p < .001, CFI = 0.94, TLI = 0.92, RMSEA = 0.10, 90% CI[0.093, 0.109], SRMR = 0.04).

Analysis plan

Items removal

Goldbricker in the networktools R package⁴⁵ was used to detect and remove redundant variables by computing every possible correlation in the network. We set the minimum correlation to 0.50 as this is the threshold for strong correlation⁴⁶. The p-value and threshold of significant proportion were set at 0.01 and 0.25, respectively⁴⁷. Variables correlated with coefficient at least r = .60 with less than 25% overlapping correlations (i.e., significantly different correlations with α = 0.01) were considered redundant. Redundant pairs of variables were merged by choosing the more unique variable and deleting the other.

We also removed subscales with low internal reliability (α < 0.70). Three subscales were removed: (1) Intrapersonal Intelligence from the Emotion Quotient Inventory (α = 0.27), and (2) Flexibility (α = 0.65), and (3) Personal Control (α = 0.42) of the Singapore Youth Resilience Scale.

Undirected networks

Network analysis, grounded in network theory, has been increasingly utilized to study various psychological constructs or phenomena, ranging from psychopathology⁴⁸, personality⁴⁹, effect of environment on subjective well-being⁵⁰, to attitudes⁵¹, intelligence⁵² and openness to experience⁵³. In network analysis, a psychological construct is not viewed as a latent phenomenon but as an emergent property of a network of elements (or nodes) that interact with one another^54,55. This approach enables a more nuanced understanding of adolescent well-being, characterizing it as a complex system of interconnected elements or facets of positive mental health rather than a homogeneous construct.

Network estimation. We estimated the undirected Gaussian graphical network using the bootnet R package⁵⁶ with stepwise estimation using the ggmModSelect algorithm recommenced for reliable estimation of centrality indices as well as high precision and specificity of edges⁵⁷. Missing data was accounted for by full-information maximum likelihood. The edges between nodes represent partial correlations after controlling for all the network nodes, with thicker edges indicating stronger relationships and blue or red edges indicating positive or negative relationships, respectively. The network was visualized using the qgraph R package⁵⁸.

Expected influence of positive mental health nodes. We focused solely on the positive mental health variables to identify highly influential or central nodes that could facilitate the maintenance or (de)activation of positive mental states. As such, we first estimated an undirected network that was limited to positive mental health nodes. The nodes related to negative mental health or psychopathology were excluded. These negative nodes included the PHQ-9, perceived stress, alienation from peers and parents, and poor stress management. To find the most influential positive mental health nodes, we computed the expected influence of the nodes in the positive mental health network. The expected influence of a node calculates the strength and direction of connections it is to other nodes in the network. A highly influential node has a greater likelihood of either activating other nodes following its activation as it has stronger positive connections to the other nodes within the network⁵⁹.

Expected influence of bridging nodes. Positive mental health serves to protect against the development of psychopathological symptoms. However, given the diversity of positive mental health, we lack clarity on the specific mechanisms or elements that confer this protective effect. As such, we estimated a second undirected network that incorporated the negative mental health nodes that were previously excluded. In this network, we were primarily interested in the bridge expected influence of the nodes. The bridge expected influence indicates the strength and direction of the connection a node has to other nodes in the other community (i.e., between positive and negative mental health communities). The bridge expected influence (bEI) allows us to identify specific positive components more strongly connected to the community of negative mental health symptoms⁶⁰. Specifically, positive mental health nodes with strong protective associations (i.e., negative edges) with negative mental health would have higher negative expected influence values⁵⁹.

We were interested in nodes with the strongest negative bridge expected influence. One-step bEI measures the influence of a node on its immediate neighbors while two-step bEI measures the overall influence of the node. A node may have low one-step bEI but high two-step bEI as it has minimal direct association with other nodes but its immediate neighbors are highly influential within the network. A node may have few but highly influential immediate neighbors, thus allowing it to exert high influence on the network through its neighbors⁵⁹.

The stability of the expected influence coefficient was assessed using 2,500 case-dropping bootstrap samples. The stability was quantified by correlation stability (also known as ‘CS coefficient’) which indicated how much data could be dropped and yet retain a correlation of ≥ 0.70 with the original centrality coefficients with 95% confidence. The CS coefficient should be > 0.50 for the expected influence coefficients to be considered stable and interpretable⁵⁶.

Bayesian directed acyclic graph

While Gaussian graphical models allow us to compute the centrality indices, the edges in these models are undirected, making it difficult to discern whether central nodes are the common cause or consequence within the network⁶¹. To overcome this limitation, we also employ Bayesian directed acyclic graphs (DAGs) to provide a comprehensive network analysis of adolescent mental health. In a DAG, the edges represent probabilistic dependencies between nodes, with the direction indicating plausible causation from one node to another⁶². Although we cannot make definitive causal inferences from observational data, DAGs offer insights into plausible causal relationships within the network⁶³. Using DAGs, we can place components of positive mental health within a putative causal hierarchy, offering a deeper understanding of how positive mental health can be enhanced and how it protects against poor mental health.

The bnlearn R package⁶⁴ was used to estimate the Bayesian directed acyclic graph (DAG) of negative and positive mental health nodes. We used the score-based structural expectation-maximization algorithm⁶⁵ that would impute missing values based on likelihood weighting, estimate network parameters using maximum likelihood⁶⁶, and use a hill-climbing algorithm to learn the network structure. We also bootstrapped the network 10,000 times. The final network only included edges that appeared in more than 85% of the bootstrapped networks with the direction appearing in more than 50% of the bootstrapped networks⁶⁷. The final DAG reveals the most likely directional links between the variables with the edge thickness, indicating the estimated effect’s statistical confidence⁶².

Sensitivity analyses

As there was an unequal gender distribution in our sample, we conducted sensitivity analyses to examine if there was a gender bias in our results. We created a subset dataset of equal gender distribution by randomly selecting 403 female participants and merging with the 403 male participants we had. We estimated the undirected networks with this subset data with the same protocols and compared them to the original networks with the Network Comparison Test R package⁶⁸.