Hello Darkness, My Old Friend: Moderating a Random Intercept Cross-lagged Panel Model of Loneliness and Symptoms of Anxiety and Depression

According to the evolutionary theory of loneliness by Cacioppo and Cacioppo (2018), perceived social isolation (i.e., loneliness) functions as a warning signal that brings attention to the possible deterioration of the body. Humans are regarded as inherently social beings (Cacioppo & Patrick, 2008), and when the innate, strong desire to connect with others is thwarted, people become motivated to fix the perceived deficiencies in their social relationships and thus avoid the negative emotions (Cacioppo & Hawkley, 2009). Paradoxically, people can also be motivated to be alert and avoid possible social dangers to ensure their self-preservation (Cacioppo & Cacioppo, 2018), or in other words, socially withdraw. The evolutionary theory proposes that the signal to self-preserve (i.e., feeling lonely) triggers several behavioural and physical adjustments to manage and deal with faulty social relations, thus avoiding premature mortality. One adjustment is increased depressive symptomatology (Cacioppo & Cacioppo, 2018). Several studies have found support for this theorised effect (for an overview, see e.g., Loades et al., 2020).

The interpersonal theory of depression (Coyne, 1976) argues that depressed and depressed-prone individuals have certain characteristics and behave in manners that impede social relationships by eliciting negative reactions from others (Giesler et al., 1996; Gotlib & Hammen, 1992; Joiner et al., 1992; Swann et al., 1992) and produce interpersonal stress and conflict (Hammen, 2020; Rudolph et al., 2000). For example, depressed people may attempt to decrease feelings of personal guilt and low self-worth through excessive reassurance from people they are close to (Coyne, 1976). Initially, people might offer support. However, because the depressed person is uncertain about the genuineness, they continue to demand reassurances which causes others to become agitated and reject them (Starr & Davila, 2008). Indeed, studies indicate that, compared to non-depressed people, depressed individuals are increasingly likely to experience social dysfunctions (Gotlib & Lee, 1989), less enjoyment and intimacy from social interactions (Nezlek et al., 2000), and withdraw from social interactions (Schaefer et al., 2011).

The cumulative interpersonal risk model (Epkins & Heckler, 2011) argues that although symptoms of anxiety (particularly social anxiety) and depression have similar etiological influences, most researchers overlook the considerable overlap in or comorbidity of anxiety and depressive symptoms. For instance, recent findings imply that certain genomes are related to the co-occurrence of anxiety and depressive symptoms (Jami et al., 2022). Thus, in this study, we combine symptoms of anxiety and depression in one construct, examining the shared and overlapping risk factors and/or consequences that symptoms of anxiety and depression pose in relation to loneliness during mid-late adolescence. This might advance research that is theoretically and empirically relevant to prevention and treatment (Epkins & Heckler, 2011).

The developmentally based interpersonal model of depression (Rudolph et al., 2008, p. 80; see Appendix A for details) suggests that gender and social cognitive factors (e.g., social self-efficacy) might moderate the extent to which relationship disturbances heighten the risk for depression (Hankin & Abela, 2005), as well as the extent to which depression influences interpersonal functioning (Coyne et al., 1998). Empirical findings indicate that the association between loneliness and symptoms of anxiety and depression could be stronger for girls than for boys (Chang, 2018; Rudolph et al., 2008). Girls, compared to boys, are increasingly likely to define themselves based on their interrelationships, be more dependent on social connections, prioritise goals in line with their relationships, and worry about others’ evaluations of them (Rose & Rudolph, 2006). Furthermore, girls tend to increase in depressive symptoms due to their reactions to interpersonal stress, such as rumination (Rose & Rudolph, 2006).

People with a cognitive vulnerability are increasingly likely to experience a pattern of information processing that is inherently negatively biased, facilitating a descent into depression (Abela & Hankin, 2008). Social self-efficacy (i.e., social capability beliefs; Bandura, 1977, 1997) may function as a vulnerability/protective factor in the association between loneliness and symptoms of anxiety and depression. It is assumed that socially efficacious people pursue and maintain social connections that help the individual be in control of difficult situations and dampen the effect of negative life events. In contrast, people with low social self-efficacy have increased vulnerability to depression through social isolation (Bandura, 1994). Indeed, social self-efficacy is negatively associated with loneliness (Hermann & Betz, 2006; Tsai et al., 2017; Wei et al., 2005) and symptoms of anxiety and depression (Hermann & Betz, 2004; Kristensen et al., 2021; McFarlane et al., 1995) in adolescence and young adulthood.

Loneliness was measured using a slightly modified six-item short form of the UCLA loneliness scale, developed for population-based studies in Western Norway (Kraft & Loeb, 1997; Mittelmark et al., 2004). An example indicator is ‘I feel lonely even when I am around other people’. The participants rated their answers on a scale ranging from 1 ‘Not at all true’ to 5 ‘Very true’. The adapted short form has shown adequate reliability, with Cronbach’s alpha values above 0.77 (Mittelmark et al., 2004).

To measure anxiety and depressive symptoms, a short Norwegian version of the Symptom Check List-90-R (SCL-5; Tambs & Moum, 1993) was used, with items from the anxiety and depression subscales. The participants were asked to rate how bothered or distressed they had felt during the last 14 days on a scale ranging from 1 ‘not at all’ to 4 ‘very much’. Indicator examples of anxiety and depression are ‘nervousness or shakiness inside’ and ‘feeling hopeless about the future’, respectively. Studies have found the Cronbach’s reliability of the SCL-5 to be higher than 0.83 (Gjerde et al., 2011; Skrove et al., 2013; Strand et al., 2003; Tambs & Moum, 1993). Of note, this scale is not a clinical assessment of or diagnostic tool for anxiety and depression but instead a measure of symptoms of mixed anxiety and depression (Siqveland et al., 2016). The cut-off value to best predict the presence of mental disorders and/or belonging to a high-risk group is 2.0 for the SCL-5 (Strand et al., 2003).

We used the social subscale from the Self-efficacy Questionnaire for Children (SEQ-C: Muris, 2001) to measure social self-efficacy. The scale was adapted to fit the adolescent age group, that is, wordings like ‘children’ were replaced with ‘peers’ and so on. The social SEC-Q consists of seven indicators that participants rated on a scale ranging from 1 ‘not at all’ to 5 ‘very well’. A sample item is ‘How well can you become friends with peers?’. Previous research has found Cronbach’s alpha values above 0.81 in adolescent samples (Minter & Pritzker, 2015; Muris, 2001, 2002). Because social self-efficacy was tested as a moderator, we created a dummy variable by doing a median split on the personal mean level of social self-efficacy across all measurement waves. Thus, adolescents were, on average, either in the low (coded as 0) or high (coded as 1) social self-efficacy group.

Gender was coded as 0 (boys) and 1 (girls).

We included several time-invariant covariates in the model. The participant’s socioeconomic position was assessed by a question on how well off economically they perceived their family to be (Iversen & Holsen, 2008). Individuals rated their perceived family wealth on a scale ranging from 1 (not at all well off) to 5 (very well off). Baseline socioeconomic position was dummy coded as 0 (low) and 1 (high) by a median split. Ethnicity was coded as 0 (Norwegian-born) and 1 (non-ethnic Norwegian). The two dummy variables for intervention conditions were coded as 0 (control group) and 1 (intervention condition).

Initially, we investigated the omega reliability and longitudinal measurement invariance of the study’s constructs. Next, we followed the procedure by Snedecor and Cochran (1980) to investigate the difference in correlation coefficients between subgroups across time. We used SPSS version 25 for data cleaning and correlation analysis and Mplus version 8 (Muthén & Muthén, 1998‐2017) for structural equation modelling. We used the following criteria to evaluate if the structural equation models achieved acceptable fit: CFI (> 0.90), RMSEA (< 0.08), and SRMR (< 0.08) (Hu & Bentler, 1999). When performing longitudinal measurement invariance testing, we used the effects-coding approach by (Little et al., 2006). Thus, the latent factors’ means and variances were constrained to 0.0 and 1.0, respectively. We specified configural models for the loneliness and symptoms of anxiety and depression constructs and applied factor loading constraints to establish weak (metric) measurement invariance across time and the moderation subgroups. Acceptable differences between the nested models and the comparison models were evaluated using the following criteria: ΔCFI < 0.010, ΔRMSEA < 0.015, and ΔSRMR < 0.030 (Chen, 2007). The invariance constraints were retained for further modelling. For space considerations, measurement invariance results are presented in the appendices (see Appendix B for details).

For the main analysis, we specified a random intercept cross-lagged panel model (RI-CLPM) between loneliness and symptoms of anxiety and depression using maximum likelihood estimation. This was achieved by creating corresponding latent factors for each first-order latent factor at all time points (four variables in each construct), with factor loading constrained to 1 and a random intercept for each construct, regressed by all first-order latent factors, with factor loadings constrained to 1 (Hamaker, 2018; Hamaker et al., 2015; Mulder & Hamaker, 2021). The variance of the first-order latent variables was constrained to 0.0 to ensure all variance is captured by the between-person variables (intercepts) and within-person variables. Next, we included gender, social self-efficacy, socioeconomic position, ethnicity, and intervention conditions as time-invariant covariates regressed by the random intercepts. Lastly, we investigated whether the correlation and regression coefficients in the RI-CLPM were time-invariant by constraining the within-person correlation coefficients and autoregressive- and cross-lagged regression coefficients to be equal over time and comparing the model fit of the nested model to the comparison model. A significant deterioration of model fit would indicate that the effects were not invariant over time and the constraints would be removed for further modelling.

Next, we tested if gender or social self-efficacy moderated the between- and within-person associations between loneliness and symptoms of anxiety and depression in the RI-CLPM. Preliminarily, we investigated RI-CLPMs of all groups without constraints and with time-invariant correlation and regression coefficients. If chi-square did not significantly deteriorate, we would keep the constraints in the moderation analyses. Lastly, we performed two multi-group analyses (one for gender and one for social self-efficacy) on the RI-CLPM with 1000 bootstrap replications and compared the parameters across groups using model constraints. We included gender as a time-invariant covariate in the social self-efficacy moderation model and social self-efficacy in the gender moderation model.

When comparing unconstrained to constrained models within each gender, the model fit did not significantly deteriorate for either girls or boys (p > 0.05). Thus, we tested the difference between six parameters between boys and girls in the RI-CLPM of loneliness and symptoms of anxiety and depression: the time-invariant correlations, autoregressive and cross-lagged regression coefficients at the within-person level and the correlation between the random intercepts. The gender multi-group RI-CLPM of loneliness and symptoms of anxiety and depression achieved satisfactory fit: χ² = 4065.90, df = 2163, p < 0.001, RMSEA = 0.04, 95% CI [0.04, 0.04], CFI = 0.92, SRMR = 0.06. The model results are presented in Fig. 3, and further details are provided in Appendix D. We found that gender differed on two parameters. One, the concurrent state-like associations between loneliness and symptoms of anxiety and depression were significantly stronger for girls across all measurement waves compared to boys (r_diff = 0.07, p < 0.05). Two, the within-person effects from symptoms of anxiety and depression to later loneliness throughout the study were significantly higher for girls compared to boys (B_diff = 0.30, p < 0.001).

The model fit did not significantly change when comparing nested, fully time-invariant models to completely unconstrained models within the low and high social self-efficacy groups (p > 0.05). As such, we tested the difference between the same six parameters as in the gender moderation model: the time-invariant correlations, autoregressive and cross-lagged regression coefficients at the within-person level and the correlation between the random intercepts. The results from the social self-efficacy moderation analysis on the loneliness and symptoms of anxiety and depression RI-CLPM produced acceptable model fit: χ² = 4002.76, df = 2163, p < 0.001, RMSEA = 0.04, 95% CI [0.04, 0.04], CFI = 0.91, SRMR = 0.07. The standardised results from the model are presented in Fig. 4, and unstandardised and standardised estimates, standard errors, and bootstrapped 95% confidence intervals are presented in Appendix E. Even though the low and high social self-efficacy groups have somewhat different coefficients, these differences were not significant (p > 0.05). This indicates that high social self-efficacy does not act as a protective factor in the within- and between-person association between loneliness and symptoms of anxiety and depression.

In support of hypothesis 1, our results indicate that adolescents who, in general, experience symptoms of anxiety and depression are highly likely to also feel lonely during three years in the mid-late adolescent period. In other words, the trait-like components of anxiety and depression symptoms and loneliness were strongly related. This is largely in line with previous research, which has found positive associations between loneliness, anxiety, and depression (Beutel et al., 2017; Erzen & Çikrikci, 2018; Park et al., 2020). The trait-like association between loneliness and symptoms of anxiety and depression might, in part, be related to the conceptualisation of the constructs. For example, Peplau and Perlman (1982) illustrate that loneliness taps into feelings that are central to anxiety and depression, such as distress, sadness, lacking care, psychological discomfort, and boredom. Individuals who feel lonely on a trait level are more likely to be less socially competent, believe their loneliness is due to ever-lasting personal qualities, and struggle to overcome social shortages (Perlman & Peplau, 1998). Adolescents with these perceptions and beliefs are arguably expected also to experience feelings of sadness, apprehension, worry, nervousness, and hopelessness simultaneously. This is further supported by the strong state-like associations between loneliness and symptoms of anxiety and depression across time. Consistently in mid-late adolescence, individuals that were lonelier than usual on one occasion, were increasingly likely to experience higher than average symptoms of anxiety and depression at the same time point.

The positive carry-over stability effects in loneliness might be related to the major social transitions that take place during mid-late adolescence. First, during adolescence, peer relationships become increasingly important, while parental influence diminishes (Larson & Richards, 1991; Prinstein & Dodge, 2008), causing several social expectancies and alterations that can be challenging (or exceedingly easy) for some individuals. Second, most people begin an upper secondary school education at age 15–16, which is a major shift in adolescents’ social lives (Eccles & Roeser, 2009, 2011; Wigfield et al., 1991). Third, during adolescence, individuals experience sexual maturation and an increased interest in pursuing romantic relationships with others (Collins et al., 2009), which could both alter individuals’ perception of what being alone means and bring about several time periods of unusual relationship satisfaction or dissatisfaction.

We found positive carry-over stability effects within symptoms of anxiety and depression, implying that adolescents who experienced an unusual level of anxiety and depression at one time point likely experienced the same unusual level one year later. This finding has been discussed in light of helplessness-hopelessness theory (Kristensen et al., 2021), which states that periods of increased symptoms of anxiety and depression are likely to be followed by the same deviations of anxiety and depressive symptoms due to a vicious cycle of exacerbating symptoms over time (Alloy et al., 1990).

In contradiction to hypothesis 2 and previous research (see Cacioppo & Cacioppo, 2018 for an overview; Lim et al., 2016; Wei et al., 2005), our results indicate that feelings of unusually high or low loneliness did not predict unexpectedly high or low symptoms of later anxiety and depression. Because previous research has largely failed to separate the within-person effects from between-person effects, the assumption that loneliness increases subsequent mental health issues might be overestimated (Hamaker et al., 2015). Thus, although reducing feelings of being socially isolated is an important goal in itself, our findings question whether modifying unusual perceptions of social isolation would improve unexpected levels of adolescent anxiety and depressive symptoms. Interventions aimed at reducing loneliness have not separated implementation strategies between persistent (trait) and transient (state) loneliness (Eccles & Qualter, 2021), which could be important in future endeavours aimed at decreasing unusual loneliness.

We found that higher-than-normal symptoms of anxiety and depression were predictive of unusually high feelings of subsequent loneliness throughout the study. This result could imply that initiatives aimed at reducing anxiety and depressive symptoms in adolescence (see e.g., Das et al., 2016 for an overview) may decrease loneliness as a result. However, it is also important to identify adolescents who experience symptoms of anxiety and depression that are out of the ordinary (i.e., deviating from their personal norm). Significant adults that are close to adolescents, such as parents (Logan & King, 2001), teachers (Rothì et al., 2008), school healthcare professionals (Levinson et al., 2019), and guidance counsellors (Collins, 2014) or school administrators (Green et al., 2013) need resources and competence to identify and deal with negative fluctuations in young people’s mental health to avoid unusual escalations of loneliness.