Temporal dynamics in mental health symptoms and loneliness during the COVID-19 pandemic in a longitudinal probability sample: a network analysis

The study investigates which specific mental health symptoms were most central in driving symptom escalation during the COVID-19 pandemic, distinguishing symptoms that influence others from those that are influenced. Against the backdrop of widespread increases in psychological distress during lockdowns, prior work largely relied on composite scores, obscuring symptom-level dynamics. The authors applied a psychopathology network approach to examine temporal dynamics from pre-pandemic (2019) to the first pandemic incidence peak (April 2020) and across subsequent peaks (November 2020 and January 2021). They hypothesized, based on prior reports, that loneliness would centrally influence other symptoms over time and explored the roles of other GHQ-12 symptoms (e.g., concentration, sleep problems, decision-making, self-confidence, happiness, enjoyment of activities). The goal was to identify actionable targets for prevention and intervention by clarifying temporally central symptoms during maximal and repeated stress.

Large longitudinal studies show that approximately 40% of populations experienced increases in psychological distress during the pandemic, with peaks aligning with lockdowns and declines after easing restrictions, a pattern replicated across countries. Network-analytic research during COVID-19 suggested loneliness often emerges as a central variable linking to depression, anxiety, stress, and sleep problems, though findings are not uniform across all samples. Worthlessness has also been highlighted as influential, including acting as a bridge symptom between depression and parental stress. However, most prior work is cross-sectional and cannot disentangle whether symptoms are influential drivers or downstream recipients. Limited longitudinal or ecological momentary assessment work indicates that loneliness may be more central during lockdown than in non-lockdown periods, but prior studies often used small convenience samples and lacked pre-pandemic baselines, leaving a gap in understanding temporal symptom cascades in representative samples.

Design and data: Secondary analysis of Understanding Society, a nationally representative UK longitudinal household study. The analysis used the last pre-COVID survey (2019/2020), April 2020 (wave 1; N = 17,761), November 2020 (wave 6; N = 12,035), and January 2021 (wave 7; N = 11,968), chosen to represent maximum stress (pre-COVID to pandemic onset) and repeated stress (peak to peak). Measures: GHQ-12 (0–3 per item; higher = poorer mental health; α = 0.90–0.92) and a single-item loneliness measure (3-point scale; harmonized GSS wording adjusted for pre- vs. during-COVID). A GHQ-12 cutoff of 11 indexed probable mental distress. Handling missingness: Data assumed missing at random. Given incompatibility of multiple imputation pooling with the network approach, one imputed dataset was created using MICE with predictive mean matching, including sex and ethnicity as auxiliary variables. A complete-case sensitivity analysis yielded similar results. Analytic approach: Cross-lagged panel network models (CLPN) were estimated for three intervals: (1) pre-COVID to wave 1, (2) wave 1 to wave 6, (3) wave 6 to wave 7. For each model, symptom-level autoregressive paths and cross-lagged paths were estimated, adjusting for all other symptoms. LASSO regularization (glmnet) penalized spurious edges; visualization was done with qgraph using a common layout; a threshold of 0.05 excluded weaker relations. Centrality: In- and out-expected-influence indices quantified, respectively, how much a symptom was predicted by others and how much it predicted others at the next wave. Accuracy and stability: Bootstrapped CIs around edge weights were small-to-moderate, and case-drop bootstrapping indicated high stability of centrality. Network comparisons: Non-zero edge counts were similar across networks (125–130), edge-list correlations were moderate-to-strong (r = 0.75–0.83), and centrality correlations high (out EI r = 0.89; in EI r = 0.84). Edge- and centrality-difference tests assessed significance.

Descriptive change: GHQ-12 scores increased from pre-COVID to April 2020, t(15,464) = -11.22, p < .001, Hedges’ g = 0.185. From April to November 2020, distress increased further, t(10,892) = -6.91, p < .001, g = 0.027; January 2021 did not differ significantly from November 2020, t(9,738) = -1.90, p = 0.058. Proportions above GHQ-12 cut-off rose from 37.86% pre-COVID to 50.08% (wave 1), 53.43% (wave 6), and 53.95% (wave 7). Network structure: Networks were highly consistent over time (edge-list r = 0.75–0.83; centrality r: out EI = 0.89; in EI = 0.84). Strongest temporal edges across networks included thinking of self as worthless → losing confidence (β ≈ 0.15–0.20) and loneliness → feeling unhappy and depressed (β ≈ 0.12–0.17). Additional robust edges included loneliness → losing confidence (β ≈ 0.08–0.15), loneliness → thinking of self as worthless (β ≈ 0.10–0.14), and thinking of self as worthless → could not overcome difficulties (β ≈ 0.08–0.13). Centrality: Out-expected-influence was highest for loneliness and thinking of self as worthless, indicating strong predictive influence on other symptoms at subsequent time points. In-expected-influence was highest for feeling unhappy and depressed, could not overcome difficulties, and losing confidence, indicating these were downstream recipient symptoms.

The findings confirm that loneliness and worthlessness are temporally influential drivers in symptom networks during both the pandemic onset and subsequent peaks, supporting the hypothesis that loneliness centrally influences other symptoms. Feeling depressed and difficulties overcoming problems emerged as recipient symptoms, downstream in symptom cascades. The stability of network structure and centrality across intervals underscores robust temporal dynamics. These results align with sociometer theory, linking perceived social inclusion to self-esteem, and suggest a sense-of-belonging cluster (loneliness and worthlessness) as key levers in distress escalation. Clinically and at a policy level, preventive and early interventions to address loneliness and self-worth—particularly at lockdown onset and peaks—may interrupt symptom cascades. Despite pandemic-related restrictions, enjoyment of activities did not emerge as a key driver, implying that cognitive evaluations and coping may play larger roles than activity restrictions per se.

Using longitudinal probability data and cross-lagged panel network models, the study identifies loneliness and worthlessness as central predictors of subsequent symptom activation during COVID-19, with depression and difficulty overcoming problems functioning as downstream endpoints. These insights highlight loneliness and self-worth as promising targets for prevention and intervention to mitigate cascading mental health deterioration under prolonged stress. Future research should test causal effects experimentally and clinically, evaluate targeted single-session and scalable interventions for loneliness and self-worth, and determine how these insights can inform policy to maintain social connectedness during restrictions.

Loneliness was measured by a single item, potentially limiting precision relative to multi-item scales. GHQ-12 is a screening instrument lacking diagnostic specificity, though it captures transdiagnostic symptoms. Survey mode changes (face-to-face to online/telephone) could introduce response bias; dropout was higher among older individuals, those living alone, and those with lower education. Although data were treated as missing at random and a single imputed dataset was used (due to methodological constraints), attrition and imputation choices may affect generalizability. Nevertheless, psychometric analyses indicated longitudinal measurement invariance for GHQ-12 in this sample, and complete-case sensitivity analyses yielded similar results.