We conducted a secondary data analysis by drawing on data from the Understanding Society UK Household Longitudinal Study, a nationally representative, prospective cohort study in the United Kingdom (UK) with about 40,000 households assessed since 2009 [26, 27]. Since April 2020, all members of the main Understanding Society samples aged 16 years and older in April 2020 were invited to participate, if they participated in at least one of the last two waves of data collection investigating various questions pertaining to individuals’ life’s during COVID-19 (Understanding Society: COVID-19 Study). Until now, data for eight consecutive COVID-19 waves were collected in April, May, June, July, September, November 2020, as well as January 2021, and March 2021. All participants of the Understanding Society: COVID-19 Study subsample were assessed using online and telephone surveys.

In line with our research questions, we focused on the analysis of the most recent measurement occasion before the onset of the pandemic (2019/2020), the first time-point during the pandemic (April 2020, wave 1, n = 17,761), and the subsequent peaks, i.e., wave 6 (November 2020, n = 12,035) and wave 7 (January 2021, n = 11,968), as defined by COVID-19 incidence and stringency index (see supplement Fig. S18–S20 for the course of the incidence of COVID-19 cases, for confirmed deaths, and for the degree of the stringency of lockdown measures). In total, 17,761 participants participated in the first wave during the COVID-19 pandemic of the Understanding Society: COVID-19 and were the target of the present analyses. Among the 34,318 participants who participated in the last assessment of the pre-COVID-19 survey (i.e. before the COVID-19 pandemic), the response to the first wave during COVID-19 was ~51,75%. This is consistent with the individual-level online response levels of the main study, which was 50–55%, which increased to 85–90% through face-to-face surveys [28]. Full demographic characteristics of our sample including the sample sizes for each sampling time are depicted in Table 1. Ethical approval was granted by the University of Essex Ethics Committee (ETH1920-1271). The data are openly available to researchers via the UK Data service (https://ukdataservice.ac.uk). Detailed information on the Understanding Society UK Household Longitudinal Study and Understanding Society: COVID-19 Study, including the sample structure, subsamples, and panel attrition, has been previously presented [6, 29]. All procedures and measures collected in the Understanding Society Study are described at https://www.understandingsociety.ac.uk/about/study-content. For our secondary data analysis additional ethical approval was not required. The authors assert that all procedures contributing to this work comply with the Helsinki Declaration of 1975, as revised in 2013.Table 1

Demographic characteristics for each sampling time.

Sample characteristics of the participants who dropped out of the study between April 2020 and January 2021 compared to those who stayed have been described elsewhere (see supplemental material in refs. [6, 29]). Briefly, characteristics that were associated with missingness were ethnicity (non-white had higher values of missingness than white participants), sex (males had higher levels of missingness than females), and age (older participants displayed higher levels of missingness than younger participants). Data were thus treated as missing at random (MAR) [34]. Excluding participants that did not complete all relevant questions would result in a final sample of N = 7815 complete cases. Thus, missing data were imputed. As multiple imputations [35] are not compatible with the present network analytical approach, we used one imputed dataset as recommended under such circumstances [36, 37]. To this end, we imputed the data with the MICE package in R [38] by using predictive mean matching [34]. In our imputation model, demographic characteristics that were associated with missingness, i.e. sex and ethnicity, were included as auxiliary variables. We performed a sensitivity analysis using full-cases only (see Supplemental Figs. S1 and S2). The results essentially remained the same compared to the imputed data set, increasing confidence in our results and their interpretation.

We performed all analyses with the R Version 4.0.3 [39]. The R code to reproduce the current results is openly available on the OSF (https://osf.io/cusf3/?view_only=4edd973d5da24a20bbc2895a8288e2bd). We used cross-lagged panel network models (CLPN) to analyze the unfolding networks over time [37]. We analyzed three longitudinal networks: (1) pre-COVID survey to wave 1, (2) wave 1 to wave 6, (3) wave 6 to wave 7. In a first step, we calculated the regression coefficients for the models: We computed autoregressive pathways, where a symptom at one time-point predicts itself at the next time-point while adjusting for all other symptoms. Then, we computed cross-lagged pathways, where a symptom predicts another symptom while adjusting for the autoregressive effects and all other symptoms. For the estimation of the regression coefficients, we applied the least absolute shrinkage and selection operator (LASSO), which applies a penalization to avoid estimating spurious edges [40, 41]. We calculated our directional CLPNs from pre-measurement to first incidence peak and incidence peak to incidence peak using the glmnet package [37, 42]. The qgraph package [43] was used for the visualization of the networks using an average layout for the three different networks over time.

To quantify the centrality of symptoms in our directed CLPNs, we examined two centrality indices: in- and out-expected-influence. These indices are calculated by all outgoing symptom associations one respective item has with all other items in the following wave. This centrality measure is coined out-expected-influence (the degree to which a symptom predicts other symptoms). The other centrality metric is quantified by the ingoing symptom associations one item has with all other items and is labeled in-expected-influence (the degree to which a symptom is predicted by other symptoms). The edge weight difference test and centrality difference tests were used to determine possible significant differences in edges and centrality indices, respectively [44].

The GHQ items with the highest mean values (i.e., >1 in each sampling time) were able to concentrate, playing a useful part, capable of making decisions, enjoy normal activities, can face up to problems, and feeling reasonably happy (see Table 2). GHQ values increased from the pre-COVID-19 assessment compared to the first wave during COVID-19 in April 2020, t(15,464) = −11.22, p < .001, Hedges’s g = 0.185, indicating a deterioration of mental health. Comparing incidence peak to incidence peak, mental health problems in November 2020 (wave 6) were significantly higher than in April 2020 (wave 1), t(10,892) = −6.91, p < .001, Hedges’s g = 0.027 but January 2021 (wave 7) was not significantly different from November 2020 (wave 6), t(9,738) = −1.90, p = 0.058. Detailed descriptive statistics can be found in Table 2.Table 2

Mean and standard derivation of each GHQ-12 and the loneliness item, sum scores for the GHQ-12 for each sampling time and internal consistencies (N = 7815).

The bootstrapped confidence intervals around the edge weights are shown in Figs. S3–S5 (see supplemental material). The bootstrapped confidence intervals were small to moderate, indicating good accuracy of our networks. In addition, the case-drop bootstrapping results indicate a high stability of the centrality indices (Figs. S6– S8 in the supplemental material).

The edge lists of all networks are presented in the supplemental materials edg1. The number of non-zero edges was consistent across networks [range: 125 (pre-COVID to first incidence peak) – 130 (second incidence peak to third incidence peak)]. The correlation of the edge lists between the networks was moderate to strong (range: r = 0.75 – r = 0.83). The overall correlation of the out-expected-influence (r = 0.89) and the overall correlation of the in-expected-influence (r = 0.84) between networks was high.

All three CLPN networks are depicted in Fig. 1: The symptoms loneliness, feeling unhappy and depressed, thinking of self as worthless, losing confidence, could not overcome difficulties, under stress, and lost much sleep formed a consistent cluster with strong connections over time. The symptoms can face up to problems, capable of making decisions, feeling reasonably happy, able to concentrate, and enjoy normal activities were less consistently connected across time.Fig. 1

The relationship of the symptoms is indicated by the arrow’s color (green = positive, red = negative), the strength of the relationship is indicated by the arrow’s thickness (thicker = stronger). In these networks, autoregressive effects are excluded. Threshold was set to 0.05, excluding all relationships < 0.05. Final N = 17,761.

Table 3 depicts the five strongest edges for all three networks. Overall, the strongest edge connection in the three networks was the connection thinking of self as worthless → losing confidence (range: β = 0.15 – 0.20). In all three networks, the edge loneliness → feeling unhappy and depressed displayed the second strongest edge connection (range: β = 0.12 – 0.17). In the first network, the connection loneliness → losing confidence was the third strongest one, while it was the fifth strongest in the second network, and not as strong in the third CLPN network (range: β = 0.08 – 0.15). The connection loneliness → thinking of self as worthless was the fourth strongest in the first network, the third strongest in the second and the fifth strongest in the third CLPN network (range: β = 0.10 – 0.14). Moreover, the connection thinking of self as worthless → could not overcome difficulties was the fifth strongest in the first, the seventh strongest in the second and the third strongest connection in the third network (range: β = 0.08 – 0.13). Finally, the connection losing confidence → thinking of self as worthless was the fourth strongest in the second network, but not as strong in the first and third network (range: β = 0.06 – 0.10). The edge weights difference tests for each of the 3 networks indicated significantly stronger connections between the aforementioned edges compared to most other edges (see Figs. S9–S11 in the supplemental material).Table 3

Table depicting the 5 strongest edges for all three networks.

The standardized centrality parameters revealed consistent patterns over time and can be seen in Fig. 2: We found the strongest in-expected-influence for feeling unhappy and depressed, followed by could not overcome difficulties and losing confidence, showing a significantly higher in-expected influence than many other symptoms across sampling times. Lower in-expected-influence values were found for the symptoms capable of making decisions, enjoy normal activities and can face up to problems in all networks and loneliness in the CLPN network second incidence peak to third incidence peak (see supplemental material Figs. S12–S14 for in-expected-influence difference tests). The strongest out-expected-influence was found for loneliness and thinking of self as worthless. The lowest out-expected-influence was found for feeling reasonably happy, can face up to problems, playing a useful part (see supplemental material Figs. S15–S17 for out-expected influence difference tests).

We found consistent patterns of symptom associations over time as thinking of self as worthless → losing confidence and loneliness → feeling unhappy and depressed displayed the strongest connections in all networks. Similarly, the standardized centrality parameters indicated that feeling unhappy and depressed and could not overcome difficulties and losing confidence had the strongest in-expected-influence, and loneliness and thinking of self as worthless had the strongest out-expected-influence in all networks.

The connections and the central symptoms with their connections can be used to investigate, both through experimental and clinical studies, whether these are indeed causal for symptom escalation under stress. Subsequently, it is to be examined whether this symptom escalation, i.e., this symptom spread can be prevented by addressing these nodes [14, 54, 55].

Together with previous findings on loneliness [48, 56, 57], our study reemphasizes the importance of loneliness for the escalation of other mental health symptoms, thus constituting a key target for intervention [58]. Prevention and intervention programs could be implemented on societal level (e.g., in educational institutions) as regular courses on stress reduction, emotional regulation, empathy, and self-compassion after proving their effectiveness. On an individual level, home visits and daily contact programs were proposed to target loneliness [58]. Hence, at any time in the pandemic, especially at the relevant times of onset and incidence peaks, it is advisable and important to address the social connectedness and potential social isolation of individuals and to implement appropriate programs to counteract the impairment of mental health at an early stage. For instance, when implementing contact restrictions, policy makers need to consider how to adequately enable social support groups or small in-person group activities to stop the manifestation of loneliness. For example, single-session growth mindset interventions for adolescent anxiety and depression could be helpful to reduce mental health deterioration [59].

The results of our study revealed that, prior to the COVID-19 pandemic, GHQ-12 mean levels were slightly above the clinical cut-off of 11 as established by Goldberg et al. [31]. In the following three waves during the pandemic, the threshold was likewise exceeded with an increase of one point from the pre-COVID-19 survey (see [60] for a more detailed analysis on clinical change in this sample). Despite the small effect sizes of the observed changes, it is important to note that this is a population-based study and small effects can have a significant impact on a large population [61]. The present study provides insights into the dynamics of symptom levels, and sheds light on the activation of symptom cascades that may have been involved in such population level shifts. It appears that the symptom cascades and the findings as a whole have a high relevance in our sample, and we found that this relevance is not only for the subjects who were already in the pathological range (measured with the GHQ-12), but also for the subjects with subclinical mental health issues.

In addition, it is important for prevention and intervention to disentangle how to address the needs for individuals who were already lonely before the pandemic compared to persons who became lonely in light of the restrictions [56]. From our findings it can be deduced that loneliness does not have to be addressed only during the pandemic, but already before, so that one is better prepared for such stressors [58]. As loneliness was an important variable during all stages of the pandemic, it needs to be investigated whether different groups of individuals felt lonely during the pandemic. As several studies found younger individuals to be at a higher risk (e.g. [57, 62]), prevention and intervention programs should address these groups in particular. Li & Wang [63] analyzed the first wave of the Understanding Society dataset with weighted univariate analyses. They found significantly higher scores in GHQ-12 and for loneliness among women, younger people, participants without a partner, and unemployed individuals. These findings suggest that these subgroups represent populations at higher risk, warranting the implementation of specific intervention and preventive measures tailored to these populations. Intervention programs that take place online are already available for this purpose (e.g. [64, 65]), reducing loneliness and ameliorating feelings of belonging (for an extended review see ref. [66]). Single-session interventions are also available, and first studies found evidence of reducing hopelessness and increasing agency, among other outcomes [67]. There are also targeted programs aimed at reducing loneliness, such as the digital intervention “Happify”, which has been shown to have a significant positive impact on feelings of loneliness [68]. Also, group interventions conducted via Zoom have shown positive significant effects on loneliness and depression [69]. Similarly, interventions can be implemented that can reduce worthlessness and enhance self-worth [70]. Overall, there are different opportunities to implement appropriate specific interventions to counteract the deterioration of symptom cascades at an early stage. These could be implemented as early interventions or, at the onset of a pandemic, in terms of indicated prevention as a preventive measure for potential at-risk groups.

Targeting specific reasons of loss of self-worth may be advisable for intervention efforts, as loss of self-worth may be attributable to a variety of reasons such as inability to engage in hobbies or being socially less active [46, 47]. The question of appropriate interventions here is a particularly individual one that depends on the cause of the loss of self-worth. In the case of the loss of employment, support for existential issues would be more likely helpful. The inability to engage in hobbies might be more likely to benefit from motivation to resume previous hobbies or to take up new ones. Reduced social activity with its negative impact on self-worth can be addressed closely in conjunction with the interventions mentioned above for loneliness.

A major advantage of our study is the use of a representative, large dataset that includes both pre-pandemic surveys and waves during the pandemic. It is particularly advantageous that these waves took place during pandemic-relevant assessment intervals, so that, based on the stringency index, particularly relevant time points could be examined.

The network analytic approach allowed us to perform an analysis with a high degree of resolution, as we can describe dynamic changes over time and differentiate these changes in incoming and outgoing connections separately (influence of vs. on a symptom). This is important, as the common cause perspective of mean value differences of GHQ-12 suggests targeting psychological distress holistically, which can be complicated, while we provide specific target points by identifying constitutive elements of psychological distress.

One limitation results from the necessary use of imputation. Missingness was a concern in the present study. Our conclusions hold only under the missing at random (MAR) assumptions, which is based on our own missingness analyses and previous studies on this dataset that identified important variables associated with missingness (e.g. [6]). Although our results point to plausible intervention targets, within- and between person variance remains entangled in the present study [37], and intervention studies are needed first to showcase the effects of targeting central symptoms (e.g. refs. [14, 54]). The results obtained here relate specifically to pandemic events and conditions in the UK. Because the pandemic event was a country-specific and individual event for which country-specific measures were taken, our results cannot be reliably extrapolated to other countries or populations. As the GHQ-12 is primarily tailored to depressive symptoms, it would be useful to broaden the range of symptoms in further studies by including anxiety symptoms or more fine-grained loneliness measures to unravel specific dynamics among these symptoms. In the Understanding Society study, loneliness was assessed with a single item, limiting the construct validity of loneliness. While this item presents an indicator of the subjective feeling of loneliness, important nuances underlying loneliness and their role in an evolving network remains masked. We cannot discern differential effects of important aspects of social loneliness like a lack of social contact and emotional loneliness or a lack of emotional closeness and intimacy within a network perspective. These other facets of loneliness should be considered in future studies, for instance by the means of the UCLA 3-item loneliness scale as an approximate measure or other more extensive scales [71, 72].

Furthermore, it is important to note that the GHQ-12 is a screening tool to indicate caseness but does not possess psychiatric diagnostic-specific validity. Nevertheless, the GHQ-12 is useful for capturing important transdiagnostic symptoms commonly found across several mental disorders. This allowed us to examine important associations between these transdiagnostic symptoms and loneliness. However, the change in survey method from face-to-face surveys to online surveys or telephone surveys must be critically considered [73]. Dropout in 2020 was particularly higher in the older age groups, among people living alone and among people with a lower level of education. In addition to the reduced response rate, it cannot be ruled out that changes in the survey format induced response bias. However, a thorough psychometric analysis of the GHQ-12 in the present sample indicated longitudinal measurement invariance across time before and during COVID-19 [73].