Study design, setting and participant recruitment

Participant recruitment for the UK COVID-MH study was conducted by Taylor McKenzie, a social research company. A non-probability quota sample of adults (aged 18 years or older) from across the UK was recruited in March 2020. A quota sampling methodology was employed (see O'Connor et al^{Reference O'Connor, Wetherall, Cleare, McClelland, Melson and Niedzwiedz3} for more detail) to recruit a stratified sample during the early phase of lockdown, with quotas based on age, gender, SEG and UK region (12 regions). This study had a longitudinal design that included at least six follow-ups. Because of COVID-19 restrictions, data were collected online in all waves.

At wave 1 (31 March to 9 April 2020), participants of an online UK panel called Panelbase.net were invited by email to participate in an online survey tracking the health and well-being of people in the UK. Eligibility for the study was assessed with demographic questions based on the quotas. Eligible participants then provided informed consent online, and completed a range of psychological and social measures. Participants completed two further waves within the following 6 weeks, wave 2 (10–27 April 2020) and wave 3 (28 April to 11 May 2020), reported previously.^{Reference O'Connor, Wetherall, Cleare, McClelland, Melson and Niedzwiedz3} The current paper primarily reports analysis from wave 4 (27 May to 15 June 2020), wave 5 (17 July to 17 August 2020), wave 6 (1 October to 4 November 2020) and wave 7 (4 February to 2 March 2021). Findings reported here are for depressive symptoms, anxiety symptoms, suicidal ideation, self-harm history, defeat, entrapment, loneliness and well-being.

Those who took part in wave 1 were then invited by email to take part in the follow-up waves, with the exception of waves 4 and 5. Because of a technical error, only those who had completed the previous wave (wave 3) were asked to take part in waves 4 and 5, but all respondents were once again invited to complete waves 6 and 7; 15.4% (n = 473) of the wave 1 sample did not take part in wave 3, and 24.3% (n = 115) of that group completed wave 6 when invited. Those who did not complete wave 6 when re-invited (n = 358; 75.7% of those who did not complete wave 3) were younger (χ ² = 7.44, P = 0.024, φ = 0.13) and reported higher anxiety symptoms (χ ² = 5.57, P = 0.018, φ = −0.11), although there were no differences between gender, SEG, pre-existing mental health condition, depressive symptoms or suicidal ideation. Therefore, as the effect sizes of these differences were small, we have retained these participants in the analyses. Respondents were included in the analyses if they had taken part in at least one wave of data collection from wave 4 to wave 7. The sample characteristics are presented in Table 1 (n = 2691).

Table 1 Demographic characteristics of the waves 4–7 sample (n = 2691) compared with the wave 1 sample (n = 3077)

a. Wave 1, n = 3073; waves 4–7 n = 2689.

b. Categories A, B and C1 indicate high socioeconomic group; categories C2, D and E indicate low socioeconomic group.

Figure 1 illustrates the changes in public health restrictions in the UK occurring during each wave. The first three waves occurred within the first 6 weeks of the UK lockdown, and the subsequent four waves were roughly every 2–3 months, with the interval between waves increasing over time. During waves 1–3, the initial lockdown was underway. Wave 4 (May/June 2020) coincided with the easing of some restrictions across the UK (e.g. England on 13 May). During wave 5 (July/August 2020), restrictions had mostly been relaxed across the UK. Wave 6 (October 2020) coincided with the increasing of restrictions again across the UK, with cases of COVID-19 on the rise. During wave 7 (February 2021), there was a UK-wide lockdown. To ensure that it is clear when each wave occurred, they will be referred to by the main month and year when they occurred (e.g. wave 4, May/June 2020), where appropriate.

Fig. 1 Overview of the waves of the UK COVID-19 Mental Health and Wellbeing study and key events during the COVID-19 pandemic in the UK.

In the interests of completeness, data for all waves are included in Table 2 and Figs. 2 and 3.

Table 2 Changes in primary outcome variables over waves 1–7 of the UK COVID-19 Mental Health and Wellbeing study, with odds ratios and 95% confidence intervals

Data are provided for all waves but analysis is for waves 4–7 only; analysis for waves 1–3 can be found in O'Connor et al.^{Reference O'Connor, Wetherall, Cleare, McClelland, Melson and Niedzwiedz3} PHQ-9, nine-item Patient Health Questionnaire; GAD-7, seven-item Generalised Anxiety Disorder assessment.

a. Reference group wave 4.

b. Reference group wave 5.

c. Reference group wave 6.

Fig. 2 Percentages and 95% confidence intervals (error bars) of suicidal ideation, depressive symptoms and anxiety symptoms (percentage of participants scoring ≥10) over waves 1–7 of the UK COVID-19 Mental Health and Wellbeing study (N = 3077). GAD-7, seven-item Generalised Anxiety Disorder assessment; PHQ-9, nine-item Patient Health Questionnaire.

Fig. 3 Means and 95% confidence intervals of defeat, entrapment and loneliness scores over waves 1–7 of the UK COVID-19 Mental Health and Wellbeing study (N = 3077).

The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008. All procedures involving human participants were approved by the University of Glasgow's Medical, Veterinary and Life Sciences Ethics Committee (approval number: 200190146). The study was preregistered at AsPredicted.org (identifier: 41910). Study participants were compensated approximately £1.50 for the completion of each survey, and were entered into prize draws. Details of mental health support organisations were made available to participants.

Measures

All suicide-related items were assessed within the time frame of the past week. Suicidal ideation was measured with ‘How often have you thought about taking your life in the past week? (‘one day’, ‘several days’, ‘more than half the days’, ‘nearly every day’, ‘never’, ‘I would rather not answer’)’; ‘one day’ to ‘nearly every day’ were coded as yes, and ‘never’ was coded as no. Suicide attempt and self-harm were measured with ‘In the past week, have you made an attempt to take your life, e.g. by taking an overdose of tablets or in some other way?’ and ‘In the past week, have you ever deliberately harmed yourself in any way but not with the intention of killing yourself?’ (‘yes’, ‘no, ‘I would rather not answer’).

Depressive symptoms were assessed with the nine-item Patient Health Questionnaire (PHQ-9^{Reference Löwe, Kroenke, Herzog and Gräfe9}), and anxiety symptoms were assessed with the seven-item Generalised Anxiety Disorder (GAD-7) scale.^{Reference Spitzer, Kroenke, Williams and Lowe10} Both tools ask how often symptoms have bothered the respondents in the previous 2 weeks. A score of ≥10 on both measures is suggested to indicate moderate levels (or more) of depression and anxiety, and this cut-off is used in the current study.^{Reference Spitzer, Kroenke, Williams and Lowe10,Reference Kroenke and Spitzer11}

Feelings of defeat (perceived failure and loss of rank) were assessed with the four-item defeat subscale of the Short Defeat and Entrapment Scale by Griffiths et al.^{Reference Griffiths, Wood, Maltby, Taylor, Panagioti and Tai12} Perceptions of entrapment (feeling trapped by thoughts and feelings or circumstances) were measured with the Entrapment Scale Short-Form.^{Reference De Beurs, Cleare, Wetherall, Eschle-Byrne, Ferguson and O'Connor13} Mental well-being was assessed with the seven-item Short Warwick–Edinburgh Mental Well-Being Scale.^{Reference Fat, Scholes, Boniface, Mindell and Stewart-Brown14} Loneliness was measured with the Three-Item Loneliness Scale, developed from the Revised UCLA Loneliness Scale.^{Reference Hughes, Waite, Hawkley and Cacioppo15} Percentages are reported for all binary outcome variables (suicidal ideation, depressive symptoms and anxiety symptoms) and mean scores are reported for all continuous outcome variables (defeat, entrapment, well-being and loneliness).

SEG was assessed according to the National Readership Survey social grade:¹⁶ high (categories A, B and C1) versus low (categories C2, D and E). Participants were asked whether they had any long-standing physical or mental impairment, illness or disability, and this was coded to create a dichotomous variable for presence or absence of a pre-existing mental health condition. Specifically, participants were asked to select their mental or physical impairment from a list of options, which included mental health conditions, neurodivergent disorders and alcohol and drug problems.

Statistical analysis

Analysis of data from waves 4–7 was consistent with data from the previous waves 1–3 where possible, to aid comparison. All analyses were conducted with Stata MP 16 (Windows), using an imputed data-set of the 2691 participants who completed at least one survey from waves 4 to 7.

We used multiple imputation to generate 50 data-sets for each outcome variable. Multiple imputation is a statistical technique whereby an imputation represents one set of plausible values for missing data, and the imputation model for deriving these imputations includes predictors relevant to the missing-data mechanism. In analysis, the results from the multiple imputations are pooled into a single result.¹⁷ The binary variables (suicidal ideation, depressive symptoms, anxiety symptoms) were imputed with logistic regression (mi impute logit), and continuous variables (defeat, entrapment, loneliness, well-being) were imputed with linear regression (mi impute regress). The imputation model included age, gender, SEG, history of mental health condition and the region of residence in the UK.

Participants (n = 2691) were included in the main analysis if they completed at least one survey from waves 4 to 7; therefore, there was missing data at each wave, and this varied across waves. Across these four waves, when totalled up, each variable had approximately 16.1% (n = 1729/10 764) missing cases (percentage missing at each wave: n = 307, 12.9% at wave 4; n = 547, 25.5% at wave 5; n = 4.8, 15.2% at wave 6; n = 467, 17.4% at wave 7). Suicidal ideation had 2.7% (n = 299) more missing cases (18.8%, n = 2028/10 764), as an option for ‘would rather not answer’ was included with this item (percentage answering ‘would rather not answer’ at each wave: 2.4% at wave 4, 2.4% at wave 5, 3.2% at wave 6, 3.0% at wave 7). Full details of missing data (‘would rather not answer’) and the frequencies of respondents who reported suicidal ideation, suicide attempt and self-harm are included in the supplementary materials (Supplementary Table 1 available at https://doi.org/10.1192/bjo.2022.58). For suicide attempts and self-harm, often there was more missingness than the prevalence of that outcome, but no inferential statistics are applied to these outcomes.

Multiple imputation generalised estimating equations (GEE) models were then constructed to test the changes in the variables across waves 4–7 for the whole sample. This approach is suitable for multilevel longitudinal panel data.^{Reference Gardiner, Luo and Roman18} In a sensitivity check, all analyses were conducted with and without multiple imputation, and a similar pattern was found for both, with the multiple imputation analysis more conservative and reported here. GEE models apply a multilevel method,. Binomial logit modelling was conducted for each of the binary outcome variables (suicidal ideation, depressive symptoms and anxiety symptoms), and linear Gaussian identity modelling was conducted for the continuous outcome variables (defeat, entrapment, loneliness and well-being).

To test for subgroup differences, additional GEE models were applied to the outcome variables, investigating differences in age (18–29, 30–59, ≥60 years), gender (men, women), SEG (higher, lower) and a pre-existing mental health problem (no, yes). Unlike the waves 1–3 analysis, ethnicity was not included because of reduced numbers. Interactions between the subgroups and changes in the outcomes over the waves were then tested, with significant findings only reported in the results.

The region of the UK in which a participant lived (South England, Midlands, North England, Scotland, Wales and Northern Ireland) was controlled for in each analysis. The temporal covariation was modelled with an unstructured correlation matrix, as the pattern of associations was neither fully exchangeable nor had a first-order autoregressive structure. Additionally, because of the large number of analyses, a false discovery rate procedure^{Reference Benjamini and Hochberg19} was applied to the between, within and interaction P-values for all analyses. The false discovery rate procedure places all P-values in ascending order and assigns ranks (e.g. smallest is ranked 1), then the Benjamini–Hochberg critical value for each P-value is calculated with the formula (i/m)Q (with i being the individual P-value's rank, m being the total number of tests and Q being the false discovery rate (0.05)), and the largest P-value that is smaller than its critical value (P < (i/m)Q) is significant, along with all smaller P-values. This method adjusts for type 1 errors in null hypothesis testing when conducting multiple comparisons, as it controls for the expected proportion of ‘discoveries’ that are false.^{Reference Verhoeven, Simonsen and McIntyre20}