The study used a cross-sectional study design.

The study used an online survey. The recruitment and data collection period took place during the second COVID-19 wave (Jan–April 2021).

Eligible participants were health staff members currently employed by the NHS, aged 18 and over, and working within the geographic area covered by nine NHS Trusts in NWL. Potential participants were identified and recruited via trusts’ internal human resource systems, with existing dedicated emails distributing explanatory invites to staff at the respective sites and providing a link to the study website for the survey. A specific study email was set up for participants’ questions. Website information reinforced the fact that participation was entirely voluntary, and that no study data would be identifiable to anyone outside the immediate research team. The website was advertised by posters around the sites, verbally promoted during staff support sessions offered, in email signatures, via social media, Primary Care Networks and GP practice manager communication, trade union organisations, professional organisations and Royal Colleges, newsletters and magazines and ethnic minority networks. As a result of the range of distribution methods used across the participating organisations, estimating the exact number of individuals invited to participate is difficult. The organisations collectively employ around 40,000 staff who may have been eligible to participate.

The study was approved by the Imperial College Research Ethics Committee (ICREC reference number: #20IC6241) and Health Regulator Authority (IRAS reference number: 20/PR/0828). Participants provided informed online consent, including for publication, after the survey had been explained to them in an online document.

The survey was comprised of two parts (a short and long survey). (i)

Short baseline survey

This survey had two components: socio-demographic and psychological. All participants were asked to complete a 5-10 minute questionnaire and given the option of completing a supplementary longer questionnaire.

The short baseline survey included information on respondents’ contact details (e.g., name, work and personal email address, phone number), occupational information (e.g., occupational group, length of professional registration, usual place of work), demographics (e.g., age, gender, relationship status), and a few short questionnaires concerning occupational and social support available, general health, and information on direct and indirect experience of the pandemic – e.g., participant or family members becoming ill, caring responsibilities and institutional support during the pandemic. The psychological component of the survey included the 12-item General Health Questionnaire (GHQ-12), to screen for general (non-psychotic) psychiatric morbidity ( Goldberg and Williams, 1988).

Study tools included standard question sets with established scoring rules that generate validated measures of mental health morbidity and related study areas of concern: 12-item General Health Questionnaire (GHQ) ( Goldberg and Williams, 1988); 6-item Brief Resilience Scale (BRS) ( Smith et al., 2008); 7-item Generalised Anxiety Disorder (GAD-7) ( Spitzer et al., 2006); 9-item Patient Health Questionnaire (PHQ-9), measuring depression ( Kroenke et al., 2001); 14-item Warwick-Edinburgh Mental Well-being Scale (WEMWBS) ( Tennant et al., 2007); 9-item Moral Injury Event Scale (MIES) ( Nash et al., 2013); 12-item Burnout Assessment Tool ( Schaufeli et al., 2019); 3-item Clinical Interview Schedule-Revised (CIS-R) Lamb et al., 2021); 10-item Alcohol Use Disorder Identification Test-C (AUDIT) ( Babor et al., 2001); 6-item PCL-6 civilian version to explore PTSD ( Lang and Stein, 2005). This paper omits data from the BRS, MIES and WEMWBS.

As the sample is self-selecting, it is possible that participation may be correlated with outcome and/or exposure variables, adversely affecting the generalisability of results to the eligible population. The direction of any such selection bias is difficult to predict, as a plausible case can be made for either under- or over-representation of mental health morbidity in the sample. Efforts to improve the generalisability of the sample and minimise selection bias include the range of methods by which the study was advertised, the information provided to participants on the study website, and the splitting of the survey into the core “short” and optional “long” versions.

To control for potential bias due to confounding of the associations of interest with other variables, covariate-adjusted models were fitted (see below under Multiple Regression Model).

The primary outcome was common mental disorder, as measured by a GHQ-12 score greater than 4. Mean GHQ-12 scores in the UK increased from 1.8 to 3.8 during the pandemic, and the average was higher still for key workers and individuals from ethnic minority backgrounds ( Fujiwara et al., 2020). Preliminary results from the NHS Check study ( Lamb et al., 2021), using the same survey in a similar population, suggested the prevalence of common mental disorders in the study sample may be as high as 60%.

We estimated that a sample size of 650 would provide at least 80% power at the 95% significance level to detect an absolute difference of 12% in prevalence between groups based on survey responses; 70% power to detect a difference of 11%; and 50% power to detect a difference of 9%. These figures assume that groups may differ in size by a factor of 3; to the extent that numbers in the groups are more evenly matched, power will be somewhat improved.

Objectives

The three analysis objectives were to: 1.

Describe the baseline demographic characteristics of the sample cohort;

Employing statistical software (R 4.0.3), descriptive statistics were used for socio-demographic data. Summarized distributions of each outcome in the overall study sample are presented. Mean, standard deviation and 95% confidence interval (CI) are presented for continuous/numerical variables, including all validated and novel composite scores (the latter is discussed below). The number of cases and their prevalence are reported for binary variables, and the number and proportion in each category for ordinal or other categorical variables. For the validated mental health measures, both the numerical score and the derived binary variable are shown.

Outcomes were compared cross-sectionally in accordance with the following contrasts: •

Participants who came into contact with those with suspected COVID-19 symptoms vs participants who did not;

Comparisons were performed for the following outcomes: all binary outcomes derived from validated mental health scores, and a binary indicator for use of any staff support programme. Missing responses were excluded from the analysis on an outcome-by-outcome basis.

Unadjusted comparisons

For each outcome and contrast, an unadjusted odds ratio and 95% CI is presented.

Multiple regression model

A multiple logistic regression model was fitted for each outcome, using as covariates all of the contrasts listed above (except age, which was included as a continuous covariate) together with the demographic variables collected in the short-form baseline survey and the novel scores for quality-of-care provision and social disconnection.

The quality-of-care provision score was based on the set of questions: “In the past 2 weeks … - I have had to do work I do not feel sufficiently competent or trained to do”, “In the past 2 weeks … - I or my team have had to provide significantly worse care than usual to our patients or deny them treatment that would normally be available”, and “In the past 2 weeks … - I have felt let down because I am working with insufficient staffing or resources.” Each question took the possible answers “Never”, “Rarely”, “Sometimes”, “Often”, or “Always”.

The social disconnection score was based on the set of questions: “Since the COVID-19 (coronavirus) pandemic - How well do you feel supported by your family and/or friends?”, “Within the last few weeks: - Do you feel more isolated from family/friends because of your work role?”, “Within the last few weeks: - Do you feel like friends, family or strangers are treating you differently (positively) because you work in a hospital/care setting?”, and “the last few weeks: - Do you feel like friends, family or strangers are treating you differently (negatively) because you work in a hospital/care setting?” Each question took the possible answers “Not at all”, “A little bit”, “Moderately”, “Quite a bit”, or “Extremely”.

In both cases the composite score was derived by applying Likert scoring (0-4, where a higher score indicates a less favourable response) to each question and summing the result within each set. Social connectedness was therefore scored between 0 and 16, and quality of care provision between 0 and 12. Categorical variables based on these ranges were then included as predictors in the model (see Table 5 for details).

The full list of regression covariates is shown in Table 5. We explored the use of a linear mixed model with a random intercept for each participating employing organisation (and fixed effects for the covariates) to allow for additional variability of outcomes between organisations; however, the random effects were not significant, and therefore a fixed intercept model was used instead. In the multiple regression analysis, for each covariate a separate category was included for responses indicating uncertainty or unwillingness to indicate a category, such as “Don’t know”, “Unsure”, “Other” or “Prefer not to say”. Missing values for the social disconnection and quality-of-care scores were also treated as a separate category. For other covariates the level of missingness was very low and participants with missing values were excluded.

Significance testing

A 95% confidence interval is presented for each studied mean score, prevalence and relative effect measure. Regression coefficients were annotated to indicate the statistical significance according to the magnitude of the P-value, i.e., *=P<.05, **=P<.01, ***=P.001.

A total of 1,065 participants completed the baseline survey, with 159 (15%) completing the short survey only, and 906 (85%) also providing responses to the long survey.

Participants’ socio-demographic and role profile

The characteristics of the study sample are shown in Table 1. A greater proportion of participants were aged 30-59 years (72%; n=755), female (76%; n=806), in some form of relationship (68%; n=726), with a greater proportion on bands 7-9 of the Agenda for Change pay rate scale (26%; n=276) and had been working for their current employer for between 1-5 years (43%; n=454).

Prevalence of adverse mental health outcomes

The prevalence of mental health scores exceeding the cut-off for probable morbidity are shown in Table 2 along with the mean and standard deviation of the raw scores. Over half (54.1%; n=543) of respondents exceeded the cut-off score on the GHQ-12 for probable mental disorders, and approximately a quarter for anxiety (GAD-7, 25.8%; n=191). Approximately 30% exceeded the cut-off point for probable depression (PHQ-9, 29.8%; n=219), and similarly for probable burnout (31.1%; n=222) and PTSD (31.7%; n=189). Just over a tenth (10.6%; n=78) recorded probable alcohol misuse.

From the Clinical Interview Schedule ( Goldberg and Williams, 1988), nearly three-tenths (28.6%; n=214) of respondents had experienced suicidal thoughts, with one in ten (9.9%; n=74) reporting suicidal ideation in the last two months; 1.7% (n=13) had made an actual attempt on taking their own life in the previous two months. Over one in ten (11.5%; n=86) reported ever harming themselves but not with the intention of killing themselves, with 2.5% (n=19) reporting doing so in the last two months.

Table 3 summarises the responses to three survey items on fatigue. Fifty-eight percent (n=429) of respondents scored 5-8 when grading their emotional and physical exhaustion. Approximately 1 in 6 (16.8%; n=125) scored 9-10, the extreme end of the severity scale. Respondents also reported not only was this fatigue significantly greater than their usual tiredness (67.2%; n=477), but over half (51.4%; n=376) said it interfered with their ability to do things.

While more than 1 in 6 (17.1%; n=110) respondents reported drinking more alcohol in the past week than they would have typically before the pandemic, 82.9% (n=533) reported consumption levels that were either the same (41.7%; n=268) or less than usual (41.2%; n=265).

Table 4 shows the number and proportion of survey respondents indicating they had used various types of staff mental health support measures, as well as the number and proportion giving a “Not applicable” response.

Nearly a third (n=328) of respondents reported using some kind of staff support service. Of these, 61.0% (n=200) used wellbeing services, 38.7% (n=127) used relaxation/timeout areas, and only 22.0% (n=72) used employee assistance programmes.

A logistic regression model was fitted for each binary indicator of probable mental health disorder (as measured by GHQ-12, GAD-7, PHQ-9, BAT, AUDIT and PCL-6), and also an indicator for having used any staff support measures. For each of these outcomes, the covariates and their associated odds ratios and confidence intervals are shown in Table 5.

Older staff were less likely to have anxiety (Odds Ratio [OR] for 1-year increase in age: 0.98 [0.96,1.00]), or burnout (OR for 1-year increase in age: 0.98 [0.96,0.99]) than younger staff. Male staff were significantly less likely to access staff support programmes (OR 0.61 [0.43,0.88]) than females. Minority ethnic staff were significantly less likely to misuse alcohol (OR 0.19 [0.08,0.45]).

Staff in pay bands 4-6 (for example, ward nurses) were significantly more likely to have common mental health disorders (OR 2.42 [1.20,4.88]) and PTSD symptoms (OR 4.18 [1.44,12.14]) than lower banded staff, and were also more likely to access staff support (OR 2.66 [1.29,5.48]) than lower banded staff. The same outcomes were also significantly associated with bands 7-9 (senior clinical and managerial), albeit with marginally lower odds ratios.

Staff working in acute care settings were more likely to experience common mental health problems (OR 1.50 [1.08,2.07]) and PTSD (OR 1.57 [1.03,2.41]) than those in non-acute settings (e.g., community/primary care). Those with direct clinical facing duties had higher odds of PTSD symptoms than those in non-clinical roles (OR 1.60 [1.02,3.51]).

Social connectedness and perceived delivery of poor-quality care were highly associated with risk of all mental health problems apart from alcohol misuse (OR ranging from 1.56–18.34).

Psychosocial impact

The prevalence of mental health morbidity in this study was high. A possible reason for this is the self-selecting, rather than probabilistic nature of the sample. Staff with existing mental health issues arising from the pandemic may be more likely to be attracted to the survey’s subject matter, thereby inflating the prevalence of mental ill-health findings. Nonetheless, the finding that nearly 55% of staff could probably be diagnosed with a mental disorder, and that over 25% could be diagnosed with anxiety and depression, is comparable to other studies. The NHS Check study reported a similar rate of probable mental disorders from the first pandemic wave ( Lamb et al., 2021), while a study on intensive care unit (ICU) staff reported 45% met the threshold for a clinically significant mental disorder ( Greenberg et al., 2021). Ike et al’s (2021) study of clinical staff in a tertiary London hospital reported higher rates of both anxiety (41.1%) and depression (42.7%). This difference could be accounted for by the inclusion of both clinical and non-clinical staff in the study, or the different period of the pandemic in which the data were collected.

Interestingly, unlike previous studies on the mental health impact of the pandemic ( Lamb et al., 2021; Ike et al., 2021; Cipolotti et al., 2021) our study did not find female sex to be associated with increased mental problems. Particularly concerning is the proportion of respondents (nearly 30%) reporting suicidal thoughts, the one in ten reporting suicidal ideation in the last two months, and the ten staff members reporting having tried taking their life in the previous two months. This rate lies between the figures reported by Lamb et al. (2021) (8.5% in the last 2 months) and by Greenberg et al. (2021), who reported a rate of 13% among ICU staff over the last two weeks, with a higher rate among nurses compared to doctors or other clinical staff. Similarly, self-harm, with no suicidal intent, was found among 11.5% of respondents, but less common (2.5%) over the last two months.

Of less concern—in the absence of baseline data as to ‘typical’ staff levels of alcohol consumption—is the fact that the vast majority of respondents (82.9%) reported alcohol consumption levels that were either the same or less than usual. This could be accounted for by restricted opportunities for social drinking (in the company of others in a public setting), especially during lockdown periods. It might also be affected by social desirability bias and the fact that the sample had a high proportion of people from ethnic minority backgrounds (39%), in whom we found very low likelihood of alcohol misuse. The proportion of participants reporting drinking more frequently is lower than that in the general population in the UK, among whom 30% reported an increased frequency of drinking ( Oldham et al., 2021). Nonetheless, overall, 10.6% of the sample met the threshold for probable alcohol misuse. This is consistent with findings from the NHS Check study, which reported a rate of 10.5% ( Lamb et al., 2021), and lower than the rates of problem drinking among UK adults overall. According to one study, 23% of UK adults reported drinking ≥ 4 times a week, and 16.6% reported heavy episodic drinking at least once weekly during the COVID-19 lockdown in April 2020 ( Daly and Robinson, 2020).

We created novel composite scores for social disconnection and self-perception of poor-quality care provision which were very strongly associated with most of the validated mental health outcomes. This is consistent with Cipolotti et al’s (2021) finding that social change concerns, which include distancing from family and friends, is predictive of psychological distress among HCWs. Missing or incomplete responses to the quality-of-care questions were also very strongly associated with poor mental health, possibly suggesting reluctance to provide responses criticising care provision. These findings are important for those assessing staff at risk of, or currently experiencing, problems with mental wellbeing, as both are potentially modifiable. For example, employers should consider screening for social support and job satisfaction, and signpost staff to community social prescribing interventions or offering such interventions within employing organisations for staff with low social support. The importance of feeling that the care staff offer is high quality is intriguing. It is unclear from our study whether this reflects a generic sense of purpose among health care staff that is an important determinant of wellbeing, or whether there has been a direct causal effect of COVID-19 on perception of quality care.

We found strong evidence that younger staff were more anxious than their older colleagues. While the association between adverse outcomes and younger age was also found in Lamb et al’s study (2021), age was not found to have a significant effect on psychological distress scores in Cipolotti et al’s (2021) study of neuroscience HCWs in the UK.

Salary pay-scale bands 4-6 tended to have the worst mental health outcomes and to use support most often, although bands 7-9 are not far behind in both respects. This is in contrast to those on bands 1-3, who had the best outcomes and were least likely to use support. This is consistent with previous research indicating that nurses (who are generally in bands 4-6) are more likely to meet clinical thresholds for depression, PTSD, and anxiety ( Lamb et al., 2021; Greenberg et al., 2021; Ike et al., 2021), and that staff with higher levels of control and authority may be less impacted. Most concerning is that nearly one third of respondents had probable burnout, and almost as many reported probable PTSD. The long-term consequences of burnout and fatigue are significant, particularly if they become chronic. The factors contributing to burnout appear to be modifiable, however, and include working in a COVID-19 role, having access to PPE and the ability to rest and recover during breaks ( Gemine et al., 2021). Working in acute care, or in a clinical role, was associated with an elevated risk of PTSD, while acute care settings were associated with common mental health disorders.

Existing research indicates NHS staff consider psychological support and clear updates as the most useful forms of support for protecting their mental health ( Cipolotti et al., 2021; Cubitt et al., 2021). Crucially, the types of support endorsed by staff seem to vary based on age, professional role, and exposure to COVID-19 patients ( Cipolotti et al., 2021). A number of organisations have called for better mental health and wellbeing support for health and social care staff following the COVID-19 pandemic ( Health and Social Care Committee, 2021). However, there appears to be discordance between guidelines and staff perspectives of what the focus of support interventions should be, as the recommended interventions do not always address staff lived experiences ( San Juan et al., 2020). It is therefore important to examine the impact and effectiveness of the interventions being offered by organisations.

A number of limitations should be acknowledged. First, the sample may not fully represent the NHS workforce. Despite a comprehensive and inclusive recruitment strategy across all health care provider organisations in NWL and adoption by the Clinical Research Network, giving the many thousands of potential recruits, at just over one thousand the number of responses was disappointing. This can be explained by multiple factors, including: a surfeit of concurrent staff surveys—research, audit, routine annual surveys (e.g., the national NHS Pulse survey ( https://www.england.nhs.uk/nhs-people-pulse/)—that contributed to ‘survey fatigue’; occupational burnout among staff members striving to combat the impact of COVID-19 upon those they cared for; and a benign paternalism among protective senior management anxious not to further ‘burden’ staff by asking about their wellbeing.

Second, the study sample was self-selecting rather than probabilistic. This could either over- or under-represent the prevalence of the studied psychosocial, occupational and support programme domains. However, most levels of key socio-demographic variables were well represented, and the overall data distribution appears broadly in alignment with expectations for the target population.

Third, there were some differences between those completing the short and long surveys (see Table 6). Respondents choosing to withdraw after the short survey were more likely to be male, from an ethnic minority background, be born overseas, and have a more favourable GHQ-12 score. Those continuing to engage with the survey to its end were also more likely to be in a clinical role. Additionally, since most outcomes—other than the GHQ-12—are only measured in the long survey, this may further affect the generalisability of those results. However, it is important to note that given we cannot observe the earlier self-selection determining who is in the sample to begin with, there is no guarantee it operates in the same direction; i.e., it is possible those who completed the long-survey could actually be more representative of the population as a whole than the short-survey respondents.

Fourth, given the number of potential associations being investigated, it is highly likely that some of the significant findings (especially significance codes * and **) are false positives ( Fisher, 1971). However, some variables, particularly pay band, perception of care quality and social disconnection, were associated with more than one mental health outcome and showed evidence of a consistent trend across categories, suggesting that these findings are especially credible.

Lastly, our study does not have any pre-pandemic comparative mental health data for NHS staff or for non-NHS staff during the pandemic with which to further contextualise the impact of COVID-19 on our study population.

This study found ongoing evidence of significant psychosocial morbidity and adverse occupational impacts upon NHS staff members from the second wave of the COVID-19 pandemic that echoes that reported from the first wave, identified staff groups at risk factors and suggests areas for potential intervention.

Research is needed to explore how this group’s psychological burden changes, if at all, over time, and if that change is attributable to factors around personal resilience or preferred and effective local and/or national mental health support interventions and resources. It is only through such insights that effective preventive measures and reactive interventions can be designed that are tailored to the needs of HCW groups.

The sample was a large, ethnically diverse group of NHS staff from across the workforce, in clinical and non-clinical roles, working in acute and community settings and from low grade to senior roles. The proportion of respondents completing the full survey once they had enrolled was high. The protocol was subject to extensive peer review and measured a range of mental disorders, not just common mental health problems of anxiety and depression. The majority of data collection took place during the second wave lockdown, when rates of infection and death were particularly high (winter 2020/21). It is likely therefore to reflect the worst of the pandemic in terms of mental health impact. The self-selecting nature of our sample may mean that those with a high mental health burden are either under- or over-represented. However, our findings echo many hypotheses from the literature, add novel findings and have important workforce implications.