Measures.

At baseline and the six- and 12-months follow-ups, in addition to the main questionnaire, participants were invited to complete an additional longer survey. Splitting the survey offered participants the option to reduce the time taken to contribute, while still collecting some key data. As a result, sample sizes for measures included in the ‘long’ survey are smaller than the overall sample size. At 32-months all participants received the full set of survey questions. While the majority of the survey at each time used the same measures, to account for changing research and policy priorities the surveys were not identical. See the Supplementary files for information on which measures were used in each survey.

Surveys collected information on participants’ socio-demographic characteristics, their occupational role and context, COVID-19 experiences, and the validated Moral Injury Events Scale (MIES) was used to collect data on exposure to potentially morally injurious events (PMIEs) [14], with endorsement of one or more item indicating exposure (i.e., selecting moderately/strongly agree).

The primary outcome was prevalence of probable common mental disorders (CMDs), ascertained using the 12-item General Health Questionnaire (GHQ-12). The GHQ scoring method was used (where each item is scored 0-0-1-1, resulting in a total score of 0–12 for the scale), with a cut-off score of 4 or more indicating ‘caseness’ of a CMD (indicating increased probability of experiencing a recognised mental disorder) [15].

Secondary outcomes included:

1. Anxiety, assessed using the Generalised Anxiety Disorder scale (GAD-7) with higher scores indicating higher levels of anxiety, and a cut-off of ≥10 indicating caseness [16]
2. Depression, assessed using the Patient Health Questionnaire (PHQ-9) with higher scores indicating higher levels of depression, and a cut-off of ≥10 indicating caseness [17].
3. Problem drinking, assessed using the Alcohol Use Disorder Identification Test (AUDIT-C), with higher scores indicating more problematic drinking, and a cut-off of ≥8 indicating caseness [18].
4. Post-Traumatic Stress Disorder (PTSD), assessed using the Post-Traumatic Stress Disorder checklist (PCL–6), with higher scores indicating higher levels of PTSD, and a cut-off of ≥14 indicating probable caseness [19].
5. Burnout, assessed using the Burnout Assessment Tool (BAT-12), with higher scores indicating higher levels of burnout, and a cut-off of 2.96 indicating probable caseness [20].
6. Wellbeing, assessed using the Warwick Edinburgh Mental Wellbeing Scale (WEMWBS), with higher scores indicating better mental health, and scores of ≥43 indicating good mental health [21].
7. Resilience, assessed using the Brief Resilience Scale (BRS), with higher scores indicating higher levels of resilience [22].
8. Growth after trauma, assessed using the Post-Traumatic Growth Inventory Short Form (PGTI-SF), with higher scores indicating higher levels of post-trauma growth [23].

There are no accepted cut-off scores for the BRS or PTGI, so, for the purposes of consistent analysis, for this paper we generated tertiles for each measure and used the cut point between the top and middle tertiles in each case to indicate presence of high resilience or growth respectively. Using this method, for the BRS a score of ≥24 indicates higher levels of resilience, and for the PTGI a score of ≥17 indicates higher levels of growth.

Survey period.

To account for time-varying pressures on NHS Trusts and staff, we included a variable measuring survey period. We hypothesised that reporting of mental ill-health symptoms would be greater during periods of higher pressure and that this could impact on outcomes at later timepoints. In line with previous research [24], we identified pressure as: High early in the pandemic (T1: April-June 2020), when little was known about COVID-19 and swift changes were needed in healthcare settings and in general life; Lower in the summer/early autumn of 2020 (T2: July-October 2020), when the initial wave of infections lessened and lockdown restrictions eased; before increasing again over the winter months (T3: November 2020-February 2021). These time periods correspond to periods with higher and lower COVID-19 death rates: T1 COVID-19 deaths = 53,389; T2 COVID-19 deaths = 7,157; T3 COVID-19 deaths = 82,083 [25]. This enabled us to explore outcomes and associated factors, taking into account the effects of differing levels of pressure on the NHS when participants completed their baseline and some of the follow-up surveys.

Data cleaning.

Using email address as a unique identifier, where participants erroneously completed the survey more than once (at any timepoint), we retained their first response only, with any missing demographic data available in dropped responses added to the retained response.

As specified in the protocol [26], we excluded any Trust where the response rate was below five percent.

Weighting.

We derived two sets of weights. First, we generated baseline response weights using aggregate demographic information provided by Trust Human Resources (HR) departments as of April 2020. We derived weights for the full baseline cohort using a raking algorithm (citation) based on age, sex, ethnicity, job role For weighting purposes only, missing demographic data were imputed using k-Nearest Neighbour (kNN) imputation (k = 5; [27]). At baseline, missingness was ≥ 5% for age, sex, ethnicity, and job role. Individual person level strata were weighted by Trust size as well as composition of the four main proposed prognostic factors of ethnicity, categorical age, role and sex. Post strata and post weight specifications were included but no finite population correction was made.

Second, we applied inverse probability weighting to account for non-response at each follow-up timepoint. A binary logistic regression model was used to model the probability of non-response based on age, sex, ethnicity, and job role at each timepoint.

For each timepoint, we multiplied these two weights to create a single combined weight that was applied to all analyses.

Descriptive analyses.

We explored the data in several stages. First, we described the characteristics of those completing the surveys at each timepoint. We then summarised the weighted prevalence and mean scores of the primary (GHQ-12) and secondary outcomes (GAD-7, PHQ-9, AUDIT-C, PCL-6, BAT-12, WEMWBS, BRS, PTGI) cross-sectionally by timepoint. Finally, we checked for collinearity between the variables.

Regression analyses.

We used multilevel models to account for clustering by Trust. Our primary interest was in exploring factors associated with meeting cut-off scores on our outcome variables, so we used multivariable logistic regression models. The factors included in the models were identified a priori, based on existing evidence and research team discussions. The baseline factors included in the models were agreed upon by the research team based on existing evidence and expert opinion: age category; sex; ethnicity; relationship status; job role; pay grade; job setting; had good access to PPE; felt supported by colleagues; felt supported by managers; experienced moral injury; baseline measure of the outcome variable (in models for 6, 12, and 32 months); the survey period during which participants completed their baseline survey. We modelled outcomes adjusting for all other variables in the model, and include unadjusted models in Table S3 in S1 File.

This paper presents descriptive data at multiple time points, and we provide longitudinal trajectory modelling in a separate paper [accepted for publication in Psychological Medicine – reference to be provided ASAP].

All analyses were conducted using Stata version 18.0.

Cohort demographics.

The demographic composition of the sample can be seen in Table 2, which shows the total number of respondents in each category, with total unweighted and weighted proportions given, for the four time periods.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Descriptive characteristics of the cohort for all time periods, with weighted and unweighted proportions.

https://doi.org/10.1371/journal.pone.0350918.t002

On checking for collinearity between the variables, the mean VIF was 1.16, and no VIF was found above 1.5, indicating that collinearity is not a concern for these variables.

Outcome prevalence.

At baseline, over half of participants reported symptoms of common mental disorders (CMDs) (52%, 95%CI 51%, 53%), and this was very similar at six months (51%, 95%CI 49%, 52%), 12 months (47%, 95%CI 46, 49), and 32-months (50%, 95%CI 48%, 52%). The proportion of those reporting symptoms of anxiety and depression remained stable over the four time periods, at between 21–29%, and the proportion reporting good wellbeing also showed little change, at around 50% at each time period. Proportions reporting alcohol misuse and PTSD showed a spike from baseline to six months (around 1/5 and 1/3 reporting symptoms respectively) before dropping again at 12 and 32 months, although notably PTSD remained at nearly 1/3 at 32 months. Proportions reporting burnout rose over the first three time periods, from 16% (95%CI 14%, 17%) at baseline to 26% (95%CI 24%, 27%) at 12 months, dropping to 20% (95%CI 19%, 22%) at 32 months. Levels of resilience remained at around 1/3 reporting resilience at the first three timepoints. PTGI was collected at 6,12, and 32 months, and remained at just over 1/3 reporting growth at each timepoint.

Fig 1 shows proportions and confidence intervals for each outcome measure (this data is available in Table S2 in S1 File).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Dot plot of outcome variables across timepoints with 95% confidence interval whiskers.

https://doi.org/10.1371/journal.pone.0350918.g001

Common mental disorders (primary outcome).

A multilevel multivariable logistic regression found that at baseline, six and 32 months, compared to those aged 30 or younger, those over 40 had statistically significantly lower odds of reporting symptoms of CMDs, but this was not the case at 12 months, where 31–40 year olds had higher odds than those ≤30. At baseline, six months, and 32 months, compared to women, men had lower odds of reporting symptoms of CMDs. Compared to White participants, at baseline Asian participants had lower odds of reporting symptoms of CMDs, and at 32 months those in a relationship had lower odds of reporting symptoms than those who were single. At baseline and 12 months, compared to non-clinical staff, those in other clinical roles had higher odds of reporting symptoms, while at six months doctors had lower odds of reporting symptoms than non-clinical staff. At six months, those working in ICU/Critical Care had higher odds of reporting symptoms. At baseline and 32 months, those who often or always had access to adequate PPE had lower odds of reporting symptoms than those who never, rarely, or sometimes had access. At baseline, those who felt supported by colleagues and by managers had lower odds of reporting symptoms than those who did not feel supported, and this was true for colleagues at 12 months too. At all timepoints, those who had experienced moral injury had higher odds of reporting symptoms than those who had not. Those who completed their baseline survey later, during a high pressure period (Time 3, Nov 20-Feb 21) had higher odds of reporting symptoms at baseline and 12 months, compared to those who completed their baseline survey during a lower pressure period (Time 2, Jul-Oct 2020). Full results are available in Table 3.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Weighted, multilevel, multivariable binary logistic regression models of factors associated with meeting GHQ-12 cut-off at each timepoint.

https://doi.org/10.1371/journal.pone.0350918.t003

Secondary outcomes.

The full results of the secondary outcome models are available in Supplementary files (Tables S4-S11 in S1 File). A summary of the results is available in Table 4 below.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Summary of primary and secondary outcome results – those with higher odds of meeting cut-off (unless otherwise specified).

https://doi.org/10.1371/journal.pone.0350918.t004

The strongest and most consistent finding was that having met cut-off on a measure at baseline was associated with meeting cut-off on that same measure at the current time. This was the case for nearly all timepoints and all measures, except that those reporting more resilience at baseline did not have higher odds of reporting more resilience at 6 months.

Several factors were associated with nearly every measure, but not at every timepoint. Being younger was associated with worse outcomes across every measure at some timepoint, particularly at baseline and 6 months. Older age was associated (at baseline) with alcohol misuse. Moral injury was consistently associated with worse outcomes across all measures except alcohol misuse. Feeling unsupported by colleagues/managers was linked to worse outcomes for most measures, excluding alcohol misuse, resilience, and post-traumatic growth. Being single was also associated with worse outcomes across most measures.

Black, Asian, Mixed and Other ethnicity participants generally had better outcomes than White participants, across multiple measures and timepoints. Exceptions included worse alcohol outcomes for White participants at all timepoints except 12 months.

Female participants had worse outcomes across most measures and across timepoints, except for alcohol misuse (more in males), and greater post-traumatic growth. Notably, females did not report worse depression than males at any time.

Findings were less consistent for other factors. Job setting showed mixed results, with worse outcomes for A&E and ICU/Critical care on some measures at various timepoints, and those working in other hospital settings, community, and non-patient-facing settings having worse outcomes on other measures/timepoints.

There was some evidence that inadequate PPE was associated with worse outcome. Similarly, nurses, other clinical staff, and non-clinical staff reported reported worse outcomes than doctors on some measures at some timepoints.

Those who completed their baseline survey at a higher pressure timepoint (Time 1 and 3) tended to have worse outcomes than those at Time 2.

Lower pay grade was associated with worse outcomes on some measures at different timepoints. However, higher paid was associated with higher alcohol misuse at 6 months, lower resilience at 12 months, and higher burnout at 32 months.

In the summary table below, the characteristics listed are those with higher odds of having poor outcomes, unless otherwise specified.