Ethics

Ethical approval for the study was granted by the Health Research Authority (reference: 20/HRA/210, IRAS: 282686) and local Trust Research and Development approval. Cohort data are collected via Qualtrics online survey software, pseudonymised and held on secure university servers. Participants are aware that they can withdraw from the study at any time, and there is signposting to support services if participants feel they need it. Only those consenting to be contacted about further research will be invited to participate in further components.

Data collection

The data used in this analysis comes from the NHS CHECK study (for more information about the wider study, see [12]), a longitudinal cohort study of the mental health and wellbeing of HCWs during the COVID-19 pandemic in England. All staff members in 18 participating NHS Trusts were invited to complete an online survey once between April 2020 and January 2021. At the end of the survey was a free-text box where participants were asked, “Is there anything else you think we should know about your experiences of the COVID-19 pandemic?”. The responses to this question form the data used in this analysis.

Data analysis

We performed STM, implemented with the stm R package [13] to extract topics from responses. We performed STM using unigrams (i.e. single words). STM treats documents as a probabilistic mixture of topics, and topics as a probabilistic mixture of words. STM is a “bag of words” approach that uses correlations between word frequencies within documents to define topics. An advantage of STM is that it allows for inclusion of covariates in the estimation model, so the estimated proportion of a text devoted to a topic can differ according to document metadata (e.g., participant characteristics). We included participant’s age (modelled with natural splines with 3 degrees of freedom to account for potential non-linearities in the association: [14], date of response (natural splines, degrees of freedom 3), sex, marital status (married, co-habiting, single, divorced, widowed), ethnicity (White, Mixed Race, Black, Asian, Other) and NHS trust (18 categories) in our models, each collected at baseline. There was only a small amount of item missingness (6.1% for age; 2.2% for sex; 2.1% for marital status; 1.9% for ethnicity; 3.8% for role), so we used complete case data.

Prior to performing the analysis, we cleaned free-text response using an iterative process. We excluded responses containing fewer than five words and removed words which appeared in fewer than five responses. We also removed common “stop” words (“the”, “and”, “I”, etc.) from the analysis. Spelling mistakes were identified with the Hunspell spellchecker [15], amended manually if they had five or more occurrences and replaced using the Hunspell suggested word function if the number of occurrences was fewer than five. Where the algorithm provided multiple suggestions, the word with the highest frequency across responses was used. To reduce data sparsity, in the structural topic models, we used word stemming using the Porter [16] algorithm (e.g., help, helping and helped become help). Data cleaning was carried out in R version 3.6.3 [17].

We ran STM models, which included between two to 50 topics and selected the final models based on visual inspection of the residuals and lower bound statistics of the model solutions. After selecting a final model (with 35 topics), we carried out two analyses, a thematic analysis of the topics and qualitative text excerpts, and a regression analysis of participant characteristics and generated themes. In line with previous work using this methodology [18], for the thematic analysis, we generated short descriptive titles for the 35 topics, firstly individually, and then via collaborative discussion. We dropped one topic due to incoherence between the exemplar texts, and merged two topics due to similarity. The remaining 33 topics were grouped, again, initially individually, and then in an agreed thematic structure via a collaborative process involving interpretation of the content of the topics. The exemplar responses for each topic were selected based on the ‘Theta’ assigned to them. Theta is the estimated proportion of text the response contains on a given topic. The closer Theta is to 1, the more likely the responses are to form a single semantically coherent topic. Only responses with a Theta greater than 0.3 were included in the analysis. Descriptive summaries were written for each theme and subtheme, and the thematic analysis proceeded via an iterative process of reading and re-reading exemplar responses. We identified responses that captured the essence of each of the themes generated, and reflected on meanings and interpretations of these in terms of the overarching narrative of the themes and data, and in terms of our own perspectives. While the generation of themes was supported by STM, rather than initial familiarisation with every text excerpt (which, while usual in reflexive thematic analysis, was not possible with over 7,000 responses), the subsequent stages of meaning-making and discussion of our own positionality was in keeping with recent guidance regarding reflexive thematic analysis [19].

Reflexivity

The researcher team comprised women and men, researchers and clinicians, and diversity in terms of age and ethnic background. The non-clinical researchers on the research team have extensive experience of conducting qualitative mental health research with occupational groups, particularly healthcare workers, and all have colleagues, friends, or family members who worked in healthcare roles through the pandemic. Several of the clinical researchers worked on the frontline during the pandemic in mental health roles. As such, all authors had some existing perspectives on the pressures faced by healthcare workers during this time (and before), whether first or second hand. We therefore undertook extensive discussion within the team throughout the analysis process of our subjective interpretations of the data, and how these may have been influenced by our own experiences and perspectives.

Patient and public involvement

Frontline NHS staff proposed this research, and we tested the proposal’s acceptability and approach with a small informal reference group of front-line staff (psychologists, managers, intensivists and trainee psychiatrists) and refined it accordingly. We have developed an advisory group of NHS staff (clinical, managerial, auxiliary, students), who meet online and contribute virtually to provide input on methods, recruitment strategy, communications, and interpretation of findings.

Descriptive statistics

Sample descriptive statistics are displayed in Table 1. There were 7,412 participants (35% of total sample) who provided a valid free-text response. The distribution of responses according to age, sex, marital status, ethnicity, role, and NHS trust type was similar among those who provided a free-text response and those who did not.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Descriptive statistics.

https://doi.org/10.1371/journal.pone.0275720.t001

Thematic analysis

From the 33 topics, two overarching domains were identified, personal (45.8% of text), and professional (52.9% of text), each containing four themes (see Table 2), although many themes contained elements of both the personal and the professional, and were often interlinked.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Theme table.

https://doi.org/10.1371/journal.pone.0275720.t002

The personal domain contained themes related to physical health (15.4% of text), mental health (12.4% of text), social impact (10.0% of text) and restrictions and rules (8.0% of text). The professional domain contained themes related to home working (17.1%) of text), workplace challenges (13.7% of text), workplace culture (11.3% of text), and new roles (10.9% of text). See Fig 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Proportion of text devoted to each theme.

Source: Drawn by the authors from data analysed for this paper.

https://doi.org/10.1371/journal.pone.0275720.g001

After each quote used to illustrate the themes below is the role of the participant, grouped by job role (doctors, nurses, Allied Health Professionals (AHPs), and non-clinical staff). Sub-themes are indicated by bold text.

Personal

Professional

Regression analyses

The results of the linear regression models are displayed below. There were some differences in themes raised by age (Fig 2). Notably, older participants were more likely to discuss challenges related to physical health and to home working, while younger participants were more likely to discuss new roles.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Marginal effect of age on estimated proportion of text devoted to theme.

Derived from linear regression models including adjustment for age, date, sex, marital status, ethnicity and NHS trust. Source: Drawn by the authors from data analysed for this paper.

https://doi.org/10.1371/journal.pone.0275720.g002

There were also differences according to the date free-text responses were recorded (Fig 3). As may be expected, discussion of home working decreased across the pandemic, while discussion of workplace challenges increased. Discussion of mental health was lowest between September-November 2020, the period between the first and second waves of COVID-19 in the UK.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Marginal effect of date on estimated proportion of text devoted to theme.

Derived from linear regression models including adjustment for age, date, sex, marital status, ethnicity and NHS trust. Source: Drawn by the authors from data analysed for this paper.

https://doi.org/10.1371/journal.pone.0275720.g003

Differences according to sex, marital status and ethnic group were also observed (Fig 4). Notably, women were more likely to discuss home working, new roles and social impact, and were less likely to discuss restrictions and rules. There were also differences by NHS trust, most notably regarding positive and negative experiences of home working and physical health (results not shown for privacy reasons).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Marginal effect of sex, marital status and ethnicity on estimated proportion of text devoted to theme.

Derived from linear regression models including adjustment for age, date, sex, marital status, ethnicity and NHS trust. Source: Drawn by the authors from data analysed for this paper.

https://doi.org/10.1371/journal.pone.0275720.g004

Summary of results

Responses from the 7,412 participants fell broadly into two domains, personal and professional. though there was overlap in the content of several themes, with important connections between home working, isolation, and HCW mental health. Furthermore, we found evidence that demographic factors affected the way the respondents reported their experiences of the pandemic, and that experiences also varied depending on the nature of national picture of the pandemic at the time the data was collected. This paper presents qualitative data from the largest UK survey of HCW wellbeing, with demographics broadly representative of the wider NHS workforce [20]. It is likely that participants reported the most salient experiences for them, and thus the responses provide insight into important factors that may not have been picked up by other research using pre-defined survey questions or fixed interview schedules.

One, perhaps unsurprising, finding was that the personal challenges most commonly reported by HCWs overlapped considerably with those faced by the general population (e.g. bereavement, social isolation, caring responsibilities). However, we also identified specific workplace challenges for HCWs that are less relevant for many other types of workers (e.g. facing significant risk of infection, redeployment), and these in turn impacted home life. That said, our sample was not homogenous, and variations between participants’ experiences within the workplace suggest that guidance around infection control, vulnerable staff, and redeployment differed between Trusts and teams. This also demonstrates important differences in personal preferences for ways of working, and underlines the challenges for Trusts in offering support that is appropriate for all members of staff.

Results in relation to previous research

Some of our findings echo the results of previous quantitative work in this area especially in that increased workloads were a key concern for HCWs [2]. Many of our participants reported increased workloads, with a vicious cycle of this leading to burnout and sickness absence, which in turn led to further increased workloads for those remaining; this has been noted in other qualitative research [21, 22]. As our regression analyses showed, discussion of workplace challenges increased over time, indicating the growing importance of these as the pandemic progressed. However, concerns about heavy workloads would not have been solely due to the pandemic, though it has clearly exacerbated them. The 2018 NHS Staff Survey found that 40% of HCWs reported work-related stress, burnout, and dissatisfaction due to increased workload, which was caused by staff shortages [23]. In our findings and in other research, inconsistency and uncertainty at work are repeatedly mentioned as stressful, and causes of negative wellbeing [24–27], as were redeployment and changes to working patterns. In contrast, and also found in quantitative studies, peer support and effective leadership help to reduce negative outcomes [21, 26, 28].

Lack of PPE was a concern for our participants, and has also been found to negatively impact of employee mental health in non-healthcare settings [29]. A number of qualitative studies have found similar results, that access to appropriate PPE is an important factor for HCWs’ mental and physical wellbeing [26, 27, 30], though our evidence suggests a temporal trend, with more concerns about this in the early months of the pandemic. There is evidence that access to PPE has been an issue internationally [28] and in previous pandemics [22]. Lack of PPE, or access to inadequate PPE, as well as changing guidance about its use, have impacted HCW trust in leaders and politicians, and there have been particular concerns about differential access to PPE, e.g. by women and Black, Asian and minority ethnic HCWs [30].

Fear of infecting loved ones was frequently reported as an important source of stress. This has been reported previously [21, 22], as have the challenges of balancing personal and professional stressors. For example, home working was experienced positively by some, providing time to slow down and connect with family, but others reported this to have been a negative experience, with social isolation, home schooling while trying to work, and physical health impacts all issues of concern [31, 32]. These findings were somewhat gendered, with women more likely to discuss home working than men, and this echoes previous qualitative research looking at female HCWs’ experiences of the pandemic [33].

Implications for policy and practice

Whilst the provision of clear and appropriate guidance on working practices is needed, employers of HCWs need to provide flexibility to cater to differing social needs and preferences. We identified the highly detrimental effects of uncertainty on staff wellbeing, and we suggest that proactive, clear communication from senior managers, and Government, is likely to positively impact on the mental health of staff.

We also identified that peer support and compassionate leaders were viewed as mitigating stress, and may prevent the onset of formal mental health problems. These views correlate well with findings from quantitative studies that have examined other trauma-prone occupational groups, which show that supportive junior leadership and effective peer support are associated with better employee mental health [34]. Training for managers may help to reduce the highly variable support participants reported in this study, and could be particularly beneficial given the fact that HCWs seem to want immediate support from those around them, rather than from well-being counsellors on a help line. Brief training for managers to improve their confidence to speak to their staff about mental health has been shown to be effective [35]. Alongside strong employee social bonds, our results also suggest Trusts should invest in ensuring that staff are able to access both wellbeing support measures (e.g. wellbeing spaces) as well as access to professional support from occupational health teams and, where needed, primary and secondary care services.

One aspect of our findings that has been consistently echoed in other research during the COVID-19 pandemic, and of previous pandemics, is the necessity for HCWs to be provided with appropriate and adequate PPE. While this is an issue that also needs to be addressed at the political level, healthcare leaders should be aware of the urgency in advocating for their staff on this issue. Another aspect repeatedly reported in other research is the finding that women appear to experience worse outcomes on many measures than men. Given that the NHS workforce is comprised predominantly of women (76%) [20], this finding should be considered by national and local NHS systems in order to redress inequalities. Existing qualitative research discusses the ways in which gendered organisational norms and structures can reinforce pre-pandemic inequalities (for example, greater unpaid care work undertaken by women) [33], and suggests that policies that support more equitable distribution of such labour are needed.

Strengths and limitations

Firstly, the most important strength of this study is that the data analysed are drawn from the largest UK study of HCWs, with a relatively representative sample from 18 English Trusts. Secondly, and relatedly, the STM methods used enabled us to include the views of a much larger sample than is typically possible in qualitative work, and allowed us to statistically examine how the topics raised differed according to participant characteristics. Thirdly, the open phrasing of the question allowed us to capture issues that are important to HCWs, and as such, we have generated a more holistic picture of the key issues than more tightly focussed qualitative studies can offer. This includes the interplay between the personal and the professional, such as the influences of personal circumstance, organisational culture and support, and the role of Government and wider society on respondents’ perceptions. Fourthly, the anonymity offered by the online data collection method means some participants may have felt able to be more honest about challenging topics than they would have done in a face-to-face or virtual interview.

However, the online nature of the survey highlights the first limitation of this research, that technical capability and access were required to participate, as well as English language skills, which may have excluded the most vulnerable. Secondly, as the free-text question from which the data was drawn was not compulsory, and was in addition to survey that covered a range of wellbeing topics, we might have been more likely to capture the views of participants with stronger or more extreme experiences. Conversely, the preceeding questions included measures of positive wellbeing as well as ill-health, meaning there was not a wholly negative framing. In addition, the wording of the free text question was neutral, inviting any and all perspectives. Thirdly, while the STM enabled inclusion of a large sample, the modelling is no more objective than the researchers interpreting findings; some topics covered a number of different areas meaning that researchers had to use their own judgement about which comments most clearly captured the main point of a topic, and about how to group the topics into meaningful themes. To mitigate this as far as possible, four researchers collaborated in an iterative process of individual and group interpretation. Fourthly, the data were collected over a relatively long time-period (April 2020-January 2021) so do not capture the mood of a specific era of the pandemic, though conversely this does mean we were able to map the themes identified over time.

Implications for future research

The method used in this study clearly demonstrates the benefits of large-scale qualitative data collection. Such methods can capture factors that survey designers might have missed as important to participants, and can add context, depth, and nuance to quantitative results. STM has not been widely used in health research to date, despite the frequent inclusion of free-text responses in large surveys. We advocate the use of these methods. Future research could helpfully address how to mitigate some of the limitations of these methods, such as how to capture the views of those with limited access to online surveys, or cultural differences in interpretations.