This study aimed to investigate key predictors of psychological distress and subjective wellbeing among FHWs during the COVID-19 pandemic. Specifically, this study aimed to examine to what extent did demographics, health factors, COVID-infection factors, occupational factors, lockdown stressors and relational factors have an effect on psychological distress and subjective wellbeing in FHWs during the COVID-19 pandemic.

This study used cross-sectional data collected in the third timepoint of a longitudinal cohort study on FHWs’ working in Victoria, Australia. The data used in this study was collected between late January 2022 to early March 2022 using an online survey, administered on RedCap.

Participants were recruited state-wide in Victoria (Australia) from a large metropolitan health service (Eastern Health), and five major Australian healthcare associations: Australian medical association Vic (AMAVic), Australian Nursing and Midwifery Federation Vic (ANMFVic), Aged & Community Care Providers Association (ACCPA), Victorian Healthcare Association (VHA) and Health Services Union (HSU). Participants were eligible to participate if they worked in any capacity in Victorian (Australia) hospitals and in any of the following departments: the Emergency Departments, Intensive Care Units, COVID Wards, Hospital in the Home or Aged Care. These departments were chosen to represent frontline healthcare working sites and to recruit FHWs, as they posed the highest risk of COVID-19 infection in Victoria, Australia.

Study and ethics approval (HEAG 2020-296) was obtained from Deakin University High Risk Ethics Committee, Eastern Health’s Ethics Committee, and all partnering healthcare associations prior to the commencement of recruitment and data collection. Participants provided their consent and voluntarily participated. To ensure confidentiality, participant’s data were all de-identified prior to data-analysis.

During the data collection period, the general community were experiencing the easing of COVID-19 restrictions in Victoria, Australia (Premier of Victoria, 2022a). However, Victoria (Australia) was just coming out of the Omicron wave and largest ever surge of positive COVID-19 cases and hospitalisation since the start of the COVID-19 pandemic, which started in November and peaked in mid-January 2022 (State Government of Victoria, 2023). As such, the Victorian government declared a code brown emergency due to the influx of COVID-19 patients in hospitals (Premier of Victoria, 2022b). This meant that hospitals could configure or shutdown non-essential services to free staff, redeploy staff to higher priority departments, and ask staff to return from leave.

The survey collected data using individual items and scales from the Coronavirus Health Impact Survey tool (CRISIS) (Nikolaidis et al., 2021). Individual CRISIS items used in the survey included two demographic items: age and gender, two health related items: self-reported physical health and pre-COVID-19 mental health, two lockdown related items: self-reported stress with lockdown restrictions and time spent outdoors, and three COVID-infection related items: self-reported risk of infection from work, risk of infection from the community, and COVID-19 diagnosis.

A modified version of the COVID-19 infection worries scale from the CRISIS tool was also used in the survey. COVID-19 infection worries was measured using a validated scale score of four items measuring to what extent participants were worried about (1) being infected, (2) their family being infected, and the impacts of infection on their (3) mental health and (4) physical health. The original scale was found to have high internally consistency with a coefficient Omega of >0.8 and good unidimensionality (CFI > 0.95) (Nikolaidis et al., 2021). In our study, to reflect worries of infection more accurately, we removed the items that related to reading and talking about COVID-19, and hope that the pandemic will end soon. Our modified version had good internal consistency (McDonald’s ω = 0.88) and unidimensionality (CFI = 0.96).

To measure relational factors, two items from the CRISIS tool measuring (1) changes in relationship quality with regards to family and (2) changes in relationship quality with regards to social contacts were combined to avoid multi-collinearity in regression models and derive a single scale score of overall changes in relationship quality. This scale showed moderate internal consistency (McDonald’s ω = 0.60). Overall relationship stress was measured using a single scale score derived from the mean of two items measuring (1) relationship stress with regards to family and (2) relationship stress with regards to friends (5-point Likert scale, 1 = Not at all to 5 = Extremely). This scale showed good internal consistency (McDonald’s ω = 0.76).

To measure occupational factors, participants were asked to specify their occupation, whether they provided direct care to COVID-19 patients in the last 2 weeks, and a self-report rating of their supervisor’s support for mental health and wellbeing during the COVID-19 pandemic.

To measure psychological distress participants responded on a Likert scale from 1 (none of the time) to 4 (all of the time) to items on the Kessler Psychological Distress Scale- Six item (K6) scale, which has good internal consistency and reliability with a Cronbach’s alpha of 0.89 (Kessler et al., 2003). In this study Cronbach’s alpha and McDonald’s Omega was both 0.88 and CFI was 0.95 showing good internal consistency and unidimensionality. Scores were summed to produce a final score. A cut-point of 19 was used to indicate the presence of a probable mental health disorder. This cut-point has been shown to have high specificity (96%) but lower sensitivity (36%) in detecting health disorder diagnosed through clinical interviews (Kessler et al., 2003). The Kessler Psychological Distress Scale has also shown to have good measurement invariance across different age groups (Sunderland et al., 2013) and gender (Drapeau et al., 2010), as well being validated across a wide range of populations from different countries and culture (Donker et al., 2010; Oakley Browne et al., 2010; Andersen et al., 2011; Fernandes et al., 2011; Bu et al., 2017), including Australia (Slade et al., 2011).

To measure Subjective wellbeing participants completed the Personal Wellbeing Index (PWI), which has shown to have good internal consistency and reliability with a Cronbach’s alpha between 0.7 to 0.85 in the Australian population. Test–retest reliability over 1 to 2 weeks have shown to be good with an intra-class correlation of 0.84. In this study, internal consistency (Cronbach’s alpha = 0.92, McDonald’s Omega = 0.92) and unidimensionality of the PWI was good (CFI = 0.98). The PWI has seven domains that consistently form one factor, explaining 50% of the variance in the domain “satisfaction as a whole” in the Australian population (International Wellbeing Group, 2013). The seven domains are measured on a Likert scale from 0 (no satisfaction at all) to 10 (completely satisfied), which were totaled, averaged, and multiplied by 10 to produce the single mean PWI score. The PWI has been validated in Australia (International Wellbeing Group, 2013) and across a wide range of countries and cultures, showing good measurement invariance (Żemojtel-Piotrowska et al., 2017; Jovanović et al., 2019).

To address missing values observed in the data, missing data were multiply imputed in R using the MICE package (van Buuren and Groothuis-Oudshoorn, 2011). Fifty imputations were conducted and 5 iterations, which was conservative given that missingness did not exceed 11%. To assist imputations, auxiliary variables were used in the imputation. Outcome variables (i.e., K6 and PWI total scores) were not imputed and were only computed after imputation to reduce any potential bias during the imputation process. All analyses were conducted and pooled using Rubin’s rule (Rubin, 2004).

T-tests were used to compare sample means with normative data that was nationally representative of the Australian population and collected at similar timepoints as the current study. Normative data was taken from the Australian Unity Wellbeing Index national 2022 report (Khor et al., 2022) for PWI scores and the Australian National University national COVID-19 tracking poll (Biddle and Gray, 2022) for K6 scores. Univariate associations between predictors and outcome variables were analysed using unadjusted linear regressions.

Multivariable linear regression was used to estimate the total effect (i.e., includes both the direct and indirect effects) of each predictor on psychological distress and subjective wellbeing while adjusting for covariates. For each predictor, a minimally sufficient adjustment set (MSAS) was identified. MSAS is the minimum set of covariates to adjust in models to estimate the total effect and avoid distorted and biasing associations that can occur in typical regression models where covariate adjustment is based only on significance of results (Rohrer, 2018; Griffith et al., 2020). To identify MSAS for each predicter, directed acyclic graphs (DAG) were developed (Rohrer, 2018), using DAGitty¹ (Textor et al., 2016). The DAG is a graph that represents a theoretical framework around the causal relationships between variables, which is represented by arrows that are “directed” and imply a casual sequence. Once the theoretical framework was developed, Daggity applied the Pearl’s single and back door criterion (Pearl, 2009; Rohrer, 2018) to find the MSAS for each predictor’s model. When developing the DAG, the authors followed current recommendations (Tennant et al., 2021) and to ensure that omitted relationships were justified, all assumed relationships were tested for conditional independence in the data using polychoric correlations with the Lavaan package (Rosseel, 2012) in R based on current protocols (Ankan et al., 2021). The final relationships assumed in our DAG are presented in the supplements (Supplementary File S2). Multivariate linear regressions were then conducted individually, controlling for each predictors’ MSAS (see Table 1 for MSAS for each predictor).

Regression models were tested for heteroscedasticity using Breusch-Pagan test. Only subjective wellbeing models were found to be significantly heteroscedastic. Thus, utilising the sandwich package in R (Zeileis, 2004; Zeileis et al., 2020), Heteroscedasticity-Consistent (HC) standard errors, specifically the recommended HC3 estimator (Long and Ervin, 2000), were used in subjective wellbeing regression models. Using Gpower 3.1, post hoc power analysis showed that at alpha level 0.05, the sample size was large enough to detect statistical significance in effects ( R 2 ) above 4.7% for psychological distress and 4.6% for subjective wellbeing.

To identify key predictors and compare the unique contribution of each predictors’ total effect on K6 and PWI scores, relative importance analysis was conducted using relaimpo R package using the Lindemann-Merenda-Gold (LMG) method (Groemping, 2006). This analysis partitions variance explained in outcome variables by averaging over orderings and accounting for intercorrelations among covariates. This produces an unbiased R 2 for individual predictors that is decomposed and adjusts for other covariates, which is not typically accounted for in common estimates of decomposed R 2 that are biased by multi-collinearity (Groemping, 2006; Tonidandel and LeBreton, 2011).

Significance thresholds were all set at 0.05 for all analyses. Confidence intervals (95% CI) were computed by bootstrapping based on 1,000 bootstrap resamples from each of the 50 imputed datasets in R (Schomaker and Heumann, 2018) and pooling them together using Rubin’s rule (Rubin, 2004).

One hundred and seventy-two Victorian frontline healthcare workers completed the survey. Two participants that noted “other” as their gender and three responses that did not identify their gender were removed due to extreme uneven distribution in analyses with gender as a predictor, leaving a total of 167 responses. Details of participants’ characteristics are displayed in Table 2. In brief, the majority of respondents were female (88.6%), nurses and midwives (70.1%), and working in ICU (32.9%). When compared to previous data obtained from a large state-wide health service (Holton et al., 2020), the sample shows a good representation of the nursing population but an overrepresentation of allied health staff and under representation of physicians. Most respondents worked in public hospitals (92.3%). Only 17.4% of participants had been diagnosed with COVID-19 since the start of the pandemic. Approximately half of participants (55.1%) worked directly with COVID-19 patients.

The K6 measure was completed by 161 participants. Mean psychological distress among participants was 13.70 (SD = 4.96), which was significantly higher (t = 5.4, p < 0.001) compared to normative data (Normative mean = 11.6) that was collected during the same period as this study (Biddle and Gray, 2022). Using the recommended cut off of 19, 18.1% of the sample screened positive for a probable mental health disorder.

The PWI measure was completed by 163 participants. Mean SWI was 67.17 (SD = 18.22), significantly lower (p < 0.001) than aggregated normative data (normative mean = 75.5; Khor et al., 2022). Overall, 25 participants had SWI mean scores lower than 50, indicating that 15.3% of participants were experiencing concerningly low levels of wellbeing. Sample means were also consistently significantly lower (p < 0.05) than normative data in all sub-domains except for standard of living (p < 0.05; Figure 1). Based on Cohens’ D effect sizes, these significant mean differences were small (Cohens’ D: 0.2–0.5) for all sub-domains except for future security (Cohens’ D < 0.2), which was negligible.

The results for the adjusted and unadjusted regression analyses, and relative importance analysis for psychological distress are displayed in Table 3. In the multivariate models adjusting for the MSAS, significant total effects were found for age, both health factors, supervisor support, all COVID infection related factors and relationship stress. Results show that younger age, higher perceived work infection risk, high perceived community infection risk, more COVID infection worries, a previous positive COVID-diagnosis and more relationship stress were significant risk factors and associated with higher levels of psychological distress. Better supervisor support, better pre-covid mental health, and better physical health were found to be protective and significantly associated with lower psychological distress.

Based on the results of the relative importance analysis, COVID worries ( R 2 =17.66%) explained the most unique variance in psychological distress and was considered a large effect based on Cohen’s proposed magnitude for R 2 effect sizes. All other significant predictors explained moderate amounts of unique variance in psychological distress, R 2 ranging from 2.76 to 8.63% (Predictor ranking displayed in Figure 2).

The results for the PWI and its sub-domains’ regression analyses are presented in Tables 4–6. After adjusting for the MSAS, significant total effects on PWI were found for both health factors, occupation, supervisor support, work infection risk and changes in relationship quality. Results show that a higher rating of pre-COVID mental health, physical health and supervisor support were all significantly associated with a higher overall subjective wellbeing. Positive changes in relationship quality were also associated with a higher subjective wellbeing. Being a nurse was associated with a lower subjective wellbeing, however, only when compared to allied health staff. Higher perceived work infection risk was associated with lower subjective wellbeing. When analysed within the PWI sub-domains, pre-COVID mental health, physical health and supervisor support were consistently significant predictors in all domains except for personal relationships, where supervisor support had no significant effect. In terms of occupation, being a nurse was associated with a lower satisfaction in standard of living, and future security when compared to allied health, and a lower satisfaction of health when compared to non-medical staff. Nurses also had lower satisfaction in their achievement in life when compared to physicians. Work infection risk was only a significant predictor of standard of living, personal safety, and future security. Relationship quality was only predictive of standard of living, personal relationships, and community connectedness.

Based on the relative importance analysis, physical health had a large effect ( R 2 = 20.60%) and explained the most variance in overall subjective wellbeing. The other significant predictors explained moderate amounts of variance in subjective wellbeing, ranging from 2.67 to 11.41% in R 2 . Physical health also consistently explained the most variance in all domains, except for achieving in life, where pre-COVID-mental health explained the most variance. Ranking of predictors variance explained in overall subjective wellbeing and in its sub-domains are shown in Figure 3.

When interpreting findings in this study, several limitations should be considered. Firstly, as a cross-sectional study, it is not evident that the distress observed in this study is indicative of an acute reaction or persisting distress, which should be investigated further in longitudinal studies. Additionally, causal links between variables should be interpreted with caution. While the use of DAGs in this study provides a framework around causality between investigated variables, it relies on the assumptions in the DAGs. Other models may exist and the model in this study is not intended to be a proposal for a theoretical framework around FHWs’ mental health and wellbeing. The use of the DAG in this study is intended to be a way to systematically adjust for covariates to estimate effects and provide transparency around assumed relationships (Ferguson et al., 2020). Secondly, mediation analysis was beyond the scope of this study, however, the results points to several mediating relationships among the investigated variables and should be considered when interpreting results and in future research. It is thus recommended that future research investigate these mediating relationships further through structural equation modelling or mediation analyses. Thirdly; there was a low representation of physicians, which may have impacted the generalisability of results for this cohort. Lastly, due to the small sample size, precision of estimates may be weak, thus effects and mean differences with population norms should be interpreted and generalised with caution.