Study setting, design, and period
We used a cross-sectional study design in nine selected hospitals located in six zones of the Central and Southern Ethiopia Regional States from 16th January to 28th February 2023. The healthcare facilities serve multiethnic, multicultural, and lingual populations located in six zones: Hadiya, Halaba, Kembata, Wolaita, Gurage, and Silitie zones. These zones are 132 to 328 away from the capital city of the country, Addis Ababa. The total population residing in the six zones is estimated to be more than approximately 9, 201,127.
The delivery of Ethiopian health care services involves three-tier systems: primary healthcare units, secondary healthcare units, and tertiary care. Primary care includes health posts, health centres, and primary hospitals. Health posts are healthcare facilities where health extension workers offer preventive universal health services at the Kebele level. Health centres additionally provide curative services. Primary hospitals offer inpatient and ambulatory care including emergency surgical services. Secondary healthcare units include general hospitals, receiving referrals from primary hospitals. Tertiary care hospitals include referral hospitals such as comprehensive specialized hospitals and teaching referral hospitals. Comprehensive specialized hospitals are referral hospitals that can provide advanced regional healthcare services and receive referrals from general hospitals and in some cases primary hospitals39.
During the data collection period, there were 317 public health centres and above public health facilities (health centres, primary, general, and referral(comprehensive specialised and teaching referral hospitals): 293 functional health centres, 16 primary public hospitals, 4 general hospitals, and 4 tertiary care hospitals (1 comprehensive specialised hospital, and 3 teaching and referral hospitals) during the previous southern Ethiopia administrative regions. Our study selected primary and those above based on two reasons:1 assuming that work-related psychosocial factors such as job demands, job demands, job control social support (resources) and workload are more prevalent and may increase the vulnerability of healthcare workers to common mental health symptoms;2 potential resource constraints to include health posts, and health centres due to large and dispersed geographic coverage and a large number of health facilities About 7,790 healthcare professionals were employed in the public health facilities in the six Zones from former central southern Ethiopia during the data collection period. Of these, nurses, midwives, public health officers, general practitioners, and medical laboratory professionals were 3905, 1467, 1138, 764, and 516, respectively.
Source and study population
The source populations were all healthcare professionals directly engaging in patient care and paramedic healthcare services in public primary hospitals, general hospitals, and tertiary hospitals. Healthcare workers who were included in the study from randomly selected healthcare facilities in each stratum composed of the study population.
Sample size determination
We calculated sample sizes for two PhD study objectives:1 to determine the prevalence of the three outcomes (occupational stress, occupational depression, and job anxiety) and their sociodemographic, health and behaviour, and work environment factors, and2 to assess the structural causal link between work-related psychosocial factors and the three mental health outcomes (symptoms of occupational stress, occupational depression, and job anxiety) using a structural equation modelling (SEM). Accordingly, we calculated a sample size for both study objectives and took the sample size calculated for the second objective because it provided a larger sample size. Before performing the logistic regression and analyzing the proportion of healthcare workers with occupational mental health symptoms, we planned to check the validity and composite reliability indices for our latent constructs (scales of occupational stress, occupational depression inventory, and job anxiety), which require a larger sample size. Therefore, we used the larger sample size estimated for the second PhD objective.
Following, the assumptions of structural equation modelling, a sample was calculated using a lower bound sample size for SEM developed by Westland JC40. Accordingly, in sample size calculation for the second PhD work, we used 86 observed variables and 10 latent variables with a statistical power of 80%, a medium anticipated effect size of 0.3, and a probability level (α) of 0.05 before our study. We found a minimum recommended sample size of 928 healthcare workers. We used a design effect of 1.5 to accommodate the error variation in the multistage sampling and a maximum of 10% of the nonresponse rate was considered from a nonresponse rate reported by 14 studies conducted among healthcare workers elsewhere in Ethiopia32. Finally, the minimum recommended a priori sample size for the study was 1531 healthcare workers and for the current study, the sample size was larger. Given that the sample size was estimated using a large sample assumption i.e. structural equation modelling, assures a more representative estimate of the target population and the conclusions drawn from our findings.
Variables and measures
Dependent variables
We measured occupational stress using a brief perceived occupational stress scale (POS-Scale)41. This tool assesses occupational stress in a cause-specific manner to determine whether psychosocial work stressors increase the likelihood of causing mental distress at work. The tool has been validated and has good psychometric properties for the public workforce including healthcare workers in the European context41(Cronbach’s alpha of 0.82). The scale has four items rated on a 5-point Likert scale, ranging from 1 (strongly disagree) to 5 (strongly agree). The item answers are averaged to provide ranges from 1(lowest perceived occupational stress) to 5(highest perceived occupational stress). Participants are asked to provide ratings that reflect their work stress in the last 6 months.
We used an occupational depression inventory24 to capture depressive symptoms in a work-ascribed manner. This tool has been validated to measure depressive symptoms from cause-specific perspectives or via etiological approaches adapted from clinical diagnostic criteria21. Participants rated the extent of the symptoms they had experienced in the past two weeks on a 4-point Likert scale from 0(never or rarely) to 3(nearly every day). Instead of assessing depressive symptoms in a “cause-neutral” manner, each ODI item assessed causal attributions to respondents’ work/job (e.g., “My experience at work made me feel like a failure”). The tool presented excellent reliability (Cronbach’s α of 0.916 )24 and showed strong reliability among South African employees including healthcare workers (Cronbach’s α was 0.926)42.
We used a short version of the Job Anxiety Scale (JAS-15)43 to measure work-ascribed general anxiety. The JAS-15 has five subscales each with three items: stimulus-related anxiety and avoidance behaviour (SAA), social anxiety and impairment cognition (SAIC), health and body-related anxiety (HBA), cognition of insufficiency (CI), and general job-related worrying (JRW). The tool has shown excellent internal consistency (Cronbach’s α was 0.942) in other studies43. After checking responses during the adaptation and pretest phases, we used a 5-point Likert scale ranging from 0 (no agreement at all) to 5 (total agreement) with no reverse-scored items. Like occupational depression, each JAS item indicated work factors as causal attributes to respondents’ jobs instead of assessing anxiety symptoms from a “cause-neutral” perspective (e.g. “After work, I hurry up more than others just to get away from that place.”).
Independent variables
We measured sociodemographics such as sex, age, marital status, educational level, monthly income, and other work environment characteristics such as work experience, violence at work, hours of work per week, and individual behavioural factors such as cigarette smoking, Chat or Khat (Catha Edulis) consumption, alcohol consumption, coffee consumption before sleeping or at night, and experience of planned physical exercise were also collected using standardized questionnaire adapted from previous literature.
Health and behavioural-related variables such as stressful life- threatening events (LTEs) in the previous 12 months, current general perceived health, history of disease and or injury in the previous 12 months, history of taking anti-depressants or painkillers in the previous 6 months, and sleep quality were collected. We assessed LTEs using an LTEs questionnaire containing 10 items with a “Yes” or “No” response, and validated elsewhere44 with a good reliability index (Cronbach’s alpha of 0.86). General perceived health was assessed with single-item questions adapted from validated tools elsewhere45, evaluating self-health by asking “How would you evaluate your health in general?”. Then, participants responded on a 5-point Likert scale ranging from 1 (poor) to 5 (excellent).
Sleep disorders (insomnia) were assessed using 3-item questionnaires adapted from the insomnia severity index (ISI)46 assessing sleep dissatisfaction, sleep interference with daily function, and sleep difficulty. This tool has good psychometric properties with reliability indices, and its Cronbach’s alpha ranges from 0.89 to 0.90. Accordingly, participants were asked to rate their extent of sleep in the previous 2 weeks from two items of sleep dissatisfaction and interference with daily function ranging from 1 (very satisfied) to 5 (very dissatisfied). One of the items for example was “Over the past two weeks, how satisfied are you with your current sleep pattern? Whereas, we assessed sleep difficulty by asking participants to rate their experience with at least one issue, such as difficulty falling asleep, staying asleep, or waking up early, on a scale from 0 (none) to 5 (very severe) .
Adaptation and pretesting procedures
Despite assuming that healthcare workers have an understanding of items of constructs when compared to the general population, the perceived and actual experience of common mental symptoms as a result of exposure to work-related psychosocial stressors may share a unique historical, social, and cultural phenomenon. Therefore, we planned to make sure that healthcare workers understand each item by clearly associating mental health symptoms with work-related issues and responding to them accordingly.
Accordingly, we adapted self-report measures for our setting following the cross-cultural adaptation steps of measuring instruments to reduce construct bias, method bias, and item bias47,48. First, two bilinguals (one MSc in community psychiatry and the other from nonclinical professions, MSc in hygiene and environmental health) translated the instruments from English into Amharic. Second, two translators and one of the authors (PhD candidate) discussed this through cross-checking with the original language of each item and reached a consensus. Third, one translated version was provided to an English university lecturer with an MA who was fluent in Amharic, and experienced in the study area. Another copy was given to another MA in English, who was also fluent in Amharic and well-versed in the study area, for back translation to English. Fourth, together with 4 translators, one of the authors and 3 invited experts (1 MSc in psychology, 1 MPH in epidemiology, and 1 MA in anthropology and protestant religion preachers of the areas) to check local cultural and language, experiential, and conceptual equivalence and produced a final questionnaire for pretesting. Finally, the questionnaire was pretested among 80 healthcare workers who were working in public health centres not included in the study by instructing them to document and provide feedback. After receiving the feedback, many points creating confusion were corrected. For example, the left-and right-hand responses of the 7-point Likert scale of the JAS were confusing for pretested participants. Hence, we changed it to a 5-point Likert scale and the participants were instructed to complete the scale from 0 (no agreement at all) to 5 (total agreement).
Inclusion and exclusion criteria
Healthcare workers who had worked for at least 6 months in clinical activities and paramedic activities at the current hospital or who were transferred from other hospitals were included in the study. We defined a minimum of six months because one of our outcome variables (occupational stress) requires at least 6 months of experience in any clinical and paramedical activities, and to reduce a recall bias. Those medical residents who were on speciality training for more than six months were included in the study. Those who were on annual leave or who changed professional activities for any reason during the data collection period were also excluded from the study. Healthcare workers who had been out of work for more than 2 weeks were also excluded from the study.
Sampling procedures
During the data collection period, there were 4 referral hospitals (1 comprehensive specialized teaching hospital, 3 teaching and referral hospitals), 4 general hospitals, and 16 primary hospitals. We selected 9 hospitals following the following procedures. First, we stratified hospitals into three groups: primary, general, and tertiary (referral hospitals include comprehensive specialised hospital, teaching and referral) hospitals. Within each stratum, we considered hospitals as clusters assuming that most of the working units and healthcare activities share nearly similar characteristics regarding work-related stressors. Second, we randomly selected 75%, of tertiary hospitals, 50% of general hospitals, and 25% of primary hospitals assuming that the complexity of work-related stressors, diverse units, and activities across specialities increase variability. Accordingly, 3 tertiary hospitals, 2 general hospitals, and 4 primary hospitals were randomly selected from each stratum. Third, the sample size was proportionally allocated to the size of each selected hospital. To ensure the representativeness of the sample, the sample size allocated for each hospital again was allocated to each working unit (the smallest unit of hospitals), which was stratified by professional category or speciality area obtained from the hospital’s directorate and human resource office. Whenever applicable, samples allocated for the work units were also allocated to speciality areas. Finally, all healthcare workers selected in each work unit were invited to participate until the allocated sample size was met. The schematic representation of the process of sampling procedure with proportionally allocating process and total samples before the data collection is displayed in Fig. 1.
Schematic representation of multistage sampling procedures among healthcare workers, central and southern Ethiopia, 2023.
Data collection procedures
The data were collected using a self-administered structured questionnaire technique. We trained eleven data collectors about the purpose, content of the measurement tool, sampling procedures, how to support participants with inquiries about filling out the questionnaire, precautions during the distribution and returning it upon completion, and other ethical aspects. After the onsite training, each data collector took a minimum of 8 questionnaires to be pretested among healthcare workers who were working near each public health facility selected for the study. A week later, we collected feedback from each data collector. Finally, we duplicated and distributed a questionnaire for each study facility.
Following the sampling procedures and exclusion criteria, data collectors reached out to each health worker in each working unit of each selected hospital during working hours. Upon reaching out, participants were briefed about the purpose, benefits, and risks of the study, confidentiality, extent of autonomy, and justice for participating in the study. Each unit manager was individually informed of the purpose and the aim of the study. The data collectors also received written consent from each participant on the first page of the questionnaire. For those voluntary healthcare workers who were unable to fill out the questionnaire during the same day and time, the data collectors provided adequate time based on the consensus between the data collectors and the participants. We trained and assigned 3 collaborators for referral hospitals and 1 for primary hospitals from the facility’s unit heads or representatives to support the data collectors. In cases where healthcare workers took a questionnaire and lost it due to any inconvenience, the data collectors redelivered the reserve questionnaire to be completed. Finally, a completed questionnaire was returned to the data collectors to check its completeness.
Data processing, and analysis
Before the data entry, the questionnaires with incomplete responses on items were discarded. We used EPI-info version 7 to enter the data and exported it to SPSS version 25 for descriptive and ordinal logistic regression analysis. The SPSS file was exported to JAMOVI software for confirmatory factor analysis. The data were cleaned and checked for consistency and completeness. Initial data exploration was conducted for all variables ranging from socioeconomic and other work environment variables to each of our outcome variables. Outliers and items missing were checked. Outliers were controlled during the data entry by restricting the minimum and maximum values of the data, when necessary. Missing values of responses are substituted by means or medians depending on the nature of data distribution assuming that missing is completely at random (e.g. age). We performed univariate analysis to describe our study population. For sleep disorders, the sum of the mean scores of the three items was categorized into four from none to severe sleep disorders, and the sum score of LTEs was categorized into four for ease of data analysis.
We performed post-hoc measurement validity, reliability, and confirmatory factor analysis (CFA) to check whether items that measured outcomes of interest were in alignment with previously validated tools elsewhere24,41,42,43. The main purpose of performing CFA for each of our latent outcome variables was to check whether the measurement model fits the data, and to check the validity and reliability of the scales before determining the prevalence of mental health symptoms and their associated predictors. We performed the CFA using a structural equation modelling module of JAMOVI version 2.3.8 to allow us to use various robust estimation approaches and additional composite reliability and validity statistics. To handle data of skewed distribution with ordinal item responses, we used diagonal weighted least squares (DWLS) to estimate our parameters in the CFA. We fit POS and ODI as a unidimensional scale. Whereas, we fit the JAS with its five subscales to check the factor loadings of each subscale item. After checking each subscale item, we used the JAS as a unidimensional measurement of categorizing the score to determine the overall prevalence of job-related anxiety symptoms and its factors. Accordingly, all the CFA measurement models with three outcome variables had acceptable fit indices confirming the previous models for measuring the perceived occupational stress scale (POS), occupational depression inventory (ODI), and Job anxiety (JAS) fit for our data.
The reliability statistics also show that our measurements have good reliability scores. The ordinal alpha (α) (internal consistency based on polychromic correlations among items) was reported to account for unequal factor loadings and unequal error variances. The ordinal alpha (α) value for POS was 0.922 and for ODI was 0.958. The ordinal alpha (α) for subsequent JAS subscales ranged from 0.891 to 0.956. Omega hierarchy (Ѡ2) was used for internal consistency, accounting for item-specific weights for POS and ODI of 0.884 and 0.917. Regarding convergent validity, the average variance extracted (AVE) values for the indicators explained by POS was 0.758 and by ODI was 0.734. The average proportion of variations accounted for in a set of indicators by each subscale of JAS ranged from 0.746 to 0.878. For discriminant validity of the JAS subscales, the heterotrait-monotrait (HTMT) ratio of correlations between two pairs of subscales was less than 0.955. Specifically, the HTMT ratio was 0.954 for “job-related worrying and cognition of insufficiency”, and 0.929 for “health and body-related anxiety and cognition of in-sufficiency”., The values are larger than the recommended indicating one item was not well discriminated the concept than the other. However, the problem of this correlation could not be necessarily a problem for our estimation of the magnitude of overall job anxiety because the overall JAS has good reliability and convergent validity.
The univariate descriptive summaries, unstandardised, standardised factor loadings, the reliability and validity statistics, additional CFA information, and the covariances and correlation matrices of the measurement model for the three outcome variables are displayed in the supplementary Table S1-S4. The univariate descriptive summaries are displayed for the latent constructs of our outcome variables in supplementary Table S1. The standardised factor loadings, reliability and validity information are displayed in supplementary Table S2. Additional measurement model summaries are displayed in supplementary Table S3. Covariances and correlations matrices are also displayed in supplementary Table S4.
Following the confirmatory, reliability, and validity analyses, we made a dimensional severity cut-off of points for all the outcome variables. Accordingly, as the previously validated ODI-9 is based on the patient-health questionnaire, the scoring system is similar. Therefore, for occupational depressive symptoms, sum scores of 0 to 4, 5 to 9, 10 to 14, 15 to 19 and 20 to 27 were used to indicate “almost no symptom”, “minimal”, “mild”, “moderate” and “severe” symptoms, respectively. Because we could not find standard cut-off values for the job anxiety scale and occupational stress scale, we used percentile rank cases to show the severity of each disease. We followed the same procedure for job anxiety with similar labelling for job anxiety, 0, 16 to18, 19 to26, 27 to39, and 40 to 75 were categorized “no symptom”, “minimal”, “mild”, “moderate” and “severe” symptoms, respectively. For occupational stress, we used four classifications minimal (4 to 8), mild (9 to 10), moderate (11 to 12), and severe distress (13 to 16) because the “almost no stress” classification is not possible in stress theory.
Multiple variable ordinal logistic regression was done to assess independent sociodemographic, health, behavioural, and work environment factors with each common mental symptom (occupational stress, occupational depression, and job anxiety). We used multinomial probability distribution with a link function of cumulative logit from a generalized linear model because of failure to parallel lines test assumptions. Variables with a p-value of 0.25 or less for binary ordinal logistic regression were eligible for multiple ordinal logistic regression. Finally, model fit statistics (chi-square with degrees of freedom and p-value), tests of parallel lines, and an adjusted odds ratio of 95% are reported. The odds of moving from one category to the next were all checked for the explanatory variables (p-value < 0.05) based on model threshold values.
link