Types of studies

Consistent with the previous versions of this review, we limited inclusion to randomised controlled trials (RCTs) to evaluate intervention effectiveness. We only included completed studies published in peer‐reviewed scientific journals; abstracts without accompanying full texts and dissertations were excluded. 

Types of participants

We included studies in which the interventions were directed at healthcare workers who had not actively sought help for conditions such as burnout, depression, or anxiety disorder. This included all healthcare workers and trainees in any healthcare setting engaged in clinical work. We excluded studies in which any portion of participants were not doing clinical work, e.g. administrators, receptionists or when the outcomes were not reported separately for the participants who were doing clinical work. Personal caregivers who were family members or friends were excluded from this review.

Types of interventions

We included RCTs that evaluated the effectiveness of any type of intervention for individual healthcare workers aimed at preventing or reducing symptoms of stress. We excluded interventions targeting healthcare organisations because they are covered by the Giga 2018 review. Generally, four approaches to managing work‐related stress at the individual level can be distinguished:

1. focus one’s attention on the (modification of the) experience of stress (thoughts, feelings, behaviour); 

2. focus one’s attention away from the experience of stress by various means of psychological disengagement; 

3. alter work‐related risk factors on an individual level; and

4. combine two or more of the above.

Interventions such as mindfulness‐based stress reduction which focus on increasing awareness, regulation, cognitive flexibility and goals‐based behaviour directly related to stress were classified as type 1, whereas mindfulness ‐based meditation (Hofmann 2017; Tang 2015) that aim to shift attention away from the experience of stress and unpleasant thoughts was in type 2. 

We included all trials that compared the effectiveness of an active intervention with no intervention (including usual care), wait list, a placebo intervention, no stress‐reduction intervention or to another type of stress‐reduction intervention. 

The distinction between no intervention, wait list, placebo intervention, and no stress‐reduction intervention is not always apparent. We considered the comparison with a placebo intervention when participants were blinded to group assignment and both groups were told that they received a stress reduction intervention and the placebo intervention has no ‘active ingredient’. For instance when transcranial magnetic stimulation is compared to sham transcranial magnetic stimulation (Kim 2016). Trials with placebo arms were combined with those with no‐intervention controls, wait list controls and no stress‐reduction intervention controls in the meta‐analysis.

We considered the comparison with another type of stress‐reduction intervention when both groups received some kind of stress reduction intervention that was not part of regular care. In this comparison participants may or may not be blinded to group assignment. The comparison could include only different types of interventions, for instance, type 1 versus type 2 (psycho‐educational stress management (SMC) vs mindfulness‐based stress reduction (MSBR) (Errazuriz 2022). 

Types of outcome measures

We included studies that evaluated the effectiveness of interventions using validated and standardised self‐report questionnaires measuring symptoms of work‐related stress or burnout. We deemed all other outcomes that do not measure stress or its effects on individuals beyond the scope of this review. Examples of excluded outcomes are: risk factors for stress (such as workload, conflicts, support), coping skills, knowledge or attitude change, work performance, patient satisfaction and claims from clients, employee absenteeism and turnover.

We considered the following follow‐up times for outcome measurement:

• short term defined as up to and including three months after the intervention has been completed;

• medium term defined as more than three months up to 12 months; and

• long term defined as 12 months or longer.

Electronic searches

1. Cochrane Central Register of Controlled Trials (CENTRAL) (The Cochrane Library, 2013 to February 2022)

2. MEDLINE/PubMed (2013 to February 2022)

3. Embase (2013 to February 2022)

4. PsycINFO (2013 to February 2022)

5. CINAHL/EBSCO (2013 to February 2022)

6. Web of Science (2013 to February 2022)

Searching other resources

We examined the reference lists from included articles and reviews for any additional eligible studies.

Selection of studies

We used Covidence (Covidence 2022) for screening. Six review authors (ST, LE, AL, AT, KN, HM) independently screened titles and abstracts followed by full‐texts against the inclusion criteria. If there was any disagreement, the two review authors involved discussed this until disagreement was resolved.

Data extraction and management

Three review authors conducted the extraction of data by using a made‐to‐measure data extraction form in Covidence (ST, LE, AL) (Covidence 2022). Data extraction of the outcomes was done independently by the three review authors or researchers and students from the medical faculty of the University of Amsterdam. One review author checked all data extraction and reached consensus in cases of conflict. All questions concerning data extraction processes were resolved by discussion with all review authors.

Assessment of risk of bias in included studies

We used the Cochrane risk of bias tool (Higgins 2011) to assess the risk of bias in included studies. The tool includes the following assessment items: adequate sequence generation, allocation concealment, blinding, incomplete outcome data addressed, selective outcome reporting, and other bias. 

Measures of treatment effect

We plotted the results of each trial as means and standard deviations (SDs) for continuous outcomes. Because in many cases different instruments were used to measure stress, we transformed the means into standardised mean differences (SMDs). 

In many cases multiple similar outcome measures were used, or an instrument had several subscales but no summary measure. In case of multiple similar outcomes, we chose the outcome which we deemed to best represent a measure of stress symptoms in healthcare workers, such as the PSS (Cohen 1983). When study authors used subscales such as with the MBI (Maslach 1996), we chose the subscale that in our view best represented stress, such as the emotional exhaustion scale of the MBI (Maslach 1996). 

Unit of analysis issues

For studies that employed a cluster‐randomised design and that reported sufficient data to be included in the meta‐analysis and that did not make an allowance for the design effect, we calculated the design effect based on a fairly large assumed intra‐cluster correlation (ICC) of 0.10. Even though we did not find information for the ICC)for these types of studies we assumed that 0.10 would be a realistic estimate. We used studies from implementation research to support this assumption (Campbell 2001). We followed the methods stated in the Cochrane Handbook for Systematic Reviews of Interventions (Cochrane Handbook, Higgins 2022) for the calculations: design effect = 1+(M‐1)*ICC, where M is the average cluster size and ICC is the intra‐cluster correlation coefficient.

For studies with multiple study arms and one control condition, we combined groups to create a single pair‐wise comparison with the control condition.

For studies with multiple study arms and no control condition, we entered the first two study arms in the meta‐analysis. 

Dealing with missing data

 Where necessary, we sought missing data (means and standard deviations (SDs)) from authors. In total, 16 study authors either provided data that had not been published in their articles which enabled us to enter these studies into the meta‐analyses, provided clarification on their published article, or referred us to supplementary information (Barattucci 2019; Cohen‐Katz 2005; Dunne 2019; Dyrbye 2019; Errazuriz 2022; Gärtner 2013; Jensen 2006; Kline 2020; Moody 2013a; Oman 2006; Ozgundondu 2019; Pehlivan 2020; Sampson 2019; Sawyer 2021; West 2014; West 2021).

Where necessary and possible, we used WebPlotDigitizer (Rohatgi, 2022) to retrieve means and SDs from figures for the following studies: CezardaCosta 2019; Cheng 2015; Luthar 2017; Kline 2020 (control group only). 

When SDs were not reported we calculated them from other reported values according to the methods stated in the Cochrane Handbook (Higgins 2022). 

For West 1984 we took the means and SDs that resulted from the post‐hoc comparisons in the repeated measures analyses. For Norvell 1987, we took the post‐treatment values and calculated SDs based on the P value. We calculated a t‐value from this P value even though the authors used a Mann‐Whitney U test. For Shapiro 2005, we took the post‐treatment values and the F‐value reported by the authors. We calculated a t‐value and subsequent SDs by taking the square root of the F‐value as the t‐value. For Tsai 1993, we took the post‐treatment values from the figure reporting the results of the repeated measures' analysis. We took the reported P value belonging to the repeated measures' analysis as if it had resulted from a t‐test and calculated the SDs based on this t‐value. For Ewers 2002, we took the post‐treatment scores and the P values belonging to the independent t‐tests to calculate a t‐value and subsequently SDs. 

For Dahlgren 2022 and Gunasingam 2015, we calculated SDs based on 95% confidence intervals (CIs). For CezardaCosta 2019, Mao 2021 and Riley 2017 we calculated SDs based on the standard error (SE.) For Seidel 2021, the N per group was not reported, we assumed that there were an equal number of participants in the two study groups, i.e. 41 and 42. 

Lee 2021, Mealer 2014 and Ozgundondu 2019 reported their stress outcomes with a median and interquartile range (IQR). In accordance with the Cochrane Handbook (Higgins 2022) we requested mean and SDs. In the case we did not receive a response, we entered the median and IQR in the meta‐analysis, and we assumed that outcome data were normally distributed. 

Participants are included in the groups to which they were originally randomised, but missing data for participants were not included in the denominator.

In the case missing SDs were either not provided by the study authors or could not be calculated, these missing data were not imputed. 

Assessment of heterogeneity

We assessed heterogeneity in line with GRADE guidance (Schünemann 2013). We deemed an I² value of more than 50% to indicate considerable heterogeneity. When we identified heterogeneity, we tried to understand the reasons for the heterogeneity by exploring the options outlined in the Cochrane Handbook (Higgins 2022) and we investigated the presence of outlying studies. When the heterogeneity could not be explained, we downgraded the certainty of the evidence. In addition, we calculated the prediction intervals, to provide information about how much the true effect size varies across studies calculated with CMA Prediction Intervals.

Assessment of reporting biases

We avoided reporting bias by including studies and not articles. If multiple articles reported results from a single study, we consolidated all the data from all articles under one study ID only. We avoided language bias by including studies in any language. Because standardised mean differences (SMDs) are related to their standard error (SE) (Zwetsloot 2017), we did not use the SEs to generate a funnel plot instead we used the sample size as recommended by Zwetsloot 2017. The funnel plots were generated in STATA 17 (STATA 2022).

Data synthesis

We combined studies that we deemed sufficiently similar regarding participants, intervention, control, outcome and follow‐up time in one comparison.

We pooled the results statistically when the outcomes were similar concepts, such as perceived stress symptoms. Because many instruments were used, we used SMDs to combine the stress‐related outcomes using meta‐analysis. Not all instruments used one summary score, but presented the results of various subscales. In cases where there was no summary measure, we chose the subscale that best represented a measure of stress. For example, for this analysis, we used only the emotional exhaustion subscale of the Maslach Burnout Inventory (MBI). In this way, we considered the various stress symptoms scales to measure the same concept. We pooled the results using a random‐effects model. 

To interpret the effect size, the mean (23.6) and SD (12.2) on the MBI of the control healthcare worker population in Fiol DeRoque 2021 was used. The MBI‐Emotional exhaustion scale has a total score of 54. A score below 18 points is regarded as a low score on emotional exhaustion and a score above 36 as a high score on emotional exhaustion (Maslach 1996).

Subgroup analysis and investigation of heterogeneity

We conducted the following subgroup analyses and incorporated them in all comparisons:

• type of intervention (see Types of interventions for more details);

• length of follow‐up (see Types of outcome measures for more details);

• type of outcome (see Primary outcomes; Secondary outcomes for more details).

On top of those subgroups, we considered the subgroups mentioned in the original protocol, i.e. type of healthcare worker and duration and intensity of the intervention (Marine 2000). When considering those subgroups, we took into account that subgroup effects on top of the current subgroups in interventions and outcomes may prove spurious and may not explain all the variability in the extent of inconsistency, as most putative subgroup effects ultimately prove spurious (Schünemann 2013). 

Compliance

de Wijn 2022 found that stress management interventions for nurses in which the sample was exposed to the majority of the planned sessions reached greater effect sizes compared to interventions in which the compliance to the intervention/attendance to the planned sessions was lower. This finding should be interpreted with caution due to a lot of missing data (de Wijn 2022). However, we aimed to explore if the effect sizes based on studies in which participants attended 80% or more of the scheduled sessions would differ from the studies where participants attended less than 80% of the scheduled sessions. We added this in the "Differences between protocol and review". 
 

Sensitivity analysis

To assess the effect of risk of bias on the pooled results, we performed a sensitivity analysis in which we excluded studies with a high risk of bias and assessed whether this changed the results appreciably. We defined a study having a high risk of bias overall when we judged it to have a high risk of bias in three or more domains. 

Summary of findings and assessment of the certainty of the evidence

We used the GRADE approach to assess the certainty of the body of evidence for the intervention categories and comparisons most important for health decision‐making (Guyatt 2011). A priori, we decided that the comparisons of an intervention with no intervention are most important for decision‐making for the primary outcome of stress symptoms only for all three follow‐up times. Comparisons of one intervention versus another intervention were considered to be less informative. We downgraded the certainty of the evidence by one to three levels depending on the seriousness of the violations in each domain. We considered the risk of bias tables for each study in that intervention category to assess the risk of bias for an intervention category. We downgraded the certainty of the evidence if there were one or more limitations in the following domains: risk of bias, consistency, directness of the evidence, precision of the pooled estimate and the possibility of publication bias. All statements on the effects of interventions, such as in the summary of finding tables and the conclusion were worded in line with the recommendations on communicating findings when using the GRADE approach (Santesso 2020). Review authors ST and JB undertook GRADE, which was then also discussed with JR, RS, KN, LE and HM until consensus was reached.