Participants and procedure
This observational Ecological Momentary Assessment (EMA) study was conducted in Switzerland between October 01, 2019 and July 31, 2020. Due to the restrictions on research with human subjects at the beginning of the SARS-CoV-2 pandemic, the recruitment of participants had to be interrupted in Spring 2020. To be eligible for participation in this study, participants had (a) to be 18 years or older, (b) to live in Switzerland, (c) to work as or to study to become a health care worker, (d) to be in daily contact with patients, and (e) to have access to a smartphone running the iOS or Android operating system. Members or students of the following professions were defined as health care workers: physicians, psychotherapists, nursing staff, physio—, occupational—or speech therapists, paramedics, or medical admin staff with patient contact. Participants were recruited through a study website, posters and flyers distributed at health care facilities and universities, and through personal contacts of the study team members. Subjects willing to participate in this study underwent a telephone interview, during which the details of the study were explained, and eligibility criteria were assessed. Informed consent for participation was obtained from all participants. Participation in this study was reimbursed 20 CHF (approx. USD 20) when at least 50% of the EMA and all questions from the first and the last test day were completed. An additional CHF 20 were reimbursed when at least 80% of the EMA ratings were completed. The study protocol was approved by the ethics committee of the canton Zurich (BASEC-Nr. 2019-01020).
The study lasted a total of 18 days and was split in three phases. On Day 1, cross-sectional data and demographics were assessed. Subsequently, the 17-day long EMA period lasted from Day 2 to the evening of Day 18. Finally, the level of burnout was again assessed cross-sectionally on the evening of Day 18. During the EMA period, participants were prompted to rate their momentary emotions on their smartphone five times a day (between 6am and 6 pm on working days and 8am and 8 pm on weekends). These prompts were delivered pseudorandomly, meaning that each prompt was delivered randomly during a predefined 30-min period. Once prompted, participants had 90 min to provide their ratings. All data was collected using the LifeData smartphone application and platform.
The sample size rationale was built upon our hypothesis and methodological considerations. First, the number of observations needed to estimate the planned subject-specific network models cannot readily be obtained. Still, a simulation study indicated that 50 repeated measurements per individual is supposed to be sufficient for such a network consisting of 8 nodes . Thus, based on the analytic procedure (resulting in a loss of 20% of the daily measures due to the exclusion of the overnight lag, see below) and an assumed missing data rate of 20%, we set the length of the EMA period to 17 days, resulting in a total of 85 assessments of emotions (and 54 assessments when assuming a 20% missing rate). Second, we calculated the required sample size to test the correlation between network density and burnout based on the following assumptions: (a) a medium effect size (Spearman’s ρ = 0.3), a conservative assumption given the previously reported large effects [18, 19], (b) a one-tailed, positive correlation, (c) an alpha level of 0.05, and (d) 80% power. This resulted in a target sample size of N = 64, using G*Power .
Basic demographic data was obtained from all participants including gender (female, male, and other), age (in years), profession (physician, psychotherapist, nursing staff, physiotherapists, occupational therapists, speech therapists, paramedics, and medical admin staff), average weekly work hours, and hours working overtime.
Maslach burnout inventory: general survey
We used the Maslach Burnout Inventory-General Survey (MBI-GS) to determine burnout severity. The MBI-GS consists of 16 items, structured in three dimensions. Five items each refer to emotional exhaustion (dimension 1) and depersonalization (dimension 2), whereas professional efficacy (dimension 3) was assessed with six items. All items were rated on an 8-point Likert scale (ranging from 0 indicating never and 7 indicating very strong) with the items assessing professional efficacy being coded reversely. Following common practice (e.g., , burnout severity was assessed with a total score, accounting for the different weights of the burnout dimensions and the number of items resulting in a total score ranging from 0 to 6.Footnote 1 The questionnaire was adapted to cover the study period (i.e., 17 days) instead of a full year. Although several German translations of the MBI-GS do exist, none has officially been validated. Nevertheless, the chosen translation has been widely used in previous research (e.g., [27, 28] and demonstrated good internal consistency in this study with Cronbach’s alpha of 0.91 for the MBI-GS total score.
Ecological momentary assessment (EMA)
The emotions assessed included four with positive (happy, satisfied, relaxed, being full of energy) and four with negative valence (frustrated, stressed, worried, exhausted). All emotions were assessed using single items (e.g., How frustrated are you at the moment) and a visual analogue scale ranging from 0 (not at all) to 100 (absolutely). Only the emotions with negative valence were included into the primary analysis. However, network density of a network consisting of only positive emotions was calculated in a secondary, not preregistered analysis as well.
The analytic plan to test our hypothesis was structured in three parts: (a) data preparation, (b) network estimation and density calculation, and (c) correlating density and burnout severity. The three steps followed the analyses as suggested by the developers of these methods [18, 29], and by the authors of previous studies [19, 21, 30].
Data preparation and missing values
In a first step, the data obtained with the LifeData app was relabelled and reorganized into two datasets, one containing the data of the assessments of the first and last day and the other the EMA data. Next, participants with missing data in the burnout assessment at Day 18 and those with more than 20% missing EMA data were excluded from the analysis. Finally, after the conductance of the analysis outlined below, the density and burnout severity measures were checked for outliers (defined as diverging more than 3 SD from the sample mean), which were excluded from further analysis. All data analysis was conducted in the R environment using, among others, the packages mlVAR , and tidyverse . The data and the analytic code are available in the Additional files 1, 2 and 3.
Network estimation and density calculation
Person-specific temporal networks were estimated for all individuals using the two-step approach of the multilevel vector auto-regression model (mlVAR; [18, 29]. The mlVAR is an extension of the vector auto-regression model (VAR) to model the individual networks of a group of subjects. A lag-1 VAR model estimates to what extent the rating of each emotion at one timepoint (t0) is predicted by the ratings of all emotions (including itself) at the previous rating (t−1). Formally, each emotion is regressed on the lagged values of all other emotions (cross-lagged effects) and its own lagged values (autoregressive effects) using a series of univariate regression models [18, 29].
The mlVAR model extends this VAR model into a multilevel modelling framework, in which the average auto- and cross-lagged effects are obtained for the whole sample (fixed effects) but are also allowed to vary across the subjects of a population (random effects; [18, 29]. For this purpose, the variables included into the network were within-subject (i.e., person-mean) centred prior to the analysis. Furthermore, given that the lag, the time between two consecutive prompts, is assumed to be constant, the lag from the first measurement on one day onto the last measurement of the day before was removed from the analysis . Following the definition of previous studies, network density was determined as the average of the absolute values of the cross- and auto-lagged effects of emotions [18, 30, 33]. Therefore, the effects of each subject’s network were extracted, their absolute value was summed and then divided by 20, the total number of the effects in each network. Missing data was handled with listwise deletion.
Correlation between density and burnout severity
The correlation between the density of the person-specific networks and burnout severity was quantified using both Pearson's r and Spearman's ρ. Both correlations were tested for significance with the help of a one-sided test (assuming a positive correlation). The alpha level was set to 0.05.