To test our hypotheses, we use unique multilevel data from the European Sustainable Workforce Survey (ESWS)  This survey, first conducted in 2015-16, contains data on employees, their teams, and the organisations they worked for in nine European countries (Bulgaria, Finland, Germany, Hungary, the Netherlands, Portugal, Spain, Sweden, and the UK). Organisations were approached using stratified random sampling based on sector (manufacturing, health care, higher education, transport, financial services, and telecommunications) and size (1–99 employees, 100–249 employees, and 250 + employees). When the random sample did not yield enough participants in a stratum, referrals and personal connections were used to complement the selection. Within each organisation, a contact person (usually the human resources manager) decided on whether the organisation wanted to join the study. Upon a positive response, at least three teams were selected in consultation with the HR manager. All employees, and the manager, of those teams were addressed at work to fill out the survey in their own language. The HR manager provided information about the organisation as a whole. This data structure enabled us to construct the networks of employees who worked together in the same teams, which is necessary for our purposes.
Our study incorporates data from the second round of the ESWS, due to its detailed information on employees’ lifestyle choices, which is not included in the first round. Data for the second wave was collected in 2018-19. Organisations from the first round were invited to participate once again, and 13 new organisations also joined the study under the same selection and survey completion procedures as the first round. All participants provided written informed consent prior to filling out the survey. The response rate in the second wave was 89% among HR managers, 68% among team managers, and 54% among employees, resulting in a sample of 4345 employees working as part of 402 teams in 113 organisations.
Because our study addresses three different behaviours, we created three analytical samples: one for fruit consumption, one for vegetable consumption, and one for physical activity. For each analytical sample, we first excluded employees who had missing values on any of the variables included (Nfruit consumption=1197, Nvegetable consumption=1162 and Nphysical activity=1314). Most of these missing values were for the dependent variables.Footnote 1 Since we are interested in employees’ networks we excluded employees who had no colleagues (Nfruit consumption=37, Nvegetable consumption=38 and Nphysical activity=39). Our final analytical samples were N = 3111, N = 3145 and N = 2992 for fruit consumption, vegetable consumption and physical activity, respectively.
The measurement of our dependent variables is similar to questions in the European Social Survey . Fruit and vegetable consumption were measured by asking respondents how often they ate fruits, including frozen fruits but excluding juice, and how often they ate vegetables or salads, including frozen vegetables but excluding potatoes. For both fruit and vegetable consumption, response categories were 1 = three times a day or more, 2 = twice a day, 3 = once a day, 4 = less than once a day, but at least four times a week, 5 = less than four times a week, but at least once a week, 6 = less than once a week and 7 = never. Answers were recoded so that a higher score indicated higher fruit or vegetable consumption. Physical activity was measured by asking participants on how many of the last 7 days they walked quickly, did sports or other physical activity for 30 min or longer. This is in line with European recommendations for sufficient physical activity . A higher score indicates engaging in physical activity on more days.
The independent variable, perceived encouragement of healthy behaviours by colleagues, was measured separately for healthy eating and exercise. For healthy eating, the item was “My colleagues encourage me to eat healthy food” and for exercise the item was “My colleagues encourage me to exercise frequently”. We created two variables, one for healthy eating encouragement and one for that of physical activity, as the correspondence principle holds that specific encouragement is likely more influential than generic encouragement . Answer options ranged from (1) always to (5) never, and were reversed so that a higher score indicated more perceived encouragement.
We added several control variables to our analysis. Female, younger, and higher educated people reportedly eat healthier , while men, younger and higher educated tend to engage in physical activity more . Therefore, our models controlled for gender (female = 1), age and years of education. We further controlled for self-rated health, as health and healthy behaviours are interlinked . According to previous research, people with a partner tend to behave healthier than those without, so we added a control for having a partner .
Moreover, we included several variables related to the work context. Since employees who work more hours tend to have more contact with their colleagues, we included working hours. Employees who have been part of the same team for longer have had more opportunities to be influenced by their colleagues there, so we added tenure in years in the team. Physical activity in the workplace may also contribute towards total physical activity . We therefore controlled for physical work demands, measured by how often employees’ duties involved standing, walking, or other physical activities. Additionally, how often employees worked from home — ranging from 1=(almost) never to 7 = four or five days a week was included, as employees tend to have less contact with their colleagues when doing home office often . Whether the employer had worksite health promotion policies (WHP), and if employees used them, as this has been related to healthier behaviour [48, 49], and colleagues may affect one another’s lifestyle choices by participating in WHP together . For fruit and vegetable consumption, this relates to catering or cafeteria menus offering healthy nutrition, and for physical activity, to sport facilities at work or a financial contribution towards a sport activity outside the workplace. Finally, we controlled for team size, sector, and country.
The pairwise correlations between the three outcome variables were low to moderate: rvegetable consumption, physical activity=0.12, rfruit consumption, physical activity=0.17 and rfruit consumption, vegetable consumption=0.49. We therefore fitted separate models for each outcome.
Because we expected employees’ healthy behaviours to be related, ordinary least squares regression models were not suitable: these models require observations to be independent – meaning that employees’ behaviours within a team may not correlate . Indeed, a test using Moran’s I found autocorrelation for all the dependent variables: fruit consumption (χ2 = 129.39(1), p < 0.001), vegetable consumption (χ2 = 150.36(1), p < 0.001) and physical activity (χ2 = 21.22(1), p < 0.001). We thus used network autocorrelation models (also known as spatial lag models or network effects models), which account for the interdependency of observations, and are therefore commonly used in social network analysis.[26, 28] The model builds upon standard linear regression models and takes the form of Y = ρWY + βX + ε, where Y is the vector of the outcome variable, W the adjacency matrix denoting which observations are part of the network, X a matrix of independent variables, β the vector of associated coefficients and ε a vector with error terms. As can be seen from the equation, the network autocorrelation model allows for the outcome of an employee (Y) to be directly associated with the outcomes of their colleagues (ρWY). Due to the nested data structure, we know which employees work together in the same team, and these are the colleagues whose outcomes we consider.
A relevant feature of the network autocorrelation model is that it includes a parameter ρ, which estimates the strength of the network effect. The network effect tests whether employees’ behaviours are related to their colleagues. The parameter ρ is a measure of the degree to which an employee behaves similarly to their colleagues, and ranges between − 1 and + 1 . For example, in the analysis for physical activity, ρ can be interpreted as the expected increase in the number of days an employee engages in physical activity if their colleagues increase their physical activity by an average of one day.
Central to a network autocorrelation model is the weight matrix W, which represents the influence mechanism in the network . In our study, we constructed W in such a way that only employees who worked in the same team were seen as influencing one another’s behaviours, as these were direct colleagues. Hence, the resulting adjacency matrix recorded a link between observations if employees worked within the same team, but no link if they worked in different teams or organisations. To account for differences in team sizes, we employed row normalisation, a common practice when using network autocorrelation models . In this procedure, each colleague has the same amount of influence, irrespective of team size. We created three separate weight matrices, to account for the different numbers of missing variables for our three outcomes.
We used a GS2SLS estimator for fitting the models because the alternative ML estimator reportedly produces biased estimates . For the hypotheses on encouragement, we examined direct and spill-over effects. The direct effect estimated the association between encouragement and a dependent variable. However, spill-over effects may be present due to interdependency: if one employee changes her fruit consumption because her colleagues encourage her to do so, this also affects the fruit consumption of other colleagues based on the network effect. As explained earlier, we examined the network effort for the hypotheses on employee behaviour.