*3.2. Methods*

For ease of interpretation, we performed Ordinary Least Squares (OLS) regression analyses with robust standard errors. We thus applied linear probability models, given that the cognitive demands, as well as the health outcome measures, were coded as indicator variables. Coefficients may thus be interpreted as differences in the probability of the outcome variable—in this case facing a specific cognitive demand or a certain health condition. Given that linear and logistic models often render very similar results, we preferred this model, as this interpretation is much more intuitive and fits the research questions studied somewhat better compared to the interpretation of other methods (e.g., odds ratios derived from logistic regressions) [55]. In order to adjust for the violation of homoscedasticity, heteroscedasticity-consistent standard errors were applied.

In a first step, we explored the determinants of cognitive demands by regressing the cognitive demands on the control variables to get an idea of the groups frequently facing cognitive demands at work. Second, we explored whether, and to what extent, cognitive demands are related to employee well-being. We additionally exploited the information on whether or not the specific cognitive demand is perceived as stressful by restricting the analyses to those reporting to be frequently facing new tasks or doing unlearned things. In doing so, we aimed to make an effort in testing the P–E fit (see Section 2) and assess whether the cognitive demands are related to health or whether it depends on the individual's characteristics. Moreover, these analyses enabled us to estimate the relationship between cognitive demands and health at the intensive margin, and thus mitigate the potential bias resulting from selection into the extent of facing cognitive demands. Previous studies suggest that cognitive demands might be harmful if they co-occur with work overload [28]. Therefore, in the final analyses we included interaction terms between one specific cognitive demand (i.e., doing unlearned things) and two working conditions (work intensity and autonomy) in the regressions of well-being. As a result, we were able to assess whether or not facing high work intensity or autonomy moderates the relationship between this cognitive demand and employee well-being. We focused on this cognitive demand as our analyses revealed that it is strongly related to adverse health or well-being, and, therefore, can be interpreted as a stressor. Work intensity and job autonomy were chosen, as according to common work–stress theories, their importance for employee well-being is well explored and widely accepted. Both the measures for cognitive demands and working conditions were operationalized as indicator variables, and thus the interaction terms between these two variables can be interpreted as follows: a positive interaction suggests that the working condition strengthens the association between the specific cognitive demand and employee well-being, while a negative interaction term mitigates the association.
