Experimental Design Details
We base our analysis on a survey experiment that will be implemented in a large-scale representative online survey among the resident population of Switzerland in 2023. The survey will be carried out by a professional survey institute among Swiss residents between the age of 25 and 60.
Prior to the discrete choice experiment, respondents are assigned to two subsamples by stratified random sampling based on respondent sex, age, and language region. In subsample 1, respondents will be asked about their hypothetical daughter, and in subsample 2 about their hypothetical son. Each of the two subgroups is additionally subdivided into three subgroups, again by stratified random sampling based on the same respondent characteristics. This second subdivision is necessary due to the large number of degrees of freedom in our model which require a large number of choice situations to be answered which must be distributed to multiple respondents in order to maintain response efficiency. Respondents are asked to imagine having a 40-year-old daughter or son today (depending on which of the two subsamples they belong to), and in a fixed selection of subsequent choice situations (the selection depending on which of the three subgroups they belong to) which of two alternatives (“child’s careers”) they would prefer for their child.
The alternatives are defined by four attributes:
• Highest educational attainment (qualitative, 3 levels): University degree, University of Applied Sciences degree, apprenticeship degree
• Hierarchical position (qualitative, 2 levels): High (top management), low (no management function)
• Yearly gross wage (quantitative, 4 levels): 75’000 CHF, 100’000 CHF, 115’000 CHF, 130’000 CHF
• Job automation risk (quantitative, 3 levels): 30%, 45%, 60%
Job automation risk is defined as the risk that the job could be completely replaced by technologies such as robots or artificial intelligence within 10 years.
The primary outcomes of interest are the coefficients corresponding to our four attributes and, in the case of educational attainment and hierarchical position, the corresponding levels, as well as all their two-way interactions. These coefficients are estimated using mixed multinominal logit models (MXL) with the likelihood of an alternative being chosen based on given attribute levels as the dependent variable depending on whether individuals are asked about their hypothetical daughter or their hypothetical son.