What does a good day look like?: An interpretable machine learning approach to the American Time Use Survey

Area of Science:

• Behavioral Science
• Psychology
• Data Science

Background:

• Understanding the factors contributing to a "better than typical" day is crucial for well-being research.
• Previous studies have explored happiness correlates, but detailed time-use analysis is less common.

Purpose of the Study:

• To identify specific activities and their durations associated with reporting a "better than typical" day versus a "typical" day.
• To leverage interpretable machine learning to analyze large-scale time use data.

Main Methods:

• Utilized interpretable machine learning, specifically random forest models, to analyze data from the American Time Use Survey (2013 and 2021).
• Assessed the relationship between time spent on over 100 activities and daily happiness ratings.
• Model accuracy for distinguishing better than typical days was 62-63% balanced accuracy.

Main Results:

• Socializing significantly increased the likelihood of a better than typical day, but benefits plateaued after two hours.
• Working up to six hours showed no negative impact on daily happiness ratings.
• Working beyond six hours was strongly associated with a decreased probability of reporting a better than typical day.

Conclusions:

• Specific time allocations to activities, particularly socializing and work, are key differentiators of happy days.
• The findings provide empirical insights into the daily routines that promote subjective well-being.
• Machine learning offers a powerful tool for dissecting complex relationships in time use and happiness data.