Using Machine Learning to Characterize Adolescents’ Time Well Spent on Social Media

• In Event: Methodological Innovations to Advance the Science of Social Media Use and Adolescent Development and Well-being

Abstract

What is “time well spent” on social media and how can we measure it? Do adolescents maximize time well spent relative to other activities? How is time well spent associated with subjective well-being?
Much research examining the impact of social media on adolescents’ lives focuses on how much time they spend (quantity) and, to a lesser degree, how they spend it (quality). Yet these approaches are limited. Measuring quantity of social media use ignores the specific behaviors or experiences youth have online, and measures of quality of social media often impose researchers’ assumptions about what is “positive” vs. “negative” engagement.
This study uses topic modeling, an unsupervised machine learning technique, to identify patterns in adolescents’ free-response statements regarding their social media use. The goal of this approach is to characterize adolescents’ consensus of time well spent on social media, and further, to understand whether youth who align their social media behaviors with this consensus are happier and less stressed.
Data for this study were collected by the Character Lab Research Network in January 2021, who shared a deidentified dataset with our research team. The sample included roughly 24,500 sixth through twelfth grade students in the U.S. The sample was approximately 58% female, 71% White, 18% Black, 8% Asian; approximately 38% of the sample were Hispanic; and about 40% qualified for free or reduced-price school meals.
Students responded to four open-ended questions using the prompt, “Tell us ONE specific thing you do on social media that…” (stresses you out; gives you lasting happiness; makes you feel good in the moment but doesn’t give you lasting happiness; and is hard to stop doing). Students were then asked to rank-order each of the four behaviors they had provided, from most to least time spent doing them. They also completed several items assessing subjective well-being (e.g., happy; sad; stressed; relaxed) using Likert scale response options.
We will use latent Dirichlet allocation (LDA) to identify 200 clusters of words (i.e., “topics”) in adolescents’ responses across all four prompts. Analysis of the data remains ongoing, but preliminary results indicate that topics are differentially associated with the four prompts. For example, words from a topic relating to artwork (including words like “create,” “artwork,” “artists,” and “inspiration”) are more strongly associated with responses to the “lasting happiness” prompt, whereas words from a topic relating to politics (“political,” “issues,” “racial,” and “rights”) are more strongly associated with responses to “stress” prompt. See Figure 1.
To identify adolescents’ consensus of what constitutes time well spent on social media, we will compute a mean score for the top ten topics most strongly correlated with the “lasting happiness” prompt. The association between each individual adolescent’s distance from this mean and their subjective well-being will be used to examine if aligning behavior with the “wisdom of the crowd” is associated with higher well-being. These analyses will be preregistered and conducted prior to presentation.

Authors

• Brian Galla, University of Pittsburgh
  Presenting Author

• Anne Maheux, University of Delaware

• Sophia Choukas-Bradley, University of Delaware

• Jacqueline Nesi, Brown University

• Dipen Rupani, Stony Brook University

• Adithya Ganesan, Stony Brook University

• H. Andrew Schwartz