What Do People Watch under Adversity? Testing Interactions of Semantic Affinity and Coping Style Using Netflix Data Donations

Rebekka Kreling, Felix Dietrich, Alicia Gilbert, Leonard ReineckePsyArXiv 2024 · DOI

TL;DR

This study challenges a long-standing psychological theory—Mood Management Theory—by showing that people don’t actually avoid media content that reminds them of their real-life stressors. By linking daily diaries with actual Netflix viewing histories and using Large Language Models (LLMs) for content classification, the researchers found that "semantic affinity" (the overlap between your problems and what you watch) has little to no impact on media choice compared to personal habits and situational coping goals.

What Problem Does This Paper Address?

When we are stressed, we often turn to media to cope. For decades, Mood Management Theory (MMT) has suggested that we should instinctively avoid content with high "semantic affinity"—meaning if you just had a fight with your partner, you should avoid romantic dramas to prevent ruminating on your own misery.

However, most evidence for this comes from controlled laboratory experiments using artificial stressors (like a fake test) and limited media choices. This paper addresses the "ecological validity" gap: what do people actually choose to watch in the wild when facing real-world "daily hassles" like work deadlines or family friction? The authors seek to bridge the gap between theoretical media psychology and the reality of high-choice streaming environments like Netflix.

Methods

The researchers employed a sophisticated linkage approach, combining subjective self-reports with objective behavioral data:

  1. Diary Study: 122 participants completed "measurement bursts" (5 consecutive days, 3 times a semester) reporting their daily stress levels across five life domains: family, friendships, romance, university, and work.
  2. Data Donations: Participants "donated" their actual Netflix viewing histories (CSV files downloaded from their accounts), resulting in a dataset of 2,122 unique titles.
  3. LLM-Based Content Classification: To determine the "semantic affinity" of the Netflix titles, the team used a zero-shot classification model (deberta-v3-large-zeroshot-v2.0). The model analyzed short descriptions of the shows/movies to calculate the probability that a title covered specific life domains (e.g., "Does this movie cover the life domain: work?").
  4. Statistical Modeling: They used Bayesian random effects within-between (REWB) models to account for both daily fluctuations (within-person) and general personality traits (between-person).

Datasets

  • Primary Survey Data: 15 days of diary entries from $n = 122$ university students.
  • Netflix Viewing History: $n = 2,122$ titles matched to specific days of reported stress.
  • JustWatch API: Used to scrape descriptions, genres, and ratings for the donated Netflix titles to provide the "premise" for the LLM classifier.

Key Insights and Ideas

  • The "Semantic Affinity" Myth: The core finding was a series of null results. There was no significant evidence that people avoid content related to their stressors. Whether a student was stressed about school or work, it didn't reliably predict whether they watched a show about school or work.
  • Genre Matters More Than Topic: While the specific topic (domain) didn't drive choice, the coping strategy did. People chose specific genres based on how they wanted to feel. For example, those using media for "active coping" gravitated toward comedy and romance, while those seeking "comfort" leaned toward action and thrillers.
  • Habit Trumps Stress: The study found high Intraclass Correlation Coefficients (ICCs), suggesting that Netflix use is highly habitual. You are more likely to watch what you usually watch (or what you started last week) than to change your viewing habits based on a bad day at the office.
  • LLMs as Research Assistants: The paper demonstrates that zero-shot classification via LLMs is a highly effective, reproducible way to categorize massive amounts of media content, correlating strongly ($r = .67–.82$) with human coders.

Key Results

  • Hypothesis 1 & 2 (Avoidance of Related Content): Unsupported. There was no negative relationship between adversity in a life domain and the prevalence of that domain in selected media, regardless of whether the person had an "approach" or "avoidant" coping style.
  • Situational Coping (RQ1): The only significant link to semantic affinity was "self-distraction," which—counter-intuitively—was positively related to watching related content.
  • Exploratory Genre Findings:
    • Active Coping: Positively associated with Comedy and Romance; negatively with Thrillers.
  • Denial: Positively associated with Action.
  • Self-Distraction: Positively associated with Drama.

Limitations and Future Work

  • The "Deselection" Problem: Netflix data shows what people watched, but not what they stopped watching. It’s possible people start a show with high semantic affinity and turn it off after five minutes because it hits too close to home—a nuance the current data can't capture.
  • Sample Bias: The participants were university students. Their primary stressor was "university/education," but "university" is a relatively rare theme in Netflix’s catalog compared to "work" or "romance," potentially skewing the affinity scores.
  • Granularity: The authors suggest future work should use Experience Sampling Methods (ESM) to measure stress and media use in closer proximity (e.g., hour-by-hour) rather than aggregating by day.
  • Conceptual Clarity: The authors call for a theoretical "reboot" of the semantic affinity concept, noting that it remains unclear what "similarity" actually means in a real-world context (e.g., does a fight with a friend mean you avoid all social dramas or just shows about that specific topic?).

Summary updated

Abstract

Media are frequently used for coping with everyday stressors, but little is knownabout what content individuals choose under such adversity. Building on mood managementtheory’s propositions about semantic affinity of content, we test how adversity in specific lifedomains is reflected in choice of entertaining media content surrounding the same lifedomain. In combining diary data from n = 122 Netflix users and data donations of theirviewing histories with n = 2,122 titles, our study provides a novel methodological approach.Our findings show no effects of daily adversity on selecting content with more or lesssemantic affinity, but exploratory analyses reveal that situational media coping strategiesdifferentially relate to the selection of different entertainment genres. These results emphasize(a) a lack of conceptual clarity surrounding semantic affinity in real-life media use, and (b)the importance of situational approaches in media coping research.
BibTeX
@misc{kreling2024what,
  author = {Rebekka Kreling and Felix Dietrich and Alicia Gilbert and Leonard Reinecke},
  title = {{What Do People Watch under Adversity? Testing Interactions of Semantic Affinity and Coping Style Using Netflix Data Donations}},
  year = {2024},
  doi = {10.31234/osf.io/utjre},
  howpublished = {PsyArXiv},
}
Figure