Content matters, context matters: unraveling behavior dynamics in an online health community for tobacco cessation

The objective of this research was to examine the content and context-specific information diffusion patterns underlying communication pertaining to tobacco use from online health communities (OHCs).

We utilized a mixed-methods approach comprising multidimensional qualitative coding to identify themes and communication attributes, automated text analysis leveraging advances in large language models (LLMs) to classify message content and context, and social network analysis to examine the dynamics of peer interactions in this study. Using QuitNet, an online tobacco cessation forum (n = 64 632 members, n = 2.39 million forum messages spanning 2000-2015), we extracted message-level features (eg, topic, theory) and context-related communication attributes underlying tobacco use behaviors as manifested in peer interactions. We then utilized stochastic actor-oriented models (SAOMs) to examine how communication content and context impact social network topologies and behavior dynamics [n = 3055 members (Wave 1), 2475 members (Wave 2), and 2289 members (Wave 3)].

OHC members expressed themselves using a variety of content and context categories such as social support (communication themes), feedback and monitoring (behavior change techniques), and emotion (pragmatic context). For the classification of communication attributes, LLMs trained on domain datasets outperformed other deep learning models [average F1-score = 0.91 (communication themes), 0.81 (behavior change techniques), and 0.79 (pragmatic context)]. Content-specific SAOMs revealed that engaging in dense triads with specific content-context patterns [comparison of behavior (content) with questions (context)] significantly affected the abstinence status of community members (P < .05).

These findings indicate that specific combinations of communication content and interaction context are associated with peer influence and abstinence outcomes in online health communities.

Novel behavior modeling approaches can identify latent peer interaction patterns in OHCs and advance the science of just-in-time digital behavioral interventions. Theory-enriched large language models combined with network analysis provide scalable and actionable insights for individual- and network-level strategies to support risky behavior modification such as tobacco cessation.