LLM analysis of 72K Reddit posts reveals self-stigma patterns in substance use communities

Researchers from multiple institutions developed a ten-indicator codebook for self-stigma across cognitive, affective, and behavioral domains, then scaled classification using a large language model (LLM) validated against expert coding (Cohen's k=0.73, F1=0.80). They analyzed 72,115 thread-initiating posts from 1,660 English-language Reddit users (2006–2025). Results showed 3,838 posts (5.3%) from 1,228 users (74.0%) contained self-stigma, with all ten indicators discriminatory (relative risk 3.6 to 86.2). Leading indicators were self-labeling (56.0%) and despair/hopelessness (48.5%). Core and behavioral indicators were strongly associated at the user level (OR=4.65, 95% CI 3.12–6.94, p<0.001), and 87% of posts with behavioral indicators also contained a core indicator.

Contrary to progressive stage models, behavioral indicators like desire to quit emerged earlier than core indicators like shame (median position 0.08 vs. 0.38). Nine of ten indicators remained stable across posting trajectories; only pessimism increased over time (OR=1.62, 95% CI 1.25–2.10). This suggests self-stigma is integrated rather than sequential, with behavioral expressions rarely appearing without internalized ones. The findings highlight pessimism as a specific target for early digital interventions and demonstrate that textual disclosure does not follow traditional progressive stage models.