Early Posting Behaviour and Later Deep Engagement in Incel Communities on Reddit: An Exploratory Observational Study

radpath is an exploratory observational research project examining whether early posting behaviour is associated with later sustained engagement in selected incel-related Reddit communities. The current release uses posting metadata only, excluding text, archived scores and removal status. Version 0.1.0 includes a six-page preprint, manuscript sources and supporting code on Zenodo. The deposit is not peer reviewed.

To what extent can early posting metadata predict later sustained participation in selected incel-related Reddit communities under temporal user-level evaluation? A separate planned analysis of gateway activity before entry is unsupported by the available archive coverage.

Developed the radpath research pipeline for data ingestion, cohort construction, coverage-aware outcome labels, temporal evaluation, aggregate exploratory analysis and metadata-only models. Prepared the manuscript, release configuration and supporting code, including synthetic tests, provenance checks and aggregate reporting.

Early predictors use target-community contributions from first observed entry through the earlier of the fifth contribution and 14 days. Sustained engagement requires at least 20 target contributions in four distinct seven-day bins after the predictor cutoff and within the following 90 days. The analysis deduplicates records, uses temporal user-level splits with a 90-day label embargo, and selects hyperparameters and classification thresholds on validation data. Free text, scores and removal status are excluded from corrected predictors.

A locally held partial archive of selected Reddit communities: incels, Braincels and IncelsWithoutHate, with MGTOW as a separate comparison community. The final cohort restricts entry to 19 September 2018 or earlier, and the final test population consists of 2018 entrants. Nominal train, validation and test fractions are 50%, 25% and 25%. Coverage-truncated outcome windows receive null labels. Authors are pseudonymized at ingestion; free text is discarded before persistence. Raw archives, user-level tables and individual predictions are not distributed.

The released preprint reports aggregate class balance, AUROC, AUPRC, F1 and calibration for the corrected metadata-only analysis. These results are exploratory because earlier test inspection informed the corrected design. The available coverage does not support a conclusion for the planned gateway-activity analysis. Numerical reproduction requires access to the same locally held partial snapshot, and independent replication is required.

The archive is partial, observed entry may occur after true entry, and missingness may vary across communities and periods. Coverage endpoints do not establish complete observation. Earlier test inspection informed the final cohort and predictor choices, so this is not an independent confirmatory evaluation. Historical text and embedding experiments do not establish whether content adds signal in the corrected analysis. Community participation does not establish ideology, radicalization, intent or dangerousness. The models are not suitable for individual risk assessment, flagging or moderation.

Raju Sah, sole author and independent researcher (ORCID 0009-0006-0412-6139). OpenAI Codex and Anthropic Claude assisted with development, analysis support, manuscript preparation and review; the author retains responsibility. No formal ethics review was sought, and no institutional approval or exemption is claimed. No external funding or competing interests are declared. Manuscript and aggregate figures: CC BY 4.0. Supporting code: MIT.