Platform Governance and the Heterogeneity of Online Discourse Hostility: A Matched Community-Month Panel Study of Two Social Platforms, 2015–2020

Abstract

Background: Service-wide toxicity figures circulate in regulatory debate as though they described a platform. Whether such a figure is stable, or an average over communities that differ, is untested.

Objective: To estimate the difference in discourse hostility between a strongly and a minimally moderated platform in matched communities, and to test whether it changes over time, varies across communities, and rests on equivalent measurement.

Methods: A matched observational panel of 914 community-platform-month cells, of which 820 formed 410 matched pairs across six identically named communities on Reddit and Voat over 71 overlapping months to December 2020, was analysed. Hostility was modelled as a latent construct with three indicators and tested for cross-platform invariance. Month-clustered regressions, random-effects pooling, Newey-West trends, and bias-corrected bootstraps were estimated with Holm adjustment across nine directional tests.

Results: Matched cells on the minimally moderated platform were more hostile (+0.066 in toxicity, 95% CI 0.053 to 0.078; p < 0.001) and more negative in valence (-0.064, 95% CI -0.076 to -0.053). The difference widened by +0.002898 units per month (p < 0.001), reversing sign at month 10.2, and was null only between months 9.1 and 13.6. Between communities, I-squared reached 96.5% and Hedges g ranged from 0.36 to 1.70, so the prediction interval for an unmeasured community (-0.0363 to 0.1676) included zero. Metric invariance held (p = 0.174); scalar invariance did not (p < 0.001).

Conclusion: The platform difference is real, widening, and community-conditional rather than constant, and raw levels are not comparable across services. Comparative metrics need a heterogeneity statistic, a prediction interval, and an observation window.