Thread by @ashton1anderson

Thrilled to share our paper, out today in @nature, which proposes a new paradigm for the analysis of online platforms and applies it to study political polarization over the complete history of Reddit
w/ @isaacwaller
Paper: nature.com/articles/s4158…
🧵1/n

We show how to quantify online communities along interpretable social dimensions (e.g. “left-wing” to “right-wing” or “younger” to “older”) in a completely behavioural way. A community is “left-wing” only if the people who comment in it are more active on the left than the right.

Looking at how communities are distributed along the age, gender, and US political partisanship dimensions offers a high-resolution and large-scale picture of how the platform is socially organized.

We apply this method to study political polarization on Reddit. Are people increasingly selecting into politically homogeneous communities? The closer to the partisan pole a community is, the more of a political echo chamber it resembles.

Examining political activity, we see that Reddit underwent a significant polarization event in 2016: activity in communities near the partisan extremes rose sharply around the 2016 US presidential election.

What happened in 2016? To determine the mechanisms and dynamics behind this shift, we decomposed users by cohort based on when they were first active in political communities on Reddit. Several important insights emerge.

1) Individual users generally do not polarize over time. Aside from 2016, within-cohort polarization is either stable or decreases over time.
2) During 2016, all active cohorts were remarkably synchronized in their polarization trends.

3) The intense increase in polarization in 2016 was disproportionately driven by new (and newly political) users. The 2016 cohort was by far the most polarized group.
4) Individual polarization level is unrelated to previous activity on the platform.

In short, we don’t see much evidence for the platform polarizing people. Instead, new people polarized the platform in 2016.

We also observe a stark ideological asymmetry. Activity on the right shifted substantially towards the right-wing pole, whereas activity on the left and in the center did not polarize at all.

Read the full paper: nature.com/articles/s4158…
All the code and data—the community embedding, social dimensions, and community scores—are freely available at github.com/CSSLab/social-…

This project has been a herculean effort. Congratulations to @isaacwaller for all the hard work he put in over the last 2 years!

We are indebted to many people for their help, including the editors and reviewers for their invaluable feedback, many mentors, colleagues, and friends for their advice, and @jasonbaumgartne of Pushshift for the dataset analyzed in this paper.

The open access PDF is here: rdcu.be/cCqxr