Shared partisanship dramatically increases social tie formation in a Twitter field experiment

The study investigates whether shared partisanship causally drives social tie formation, beyond correlational assortment observed offline and on social media. Prior observational work shows Americans are more likely to interact with and connect to copartisans, potentially creating echo chambers. However, such homophily may be due to confounding factors (e.g., age, race, geography, interests) or differing opportunities for contact, including algorithmic recommendations. Experimental evidence on actual tie formation is limited, often relying on survey hypotheticals, with rare field tests outside romantic contexts. This paper tests, in a naturalistic Twitter field experiment, whether users are more likely to form ties with copartisans than counterpartisans and whether this varies by strength of partisan identification or differs between Democrats and Republicans.

The paper builds on research showing partisanship as a core social identity associated with distrust and dislike of the opposing party and reluctance to befriend counterpartisans (e.g., Huddy et al., Iyengar et al.). Observational studies report greater face-to-face interactions and online connections among copartisans (Gentzkow & Shapiro; Colleoni et al.), contributing to filter bubbles and echo chambers (Flaxman et al.; Stewart et al.; Sinclair). Yet, these studies are correlational and cannot cleanly identify causal effects due to confounds and opportunity structures (DellaPosta et al.; Currarini et al.). Experimental evidence on real tie formation is scarce, with notable work in online dating showing copartisan preference (Huber & Malhotra). There are debates about partisan asymmetries in homophily and bias, including claims that conservatives are more homophilous on Twitter (Boutyline & Willer) and broader questions of symmetric versus asymmetric bias (Baron & Jost; Ditto et al.). This study addresses these gaps by using a field experiment on Twitter.

Design and stimuli: The researchers created eight human-like, identical-looking Twitter bot accounts that varied on two dimensions: political partisanship (Republican vs. Democrat) and strength of partisan identification (stronger vs. weaker). All bots followed elite accounts matching their political partisanship and retweeted randomly from those accounts daily. Strong-identification bots included a background supporting the party’s presidential candidate and incorporated the candidate’s name in the profile; weaker-identification bots did not.

Subject recruitment and classification: The team identified Twitter users who had retweeted MSNBC or Fox News. For each user, they collected up to the last 3,200 tweets and classified partisanship based on the balance of shared content from left- versus right-leaning websites (per Eady et al., 2019), with the absolute value used as a measure of partisanship strength. Users with more than 15,000 followers or lacking an ideology score were removed. A politically balanced subject pool was constructed.

Blocking and random assignment: Users were grouped into homogeneous blocks based on estimated partisanship, partisanship extremity, number of followers, recent activity (number of days with at least one tweet in the past 14 days), and reciprocity tendency (mutual friendships divided by total followers). Within blocks, users were randomly assigned to be followed by a bot varying in partisanship concordance (copartisan vs. counterpartisan) and identification strength (stronger vs. weaker).

Bot credibility and exposure: Each bot initially had approximately 1,000 politically neutral followers and retweeted 10 political tweets aligned with the bot’s ideology to appear credible. The plan was to follow 6,000 users over 14 days, but Twitter restricted the bots’ follow capabilities after 2 days; the experiment concluded after following n = 842 users (median 64.5 followers, 218 followed accounts, 4,416 total tweets; 46% Republican; 45% female; mean age 45.8 years).

Outcome and analysis: The primary outcome was whether the user reciprocally followed back the bot. Analyses used a linear probability model predicting follow-back from copartisanship (bot-user match), bot partisanship extremity, user partisanship, and their interactions. Exact P-values (P_FRI) were computed via Fisherian randomization inference with 10,000 permutations. The study was IRB-approved with a waiver of informed consent (MIT COUHES Protocol 910465) and preregistered (https://aspredicted.org/ca3nm.pdf). Data and code are available at OSF (https://osf.io/s5e6j/).

• Strong causal effect of shared partisanship on tie formation: Users were nearly three times more likely to follow back a copartisan bot than a counterpartisan bot (b = 0.093 [0.051, 0.135], t(840) = 4.381, P < 0.001, P_FRI < 0.001).
• No main effect of bot extremity and no interaction with copartisanship: Interaction between copartisanship and bot partisanship strength was not significant (P = 0.465, P_FRI = 0.469). Bot extremity main effect was not significant (P = 0.754, P_FRI = 0.748).
• Exploratory post hoc evidence: A positive three-way interaction among copartisanship, bot partisanship strength, and user partisanship strength (P = 0.051, P_FRI = 0.037) suggested that more partisan users were particularly likely to follow back strong copartisan bots.
• No partisan asymmetry: Democrats and Republicans did not differ significantly in their preferential follow-back of copartisans (interaction b = 0.012 [−0.071, 0.096], P = 0.771, P_FRI = 0.784). No significant main effect of user partisanship (P = 0.563, P_FRI = 0.583). No significant three-way interaction with bot extremity (P = 0.886, P_FRI = 0.881).

The findings provide direct causal evidence that shared partisanship substantially increases the likelihood of forming social ties on Twitter, validating that preferences documented in survey experiments generalize to real-world behavior. Contrary to some observational claims of conservative-specific homophily, the effect was symmetric across parties: both Democrats and Republicans were similarly more likely to reciprocate follows from copartisans. These results inform theories of echo chambers and partisan assortment by showing that network segregation is not solely driven by offline assortative structures or algorithmic recommendations; individuals actively prefer connections with copartisans, even among strangers. Practically, efforts to reduce partisan assortment on social platforms may need to counteract users’ psychological biases favoring copartisan ties.

This study demonstrates a large, causal effect of shared partisanship on the formation of social ties in a naturalistic Twitter field experiment and shows no partisan asymmetry in this preference. It advances political psychology by linking survey-based findings to actual behavior, and it clarifies the microfoundations of online echo chambers. Future research should test generalizability to more representative populations and platforms, explore alternative and subtler signals of partisanship beyond candidate support, examine longitudinal dynamics of tie formation, and evaluate algorithmic interventions designed to diversify users’ connections.

• External validity: The sample comprised politically active Twitter users identified via engagement with MSNBC or Fox News; results may not generalize to the broader Twitter population or the general public.
• Signaling mechanism: Bots primarily signaled partisanship through presidential candidate support. Some copartisans may not support their party’s nominee, potentially underestimating the true effect of shared partisanship; effects may differ across parties or contexts.
• Early termination and sample size: Twitter restricted the bots’ ability to follow more users after 2 days, reducing the planned sample (n = 842 instead of 6,000) and study duration, which may affect precision and the detection of smaller effects.
• Outcome scope: The measure focused on immediate follow-back behavior; other forms of social tie quality or sustained interaction were not assessed.
• Platform-specific context: Results pertain to Twitter’s ecosystem and design during the study period; effects may vary on other platforms or over time.