The Engagement Effects of Partisan Negativity in Social Media Posts: A Randomised Field Experiment

We test whether partisan negativity increases the reach and engagement of political social media posts. Using a randomised field experiment across major social media platforms, we compare substantively equivalent positive and partisan-negative versions of political posts to estimate the causal effect of negative partisan framing on post performance.

2026-09-16

2027-09-16

Primary outcome
The primary outcome is the platform-specific measure of the number of times a post was viewed or displayed, measured at the end of the experiment.
Because social media platforms differ in the metrics they make available, the following measurement rules will be specified in advance.
X
The primary measure will be the post's impressions.
X reports the total number of times a post has been viewed on a screen. This measure is not a count of unique users and can include multiple views by the same user.
If the standard X impression count is unavailable for a particular experimental post because of a change in platform functionality, the closest available exposure measure supplied by X or its analytics interface will be used.
As a secondary measure, engagement rate will be calculated by dividing total observable engagements (likes, comments, bookmarks and retweets/quote tweets) by post impressions.
Facebook
The preferred measure will be the post-level views measure available through Facebook or Meta Business Suite. Views can be used consistently across text, image, link and video posts.
If a directly comparable post-level views measure is unavailable, post reach will be used.
Reach measures displays of the content and is therefore conceptually closer to the X impression measure. Reach measures the number of unique accounts or users exposed to the post and will be used only where a display-based measure is unavailable.
The metric used will be recorded consistently for all Facebook experimental posts wherever platform data availability permits.
As a secondary measure, engagement rate will be calculated by dividing total observable engagements (reactions, comments, and shares) by post views.
LinkedIn
The primary measure will be post impressions.
LinkedIn defines impressions as the estimated number of times a post was displayed on the platform. This is the closest available analogue to the X view count and will be used for all content formats.
If post impressions become unavailable, members reached, which measures distinct members or Pages that saw the post, will be used as the fallback measure.
For video posts, format-specific video-view counts will not replace impressions where impressions remain available, since impressions provide a measure that can be used consistently across text, image, link and video posts.
As a secondary measure, engagement rate will be calculated by dividing total observable engagements (reactions, reposts, saves and comments) by post impressions.
Bluesky
At the time of preregistration, Bluesky does not provide post-level views, impressions or reach.
The primary Bluesky outcome will therefore be total observable engagement, defined as:
likes + replies + reposts + quote-posts.
This is an engagement measure rather than a direct exposure measure. It is preregistered as the closest consistently observable post-level indicator of performance available on Bluesky.
If Bluesky introduces a native post-level views or impressions measure during the experiment, the measure available at the commencement of the experiment will remain the preregistered primary outcome for Bluesky. Any analysis using a subsequently introduced exposure measure will be reported separately as a supplementary analysis.

Where available, secondary outcomes measured at the end of the experiment will include:
• likes or reactions;
• comments or replies;
• shares, reposts or quote-posts; and
• total observable interactions.
For X, Facebook and LinkedIn, an engagement rate may also be calculated by dividing observable interactions by the relevant exposure measure.
Because Bluesky does not provide an exposure denominator, a directly comparable engagement rate will not be calculated for Bluesky.
These outcomes will be clearly identified as secondary. The preregistered platform-specific exposure or fallback measure remains the sole confirmatory primary outcome.

We test whether partisan negativity increases the reach and engagement of political social media posts. Using a randomised field experiment across major social media platforms, we compare substantively equivalent positive and partisan-negative versions of political posts to estimate the causal effect of negative partisan framing on social media post performance.

Not available

Randomisation will be conducted by computer before the first experimental post.
The underlying content event will be the randomisation block. For every content event, exactly two of the four platforms will receive the partisan-negative treatment and two will receive the positive control.
The experiment will contain 100 complete content-event blocks.
Randomisation will be constrained so that:
• every content event contains exactly two treatment platform-posts and two control platform-posts;
• each platform receives exactly 50 treatment assignments and 50 control assignments; and
• the experiment overall contains exactly 200 treatment platform-posts and 200 control platform-posts.
A random allocation schedule satisfying these constraints will be generated before the experiment begins. Its order will be randomly determined using a recorded random seed.
The randomisation code, seed and resulting allocation schedule will be stored before the intervention begins.
For each content event, the two versions of the post will be finalised before the relevant treatment allocation is revealed.

The unit of randomisation is the platform-post.
Randomisation is blocked by content event.

No

N/A

400 platform-post observations. These consist of 100 eligible content events published on four platforms.

Positive framing control: 200 platform-posts.
Partisan-negative framing treatment: 200 platform-posts.
Within each platform there will be 50 treatment platform-posts and 50 control platform-posts.

With 400 platform-post observations, divided equally between treatment and control, the approximate minimum detectable treatment effect is 0.28 residual standard deviations at 80 per cent power and a two-sided 5 per cent significance level. This is an approximate calculation. Actual precision will depend on the amount of variation explained by the content-event and platform fixed effects and on the distribution of the platform-specific outcome measures.