Feeling all the (Partisan) Feels: Exploring the Drivers of Affective Polarization at the Individual Level

The aim of this research is to capture how a particular variable of interest (social media echo
chamber) interacts with human behavior (extremism). While an observational study may be used to infer association between these two parameters, it would make it difficult to isolate and generalise the impact of echo chambers themselves, while controlling for other factors. Therefore, I propose instead to design an online laboratory experiment.
The research will proceed in three stages:
1. I will first need to operationalise both explanatory and response variables by defining them into measurable factors.
2. I will then formulate sharp hypotheses on the potential effects of echo chambers on political extremism.
3. Finally, I will test these hypotheses in an experimental setting.
While there will be several ways to operationalise the concept of social media echo chamber, one way would be to simulate a situation in study participants are exposed to ideologically homogenous media messages inside a closed system (news feed). To develop such ideologically homogenous news feeds, I will first select a sample of Facebook8 users from a target population and infer their political orientations using discrete behavioral data such as publicly expressed preferences, as it is thought more robust and less biased then self-reported data (Bond and Messing, 2015). This should lead to a comprehensive topology of said users' political orientations, ranging from very liberal to very conservative. A good sampling frame will be crucial to minimise sampling error (Neuman, 2014: 252). I will then extract data on media content shared by users included in the sample over a specified period of time (e.g: a political event such as the 2016 U.S. presidential election).
To circumvent data privacy issues, in both cases the data collection process will be carried out using Digital Footprints - a software that lets researchers collect private Facebook data (user data, posts and news feeds) normally unavailable to them using the Facebook's API, with explicit user consent9. While providing unrivalled access to users' private information and actual behaviour on the platform, this collection technique prevents important limitations. Due to the voluntary nature of participation in the data collection process, the sample may be non-representative of the target population (Neuman, 2014: 259). However, this could be remedied by supplementing the dataset with public data gleaned from publicly available Facebook profiles. Moreover, Facebook's API is not only less stable than live data feed, but there is still little documentation of its structure.
When going back in time in data collection, it might thus be difficult to discern if data patterns occur due to changes in the user interface or in the API structure itself (Bechmann and Vahlstrup, 2015). Drawing on the methodology employed in a study conducted by Facebook scientists (Bakshy, Messing and Adamic, 2015), which looked at how a subset of the social network's users reacted to the news appearing in their feeds, I will then classify news articles according to their ideological alignment on specific issues. This will involve tracking which articles were most shared by each category of users and calculate a political "alignment score" for each link (ibid.) Using that score, links will be grouped into ideological categories and sorted by topical issues. Here, text classification methods such as N-gram based text categorisation will be useful. To equip myself with the required technical skills to formalise this research design, I plan to audit both the Experimental Approaches and Accessing Research Data from the Social Web graduate courses at the OII. The second stage of the research will be to test the effect of this media stimulus on attitude extremism.