This paper came about when my clever colleague Toby Hopp asked me about a dataset I had collected. It was on the 2012 election and it included tweets that mentioned candidates Obama and Romney. I used it to investigate agenda-setting theory. He was interested in different theoretical areas — incivility and social capital. This lens led to interesting questions.
The premise — that different areas of the country would differ in how uncivil they were toward political candidates — was interesting. I knew incivility was in there, after all the #1 non-stop word associated with Obama was F@#$. Of course, I expected there to be variance in areas across the country, but to what extent would that variance be explained by the demographic status, economic status and political partisanship of a particular area?
[iframe src=”http://www.chrisjvargo.com/incivility/” width=”100%” height=”400″]
The genius piece of programming that allowed us to embark on this investigation was the Sunlight API. My deepest thanks to Sunlight Foundation for making the API freely available to the public. It made resolving a GPS coordinate to a congressional district as easy as using requests in python to fetch some JSON. Once we had that, getting the census data and voting results for each congressional district was straight forward.
Developing a computational score for incivility was more — colorful. We relied on Cluebot and Google’s “bad word” lists. It worked well for a recent Kaggle competition aimed at detecting insults in online commentary. We show in the paper that it was valid for political candidate incivility as well. We manually added in candidate specific insults as we saw them, of course. I encourage you to try to read through these lists with a straight face.
The result? We found that incivility on Twitter was highest in districts that had:
- Low Socio-Economic Status
- Low Levels of Social Capital
- Low Levels of Partisan Polarity
As with all big data studies, there is some nuance and exceptions to these results. For a detailed breakdown, see the full paper here >>