Using Twitter to Observe Election Incidents in the United States

• In Event: Research Methodologies Using Twitter and Facebook

Abstract

No institution exists in the United States to collect and report citizen complaints about election processes or administration on a nationwide basis. Federally mandated procedures and many state-level procedures are ineffectual, producing few or no reported complaints. The lack of reports should does not mean that problems do not occur but rather that reporting procedures are ineffective. We describe infrastructure built using statistical learning tools to gather information from Twitter that contains observations of election problems by individuals all across the United States. We use data from the presidential primary elections and caucuses held across the country in 2016 as well as from the general election. To classify Tweets for relevance and by type of election incident, we use machine classification methods in an active learning framework. Only a small percentage (less than two percent) of Tweets captured using our keywords and location criteria are observations of problematic incidents, but our corpus includes millions of Tweets. One technical challenge for which we develop new classification techniques is that different types of complaints occur with very different frequencies, and an individual Tweet may exhibit several different types of complaints. Our solution to such complications involves binarizing attributes, sampling and weighting in cross-validations. Another challenge is to determine precisely where each observation occurred; while a Tweet must include some "location" information to be included in our corpus, few Tweets include geotagging information. We obtain data from every state. For primary election day in at least one state (California), the distribution of types of incidents revealed by data developed from Twitter roughly matches the distribution of complaints called in to a hotline run on that day by the state. We check whether the pattern of Twitter-reported incidents differs in closely contested states from other states, and we use the timelines of Twitter users who reported an incident to try to determine whether there are partisan biases among complainers.

Authors

• Walter R. Mebane, University of Michigan, Ann Arbor

• Blake Miller, University of Michigan

• Alejandro Pineda, University of Michigan

• Joe Klaver, University of Michigan

• Patrick Y. Wu, University of Michigan, Ann Arbor

• Logan Woods, University of Michigan