Predicting Voter Behavior: Twitter vs Polls
The results of the 2016 US presidential election came as a shock to many people. Why is that? The media was conducting the same polls and surveys that they always do during elections in an attempt to gauge the support for each candidate at various points of their campaigns, which have been fairly representative in the past. What was so different this time? Pew Research Center has some predictions about what may have gone wrong, pointing out that uneducated voters are consistently hard for pollers to get responses from, and that demographic was one of Trump’s main sources of support. While these national and state polls are equal opportunity, that does not necessarily mean that there will be a truly representative sample that chooses to respond. This is a weakness of survey data collection and its ability to accurately predict social behavior.
However, there was another form of media prediction which was slightly more successful in predicting the 2016 presidential election results, and that was content analysis on social media. Media experts have found Twitter to be a useful social media platform in predicting election results, not only in the US, but also in Ireland and Germany. This is largely due to the fact that people often use hashtags that contain keywords that can be tracked to see which candidate people are talking about more, and algorithms that can detect positive and negative code words in conjunction with candidate names. This allows researchers to collect large randomized samples from many different demographics.
While this may seem like the ideal solution, it is limited by the fact that codes and algorithms and hashtags are generally only capable of revealing manifest content, which requires minimal interpretation. Data collected from Twitter would be even more accurate, although much more time consuming, if political tweets were sifted through and analyzed by actual people. This is because systematic and robotic content analysis can’t pick up on or interpret things like sarcasm or vaguely worded tweets. If media analysts were to analyze the same tweets, they may be able to interpret them more accurately and pull more useful information from them. They are able to find not only the manifest content, but also the latent content, which requires more interpretation that can not yet be performed by computers and coding.
In Social Media Today, an online newsletter that covers top headlines concerning social media, there is a feature article about how media analysts used Twitter to make educated inferences about the 2016 presidential election. The 3 primary measures outlined in this article were share of voice, audience growth, and sentiment. These measures, when used together, support the idea that latent content is much more valuable in studying people’s positions on competitive elections. The first measure, share of voice, relies on a general manifest content analysis, and tallies the number of mentions each candidate receives. While this doesn’t directly measure people’s opinions of the candidate, it does give a general idea of how popular they are among the general population and how many people are talking about them. Audience growth gives somewhat of a better idea of how many supporters each candidate has, because people generally follow accounts that are consistent with their personal opinions and beliefs. The final measure, sentiment, is what provides the latent content for this analysis.
This tri-layered process is much more accurate in predicting voting behavior than traditional polling and surveys because of its ability to reach a wider demographic of people without social desirability and conformity affecting the way people choose to respond. However, analysts must be sure to put in the effort to find and interpret latent content, not just manifest content. The data collected by surveys may not be the most effective form of prediction in terms of voter behavior, but now that we have access to people’s unfiltered opinions via social media sites such as Twitter we may be able to form more accurate predictions in future elections.
