Predicting and Interpolating State-level Polling using Twitter Textual Data

Presidential, gubernatorial, and senatorial elections all require state-level polling, but continuous real-time polling of every state during a campaign remains prohibitively expensive, and quite neglected for less competitive states. This paper employs a new dataset of over 500GB of politics-related Tweets from the final months of the 2012 presidential campaign to interpolate and predict state-level polling at the daily level. By modeling the correlations between existing state-level polls and the textual content of state-located Twitter data using a new combination of time-series cross-sectional methods plus bayesian shrinkage and model averaging, it is shown through forward-in-time out-of-sample testing that the textual content of Twitter data can predict changes in fully representative opinion polls with a precision currently unfeasible with existing polling data. This could potentially allow us to estimate polling not just in less-polled states, but in unpolled states, in sub-state regions, and even on time-scaled shorter than a day, given the immense density of Twitter usage. Substantively, we can also examine the words most associated with changes in vote intention to discern the rich psychology and speech associated with a rapidly shifting national campaign. ∗Email: [email protected]; web: nickbeauchamp.com.