Ask and Ye Shall Receive? Automated Text Mining of Michigan Capital Facility Finance Bond Election Proposals to Identify Which Topics Are Associated with Bond Passage and Voter Turnout.

Abstract

The purpose of this study is to bring together recent innovations in the research literature around school district capital facility finance, municipal bond elections, statistical models of conditional time-varying outcomes, and data mining algorithms for automated text mining of election ballot proposals to examine the factors that influence the probability of school districts in the state of Michigan passing or failing capital facility finance bond elections from 1998– 2014. Automated text mining is a data mining technique that identifies latent topics from a corpus of documents. We used an unsupervised correlated topic model to analyze the full text wording of all 1,210 school district capital facility bond election ballot proposals in Michigan over 16 years. The model identified 9 different latent topics across the bonds, including requests to purchase new buildings, renovations, and athletic facilities. Interestingly, equipment purchases appear to be a distinct category of bond proposal topics. We then examined the independent effect of the bond topics on the probability of passing the bond and voter turnout using modeling techniques and control variables from the recent literature. Bonds that focused exclusively on athletic facilities were 4.35 times less likely to pass than bonds that request new construction or omnibus ballot proposals. This work extends previous research to show that capital facility bond proposals that pass the most often include all facility needs in a single ballot proposal, are the first attempt at the polls, are at the top of the ballot, and request lower amounts of spending.