Pre-Election Polling: Identifying Likely Voters Using Iterative Expert Data Mining

Abstract

One often-noted difficulty in pre-election polling is the identification of likely voters. Our objective is to build a likely voter model for presidential elections that efficiently balances accuracy and number of questions used. We employ the Iterative Expert Data Mining technique and data from the American National Election Studies to identify a small number of survey questions that can be used to classify likely voters while maintaining or surpassing the accuracy rates of other models. Specifically, we propose two survey items that together correctly classify 78 percent of respondents as voters or nonvoters over a multielection, multidecade period. We argue that our proposed model compares favorably to competing models by capturing the successful elements of those models while ignoring other elements that constrain identification. We end by suggesting that our model offers a new approach to identifying and evaluating likely voters that may maintain or increase accuracy without also increasing cost.