Forecasting the dynamics of the undecided population in a public opinion poll: A neural network approach

Abstract

One of the unsettled issues in any survey concerns the handling of the undecided respondents (UR). This problem can be safely ignored if the UR constitute only a small percentage of the population and their number is not enough to substantially sway the outcome of the survey. However, if the UR population is significant, say 30%, and the decided respondents (DR, in a two-state poll) split 40% : 30%, then the results of the survey have to be taken as inconclusive.
The conventional way of dealing with this problem by national survey groups (e.g. SWS, Pulse Asia) as publicized in mass media is to use net ratings which simply splits the UR in proportion to the DR. Such undertaking merely 'erases' the possible significance of the UR. Another approach is to allocate the UR among the different responses, based on such criteria as geographical distributions (e.g. residents of a given area has traditionally gone with candidate X), or socio-economic distributions (because, perhaps, affluent voters tend to go to Republican in the US). Such may be termed an exogenous method, inasmuch as it relies on considerations other than the data at hand.
Still another method is to use statistical analysis tools, such as discriminant analysis, which tries to look at patterns in statistical distributions to forecast the most probable separation of the undecided respondents. This has been used in the past, with claims of up to 86% success rate. Such a method we would term endogenous since it looks only at the given data and infer its conclusions only from them, with minimal or no heuristics based on outside considerations.
In this study commissioned by Pulse Asia, Inc. (PA), a public opinion poll organization, we demonstrate that an artificial neural network (ANN) can forecast up to 94% accuracy the most apparent sentiment of the UR when they eventually decide or is forced to decide. Specifically, an ANN is tasked to determine how the electorate rates the performance of Joseph Estrada, the incumbent president of the Philippines. The poll questionnaire consists of one direct question and 291 (pro-rated) indirect ones that probe into the opinions of the respondent on specific socioeconomic and political issues and ratings of other government executives, legislators, and institutions. The answers to the direct question can be divided in three parts: one, those who approved of the Estrada administration, second those who disapproved, and third, those who are undecided. A total of 1200 respondents were chosen randomly all over the Philippines with an uncertainty of ±6% and ±3% in the national and regional level, respectively.