Swing Voters’ Vote Choice Prediction Using Multilevel Logit Model to Improve Election Survey Accuracy

Abstract

Public opinion surveys are often used to predict an election result. However, the predictions are not always accurate due to many factors. The presence of swing voters at the time of survey is one of the sources of the inaccuracy. On the other hand, election surveys are also often conducted by using multi-stage random sampling method so that ordinary models such as logit model generally do not provide satisfactory results. The data, hence, is complex and may be approached by multilevel models. The study is conducted to assess the extent to which a prediction of swing voters’ vote choice through a multilevel logit model can improve survey accuracy. The data used in this study was the result of a survey conducted using stratified multistage random sampling method two weeks before the 2019 presidential election. The model with 15 predictors and random effects for villages and neighborhood providing 96.3% accuracy and AUC reached 99.1% in the validation process. Based on the final model, the swing voters in this survey were predicted to vote more for Candidate B (10.4%) than Candidate A (7.5%). The direction of the swing voters’ support different from the loyal voters who prefer Candidate A (49.1%) than Candidate B (33.0%). The prediction of swing voters’ vote choice using multilevel logit model significantly improved the survey accuracy. Before the swing voters’ support was predicted the absolute deviation between the survey result and the election result was quite large, around 6.4%-11.5%. After swing voters’ support was predicted the absolute difference shrank to 1.1%.