民意調查者通常是以受訪者回答的投給某候選人作為受訪者的投票意向，然後據以預測選舉，然而，本研究指出簡單的將選民二分為回答投票意向者或不回答投票意向者，以及二分為投給某候選人或不投給某候選人，則選舉預測的錯誤機會相當高，這一個錯誤一方面來自於受訪者本身的不確定，而另一方面則來自於受訪者本身的不確定因而導致的測量誤差。本研究以為選民投給某個候選人的機率不是二分法的1或0，而是介於1與0之間，根據選民的政治態度與背景，作者使用多元洛基模型（multinomial logit model）估算出選民可能投票給每一個候選人的機率，並且發現選民如果投給某一個候選人的機率高，則他投票給投給該候選人的可能性就大，且不易變更。本研究據此原則預測選舉結果，結果有兩次選舉-民國八十三年的台北市長與民國八十六年的桃園縣長補選--預測準確，但另有兩次選舉-民國八十三年的台灣省長與民國八十五年的總統選舉--預測不準。 This paper examines the process of voters’ voting decision and develops a way of predicting voters’ voting decision. Most surveys on election predication are based on respondents' answers on their vote choices. However, this paper indicates two serious problems in this way of election prediction. One problem comes from the indecision of most voters, and the other problem comes from measurement errors. To deal with this problem, this paper provides a way of estimating each voter’s probabilities of voting for each candidate. The author indicates that the probability of voting for a particular candidate is not 0, neither is 1. Instead, the probability of voting for a particular candidate range from 0 to 1. The author bases on four major variables-- candidate evaluations, party identification, incumbent government satisfaction, and ethnicity-- to construct a vote choice model, and estimates the model by the way of a multinomial logit model. Then, the author calculates individual voters’ probabilities of voting for each of the candidates. The research finding shows that the chance of voting for a particular candidate is large if a voter has a high probability of voting for the candidate. Based on this principle, the paper predicts the outcomes of four elections and two of them are quite accurate.