在從事選舉預測時，研究者常面臨受訪者不告知其投票對象的問 題，若僅以告知投票對象的受訪者作選舉預測，將無可避免地造成選 樣偏誤的問題。本文主要目的在於評估選樣偏誤對於投票模型估計所 造成的影響，並且試圖藉由矯正選樣偏誤所造成的問題，得到較正確 的參數估計值，並進而作更精確的選舉預測。在本文中，我們採取 Dubin與Rivers戶斤發展出來的二變量選樣偏誤模型’（bivariateselection biasmodel)為研究方法，為了檢視選樣偏誤模型在選舉預測上的穩定 性，我們將之應用在五次不同的選舉中。結果發現在五次選舉中，未 校正選樣偏誤（也就是只以願意回答投票對象者加以預測），都會造 成高估自變數對應變數的影響，因為願意回答投票對象者往往是政治 偏好較強或較確定的受訪者，也因此會造成選舉預測的偏誤。當我們 校正選樣偏誤後，在四次選舉中都發揮了極好的效果，預測的誤差都 比原本不校正選樣偏誤來得更小，且誤差都不超過1.16%，可謂相當 地準確。唯有在一次選舉無法發揮校正的效果，但是即便如此，也並 不會比不校正更差。我們認為這樣的效果顯示，選樣偏誤模型是一個 相當可以信賴的選舉預測工具。

When predicting elections, researchers always face the difficulty that some respondents do not answer the question of vote choice. However, including only those who give a clear answer on vote choice, researchers might have the problem of selection bias. This article tries to assess the effects of selection bias on vote choice model, correct the wrong estimates resulting from the bias, and predict elections accurately. In this article, we apply a bivariate selection bias model-- developed by Dubin and Rivers-- to five different elections. The research findings show that if researchers include only those respondents who give a clear answer on vote choice question, they might take a risk of overestimating the effects of independent variables on vote choice. This is because respondents who give a clear answer on the question of vote choice may also have definite and strong political preferences, and they are quite different from those who do not give a clear answer. After correcting the estimating errors resulting from selection bias, we might predict election outcome accurately. The largest predicting error in four elections is I. I 6%. It is less than sampling error. The model cannot do a better job in only one election. Fortunately, it does not do a worse job, either. The overall result shows that the selection bias model is a reliable tool in predicting elections.