廣義之潛在變數模型之探討及其在選舉研究之應用

Abstract

政治制度在政治學研究中向來具有舉足輕重的地位，由於其在政治實務的重要性，也是歷久不衰 的研究主題，而制度研究的範圍也逐漸擴展到探究制度對個體行為以及其形成的總體現象的影響。在 台灣以選舉制度最受到一般人民與研究者的重視，因此，本計畫關心的正是台灣立法委員選舉制度對 政黨競爭與選民行為的影響。在目前並立式單一選區兩票制的政治環境下，選民在投票時，會審慎思 考如何分別投下手中的兩張選票。由於兩張選票的選區大小、選票結構與當選規則皆完全不同，選民 在投票時所考慮的因素也不相同。但是因為選民是同時投下兩張選票，所以這些背後的因素會互相影 響。釐清這些因素如何影響選民在兩張選票的決定便是此計劃的主要目標，為達成此目標，此計畫將 採用一般化的潛在變數模型來分析個體層次的調查資料，並以貝氏統計估計之。此統計模型的優點是 能夠適切地呈現資料產生過程的主要特徵。

From the perspective of institutional scholars, political institutions play a critical role in the political process, in which political institutions link public preferences to policy decisions. Among democratic institutions, electoral systems attract the most attention of the citizens and researchers in Taiwan. One of the reasons why Taiwan’s political scientists pay much attention to electoral systems is due to the legislative electoral reform in 2005. After then, Taiwan’s legislators are elected by a mixed-member majoritarian (MMM) system, which is a combination of a plurality with single-member-district (SMD) system and a closed-list proportional representation (PR) system, rather than a single non-transferable vote with multimember district (SNTV-MMD) system. One of the characteristics of MMM systems is that each voter has two ballots: one for a constituency representative and the other for a party. In addition, the district magnitude and electoral formula for these two tiers are also different. As a result, voters’ voting decisions for these two ballots depend on different information. In specific, voters rely more on candidates’ personal attributes than party brands when voting for constituency representatives while they rely more on the information of parties than on the information of individual candidates when voting for party lists. However, since voters simultaneously cast the two ballots, voters’ voting decisions for these two ballots are somewhat correlated. The correlation between the two voting decisions may induce spurious relationship if only one voting decision is analyzed. To deal with this problem, I apply generalized latent variable models to analyzing survey data, which is estimated by Bayesian methods.