不完全資料常發生於長期追蹤研究，在產生遺失值的情況下，資料分析過程將易 趨於繁複，因此如何找出適當的分析方法成為重要的議題之一。Little and Rubin (2002) 指出插補法為解決不完全資料問題的其中一種方法，大多數有關多重插補法之文獻則 著重於討論長期追蹤連續型反應變數以及長期追蹤二元資料。Demirtas and Hedeker (2008)提出針對長期追蹤順序資料之擬插補(quasi-imputation)策略，先將順序型類別分 解成二元型態，經由二元反應變數之相關結構轉換成多變量常態，再反轉換成二元型 態，進而還原到順序型態。在此計劃中，將提出另一種較簡易之插補策略，主要是依 據隨機數生成，並藉由模擬研究來討論比較Demirtas-Hedeker 方法與所提出方法之表 現。此外，利用Lang et al. (1999)探討分析青少年使用大麻之研究資料來闡述所提出方 法之應用。 Incomplete data are common problems in longitudinal studies. An appropriate analytical procedure in the presence of incomplete observations is a crucial issue due to the additional complexity that arises through incomplete data. Little and Rubin (2002) pointed out that imputation-based procedure is one of the methods for dealing with incomplete data. Most of the literature related with multiple imputation focus on the strategies for longitudinal continuous responses and for correlated binary outcomes. A quasi-imputation strategy for incomplete longitudinal ordinal data was proposed by Demirtas and Hedeker (2008), which collapses ordinal levels to binary ones and converts correlated binary outcomes to multivariate normal outcomes so that re-conversion to the binary and then ordinal setting. In this project, alternative imputation strategy based on random number generation for incomplete longitudinal ordinal data is proposed, and the performance between Demirtas-Hedeker method and the proposed procedure is compared by simulation studies. A study of marijuana use analyzed by Lang et al. (1999) will be used to demonstrate the application of the proposed method.
