遺失值常發生於長期追蹤研究中，多重插補法即為解決遺失值問題的方法之一。 Demirtas and Hedeker (2008)提出針對長期追蹤順序資料之多重插補策略，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性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。在目前進行的研究計畫中，已應用模擬研究針對 此方法，就標準化偏誤、平均平方根與涵蓋率來衡量在完全隨機遺失和隨機遺失型態 下之表現。模擬研究顯示，此插補方法在許多不同模擬條件下，其參數估計表現良好。 然而此方法過程頗為繁瑣，發展出另一種新的插補方法，則為此申請研究計畫之主要 研究目標，透過直接生成順序型插補值，而非經由繁雜的轉換。除此之外，藉由模擬 研究以比較Demirtas-Hedeker 方法與所提出的新方法之差異。 In longitudinal studies, missing data are common occurrence. The multiple imputation method is one of the approaches for dealing with missing data. A strategy proposed by Demirtas and Hedeker (Statistics in Medicine, 27, pp.4086-4093, 2008) is for imputing incomplete longitudinal ordinal data, which converts discrete outcomes to continuous outcomes by generating normal values, employs the multiple method based on normality, and reconverts to binary scale as well as ordinal one. In our working project, the performance of multiple imputation in terms of standardized bias, root-mean-squared error and coverage percentage under missing completely at random (MCAR) as well as missing at random (MAR) was evaluated by various configurations. The simulated results indicated this imputation strategy is suitable for most of incomplete data under these two missing-data mechanisms. However, the imputation procedure provided by Demirtas and Hedeker is somehow tedious. The aim of the proposed project is to develop a new imputation method based on the generation of correlated ordinal responses rather than binary ones. Also its performance compared with the method of Demirtas and Hedeker will be discussed by simulations.
