長期追蹤資料分析常用來研究隨著時間改變，重複測量特性之變化情形。對於平 均母體效果模式而言，Liang 與Zeger (1986)所提出的廣義估計方程組方法是最常被引 用於估計模式參數以及重複測量值之間的相關性。然而廣義估計方程組估計量之一致 性僅限於反應變數之遺失值型態為完全隨機。針對長期追蹤二元資料，Spiessens 等人 (2003)應用模擬研究分別比較在固定樣本和群序逐次檢定方法下，邏輯斯隨機效果模式 與GEE 模式之型I 誤差與檢定力。其模擬結果顯示，當遺失值為隨機產生時，使用邏 輯斯隨機效果模式之檢定力較高。為改善GEE 方法之缺乏，數種推廣型之GEE 方法 陸續被提出。Hines and Hines (2005)比較在相關結構錯誤指定之下，未加權與加權後之 GEE 方法在估計偏差、變異數與收斂速度之差異。在此研究計畫中，我們將進一步應 用模擬研究來比較，在長期追蹤順序資料具有隨機遺失型態下，GEE 模式、推廣型GEE 模式與廣義線性混合模式之估計偏差，以及收斂速度等問題。此外，亦將考慮固定樣 本與群序逐次檢定方法之差異。 Longitudinal studies are often designed to investigate the change of a specific characteristic which is measured repeatedly over time. A popular approach for estimating both the regression parameters and correlations in population-averaged models is the generalized estimating equations (GEE) methodology proposed by Liang and Zeger (1986). However, the consistency of GEE estimators for the regression parameters and correlations only hold when the responses are missing completely at random. For longitudinal binary data, Spiessens et al. (2003) compared the two models based on logistic random-effects and GEE with respect to their type I error and power in both fixed sample and group sequential methods, and pointed out that the GEE model had less power when the responses were missing at random. To improve the drawback of the GEE method, some modified GEE procedures are developed. Hines and Hines (2005) compared the unweighted and weighted GEE methods in terms of biases, variance and convergence rates when the correlation structure is misspecified. In this project, the performances of GEE models, modified GEE models and generalized linear mixed model for the analysis of longitudinal ordinal data with MAR drop-outs in terms of bias in estimation, their operation characteristics and convergence rate are compared by simulation. In addition, the testing procedures of fixed samples and group sequential analysis of incomplete longitudinal ordinal data are also discussed
