GEE之敏感度分析－偵測高影響之觀察值

Sensitivity Analysis in Gee-Identification of High Influential Observations

10.6288/CJPH1996-15-05-01

張玉坤(Yue-Cune Chang)

敏感度分析 ； 高影響觀察值 ； 廣義線性模型 ； 長期資料 ； sensitivity analysis ； high influential observations ； generalized linear models ； longitudinal data

中華公共衛生雜誌

15卷5期（1996 / 10 / 01）

403 - 410

繁體中文

高影響觀察值(high influential observations)的確認(identification)在統計迴歸模型(regression model)的應用上有其不容置疑重要性。在早期的一般線性模型(general linear model)及近幾年來被多位學者廣泛探討的廣義線性模型(generalized linear model)中，對此問題已有多篇論文發表。但是，處理長期資料(longitudinal data)線性模型的統計方法－GEE(generalized estimating equation)[1]，對此問題至今尚未見任何有開之論文刊載。本文對此問題提出一個簡單可行的圖形判讀法，並將原來的SAS/IML Macro程式，GEE1，加以修改後納入此項功能，以利原使用者之應用。我們也成功地將此方法應用在台灣省立新竹醫院眼科的一組資料上。

The importance of identification of high influential observations in the applications of regression model is indubitable. There are a lot of related papers published for the general linear model and the generalized linear model as well. However, for the longitudinal data analysis, we haven't seen any literature published yet. In this paper, we proposed a simple graphic method to handle this sensitivity analysis problem. We also modified the original longitudinal data analysis SASIIML macro program, GEE1, to include the proposed graphic method. For those GEE1 user, the modified macro program is easy to use. We successfully applied this graphic method to analyze a real data set which was conducted by the Provincial Hsin-chu hospital in Taiwan.