| Abstract: | 本論文旨在“利用肥胖指標預測代謝症候相關疾病最適切點” 的探討。以邏輯斯迴歸模型(logistic regression model)，使得敏感度(sen.)和特異度(spe.)的勝算比值(odds ratio)最大，依據此方法找最敏感的合適預測切點。
根據國民健康訪問(NHIS)調查，以「家戶」為最終抽樣單位，我們蒐集年齡層從35歲到65歲的研究樣本，共3404人。分別記錄每個個體BMI值、腰圍(WC)、腰臀比(WHR)，並調查有否吸菸、飲酒，是否罹患高血壓、高膽固醇、高三酸甘油脂、高尿酸、糖尿病、以及是否為代謝症候群。
由於樣本之間有家族成員關係，不符獨立假說，因此採取GEE(generalized estimating equations )分析方法，考慮抽菸、飲酒健康危險因子以及年齡層，利用肥胖指標對各項疾病建立邏輯斯迴歸模型(logistic regression model)，有別於先前文獻中，以ROC曲線中敏感度(sen.)和特異度(spe.)的總和最大值，和以樣本中肥胖指標的第75百分比為找切點的方法。
分析結果顯示，以本研究方法建立的模型，與以往的方法相比較，產生結果最敏感的合適切點有等級的現象；另外，此方法不會受探討變項的多寡與性質而使得分析模型複雜化。最後，兩種方法生的預測切點結果有一致性。
This thesis aims at finding the optimal cutoffs of common obesity indices (BMI, WC, WHR) in predicting metabolic related disorders. Given a cutoff of a specified obesity index and disease status, the odds of sensitivity and specificity are obtained from logistic regression model using generalized estimating equations (GEE). Individual smoking status, age group, gender, and disease status were adjusted as covariables in the regression model. The odds ratio of sensitivity and specificity which terns out to be the coefficient of the disease status can then be estimated and be applied as an indicator to find the optimal cutoff as one moves across the support possible cutoffs. Individual smoking status, age group, gender, and disease status.
The data sets for analysis is the outcomes of the 2001 National Health Interview Survey (NHIS) and the 2002 three high surveillance study, which has all members residing within the same household as a sampling unit. There were totally 3404 individuals aged 35-65 in the study. Individual BMI, waist circumference (WC), and waist-to-hip ratio (WHR) and the status of metabolic disorders (high blood pressure, diabetes, high cholesterol, hypertriglyceridemia, hyperuricemia ) were obtained from these individuals.
The proposed method is compared with the commonly adopted method by finding the maximum sum of sensitivity (sen.) and specificity (spe.) from empirical ROC. The result showed that the proposed method could automatically adjust for other covariables and generate several peaks as candidate sensitive zones of optimal cutoffs, which the maximum sum method could only yield one peak and is dependent on covariates. In general, the outcomes of the two methods are in consistent. |
