近幾年來，國人自殺死亡率不斷提高，且自殺死亡從1997年起已連續多年列於國人十大死亡原因之一，所以自殺防治工作刻不容緩。本研究採用自殺防治中心在桃園縣六家居家服務單位(龍祥、中國、仁愛、紅十字、家輔及寬福)所做之問卷調查資料，目的在於找出何種特性者，BSRS5 (The Five-Item Brief Symptom Rating Scale)分數及自殺意念分數可能較高。本研究屬於時序性追蹤資料，自民國96年5月份起，由居服人員針對受測對象進行訪談，大約每隔兩週收集一次，總共進行四次。 針對問卷進行基本敘述性統計、單項排名分析以及交叉分析後發現，在人口特質方面，男女性比例相當，年齡層主要皆在65~84歲，教育程度以不識字及國小為主；在BSRS5五題排名方面，以第一題「睡眠困難(難以入睡或早醒)」的平均分數最高，第四題「覺得比不上別人」平均分數最低；由交叉分析的結果發現身體狀況為一個重要的變數，身體狀況差的人BSRS5總分6分以上或自殺意念2分以上明顯較多。 對資料配適廣義估計方程式及Alternating Logistic Regressions的結果，發現在反應變數為BSRS5總分時，女性、身體狀況差及曾經看過精神科者BSRS5分數達到6分以上的可能性較高。若反應變數為自殺意念時，無論是利用廣義估計方程式或Alternating Logistic Regressions，從模型配適的結果發現只有BSRS5的效應顯著。不管利用BSRS5總分或是各題分開來看，BSRS5對自殺意念是一個相當有效的檢測工具，BSRS5分數愈高則自殺意念2分以上的機會也愈高。此外利用多層結構分析方法配適廣義估計方程式，針對BSRS5與受測次數間的關聯性分析，發現與配適傳統unstructured相關性矩陣的估計結果差異不大，但是可以減少許多參數估計，並且在電腦計算時間上是較快速的。 / In Taiwan, suicide has been among the top ten causes of death since 1997, and suicide prevention has thus attracted much attention since. Using the data provided by Taiwan Suicide Prevention Center (TSPC), this study is aimed to find out possible personal characteristics that might have some impacts on the BSRS5 (the Five-Item Brief Symptom Rating Scale) and suicide ideation scores The data come from a longitudinal study in which subjects from six elderly home service centers in Taoyuan County, Taiwan were visited four times between May and July, 2007, about two weeks between each visit. The total number of subjects is 1981. The proportions of male and female are nearly the same, the age range is from 65 to 84, and most of them have only an elementary school degree. Preliminary analyses indicate that among the five items in BSRS5, insomnia (the first item) is ranked the highest, and inferiority (the fourth item) is the lowest. In addition, health status is highly correlated to the BSRS5 and suicide ideation scores, the worse the health status, the higher the BSRS5 and suicide ideation scores. Fitting the data with Generalized Estimating Equation (GEE) and Alternating Logistic Regressions models with respect to the BSRS5 score, we further find that female, those who have bad health status, and those who have ever consulted a psychiatrist have higher probability that the BSRS5 score is greater than 6. As far as the suicide ideation score is concerned, the BSRS5 score is the only covariate that is statistically significant, an indication that BSRS5 is a useful tool for screening subjects at risk of committing suicide. While the conclusions stay the same whether the data are analyzed through GEE with commonly used unstructured correlation structure or newly developed multiblock and multilayer correlation structure, the latter approach reduces the computer time significantly.

在測驗含有題組(testlet)結構時，由於違反了試題反應理論(Item Response Theory, IRT)中局部獨立性的假設，使得IRT的估計方法產生偏誤，過去研究的解決方式為在IRT模型中多加入一個參數，將題組的影響力納入模型中，此即為題組反應理論(Testlet Response Theory, TRT)，在貝氏(Bayesian)的架構下，此方法的計算則可透過SCORIGHT軟體來達成。本研究旨在透過另一種方法，即廣義方程式(Generalized Estimation Equation, GEE)去處理測驗中的題組效果。GEE過去常被使用於分析縱貫式(longitudinal)的資料，本研究使用此方法來捕捉題組測驗下作答結果的相關性，並經重新參數化調整係數後使其能對受試者能力值進行估計。 電腦模擬的結果顯示GEE能有效的處理題組效果帶來的影響。在GEE和貝氏題組模型的比較上，GEE對於程度好和程度差的受試者有較佳的估計效果；而貝氏題組模型則對於程度中等的受試者表現較好，此外我們也針對GEE的估計效率進行了實驗，結果顯示先將受試者依能力分組再進行GEE估計能提升GEE的估計效率。 在文章中，我們也展示了使用GEE計算題組訊息量的方式，做為題組式測驗下評估該測驗對於各能力區間的受試者在估計準確度上的參考。 / If the tests have testlet structure, the bias may arise when using traditional Item Response Theory(IRT) estimation methods due to the violations to the assumption of local independence. To deal with the testlet effect, previous studies introduced a new parameter to the classical IRT model which called Testlet Response Theory(TRT). Under the Bayesian framework, the estimation can be accomplished on the SCORIGHT program. The purpose of this paper is to use another method named Generalized Estimation Equation(GEE) to model testlet response data. GEE was commonly used to analyze the longitudinal data. We use this method to capture the information from the correlated items and estimated ability of the examinees through re-parametrization. Simulation results indicate that GEE can deal with the testlet effect effectively. On the comparison between GEE and Bayesian testlet model, GEE does better on estimation of the examinees who have high or low ability level. In contrast, Bayesian testlet model does better on estimation of medium ability level. In addition, we design the experiment to test the efficiency of GEE. The results show that group the examinees according to their ability before doing the GEE estimation can improve the efficiency of GEE. In this paper, we also demonstrate the method to calculate testlet information using GEE which can be taken as reference for assessing estimation accuracy of each ability level in testlet-based testing.

當資料屬於類別型的長期追蹤資料（Longitudinal categorical data）時，除了可以透過廣義估計方程式（General estimate equation, GEE）來求解模型參數估計值外，潛在移轉分析（Latent transition analysis, LTA）法也是一種可行的資料分析方法。若資料的期數不多，也可以選擇將資料適度分群後使用羅吉斯迴歸分析（Logistic regression）法。當探討的反應變數為二元（Binary）型態，且觀察對象於每一期提供多個測量變數值的情況之下，廣義估計方程式與羅吉斯迴歸分析法的使用，文獻上常見先將所有的測量變數值加總後，以「中位數」作為分類的切割點。不同於以上兩種方法，潛在移轉分析法則是直接使用原始資料來取得觀察對象的潛在狀態相關訊息，因此與前二者的作法不同，可能導致後續的各項分析結果有所差異存在。 為了能夠了解造成中位數分類法與移轉分析法差異的可能因素，我們架構在潛在移轉分析法的模型下，以不同的參數設定來進行電腦模擬，比較各參數條件下的兩分類方法差異。結果發現各潛在狀態下的測量變數反應機率形式、第一期潛在狀態的組成比例等皆會對兩分類方法是否具有相同分類有所影響。另外，透過分析「青少年媒體使用與健康生活調查」的實際資料得知，潛在移轉分析會將大部分的觀察對象歸屬於「網路成癮」，而中位數分類法則是將大部分的觀察對象歸屬於「無網路成癮」。此外，可以注意到「沮喪」、「線上情色每星期平均使用天數」、及「父母相處狀況」這幾個控制變數與各分組結果的關聯性，於上述三種資料分析方法中有所不同。 / Several methods can be used to analyze longitudinal categorical data, as among them Latent Transition Analysis (LTA), and Generalized Linear Models estimated by Generalized Estimating Equations (GEE) probably the most popular. In addition, if the number of periods is two, then with certain grouping of data, the Logistic Regression can also be applied to perform the analyses. When there are more than one manifest response variable for each study subject, LTA is able to classify the subjects in terms of the original manifest response variables and proceeds with necessary analyses. On the other hand, GEE method and Logistic Regression lack the flexibility, and require certain transformation to transform the manifest response variables into a categorical response variable first. One common way to form a binary response is to sum all manifest variables, and then taking median as a cut-point. In this study, we explore the differences of the classification resulted from LTA directly and using median as a cut-point through simulations. An empirical study is also provided to illustrate the classification differences, and the differences on the subsequent analyses using LTA, GEE method, and Logistic Regression approach.