請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/72636
標題: | 應用NRI衡量臺灣自殺企圖者受訪意願之預測模型 Application of Net Reclassification Improvement to Evaluate Aftercare Willingness Predictive Models for Suicide Attempters in Taiwan |
作者: | Shao-Kai Lee 李韶凱 |
指導教授: | 劉力瑜(Li-Yu Daisy Liu) |
共同指導教授: | 吳佳儀(Chia-Yi Wu) |
關鍵字: | 自殺企圖,預測模型,通報單位,接收者操作特徵曲線,陰性結果, suicide attempt,predictive model,reporting organization,ROC curve,negative outcome, |
出版年 : | 2019 |
學位: | 碩士 |
摘要: | 本研究以2016年自殺通報數據為例,經由變數選擇方法篩選具統計意義之變數,旨在建立模型以預測通報個案是否願意接受訪視。最初以羅吉斯迴歸建立預測模型時,其接收者操作特徵 (Receiver Operating Characteristic, ROC)曲線下面積 (Area Under Curve, AUC) 值為0.65,並未能有效區分企圖者接受訪視的意願。研究者納入「通報單位別」變數於羅吉斯迴歸模型以建立新的風險預測模型,採方法比較模型(Net Reclassification Improvement, NRI)比較表現差異;風險預測模型的判定閾值則由事發率以及ROC曲線方法決定之。結果顯示當採用切點型NRI加入「通報單位別」後,對陰性正確結果有0.7%的提升,採連續型NRI則呈現75%的不願接受訪視者其預測事發機率降低,兩方式有一致的結果;而加權後NRI (weighted NRI, wNRI) 能提供不同陰性(即不願受訪)與陽性(即願意受訪)的權重比例,會有不同的判定閾值設定,若陰性結果重要性為陽性結果的兩倍時,「通報單位別」的納入對模型預測有幫助;當重視陽性結果過於陰性時,wNRI則為負值,代表新變數的納入反倒使模型預測效果降低。此通報檔的紀錄並未代表自殺企圖者真實受訪情形,因此本研究結果僅作為統計方法的測試,針對於陰性預測結果的探討,可作為同為注重陰性預測研究之參考。 Our research used the National Suicide Surveillance System records reported in 2016 and selected the statistically significant variables to build up aftercare willingness predictive models for suicide attempters. The area under curve (AUC) equals to 0.65 by a logistic regression analysis indicating that the model could not well predict attempters’ willing of being visited. While the variable “rpt_unit_cat” recording the organization being responsible to report the suicide attempt was taken into concerned and built a new model to compare the difference of performance between the old and new models by the methods called Net Reclassification Improve (NRI). The discriminant threshold of risk prediction models was set as the incidence observed from the data and the receiver operating characteristic (ROC) curve method. The negative part of categorical NRI shows 0.7% increasing in the true negative outcome after adding “rpt_unit_cat”. Similarly, the negative part of continuous NRI had 75% improvement of prediction for people who rejected to be visited. On the other hand, wNRI is capable to provide different weights on negative and positive outcome depending on discriminant threshold. In this study, setting negative outcome is two times more important than the positive outcome and prediction is improved after adding “rpt_unit_cat”. When paying more attention to positive outcome, wNRI is negative meaning the variable is not good for prediction of positive outcomes. The system records do not present the visited circumstance of attempters in reality. Therefore, the testing result of our research is used to discuss the prediction of negative outcome, which can be served as a reference for the study focusing on negative prediction. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/72636 |
DOI: | 10.6342/NTU201902227 |
全文授權: | 有償授權 |
顯示於系所單位: | 農藝學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-108-1.pdf 目前未授權公開取用 | 1.8 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。