預測地震後倖存者持續性創傷後壓力症狀的預測模式：比較分類回歸樹與邏輯氏回歸差異

Models to Predict Unresolved Post-traumatic Stress Symptoms in Earthquake Survivors: Comparison of Classifi cation and Regression Tree and Logistic Regression

目的：研究是評估以分類回歸樹與邏輯氏回歸模式去預測在台灣大地震後半年發現的案例，追蹤三年後仍然具有持續性創傷後壓力症狀者與康復者的差異。方法：在地震後半年共有4223的反應者被篩檢，其中有875(20.7%)具有創傷後壓力症狀。三年後此875人，共有418(47.8%)人被追蹤到，其中，38(9.1%)人的症狀仍存在。根據基本資料、生活品質、可能的危險因素與創傷後壓力疾患症狀等因素進行邏輯氏回歸分析，同時也對上述資料進行決定樹分析。使用決定樹分析，我們發現個案如果有明顯經濟損失、心理生活品質分數低於36.0分、以及重複出現創傷經驗，將有53.6%的人無法從創傷後壓力症狀恢復健康。這三項獨立預測因子也包含在邏輯氏回歸結果的六個因子裡面。結論：回歸樹分析證實了某些邏輯氏回歸的結果，研究人員可以從決定回歸樹的方法去學習一些新的知識。

Objective: This study was to evaluate the implementation and performance of a classifi cation and regression tree (CART) and a logistic regression model to predict unresolved post-traumatic stress symptoms (PTSS) in survivors three years after the Taiwan Chi-Chi Earthquake from multivariate data presented at 0.5 year. Methods: We surveyed 4,223 respondents 0.5 year after the earthquake, and 875 (20.7%) of them were found to be positive for PTSS. Three years later, we followed up 418 (47.8%) of the 875 participants, and in 38 (9.1%) of these cases were found to have their symptoms unresolved. Verifi ed values falling outside threshold limits were analyzed according to demographic data, quality of life (QOL), putative risk factors, and post-traumatic stress disorder (PTSD)-related symptoms with the aid of logistic regression. A decision tree was automatically produced from the root nod to target classes (remissive or unresolved PTSS). Result: With CART, we found that the predicted probability for unresolved PTSS was 53.6%, if the respondents had “prominent fi nancial loss,” “mental component summary (MCS) score ≦36.0”, and “reliving the traumatic experience”. These three factors were also included in the six signifi cant independent variables identifi ed in logistic regression. Conclusion: Decision tree analysis confi rmed some of the results of logistic regression. This investigation shows there is knowledge to be gained from analyzing observational data with the aid of decision tree analysis.