臺大學術期刊資料庫NTU Journal Database

模型設定、參數估計與單根檢定量的關係──模擬分析

Model Specification, Estimation and Unit Root Tests--Simulation Analysis

許怡隆(Yilong Xu);汪義育(Yi-Yu Wang)

國立政治大學國貿研究所博士候選人(Department of International Business, National Chengchi University);國立政治大學國貿研究所(Department of International Business, National Chengchi University)

卷期
28卷2期
出版
2000年06月
頁碼
203–239
編號
NTUEJ00007002
外部連結本篇全文由原出版單位的網站提供。

連結出處:國立台灣大學經濟學系。連結會開啟本資料庫以外的網站,能否閱讀由該網站決定。

摘要

本研究之主要目的在探討影響單根檢定量SIZE之原因。應用參數化短期動態過程與效率估計確定趨勢係數值之概念,本文分別建立PSLH與TDK-AR兩個新的檢定量。模擬結果顯示:PSLH檢定量在MA(1)序列相關模型之SIZE表現相對較ADF-GLS、PSL檢定量為佳,在小樣本下,PSLH檢定量與MZ-GLS檢定量有相近之SIZE績效;其整體SIZE表現則相對較這些檢定量為佳。而TDK-AR檢定量除了在低序列相關模型有很好之SIZE表現外,其調整SIZE後之檢定力普遍高於ADF-GLS、MZ-GLS檢定量。本文其它模擬結果顯示:確定趨勢項係數值之估計不僅影響單根檢定量之檢定力,同時也影響檢定量之SIZE。確定趨勢項係數值之估計對長期變異數W2AR,修正統計量之SIZE與檢定力也有很重之影響。以W2RMLE估計式所建立之修正統計量SIZE扭曲較小,但檢定力較差,而以W2GLS估計式所建立之修正統計量SIZE扭曲較大,但檢定力較佳。以參數化短期動態過程處理序列相關所建立之檢定量,SIZE表現優於未參數化短期動態過程之修正統計量。現有的單根檢定量多可視為是修正AR(1)係數估計偏誤之檢定量。同時處理AR(1)係數估計與序列相關偏誤之檢定量會較只考慮序列相關偏誤之檢定量會較只考慮序列相關偏誤之檢定量有較好之SIZE表現。

Abstract

It is well know that majority of the unit root tests generally suffer from severe size distortions and have low power in a high serial correlation model. In this paper, we study the effects of model specification and parameter estimation for unit root tests. Applying the concepts of parametric dynamic process and efficient estimation of deterministic trend coefficients, two new unit root tests: PSLH and TDK-AR are constructed. Our Monte Carlo study results show that the PSLH test has better size performance than ADF-GLS and PSL tests and comparable with the MZ-GLS test in an MA(1) correlation model. Overall, the PSLH test hasa more satisfactory size performance than all tests in the study. The TDK-AR test has reasonalbe size performance in a low serial correlation model and in general has higher size-adjusted power than ADF-GLS and MZ-GLS tests. Other simulation results also show that deterministic trend estimated methods not only affect the power but also the size of the unit tests. Different detrending methods affect the long run varince estimation and the size of unit root tests. In a serial correlation model, tests using parametric methods seem to have better size performance results than semi-parametcic methods. We also find that most unit root tests can be An econometrician will have to look at this, I wonld "writtan as a modified AR(1) with coefficient bias," but I am not sure. I may be twisting the meaning and tests which consider the AR(1) coefficient bias and serial correlation bias at the same time may have better size properties.

關鍵詞

程式與 AI 助手可以改用資料介面(JSON、MCP),查詢比讀頁面快。給程式與 AI 代理的存取說明