Estimation and model selection for left-truncated and right-censored data: Application to power transformer lifetime modeling

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/60085

Title:	Estimation and model selection for left-truncated and right-censored data: Application to power transformer lifetime modeling
Authors:	徐紹凱;Shiu,Shau-Kai
Contributors:	統計研究所
Keywords:	左截略;右設限;對數常態分佈;韋伯分佈;牛頓-拉弗森演算法;最大期望演算法;赤池信息量準則;Left truncation;right censoring;lognormal distribution;Weibull distribution;Newton-Raphson algorithm;EM algorithm;Akaike's information criterion
Date:	2013-06-26
Issue Date:	2013-07-10 12:05:01 (UTC+8)
Publisher:	國立中央大學
Abstract:	電力變壓器壽命資料經常發生左截略(left truncation)與右設限(right censoring)的情況。Balakrishnan and Mitra (2011 JSPI, 2012 CSDA)針對截略與設限資料，分別以對數常態(lognormal)與韋伯(Weibull)為模型，提出利用最大期望演算法(EM algorithm)得到最大概似估計量(maximum likelihood estiamtion)。我們以模擬研究來顯示牛頓-拉弗森演算法(Newton-Raphson algorithm)與最大期望演算法在參數估計上的優劣。我們發現以對數常態分配為模型，當樣本數較小與設限比率較高時，使用牛頓法估計參數容易發生發散的情形。然而，在樣本數較大的情況下，牛頓法比最大期望演算法有較快的收斂速度並能得到較準確的估計量。另外，我們使用赤池信息量準則(Akaike's information criterion)來選擇最合適的模型。 Left truncation and right censoring often occurs in power transformer lifetime data. Suitably adjusted for censoring and truncation, the maximum likelihood estimation has been proposed with the EM algorithm under the lognormal and Weibull models (Balakrishnan and Mitra, 2011 JSPI, 2012 CSDA). In this thesis, we compare the performance of the Newton-Raphson algorithm with their EM algorithm by simulations. Our comparison based on Monte Carlo simulations shows that the Newton-Raphson method for lognormal distribution fails to converge frequently when the sample size is small and the percentage of censoring is high. However, we observe that the Newton-Raphson method has a faster rate of convergence and give more accurate standard error estimates than the EM with missing information principle for moderate sample sizes. In addition, we examine the performance of the Akaike's information criterion (AIC) for selecting a best distribution among candidate models. Finally, these methods discussed here are illustrated through real data examples.
Appears in Collections:	[Graduate Institute of Statistics] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	812	View/Open

社群 sharing

Loading...