陳麗曲,黃介武
對(duì)數(shù)正態(tài)分布序列單均值變點(diǎn)的識(shí)別和估計(jì)
陳麗曲,黃介武
(貴州民族大學(xué) 數(shù)據(jù)科學(xué)與信息工程學(xué)院,貴州 貴陽(yáng) 550025)
針對(duì)數(shù)據(jù)呈現(xiàn)偏態(tài)分布且存在變點(diǎn)的情況,構(gòu)建對(duì)數(shù)正態(tài)分布的單均值變點(diǎn)模型,給出分布的均值單變點(diǎn)模型的似然函數(shù),并采用極大似然方法和貝葉斯方法對(duì)變點(diǎn)位置進(jìn)行識(shí)別和估計(jì).通過模擬比較研究,這兩種方法都能有效地估計(jì)變點(diǎn)位置,在標(biāo)準(zhǔn)差和相對(duì)誤差準(zhǔn)則下,貝葉斯方法比極大似然方法效果更理想.其中共軛先驗(yàn)分布下的貝葉斯方法較無信息先驗(yàn)下的貝葉斯方法識(shí)別和估計(jì)變點(diǎn)位置表現(xiàn)更優(yōu).
單均值變點(diǎn);對(duì)數(shù)正態(tài)分布;貝葉斯方法;極大似然方法
變點(diǎn)是指一隨機(jī)序列數(shù)據(jù)在某一個(gè)位置或時(shí)間點(diǎn)發(fā)生了分布或數(shù)字特征突然變化的點(diǎn)[1],如果不考慮對(duì)變點(diǎn)進(jìn)行識(shí)別和估計(jì)就進(jìn)行統(tǒng)計(jì)分析則很可能導(dǎo)致結(jié)果具有誤導(dǎo)性.變點(diǎn)問題多集中于基于正態(tài)分布的變點(diǎn)模型,包括正態(tài)均值、方差變點(diǎn)模型,正態(tài)單變點(diǎn)、多變點(diǎn)模型,正態(tài)線性回歸模型中回歸系數(shù)以及誤差方差的變點(diǎn)模型.文獻(xiàn)[2-4]運(yùn)用似然比方法和貝葉斯方法詳細(xì)討論了一元正態(tài)以及多元正態(tài)均值或方差變點(diǎn)模型、正態(tài)回歸變點(diǎn)模型中回歸系數(shù)變點(diǎn)位置的識(shí)別和估計(jì).文獻(xiàn)[5]介紹了正態(tài)以及其他常見變點(diǎn)模型識(shí)別和估計(jì)的理論和方法,以及變點(diǎn)模型在各領(lǐng)域中的應(yīng)用.文獻(xiàn)[6]針對(duì)具有異質(zhì)方差的正態(tài)模型的多變點(diǎn)估計(jì)問題,提出了一種基于貝葉斯模型選擇過程的變點(diǎn)估計(jì)方法.趙江南[7]等運(yùn)用ASAMC方法討論了變點(diǎn)個(gè)數(shù)未知情況下正態(tài)分布變點(diǎn)的個(gè)數(shù)及位置識(shí)別.
針對(duì)數(shù)據(jù)呈現(xiàn)偏態(tài)且存在變點(diǎn)的情況,本文建立對(duì)數(shù)正態(tài)分布的單均值變點(diǎn)模型,探討極大似然方法和貝葉斯方法在標(biāo)準(zhǔn)差和相對(duì)誤差準(zhǔn)則下識(shí)別和估計(jì)變點(diǎn)位置的優(yōu)良性.
模型(2)的似然函數(shù)為
當(dāng)使用貝葉斯方法分析問題時(shí),為了獲得更準(zhǔn)確和可靠的統(tǒng)計(jì)推斷結(jié)果,往往需要引入恰當(dāng)?shù)南闰?yàn)分布.然而,在實(shí)際應(yīng)用中,有時(shí)缺乏關(guān)于先驗(yàn)分布的知識(shí)或信息,并且即使擁有這些知識(shí),也可能不足以提供全面的先驗(yàn)分布.因此,引入無信息先驗(yàn)分布、共軛先驗(yàn)分布往往是一種合理可靠的選擇.
取無信息先驗(yàn)分布
于是各參數(shù)的聯(lián)合后驗(yàn)分布為
從而
類似地,有
從而
類似地,有
運(yùn)用R軟件進(jìn)行對(duì)數(shù)正態(tài)分布極大似然方法的隨機(jī)模擬,結(jié)果見表1.
表1 對(duì)數(shù)正態(tài)分布各參數(shù)的極大似然估計(jì)
圖1 極大似然估計(jì)分布
圖2 極大似然方法下變點(diǎn)k的軌跡
表2 兩種先驗(yàn)分布下對(duì)數(shù)正態(tài)分布各參數(shù)的貝葉斯估計(jì)
表3 兩種先驗(yàn)分布下不同變點(diǎn)位置的貝葉斯估計(jì)
圖3 無信息先驗(yàn)下的Gibbs抽樣迭代過程及多層迭代鏈軌跡
圖4 共軛先驗(yàn)下的Gibbs抽樣迭代過程及多層迭代鏈軌跡
4.3.1極大似然方法數(shù)和貝葉斯方法的比較由表1~2可以看出,本文采用的極大似然方法數(shù)和貝葉斯方
本文通過構(gòu)建對(duì)數(shù)正態(tài)單均值變點(diǎn)模型,采用極大似然方法和貝葉斯方法探討了數(shù)據(jù)呈現(xiàn)偏態(tài)分布且存在變點(diǎn)的問題.通過研究比較可知,相較于極大似然方法,貝葉斯方法估計(jì)變點(diǎn)位置更精準(zhǔn).特別是在共軛先驗(yàn)分布下對(duì)數(shù)正態(tài)單均值變點(diǎn)的貝葉斯方法估計(jì)變點(diǎn)位置效果是最優(yōu)的.
[1] PAGE E S.Continuous inspection schemes[J].Biometrika,1954,41(1/2):100-115.
[2] CHEN J.Inference About the Change Points in a Sequence of Gaussain Random Vectors Using Information Criterion[D].Bowling Green:Bowling Green State University,1995.
[3] CHEN J,Gupta A K.Testing and locating variance changepoints with application to stock prices[J].Journal of The American Statistical Association,1997,92(438):739-747.
[4] CHEN J,Gupta A K.On change point detection and estimation[J].Communications in Statistics Simulation and Computation, 2001,30(3):665-697.
[5] Gurevich G,Vexler A.Retrospective change point detection:from parametric to distribution free policies[J].Communications in Statistics Simulation and Computation,2010,39(5):899-920.
[6] KANG S G,Lee W D,Kim Y.Bayesian Multiple Change-Points Detection in a Normal Model with Heterogeneous Variances[J].Computational Statistics,2021,36(12):1365-1390.
[7] 趙江南,周菊玲.正態(tài)分布序列變點(diǎn)識(shí)別問題探討[J].統(tǒng)計(jì)與決策,2019,35(15):13-17.
[8] Fisher R A.Statistical Methods in Scientific Inference[M].London:Oliver and Boyd,1956.
[9] Bickel P J,Doksum K A.Mathematical Statistics:Basic and Selected Topics(Vol 1)[M].San Francisco:Holden-Day,1977.
[10] ZHANG Fengqing,GOU Jiangtao.A unified framework for estimation in lognormal models[J].Journal of Business & Economic Statistics,2022,40(4):1583-1595.
[11] Lyu J,Nadarajah S.Discrete lognormal distributions with application to insurance data[J].International Journal of System Assurance Engineering and Management,2021(1):1-15.
[12] 陶鶴,劉偉,付晶園.對(duì)數(shù)正態(tài)分布可靠性模型及應(yīng)用[J].統(tǒng)計(jì)與決策,2019,35(3):89-92.
[13] Jandhyala V K, Fotopoulos S B. Estimating the unknown change point in the parameters of the lognormal distribution[J]. Environmetrics,2007,18(2):141-155.
Identification and estimation of single mean change points in lognormal distribution sequences
CHEN Liqu,HUANG Jiewu
(School of Data Science and Information Engineering,Guizhou Minzu University,Guiyang 550025,China)
In response to the situation where the data presents a skewed distribution and there are change points,a single mean change point model with a lognormal distribution is constructed.The likelihood function of the mean single change point model for this distribution is given,and the maximum likelihood method and Bayesian method are used to identify and estimate the position of the change points.Through simulation and comparative research,both methods can effectively estimate the position of the change point.Under the criteria of standard deviation and relative error,the Bayesian method is more effective than the maximum likelihood method.The Bayesian method under conjugate prior distribution performs better in identifying and estimating the position of change points compared to the Bayesian method without prior information.
single mean change point;lognormal distribution;Bayesian method;maximum likelihood method
1007-9831(2023)12-0015-07
O212.1
A
10.3969/j.issn.1007-9831.2023.12.003
2023-05-28
貴州省科技計(jì)劃基金項(xiàng)目(黔科合基礎(chǔ)[2017]1083號(hào));貴州省教育廳自然科學(xué)研究項(xiàng)目(黔教技[2022]015號(hào))
陳麗曲(1997-),女,貴州道真人,在讀碩士研究生,從事統(tǒng)計(jì)模型與統(tǒng)計(jì)計(jì)算研究.E-mail:2290167300@qq.cm
黃介武(1977-),男,湖南益陽(yáng)人,教授,博士,從事統(tǒng)計(jì)模型與統(tǒng)計(jì)計(jì)算研究.E-mail:846221886@qq.com