查看“︁估计量的偏差”︁的源代码
←
估计量的偏差
跳转到导航
跳转到搜索
因为以下原因,您没有权限编辑该页面:
您请求的操作仅限属于该用户组的用户执行:
用户
您可以查看和复制此页面的源代码。
{{about|统计估计量的偏差|统计学中其他用法|偏差}} {{NoteTA |G1 = Math |T=zh-tw:估計量的偏誤;zh-cn:估计量的偏差 |1 = zh-cn:方差; zh-tw:變異數; |2= zh-cn:无偏; zh-tw: 不偏 |3= zh-cn:有偏; zh-tw:偏誤 |4= zh-cn: 贝塞尔修正; zh-tw:自由度修正 |5= zh-cn:参数; zh-tw:母數 |6= zh-cn:变量;zh-hk:變量;zh-tw: 變數 |7=zh-cn: 偏差;zh-tw:偏誤 }} 在[[统计学]]中,[[估计量]]的'''偏差'''(或'''偏差函数''')是此估计量的[[期望值]]与估计参数的真值之差。偏差为零的估计量或决策规则称为'''无偏的'''。否则该估计量是'''有偏的'''。在统计中,“偏差”是一个函数的客观陈述。 偏差也可以相对于[[中位數]]来衡量,而非相对于均值(期望值),在这种情况下为了与通常的“均值”无偏性区别,称作“中值”无偏。偏差与[[一致估计量|一致性]]相关联,一致估计量都是收敛并且''渐进''无偏的(因此会收敛到正确的值),虽然一致序列中的个别估计量可能是有偏的(只要偏差收敛于零);参见[[一致估计量|偏差与一致性]]。 当其他量相等时,无偏估计量比有偏估计量更好一些,但在实践中,并不是所有其他统计量的都相等,于是也经常使用有偏估计量,一般偏差较小。当使用一个有偏估计量时,也会估计它的偏差。有偏估计量可能用于以下原因:由于如果不对总体进一步假设,无偏估计量不存在或很难计算(如{{le|标准差的无偏估计|unbiased estimation of standard deviation}});由于估计量是中值无偏的,却不是均值无偏的(或反之);由于一个有偏估计量较之无偏估计量(特别是{{le|收缩估计量|shrinkage estimator}})可以减小一些[[损失函数]](尤其是[[均方差]]);或者由于在某些情况下,无偏的条件太强,这种情况无偏估计量不是必要的。此外,在非线性变换下均值无偏性不会保留,不过中值无偏性会保留(参见[[#变换的效应|变换的效应]]);例如[[方差|样本方差]]是总体方差的无偏估计量,但它的平方根[[標準差]]则是总体标准差的有偏估计量。下面会进行说明。 ==定义== 设我们有一个参数为实数 ''θ'' 的[[概率模型]],产生观测数据的概率分布 <math>P_\theta(x) = P(x\mid\theta)</math>,而统计量 <math>\hat\theta</math> 是基于任何观测数据 <math>x</math> 下 ''θ'' 的[[估计量]]。也就是说,我们假定我们的数据符合某种未知分布 <math>P_\theta(x) = P(x\mid\theta)</math>(其中 ''θ'' 是一个固定常数,而且是该分布的一部分,但具体值未知),于是我们构造估计量 <math>\hat\theta</math>,该估计量将观测数据与我们希望的接近 ''θ'' 的值对应起来。因此这个估量的(相对于参数 ''θ''的)'''偏差'''定义为 :<math> \operatorname{Bias}_\theta[\,\hat\theta\,] = \operatorname{E}_\theta[\,\hat{\theta}\,]-\theta = \operatorname{E}_\theta[\, \hat\theta - \theta \,],</math> 其中 <math>\operatorname{E}_\theta</math> 表示分布 <math>P_\theta(x) = P(x\mid\theta)</math> 的[[期望值]],即对所有可能的观测值 <math>x</math> 取平均。由于 ''θ'' 对于条件分布 <math>P(x\mid\theta)</math> 是可测的,就有了第二个等号。 对于参数 ''θ'' 的所有值的偏差都等于零的估计量称为'''无偏'''估计量。 在一次关于估计量性质的模拟实验中,估计量的偏差可以用{{le|平均有符号离差|mean signed difference}}来评估。 ==例子== ===样本方差=== {{main|样本方差}} 随机变量的[[方差|样本方差]]从两方面说明了估计量偏差:首先,自然估计量({{lang|en|naive estimator}})是有偏的,可以通过比例因子校正;其次,无偏估计量的[[均方差]](MSE)不是最优的,可以用一个不同的比例因子来最小化,得到一个比无偏估计量的MSE更小的有偏估计量。 具体地说,自然估计量就是将离差平方和加起来然后除以 ''n'',是有偏的。不过除以 ''n'' − 1 会得到一个无偏估计量。相反,MSE可以通过除以另一个数来最小化(取决于分布),但这会得到一个有偏估计量。这个数总会比 ''n'' − 1 大,所以这就叫做{{le|收缩估计量|shrinkage estimator}},因为它把无偏估计量向零“收缩”;对于正态分布,最佳值为 ''n'' + 1。 设 ''X''<sub>1</sub>, ..., ''X''<sub>''n''</sub> 是[[期望值|期望]]为 ''μ''、[[方差]]为 ''σ''<sup>2</sup> 的[[独立同分布]](i.i.d.)随机变量。如果[[样本均值]]与未修正[[方差|样本方差]]定义为 :<math>\overline{X}=\frac{1}{n}\sum_{i=1}^nX_i, \qquad S^2=\frac{1}{n}\sum_{i=1}^n\left(X_i-\overline{X}\,\right)^2,</math> 则 ''S''<sup>2</sup> 是 ''σ''<sup>2</sup> 的一个有偏估计量,因为 :<math> \begin{align} \operatorname{E}[S^2] &= \operatorname{E}\left[ \frac 1 n \sum_{i=1}^n \big(X_i-\overline{X}\big)^2 \right] = \operatorname{E}\bigg[ \frac 1 n \sum_{i=1}^n \bigg((X_i-\mu)-(\overline{X}-\mu)\bigg)^2 \bigg] \\[8pt] &= \operatorname{E}\bigg[ \frac 1 n \sum_{i=1}^n \bigg((X_i-\mu)^2 - 2(\overline{X}-\mu)(X_i-\mu) + (\overline{X}-\mu)^2\bigg) \bigg] \\[8pt] &= \operatorname{E}\bigg[ \frac 1 n \sum_{i=1}^n (X_i-\mu)^2 - \frac 2 n (\overline{X}-\mu) \sum_{i=1}^n (X_i-\mu) + \frac 1 n (\overline{X}-\mu)^2 \sum_{i=1}^n 1 \bigg] \\[8pt] &= \operatorname{E}\bigg[ \frac 1 n \sum_{i=1}^n (X_i-\mu)^2 - \frac 2 n (\overline{X}-\mu)\sum_{i=1}^n (X_i-\mu) + \frac 1 n (\overline{X}-\mu)^2 \cdot n\bigg] \\[8pt] &= \operatorname{E}\bigg[ \frac 1 n \sum_{i=1}^n (X_i-\mu)^2 - \frac 2 n (\overline{X}-\mu)\sum_{i=1}^n (X_i-\mu) + (\overline{X}-\mu)^2 \bigg] \\[8pt] \end{align} </math> 换句话说,未修正的样本方差的期望值不等于总体方差 ''σ''<sup>2</sup>,除非乘以归一化因子。而样本均值是总体均值 ''μ'' 的无偏<ref name="JohnsonWichern2007">{{cite book|author1=Richard Arnold Johnson|author2=Dean W. Wichern|title=Applied Multivariate Statistical Analysis|url=https://books.google.com/books?id=gFWcQgAACAAJ|accessdate=10 August 2012|year=2007|publisher=Pearson Prentice Hall|isbn=978-0-13-187715-3|archive-date=2016-05-29|archive-url=https://web.archive.org/web/20160529154017/https://books.google.com/books?id=gFWcQgAACAAJ|dead-url=no}}</ref>估计量。 ''S''<sup>2</sup> 是有偏的原因源于样本均值是 ''μ'' 的{{le|普通最小二乘|ordinary least squares}}(OLS)估计量这个事实:<math>\overline{X}</math> 是令 <math>\sum_{i=1}^n (X_i-\overline{X})^2</math> 尽可能小的数。也就是说,当任何其他数代入这个求和中时,这个和只会增加。尤其是,在选取 <math>\mu \ne \overline{X}</math> 就会得出, :<math> \frac{1}{n}\sum_{i=1}^n (X_i-\overline{X})^2 < \frac{1}{n}\sum_{i=1}^n (X_i-\mu)^2, </math> 于是 :<math> \begin{align} \operatorname{E}[S^2] &= \operatorname{E}\bigg[ \frac{1}{n}\sum_{i=1}^n (X_i-\overline{X})^2 \bigg] < \operatorname{E}\bigg[ \frac{1}{n}\sum_{i=1}^n (X_i-\mu)^2 \bigg] = \sigma^2. \end{align} </math> 注意到,通常的样本方差定义为 :<math>s^2=\frac{1}{n-1}\sum_{i=1}^n(X_i-\overline{X}\,)^2,</math> 而这时总体方差的无偏估计量。可以由下式看出: :<math>\operatorname{E}\big[ (\overline{X}-\mu)^2 \big] = \frac{1}{n}\sigma^2 .</math> 方差的有偏(未修正)与无偏估计之比称为{{le|贝塞尔修正|Bessel's correction}}。 ==参见== *[[点估计|點估計]] * {{le|忽略变量偏差|Omitted-variable bias}} * [[一致估计量]] * [[估计理论]] * {{le|期望损失|Expected loss}} * [[期望值]] * [[损失函数]] * [[中位數]] * [[决策论]] * {{le|乐观偏差|Optimism bias}} {{portal box|科学|统计学}} ==参考文献== * [https://web.archive.org/web/20110310043642/http://www.universityofcalifornia.edu/senate/inmemoriam/georgewbrown.htm Brown, George W.] "On Small-Sample Estimation." ''The Annals of Mathematical Statistics'', vol. 18, no. 4 (Dec., 1947), pp. 582–585. {{JSTOR|2236236}}. * {{tsl|en|Erich Leo Lehmann||Lehmann, E. L.}} "A General Concept of Unbiasedness" ''The Annals of Mathematical Statistics'', vol. 22, no. 4 (Dec., 1951), pp. 587–592. {{JSTOR|2236928}}. * {{tsl|en|Allan Birnbaum||Allan Birnbaum}}, 1961. "A Unified Theory of Estimation, I", ''The Annals of Mathematical Statistics'', vol. 32, no. 1 (Mar., 1961), pp. 112–135. * Van der Vaart, H. R., 1961. "Some Extensions of the Idea of Bias" ''The Annals of Mathematical Statistics'', vol. 32, no. 2 (June 1961), pp. 436–447. * Pfanzagl, Johann. 1994. ''Parametric Statistical Theory''. Walter de Gruyter. * {{cite book | last1 = Stuart | first1 = Alan | first2 = Keith | last2 = Ord | first3=Steven [F.] | last3=Arnold | title=Classical Inference and the Linear Model | series = Kendall's Advanced Theory of Statistics | volume = 2A | year = 2010 | publisher = Wiley | isbn=0-4706-8924-2 }}. * {{cite book |last1 = Voinov |first1 = Vassily [G.] |last2= Nikulin |first2= Mikhail [S.] |title= Unbiased estimators and their applications| volume = 1: Univariate case |year = 1993 |publisher = Kluwer Academic Publishers | location= Dordrect | ISBN= 0-7923-2382-3}} * {{cite book |last1 = Voinov |first1 = Vassily [G.] |last2= Nikulin |first2= Mikhail [S.] |title= Unbiased estimators and their applications| volume = 2: Multivariate case |year = 1996 |publisher = Kluwer Academic Publishers | location= Dordrect| ISBN= 0-7923-3939-8}} * {{cite book | last1 = Klebanov |first1 = Lev [B.] | last2 = Rachev |first2 = Svetlozar [T.] | last3 = Fabozzi |first3 = Frank [J.] | title = Robust and Non-Robust Models in Statistics | year = 2009 |publisher = Nova Scientific Publishers | location = New York | ISBN = 978-1-60741-768-2}} ==外部链接== * {{springer|title=Unbiased estimator|id=p/u095070}} {{認知偏誤}} [[Category:統計理論]] [[Category:估计理论]] [[category:認知偏誤]]
该页面使用的模板:
Template:About
(
查看源代码
)
Template:Cite book
(
查看源代码
)
Template:JSTOR
(
查看源代码
)
Template:Lang
(
查看源代码
)
Template:Le
(
查看源代码
)
Template:Main
(
查看源代码
)
Template:NoteTA
(
查看源代码
)
Template:Portal box
(
查看源代码
)
Template:Springer
(
查看源代码
)
Template:Tsl
(
查看源代码
)
Template:認知偏誤
(
查看源代码
)
返回
估计量的偏差
。
导航菜单
个人工具
登录
命名空间
页面
讨论
不转换
查看
阅读
查看源代码
查看历史
更多
搜索
导航
首页
最近更改
随机页面
MediaWiki帮助
特殊页面
工具
链入页面
相关更改
页面信息