第 13 章补充材料
S13.1 随机模型的期望均方 ¶ 考虑正文式(13.1)的两因子随机效应平衡方差分析模型:
y i j k = μ + τ i + β j + ( τ β ) i j + ϵ i j k , i = 1 , … , a , j = 1 , … , b , k = 1 , … , n . y_{ijk}=\mu+\tau_i+\beta_j+(\tau\beta)_{ij}+\epsilon_{ijk},\qquad i=1,\ldots,a,\quad j=1,\ldots,b,\quad k=1,\ldots,n. y ijk = μ + τ i + β j + ( τ β ) ij + ϵ ijk , i = 1 , … , a , j = 1 , … , b , k = 1 , … , n . 正文式(13.3)列出了期望均方,但没有正式推导。直接应用期望算子即可求得。例如,
E ( M S A ) = E ( S S A a − 1 ) = E ( S S A ) a − 1 . E(MS_A)=E\left(\frac{SS_A}{a-1}\right)=\frac{E(SS_A)}{a-1}. E ( M S A ) = E ( a − 1 S S A ) = a − 1 E ( S S A ) . S S A SS_A S S A 是行因子的平方和。模型中的 τ i , β j , ( τ β ) i j \tau_i,\beta_j,(\tau\beta)_{ij} τ i , β j , ( τ β ) ij 独立服从均值为零、方差分别为 σ τ 2 , σ β 2 , σ τ β 2 \sigma_\tau^2,\sigma_\beta^2,\sigma_{\tau\beta}^2 σ τ 2 , σ β 2 , σ τ β 2 的正态分布。平方和及其期望为
S S A = 1 b n ∑ i = 1 a y i ⋅ ⋅ 2 − y ⋯ 2 a b n , E ( S S A ) = 1 b n E ( ∑ i = 1 a y i ⋅ ⋅ 2 ) − 1 a b n E ( y ⋯ 2 ) . SS_A=\frac1{bn}\sum_{i=1}^ay_{i\cdot\cdot}^2-\frac{y_{\cdots}^2}{abn},\qquad
E(SS_A)=\frac1{bn}E\left(\sum_{i=1}^ay_{i\cdot\cdot}^2\right)-\frac1{abn}E(y_{\cdots}^2). S S A = bn 1 i = 1 ∑ a y i ⋅⋅ 2 − abn y ⋯ 2 , E ( S S A ) = bn 1 E ( i = 1 ∑ a y i ⋅⋅ 2 ) − abn 1 E ( y ⋯ 2 ) . 由于
y i ⋅ ⋅ = b n μ + b n τ i + n β ⋅ + n ( τ β ) i ⋅ + ϵ i ⋅ ⋅ , y_{i\cdot\cdot}=bn\mu+bn\tau_i+n\beta_\cdot+n(\tau\beta)_{i\cdot}+\epsilon_{i\cdot\cdot}, y i ⋅⋅ = bn μ + bn τ i + n β ⋅ + n ( τ β ) i ⋅ + ϵ i ⋅⋅ , 利用随机分量均值为零以及相互独立的性质,有
1 b n E ( ∑ i y i ⋅ ⋅ 2 ) = 1 b n [ a ( b n μ ) 2 + a ( b n ) 2 σ τ 2 + a b n 2 σ β 2 + a b n 2 σ τ β 2 + a b n σ 2 ] = a b n μ 2 + a b n σ τ 2 + a n σ β 2 + a n σ τ β 2 + a σ 2 . \begin{aligned}
\frac1{bn}E\left(\sum_i y_{i\cdot\cdot}^2\right)
&=\frac1{bn}\left[a(bn\mu)^2+a(bn)^2\sigma_\tau^2+abn^2\sigma_\beta^2+abn^2\sigma_{\tau\beta}^2+abn\sigma^2\right]\\
&=abn\mu^2+abn\sigma_\tau^2+an\sigma_\beta^2+an\sigma_{\tau\beta}^2+a\sigma^2.
\end{aligned} bn 1 E ( i ∑ y i ⋅⋅ 2 ) = bn 1 [ a ( bn μ ) 2 + a ( bn ) 2 σ τ 2 + ab n 2 σ β 2 + ab n 2 σ τ β 2 + abn σ 2 ] = abn μ 2 + abn σ τ 2 + an σ β 2 + an σ τ β 2 + a σ 2 . 另一方面,
y ⋯ = a b n μ + b n τ ⋅ + a n β ⋅ + n ( τ β ) ⋅ ⋅ + ϵ ⋯ , y_{\cdots}=abn\mu+bn\tau_\cdot+an\beta_\cdot+n(\tau\beta)_{\cdot\cdot}+\epsilon_{\cdots}, y ⋯ = abn μ + bn τ ⋅ + an β ⋅ + n ( τ β ) ⋅⋅ + ϵ ⋯ , 所以
1 a b n E ( y ⋯ 2 ) = 1 a b n [ ( a b n μ ) 2 + a ( b n ) 2 σ τ 2 + b ( a n ) 2 σ β 2 + a b n 2 σ τ β 2 + a b n σ 2 ] = a b n μ 2 + b n σ τ 2 + a n σ β 2 + n σ τ β 2 + σ 2 . \begin{aligned}
\frac1{abn}E(y_{\cdots}^2)
&=\frac1{abn}\left[(abn\mu)^2+a(bn)^2\sigma_\tau^2+b(an)^2\sigma_\beta^2+abn^2\sigma_{\tau\beta}^2+abn\sigma^2\right]\\
&=abn\mu^2+bn\sigma_\tau^2+an\sigma_\beta^2+n\sigma_{\tau\beta}^2+\sigma^2.
\end{aligned} abn 1 E ( y ⋯ 2 ) = abn 1 [ ( abn μ ) 2 + a ( bn ) 2 σ τ 2 + b ( an ) 2 σ β 2 + ab n 2 σ τ β 2 + abn σ 2 ] = abn μ 2 + bn σ τ 2 + an σ β 2 + n σ τ β 2 + σ 2 . 收集各项并除以 a − 1 a-1 a − 1 ,得到
E ( M S A ) = ( a − 1 ) σ 2 + n ( a − 1 ) σ τ β 2 + b n ( a − 1 ) σ τ 2 a − 1 = σ 2 + n σ τ β 2 + b n σ τ 2 , \begin{aligned}
E(MS_A)&=\frac{(a-1)\sigma^2+n(a-1)\sigma_{\tau\beta}^2+bn(a-1)\sigma_\tau^2}{a-1}\\
&=\sigma^2+n\sigma_{\tau\beta}^2+bn\sigma_\tau^2,
\end{aligned} E ( M S A ) = a − 1 ( a − 1 ) σ 2 + n ( a − 1 ) σ τ β 2 + bn ( a − 1 ) σ τ 2 = σ 2 + n σ τ β 2 + bn σ τ 2 , 与正文式(13.3)的第一个结果相同。
S13.2 混合模型的期望均方 ¶ 正文第 13.3 节介绍了若干混合模型,其期望均方依赖于模型假定。这里考虑受限模型,下一节讨论非受限模型。受限模型对固定因子 A 的约束为
τ ⋅ = 0 , ( τ β ) ⋅ j = 0 , V [ ( τ β ) i j ] = a − 1 a σ τ β 2 . \tau_\cdot=0,\qquad(\tau\beta)_{\cdot j}=0,\qquad
V[(\tau\beta)_{ij}]=\frac{a-1}{a}\sigma_{\tau\beta}^2. τ ⋅ = 0 , ( τ β ) ⋅ j = 0 , V [( τ β ) ij ] = a a − 1 σ τ β 2 . 求随机因子 B 的期望均方:
E ( M S B ) = E ( S S B ) b − 1 , E ( S S B ) = 1 a n E ( ∑ j = 1 b y ⋅ j ⋅ 2 ) − 1 a b n E ( y ⋯ 2 ) . E(MS_B)=\frac{E(SS_B)}{b-1},\qquad
E(SS_B)=\frac1{an}E\left(\sum_{j=1}^by_{\cdot j\cdot}^2\right)-\frac1{abn}E(y_{\cdots}^2). E ( M S B ) = b − 1 E ( S S B ) , E ( S S B ) = an 1 E ( j = 1 ∑ b y ⋅ j ⋅ 2 ) − abn 1 E ( y ⋯ 2 ) . 利用约束,
y ⋅ j ⋅ = a n μ + a n β j + ϵ ⋅ j ⋅ , y_{\cdot j\cdot}=an\mu+an\beta_j+\epsilon_{\cdot j\cdot}, y ⋅ j ⋅ = an μ + an β j + ϵ ⋅ j ⋅ , 因而
1 a n E ( ∑ j y ⋅ j ⋅ 2 ) = b ( a n μ ) 2 + b ( a n ) 2 σ β 2 + a b n σ 2 a n = a b n μ 2 + a b n σ β 2 + b σ 2 . \frac1{an}E\left(\sum_jy_{\cdot j\cdot}^2\right)
=\frac{b(an\mu)^2+b(an)^2\sigma_\beta^2+abn\sigma^2}{an}
=abn\mu^2+abn\sigma_\beta^2+b\sigma^2. an 1 E ( j ∑ y ⋅ j ⋅ 2 ) = an b ( an μ ) 2 + b ( an ) 2 σ β 2 + abn σ 2 = abn μ 2 + abn σ β 2 + b σ 2 . 又因 y ⋯ = a b n μ + a n β ⋅ + ϵ ⋯ y_{\cdots}=abn\mu+an\beta_\cdot+\epsilon_{\cdots} y ⋯ = abn μ + an β ⋅ + ϵ ⋯ ,有
1 a b n E ( y ⋯ 2 ) = ( a b n μ ) 2 + b ( a n ) 2 σ β 2 + a b n σ 2 a b n = a b n μ 2 + a n σ β 2 + σ 2 . \frac1{abn}E(y_{\cdots}^2)
=\frac{(abn\mu)^2+b(an)^2\sigma_\beta^2+abn\sigma^2}{abn}
=abn\mu^2+an\sigma_\beta^2+\sigma^2. abn 1 E ( y ⋯ 2 ) = abn ( abn μ ) 2 + b ( an ) 2 σ β 2 + abn σ 2 = abn μ 2 + an σ β 2 + σ 2 . 所以
E ( M S B ) = ( b − 1 ) σ 2 + a n ( b − 1 ) σ β 2 b − 1 = σ 2 + a n σ β 2 . E(MS_B)=\frac{(b-1)\sigma^2+an(b-1)\sigma_\beta^2}{b-1}
=\sigma^2+an\sigma_\beta^2. E ( M S B ) = b − 1 ( b − 1 ) σ 2 + an ( b − 1 ) σ β 2 = σ 2 + an σ β 2 . 其他期望均方可类似推导。
S13.3 受限与非受限混合模型 ¶ 考虑非受限模型
y i j k = μ + α i + γ j + ( α γ ) i j + ϵ i j k , i = 1 , … , a , j = 1 , … , b , k = 1 , … , n , y_{ijk}=\mu+\alpha_i+\gamma_j+(\alpha\gamma)_{ij}+\epsilon_{ijk},\qquad
i=1,\ldots,a,\quad j=1,\ldots,b,\quad k=1,\ldots,n, y ijk = μ + α i + γ j + ( α γ ) ij + ϵ ijk , i = 1 , … , a , j = 1 , … , b , k = 1 , … , n , 假定 α ⋅ = 0 \alpha_\cdot=0 α ⋅ = 0 、V [ ( α γ ) i j ] = σ α γ 2 V[(\alpha\gamma)_{ij}]=\sigma_{\alpha\gamma}^2 V [( α γ ) ij ] = σ α γ 2 ,且所有随机效应互不相关。与受限模型不同,不要求交互作用对固定因子水平求和为零。受限模型允许更一般的相关结构,但现代软件有的提供两种选择,有的只采用非受限模型,因此两种形式都值得关注。
下面推导随机因子 B 的期望均方。它与受限模型不同,而差异的关键正是交互作用假定。仍有
E ( M S B ) = 1 b − 1 [ 1 a n E ( ∑ j y ⋅ j ⋅ 2 ) − 1 a b n E ( y ⋯ 2 ) ] . E(MS_B)=\frac1{b-1}\left[\frac1{an}E\left(\sum_jy_{\cdot j\cdot}^2\right)-\frac1{abn}E(y_{\cdots}^2)\right]. E ( M S B ) = b − 1 1 [ an 1 E ( j ∑ y ⋅ j ⋅ 2 ) − abn 1 E ( y ⋯ 2 ) ] . 先考虑
y ⋅ j ⋅ = a n μ + n α ⋅ + a n γ j + n ( α γ ) ⋅ j + ϵ ⋅ j ⋅ = a n μ + a n γ j + n ( α γ ) ⋅ j + ϵ ⋅ j ⋅ . \begin{aligned}
y_{\cdot j\cdot}&=an\mu+n\alpha_\cdot+an\gamma_j+n(\alpha\gamma)_{\cdot j}+\epsilon_{\cdot j\cdot}\\
&=an\mu+an\gamma_j+n(\alpha\gamma)_{\cdot j}+\epsilon_{\cdot j\cdot}.
\end{aligned} y ⋅ j ⋅ = an μ + n α ⋅ + an γ j + n ( α γ ) ⋅ j + ϵ ⋅ j ⋅ = an μ + an γ j + n ( α γ ) ⋅ j + ϵ ⋅ j ⋅ . α ⋅ = 0 \alpha_\cdot=0 α ⋅ = 0 ,但交互作用和不为零。于是
1 a n E ( ∑ j y ⋅ j ⋅ 2 ) = a b n μ 2 + a b n σ γ 2 + b n σ α γ 2 + b σ 2 . \frac1{an}E\left(\sum_jy_{\cdot j\cdot}^2\right)
=abn\mu^2+abn\sigma_\gamma^2+bn\sigma_{\alpha\gamma}^2+b\sigma^2. an 1 E ( j ∑ y ⋅ j ⋅ 2 ) = abn μ 2 + abn σ γ 2 + bn σ α γ 2 + b σ 2 . 由于
y ⋯ = a b n μ + a n γ ⋅ + n ( α γ ) ⋅ ⋅ + ϵ ⋯ , y_{\cdots}=abn\mu+an\gamma_\cdot+n(\alpha\gamma)_{\cdot\cdot}+\epsilon_{\cdots}, y ⋯ = abn μ + an γ ⋅ + n ( α γ ) ⋅⋅ + ϵ ⋯ , 有
1 a b n E ( y ⋯ 2 ) = a b n μ 2 + a n σ γ 2 + n σ α γ 2 + σ 2 . \frac1{abn}E(y_{\cdots}^2)=abn\mu^2+an\sigma_\gamma^2+n\sigma_{\alpha\gamma}^2+\sigma^2. abn 1 E ( y ⋯ 2 ) = abn μ 2 + an σ γ 2 + n σ α γ 2 + σ 2 . 相减并除以 b − 1 b-1 b − 1 ,得到
E ( M S B ) = ( b − 1 ) σ 2 + n ( b − 1 ) σ α γ 2 + a n ( b − 1 ) σ γ 2 b − 1 = σ 2 + n σ α γ 2 + a n σ γ 2 , E(MS_B)=\frac{(b-1)\sigma^2+n(b-1)\sigma_{\alpha\gamma}^2+an(b-1)\sigma_\gamma^2}{b-1}
=\sigma^2+n\sigma_{\alpha\gamma}^2+an\sigma_\gamma^2, E ( M S B ) = b − 1 ( b − 1 ) σ 2 + n ( b − 1 ) σ α γ 2 + an ( b − 1 ) σ γ 2 = σ 2 + n σ α γ 2 + an σ γ 2 , 与正文式(13.12)一致。直接使用期望算子推导较繁琐,正文的规则可节省很多工作。还有适用于不平衡设计的其他规则和算法,参见 Milliken 和 Johnson(1984)。
S13.4 样本量不等的随机模型和混合模型 ¶ 若各单元观测数不同,方差分析通常更复杂。第 15 章简要讨论两因子固定效应不平衡设计,这里对随机或混合模型提供一些建议。
不平衡随机或混合模型通常没有平衡设计中那样的精确 F 检验,正文基于独立均方的 Satterthwaite 合成检验公式也不能直接套用。较方便的方法是极大似然或 REML,正文第 13.2 节介绍的方法及 SAS、JMP 均可用于不平衡设计。其局限是方差分量推断常依赖大样本近似,而设计试验的次数通常不多。一般讨论见 Searle(1987)。
S13.5 修正大样本方法的背景 ¶ 正文第 13.6.2 节讨论了可写为均方线性组合的方差分量的置信区间。大样本理论表明,当 min ( f 1 , … , f Q ) → ∞ \min(f_1,\ldots,f_Q)\to\infty min ( f 1 , … , f Q ) → ∞ 时,
Z = σ ^ 0 2 − σ 0 2 V ( σ ^ 0 2 ) → d N ( 0 , 1 ) , V ( σ ^ 0 2 ) = 2 ∑ i = 1 Q c i 2 θ i 2 f i . Z=\frac{\hat\sigma_0^2-\sigma_0^2}{\sqrt{V(\hat\sigma_0^2)}}\ \xrightarrow{d}\ N(0,1),\qquad
V(\hat\sigma_0^2)=2\sum_{i=1}^Q\frac{c_i^2\theta_i^2}{f_i}. Z = V ( σ ^ 0 2 ) σ ^ 0 2 − σ 0 2 d N ( 0 , 1 ) , V ( σ ^ 0 2 ) = 2 i = 1 ∑ Q f i c i 2 θ i 2 . θ i = E ( M S i ) \theta_i=E(MS_i) θ i = E ( M S i ) 是第 i 个均方所估计的方差分量线性组合,f i f_i f i 是其自由度。因此,双侧 100 ( 1 − α ) % 100(1-\alpha)\% 100 ( 1 − α ) % 大样本置信区间为
σ ^ 0 2 − z α / 2 V ( σ ^ 0 2 ) ≤ σ 0 2 ≤ σ ^ 0 2 + z α / 2 V ( σ ^ 0 2 ) . \hat\sigma_0^2-z_{\alpha/2}\sqrt{V(\hat\sigma_0^2)}
\le\sigma_0^2\le
\hat\sigma_0^2+z_{\alpha/2}\sqrt{V(\hat\sigma_0^2)}. σ ^ 0 2 − z α /2 V ( σ ^ 0 2 ) ≤ σ 0 2 ≤ σ ^ 0 2 + z α /2 V ( σ ^ 0 2 ) . 实际计算用 M S i MS_i M S i 替换 θ i \theta_i θ i 。这也是正文中 JMP 构造近似区间的依据。自由度大时效果较好,自由度小时可能不可靠。Welch(1956)提出了改善方法;Graybill 和 Wang(1980)的修正使某些特殊情形的区间成为精确区间,在其他情形中也常有较好的近似表现,其结果见正文式(13.25)。
S13.6 用修正大样本方法构造方差分量比值的置信区间 ¶ 实际常关注方差分量的比值。例如,正文例 13.1 的测量系统能力研究中,量具总方差为 σ β 2 + σ τ β 2 + σ 2 \sigma_\beta^2+\sigma_{\tau\beta}^2+\sigma^2 σ β 2 + σ τ β 2 + σ 2 ,产品方差为 σ τ 2 \sigma_\tau^2 σ τ 2 。为将量具变异表示为产品变异的百分比,可考察
σ β 2 + σ τ β 2 + σ 2 σ τ 2 . \frac{\sigma_\beta^2+\sigma_{\tau\beta}^2+\sigma^2}{\sigma_\tau^2}. σ τ 2 σ β 2 + σ τ β 2 + σ 2 . 设所关注比值为 σ 1 2 / σ 2 2 \sigma_1^2/\sigma_2^2 σ 1 2 / σ 2 2 ,两方差可分别用均方线性组合估计:
σ ^ 1 2 σ ^ 2 2 = ∑ i = 1 P c i M S i ∑ j = P + 1 Q c j M S j . \frac{\hat\sigma_1^2}{\hat\sigma_2^2}
=\frac{\sum_{i=1}^Pc_iMS_i}{\sum_{j=P+1}^Qc_jMS_j}. σ ^ 2 2 σ ^ 1 2 = ∑ j = P + 1 Q c j M S j ∑ i = 1 P c i M S i . 其 100 ( 1 − α ) % 100(1-\alpha)\% 100 ( 1 − α ) % 单侧置信下限为
L = σ ^ 1 2 σ ^ 2 2 [ 2 + k 4 / ( k 1 k 2 ) − V L 2 ( 1 − k 5 / k 2 2 ) ] , L=\frac{\hat\sigma_1^2}{\hat\sigma_2^2}
\left[\frac{2+k_4/(k_1k_2)-\sqrt{V_L}}{2(1-k_5/k_2^2)}\right], L = σ ^ 2 2 σ ^ 1 2 [ 2 ( 1 − k 5 / k 2 2 ) 2 + k 4 / ( k 1 k 2 ) − V L ] , 其中
V L = ( 2 + k 4 k 1 k 2 ) 2 − 4 ( 1 − k 5 k 2 2 ) ( 1 − k 3 k 1 2 ) , k 1 = ∑ i = 1 P c i M S i , k 2 = ∑ j = P + 1 Q c j M S j , k 3 = ∑ i = 1 P G i 2 c i 2 M S i 2 + ∑ i = 1 P − 1 ∑ t > i P G i t ∗ c i c t M S i M S t , k 4 = ∑ i = 1 P ∑ j = P + 1 Q G i j c i c j M S i M S j , k 5 = ∑ j = P + 1 Q H j 2 c j 2 M S j 2 . \begin{aligned}
V_L&=\left(2+\frac{k_4}{k_1k_2}\right)^2-4\left(1-\frac{k_5}{k_2^2}\right)\left(1-\frac{k_3}{k_1^2}\right),\\
k_1&=\sum_{i=1}^Pc_iMS_i,\qquad k_2=\sum_{j=P+1}^Qc_jMS_j,\\
k_3&=\sum_{i=1}^PG_i^2c_i^2MS_i^2+\sum_{i=1}^{P-1}\sum_{t>i}^PG_{it}^*c_ic_tMS_iMS_t,\\
k_4&=\sum_{i=1}^P\sum_{j=P+1}^QG_{ij}c_ic_jMS_iMS_j,\\
k_5&=\sum_{j=P+1}^QH_j^2c_j^2MS_j^2.
\end{aligned} V L k 1 k 3 k 4 k 5 = ( 2 + k 1 k 2 k 4 ) 2 − 4 ( 1 − k 2 2 k 5 ) ( 1 − k 1 2 k 3 ) , = i = 1 ∑ P c i M S i , k 2 = j = P + 1 ∑ Q c j M S j , = i = 1 ∑ P G i 2 c i 2 M S i 2 + i = 1 ∑ P − 1 t > i ∑ P G i t ∗ c i c t M S i M S t , = i = 1 ∑ P j = P + 1 ∑ Q G ij c i c j M S i M S j , = j = P + 1 ∑ Q H j 2 c j 2 M S j 2 . G i , H j , G i j , G i t ∗ G_i,H_j,G_{ij},G_{it}^* G i , H j , G ij , G i t ∗ 的定义沿用正文修正大样本方法中的量。更多细节见 Burdick 和 Graybill(1992)。原补充材料中若干公式及交叉引用沿用旧版编号,这里按当前正文调整;比值公式中的分母也按上下文恢复为 σ ^ 2 2 \hat\sigma_2^2 σ ^ 2 2 。[1]
补充参考文献 ¶ Graybill, F. A. and C. M. Wang(1980),“Confidence Intervals on Nonnegative Linear Combinations of Variances”,Journal of the American Statistical Association ,75,869–873。
Welch, B. L.(1956),“On Linear Combinations of Several Variances”,Journal of the American Statistical Association ,51,132–148。