线性回归模型
一、线性模型理论
1.1 定义
y
=
β
0
+
∑
i
=
1
k
f
i
(
x
1
,
⋅
⋅
⋅
,
x
m
)
β
i
+
ε
,
ε
⇔
N
(
0
,
σ
2
)
y = \beta_0 + \sum_{i=1}^{k}f_i(x_1,···,x_m)\beta_i + \varepsilon, \varepsilon \Leftrightarrow N(0, \sigma^2)
y=β0+i=1∑kfi(x1,⋅⋅⋅,xm)βi+ε,ε⇔N(0,σ2)
- “线性”是针对未知参数
β
\beta
-
E
y
=
β
0
+
x
1
β
1
+
.
.
.
+
x
k
β
k
Ey = \beta_0 + x_1\beta_1 + … + x_k\beta_k
[
x
]
[x]
y
y
-
y
=
β
0
+
x
1
β
1
+
.
.
.
+
x
k
β
k
+
ε
,
E
ε
=
0
y = \beta_0 + x_1\beta_1 + … + x_k\beta_k + \varepsilon, E\varepsilon = 0
1.2 参数的估计
Y
=
X
β
+
ε
Y = X\beta + \varepsilon
Y=Xβ+ε
1、未知参数
β
\beta
β 的估计:最小二乘估计(LSE)
-
∣
∣
Y
−
X
β
^
∣
∣
2
=
i
n
f
∣
∣
Y
−
X
β
∣
∣
2
,
β
∈
R
k
+
1
||Y-X\hat{\beta}||^2 = inf||Y-X\beta||^2,\beta \in R^{k+1}
- 求解思路:平方和分解
∣
∣
Y
−
X
β
∣
∣
2
=
∣
∣
Y
−
X
β
^
∣
∣
2
+
∣
∣
X
(
β
^
−
β
)
∣
∣
2
+
2
(
β
^
−
β
)
T
X
T
(
Y
−
X
β
^
)
||Y-X\beta||^2 = ||Y-X\hat{\beta}||^2 + ||X(\hat{\beta} – \beta)||^2 + 2(\hat{\beta} – \beta)^T X^T (Y-X\hat{\beta})
2
(
β
^
−
β
)
T
X
T
(
Y
−
X
β
^
)
=
0
2(\hat{\beta} – \beta)^T X^T (Y-X\hat{\beta}) = 0
- 正规方程:
(
X
T
X
)
β
^
=
X
T
Y
(X^TX)\hat{\beta} = X^TY
β
^
=
(
X
T
X
)
−
1
X
T
Y
=
S
−
1
X
T
Y
\hat{\beta} = (X^TX)^{-1}X^TY = S^{-1}X^TY
- 经验回归函数:
X
β
^
X\hat{\beta}
- 经验回归方程:
Y
=
X
β
^
Y = X\hat{\beta}
2、误差方差
σ
2
\sigma^2
σ2 的估计
y
i
=
β
0
+
β
1
x
i
1
+
.
.
.
+
β
k
x
i
k
+
ε
i
,
1
≤
i
≤
n
y_i = \beta_0 + \beta_1x_{i1} + … + \beta_kx_{ik} + \varepsilon_i,1\le i\le n
yi=β0+β1xi1+...+βkxik+εi,1≤i≤n
- 残差
e
i
=
y
i
−
β
0
^
+
β
1
^
x
i
1
+
.
.
.
+
β
k
^
x
i
k
e_i = y_i – \hat{\beta_0} + \hat{\beta_1}x_{i1} + … + \hat{\beta_k}x_{ik}
- 残差平方和
Q
e
=
e
1
2
+
e
2
2
+
.
.
.
+
e
n
2
=
∣
∣
Y
−
X
β
^
∣
∣
2
=
Y
T
(
I
n
−
X
S
−
1
X
T
)
Y
Q_e = e_1^2 + e_2^2 + … + e_n^2 = ||Y-X\hat{\beta}||^2 = Y^T(I_n – XS^{-1}X^T)Y
3、线性模型的最小二乘估计
-
β
\beta
L
S
E
LSE
β
^
=
(
X
T
X
)
−
1
X
T
Y
=
S
−
1
X
T
Y
\hat{\beta} = (X^TX)^{-1}X^TY = S^{-1}X^TY
-
σ
2
\sigma^2
L
S
E
LSE
σ
^
2
=
1
n
−
k
−
1
Y
T
(
I
n
−
X
S
−
1
X
T
)
Y
\hat{\sigma}^2 = \frac{1}{n-k-1}Y^T(I_n – XS^{-1}X^T)Y
4、最小二乘估计的无偏性质
-
E
(
Y
T
A
Y
)
=
(
E
Y
)
T
A
(
E
Y
)
+
t
r
{
A
[
V
a
r
(
Y
)
]
}
E(Y^TAY) = (EY)^TA(EY) + tr\{A[Var(Y)]\}
-
E
Y
=
X
β
,
V
a
r
(
Y
)
=
σ
2
I
n
EY=X\beta, Var(Y) = \sigma^2I_n
-
β
^
=
(
X
T
X
)
−
1
X
T
Y
\hat{\beta} = (X^TX)^{-1}X^TY
- 残差平方和的数学期望是:
E
(
Q
e
)
=
(
n
−
k
−
1
)
σ
2
E(Q_e) = (n-k-1) \sigma^2
1.3 估计量的分布
-
β
^
=
S
−
1
X
T
Y
\hat{\beta} = S^{-1}X^TY
N
(
β
,
σ
2
S
−
1
)
N(\beta, \sigma^2S^{-1})
-
n
−
k
−
1
σ
2
σ
^
2
=
1
σ
2
Y
T
(
I
n
−
X
S
−
1
X
T
)
Y
\frac{n-k-1}{\sigma^2}\hat{\sigma}^2 = \frac{1}{\sigma^2}Y^T(I_n – XS^{-1}X^T)Y
χ
2
(
n
−
k
−
1
)
\chi^2(n-k-1)
-
β
^
\hat{\beta}
σ
^
2
\hat{\sigma}^2
二、一元回归与相关分析
1.1 定义
1、回归分析:研究一个(或多个)自变量的变化如何影响因变量。
2、相关分析:研究这两个数值变量的相关程度。
3、回归方程
y
=
β
0
+
x
1
β
1
+
.
.
.
+
x
k
β
k
y = \beta_0 + x_1\beta_1 + … + x_k\beta_k
y=β0+x1β1+...+xkβk
1.2 一元线性回归模型
y
i
=
β
0
+
β
1
x
i
+
ε
i
,
1
≤
i
≤
n
y_i = \beta_0 + \beta_1x_i + \varepsilon_i, \,\,\,\,\, 1 \le i \le n
yi=β0+β1xi+εi,1≤i≤n
-
β
0
^
=
y
‾
−
β
1
^
x
‾
\hat{\beta_0} = \overline{y} – \hat{\beta_1}\overline{x}
-
β
1
^
=
L
x
y
L
x
x
\hat{\beta_1} = \frac{L_{xy}}{L_{xx}}
-
σ
^
2
=
1
n
−
2
(
L
y
y
−
β
1
^
L
x
y
)
\hat{\sigma}^2 = \frac{1}{n-2}(L_{yy} – \hat{\beta_1}L_{xy})
1.2 简单的相关分析
T
S
S
=
R
e
g
S
S
+
R
S
S
TSS = RegSS + RSS
TSS=RegSS+RSS
- 总(变差)平方和
T
S
S
=
∑
i
=
1
n
(
y
i
−
y
‾
)
2
TSS = \sum_{i=1}^n(y_i – \overline{y})^2
- 回归平方和
R
e
g
S
S
=
∑
i
=
1
n
(
y
i
^
−
y
‾
)
2
RegSS = \sum_{i=1}^n(\hat{y_i} – \overline{y})^2
- 残差平方和
R
S
S
=
∑
i
=
1
n
(
y
i
−
y
i
^
)
2
RSS = \sum_{i=1}^n(y_i – \hat{y_i})^2
- 相关系数
r
r
r
2
=
R
e
g
S
S
T
S
S
=
L
x
y
2
L
x
x
L
y
y
r^2 = \frac{RegSS}{TSS} = \frac{L_{xy}^2}{L_{xx}L_{yy}}
1.3 回归方程的检验与区间估计
1、回归系数的假设检验
-
H
0
:
β
1
=
0
H_0: \beta_1 = 0
-
β
0
^
\hat{\beta_0}
N
(
β
0
,
σ
2
(
1
n
+
x
‾
2
L
x
x
)
)
N(\beta_0, \sigma^2(\frac{1}{n} + \frac{\overline{x}^2}{L_{xx}}))
-
β
1
^
\hat{\beta_1}
N
(
β
1
,
σ
2
L
x
x
)
N(\beta_1, \frac{\sigma^2}{L_{xx}})
-
β
0
^
\hat{\beta_0}
β
1
^
\hat{\beta_1}
C
o
v
(
β
0
^
,
β
1
^
)
=
−
σ
2
x
‾
L
x
x
Cov(\hat{\beta_0}, \hat{\beta_1}) = -\sigma^2 \frac{\overline{x}}{L_{xx}}
-
σ
2
\sigma^2
β
0
^
\hat{\beta_0}
β
1
^
\hat{\beta_1}
n
−
2
σ
2
σ
^
2
⇔
χ
2
(
n
−
2
)
\frac{n-2}{\sigma^2} \hat{\sigma}^2 \Leftrightarrow \chi^2(n-2)
- 要检验回归关系是否显著,可以利用
t
t
β
1
^
σ
^
∑
i
=
1
n
(
x
i
−
x
‾
)
2
⇔
t
(
n
−
2
)
\frac{\hat{\beta_1}}{\hat{\sigma}}\sqrt{\sum_{i=1}^n(x_i-\overline{x})^2} \Leftrightarrow t(n-2)
- 更多的是采用
β
1
^
σ
^
L
x
x
⇔
F
(
1
,
n
−
2
)
⇔
(
n
−
2
)
L
x
y
2
L
x
x
L
y
y
−
L
x
y
2
\frac{\hat{\beta_1}}{\hat{\sigma}}L_{xx} \Leftrightarrow F(1,n-2) \Leftrightarrow \frac{(n-2)L_{xy}^2}{L_{xx}L_{yy} – L_{xy}^2}
- 否定域
F
=
(
n
−
2
)
r
2
(
1
−
r
2
)
>
F
0.05
(
1
,
n
−
2
)
F = \frac{(n-2)r^2}{(1-r^2)} > F_{0.05}(1,n-2)
2、回归系数的区间估计
β
1
^
σ
^
∑
i
=
1
n
(
x
i
−
x
‾
)
2
⇔
t
(
n
−
2
)
\frac{\hat{\beta_1}}{\hat{\sigma}}\sqrt{\sum_{i=1}^n(x_i-\overline{x})^2} \Leftrightarrow t(n-2)
σ^β1^i=1∑n(xi−x)2⇔t(n−2)
β
1
^
−
σ
^
∑
i
=
1
n
(
x
i
−
x
‾
)
2
t
α
/
2
(
n
−
2
)
—
—
β
1
^
+
σ
^
∑
i
=
1
n
(
x
i
−
x
‾
)
2
t
α
/
2
(
n
−
2
)
\hat{\beta_1} – \frac{\hat{\sigma}}{\sqrt{\sum_{i=1}^n(x_i-\overline{x})^2}}t_{\alpha/2}(n-2) —— \hat{\beta_1} + \frac{\hat{\sigma}}{\sqrt{\sum_{i=1}^n(x_i-\overline{x})^2}}t_{\alpha/2}(n-2)
β1^−∑i=1n(xi−x)2σ^tα/2(n−2)——β1^+∑i=1n(xi−x)2σ^tα/2(n−2)
1.4 回归方程的预测与控制
1、回归方程的预测
y
0
−
y
0
∗
⇔
N
(
0
,
σ
2
[
1
+
1
n
+
(
x
0
−
x
‾
)
2
∑
i
=
1
n
(
x
i
−
x
‾
)
2
]
)
y_0 – y_0^* \Leftrightarrow N(0, \sigma^2[1 + \frac{1}{n} + \frac{(x_0 – \overline{x} )^2}{\sum_{i=1}^n (x_i – \overline{x})^2}])
y0−y0∗⇔N(0,σ2[1+n1+∑i=1n(xi−x)2(x0−x)2])
β
0
^
+
β
1
x
0
−
h
^
—
—
β
0
^
+
β
1
x
0
+
h
^
\hat{\beta_0} + \hat{\beta_1 x_0 – h}——\hat{\beta_0} + \hat{\beta_1 x_0 + h}
β0^+β1x0−h^——β0^+β1x0+h^
h
=
t
α
/
2
(
n
−
2
)
σ
^
1
+
1
n
+
(
x
0
−
x
‾
)
2
∑
i
=
1
n
(
x
i
−
x
‾
)
2
h = t_{\alpha/2}(n-2)\hat{\sigma}\sqrt{1 + \frac{1}{n} + \frac{(x_0 – \overline{x} )^2}{\sum_{i=1}^n (x_i – \overline{x})^2}}
h=tα/2(n−2)σ^1+n1+∑i=1n(xi−x)2(x0−x)2
2、回归方程的控制
- 上述方程与下两个方程同时成立:
A
≤
y
0
∗
−
h
y
0
∗
+
h
≤
B
A \le y_0^* – h \,\,\,\,\,\, y_0^* + h \le B
3、注意
- 实际问题中回归模型的建立要依赖于专业知识,并且注意散点图的使用
- 即使回归模型通过了检验也只能认为所研究的变量是统计相关的
- 回归分析一般需要与相关分析结合起来
- 异方差性、序列相关性、多重共线性问题
三、多元回归分析
1.1 未知参数的估计
- 同上
1.2 回归模型的检验
-
H
0
:
β
1
=
β
2
=
.
.
.
=
β
k
=
0
H_0: \beta_1 = \beta_2 = … = \beta_k = 0
T
S
S
=
∑
i
=
1
n
(
y
i
−
y
‾
)
2
,
R
e
g
S
S
=
∑
i
=
1
n
(
y
i
^
−
y
‾
)
2
,
R
S
S
=
∑
i
=
1
n
(
y
i
−
y
i
^
)
2
TSS = \sum_{i=1}^n(y_i – \overline{y})^2, RegSS = \sum_{i=1}^n(\hat{y_i} – \overline{y})^2,RSS = \sum_{i=1}^n(y_i – \hat{y_i})^2
R
S
S
σ
2
⇔
χ
2
(
n
−
k
−
1
)
\frac{RSS}{\sigma^2} \Leftrightarrow \chi^2(n-k-1)
R
e
g
S
S
σ
2
⇔
χ
2
(
k
)
\frac{RegSS}{\sigma^2} \Leftrightarrow \chi^2(k)
F
=
n
−
k
−
1
k
R
e
g
S
S
R
S
S
↔
F
(
k
,
n
−
k
−
1
)
F = \frac{n-k-1}{k} \frac{RegSS}{RSS} \leftrightarrow F(k, n-k-1)
1.3 回归因子的挑选
- 逐步回归的想法:
H
0
i
:
β
i
=
0
⇔
H
1
i
:
β
i
≠
0
H_{0i}: \beta_i = 0 \Leftrightarrow H_{1i}: \beta_i ≠ 0
-
t
t
n
−
k
−
1
n – k – 1
T
i
=
β
i
^
c
i
i
σ
^
T_i = \frac{\hat{\beta_i}}{\sqrt{c_{ii}}\hat{\sigma}}
-
F
F
F
i
=
β
i
^
2
c
i
i
σ
^
2
F_i = \frac{\hat{\beta_i}^2}{c_{ii}\hat{\sigma}^2}