STAT 3690 Lecture 16
Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: daixieit
STAT 3690 Lecture 16
2022
What is a linear model?
● Responses are linear functions with respect to unknown parameters.
Univariate/multiple linear regression (J&W Sec. 7.2–7.5)
● Interested in the relationship between random scalar y and random g-vector [x1 , . . . , xq ]T
● Model
_ Population version: y l x1 , . . . , xq ~ ([1, x1 , . . . , xq ]β, 72 ), where β = [β0 , . . . , βq ]T , i.e., * E(y l x1 , . . . , xq ) = [1, x1 , . . . , xq ]β = β0 + j(q)=1 xj βj
* var(y l x1 , . . . , xq ) = 72
_ Sample version ψ = 太β + ε
* ψ = [y1 , . . . , yn ]T and design matrix
┌ 1 x11 . . . xq1 ┐
太 = ' . . . '
' '
● Independent realizations [yi , xi1, . . . , xiq ]T ~ [y, x1 , . . . , xq ]T , i = 1, . . . , n
● rk(太) = g + 1 < p + g + 1 < n
* ε = [e1 , . . . , en ]T ~ (〇n , 72 Ⅰn )
● Least squares (LS) estimation (no need of normality) _ LS = (太T 太)_ 1 太T ψ
_ L(2)S = (n _ g _ 1)_ 1 (ψ _ 太LS )T (ψ _ 太LS ) = (n _ g _ 1)_ 1 ψT (Ⅰ _ 工)ψ * Hat matrix 工 = [hij ]nxn = 太(太T 太)_ 1 太T
● Symmetric
● Idempotent: 工2 = 工工 = 工
● rk(工) = rk(太)
● Each eigenvalue of 工 is either zero or one
* E(L(2)S ) = 72
● Maximum likelihood (ML) estimation (in need of (conditional) normality)
_ ML = (太T 太)_ 1 太T ψ = LS
* Given 太, ML ~ dvNq+1(β, 72 (太T 太)_ 1 )
_ M(2)L = n_ 1 ψ(Ⅰ _ 工)ψ = n_ 1 (n _ g _ 1)L(2)S
* Given 太, nM(2)L /72 = (n _ g _ 1)L(2)S /72 ~ x2 (n _ g _ 1)
● Inference (in need of (conditional) normality) _ Inference on aT β , given a e 砝q+1
* Estimator aT ML
* 100(1 _ a)% confidence interval for aT β :
aT ML ± t1 _α/2,n _q _ 1 LS [aT (太T 太)_ 1 a]1/2
_ Inference on y0 = 太0(T)β + e0 with a new observation vector given 太0 = [1, x01 , . . . , x0q]T e 砝q+1
* Prediction 0 = 太0(T)ML
* 100(1 _ a)% prediction interval for y0
太0(T)ML ± t1 _α/2,n _q _ 1 LS [1 + 太0(T)(太T 太)_ 1 太0]1/2
2022-03-18