Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: THEend8_
STAT3922: Applied Linear Model (Advanced) Semester 1, 2024
Specifically, consider the linear model y = Xβ+ε, whereX is an n×p full rank design
matrix. Consider the general linear hypothesis test H0 : Lβ = 0 versus H1 : Lβ ̸= 0,
for some q × p matrix L with rank(L) = q < p. Let βˆ = (X⊤X)−1X⊤y denote the least
square estimator under H1 (i.e the unconstrained model). In the Advanced Lecture week
10, we construct
SSH = βˆ⊤L⊤
{
L(X⊤X)−1L⊤
}−1
(Lβˆ)
and build the test based on F =
SSH/q
SSE(H1)/(n− p) , where SSE(H1) = y
⊤(In−H)y is the
SSE for the full model.
Also, we are proving that this test can also be conducted from the full-reduced model
approach. To do it, we fitted the model under H0 using the Lagrange multiplier approach,
showed that the least square estimator for the model under H0 is given by
β˜ = (X⊤X)−1
(
X⊤ −K)y,
where K = L⊤
{
L(X⊤X)−1L⊤
}−1
L(X⊤X)−1X⊤. The residual vector of that model is
e˜ = y − y˜ = (I−P)y, y˜ = Xβ˜ = Py,
where P = X(X⊤X)−1
(
X⊤ −K). Fill in the details for the remaining step:
1. Show that
(i) P is idempotent.
(ii) trace(P) = p− q.
2. Let SSE(H0) = e˜
⊤e˜ be the residuals sum of squares under H0. Show that
SSE(H0)− SSE(H1) = SSH.