General Linear Test of a General Linear Hypothesis c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 1 / 35 Suppose the NTGMM holds so that y = Xβ + ε, where c Copyright 2012 Dan Nettleton (Iowa State University) ε ∼ N(0, σ 2 I). Statistics 611 2 / 35 Suppose C is a known q × p matrix and d is a known q × 1 vector. The general linear hypothesis H0 : Cβ = d is testable if rank(C) = q and each component of Cβ is estimable. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 3 / 35 Supposep×m A of rank s. Can H0 : β ∈ C(A) be written as a testable general linear hypothesis? c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 4 / 35 Suppose yijk = µ + αi + βj + γij + εijk , µ i = 1, 2; j = 1, 2; k = 1, . . . , nij . α1 α2 β1 Let β = β2 . γ11 γ12 γ 21 γ22 Write a testable general linear hypothesis for “no interaction.” c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 6 / 35 Suppose H0 : Cβ = d is testable. Find the distribution of the BLUE of Cβ. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 9 / 35 Find the distribution of (Cβ̂ − d)0 (σ 2 C(X0 X)− C0 )−1 (Cβ̂ − d). c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 11 / 35 Show that Cβ̂ and SSE are independent. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 13 / 35 Now note that independence of Cβ̂ and SSE =⇒ (Cβ̂ − d)0 (σ 2 C(X0 X)− C0 )−1 (Cβ̂ − d) and SSE y0 (I − PX )y σ̂ 2 = = σ 2 (n − r) σ 2 (n − r) σ2 are independent. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 18 / 35 We have previously shown that SSE ∼ χ2n−r . σ2 Thus, (n − r)σ̂ 2 ∼ χ2n−r σ2 and σ̂ 2 ∼ χ2n−r /(n − r). σ2 c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 19 / 35 It follows that (Cβ̂ − d)0 (σ 2 C(X0 X)− C0 )−1 (Cβ̂ − d)/q σ̂ 2 /σ 2 = (Cβ̂ − d)0 (C(X0 X)− C0 )−1 (Cβ̂ − d) ≡F qσ̂ 2 ∼ Fq,n−r (φ), where c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 20 / 35 φ= 1 (Cβ − d)0 (C(X0 X)− C0 )−1 (Cβ − d) 2σ 2 as defined previously. We can use F to test H0 : φ = 0 ⇐⇒ H0 : Cβ − d = 0 (∵ (C(X0 X)− C0 )−1 is PD.) ⇐⇒ H0 : Cβ = d. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 21 / 35 To test H0 : Cβ = d at level α, we reject H0 iff F ≥ Fq,n−r,α where Fq,n−r,α is the upper α quantile of the Fq,n−r distribution. By Result 5.13, the power of the test is a strictly increasing function of φ. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 22 / 35 Now suppose that H0 : c0 β = d is testable. By arguments analogous to the previous F case, it is straightforward to show that c0 β̂ − d t≡ p ∼ tn−r σ̂ 2 c0 (X0 X)− c c Copyright 2012 Dan Nettleton (Iowa State University) c0 β − d p σ 2 c0 (X0 X)− c ! . Statistics 611 23 / 35 We can conduct tests of H0 : c0 β = d against HA1 : c0 β < d HA2 : c0 β > d, or HA : c0 β 6= d by comparing the observed value of t to the tn−r distribution. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 24 / 35 Returning to the F-test of testable GLH H0 : Cβ = d, note that there are multiple ways to express the same null hypothesis. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 25 / 35 For example, suppose yij = µi + εij (i = 1, 2, 3; j = 1, . . . , ni ). Find different matrices C1 , C2 and C3 3 µ1 Ck µ2 = 0 ⇐⇒ µ1 = µ2 = µ3 µ3 c Copyright 2012 Dan Nettleton (Iowa State University) ∀ k = 1, 2, 3. Statistics 611 26 / 35 C1 = C2 = " C3 = " 1 −1 1 0 " 1 −1 1 0 1 −1 # 0 C1 µ = # 0 C2 µ = " # , C3 µ = µ1 − µ2 # µ1 − µ3 " , −1 1/2 1/2 −1 c Copyright 2012 Dan Nettleton (Iowa State University) , −1 0 " µ1 − µ2 # µ2 − µ3 µ 1 − µ2 µ1 +µ2 2 − µ3 # . Statistics 611 27 / 35 Suppose H01 : C1 β = d1 and H02 : C2 β = d2 are both testable and S1 ≡ {β : C1 β = d1 } = {β : C2 β = d2 } ≡ S2 . Show that the F-test of H01 is the same as the F-test of H02 . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 28 / 35 Recall that X0 = X0 PX = X0 X(X0 X)− X0 . Thus, X(X0 X)− is a GI of X0 ∀ X0 . It follows that C0k (Ck C0k )− is a GI of Ck . Thus, Sk = {C0k (Ck C0k )− dk + (I − C0k (Ck C0k )− Ck )z : z ∈ Rp } for k = 1, 2. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 29 / 35 Because S1 = S2 , C1 times any element of S2 equal d1 ; i.e., β ∈ S2 =⇒ C1 β = d1 . Thus C1 [C02 (C2 C02 )− d2 + (I − C02 (C2 C02 )− C2 )z] = d1 ∀ z ∈ Rp =⇒ C1 C02 (C2 C02 )− d2 − d1 + (C1 − C1 C02 (C2 C02 )− C2 )z = 0 =⇒ C1 = C1 C02 (C2 C02 )− C2 C1 C02 (C2 C02 )− d2 = d1 c Copyright 2012 Dan Nettleton (Iowa State University) ∀ z ∈ Rp and by Result A.8. Statistics 611 30 / 35 Now C1 = C1 C02 (C2 C02 )− C2 =⇒ C1 = C1 PC02 =⇒ C01 = PC02 C01 =⇒ C(C01 ) ⊆ C(PC02 ) = C(C02 ). Repeating the entire argument with the roles of C1 and C2 reversed gives C(C02 ) ⊆ C(C01 ) so that C(C02 ) = C(C01 ). c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 31 / 35 Because C1 β = d 1 and C2 β = d2 are both testable, C01 and C02 are both full-column rank = q. ∴ ∃ a unique nonsingular matrixq×q B3 C01 = C02 B. We have previously shown that C01 = PC02 C01 = C02 [(C2 C02 )− ]0 C2 C01 . Thus, B = [(C2 C02 )− ]0 C2 C01 c Copyright 2012 Dan Nettleton (Iowa State University) and B0 = C1 C02 (C2 C02 )− . Statistics 611 32 / 35 We have also shown previously that C1 C02 (C2 C02 )− d2 = d1 =⇒ B0 d2 = d1 . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 33 / 35 Now consider the quadratic form (C1 b − d1 )0 (C1 (X0 X)− C01 )−1 (C1 b − d1 ) = (B0 C2 b − B0 d2 )0 (B0 C2 (X0 X)− C02 B)−1 (B0 C2 b − B0 d2 ) = (C2 b − d2 )0 BB−1 (C2 (X0 X)− C02 )−1 (B0 )−1 B0 (C2 b − d2 ) = (C2 b − d2 )0 (C2 (X0 X)− C02 )−1 (C2 b − d2 ). This is true for ∀ b ∈ Rp including β̂ and the true parameter vector β. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 34 / 35 ∴ the F statistics for testing H01 : C1 β = d1 and H02 : C2 β = d2 are identical, as are the noncentrality parameters associated with those F statistics. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 35 / 35