Likelihood Ratio Test Likelihood Ratio Test I Recall: N responses x1 , x2 , . . . , xN are assumed to be independent, and the αth response is related to q covariate values, represented by the (q × 1) vector zα , through xα ∼ Np (βzα , Σ) , α = 1, 2, . . . , N. where β is a (p × q) matrix of regression coefficients. I Partition β into (p × q1 ) and (p × q2 ) blocks: β = (β 1 β 2 ) I We want to test the null hypothesis H0 : β 1 = β ∗1 for some given matrix β ∗1 . NC STATE UNIVERSITY 1 / 12 Statistics 784 Multivariate Analysis Likelihood Ratio Test I Write Σ̂Ω for the unrestricted mle of Σ: Σ̂Ω = N 0 1 X xα − β̂zα xα − β̂zα . N α=1 I The corresponding maximized value of the likelihood is − 1 N 1 1 2 e − 2 pN (2π)− 2 pN det Σ̂Ω NC STATE UNIVERSITY 2 / 12 Statistics 784 Multivariate Analysis Likelihood Ratio Test I To maximize the likelihood under H0 , write yα = xα − β ∗1 z(1) α where zα is similarly partitioned: (1) zα = I zα (2) zα ! (2) Then under H0 , yα ∼ Np β 2 zα , Σ , so β 2 is estimated by β̂ 2,ω = N X 0 yα z(2) α ! A−1 2,2 α=1 = (C2 − β ∗1 A1,2 ) A−1 2,2 with the corresponding partitioning of A and C. NC STATE UNIVERSITY 3 / 12 Statistics 784 Multivariate Analysis Likelihood Ratio Test I Then Σ is estimated by Σ̂ω , where N Σ̂ω = = = N X α=1 N X yα − β̂ 2,ω z(2) α yα − β̂ 2,ω z(2) α 0 0 yα yα0 − β̂ 2,ω A2,2 β̂ 2,ω α=1 N X xα − β ∗1 z(1) α xα − β ∗1 z(1) α 0 0 − β̂ 2,ω A2,2 β̂ 2,ω α=1 I The corresponding maximized value of the likelihood is − 1 N 1 1 2 (2π)− 2 pN det Σ̂ω e − 2 pN NC STATE UNIVERSITY 4 / 12 Statistics 784 Multivariate Analysis Likelihood Ratio Test I The generalized likelihood ratio criterion is therefore 1N 2 det Σ̂Ω λ= 1N 2 det Σ̂ω and the test procedure is to reject H0 if λ < λ0 for appropriate λ0 . I The two estimates of Σ are related by 0 N Σ̂ω = N Σ̂Ω + β̂ 1,Ω − β ∗1 A1,1·2 β̂ 1,Ω − β ∗1 where A1,1·2 = A1,1 − A1,2 A−1 2,2 A2,1 . NC STATE UNIVERSITY 5 / 12 Statistics 784 Multivariate Analysis Likelihood Ratio Test I The first term, N Σ̂Ω , is the matrix of sums of squares and products of residuals, or the Error SSCP. I The second term, 0 β̂ 1,Ω − β ∗1 A1,1·2 β̂ 1,Ω − β ∗1 , is the SSCP matrix associated with the Hypothesis. I If q1 = 1, the second term has rank 1, and λ is a monotone function of the T 2 -like statistic 0 −1 ∗ ∗ A1,1·2 β̂ 1,Ω − β 1 Σ̂Ω β̂ 1,Ω − β 1 . NC STATE UNIVERSITY 6 / 12 Statistics 784 Multivariate Analysis Likelihood Ratio Test Invariance I If D is (p × p) nonsingular, then I I the null hypothesis H0 : β 1 = 0, and the likelihood ratio criterion λ for testing it are both invariant under the transformation x∗α = Dxα . I To be precise, x∗α ∼ Np (β ∗ zα , Σ∗ ) , where β ∗ = Dβ = (Dβ 1 Dβ 2 ) and Σ∗ = DΣD0 NC STATE UNIVERSITY 7 / 12 Statistics 784 Multivariate Analysis Likelihood Ratio Test I So β ∗1 = Dβ 1 = 0 iff β 1 = 0, showing invariance of H0 . I Also (after some algebra) ∗ Σ̂Ω = DΣ̂Ω D0 and ∗ Σ̂ω = DΣ̂ω D0 whence λ∗ = λ. I But, when q1 > 1, λ is not the only invariant test statistic (more later . . . ). NC STATE UNIVERSITY 8 / 12 Statistics 784 Multivariate Analysis Likelihood Ratio Test Distribution of the Likelihood Ratio I Write U = λ2/N det Σ̂Ω = det Σ̂ω det N Σ̂Ω = 0 ∗ ∗ det N Σ̂Ω + β̂ 1,Ω − β 1 A1,1·2 β̂ 1,Ω − β 1 I We know that N Σ̂Ω ∼ Wp (Σ, N − q). I Using an appropriate orthogonal transformation, we can also show that under H0 0 β̂ 1,Ω − β ∗1 A1,1·2 β̂ 1,Ω − β ∗1 ∼ Wp (Σ, q1 ). NC STATE UNIVERSITY 9 / 12 Statistics 784 Multivariate Analysis Likelihood Ratio Test I Also these two Wishart-distributed matrices are independent. I So the criterion U has the distribution of Up,m,n = det(E) 1 = det(E + H) det(I + E−1 H) where: I I I I E ∼ Wp (Σ, n) , n = N − q; H ∼ Wp (Σ, m) , m = q1 ; E and H are independent. This distribution does not depend on Σ, so we can take Σ = Ip . NC STATE UNIVERSITY 10 / 12 Statistics 784 Multivariate Analysis Likelihood Ratio Test I Basic result: Up,m,n has the distribution of p Y Vi , i=1 where: I I I Vi has the beta density β 12 (n + 1 − i), 21 m ; V1 , V2 , . . . , Vp are independent. Also: Up,m,n has the same distribution as Um,p,n+m−p . NC STATE UNIVERSITY 11 / 12 Statistics 784 Multivariate Analysis Likelihood Ratio Test Special Cases I p = 1 (and hence also m = 1): 1 − U1,mn, n × = Fm,n . U1,m,n m I p = 2 (and hence also m = 2): p 1 − U2,m,n n − 1 p × = F2m,2(n−1) . m U2,m,n I Bartlett’s correction: for large N, approximately, " # N − q − 12 (p − q1 + 1) −2 × × log λ ∼ χ2pq1 . N NC STATE UNIVERSITY 12 / 12 Statistics 784 Multivariate Analysis
© Copyright 2026 Paperzz