Estimable Functions and Their Least Squares Estimators c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 1 / 51 Consider the GLM y =n×p X β + ε, n×1 p×1 n×1 where E(ε) = 0. Suppose we wish to estimate c0 β for some fixed and known c ∈ Rp . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 2 / 51 An estimator t(y) is an unbiased estimator of the function c0 β iff E[t(y)] = c0 β c Copyright 2012 Dan Nettleton (Iowa State University) ∀ β ∈ Rp . Statistics 611 3 / 51 An estimator t(y) is a linear estimator in y iff t(y) = d + a0 y for some known constants d, a1 , . . . , an . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 4 / 51 A function c0 β is linearly estimable iff ∃ a linear estimator that is an unbiased estimator of c0 β. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 5 / 51 Henceforth, we will use estimable as a synonym for linearly estimable. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 6 / 51 A function c0 β is said to be nonestimable if there does not exist a linear estimator that is an unbiased estimator of c0 β. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 7 / 51 Result 3.1: Under the GLM, c0 β is estimable iff the following equivalent conditions hold: (i) ∃ a 3 E(a0 y) = c0 β (ii) ∃ a 3 c0 = a0 X ∀ β ∈ Rp (X0 a = c) (iii) c ∈ C(X0 ). c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 8 / 51 Show conditions (i), (ii), and (iii) are equivalent. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 9 / 51 (i)⇐⇒(ii)⇐⇒(iii): ∃ a 3 E(a0 y) = c0 β ∀ β ∈ Rp ⇐⇒ ∃ a 3 a0 E(y) = c0 β ⇐⇒ ∃ a 3 a0 Xβ = c0 β ∀ β ∈ Rp ∀ β ∈ Rp ⇐⇒ ∃ a 3 a0 X = c0 ⇐⇒ ∃ a 3 X0 a = c ⇐⇒ c ∈ C(X0 ). c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 10 / 51 Show that any of the equivalent conditions is equivalent to c0 β estimable. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 11 / 51 c0 β estimable ⇐⇒ ∃ d, a 3 E(d + a0 y) = c0 β ⇐⇒ ∃ d, a 3 d + a0 Xβ = c0 β ⇐⇒ ∃ a 3 a0 Xβ = c0 β ∀ β ∈ Rp ∀ β ∈ Rp ∀ β ∈ Rp ⇐⇒ ∃ a 3 a0 X = c0 . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 12 / 51 Example: Suppose that when team i competes against team j, the expected margin of victory for team i over team j is µi − µj , where µ1 , . . . , µ5 are unknown parameters. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 13 / 51 Suppose we observe the following outcomes. Team 1 beats Team 2 by 7 3 1 3 3 2 14 3 5 17 4 5 10 4 1 1 c Copyright 2012 Dan Nettleton (Iowa State University) points Statistics 611 14 / 51 Determine y, X, β. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 15 / 51 1 −1 0 0 0 7 µ1 3 −1 0 1 0 0 µ2 14 0 −1 1 0 0 , µ3 , 0 1 0 −1 17 0 µ 4 10 0 0 0 1 −1 µ5 −1 0 0 1 0 1 y , c Copyright 2012 Dan Nettleton (Iowa State University) X , β Statistics 611 16 / 51 Is µ1 − µ2 is estimable? c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 17 / 51 Yes, µ1 − µ2 is estimable. µ1 − µ2 = c0 β, where c0 = [1, −1, 0, 0, 0] 1 −1 0 0 −1 0 1 0 0 −1 1 0 = [1, 0, 0, 0, 0, 0] 0 1 0 0 0 0 0 1 −1 0 0 1 0 0 0 −1 −1 0 = a0 X. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 18 / 51 Thus, a0 y = [1, 0, 0, 0, 0, 0]y = y1 is an unbiased estimator of c0 β = µ1 − µ2 . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 19 / 51 Is µ1 − µ3 is estimable? c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 20 / 51 Yes, µ1 − µ3 is estimable. µ1 − µ3 = c0 β, where c0 = [1, 0, −1, 0, 0] 1 −1 0 0 −1 0 1 0 0 −1 1 0 = [0, −1, 0, 0, 0, 0] 0 1 0 0 0 0 0 1 −1 0 0 1 0 0 0 −1 −1 0 = a0 X. ∴ −y2 is unbiased estimator of µ1 − µ3 . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 21 / 51 Is µ1 − µ5 is estimable? c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 22 / 51 Yes, µ1 − µ5 is estimable. µ1 − µ5 = c0 β, where c0 = [1, 0, 0, 0, −1] 1 −1 0 0 −1 0 1 0 0 −1 1 0 = [0, −1, 0, 1, 0, 0] 0 1 0 0 0 0 0 1 −1 0 0 1 0 0 0 −1 −1 0 = a0 X. ∴ y4 − y2 is unbiased estimator of µ1 − µ5 . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 23 / 51 Is µ1 estimable? c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 24 / 51 µ1 = c0 β, where c0 = [1, 0, 0, 0, 0]. So, does ∃ a 3 a0 X = [1, 0, 0, 0, 0]? c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 25 / 51 1 −1 0 0 −1 0 1 0 0 −1 1 0 a0 X = [a1 , a2 , a3 , a4 , a5 , a6 ] 0 1 0 0 0 0 0 1 −1 0 0 1 0 0 0 −1 −1 0 = [a1 − a2 − a6 , −a1 − a3 , a2 + a3 + a4 , a5 + a6 , −a4 − a5 ] ? = [1, 0, 0, 0, 0]. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 26 / 51 a1 − a2 − a6 = 1 (1) −a1 − a3 = 0 (2) a2 + a3 + a4 = 0 (3) a5 + a6 = 0 (4) −a4 − a5 = 0 (5) (1) and (2) imply −a2 − a3 − a6 = 1. (4) and (5) imply a4 = a6 , which together with (3) implies a2 + a3 + a6 = 0. ∴ There does not exist a 3 a0 X = [1, 0, 0, 0, 0] ⇒ µ1 is nonestimable. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 27 / 51 Result 3.1 tells us that c0 β is estimable iff ∃ a 3 c0 β = a0 Xβ ∀ β ∈ Rp . Recall that E(y) = Xβ. Thus, c0 β is estimable iff it is a LC of the elements of E(y). c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 28 / 51 This leads to Method 3.1: LCs of expected values of observations are estimable. c0 β is estimable iff c0 β is a LC of the elements of E(y); i.e., c0 β = n X ai E(yi ) for some a1 , . . . , an . i=1 c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 29 / 51 Use Method 3.1 to show that µ2 − µ4 is estimable in our previous example. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 30 / 51 µ1 − µ2 µ3 − µ1 µ3 − µ2 . E(y) = Xβ = µ3 − µ5 µ4 − µ5 µ4 − µ1 µ2 − µ4 = −(µ1 − µ2 ) − (µ4 − µ1 ) = −E(y1 ) − E(y6 ). c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 31 / 51 Method 3.2: c0 β is estimable iff c ∈ C(X0 ). Thus, find a basis for C(X0 ), say {v1 , . . . , vr }, and determine if c= r X di vi for some d1 , . . . , dr . i=1 c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 32 / 51 Method 3.3: By Result A.5, we know that C(X0 ) and N (X) are orthogonal complements in Rp . Thus, c ∈ C(X0 ) iff c0 d = 0 ∀ d ∈ N (X), which is equivalent to Xd = 0 ⇒ c0 d = 0. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 33 / 51 Reconsider our previous example. Use Method 3.3 to show that µ1 is nonestimable. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 34 / 51 Xd = 0 ⇒ d1 − d2 = 0 d3 − d1 = 0 d3 − d2 = 0 d3 − d5 = 0 d4 − d5 = 0 d4 − d1 = 0. ⇒ d1 = d2 = d3 = d4 = d5 . For example, X1 = 0. Note [1, 0, 0, 0, 0]1 = 1 6= 0. Thus, µ1 is not estimable. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 35 / 51 Now use method 3.3 to establish that c0 β = c1 µ1 + c2 µ2 + c3 µ3 + c4 µ4 + c5 µ5 is estimable iff 5 X ci = 0. i=1 c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 36 / 51 {d ∈ R5 : Xd = 0} = {d1 : d ∈ R}. Thus 5×1 c0 d = 0 ∀ d ∈ N (X) ⇐⇒ c0 (d1) = 0 ∀ d ∈ R ⇐⇒ dc0 1 = 0 ∀ d ∈ R ⇐⇒ d 5 X ci = 0 ∀ d ∈ R i=1 ⇐⇒ 5 X ci = 0. i=1 c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 37 / 51 The least squares estimator of an estimable function c0 β is c0 β̂, where β̂ is any solution to the NE (X0 Xb = X0 y). c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 38 / 51 Result 3.2: If c0 β is estimable, then c0 β̂ is the same for all solutions β̂ to the NE. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 39 / 51 Proof of Result 3.2: Suppose β̂ 1 , and β̂ 2 are any two solutions to the NE. From Corollary 2.3, we know Xβ̂ 1 = Xβ̂ 2 . Now c0 β is estimable ⇒ ∃ a 3 c0 = a0 X. Thus, c0 β̂ 1 = a0 Xβ̂ 1 = a0 Xβ̂ 2 = c0 β̂ 2 . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 40 / 51 Result 3.3: The least squares estimator of an estimable function c0 β is a linear unbiased estimator of c0 β. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 41 / 51 Proof of Result 3.3: From Result 3.2, we know c0 β̂ is the same ∀ solution to NE. We know (X0 X)− X0 y is a solution to NE. Thus, c0 β̂ = c0 (X0 X)− X0 y. This is a linear estimator. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 42 / 51 Furthermore, c0 β estimable ⇒ ∃ a 3 c0 = a0 X. Thus, E(c0 β̂) = E(c0 (X0 X)− X0 y) c Copyright 2012 Dan Nettleton (Iowa State University) = c0 (X0 X)− X0 E(y) = c0 (X0 X)− X0 Xβ = a0 X(X0 X)− X0 Xβ = a0 Xβ = c0 β. Statistics 611 43 / 51 Consider again our previous example. Recall that y1 is a linear unbiased estimator of µ1 − µ2 . Is this the least squares estimator? c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 44 / 51 It can be shown that the least squares estimator of µ1 − µ2 is 7 3 4 1 1 1 y1 − y2 + y3 − y4 + y5 − y6 . 11 11 11 11 11 11 Based on the observed y, the least squares estimate is 8.0. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 45 / 51 One of infinitely many solutions to the NE is 2.8 −5.2 β̂ = (X0 X)− X0 y = 7.8 . 2.8 −8.2 Which team is best? c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 46 / 51 Suppose y = Xβ + ε, where E(ε) = 0 and rank(n×p X) = p. Show that c0 β is estimable ∀ c ∈ Rp . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 47 / 51 Proof: rank(n×p X) = p ⇒ X0 has p LI columns and each column is an element of Rp . Thus, by Fact V4, p LI columns of X0 form a basis for Rp . ∴ C(X0 ) = Rp . This also follows from Result A.7: C(X0 ) ⊆ Rp , dim(C(X0 )) = dim(Rp ) = p ⇒ C(X0 ) = Rp . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 48 / 51 Now from Result 3.1, we know c0 β is estimable iff c ∈ C(X0 ). ∵ C(X0 ) = Rp , c0 β is estimable ∀ c ∈ Rp . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 49 / 51 Alternatively, we can prove the result by noting the following rank(n×p X) = p ⇒ rank(X0 X) = p by Corollary 2.2. Now X0 X is p × p of rank p and ∴ (X0 X)−1 exists. Thus, β̂ = (X0 X)−1 X0 y is the unique solution to NE c Copyright 2012 Dan Nettleton (Iowa State University) X0 Xb = X0 y. Statistics 611 50 / 51 Note that ∀ c ∈ Rp E(c0 β̂) = c0 (X0 X)−1 XE(y) = c0 (X0 X)−1 X0 Xβ = c0 β. Thus, c0 β̂ = c0 (X0 X)−1 X0 y is a linear unbiased estimator of c0 β ∀ c ∈ Rp . ∴ c0 β is estimable ∀ c ∈ Rp . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 51 / 51