Constraints on Solutions to the Normal Equations c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 1 / 41 If rank(n×p X) = r < p, there are infinitely many solutions to the NE X0 Xb = X0 y. For (X0 X)− any GI of X0 X, the set of all solutions is (X0 X)− X0 y + (I − (X0 X)− X0 X)z : z ∈ Rp . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 2 / 41 It is possible to place additional constraints on a solution to the NE so that ∃ a unique solution that satisfies the constraints. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 3 / 41 Example: Suppose yij = µ + τi + εij (i = 1, . . . , t; j = 1, . . . , ni ) where E(εij ) = 0 ∀ i, j. µ̂ τ̂1 Consider the following constraints on a solution to the NE β̂ = . . .. τ̂2 (Note that constraints are on β̂ not β.) c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 4 / 41 Four common choices are 1. τ̂1 = 0 (set first to zero) 2. τ̂t = 0 (set last to zero) 3. Pt 4. Pt i=1 τ̂i =0 i=1 ni τ̂i =0 (sum to zero) (weighted sum to zero) c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 5 / 41 Any one of these constraints may be imposed by insisting that a solution to the NE β̂ satisfies Aβ̂= 0 for some matrix A whose rows µ̂ τ̂1 define linear constraints on β̂ = . . .. τ̂2 c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 6 / 41 For example, 1. A = [0, 1, 0, . . . , 0]. 2. A = [0, 0, 0, . . . , 1]. 3. A = [0, 1, 1, . . . , 1]. 4. A = [0, n1 , n2 , . . . , nt ]. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 7 / 41 Note that solution β̂ satisfies both the NE and the constraint equations Aβ̂ = 0 iff " X0 X c Copyright 2012 Dan Nettleton (Iowa State University) A # β̂ = " # X0 y 0 . Statistics 611 8 / 41 Furthermore, we know that X0 Xb = X0 y ⇐⇒ Xb = PX y. Thus, we have β̂ a solution to NE satisfying the constraints iff " # X A c Copyright 2012 Dan Nettleton (Iowa State University) " β̂ = PX y 0 # . Statistics 611 9 / 41 We now know that if "we#want a unique solution to NE and constraint X equations, we need to be of full-column rank; that is, we need A " #! X = p. rank A c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 10 / 41 rank(n×p X) = r < p. Thus, we know X has r LI rows. If we can find s ≡ p − r p × 1 vectors 3, when these vectors are combined with r LI rows of X, the set of p vectors is LI, then we can use the s vectors as rows of A to get rank c Copyright 2012 Dan Nettleton (Iowa State University) " #! X A = p. Statistics 611 11 / 41 Suppose x1 , . . . , xr denote r LI rows of X (written as column vectors) or, equivalently, r LI columns of X0 . We seek a1 , . . . , as (s = p − r) so that {x1 , . . . , xr , a1 , . . . , as } is a set of p = r + s LI vectors in Rp . Then, with A0 = [a1 , . . . , as ], " # X c Copyright 2012 Dan Nettleton (Iowa State University) A will have full-column rank. Statistics 611 12 / 41 Obviously, a1 , . . . , as must be LI and ak ∈ / C(X0 ) ∀ k = 1, . . . , s. Show by example that these conditions are not sufficient to guarantee rank c Copyright 2012 Dan Nettleton (Iowa State University) " #! X A = p. Statistics 611 13 / 41 Prove the following results: Suppose x1 , · · · , xr LI in Rp . Suppose a1 , . . . , as LI in Rp . Suppose r + s = p. Then x1 , . . . , xr , a1 , . . . , as LI ⇐⇒ span{x1 , . . . , xr } ∩ span{a1 , . . . , as } = {0}. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 16 / 41 Note that the condition span{x1 , . . . , xr } ∩ span{a1 , . . . , as } = {0} ⇒ (d1 a1 + · · · + ds as )0 β is not estimable whenever d1 , . . . , ds are not all 0. Thus, the constraints that we add, as well as all nontrivial LCs of those constraints, must correspond to nonestimable functions. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 19 / 41 Note that a1 , . . . , as with the desired properties always exist. Why is that? c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 20 / 41 The result that we proved can be alternatively stated as follows: Suppose rank(n×p X) = r < p. Suppose rank(s×p A) = s where s = p − r. Then C(X0 ) ∩ C(A0 ) = {0} ⇐⇒ rank c Copyright 2012 Dan Nettleton (Iowa State University) " #! X A = p. Statistics 611 24 / 41 Lemma 3.1: For s×p A 3 rank(A) = s = p − rank(n×p X) and C(X0 ) ∩ C(A0 ) = {0}, " X0 X # b= A " X0 X A0 A " # # X0 y is equivalent to 0 " b= # X0 y 0 , which is equivalent to (X0 X + A0 A)b = X0 y. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 25 / 41 Result 3.6: Suppose rank(n×p X) = r, rank(s×p A) = s = p − r, and C(X0 ) ∩ C(A0 ) = {0}. Then (i) X0 X + A0 A is nonsingular. (ii) (X0 X + A0 A)−1 X0 y is unique solution to X0 Xb = X0 y and Ab = 0. (iii) (X0 X + A0 A)−1 is GI of X0 X. (iv) A(X0 X + A0 A)−1 X0 = 0. (v) A(X0 X + A0 A)−1 A0 = I. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 30 / 41 Returning to our example, yij = µ + τi + εij (i = 1, . . . , t; j = 1, . . . , ni ) where E(εij ) = 0 ∀ i, j. µ τ1 β = . . Let’s consider the constraint .. τt t X ni τ̂i = 0. i=1 c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 35 / 41 Find β̂ that satisfies NE and the constraint. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 36 / 41 1n 1n1 1 1n 0n 2 2 X= . .. .. . 1nt 0nt n n1 n n 1 1 0 XX= n2 0 .. .. . . nt c Copyright 2012 Dan Nettleton (Iowa State University) 0 0n1 · · · 0n1 · · · 0n2 .. .. . . 0nt · · · 1nt n2 · · · nt 0 ··· 0 n2 · · · 0 . .. . . .. . . . 1n2 .. . 0 · · · nt Statistics 611 37 / 41 nµ̂ + n1 τ̂1 + · · · + nt τ̂t n1 µ̂ + n1 τ̂1 0 X Xβ̂ = .. . nt µ̂ + nt τ̂t y·· y1· X0 y = . . .. yt· Equating X0 Xβ̂ and X0 y leads to c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 38 / 41 nµ̂ + n1 τ̂1 + · · · + nt τ̂t = y·· n1 µ̂ + n1 τ̂1 = y1· .. . nt µ̂ + nt τ̂t = yt· . This system of equations has an infinite number of solutions. However, if we insist that n1 τ̂1 + · · · + nt τ̂t = 0, the first equation becomes c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 39 / 41 nµ̂ = y·· ⇐⇒ µ̂ = ȳ·· . Substituting µ̂ = ȳ·· in the other equations yields ni ȳ·· + ni τ̂i = yi· ⇐⇒ ȳ·· + τ̂i = ȳi· ⇐⇒ τ̂i = ȳi· − ȳ·· c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 40 / 41 For the general case, we can compute β̂ = (X0 X + A0 A)−1 X0 y. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 41 / 41