Equivalent Models and Reparameterization c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 1 / 19 Two linear models y = Wα + ε and y = Xβ + ε are equivalent, or reparameterizations of each other, iff c Copyright 2012 Dan Nettleton (Iowa State University) C(X) = C(W). Statistics 611 2 / 19 Result 2.8: If C(X) = C(W), then PX = PW . Proof of Result 2.8: ∀ y ∈ Rn , y = PX y + (I − PX )y = PW y + (I − PW )y. We have written y as a sum of a vector in C(X) = C(W) and a vector in C(X)⊥ = C(W)⊥ . Such a decomposition is unique by Result A.4. Therefore, PX y = PW y ∀ y ∈ Rn ⇒ P X = P W . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 3 / 19 Corollary 2.4: If C(X) = C(W), then ŷ = PX y = PW y and ε̂ = (I − PX )y = (I − PW )y = y − ŷ. The fitted values and residuals are the same for the models that are reparameterizations of one another. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 4 / 19 Recall C(X) = C(W) ⇐⇒ W = XT for some matrix T and X = WS for some matrix S. Result 2.9: If C(W) = C(X) and α̂ solves the NE W 0 Wa = W 0 y, then β̂ = T α̂ solves the NE X0 Xb = X0 y, where T is defined by W = XT. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 5 / 19 Example: Suppose y11 y12 .. . y1n y 1 = 1 . y= y2 y21 y22 . .. y2n2 Consider 1 X = n1 1n2 1n1 0n2 0n1 1n2 c Copyright 2012 Dan Nettleton (Iowa State University) and 1 W = n1 1n2 0n1 . 1n2 Statistics 611 7 / 19 Find T 3 W = XT. Find β̂ a solution to X0 Xb = X0 y. Find α̂ a solution to W 0 Wa = W 0 y. Show that Xβ̂ = W α̂. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 8 / 19