Math 6010 Solutions to homework 4 1, p. 113. First let us start with some probability theory: Suppose a/b > 0, and define the function g(x) := 1 − 1 1 + x(a/b) (x ≥ 0). The question we ask is, what is the distribution of g(X) when X ∼ Fa,b ? As in 5010, we begin by first computing the cdf of g(X). Note that 0 ≤ g(X) ≤ 1. Therefore, we need only to compute the following for 0 ≤ y ≤ 1: 1 P {g(X) ≤ y} = P 1 − ≤y 1 + X(a/b) 1 =P ≥1−y 1 + X(a/b) a 1 =P 1+X ≤ b 1−y a y =P X ≤ b 1−y y b =P X≤ · . 1−y a We differentiate [d/dy] to find the pdf of g(X): y b d y b fg(X) (y) = fX × 1−y a dy 1 − y a y b 1 b = fX × . 1−y a (1 − y)2 a Recall from your 5080 text that a −(a+b)/2 fX (x) ∝ x(a/2)−1 1 + x . b 1 In particular, fX y 1−y (a/2)−1 −(a+b)/2 b y y ∝ 1+ a 1−y 1−y −(a+b)/2 y (a/2)−1 1 = · 1−y (1 − y)(a/2)−1 = y (a/2)−1 (1 − y)(b/2)+1 . This tells us that fg(X) (y) ∝ y (a/2)−1 (1 − y)(b/2)−1 and hence g(X) ∼ Beta a b , 2 2 (0 ≤ y ≤ 1), . (∗) Now we can solve our problem. Recall that under H0 : β1 = · · · − = βp−1 = 0, F := R2 n − p ∼ Fp−1,n−p . 1 − R2 p − 1 We write R2 in terms of F : p−1 p−1 p−1 p−1 2 2 2 2 (1 − R )F = R =⇒ F =R +R F =R 1+F n−p n−p n−p n−p p−1 F 1 n−p 2 =1− = g(F ), =⇒ R = p−1 p−1 1+F 1+F n−p n−p 2 with a := p − 1 and b := n − p. Thanks to (∗), we have p−1 n−p H R2 ∼0 Beta , . 2 2 2, p. 113. This is an easy but important exercise, as it tells us that our estimators R2 and F are unit free. To be concrete, suppose the Yi ’s are measured in pounds. Because Y − Ŷ = (I − H)Y , the coordinates of Y − Ŷ are also measured in pounds. In particular, RSS = kY − Ŷ k2 is measured in (pounds)2 . Similarly, RSSH0 is measured in (pounds)2 . This shows that Pn F has no units. Because i=1 (Yi − Ȳ )2 is also measured in (pounds)2 , it follows also that R2 has no units. 2 2, p. 117. Define Y1,1 .. . Y1,n , Y := Y2,1 . .. Y2,n x1 .. . xn X := 0 . .. β1 β := , β2 0 0 .. . 0 , x1 .. . xn ε1,1 .. . ε1,n . ε := ε2,1 . .. ε2,n Then we have Y = Xβ + ε. In other words, our problem is a generalized linear model with p = 2. Now, Pn 0 XX= i=1 0 x2i Pn0 2 i=1 xi = n X x2i I =⇒ 1 (X 0 X)−1 = Pn i=1 i=1 x2i I. We are testing H0 : β1 − β2 = 0. That is, q = 1, A = (1 , −1) and c = 0. Since 2 A(X 0 X)−1 A0 = Pn 2, i=1 xi it follows that h i0 i −1 h RSSH0 = RSS + c − Aβ̂ A(X 0 X)−1 A0 c − Aβ̂ Pn x2 = RSS + (β̂1 − β̂2 ) i=1 i (β̂1 − β̂2 ) 2 Pn 2 x (β̂2 − β̂2 )2 = RSS + i=1 i . 2 This yields the formula for RSS − RSSH0 . Also, F = (RSSH0 − RSS)/q = S2 Pn i=1 x2i (β̂2 − β̂2 )2 (β̂2 − β̂2 )2 Pn . = 2 2 2S 2S ( I=1 x2i )−1 Under H0 , F ∼ F1,n−2 . Therefore, our F -test can also be written as a t-test: β̂1 − β̂2 H0 T = s ∼ tn−2 . 2 S Pn 2 i=1 xi 3