Download Statistical Estimation: Method of Moments and Maximum Likelihood Estimation - Prof. Yi Cha and more Assignments Statistics in PDF only on Docsity! STAT610 - HWK Solution 5 2.1.1 Let N1, N2, N3 be the number of individuals in the three different types, respectively. Note that N1 +N2 +N3 = n. The corresponding probabilities are θ2, 2θ(1− θ), (1− θ)2. (a) θ̂2 = N1n , 2θ̂(1− θ̂) = N2 n ⇒ θ̂ = θ̂2 + 2θ̂(1− θ̂) 2 = N1 n + N2 2n (b) θ̂ 1− θ̂ = N1 n + N2 2n 1− (N1n + N2 2n ) = 2N1 +N2 2n− 2N1 −N2 (c) E(X) = p3 − p1 = (1− θ)2 − θ2 = 1− 2θ. Apply the mothod of moment, (1− 2θ̂) = X̄ = N3 −N1 n ⇒ θ̂ = N1 n + N2 2n = T3 2.1.3 If X ∼ β(α1, α2), E(X) = α1 α1 + α2 , E(X2) = α1(α1 + 1) (α1 + α2)(α1 + α2 + 1) . Let µ̂1 = ( ∑ Xi)/n, µ̂2 = ( ∑ X2i )/n. The method of moments estimates of (α1, α2) is given by α1 α1 + α2 = µ̂1 α1(α1 + 1) (α1 + α2)(α1 + α2 + 1) = µ̂2 ⇒ α̂1 = µ̂1(µ̂1 − µ̂2) µ̂2 − µ̂21 α̂2 = (1− µ̂1)(µ̂1 − µ̂2) µ̂2 − µ̂21 2.1.5 X1, · · · , Xn ∼ i.i.d. Bernoulli(θ). V (θ0, θ) = Eθ0ψ(X, θ) = nθ0 θ − n(1− θ0) 1− θ V (θ0, θ) = 0 ⇒ θ = θ0, so θ0 is the unique solution and ψ is the estimating equation. Solving ψ(X, θ̂) = 0 to get θ̂ = S/n. 2.1.17 The best linear predictor is given by b1 = Cov(Y, Z) Var(Z) = E(Y Z)− E(Y )E(Z) E(Z2)− (EZ)2 , a1 = E(Y )− b1E(Z). The method of moment estimate for a1 and b1 can be obtained by pluging in the sample moments, b̂1 = ( ∑ YiZi)/n− Ȳ Z̄ ( ∑ Z2i )/n− (Z̄)2 , â1 = Ȳ − b̂1Z̄. 2.2.1 The constrast used in this problem is ρ(θ) = ∑ (Yi − θ 2 t2i ) 2. d dθ ρ(θ̂) = 0 ⇒ − ∑ Yit 2 i + θ̂ 2 ∑ t4i = 0 ⇒ θ̂ = 2 ∑ Yit 2 i∑ t4i 1 2.2.2 Assume Pr(Zi = zi, Yi = yi)= 1n , i = 1, · · · , n. Then, E(Z) = Z̄, E(Y ) = Ȳ , Var(Z) = E(Z − Z̄)2 = 1 n ∑ (Zi − Z̄)2 Cov(Z, Y ) = E[(Z − Z̄)(Y − Ȳ )] = 1 n ∑ (Zi − Z̄)(Yi − Ȳ ) From Thm 1.4.3., the best linear predictor is Y = a+ bZ, where β2 = Cov(Y, Z) Var(Z) = ∑ (Zi − Z̄)(Yi − Ȳ )∑ (Zi − Z̄)2 , β1 = E(Y )− β2E(Z) = Ȳ − β2Z̄. 2.2.16 (a) θ̂ is MLE for θ, ⇒ Lx(θ̂) ≥ Lx(θ∗), ∀θ ∈ Θ. η = h(θ) is 1-1, ⇒ Lx(h−1(η)) = p(x, η) Then for any η∗ = h(θ) ∈ h(Θ), p(x, h(θ̂)) = Lx(θ̂) ≥ Lx(θ∗)) = p(x, η∗) Therefore, h(θ̂) is MLE for η. (b) Let Θ(ω) = {θ ∈ Θ : q(θ) = ω}, then Lx(ω) = sup{Lx(θ) : θ ∈ Θ(ω)} and by definition ωMLE = arg sup ω∈Ω Lx(ω) = arg sup ω∈Ω sup{LX(θ) : θ ∈ Θ(ω)}. Let ω̂ = q(θ̂), so θ̂ ∈ Θ(ω̂). Since θ̂ is MLE of θ, Lx(ω̂) = sup{Lx(θ) : θ ∈ Θ(ω)} = Lx(θ̂) ≥ sup{Lx(θ) : θ ∈ Θ(ω∗)} = Lx(ω∗) ∀ω∗ ∈ Ω. Therefore, ω̂ = q(θ̂) is MLE for ω = q(θ). 2.2.22 The likelihood of hypergeometric distribution is given by Lx(b) = ( b x )( N−b n−x )( N n ) = c · b!(N − b)! (b− x)!(N − b− n+ x)! where c is some constant which doesn’t depend on b. Lx(b+ 1) Lx(b) = (b+1)!(N−b−1)! (b+1−x)!(N−b−1−n+x)! b!(N−b)! (b−x)!(N−b−n+x)! = (b+ 1)(N − b− n+ x) (b+ 1− x)(N − b) = 1 + x(N + 1)− n(b+ 1) (b+ 1− x)(N − b) 2