Asymptotic Efficiency Recall: For two unbiased estimators T and T ∗ of γ(θ), we compare the variances of the estimators to judge their relative efficiency, i.e. RE(T, T ∗ , θ) = Varθ (T ∗ )/Varθ (T ). Similarly, we compare large-sample variances of asymptotically unbiased estimators. Definitions: Let {Tn∗ } and {Tn } be asymptotically unbiased for γ(θ). Then, define 1. The asymptotic relative efficiency of {Tn } with respect to {Tn∗ } at θ is defined as Varθ (Tn∗ ) ARE(Tn , Tn∗ , θ) = lim , θ∈Θ n→∞ Varθ (Tn ) 2. {Tn∗ } is called asymptotically efficient if ARE(Tn , Tn∗ , θ) ≤ 1, ∨θ ∈ Θ and any other {Tn } that is asymptotically unbiased for γ(θ). 3. The asymptotic efficiency of {Tn } is defined as AE(Tn , θ) ≡ ARE(Tn , Tn∗ , θ) if {Tn∗ } is asymptotically efficient. Previous example. Let X1 , X2 , . . . be iid uniform(0, θ), θ > 0. Recall • Tn = MME of θ based on X1 , . . . , Xn = 2X̄n • Tn∗ = MLE of θ based on X1 , . . . , Xn = max1≤i≤n Xi = X(n) • Eθ Tn = θ • Eθ Tn∗ = Varθ (Tn ) = n θ n+1 3θ2 n Varθ (Tn∗ ) = nθ2 (n + 1)2 (n + 2) 1 Asymptotic Properties of MLEs Main Result1 : Let X1 , X2 , . . . , Xn be iid with common pmf/pdf f (x|θ). Let θ̂n ≡ MLE of θ ∈ Θ ⊂ R based on X1 , X2 , . . . , Xn . Then, under Cramér-Rao-type regularity conditions: as n → ∞, 1. (Consistency) p θ̂n −→ θ 2. (Asymp. normality) 3. (Asymp. efficiency) √ µ d n(θ̂n − θ) −→ N ¶ µ ¶2 1 d log f (X1 |θ) 0, , I1 (θ) = Eθ I1 (θ) dθ The sequence of estimators {θ̂n } is asymptotically efficient. Discussion: Remarks: • The regularity conditions required for the validity of the Theorem above hold for distributions in the one-parameter exponential family: Binomial(∗, θ), Negative Binomial(∗, θ), Poisson(θ), N (∗, θ), N (θ, ∗), Gamma(∗, θ), Gamma(θ, ∗), etc., where ∗ indicates a given/known second parameter value. • The following “Delta Method” result, combined with the Theorem above, is useful for finding the distribution of functions of θ̂n , e.g., g(θ̂n ). Because the result has wide applicability, we state it in a general form. Delta Method: For a sequence {Tn } of real-valued random variables, suppose it holds ¡ ¢ √ d that n(Tn − a) −→ N 0, σ 2 (a) , for some a ∈ R. Then for any function g : R → R that is continuously differentiable at a with derivative g 0 (a) 6= 0, ´ ³ ´ √ ³ d n g(Tn ) − g(a) −→ N 0, [g 0 (a)]2 σ 2 (a) as n → ∞ 1 See Bickel and Doksum for details 2 Example: Let X1 , X2 , . . . , Xn be iid Poisson(θ), θ > 0. 1. Show that Tn ≡MLE of γ(θ) = θ/(1 + θ2 ) is consistent. ¢ √ ¡ 2. Find the limiting distribution of n Tn − γ(θ) . 3