关键词 > MATH96046/MATH97073
MATH96046/MATH97073 Statistical Theory: Coursework
发布时间:2022-03-14
Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: daixieit
MATH96046/MATH97073 Statistical Theory: Coursework
1. The inverse Gamma distribution with parameters a0,θ > 0, denoted IG(a0,θ), is a continuous distribution on (0, ∞) with probability density function
x −a0 − 1e −θ/x, x > 0.
Γ(a0)
Suppose that X1,...,Xn ∼iid IG(a0,θ), where a0 is known. We will consider giving θ a Gamma Γ(α,β) prior, i.e. one with density function
π(θ) = θα − 1e −βθ, θ > 0,
where Γ(α) =R0∞ tα − 1e −tdt is the usual Gamma function.
You may use the following properties in this question: for Z1,...,Zm ∼iid Γ(α,β), we have
EZi = Var(Zi) =
1/Zi ∼iid IG(α,β),
m
XZi ∼ Γ(mα,β).
i=1
(a) Show that {IG(a0,θ) : θ > 0} forms an exponential family of distributions. [1 mark]
(b) Suppose θ ∼ Γ(α,β) under the prior. What is the posterior distribution of θ given the
observations (X1,...,Xn)?
Compute the Bayesian estimator for θ (also called π-Bayes decision rule) for squared error loss. Is it an admissible estimator? Justify your answer. [3 marks]
(c) Find the MLE for θ and show that the posterior mean θ¯n converges to the MLE as α,β → 0. [1 mark]
(d) Find the Jeffreys prior for θ and the corresponding posterior distribution. Comment on the connection between the MLE and the Jeffreys prior. [2 marks]
Consider now the frequentist model where X1,...,Xn ∼iid IG(a0,θ0) for some ‘true’ θ0 > 0 and assume na0 > 2 (a0 is still known). Consider the generalized posterior mean θ¯n again in the limit β → 0 but with α ∈ R (as opposed to α > 0), i.e. take the formula in (c) but let α ∈ R. We will now treat θ¯n as a frequentist estimator.
(e) Compute the bias and variance of the resulting estimator. [3 marks]
(f) Compute the mean-squared error (MSE) of θ¯n = θ¯n,α and find the value α = α∗ ∈ R that minimizes MSE(θ¯n). [2 marks]
/ 2
(g) Is the MLE admissible? Justify your answer.
Does the existence of θ¯n,α ∗ in (f) contradict what you might have expected regarding the admissibility of the MLE given parts (b) and (d)? Explain why or why not. [2 marks]
2. Let X1,...,Xn ∼iid N(θ,θ), where θ ∈ Θ = (0, ∞).
(a) Find a minimal sufficient statistic for θ. [1 mark] (b) Recall from Problem Sheet 2 Q7(b) that the MLE in this model is
θˆn = − + r
+ X
,
where X =
P
X
(you do not need to derive this). For what values of θ > 0 (if any)
is the MLE consistent? [1 mark]
(c) Derive the asymptotic distribution of (a suitably rescaled version of) X as n → ∞. Hence or otherwise, derive the asymptotic distribution of θˆn as n → ∞. [3 marks]
You may use that for Z ∼ N(µ,σ2), EZ4 = µ4 + 6µ2σ2 + 3σ4 .
(d) Without computing the Cramer-Rao bound, state whether or not the variance of θˆn attains
the Cramer-Rao bound, justifying your answer. [1 mark] [Total 20 marks]