STAT7301 Mathematical Statistics Semester One Final Examinations, 2021
Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: daixieit
STAT7301 Mathematical Statistics
Semester 1, 2021
Question 1. [13 marks]
(a) Let t = (x1 , . . . , xn )T contain an observed random sample on a random variable X with prob- ability density function f (x; θ) specified up to an unknown vector of parameters θ . The random sample is denoted by
W = (X1 , . . . , Xn )T .
Suppose a null hypothesis H0 : θ = 0 versus an alternative hypothesis H1 : θ 0 is to be rejected if the observed value T (t) of the test statistic T (W) is sufficiently large in magnitude.
(i) Give a definition of the P-value of a test in general. [1 mark]
(ii) Give a definition via a mathematical expression for the P-value of the test proposed here. [1 mark]
(b) Let
X1 , . . . , Xn1 i. . N (µ1 , σ 1(2)),
Y1 , . . . , Yn2 i. . N (µ2 , σ2(2)),
denote two independent random samples from two classes C1 and C2 , respectively.
σ 1(2) = c σ2(2); (1)
that is, σ 1(2) = σ2 and σ2(2) = c σ2 , where c is known.
(i) Write down the likelihood equation,
∂ log L(θ)/∂θ = 0,
for this problem. [2 marks]
(ii) Give its solution, yielding the maximum likelihood (ML) estimates of µ1 , µ2 , and σ2 . Note you are not required to verify that the solution is a (global) maximizer of the likelihood function. [2 marks]
(iii) Give the distribution of the ML estimator ˆ(σ)2 and explain how your answer was obtained. [3 marks]
(iv) Consider the test of
H0 : µ1 = µ2 vs. H1 : µ1 µ2 ,
on the basis of the test statistic,
T = X - Y
2ˆ(σ)2 /n1 + c ˆ(σ)2 /n2 ,
where ˆ(σ)2 is the ML estimate of σ2 .
Give an exact result for the null distribution of this test statistic T under assumption (1), with an explanation as to how it was obtained. [4 marks]
Question 2. [17 marks]
A random sample X1 , . . . , Xn is taken from a population in which the random variable X has proba- bility function or density function f (x; ), which belongs to the regular exponential family,
f (x; ) = b(x) exp{c( )Tt(x)}/a( ), (2)
where is the vector of unknown parameters. Hence the likelihood function L( ) is given by
L( ) = h(x) exp{c( )TT (x)} {a( )}−n ,
where x = (x1 , . . . , xn )T is the observed random sample and where
T (x) = t(xj ),
h(x) = Ⅱ b(xj ).
j=1
The maximum likelihood (ML) estimate ˆ of can be shown to satisfy the equation
E{T (X)} = T (x); (3)
that is,
[E{T (X)}] ˆ = T (x),
=
where X is the random vector with particular value x.
(i) Establish the result (3) in the simplified case where the canonical parameter c( ) is equal to . [2 marks]
(ii) Using the result (3) where the canonical parameter c( ) is not necessarily equal to , show that
s (ˆ ) = I(ˆ ), (4)
where I( ) is the negative of the Hessian of log L( ) and s ( ) is the Fisher (expected) infor- mation matrix. To simplify your solution, you may take to be a scalar. [3 marks]
(iii) Why is T (X) a sufficient statistic for and why is it a minimal sufficient statistic? [2 marks]
(iv) Suppose U is an unbiased estimator of . Show that
W = E{U } T}
is also an unbiased estimator of . [2 marks]
(v) Consider for simplicity the case of only a single parameter and so where 9 , U , I , and T are scalars. Show that
var(W) < var(U), (5)
and state where equality occurs in (5). [3 marks]
(vi) Let
g(9) = f (k; 9),
where k is some specified value of the random X with probability function f (x; 9).
Give the ML estimate of g(9). [1 mark]
(vii) Show that
U (T) = pr{Xj = k } T(
is the uniform minimum variance unbiased (UMVU) estimator of g(9), where Xj is a member (before it is observed) of the random sample from which T is formed. [4 marks]
Question 3.
Consider a Markov Chain y1 , y2 , . . . , yn with joint probability density function (pdf) [5 marks]
f (y1 , y2 , . . . , yn ) = f (y1 )f (y2 }y1 )f (y3 }y2 ) . . . f (yn }yn − 1 ) .
(i) Draw the Bayesian network corresponding to this Markov Chain. [2 marks]
(ii) Show that the joint pdf can also be factorised in “reverse time” order, i.e., [3 marks]
f (y1 , y2 , . . . , yn ) = f (yn )f (yn − 1 }yn )f (yn −2 }yn − 1 ) . . . f (y1 }y2 ) .
Question 4. [11 marks]
Let x1 , x2 , . . . , xn be iid observations from a Geometric distribution, each with probability function
f (x} p) = (1 - p)zp , x = 0, 1, 2, . . . ,
where p e [0, 1] is an unknown parameter. Note that in this parametrisation, x represents the number of failures before the first success. Consider a prior distribution on p given by p ~ Beta(α, β) with pdf
f (p) = pa − 1 (1 - p)8 − 1 , p e [0, 1] ,
where B(α, β) = Γ(α)Γ(β)/Γ(α + β).
(i) Find an expression for the posterior pdf of p given the data t = (x1 , x2 , . . . , xn ) and identify its distribution. [3 marks]
(ii) Is the prior distribution p ~ Beta(α, β) conjugate for this problem? [1 mark ]
(iii) Show that the mean of any Beta(a, b) distribution is given by a/(a + b). [2 marks]
(iv) Using parts (i), (ii) and (iii), or otherwise, show that the posterior mean of p given t is
E(p} t) = . [1 mark]
(v) What happens to the posterior mean as sample size n → o? [1 mark]
(vi) What happens to the posterior mean if prior hyperparameters α, β → 0? [1 mark]
(vii) Give an interpretation of the posterior mean by completing the following sentence:
“The effect of the prior hyperparameters α, β on the posterior mean of p is like . . . ” [2 marks]
Question 5. [14 marks]
Consider the following Bayesian model:
f(µ, σ2 ) x 1/σ2 , µ e R, σ2 > 0 , x1 , . . . , xn } µ, σ2 N(µ, σ2 ) .
(i) Derive an expression for the joint posterior pdf of (µ, σ2 ) given t = (x1 , . . . , xn ). [2 marks]
(ii) Show that
n n
(xi - µ)2 = (xi - ¯(x))2 + n(µ - ¯(x))2 .
i=1 i=1 [2 marks]
(iii) Using part (ii), or otherwise, show that the conditional posterior f(µ}σ2 , t) is normal. Find the mean and variance of this normal distribution. [2 marks]
(iv) Show that the conditional posterior f(σ2 }µ, t) is Inverse-Gamma. Find the shape and scale of this Inverse-Gamma distribution. [2 marks]
(v) Describe how you would sample from the joint posterior f(µ, σ2 }t). [4 marks]
(vi) Describe how you construct posterior 90% credible intervals for the mean parameter µ and the variance parameter σ 2 . [2 marks]
2023-06-10