University of California, Los Angeles

Department of Statistics

Statistics 100B Instructor: Nicolas Christou

Final exam
18 March 2021
This is a 3-hour exam: Due by 6:15 pm on Friday, 19 March

Name: UCLA ID:

Problem 1 (20 points)

Answer the following questions:
i ∼ Bernoulli(p). We are interested here in finding the
a. Let Y1 , . . . , Yn be independent random variables with YP
n
MVUE of p(1 − p) which is the variance of Yi . Let Q = i=1 Yi . Explain why Q is a minimal sufficient statistic
for p.

1 if Y1 = 1 and Y2 = 0,
b. Refer to question (a). Let X =
0 otherwise.
Show that E[X] = p(1 − p). Is X MVUE?

q(n−q)
c. Refer to question (a). Show that P (X = 1|Q = q) = n(n−1)
.

d. Refer to question (a). Find E[X|Q] and explain how to find the MVUE for p(1 − p).

1
Problem 2 (20 points)
Answer the following questions:
a. Consider the random variables Q, R and s with
Q ∼ N (µ1 , a1 σ)
R ∼ N (µ2 , a2 σ)
2
cs
∼ χ2c ,
σ2
and assume that s2 is independent of Q, R. In addition corr(Q, R) = ρ, with ρ known. The constants a1 and
a2 are also known. We want to find a confidence interval for the parameter θ = µµ12 . By considering the random
variable Q − θR explain the steps we need to follow to find the confidence interval for θ.

1 if X1 = 0,
b. Let X1 , . . . , Xn be i.i.d. Poisson random variables with parameter λ. Let Z =
0 otherwise.
Show that Z is unbiased estimator of e−λ .

Pn
c. Refer to question (b). Find the MVUE of e−λ . Hint: Use the Rao-Blackwell theorem and note that i=1
Xi
is sufficient for λ and therefore sufficient for e−λ .

d. Consider the simpleregression model yi= β0 + β1 xi + i with i ∼ N (0, σ) and 1 , . . . ,n are independent. It is
2
(n−2)Se
given that β̂1 ∼ N β1 , pPn σ and σ2
∼ χ2n−2 . Note: Here β̂1 is the unbiased estimator of β1
(xi −x̄)2
i=1
and Se2 is the unbiased estimator of σ 2 . Suppose we are interested in testing H0 : β1 = 0 against Ha : β1 6= 0.
Please provide the details for power analysis for this hypothesis testing problem.

2
Problem 3 (20 points)
Answer the following questions:
a. Suppose X1 , . . . , Xn are i.i.d. N(0, σ). Find σ̂ 2 , the MLE of σ 2 . Is it unbiased? What is the variance of σ̂ 2 ?
Is it efficient estimator of σ 2 ? Now let Y1 , . . . , Yn be another sample also from N (0, σ). Find the minimal
sufficient statistic for σ 2 .

b. Suppose X1 , . . . , Xn and Y1 , . . . , Yn are two independent random samples from N (µ1 , 20) and N (µ2 , 15) re-
spectively. Let θ = µ1 − µ2 and suppose we are testing H0 : θ = 0 against Ha : θ > 0. We will reject H0 if
x̄ − ȳ > c. Find n and c if we are using α = 0.05 and 1 − β = 0.90 for θ = 10.

c. Suppose X1 , . . . , Xn and Y1 , . . . , Ym are two independent random samples from N (µ1 , σ1 ) and N (µ2 , σ) respec-
tively. Develop the likelihood ratio test for the hypothesis H0 : σ12 = σ22 against the alternative Ha : σ12 6= σ22
(all parameters are unknown) and show that it is a function of an F statistic.

d. Suppose X1 , . . . , Xn and Y1 , . . . , Ym are two independent random samples from exp(λ1 ) and exp(λ2 ) respec-
tively. Find the likelihood ratio test for H0 : λ1 = λ2 against H0 : λ1 6= λ2 . To which test statistic is it
equivalent to?

3
Problem 4 (20 points)
Answer the following questions:
a. Consider the random vector (Y1 , . . . , Yn )0 . Suppose it follows a multivariate normal distribution with mean
0 −1
vector µ1 and variance covariance matrix σ 2 V. Verify that the MLE of σ 2 is given by σ̂ 2 = (Y−µ̂1) Vn (Y−µ̂1) ,
10 V−1 Y
where µ̂ = 10 V−1 1
. Find E[σ̂ 2 ].

b. In lecture 3 of week 9 we discussed the prediction problem in the i.i.d. case and in the non-i.i.d. case. Let
Y1 , . . . , Yn be i.i.d. random variables with Yi ∼ N (µ, σ). Using the method of Lagrange multiplier
Pnwe minimize
var[Y0 − Ŷ0 ] (the variance of the error of prediction) subject to the unbiasedness constraint w = 1 to
i=1 i
find that the predictor of a new y value, say Y0 , is Ŷ0 = Ȳ . A different approach for this prediction problem is
to use expectation by conditioning and properties of multivariate normal (conditional pdf). It can be  shown
Y0
that var[Y0 − Ŷ0 ] is minimized when Ŷ0 = E[Y0 |Y]. To apply these results, assume that the vector
Y
follows a multivariate normal distribution. Find the mean and variance of this vector and use them to find the
predictor Ŷ0 . In your final answer, replace µ with µ̂, which in this case is Ȳ , to show that the predictor will be
the same with the predictor using the method of Lagrange multiplier from class notes. Note: Y0 is independent
of (Y1 , . . . , Yn )0 .

c. Refer to question (b). Now consider the non i.i.d. case. Let Y1 , . . . , Yn be i.i.d. random variables with
E[Yi ] = µ, var(Yi ) = σ 2 , and cov(Yi , Yj ) = σ 2 cij . Use the same result as given in question (b) to find the best
0 −1
predictor of a new y value, say Y0 . Finally, if we replace µ with µ̂ = 1 0Σ −1Y the predictor will be the same as
1 Σ 1
0
the one obtained using the method of Lagrange multiplier. Note: Y0 is not independent   of (Y1 , . . . , Yn ) . You
Y0
can assume that cov(Y0 , Yi ) = σ 2 c0i . Hint: Find the distribution of the vector and apply the results
Y
form (b).

d. Consider the simple regression model through the origin, Yi = P β1 xi + i , with i ∼ N (0, σ), 1 , 2 , . . . , n
n
xi y i
independent, and σ 2 is known. Verify that the MLE of β1 is β̂1 = Pi=1n 2
. Use the Neyman-Pearson lemma
xi
i=1
to find the most powerful test for testing
H0 : β1 = 0
Ha : β1 > 0
at the level of significance α and β1 = βa > 0. What
 is the best critical
 region? Hint: In your final answer you
will need to consider the distribution of β̂1 ∼ N β1 , pPσn .
x2
i
i=1

4
Problem 5 (20 points)
Answer the following questions:
a. Assume that Y1 , . . . , Yn follow the normal distribution with mean µy and variance σ 2 . Consider a second sam-
ple such that X1 , . . . , Xn follow the normal distribution with mean µx and variance σ 2 . The two samples are
independent. Assume that σ 2 is known. Write the test statistic for testing the hypothesis H0 : µx = µy .

b. Refer to question (a). Assume now that the data are correlated with cov(Yi , Yj ) = ρσ 2 and cov(Xi , Xj ) = ρσ 2 ,
with ρ > 0. Write the test statistic for testing the hypothesis H0 : µx = µy .

c. The p-value of a test is defined as the probability of observing the test statistic or a more extreme value under
the null hypothesis. Compare (a) with (b). Which test has a smaller p-value? Which test rejects more often
than it should?
q
d. Consider a confidence interval for µx using the correlated data of the second sample: x̄ ± 1.96σ 1+(n−1)ρ
n
.
σ
Now, suppose we fail to see the dependence in the random variables, and we decided to use x̄ ± 1.96 n . What
√

is the actual coverage of our confidence interval if n = 25, ρ = 0.2?