Statistical Inference
B. Statistical Data Science 2nd Year Indian Statistical Institute
Teacher: Soham Sarkar
Exercise Series 12
Exercise 1. For the following examples, develop the likelihood ratio test, Wald’s test and Rao’s score
test for testing H0 : θ = θ0 against Hleft : θ < θ0 , Hright : θ > θ0 and Hboth : θ ̸= θ0 . Clearly
specify the test statistic and the cut-off value. Check whether the regularity conditions required for
the asymptotic results hold. In each case, draw the power function of the three tests on the same graph
for n = 5, 10, 25, 50, 100 (fixing some specific value of θ0 ). Based on these plots, what can you say
about the relative merits/demerits of the three tests? Do your conclusions depend on the particular
value of θ0 ?
i.i.d.
(a) X1 , . . . , Xn ∼ Binomial(p), 0 ≤ p ≤ 1. θ = p.
i.i.d.
(b) X1 , . . . , Xn ∼ Poisson(λ), λ > 0. θ = λ
i.i.d.
(c) X1 , . . . , Xn ∼ Geometric(p), 0 ≤ p ≤ 1. θ = p.
i.i.d.
(d) X1 , . . . , Xn ∼ Uniform(0, θ), θ > 0.
i.i.d.
(e) X1 , . . . , Xn ∼ Normal(µ, 1), µ ∈ R. θ = µ.
i.i.d.
(f) X1 , . . . , Xn ∼ Normal(0, σ 2 ), σ > 0.
i.i.d.
(g) X1 , . . . , Xn ∼ Exponential(θ), θ > 0.
i.i.d.
(h) X1 , . . . , Xn ∼ Laplace(µ, 1), µ ∈ R. θ = µ.
i.i.d.
(i) X1 , . . . , Xn ∼ Laplace(0, σ), σ > 0. θ = σ.
i.i.d.
(j) X1 , . . . , Xn ∼ Normal(θ, θ2 ), θ ∈ R.
i.i.d.
(k) X1 , . . . , Xn ∼ Normal(θ, θ), θ > 0.
indep.
Exercise 2. Consider the linear regression model Yi ∼ Normal(α + βxi , σ 2 ), i = 1, . . . , n, where
α, β ∈ R and σ 2 > 0 are unknown parameters. We are interested in testing H0 : β = 0 vs. Ha : β ̸= 0.
(a) Formulate the likelihood ratio test, Wald’s test and Rao’s score test for this problem.
(b) Compare the power of these three tests numerically.
indep.
(c) Plot the power functions of the tests in part (a) when Yi ∼ Laplace(α + βxi , σ), i = 1, . . . , n.
What do you observe?
indep.
(d) Repeat part (c) with Yi ∼ Cauchy(α + βxi , σ), i = 1, . . . , n.
1
Note: The goal of parts (c) and (d) is to study the effect of misspecification on the inferential
procedure. The effect can be severe in many cases.
(e) Repeat parts (a) and (b) for testing H0 : (α, β) = (0, 0) vs. Ha : (α, β) ̸= (0, 0).
Exercise 3. Recall the table giving the number of persons (out of a total of 500) having four different
blood groups A, B, AB and O, which we considered in Series 2, Exercise 5 and Series 7, Exercise 6. Such
tables are known as 2 × 2 contingency tables and are generally of the following form.
B = b1 B = b2 Total
A = a1 n11 n12 n1·
A = a2 n21 n22 n2·
Total n·1 n·2 n
Here, we have two attributes A and B, each having two levels a1 , a2 and b1 , b2 , respectively. In
a collection of n items, nij denotes the number of items having attributes ai and bj , i, j = 1, 2. The
number ni· = ni1 + ni2 denotes the total number of items having attribute ai , i = 1, 2. Similarly,
n·j = n1j + n2j denotes the total number of items having attribute bj , j = 1, 2. It also holds that
n11 + n12 + n21 + n22 = n1· + n2· = n·1 + n·2 = n. If we denote pij to be the probability of a randomly
selected item having attributes ai and bj , and assume that the attributes are assigned independently
and identically among the items, then we can use a Multinomial distribution to model the data. We
already did this in Series 7, Exercise 6.
Here, we are interested in checking whether the attributes A and B are independent. That is, the
level of the attribute A (a1 or a2 ) in an item does not affect the level of the attribute B (b1 or b2 ) in
the item, and vice verse. Defining pi· = pi1 + pi2 and p·j = p1j + p2j , i, j = 1, 2, to be the marginal
probabilities, this is same as testing H0 : pij = pi· p·j ∀ i, j = 1, 2 vs. Ha : H0c . Formulate the likelihood
ratio test, Wald’s test and Rao’s score test for this problem. Clearly specify the test statistics and the
cut-off values. What are the similarities/dissimilarities among the tests?