Universidad Católica del Norte, Economic Department

Econometrics: Problem Set 2

Professor: Mauricio Sarrias

First Semester 2020

This homework is due on 11th, May at 18:00 hrs. Please, write your answers clearly.

1 Theory
1. (Unbiased or biased estimator) Consider the linear model:

yi = x i β + " i
E("i |xi ) = 0
with n observations and xi is scalar (real-values). Consider the estimator:
"n
! x3i yi
β = "i=1
n 4 (1.1)
i=1 xi

a) Is the estimator unbiased?

Solution:First, we need to find a mathematical expression for β! − β as a

function of ". Note that yi = xi β + "i , then we can rewrite β! as:

1
 "n
x3 y
β! = "i=1 i 4 i
x
"ni=1 3i
i=1"xi (xi β + "i )
= 4
i=1 xi
"n 4β +
"n 3
i=1 x"
i i=1 xi "i
= 4
i=1 xi
"n " n
β i=1 x4i 3
i=1 xi "i
= " 4 + " 4
i=1 xi i=1 xi
"n
x3 " i
= β + "i=1 i 4
i=1 xi

Now, taking expectation, we get:

# $ % "n & '

! i i=1 x3i "i &&
E β|x = β + Ex " 4 & xi
i=1 xi
1 3
= β+"
x 4 E(xi "i |xi )
i=1 i
1 3
= β+"
x 4 xi E("
( )* i |xi )
+
i=1 i
0
= β

b) What is the variance of the estimator?

Solution: Assume homokedasticity. Taking the variance over the sampling

error we get:

# $ % '2 ,
n
! i 1
Var β|x = " 4 Var(x3i "i |xi )
i=1 xi i=1
% '2 ,
n
2 1
= σ " 4 x6i
i=1 xi i=1

c) What you can say about the Gauss-Markov theorem and the above results?
Solution: Under homokedasticity, the OLS estimator
# $ is always
# BLUE.$ So, this
! !
estimator is inefficient in the sense that Var β|xi > Var βOLS |xi .

2
2. (Error variance estimator with omitted variables) Assume the following model

yi = x′1i β10 + x′2i β20 + "i

E("i |X) = 0 (1.2)
E("2i |X) = σ02

where X = (x1i , x2i ), with x1i k1 × 1 and x2i k2 × 1. Consider the regression

yi = x′1i β1 + "i (1.3)

and define the error variance estimator

n
1 , 2
!2 =
σ !
"i (1.4)
n − k1
i=1

Find E(!
σ 2 |X).

Solution: First, we write the model with matrix notation as y = X1 β10 +X2 β20 +ε
and the short regression as y = X1 β- !. Now, let M1 = I − X1 (X1′ X1 )−1 X1′ .
10 + ε
By the properties of least-squares and the fact that M1 X1 = 0,

ε! = M1 y
= M1 (X1 β10 + X2 β20 + ε) (1.5)
= M1 (X2 β20 + ε)

Thus, since M1 is idempotent

σ 2 = ε!′ ε!
(n − k1 )!
= [M1 (X2 β20 + ε)]′ [M1 (X2 β20 + ε)]
(1.6)
= (X2 β20 + ε)′ M1 M1 (X2 β20 + ε)
= ε′ M1 ε + β20
′ ′
X2 M1 X2 β20 + 2β20 X2′ M1 ε

Since X2 and M1 are functions of X, and E(ε|X) = 0

. /
σ 2 |X) = E ε′ M1 ε|X + β20
(n − k1 )E(! ′
X2 M1 X2 β20
′
(1.7)
= (n − k1 )σ 2 + β20 X2 M1 X2 β20

The second inequality since E(εε′ |X) = Iσ 2 and tr (M1 ) = rk (M1 ) = (n − k1 )

imply that

. / 0 . /1
E ε′ M1 ε|X = tr M1 E εε′ |X = tr(M1 )σ 2 = (n − k1 )σ02 (1.8)

3
 Therefore,

1
σ 2 |X) = σ02 +
E(! β ′ X2 M1 X2 β20 (1.9)
n − k1 20

3. (Different ways of testing restrictions): Consider the linear regression model:

yi = β1 + β2 xi2 + β3 xi3 + "i (1.10)

Explain how you could estimate this model subject to the restriction that β2 +β3 =
1 by running a regression that imposes the restriction. Also, explain how you could
estimate the unrestricted model in such a way that the value of one of the coeffi-
cients would be zero if the restriction held exactly for your data.

Solution: The restriction implies that β3 = 1 − β2 . Subtituting this into 1.10 gives

yi = β1 + β2 xi2 + (1 − β2 )xi3 + "i

(1.11)
yi − xi3 = β1 + β2 (xi2 − xi3 ) + "i

The second equation here has an observable variable on the left-hand side and
two observable variables, each multiplied by a parameters, on the right-hand side.
Define the new variables yi′ = yi − xi3 and zi = xi2 − xi3 . The restricted regression
is

yi′ = β1 + β2 zi + ui (1.12)

Thus running regression 1.12 estimates the model subject to the restriction. The
estimate β!2 is obtained directly, and, since the estimate of β3 must satisfy the
restriction, we see that β!3 = 1 − β!2 . We can add a third element term to the
right-hand side of equation 1.12 in order to produce a model that is equivalent to
1.10. This gives

yi′ = β1 + β2 zi + (β2 + β3 − 1)xi3 + ui (1.13)

Thus, if we run the regression:

yi′ = β1 + β2 zi + γxi3 + ui (1.14)

! − β!2 + 1. If the
we obtain estimates of β1 an β2 directly, and, for β3 , we use γ
restriction held exactly in the data (a highly unlikely event), the estimate of γ
would be 0

4
4. (Hypothesis Testing): Assume you have a random sample from the model:

yi = β1 xi + β2 x2i + "i
(1.15)
E("i |xi ) = 0

where yi is wages measured as chilean pesos per hour, and xi is age. Describe how
you test the hypothesis that the expected wage for a 40-year-old worker is $20 an
hour. You do not need to derive the theory behind your procedure.

Solution: Since E(yi |xi ) = β1 xi + β2 x2i , then E(yi |x − i = 40) = 40β1 + 402 β2 . The
hypothesis is thus

H0 : 40β1 + 402 β2 = 20 (1.16)

which is a linear restriction. If desired, this can be rewritten as

H0 : 2β1 + 80β2 = 1 (1.17)

Let β!1 and β!2 be the OLS estimates of the coefficients, and let V! denote the
estimated covariance matrix. The F statistic for this hypothesis is:
2 3−1
F = (Rβ! − r)′ RV! R′ (Rβ! − r) (1.18)

where

. /
R = 2 80 , r=1 (1.19)

Then,
# $2
2β!1 + 80β!2 − 1
F = (1.20)
RV! R′
It has an F distribution under H0 . A 5% size test is to reject H0 if F exceeds the
5% critical value. Otherwise, H0 is not rejected.

5. (Restricted regression and F ratio) We derive the F test in class. However, there
is a convenient alternative formula involving two different sum of squared residuals
that you probably learned in your introductory class of Econometrics : one is SSR,
the minimized sum of squared residuals denoted as SSRU , and the other is the
restricted sum of squared residuals, denoted SSRR , obtained from:

4 st Rβ4 = r
min SSR(β)
β

5
Finding the β4 that achieves this constrained minimization is called the restricted
regression or restricted least squared. In this exercise I ask you to show that
the F -ratio you learned in class equals:

(SSRR − SSRU )/#r

F =
SSRU /(n − K)

In the restricted least squares, the sum of squared residuals is minimized subject
to the constraint implied by the null hypothesis Rβ0 = r. Form the Lagrangian
as:

1# $′ # $ # $
L= y − X β4 y − X β4 + λ′ Rβ4 − r ,
2
where λ here is the #r-dimensional vector of Lagrange multipliers (recall: R is
#r×K, β4 is K ×1, and r is #r×1). Let β4 be the restricted least squares estimator
of β. It is the solution to the constrained minimization problem.
a) Let β! be the unrestricted OLS estimator. Show:
. /−1 ′ 2 . ′ /−1 ′ 3−1 # $
β4 = β! − X ′ X R R XX R Rβ! − r
2 . /−1 ′ 3−1 # $ .
λ = R X ′X R Rβ! − r

Solution:
From the Lagrangian:

∂L
= X ′ y − (X ′ X)β4 − R′ λ
∂ β4 (1.21)
′ 4 − R′ λ
= X (y − X β)

So that

X ′ (y − X β)4 − R′ λ = 0
4 = R′ λ
X ′ (y − X β) (1.22)
β4 = (X ′ X)−1 (X ′ y − R′ λ)
and

∂L
= (Rβ4 − r) (1.23)
∂λ

6
 Combining 1.22 and 1.23, we get

(Rβ4 − r) = 0
. /
R (X ′ X)−1 (X ′ y − R′ λ) − r = 0
R(X ′ X)−1 X ′ y − R(X ′ X)−1 R′ λ − r = 0
(1.24)
R(X ′ X)−1 X ′ y − r = R(X ′ X)−1 R′ λ
. /−1 . /
R(X ′ X)−1 R′ R(X ′ X)−1 X ′ y − r = λ
. /−1 # $
R(X ′ X)−1 R′ Rβ! − r = λ

Thus,

2. /−1 # $3
β4 = (X ′ X)−1 X ′ y − (X ′ X)−1 R′ R(X ′ X)−1 R′ Rβ! − r
. /−1 # $ (1.25)
β4 = β! − (X ′ X)−1 R′ R(X ′ X)−1 R′ Rβ! − r

b) Show that Rβ̃ = 0.

Solution: Note that:

5 6
. /−1 ′ 2 . ′ /−1 ′ 3−1
Rβ̃ = R β! − X ′ X R R XX R Rβ!
. /−1 ′ 2 . ′ /−1 ′ 3−1
= Rβ! − R X ′ X R R XX R Rβ!
( )* + (1.26)
I
= Rβ! − Rβ!
=0
# $
c) Find Var β̃|X .

Solution: Rearranging some terms:

. /−1 ′ 2 . ′ /−1 ′ 3−1

β̃ = β! − X ′ X R R XX R Rβ!
% '
. ′ /−1 ′ 2 . ′ /−1 ′ 3−1
= I− XX R R XX R R β! (1.27)
( )* +
A
= Aβ!

7
 Then,
# $
!
Var β̃|X = Var(Aβ|X)
!
= A′ Var(β|X)A (1.28)
# . /−1 $
= A′ σ 2 X ′ X A

d) Give and expression for a valid standard error for the elements of β̃.

ε!′ ε!
Solution: Let σ
!2 = n−K . Therefore,

. /−1
V! = σ
! 2 A′ X ′ X A (1.29)

The standard errors are the square roots of the diagonal elements of V! .
4 the residuals from the restricted regression. Show:
e) Let ε4 ≡ y − X β,

. /
SSRR − SSRU = (β! − β)
4 ′ X ′ X (β! − β)
4 = ε4′ P ε4.

where P is the projection matrix. (Hint: Add and subtract X(β! − β)

4 to
4
ε4 ≡ y − X β).

Solution: By using the add-and-subtract strategy, we obtain:

ε4 = y − X β4 + X(β! − β)
4 − X(β! − β)
4
(1.30)
! + X(β! − β)
= (y − X β) 4

8
 Therefore:

SSRR = ε4′ ε4
2 3′ 2 3
! + X(β! − β)
= (y − X β) 4 ! + X(β! − β)
(y − X β) 4

! ′ (y − X β)
= (y − X β) ! + 2(β! − β)
4 ′ X ′ (y − X β)
! +(β! − β)
4 ′ X ′ X(β! − β)
4
( )* +
0
! ′ (y − X β)
= (y − X β) ! +(β! − β)
4 ′ X ′ X(β! − β)
4
( )* +
SSRU

SSRR − SSRU = (β! − β)

4 ′ X ′ X(β! − β)
4
# $2 . /−1 ′ 3−1 # $
= Rβ! − r R X ′ X R Rβ! − r Using Equation 1.25
. / −1 ′
= λ′ R X ′ X R λ Using Equation 1.24
. ′ /−1 ′
′
= ε! X X X X ε! by FOC: X ′ (y − X β) 4 = R′ λ
= ε!′ P ε!
(1.31)
f) Verify that you have proved in (b) that both F -ratios are equal.

Solution: The F -ratio is defined as

# $′ 2 3−1 # $
Rβ! − r R (X ′ X)−1 R′ Rβ! − r /r
F ≡ (1.32)
!2
σ
# $′ 2 3−1 # $
As we shown Rβ! − r R (X ′ X)−1 R′ Rβ! − r = SSRR −SSRU , the
F -ratio can be rewritten as

(SSRR − SSRU )/r

F =
!2
σ
(SSRR − SSRU )/r
= (1.33)
ε!′ ε!/(n − K)
(SSRR − SSRU )/r
SSRU /(n − K)

6. (Chow Test) Assume that you are interested in whether the coefficients of one
group of the data (women, for example) are equal to those of other group (men).
For instance, consider the following regressions

yi = x′i β + "i for all observations

yi1 = x′i1 β1 + "i1 for n1 observations (group 1) (1.34)
yi2 = x′i2 β2 + "i2 for n2 observations (group 2)

9
Explain how you would test H0 = β1 = β2 . (Hint: Use the F test you derived
in previous question and think that the pooled model (model 1) is the restricted
model, whereas allowing different coefficients for each group is the unrestricted
model. )

Solution: Recall that the F statistic can be written as:

(SSRR − SSRU )/r

F = (1.35)
SSRU /(n − K)
where SSRr is the sum of squared residuals for the restricted model, whereas SSRU
is the analog for the unrestricted model.
What is the restricted and unrestricted model in this case? First, note that the
hypothesis that each coefficient is the same across the two groups involves K restric-
tions (restricted model). The unrestricted model, which we can think of as having
a group dummy variable and K interactions terms in addition to the intercept and
variable themselves, has (n1 + n2 − 2 ∗ K) degrees of freedom.
Ok, but what is the the SSR for the unrestricted model? Basically, the unrestricted
model is that model that allows the coefficients to vary for each group, so SSRU =
SSR1 + SSR2 , where SSR1 is the SSR for the first model, whereas SSR2 is the
SSR for the second model. The restricted sum of squared residuals is just the SSR
from pooling the groups and estimating a single equation, say SSRP , then:

(SSRP − (SSR1 + SSR2 ))/K

F = (1.36)
SSRP /(n − 2K)
where n is the total number of observations. It is important to note that because
the Chow test is just and F statistic, it is only valid under homokedasticity. In
particular, under the null hypothesis, the error variances for the two groups must
be equal.

10