London School of Economics Vassilis Hajivassiliou

Department of Economics Michaelmas Term 2021

MSc Econometrics (Ec402)

2021–2022
Answers to Problem Set #6

1. .

(a) Including all 4 guaranteed income dummies would cause perfect collinearity among
the regressors leading to singularity in X ′ X (the “dummy variable trap”) since
for every i, d1i + d2i + d3i + d4i = 1.
(b) The Ehi part of equations (2) and (3) are equivalent, since they differ by a non-
singular transformation:
E(hi |d1i , x2i ) formulation (2) formulation (3)
d1i = 1 β0 γ θ1 δ
d2i = 1 β0 + β2 γ θ2 δ
d3i = 1 β0 + β3 γ θ3 δ
d4i = 1 β0 + β4 γ θ4 δ
As a result, γ = δ, β0 = θ1 , β0 + β2 = θ2 , β0 + β3 = θ3 , and β0 + β4 = θ4 .

2. .
∑
(a) Let the number of observations in quarter q, q = 1, 2, 3, 4 be Tq with T = 4q=1 Tq .
 
T1 0 0 0
 0 T2 0 0 
Given the definition of the seasonal dummies, S ′ S =  

 and
 0 0 T3 0 
0 0 0 T4
∑ ∑ ∑ ∑ ′
S ′ z = ( i∈q1 zi z
i∈q2 i z
i∈q3 i z
i∈q4 i ) . Thus
∑ 
i∈q1 zi
1 1 1 1 ∑ z 
 
MS z = z − S · diag( , , , ) ·  ∑i∈q2 i  = z − z̄ (q) ,
T1 T2 T3 T4  z 
∑i∈q3 i
i∈q4 zi
where z̄ (q) is the T × 1 vector with ith element the sample average of all the
observations from the quarter i belongs to.
(b) From partitioned regression results, we know that in ŷ = S β̂1 +X2 β̂2 the estimated
vector

β̂2 = (X2′ MS X2 )−1 X2′ MS y = ((MS X2 )′ MS X2 )−1 (MS X2 )′ MS y

by the idempotency and symmetry of MS . This corresponds to regressing MS y

on MS X2 , i.e., deseasonalized y on deseasonalized X2 .

1
3. Explain what we mean by the “Normal Linear Regression Model (NLRM)”.

NLRM-A4GM: β̂ols |X ∼ N (β true , σ 2 (X ′ X)−1 )

NLRM-A4Omega: β̂ols |X ∼ N (β true , c2 (X ′ X)−1 X ′ ΩX(X ′ X)−1 )
Outline the conditions that characterize this model.
NLRM-A4GM: A1 + A2 + A3Rmi or stronger A3 + A4GM (iid) + A5Gaussian
NLRM-A4Omega: A1 + A2 + A3Rmi or stronger A3 + A4Ω + A5Gaussian
Repeat with the term “Asymptotic Linear Regression Model (ANLRM)”.
Part 1: If “A” is taken to mean “Approximate NLRM for very large, *finite*
sample size S”
ANLRM-A4GM: β̂ols |X ≈ N (β true , σ 2 (X ′ X)−1 )
ANLRM-A4Omega: β̂ols |X ≈ N (β true , c2 (X ′ X)−1 X ′ ΩX(X ′ X)−1 )
Outline the conditions that characterize this model.
ANLRM-A4GM: A1 + A2+ A3Rsru or stronger (i.e., *any* of the 7 versions of
A3) + A4GM (iid) + S very large.
ANLRM-A4Omega: A1 + A2 + A3Rsru or stronger (i.e., *any* of the 7 versions
of A3) + A4Ω + S very large.
Part 2: If “A” is taken to mean “Asymptotic NLRM for sample size S
growing unboundedly large (i.e., limits as S → ∞)”
ANLRM-A4GM:

( ( ))−1 
√ d ′
X X
S(β̂ols − β true ) → N 0, σ p lim
2 
S
S→∞

ANLRM-A4Omega:
 ( ( ))−1  ( ) ( ( ))−1 
√ d ′ ′ ′
XX X ΩX  XX
S(β̂ols −β true ) → N 0, c2  p lim  p lim p lim 
S S S
S→∞

Outline the conditions that characterize this model.

ANLRM-A4GM: A1 + A2+ A3Rsru or stronger (i.e., *any* of the 7 versions of
A3) + A4GM (iid) + S → ∞.
ANLRM-A4Omega: A1 + A2 + A3Rsru or stronger (i.e., *any* of the 7 versions
of A3) + A4Ω + S → ∞.

2
(a) Based on the NLRM, discuss which quantities will be distributed as χ2 (·) and
which as F (·, ·).
These distributional results are discussed on pages 94-96 of the Sup-
plementary Technical Notes of the Coursepack.
The three pages are appended to these answers here.
(b) Will the same two distributions be relevant for the case of the ANLRM?
As S → ∞, the random variables that are in the denominators of the t-statistic
and of the F-statistic will converges to something fixed — this is because the
denominators contain the estimated s2ols which converges asymptotically to the
true σ 2 — you can also see this from the fact that the t-statistic will be distributed
t(S − k) and the F-statistic will be F (r, S − k). As S → ∞,
t(S − k) → t(∞) = N (0, 1)
F (r, S − k) → F (r, ∞) = χ2 (r)/r
In practice, econometricians recommend still using the t(.) and F (., .) distributions
even for very large S, because the resulting rejection regions are more conservative
compared to their N (0, 1) and χ2 (r) approximations.

⃝
c Vassilis Hajivassiliou, LSE 2000-2021