ECON0064: Examination in Econometrics

2018-2019

There are two sections to the exam. Answer ALL questions from Section A
and TWO questions from Section B. The total number of points is 100. Section A
contributes 60 points and Section B contributes 40 points.

Fully justified answers are required to obtain high marks. However, answers are not
expected to exceed 10-15 lines. In cases where a student answers more questions than
requested by the examination rubric, the policy of the Economics Department is that the
students first set of answers up to the required number will be the ones that count (not
the best answers). All remaining answers will be ignored. Calculators are permitted.

Time allowed: 3:00 hours

1
Section A
Answer ALL questions from this Section.

Question A1 (20 points)

Consider the linear regression model
yi = β1 + β2 hi + ui ,
where yi is a measure of health status of individual i, hi is a dummy variable which
indicates whether individual i smokes (hi = 1) or not (hi = 0), and ui is an unobserved
error. Assume that E(ui|hi) = 0. Let xi = (1, hi ) and β = (β1 , β2 )0 . We observe an
iid sample of (yi , hi ), i = 1, . . . , n. In total we observe n = 120 individuals, of which
Pn Pn
n0 = i=1 (1 − hi ) = 80 do not smoke and n1 = i=1 hi = 40 smoke. The average
measure of health in the subpopulation with hi = 0 is n10 ni=1 (1 − hi )yi = 4 and in the
P

subgroup of individuals with hi = 1 is n11 ni=1 hi yi = 0.5.

P

Pn 0 Pn 0
(a) Calculate the 2 × 2 matrix i=1 xi xi , the 2 × 1 vector i=1 xi yi , and the OLS
−1
estimator βb = ( n x0 xi ) n
x0 yi .
P P
i=1 i i=1 i

(b) We now also assume homoscedasticity and E(u2i |hi ) = 20. Use this information to
calculate OLS standard errors for βb1 and βb2 .

For the following subquestion assume that you calculated βb2 = −2 and std(
c βb2 ) = 0.5
(these are not the numbers that you should have actually obtained).

(c) Using a large sample t-test, can you reject the null hypothesis H0 : β2 ≥ 0 against
the alternative hypothesis H0 : β2 < 0 at 95% confidence level? (Hint: The 97.5%
quantile of N (0, 1) is 1.96, the 95% quantile of N (0, 1) is 1.64, the 90% quantile of
N (0, 1) is 1.28.)
1
Pn 1
Pn
(d) Let ỹi = yi − n i=1 yi and h̃i = hi − n i=1 hi . These are demeaned versions of
the variables yi and hi . Let δ̂ be the OLS estimator obtained from regressing yi on
h̃i only. Let θ̂ be the OLS estimator obtained from regressing ỹi on h̃i only. What
is the relationship between δ̂, θ̂ and β̂2 ?

(e) It is suggested to run a regression of yi on a constant, hi and (1 − hi ), that is, to

include three regressors (1, hi , 1 − hi ) in the linear regression. Explain whether this
is a good idea, and what you would expect from the resulting OLS estimator.

2
Question A2 (20 points)
We observe an i.i.d. sample (yi , xi , zi ), i = 1, . . . , n, where n is the sample size, and yi ,
xi and zi are three variables with finite second moments. For simplicity we assume that
Eyi = Exi = Ezi = 0 and Eyi2 = Ex2i = Ezi2 = 1. We consider the model
yi = xi β + ui ,

where ui is a mean zero error term, and β is the parameter of interest. We denote
E(ziui) = ρzu, E(xizi) = ρxz and E(xiui) 6= 0, where |ρxz | < 1, |ρxu| < 1, |ρzu| < 1, and
ρxz 6= 0. We observe yi , xi and zi for an iid sample of observations i = 1, . . . , n, while ui
is an unobserved error term. The 2SLS (or IV) estimator for β reads
Pn
zi yi
β = Pni=1
b .
i=1 zi xi

(a) Show that βb →p β ∗ as n → ∞, and find an expression for β ∗ in terms of β, ρxz and
ρzu .

(b) Under what condition on ρzu is βb consistent for β? What property of the instru-
mental variable zi is described by this condition on ρzu ? Also, what property of the
instrumental variable zi is described by the condition ρxz 6= 0?
√ √ b
(c) Assume that we know ρzu = 0. Show that n(βb − β) ⇒ N (0, AsyVar( nβ)), as
√ b
n → ∞, and find an expression for AsyVar( nβ).

(d) Continue to assume that ρzu = 0. For a sample with n = 200 observations we
calculate n1 ni=1 zi yi = 0.4, n1 ni=1 zi xi = 0.2, n1 ni=1 zi2 x2i = 3, n1 ni=1 zi2 xi yi = 0.5
P P P P

and n1 ni=1 zi2 yi2 = 5. Use this information and your result in (c) to calculate an
P

asymptotically valid 95% confidence interval for β. (Hint: The 97.5% quantile of
N (0, 1) is 1.96, the 95% quantile of N (0, 1) is 1.64, the 90% quantile of N (0, 1) is
1.28.)

(e) Now assume that ui = wi γ + i , where E(i zi ) = 0, E(wi zi ) 6= 0, and (yi , xi , zi , wi ),

i = 1, . . . , n, are i.i.d. Is βb a consistent estimator for β in this case? Show your
derivations and explain your answer.

3
Question A3 (20 points)
We observe an iid sample (yi , xi ), i = 1, . . . , n. We assume that the scalar outcome yi and
the single regressor xi have zero mean, and consider the model:

y i = xi β + u i , ui | xi ∼ N (0, h(xi )),

meaning the unobserved error ui is normally distributed, conditional on xi , with mean

zero and variance h(xi ) > 0. The parameter of interest is β ∈ R. Assume that xi , h(xi )
and 1/h(xi ) have bounded range, to guarantee that all expectations in the following exist.

(a) Assume that h(xi ) is known. In that case we know that the asymptotically efficient
estimator for β (i.e. the estimator with smallest asymptotic variance within a certain
class of estimators) is the weighted least squares estimator β̂WLS , using the inverse
of h(xi ) as weights. Write down β̂WLS and show that it is a consistent estimator for
β.
√
(b) Show that as n → ∞, n(βbWLS − β) ⇒ N (0, V ) and provide an expression for V
as a function of h(xi ).

(c) We now relax the assumption that h(xi ) is known. Assume instead that h(xi ) =
3 + γx2i , where γ > 0 is an unknown scalar constant. Provide a consistent estimator
for γ (just write down the estimator, no proof of consistency required). Under this
assumption on h(xi ), write down a consistent and efficient estimator for β, denoted
by β̂ ∗ .
√
(d) Show that as n → ∞, n(βb∗ − β) ⇒ N (0, V ∗ ). Derive an expression for the
asymptotic variance V ∗ as a function of γ. Using the estimator for γ derived in part
(c), provide an estimator for V ∗ .

(e) Assume that we know ui | xi ∼ t(5), that is, conditional on xi , ui follows a t-

distribution with five degrees of freedom. In this case, is it still true that the WLS
estimator in (a) is efficient? Explain your answer, no proof required.

4
Section B
Answer TWO questions from this Section.

Question B1 (20 points)

Suppose that yi takes values 0 and 1 and that
1 1
P (yi = 0 | θ) = , P (yi = 1 | θ) = ,
1 + eθ 1 + e−θ
where θ ∈ R is a scalar parameter. Note that P (yi = 0 | θ) + P (yi = 1 | θ) = 1, because
1 eθ
1+e−θ
= 1+eθ
(just multiply numerator and denominator by eθ ). We want to do inference
on θ from an iid sample y1 , y2 , . . . , yn with sample size n.

1
Qn
(a) Write down the log-likelihood function Qn (θ) = n
log i=1 f (yi |θ) for this model,
where f (0|θ) = P (yi = 0 | θ) and f (1|θ) = P (yi = 1 | θ). Show that
n
dQn (θ) 1X
= a(θ) + yi ,
dθ n i=1

and find the function a(θ), which only depends on θ and not on y = (y1 , . . . , yn ).

(b) Calculate the maximum likelihood estimator (MLE) θb by solving the corresponding
first order condition. We do not need to check the second order condition, because
Qn (θ) is strictly concave in θ.

E
i h
d2 log f (yi |θ)
(c) Calculate the expected Hessian , evaluated at the true θ. From
dθ2
√
our general theory of the MLE we know that as n → ∞ we have n(θb − θ) ⇒
√ b
N (0, AsyVar( nθ)). Use your result on the expected Hessian to calculate the
√ b
asymptotic variance AsyVar( nθ), relying on general results from the lecture.

1
Pn
(d) Suppose you observe a sample with size n = 20 and mean n i=1 yi = 0.4. Use
this information and your result in (b) and (c) to calculate the maximum likelihood
estimator θb and an estimator for the standard error of θ.
b

i =1 | θ)
(e) We define the odds ratio parameter as λ = PP (y
(yi =0 | θ)
. Show that λ = eθ . Given the
MLE estimate for θ,
b provide a consistent estimator λ b for λ. Show that as n → ∞,
√ b
n(λ − λ) ⇒ N (0, V ) and provide an expression for V as a function of θ.

5
Question B2 (20 points)
Assume that yi ∼ N (0, σ 2 ). We observe an iid sample y1 , y2 , . . . , yn . We want to do
inference on the parameter σ > 0.

(a) Use Eyi = 0 and Var(yi ) = σ 2 to write down a moment condition for σ > 0. Provide
a formula for the corresponding Method of Moment (MM) estimator as a function
of the observed sample y1 , y2 , . . . , yn . Calculate the asymptotic variance-covariance
matrix of this estimator.

From our general GMM theory we know that for a vector of moment functions g(yi , σ)
that satisfies E[g(yi, σ)] = 0 at the true parameter σ, the corresponding GMM estimator
with optimal weight matrix W ∗ has asymptotic variance-covariance matrix
√ −1
AsyVar( n σbGMM ) = {G0 W ∗ G} ,

where G = E[ dg(ydσi ,σ) ]. In this special case where σ is only a scalar parameter G is simply
a vector. You can use this general result on the asymptotic variance without proof, and
you can assume that all required regularity conditions are satisfied.

(b) In addition to the moment condition in (a), use the moment condition E(yi4 ) = 3σ 4 .
Write down g(yi , σ) that corresponds to these moment conditions. Write down the
corresponding GMM objective function and explain how the GMM estimator can
be obtained from that objective function. You do not need to solve for the GMM
estimator.

(c) What is the optimal choice of weight matrix, W ∗ , that minimizes the asymptotic
variance-covariance matrix? Calculate the asymptotic variance-covariance matrix of
the GMM estimator in (b) that is obtained when using the optimal weight matrix.

(d) Alternatively, σ can also be estimated by MLE using the assumption that yi ∼
N (0, σ 2 ). The contribution to the likelihood in this case is given by f (yi |σ) =
i y2
√ 1 e− 2σ2 . Using results from the lecture, derive the asymptotic variance for the
2πσ 2

MLE estimator for σ

b.

(e) Compare the asymptotic variance-covariance matrix of the MLE in (d) with that
of GMM obtained in part (c) and the one obtained from the method of moments
in part (a). Is it possible to add moment conditions to GMM to improve on its
efficiency? Explain your answer.

6
Question B3 (20 points)
Consider the MA(1) model

yt = εt + θεt−1 , εt ∼ iid N (0, σ 2 ),

where θ ∈ R and σ > 0 are the parameters of the model.

(a) Calculate the variance γ0 = E(yt2) and the first and second order autocovariances
γ1 = E(yt yt−1 ) and γ2 = E(yt yt−2 ).

(b) For an observed sample y1 , y2 , . . . , yT one calculates the sample moments

T T −1
1X 2 1 X
y = 45, yt yt−1 = 18.
T t=1 t T − 1 t=1

Use these sample moments and your result in (a) to calculate the method of moments
estimators of θ and σ. There may be multiple solutions to this, calculate all of them.

(c) Are the parameters θ and σ uniquely identified from these two moments conditions?
Provide a restriction on the support of θ so that the parameters of the model are
uniquely identified. Without this restriction, is it possible to use the additional
moment conditions provided by higher order autocovariances γj = E(yt yt−j ), j =
2, 3, 4, . . ., to identify the parameters of the model uniquely?

(d) One now suspects that the true model is an AR(1) model

yt = ρyt−1 + εt , εt ∼ iid N (0, σ 2 )

with |ρ| < 1. Calculate γ2 = E(yt yt−2) as a function of γ0 = E(yt2) in this case.
Let β̂ be the OLS estimator obtained from regressing yt on yt−2 . What is the
probability limit of β̂ as T → ∞? Explain how you can use this information to
distinguish between the MA(1) and the AR(1) model in the observed sample? (no
proof required)