Chapter 8: Introduction to Panel Data

Set of Exercises
Exercise 1. Variance of the within group estimator
As in class, let us consider the model:
yi = Xi β + ηi ι + vi
where yi is T x 1. We assume:
E(vi |Xi , ηi ) = 0
Let:
!−1
N N
P 0 P 0
β̂W G = (QXi ) QXi (QXi ) Qyi
i=1 i=1
0
be the within-group estimator of β, where Q = IT − T1 ιι .
We assume that observations are iid across individuals.

1. Let
0
Σ =Var((QXi ) Qyi )

Give the expression of the limit in probability of N Var(β̂W G ), when T is kept fixed, as
a function of Σ, Q and Xi .

2. Assume in this section that

Var(vi |Xi , ηi ) = σ 2 IT

What does the previous expression become in this particular case?

3. Propose an estimator of Σ, robust to heteroskedasticity.

4. Derive from the previous questions a robust estimator of Var(β̂W G ).

Exercise 2. The Incidental Parameter Problem

Let us consider the model:
yi = Xi β + ηi ι + vi
0 0 0
where yi = (yi1 , ..., yiT ) , vi = (vi1 , ..., viT ) and ι = (1, ..., 1) are T x 1 vectors, and Xi =
0
(xi1 , ..., xiT ) is a T x K matrix.
We assume that:
vi |Xi , ηi ∼ N (0, σ 2 IT )

1
 1. Derive the joint log-likelihood of the observations: ln L(β̃, σ̃ 2 , η̃1 , ..., η̃N ).

Then we maximize the log-likelihood with respect to β̃, σ̃ 2 , η̃1 , ..., η̃N . Let β̂, σ̂ 2 ,
η̂1 , ..., η̂N be the MLE of these parameters.

2. Let
η̂i (β̃) = arg max ln L(β̃, σ̃ 2 , η̃1 , ..., η̃N )
η̃i

Show that η̂i is well defined, and prove that

η̂i (β̃) = ȳi − X̄i β̃
T T
1
P 1
P 0
where ȳi = T
yit , and X̄i = T
xit .
t=1 t=1

3. Remarking that
β̂ = arg max ln L(β̃, σ̃ 2 , η̃1 (β̃), ..., η̃N (β̃))
β̃

show that
β̂ = β̂W G ,

the within group estimator of β.

4. Show that:
!0 !
N
σ̂ 2 = N1T
P
yi − ȳi ι − (Xi − ιX̄i )β̂W G yi − ȳi ι − (Xi − ιX̄i )β̂W G
i=1

5. Interpreting η̂1 , ..., η̂N and β̂W as ordinary least squares estimators, give the value
of
!0 !
N
plim 1 P
yi − ȳi ι − (Xi − ιX̄i )β̂W G yi − ȳi ι − (Xi − ιX̄i )β̂W G
N →∞ N T − N − K i=1

6. For fixed T and N goes to infinity, prove:

plim σ̂ 2 = T T− 1 σ 2
N →∞

Conclude.

Exercise 3. Reform evaluation when there are few reforms

In this exercise, we want to measure the impact of a change in abortion laws on abortion
rates. We have annual data on the 50 US states on a 6-years period. At the beginning of
the period, abortion was prohibited in all the 50 states. Over the period, 3 states changed
that law (only once) and legalized abortion.
To analyze this problem, we consider the following model:

2
 yit = βdit + ηi + uit
where dit is a dummy variable that takes the value of one if abortion is legal in state i at
time t (zero if not), and yit is the abortion rate in state i at time t. ηi is a state fixed effect.
Lastly, uit are iid, and satisfy E(uit |ηi , di1 , ..., diT ) = 0.
We call N the number of groups and T the number of time periods. We also denote as N0
the number of groups for which abortion was legalized at one point, and order these states
such that states i > N0 never legalized abortion during the period. In our data, N is 50, T
is 6 and N0 is 3. Lastly, for each state which legalized abortion at one point, we denote as
Ti the date at which legalization took place. In those states, dit is zero if t < Ti and dit = 1
if t ≥ Ti .
1. Define the following operator .̃ which associates, to every random variable wit , the
following random variable:
T
w̃it = wit − T1
P
wit
t=1

Show that
ỹit = β d˜it + ũit
2. Show that:
N P
T N0
(Ti − 1)(T − Ti + 1)
(d˜it )2 =
P P
i=1 t=1 i=1
T

3. Let β̂ be the linear projection of ỹit on d˜it , without constant. We assume that Ti ,
i = 1, ..., N , are iid and satisfy E((Ti − 1)(T − Ti + 1)) 6= 0. Is β̂ consistent when N0
tends to infinity, for fixed T ?
4. Compute plim β̂ when T and N0 are kept fixed.
N →∞

5. Is β̂ consistent in the case when T and N0 are kept fixed and N tends to infinity? Is
there a contradiction with the results seen in class?
Exercise 4. Dynamic Panel Data
Consider the first-order autoregressive model with individual and time effects of the form:
yit − µi − δt = α (yi,t−1 − µi − δt−1
) + vit
E vit |yi0 , . . . , yi(t−1) , δ0 , . . . , δt , µi = 0
Suppose that in fact T = 2, so that for each individual we observe yi0 , yi1 , yi2 .
1. Obtain the within-groups estimate of α and discuss its properties.

2. Derive a consistent estimator of α for large N . How would your answer be modified if
T > 2?

3. Discuss the costs and benefits of assuming that the initial observations satisfy
E (yi0 |µi , δ0 ) = µi + δ0
How would you test this assumption?