Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

EC484 CLASSES: WEEK 4

KAMILA NOWAKOWICZ

Question 2, Problem Set 2

The setting is a classic errors-in-variables model. But really this is another practice on how
you derive probability limits for certain averages.
In this question zi⇤ and yi⇤ are the true values. Instead of them we observe mismeasured
zi = zi⇤ + "i and yi = yi⇤ + i . The true model is given in terms of the true (unobserved) values:

yi⇤ = ↵ + zi⇤ .

We are told that zi⇤ , i and "i are mutually independent sequences of independent random
variables, zi⇤2 , 2
i and "2i are uniformly integrable, and E( i ) = E("i ) = 0. These are very strong
assumptions on the nature of the measurement error, but even in this case we expect to see
inconsistency. In reality, you may expect that the measurement error to not be independent
and random, which would make things even worse.

Useful results: We start with listing results which will be useful in solving this question. Like
last week, we will use convergence in r-th mean to show convergence in probability:

• Conditions for convergence in second mean:


✓ˆ converges to ✓ in second mean if E[✓ˆ ✓]2 ! 0
• Conditions for convergence in first mean:
– ✓ˆ converges to ✓ in first mean, if E ✓ˆ ✓ ! 0.
– Alternatively, we can check the conditions of Theorem 11: for a sequence of
independent and uniformly integrable variables {xi }i2N :
n
1X 1st
(xi E(xi )) ! 0
n i=1
• Both of the above imply convergence in probability.

We are given information about squares of some variables being uniformly integrable. The
following will let us conclude that the variables themselves also have that property:

These solutions are adapted from solutions by Chen Qiu, which were based on Prof Hidalgo’s notes and
solutions for EC484. Their aim is to fill some gaps between notes and exercises. Prof Hidalgo’s notes should
always be the final reference. If you spot any mistakes in this file please contact me: K.Nowakowicz@lse.ac.uk.
1
Result: If {Xis } is U.I. for some s > 0 then also {Xir } is U.I. for any r < s.
Proof: By Jensen’s inequality:
r/s
sup E (|Xi |r 1(|Xi | > ))  sup E |Xi |s 1(|Xi | > )s/r
i i
✓ ◆r/s
= sup E (|Xi | 1(|Xi | > ))
s
!0
i

as ! 1.

We are given that zi⇤2 , 2


i and "2i are uniformly integrable, and by the above result we can
see that zi⇤ , i and "i must be uniformly integrable too. They are also independent, so we can
apply Theorem 11 to each of them to get that:

n
1X ⇤ p
(z E(zi⇤ )) ! 0
n i=1 i
n
1X p
i !0
n i=1
n
1X p
"i ! 0
n i=1

Another useful property is:

Theorem 10: If a sequence {xi }i2N is uniformly integrable then maxi E(|xi |) < 1.

Solution. We begin by rewriting the true (but unobservable) model in terms of the observable
variables:

yi⇤ = ↵ + zi⇤
yi i = ↵ + (zi "i )
yi = ↵ + (zi "i ) + i

= ↵ + zi + "
|i {z }i
ui

We denote the composite error i "i by ui to simplify notation. This way we get a model
which looks like a standard linear regression with a single regressor. Notice that zi and ui both
depend on "i : the regressor is endogenous, which is why we expect to see inconsistency. The
2
OLS estimator for is:

P
n
(zi z̄)(yi ȳ)
ˆ= i=1
P
n
(zi z̄)2
i=1

1
P
n
n
(zi z̄)ui
i=1
= + Pn
1
n
(zi z̄)2
i=1

and the estimator for ↵ is:


ˆ = ȳ ˆz̄

= ↵ + z̄ + ū ˆz̄
⇣ ⌘
=↵+ ˆ z̄ + ū

To get the probability limits of both coefficients we take advantage of Slutzky’s Theorem and
split them into smaller building blocks which we consider separately. We now need to find the
probability limits of the following four objects:
n
1X
A: (zi z̄)ui
n i=1
n
1X
B: (zi z̄)2
n i=1
n
1X
C : z̄ = zi
n i=1
n
1X
D : ū = ui
n i=1

C and D look simpler and will be useful in finding the limits of A and B. Start from easier
ones.
Step 1: The probability limit of D. Rewrite ui in terms of the random variables we have
explicit assumptions for:
n
1X
ū = ui
n i=1
n
1X
= ( i "i )
n i=1
n n
1X 1X
= i "i .
n i=1 n i=1
3
By Theorem 11,

n
1X p
i ! 0;
n i=1
n
1X p
"i ! 0.
n i=1

p
So D ! 0 by Slutzky’s Theorem.
Step 2: The probability limit of C. Again, we start by rewriting in terms of the original
variables:

n
1X
z̄ = zi
n i=1
n n
1X ⇤ 1X
= z + "i
n i=1 i n i=1

P
n
p P
n
p
As we have shown above (by Theorem 11): 1
n
"i ! 0 and 1
n
(zi⇤ E (zi⇤ )) ! 0.
i=1 i=1
Hence:

n n n
1X ⇤ 1X 1X
z̄ = (zi E(zi⇤ )) + E(zi⇤ ) + "i
n i=1 n i=1 n i=1
| {z } | {z }
p p
!0 !0
n
p 1X
! lim E(zi⇤ ) ⌘ µ⇤
n!1 n
i=1

Where the last line introduces shorter notation.


p
So C ! µ⇤ .
P
n
Step 3: Find the probability limit of A = 1
n
(zi z̄)ui
i=1
Notice

n n n
1X 1X 1X
(zi z̄)ui = z i ui z̄ ui
n i=1 n i=1 n i=1
n n n
1X 1X 1X
= z i ui zi ui
n i=1 n i=1 n i=1
n
1X
= z i ui C·
| {zD}
n i=1 p
!0
4
The limit of the second term follows from what we have already done. It suffices to find
P
n
probability limit of n1 zi ui . By plugging in true model zi = zi⇤ + "i , and ui = i "i :
i=1
n
X n
1 1X ⇤
z i ui = (z + "i ) ( i "i )
n i=1
n i=1 i
n n n n
1X ⇤ 1X ⇤ 1X 1X 2
= z i z i "i + "i i "
n i=1 i n i=1 n i=1 n i=1 i

We need to find a limit of each one of the four averages above.


• By Theorem 11 (given that "2i are independent and U.I.):
n n n
1X 2 1X 2 1X
"i = " E("2i ) + E "2i
n i=1 n i=1 i n i=1
n
p 1X
! lim E "2i ⌘ 2
"
n!1 n
i=1

Where the last line again introduces simplified notation.


P
n
• For n1 (zi⇤ i ) we don’t have explicit assumptions so it should be easier to show
i=1
convergence in 2nd mean.
0 !2 1
Xn n n n
@ 1 ⇤ A 1 X ⇤2 2 1 XX
E zi i = 2 E zi i + 2 E zi⇤ zj⇤ i j
n i=1 n i=1 | {z } n i=1 j=1 | {z }
indep. indep.
= E ( zi ) E ( i )
⇤2 2
i6=j = E(zi )E(zj )E( i )E(
⇤ ⇤
j )=0
=0

n
1 X
= 2 E zi⇤2 E 2
i
n i=1
n
1 X
 2 max E zj⇤2 max E 2
k
n i=1 j k
| {z } | {z }
<1 <1

1
= max E zj⇤2 max E 2
k
n j k
| {z }
=D<1
D
!0 =
n
as n ! 1. Note that the maxima of the expectations are bounded by Theorem
10 (this is one of the necessary conditions for U.I. sequences).
P
n
2nd P
n
p
So, n1 zi⇤ i ! E(zi⇤ i ) = E(zi⇤ )E( i ) = 0 and hence n1 zi⇤ i ! 0.
i=1 i=1
P
n
p P
n
p
• By identical arguments, 1
n
zi⇤ "i ! 0, and 1
n
"i i ! 0.
i=1 i=1
• Summarising:
p 2 2
A ! (0 ⇥0+0 ") µ⇤ ⇥ 0 = ".

P
n
Step 4: Find the probability limit of B = 1
n
(zi z̄)2 .
i=1
5
n n
1X 1X 2
2
(zi z̄) = z z̄ 2
n i=1 n i=1 i p
!µ2⇤

p p
We have previously shown that z̄ ! µ⇤ , by Slutzky’s Theorem we get z̄ 2 ! µ2⇤ . For the
other term:
n n n n
1X 2 1 X ⇤2 2 X ⇤ 1X 2
zi = zi + z i "i + ".
n i=1 n i=1 n i=1 n i=1 i
Using our previously obtained results:
n
1X ⇤ p
z "i ! 0
n i=1 i
n n
1X 2 p 1X
"i ! lim E "2i = 2
"
n i=1 n!1 n i=1

and by Theorem 11:


n n
1 X ⇤2 p 1X
zi ! lim E zi⇤2 = s2⇤
n i=1 n!1 n
i=1

Summarising:
p
B ! s2⇤ + 2 ⇥ 0 + 2
" µ2⇤ = s2⇤ + 2
" µ2⇤

Step 5: Collecting all from Steps 1-4, we have:


2
p
ˆ! + "
s2⇤ + "2 µ2⇤
⇣ ⌘

ˆ =↵+ ˆ z̄ + ū
2
p " µ⇤
!↵+
s2⇤ + 2
" µ2⇤
To conclude, we have consistency in the following cases:
p
• ˆ! iff: s2⇤ + 2
" µ2⇤ 6= 0 and = 0 or 2
" = 0;
p
ˆ ! ↵ iff:
• ↵ s2⇤ + 2
" µ2⇤ 6= 0 and = 0 or 2
" = 0 or µ⇤ = 0.
These conditions correspond to degenerate, unlikely situations. Either zi⇤ doesn’t enter the
model or is measured without error.

Question 3, Problem Set 2

In this question we are asked to show convergence almost surely:

• Xn converges to X almost surely or with probability 1 if

lim P r (|Xm X| > for some m n) = 0.


n!1

6
The definition of almost surely convergence is a bit awkward. A common approach to show-
ing convergence almost surely is to use a stronger notion of convergence. Since complete
convergence ) almost sure convergence, we can address this question by showing convergence
completely.

• Xn converges to X completely if:


X1
P r (|Xn X| > ) < 1.
n=1

The advantage of using convergence completely is that it involves probability terms that can
be bounded by expectations, and expectations are much easier for us to calculate.

P1 ⇣ ⌘
Solution. We show Pr ˆn
< 1, thus verify ˆn converging to
> completely
n=1
⇣ ⌘
and therefore, with probability 1 as well. We aim to bound terms P r ˆn > by
expectation terms through Markov inequality. Since:

ˆn 1
= (Z 0 Z) Z 0U
!
⇣ ⌘ 1 1X
n
= M̂ z i ui
n i=1

By Markov inequality:
✓ ◆
4
⇣ ⌘ E ˆn
Pr ˆn >  4

✓ ◆
4
We calculate E ˆn :

0 1 0 1
4 4
⇣ ⌘ 1 1 Xn ⇣ ⌘ 1 4 1 Xn
E@ M̂ zi ui A  E @ M̂ zi ui A by Cauchy-Schwarz inequality
n i=1 n i=1
0 1
4
⇣ ⌘ 1 4
1X
n
= M̂ E@ z i ui A
n i=1

= AB

⇣ ⌘ 1 4 ⇣ Pn ⌘
4
where A = M̂ ,B=E 1
n i=1 zi ui .

(1) A is innocuous, by assumption M̂ ! M > 0, so A ! c for some positive constant c.


7
p
(2) B is the key term that needs to be bounded. By definition of vector norm kak = a0 a:
0 1
n 4 n 2 2
1 X 1 X
z i ui = @ z i ui A
n i=1 n i=1

n
!0 n
!!2
1 X X
= 4 z i ui z i ui
n i=1 i=1
n X
n
!2
1 X
= zi0 zj ui uj
n4 i=1 j=1
n n n n
1 XXXX 0 0
= z z j z z l u i uj uk ul ,
n4 i=1 j=1 k=1 l=1 i k

which means
0 1
n 4 n
1 X 1 X
E@ z i ui A = 4 E (zi0 zj zk0 zl ui uj uk ul )
n i=1 n i,j,k,l=1
n
1 X 0 0
= 4 z zj z zl E (ui uj uk ul ) .
n i,j,k,l=1 i k

Since ui are independent with mean 0, if any index appears only once we can pull out
E(ui ) = 0 and the whole term becomes zero. The only situations where the last term is
not zero are:
• i = j = k = l,
• i = j 6= k = l,
• i = k 6= j = l,
• i = l 6= j = k.
In the first case we get E(u4i ), in the remaining ones we get a term of the form E(u2i )E(u2j ).
Hence:
0 1
n 4 n n n
1 X 1 X 0 0 3 XX 0 0
E @ z i ui A = 4 z z i z z i E ui + 4
4
z zi z zj E u2i E u2j
n i=1 n i=1 i i n i=1 j=1 i j
n n
!2
1 X 3 X
= 4 kzi k4 E u4i + 4 2
kzi k E u2i
n i=1 n i=1
n ✓ ◆4 n ✓ ◆2 q !2
1 X 3 X
 4 sup kzj k sup E(uk )4 + 4 sup kzj k E(u4i )
n j k n j | {z }
i=1 i=1 p
 supk E(u4k )
p
where E (u2i )  E (u4i ) by Jensen’s inequality. Let

sup kzj k  A < 1


j

sup E(u4k )  B < 1


k
8
Note that these are finite by assumptions given in the question. Then:
0 1 !2
n 4 n n
1 X 1 X 3 X p
E@ z i ui A  4 A4 B + 4 A2 B
n i=1 n i=1 n i=1

A4 B 3A4 B
= +
n3 n2
A B 3A4 B
4
 +
n2 n2
4
4A B
=
n2
C
=
n2
For some constant C.
(3) Combining the above results:
✓ ◆
ˆn
4 C C
E  M̂ 1
=
n2 n2
possibly for some other constant C. Finally:
1
X ⇣ ⌘ X 1 ✓ ◆
ˆ 1 ˆn
4
Pr n >  4
E
n=1 n=1
1
CX 1
 4 2
<1
n
|n=1{z }
2
= ⇡6 <1

This proves complete convergence of ˆn to . Almost sure convergence fallow as a


consequence.

You might also like