Advanced Econometrics I

Jürgen Meinecke
Lecture 9 of 12
Research School of Economics, Australian National University

1 / 27
Roadmap

Instrumental Variables Estimation

Weak Instruments
Strategies for Dealing with Weak Instruments
Simulation Results

2 / 27
We have seen that π = 0 (invalid instruments) leads to a
breakdown of statistical inference for the IV estimator
Now let’s look at: π 6= 0 but π ≈ 0
What I’m trying to say here:
π is not equal to zero but it is close to zero or local to zero
We will use the same setup as in the invalid instrument case
(one endogenous regressor and one instrument)
Technically, local to zero is generated by letting π = N −1/2 τ
where τ 6= 0
Where does this come from? You could guess that, once you
plug this into an asymptotic expansion, it delivers a useful rate
of convergence

3 / 27
Reminder of the setup
Yi = Xi β + ei
Xi = Zi π + vi

In other words: K1 = 0, K2 = L2 = L = 1
Recall that π = E(Xi Zi )/E(Z2i )
What happens when E(Xi Zi ) ≈ 0 so that π ≈ 0?
Let’s label this case weak instrument
To make life easy,
! let’s
! assume !
ei 1 ρ
Var |Zi =
vi ρ 1

as well as EZi = 0 and EZ2i = 1

Endogeneity, of course, implies ρ 6= 0
4 / 27
Let’s again first look at the OLS estimator
∑N 1 Xi ei
β̂OLS − β = i=
∑N 2
i = 1 Xi
N −1 ∑ Ni=1 (N
−1/2 τZ + v )e
i i i
= N
N ∑ i=1 (N
− 1 − 1/2 τZi + vi )2
p E(vi ei )
→ = ρ 6= 0
E(v2i )

which is the same as before when π = 0

Let’s turn to the IV estimator, remember
sZY ∑Ni = 1 Zi Yi β ∑N N
i=1 Zi Xi + ∑i=1 Zi ei
β̂IV = = N =
sZX ∑ i = 1 Zi Xi ∑N
i=1 Zi Xi
∑Ni=1 Zi ei
= β+ N
∑i=1 Zi Xi

5 / 27
We start by looking at
1 N 1 N
1 N
√ ∑ Zi Xi = √ ∑ Z2i π + √N ∑ Zi vi
N i=1 N i=1 i=1
N N
1 1
=
N ∑ Z2i τ + √N ∑ Zi vi
i=1 i=1
d
→ τ + ξ2

2
because we assume ! EZi = 1!for simplicity, then
!!recall
N
1 Z i ei ξ1 1 ρ
√ ∑
d
→ ∼ N 0, , therefore
N i=1 Z v
i i ξ 2 ρ 1

√1 ∑N Zi ei
N i=1 d ξ1
β̂IV − β = →
√1 ∑N Zi Xi τ + ξ2
N i=1

Again: β̂IV is inconsistent with non-normal asymptotic

distribution
6 / 27
What happens to the t test based on β̂IV under weak
identification?
Recall the generic t statistic that is based on an estimator β̂:
β̂ − β
tβ̂ ( β) =
se( β̂)

Let’s make our lives easy and consider the standard error of β̂IV
under homoskedasticity
The estimator of the asymptotic variance for β̂IV is
  ∑ N Z2
Var β̂IV |Zi = σ̂e2 N i=1 i
( ∑ i = 1 Xi Zi ) 2
therefore q
∑N 2
i=1 Zi
se( β̂IV ) = σ̂e
∑N
i=1 Xi Zi

7 / 27
Notice
N N  2
σ̂e2 = N −1 ∑ (Yi − Xi β̂IV )2 = N −1 ∑ Xi ( β − β̂IV ) + ei
i=1 i=1
N N N
= N −1 ∑ e2i − 2N −1 ∑ Xi ei ( β̂IV − β) + N −1 ∑ Xi2 ( β̂IV − β)2
i=1 i=1 i=1
 2
d ξ1 ξ1
→ 1 − 2ρ +
τ + ξ2 τ + ξ2

It follows that
∑Ni=1 Zi ei
β̂IV − β ∑N
√1 ∑N
i = 1 Z i ei
i=1 Xi Zi N
tβ̂IV ( β) = = √ = q
se( β̂IV ) σ̂e ∑Ni=1 Zi
2
1
∑N 2
N
σ̂e N i=1 Zi
∑i=1 Xi Zi
√1
N
∑N
i=1 Zi ei d ξ1
= q →r 2 =: S(ρ, τ )
1 + op (1)

σ̂e 1 − 2ρ τ +ξ 1ξ 2 + ξ1
τ +ξ 2

8 / 27
Copy and paste last line from previous slide:
d ξ1
tβ̂IV ( β) → r  2 =: S(ρ, τ )
ξ1 ξ1
1 − 2ρ τ +ξ 2 + τ +ξ 2

What does this mean?

The t statistic does NOT converge to a normal distribution
So we can’t simply compare it to the ±1.96 cutoffs
The asymptotic distribution of t depends on ρ and τ, two
parameters that we don’t know and cannot estimate

• ρ is the degree of endogeneity

• τ is the strength of the instrument

To get more intuition about what’s going on, let’s set ρ = 1

which is the worst possible case of endogeneity
9 / 27
Then ξ 1 = ξ 2 and the t statistic collapses to
ξ2
S(1, τ ) = ξ 1 + 1 ,
τ

Recall that ξ 1 ∼ N (0, 1) and ξ 12 ∼ χ21

So S(1, τ ) is a mixture of a N (0, 1) and a χ21 distribution
The degree of the mixture is controlled by the value of τ

• if τ is very large, then S(1, τ ) will be close to N (0, 1)

(strong instrument case)
• if τ is very small, then the χ21 dominates and distorts away
from normality (weak instrument case)
• in the extreme we get limτ →0 S(1, τ ) = ∞
(that’s a terrible result: very weak instruments will yield
misleadingly large t statistics suggesting significant β
regardless of the truth)
10 / 27
Roadmap

Instrumental Variables Estimation

Weak Instruments
Strategies for Dealing with Weak Instruments
Simulation Results

11 / 27
What are we to do?
In some undergraduate textbooks, the so-called rule of thumb is
often recommended
That rule originated from Staiger and Stock (1997)
Their main goal was to come up with a quick and easy and
robust way for practitioners to rule out the weak IV case
Staiger and Stock had the idea to propose a rule that is solely
based on the first stage F statistic
The first stage F statistic is the F statistic in the regression of the
endogenous regressor(s) on the instruments

12 / 27
The rule of thumb can be viewed as a decision rule:

• if first stage F > 10, then decide that instruments are strong
• if first stage F < 10, then decide that instruments are weak

If you decide that your instruments are strong, then β̂IV and
β̂2SLS can be used safely and statistical inference based on
asymptotic normality is fine
The rule of thumb turns out to be a bit crude
Stock and Yogo (2005) offer a somewhat more sophisticated
way of testing for weak IV
While more complex than the rule of thumb, their
recommendation is still quite easy to implement
Their test is still entirely based on the first stage F statistic

13 / 27
Stock and Yogo’s idea is the following:
Unless you’ve got very strong instruments, S(ρ, τ ) will be
non-normal
Statistical inference based on S(ρ, τ ) will be misleading
In particular, the actual size of the test will differ from the
nominal size
Recall: the nominal size of a well-behaved statistical test is the
probability to reject the null when it’s true
By trusting the normal approximation and setting the critical
value to 1.96 you are fixing the nominal size at 5%
Here we’re not dealing with a well-behaved statistical test, so
we admit that we may have to accept a higher actual size

14 / 27
Stock and Yogo’s idea is to pin down a worst case scenario for
the actual size and provide the corresponding critical values for
the first stage F statistic
It’s best to look at an example

• You understand that an actual size of 5% is unrealistic

• Instead you are willing to live with an actual size of 20%
• This is your worst case scenario; the largest actual size that
you are willing to tolerate
• For the case of one endogenous regressor and one
instrument, Stock and Yogo suggest to use the critical
value 6.66
• This means that your F-statistic in the first stage regression
of Xi on Zi must exceed 6.66 to achieve an actual size that
is at most 20%
15 / 27
Excerpt from Stock and Yogo (2005),
“Testing for Weak Instruments in Linear IV Regression”

Note: K2 : number of IV; n: number of endogenous variables

16 / 27
Stock and Yogo suggest to use their weak instrument test as a
decision rule:

• if the first stage F-statistic lower than the critical value,

conclude that the instruments are weak
• otherwise conclude that they are strong

What to do when you conclude that instruments are weak?

Lot’s of research still underway, but options include:
Anderson-Rubin test, Kleibergen’s K-statistic, Moreira’s
conditional likelihood test
Depending on your application the Staiger and Stock (1997)
rule of thumb that sets the critical value simply to 10 is either too
restrictive or too relaxed
If you want your actual size to be at most 10%, then the rule of
thumb is too relaxed
17 / 27
Roadmap

Instrumental Variables Estimation

Weak Instruments
Strategies for Dealing with Weak Instruments
Simulation Results

18 / 27
On the next 8 slides I will show you

• simulated distributions
• powerfunctions

for two different strengths of instruments

(motivated by Stock and Yogo’s table)

• weak instrument: F = 5.53

• strong instrument: F = 16.38

for two different degrees of endogeneity

• low degree of endogeneity ρ = 0.10

• high degree of endogeneity ρ = 0.90

These are produced by the week 9 Julia notebook

19 / 27
Strong instrument, low degree of endogeneity
well behaved distribution

20 / 27
Strong instrument, low degree of endogeneity
nice looking power function

21 / 27
Strong instrument, high degree of endogeneity
Can you spot the distortion and the resulting leftward bias?

22 / 27
Strong instrument, high degree of endogeneity

23 / 27
Weak instrument, low degree of endogeneity
Looks symmetric, but awful variance

24 / 27
Weak instrument, low degree of endogeneity
The probability to reject far away from zero is terribly low

25 / 27
Weak instrument, high degree of endogeneity
Not symmetric, and awful variance

26 / 27
Weak instrument, high degree of endogeneity
Things only get worse here

27 / 27