RS – Lecture 8

Lecture 8
Instrumental Variables

CLM: New Assumptions

• Last lecture, we presented a new set of assumptions for the CLM:
(A1) DGP: y = X  + .
(A2’) X stochastic, but E[X’ ]= 0 and E[ε]=0.
(A3) Var[|X] = σ2 IT
(A4’) plim (X’X/T) = Q (p.d. matrix with finite elements, rank= k)

• We studied the large sample properties of OLS:

- b and s2 are consistent
a
-b   N(β, (σ2/T) Q-1)
- t-tests asymptotically N(0,1), Wald tests asymptotically 2rank(ST) and
F-tests asymptotically 2rank(Var[m]) .
- Small sample behavior may be understood by simulations and/or
bootstrapping.

1
RS – Lecture 8

The IV Problem
• We start with our CLM:
y = X + . (DGP)
- Let's pre-multiplying the DGP by X'
X' y = X' X + X' .
- We can interpret b as the solution obtained by first approximating
X' by zero, and then solving the k equations in k unknowns
X'y = X'X b (normal equations).

Note: What makes b consistent when X'/T  p

0 is that
approximating (X'/T ) by 0 is reasonably accurate in large samples.

• Now, we challenge this approximation. We relax the assumption that

{xi,εi} is a sequence of independent observations. That is,
plim (X’/T) ≠ 0.  This is the IV Problem!

The IV Problem: OLS is Inconsistent

• A correlation between X &  is not rare in economics, especially in
corporate finance, where endogeneity is pervasive.

• Endogenous in econometrics: A variable is correlated with the error term.

• Q: What is the implication of the violation of plim(X’/T) = 0?

From the asymptotic CLM version, we keep (A1), (A3), and (A4’):
(A1) y = X + .
(A3) Var[|X] = σ2 IT
(A4’) plim (X’X/T) = Q

• Now, we assume (A2’’) plim(X’/T) ≠ 0.

• Then, plim b = plim  + plim (X’X/T)-1 plim (X/T)

2
RS – Lecture 8

The IV Problem: OLS is Inconsistent

• plim b = plim  + plim (X’X/T)-1 plim (X/T)
=  + Q-1 plim (X/T) ≠ 

Under the new assumption, b is not a consistent estimator of .

Note: For finite samples, we could have challenged assumption (A2)

E[|X] = 0. Then, Cov(X,) ≠ 0  E[b|X] ≠ .

• Diagram with Cov(X,) ≠ 0

X y

The IV Problem: Structural Model

• y and X are both endogenous. Suppose, we also model X as a
function of some exogenous variable Z. Then, the model becomes a
structural model (everything is modeled):
y = X + 
X=Z +V
where V &  are correlated.

The researcher is not interested in estimating the whole structural

model, it is interested on the first equation: the impact of X on y.

Now, we can rewrite the inconsistency as

plim b =  + Q-1 plim ((  Z+ V)/T)
=  + Q-1  plim (Z/T) + Q-1 plim (V/T)

 OLS inconsistency depends on relation between Z &  and V & .

3
RS – Lecture 8

The IV Problem: Example 1

• Suppose we want to study the relation between a firm’s CEO’s
compensation (y) and a CEO’s network (x).

Usually, a linear regression model is used, relating y and x, with

additional “control variables” (W) controlling for other features that
make one CEO’s compensation different from another. The term 
represents the effects of individual variation that have not been
controlled for with W or x.

The model is:

y = x + Wγ + 

If a CEO’s network is influenced by the CEO’s natural skills, we have

a problem: y and x are both endogenous –i.e., influenced by the
unobserved CEO’s skills, say S.

The IV Problem: Example 1

• y and x are both influenced by an unobservable variable. Then,
Cov(x, )≠0 (=> by LLN, plim (X’/T) ≠ 0)

• It looks like an omitted variable problem. Assuming linearity, it can be

solved by adding as a control variable “CEO’s skills,” S:
y = x + Wγ + S +
However, S is unobservable.

Note: x is endogenous. It needs a model! Say, it depends on Z:

x=Zπ+v (where σεV measures the endogeneity of x.)

• Recall: Endogeneity occurs when a variable, X, is correlated with .

4
RS – Lecture 8

The IV Problem: Example 2

• Suppose we want to study the effect of military service (x) on
earnings (y). We use a linear model, adding some control variables
(W), controlling for other features that affect y:
y = x  + Wγ + 

 would measure “the causal effect” we would get if x were randomly

assigned. But, there is selection bias by both individuals and military
recruiters.

That is, x is not randomly assigned: Unobserved factors that affect y,

also affect x  Cov(x,) ≠ 0.
.

The IV Problem: Example 3

• In this example, we introduce measurement error in X. That is, DGP:
y = x* +   ~ iid D(0, σε2)
x = x* + u u ~ iid D(0, σu2) -no correlation to 

We are interested in x*, and its marginal effect , but we

observe/measure x, which measures x* with error (u).

All of the CLM assumptions apply. Then,

y = (x – u) +  = x +  – u = x + w
E[x’w] = E[(x*+u)’( – u)] = -σu2 ≠ 0 & plim (X’w/T)≠0
 CLM assumptions violated  OLS inconsistent!

5
RS – Lecture 8

The IV Problem: Example 4

• Simple supply and demand model for some good, where quantity
(Q) and price (P) are endogenous variables –i.e., determined by the
model. In equilibrium QS = Qd = Q. We have a simultaneous equation
model (SEM):
Q   1 P   2Y   d
Q  1 P   S
where Y is income, considered exogenous, and εS and εd are the error
terms.

Suppose we are interested in estimating 1. An OLS regression with

X = P will not work, since Cov(P, εS) ≠ 0.

The IV Problem: Usual Cases

• Q: When might an explanatory variable (a regressor) be correlated

with the error term?
- Omitted variables
- Selection bias
- Measurement error
- Simultaneous equations
- Misspecification
- Correlated shocks across linked equations
- Model has a lagged dependent variable and a serially correlated
error term

6
RS – Lecture 8

Instrumental Variables: New CLM Assumptions

• New Framework:
(A1) DGP: y = X  + .
(A2’’) plim (X’/T) ≠ 0
(A3) Var[|X] = σ2 IT
(A4’) plim (X’X/T) = Q (p.d. matrix, with rank k)
 b is not a consistent estimator of .

• Q: How can we construct a consistent estimator of ?

We will assume that there exists a set of l variables, Z such that
(1) plim (Z’X/T)  0 (relevant condition)
(2) plim (Z’/T) = 0 (valid condition)

• The variables in Z are called instrumental variables (IVs). In general,

not all the X will be correlated with error .

Instrumental Variables: Endogeneity

• We can also write the new framework, emphasizing endogeneity, as:
(A1) DGP: y = Y  + U γ + .
(A2’’) plim (Y’ /T) ≠ 0 (Y: “problem,” endogenous, variables)
(A2’’) plim (U’ /T) = 0 (U: clean variables)
2
(A3) Var[|Y,U] = σ IT
(A4) Y and U have full column rank. Say kx and ku.

• We have Z, a matrix of l “excluded instruments” –the IVs. The IVs

have no impact on y except through Y. We relate Y to Z linearly by:
Y =ZП+UΦ+V – V ~ D(0, σV2 IT)

Note: When the number kx of “endogenous” variables is greater than

one, we have a system of multiple equations. The estimation of this
equation is called “first stage.”

7
RS – Lecture 8

Instrumental Variables: Endogeneity

• Concentrating on two equations (and let X=Y):
(A1) y = X  + U γ +  (called structural equation)
X=ZП+UΦ+V (first stage regression)

Replacing the second equation in (A1):

y = (Z П + U Φ + V)  + U γ +  = Z П  + U φ + ξ
This equation is called reduced form, where
φ=Φ+γ
ξ=V+
Note: Usually, V and  are N(0, σJJ I). But, they can be correlated.

• In this lecture, the parameter of interest is . OLS cannot estimate it.

But OLS works on the reduced form to consistently estimate Г=П.

Instrumental Variables: Notation

• Model:
y=X+Uγ+ - structural equation
X=ZП+UΦ+V - first stage equation
y= +Uγ+ - second stage equation
y =ZГ+Uφ+ξ - reduced form

• Variables
y, X: endogenous variables –i.e., correlated with .
U & Z: exogenous variables –i.e., uncorrelated with .
U: included instruments, clean variables (“controls”)
Z: excluded instruments, IVs –i.e., satisfies the relevant condition and
the valid condition, also referred as exclusion restriction. (Excluded =
not included in the structural equation.)

8
RS – Lecture 8

Instrumental Variables: Notation

• Parameters
: Structural parameter, usually the parameter of interest
П: 1st-stage parameter. It captures the strength of the IV, Z.
If П≈0, not very powerful.
Г: Reduced form parameter. It can show the potential of Z as
instrument.

• Equations
Structural equation: Theory dictates this relation: it relates y and X
(or Y). It measures the causal effect of X on y, ; but the effect is
blurred by endogeneity.
First stage: Regression of X on the instrument, Z (it measures a
causal effect from Z to X).
Reduced form: Regression of y on the instrument is called the
reduced form (it measures the direct causal effect from Z to y).

Instrumental Variables: Assumptions

• New assumption: we have l IVs, Z, such that
plim(Z’X/T)  0 but plim(Z’/T) = 0

• Then, we state assumptions to construct an alternative (to OLS)

consistent estimator of .

Assumptions:
{xi, zi, εi} is a sequence of RVs, with:
E[X’X] = Qxx (pd and finite) (LLN  plim(X’X/T) =Qxx )
E[Z’Z] = Qzz (finite) (LLN  plim(Z’Z/T) =Qzz )
E[Z’X] = Qzx (pd and finite) (LLN  plim(Z’X/T) =Qzx )
E[Z’] = 0 (LLN  plim(Z’/T) = 0)

9
RS – Lecture 8

Instrumental Variables: Estimation

• To construct a new estimator, we start by pre-multiplying the DGP
by W'Z’, where W l×k weighting matrix that we choose:
W'Z’y = W'Z’(X+) = W'Z’X+ W'Z’
 W helps to create a k ×k square matrix, needed for inversion.

• Following the same idea as in OLS, we get a system of equations:

W'Z’X bIV = W'Z’y

• We have two cases where estimation is possible:

- Case 1: l = k -i.e., number of instruments = number of regressors.
- Case 2: l > k -i.e., number of instruments > number of regressors.

The second case is the usual situation. We can throw l-k instruments,
but throwing away information is never optimal.

IV Estimation
• Case 1: l = k -i.e., number of instruments = number of regressors.
To get the IV estimator, we start from the system of equations:
W'Z’X bIV = W'Z’y
- dim(Z)=dim(X): Txk  Z’X is a kxk pd matrix
- In this case, W is irrelevant, say, W=I. Then,
bIV = (Z’X)-1Z’y

Note: Let Z = X. Then,

bIV = b = (X’X)-1Xy

That is, under the usual assumptions, b is an IV

estimator with X as its own instrument.

Sewall G. Wright (1889 – 1988, USA)

10
RS – Lecture 8

IV Estimators: Properties – Consistency

• Properties of bIV
(1) Consistent
bIV = (Z’X)-1Z’y = (Z’X)-1Z’(X+)
= (Z’X/T)-1 (Z’X/T)  + (Z’X/T)-1Z’ε/T
=  + (Z’X/T)-1 Z’ε/T  
p
(under assumptions)

Note: Under the context of Lecture 7 –i.e., (A2’) plim(X’ε/T) = 0–, b

is consistent. But, bIV is also consistent (though, not efficient)!

Remark: When plim(X’ε/T)  0, only the IV estimator is consistent.

Thus, we have an estimator that is consistent when b is not.

IV Estimators: Properties – Asy. Normality

• Properties of bIV
(2) Asymptotic normality
T (bIV – ) = T (Z’X)-1Z’ε
= (Z’X/T)-1 T (Z’ε/T)

Using the Lindberg-Feller CLT

T (Z’ε/T) d
 N(0, σ2Qzz)
Then,
T (bIV - ) d
 N(0, σ2Qzx-1QzzQxz-1)

11
RS – Lecture 8

IV Estimators: Asympotic Var[bIV]

• Properties of ˆ 2, under IV estimation: Consistency
T T
1 1
- We define ˆ 2 : ˆ 2 
T 
i 1
2
e IV 
T  (y
i 1
i  x ' b IV ) 2

where eIV = y – X bIV = y – X(Z'X)-1Z’y = [I – X(Z'X)-1Z']y = Mzx y

= [I – X(Z'X)-1Z'] * (X  + ) = Mzx y
- Then,
ˆ 2 = eIV'eIV/T = 'Mzx'Mzx/T
= '/T – 2 'X (Z'X)-1Z'/T + 'Z (Z'X)-1X'X(Z'X)-1Z'/T

 plim ˆ 2 = plim('/T) – 2 plim[('X/T) (Z'X/T)-1 (Z'/T)] +

+ plim('Z (Z'X)-1X’X(Z'X)-1Z'/T) = σ2

Est Asy. Var[bIV] = E[(Z'X)-1 Z''Z (Z'X)-1] = ̂ 2(Z'X)-1 Z'Z(Z'X)-1

IV Estimators: Example
Simplest case: Linear model, two endogenous variables, one IV.
y1 = y2  +  –  ~ N(0, σεε)
y2 = z π + v – v ~ N(0, σVV)
with reduced form:
y1 = z π  + v  +  = z γ + ξ.
The parameter of interest is  (= γ/π).

∑ , ̅
• We estimate  with IV:
∑ , ̅

Note: With a reasonably large T both numerator and denominator are

well approximated by Normals and if π ≠0, as T gets large, then the
ratio will eventually be well approximated by a normal distribution.

12
RS – Lecture 8

IV Estimators: Example
• To analyze the bias,
bIV = (z' y2)-1 z' y1 =  + (z' y2)-1 z' 
plim(bIV) –  = plim(z’/T)/plim(z’y2/T)  Cov(z, )/Cov(z, y2)
p

• When Cov(z, ) ≠ 0, IV estimation is inconsistent.

• If Cov(z, ) is small, but π≈0, the inconsistency can get large (π≈0)
 Cov(z, y2) = Cov(z, (zπ+v)) = π Var(z) + Cov(z, v) = π Var(z) ≈ 0

• When π = 0  Cov(z, y2) = 0, thus, the IV estimator is not defined.

When π = 0, the instrument provides no information. It is an irrelevant
instrument.

IV Estimators: Example
• When π is small, we say z is a weak instrument. It provides
information, but, as we will see later, not enough.

• Note that even when π=0, in finite samples, the sample analogue to
Cov(z, y2) ≠ 0. Not very useful fact, the sampling variation in
Cov(z,y2) is not helpful to estimate .

Note: The weak instrument literature is concerned with testing H0:=

0 when π is too close to 0. As we will see later, the normal
approximation to the ratio will not be accurate.

13
RS – Lecture 8

IV Estimators: Weak and Strong Instruments

• We assume that there exists a set of l variables, Z such that
(1) plim (Z’X/T)  0 (relevant condition)
(2) plim (Z’/T) = 0 (valid condition –or exclusion restriction)

• We are going to use the variation in Z, which is uncorrelated with ,

to explain the variation of y. Condition (1) allows to do this. Suppose
the relation between Z, X and y is given in the following diagram:

X Y Now, not all the variation in X is

Z used. Only the portion of X
which is “explained” by Z can be
used to explain y.

IV Estimators: Weak and Strong Instruments

X Y • Best situation : A lot of X is

Z explained by Z, and most of the
overlap between X and Y is
accounted for.
 Z is a strong IV.

X Y Usual situation : Not a lof of X is

Z explained by Z, or what is explained
does not overlap much with Y.
 Z is a weak IV.

14
RS – Lecture 8

IV Estimators: 2SLS (2-Stage Least Squares)

• Case 2: l > k -i.e., number of instruments > number of regressors.
- This is the usual case. We can throw l-k instruments, but throwing
away information is never optimal.
- The IV normal equations are an l x k system of equations:
Z’y = Z’X+ Z’
Note: We cannot approximate all the Z’ by 0 simultaneously. There
will be at least l-k non-zero residuals. (Similar setup to a regression!)

- From the IV normal equations  W'Z’X bIV = W'Z’y

- We define a different IV estimator
- Let ZW = Z(Z’Z)-1Z’X = PZX = Z  ˆ  Xˆ
- Then, X'PZX bIV = X'PZy
bIV  ( X ' PZ X ) 1 X ' PZ y  ( X ' PZ PZ X ) 1 X ' PZ PZ y  ( Xˆ ' Xˆ ) 1 Xˆ ' yˆ

IV Estimators: 2SLS: Properties

• It is easy to derive properties for bIV:
bIV    ( X ' PZ X ) 1 X ' PZ 
(1) bIV is consistent
(2) bIV is asymptotically normal.
- This is estimator is also called GIVE (Generalized IV estimator)

Note: In general, we think of X = ZП + V, where V~ N(0, σVV I).

In this case, we add the assumption: plim(Z’V/T) = 0.

• Interpretations of bIV
b IV  b 2 SLS  ( Xˆ ' Xˆ )  1 Xˆ ' y This is the 2SLS interpretation
b IV  ( Xˆ ' X )  1 Xˆ ' y This is the usual IV Z  Xˆ

15
RS – Lecture 8

IV Estimators: 2SLS - Interpretation

• Interpretation of bIV as a 2SLS regression -Theil (1953).
b2 SLS  ( Xˆ ' Xˆ )  1 Xˆ ' y
- First stage, an OLS regression of X on Z. Get fitted values X̂.
- Second stage, another OLS regression of y on X̂ . Get bIV = b2SLS.

Notes:
- In the 1st stage, any variable in X that is also in Z will achieve a
perfect fit (these X are clean), so that this variable is carried over
without modification to the second stage.
- In the 2nd stage, under the usual linear model for X: X = ZП + V,
y  X     Xˆ   {  ( X  Xˆ )  }
The second component of the error term is a source of finite sample
bias, but not inconsistency.

IV Estimators: 2SLS - Interpretation

• In the simplest case with one explanatory variable and one
instrument –i.e., x = z π + v– we get the simple IV estimator:
b2 SLS  ( Xˆ ' X ) 1 Xˆ ' y  (ˆ ' Z ' X ) 1ˆ ' Z ' y  ( ZX ) 1 Z ' y

• The 2SLS estimator can be interpreted as a member of the family of

GMM estimators.

• In this case the moment is E[Z’] and GMM selects  to minimize

the weighted quadratic distance:
’Z WT Z’
where WT is a weight matrix.

Henri Theil (1924-2000, Netherlands)

16
RS – Lecture 8

IV Estimators: 2SLS – Simultaneous Equations

• To check the factors that affect the behavior of IV, let's go back to a
simultaneous equation setting:
y1 = Y  +  –  ~ N(0, σεε I)
Y =ZП+V – V ~ N(0, σVV I)
Then,
b2SLS = [Y' Pz Y]-1 Y'Pz y1
= [(П'Z'+V') Pz (ZП+V)]-1 (П'Z'+V') Pz (Y  + )
b2SLS -  = [П'Z'ZП + V'PzV + П'Z'V + V'ZП]-1 (П'Z' + V'Pz)

The parameter λ = П'Z'ZП/σVV is called the concentration parameter.

• The bias depends on the behavior of Z' (correlation between Z &

), V'Z (exogeneity of Z), and ZП (correlation between Z &Y).

IV Estimators: 2SLS - Example

• Simplest case: Two endogenous variable, one IV.
y1 = y2  +  –  ~ N(0, σεε)
y2 = z π + v – v ~ N(0, σVV)
• The 2SLS bias term (Pz= zz'/Σizi ) 2

b2SLS –  = [π2 z'z + v'Pzv + 2 π z'v]-1 (z' + v'Pz)

We call λ = (π2 Σizi2)/σVV the concentration parameter.

If z is uncorrelated with v –i.e., exogenous–, then:

b2SLS –  = [π2 z'z]-1 (z')

• When Cov(z, ) ≠ 0, 2SLS is inconsistent. If λ is close to 0 the bias

term will get larger (λ ≈ 0 when π ≈ 0 –i.e., Cov(z, y2) ≈ 0).

17
RS – Lecture 8

IV Estimators: 2SLS - Example

• Subtle point: Even if Cov(z, ) = 0, in small samples b2SLS can be
misleading (biased with downward biased SEs).

Problems can be serious when π ≈ 0 and/or l is large relative to k.

IV Estimators: Identification
• Case 3: l < k -i.e., number of instruments < number of regressors.
- We cannot estimate . We do not have enough information in Z to
estimate .
- This is the identification problem. This is the case where we need to
rethink the estimation strategy.
- When we can estimate , we say the model is identified. This
happens when l ≥ k.

Note : When l ≥ k, we have two cases:

-When l = k , we say the model is just identified.
-When l > k , we say the model is over-identified.

18
RS – Lecture 8

Asymptotic Covariance Matrix for 2SLS (Greene)

• The asymptotic variance for the IV and 2SLS is given by:

E [( b IV   )( b IV   )' ]   2 ( Z ' X )  1 ( Z ' Z )( X ' Z )  1

E [( b2 SLS   )( b   )' ]   2 ( Xˆ ' X ) 1 ( Xˆ ' Xˆ )( X ' Xˆ ' ) 1
2 SLS 

  2 ( Xˆ ' Xˆ ) 1

• To estimate Asy Var[b2SLS] we need to estimate σε2:

1

T
ˆ 2  i 1
( y i  x i b 2 SLS ) 2
T

• Do not use the inconsistent estimator:

1

T
ˆ 2  i 1
( y i  xˆ i b 2 SLS ) 2
T

Asymptotic Covariance Matrix for 2SLS (Greene)

• A little bit of algebra relates the asymptotic variances of bIV & bOLS:
V [( b IV ]   2 ( Z ' X )  1 ( Z ' Z )( X ' Z )  1
  2 ( X ' X ) 1 ( X ' X )( Z ' X ) 1 ( Z ' Z )( X ' Z ) 1
 V [( bOLS ][( X ' X )( Z ' X ) 1 ]( Z ' Z )( X ' Z ) 1
 V [( bOLS ][  1 ˆ 1 ]

where we assume X = ZП + V and Φ is the coefficient in the

reverse first stage regression.

Two things to notice:

- As Z ⟶ X, V[bIV] ⟶ V[bOLS].
- As Cov(Z,X) gets smaller –i.e., Z becomes a weak instrument-, V[bIV]
gets larger.

19
RS – Lecture 8

Asymptotic Covariance Matrix for 2SLS (Greene)

• This relation between Cov(Z, X) and estimation uncertainty also
applies to the 2SLS estimators:
V [( b 2 SLS ]   2 ( Xˆ ' Xˆ )  1
  2 ( ˆ ' Z ' Z ˆ ) 1


 Weak instruments create big uncertainty about b2SLS.

2SLS Has Larger Variance than OLS (Greene)

A com parison to OLS
ˆ 'X
Asy.Var[2SLS]=  2 ( X ˆ ) -1
N eglecting the inconsistency,
Asy.Var[LS] =  2 ( X ' X ) -1
( This is the variance of LS around its m ean, not β )
Asy.Var[2SLS]  Asy.Var[LS] in the m atrix sense.
Com pare inverses:
ˆ 'X
{Asy.Var[LS]} -1 - {Asy.Var[2SLS]} -1  (1 /  2 )[ X ' X - X ˆ]
 (1 /  2 )[ X ' X - X '( I  M Z ) X ]=(1 /  2 )[ X ' M Z X ]
This m atrix is nonnegative definite. (N ot positive definite
as it m ight have som e row s and colum ns w hich are zero.)
Im plication for "precision" of 2SLS.
The problem of "W eak Instrum ents"

20
RS – Lecture 8

Asymptotic Efficiency (Greene)

• The variance is larger than that of OLS. (A large sample type of
Gauss-Markov result is at work.)
(1) OLS is inconsistent.
(2) Mean squared error is uncertain:
MSE[estimator|] = Variance + square of bias.

For a long time, IV was thought to be “the cure” to the biases

introduced by OLS. But, in terms of MSE, IV may be better or worse.
It depends on the data: X, Z and ε.

A Popular Misconception (Greene)

• A popular misconception. If only one variable in X is correlated
with , the other coefficients are consistently estimated. False.

S u p p o s e o n ly th e fir s t v a r ia b le is c o rr e la te d w ith ε
  1 
 
0
U n d e r th e a s s u m p tio n s , p lim ( X 'ε /n ) =   . Then
 ... 
 
 . 
  1   q 11 
   21 
p lim b - β = p lim ( X 'X /n ) -1  0    q 
 ...  1
 ... 
   K 1 
 .  q 
  1  tim e s th e firs t c o lu m n o f Q -1

The problem is “smeared” over the other coefficients.

21
RS – Lecture 8

Small sample properties of IV

• Now, we do not have the condition E[ε|X] = 0, we cannot get
simple expressions for the moments of b2SLS:
b2SLS =  + [Y' Pz Y]-1 Y' Pz 
by first taking expectations of conditioned on X and Z. The bias:
b2SLS - β = [Y' Pz Y]-1 Y' Pz 

• We cannot say that b2SLS is unbiased (even when Cov(z, ) = 0!), or

that it has the Var[b2SLS] equal to its Asy Var[b2SLS].

• Also, recall that the 1st stage introduces a source of finite sample
bias: the estimation of П.

• In fact, b2SLS can have very bad small-sample properties.

Small sample properties of IV - Simulation

• To study the behavior of bIV, for small T, we set up a simple Monte
Carlo experiment using a model appropriate to the context.

• Recall the asymptotic distribution of bIV

 2 
 
T bIV   

d 1
N  0, 2  2 
  X  XZ 

• We will see that the small sample behavior of bIV will depend on the
nature of the model, the correlation between X and ε, and the
correlation between X and Z.

16

22
RS – Lecture 8

Small sample properties of IV – Simulation

• We start with a simple linear model:
Y  1   2 X  
X   1 Z   2U  
with observations on Z, U, and ε are drawn independently from a
N(0,1). We think of Z and U as variables and of ε as the error term
in the model. π and π2 are constants.
• By construction, X is not independent of ε. OLS is inconsistent and
its standard errors and tests will be invalid.
• Z is correlated with X, but independent of ε. It serves as an
instrument. (U is included to provide some variation in X, not
connected with either Z or ε.)
17

Small sample properties of IV – Simulation

• To start the simulation, we set:
  π = 0.5, and π  = 2.0.
• That is,
Y  10  5 X    ~ iid N (0,1)
X  0.5 Z  2.0 U   Z ~ iid N (0,1); U ~ iid N (0,1)
• It is easy to check that plim b2,OLS = 5.19 (=5+1/(.52+22+1)). Of
course, plim b2,IV = 5.00. We draw n=25, 100 & 3,200. We do 1 million
simulations.
Sample Size b2,OLS (SE[b2,OLS ]) b2,IV (SE[b2,IV]) MSEs
n = 25 5.190 (0.080) 4.998 (0.137) .055 - .019
n = 100 5.191 (0.040) 5.000 (0.054) .038 - .003
n = 3200 5.191 (0.007) 5.000 (0.009) .036 - .0001 20

23
RS – Lecture 8

Small sample properties of IV – Simulation

10
OLS, n = 100

IV, n = 100 OLS, n = 25

IV, n = 25

0
4 5 6

• b2,IV has a greater variance than b2,OLS. For small samples (say, n = 25
or 100) OLS may be better in terms of MSE. But, as n grows, b2,IV and
b2,OLS tend to their plims (b2,IV more slowly than b2,OLS, because it has a
larger variance).
21

Small sample properties of IV – Simulation

n = 25
n = 100
limiting normal distribution 0.2
n = 3,200

0.1

0
-6 -4 -2 0 2 4 6

• We have the distribution of √n (b2,IV – b2) for n = 25, 100, and 3,200.
It also shows, as the dashed red line, the limiting normal distribution.
For n = 3,200 is very close to the limiting distribution. Inference would
be OK with samples of this magnitude. 24

24
RS – Lecture 8

Small sample properties of IV – Simulation

n = 25
n = 100
limiting normal distribution 0.2
n = 3,200

0.1

0
-6 -4 -2 0 2 4 6

• For n=25, 100, the tail are too fat. Inference would give rise to excess
instances of Type I error (under rejection). The distortion for small
sample sizes is partly attributable to the low ρxz corr(X,Z)=0.22
(=.5/sqrt(5.25)) (or weak instruments; common in IV estimation). 24

Small sample properties of IV – Simulation

• To check the effect of ρxz on the estimation, we lower π to 0.01,
which brings ρxz corr(X,Z) = 0.0045 (=.1/sqrt(5.01)) and we
increase π to 4, with ρxz 0.8729.

Empirical Type I Error

Sample Size ρ = .8729 ρ = .2182 ρ = .0045
n=25 0.0698 0.0717 0.0752
n = 100 0.0610 0.0628 0.0628
n = 3200 0.0511 0.0508 0.0527

• Some size problem for small n. Low ρxz slightly increases the size
problem.

24

25
RS – Lecture 8

Cornwell and Rupert Data (Greene)

Cornwell and Rupert Returns to Schooling Data, 595 Individuals, 7 Years
Variables in the file are
EXP = work experience
WKS = weeks worked
OCC = occupation, 1 if blue collar,
IND = 1 if manufacturing industry
SOUTH = 1 if resides in south
SMSA = 1 if resides in a city (SMSA)
MS = 1 if married
FEM = 1 if female
UNION = 1 if wage set by union contract
ED = years of education
BLK = 1 if individual is black
LWAGE = log of wage = dependent variable in regressions

These data were analyzed in Cornwell, C. and Rupert, P., "Efficient Estimation with Panel
Data: An Empirical Comparison of Instrumental Variable Estimators," Journal of Applied
Econometrics, 3, 1988, pp. 149-155. See Baltagi, page 122 for further analysis. The data
were downloaded from the website for Baltagi's text.

Application: Wage Equation (Greene)

• Are earnings affected by education? In a linear regression, we
expect the education coefficient to be positive (and significant, if
human capital theory is correct).

• Linear regression model:

logWage = y = Xβ + ε
X = one, exp, occ, ed (education), wks
- We expect Wks -weeks worked- to be endogenous
- Instruments: Z = one, exp, occ, ed, ind, south, smsa, ms, fem

• Q: How do we know when a variable is exogenous?

26
RS – Lecture 8

Estimated Wage Equation (Greene)

+----------------------------------------------------+
| Ordinary least squares regression |
+----------------------------------------------------+
+--------+--------------+----------------+--------+--------+----------+
|Variable| Coefficient | Standard Error |b/St.Er.|P[|Z|>z]| Mean of X|
+--------+--------------+----------------+--------+--------+----------+
|Constant| 5.30277*** .07406 71.605 .0000 |
|EXP | .01294*** .00058 22.393 .0000 19.8538|
|OCC | -.08511*** .01575 -5.403 .0000 .51116|
|ED | .06694*** .00288 23.204 .0000 12.8454|
|WKS | .00641*** .00120 5.330 .0000 46.8115|
+--------+------------------------------------------------------------+
+----------------------------------------------------+
| Two stage least squares regression |
+----------------------------------------------------+
+---------------------------------------------------------------------+
|Instrumental Variables: |
|ONE EXP OCC ED IND SOUTH SMSA |
|MS FEM |
+---------------------------------------------------------------------+
|Constant| -6.60400*** 1.81742 -3.634 .0003 |
|EXP | .01735*** .00205 8.457 .0000 19.8538|
|OCC | -.04375 .05325 -.822 .4113 .51116|
|ED | .07840*** .00984 7.968 .0000 12.8454|
|WKS | .25530*** .03785 6.745 .0000 46.8115|
+--------+------------------------------------------------------------+

Endogeneity Test (Hausman)

Exogenous Endogenous
OLS Consistent, Efficient Inconsistent

2SLS Consistent, Inefficient Consistent

• Base a test on d = b2SLS – bOLS

- We can use a Wald statistic: d’[Var(d)]-1d
Note: Under H0 (plim (X’/T) = 0) bOLS = b2SLS = b
- Also, under H0: Var[b2SLS ]= V2SLS > Var[bOLS ]= VOLS
 Under H0, one estimator is efficient, the other one is not.

• Q: What to use for Var(d)?

- Hausman (1978): V = Var(d) = V2SLS – VOLS
H = (b2SLS – bOLS)’[V2SLS – VOLS ]-1(b2SLS – bOLS)  χ2rank(V)
d

27
RS – Lecture 8

Endogeneity Test (Hausman)

Q: What to use for Var(d)?

- Hausman (1978): V = Var(d) = V2SLS – VOLS
H = (b2SLS – bOLS)’[V2SLS – VOLS ]-1(b2SLS – bOLS)

• Hausman gets Var(d) by using the following result:

"The covariance between an efficient estimator (bE) and its difference from an
inefficient estimator (bE – bI) is zero." That is,
Cov(bE, bE – bI) = Cov(bE, bE) – Cov(bE, bI)
= Var(bE) – Cov(bE, bI) = 0
 Var(bE) = Cov (bE, bI)

• Hausman's case: aVar(bOLS) = aCov (bOLS, b2SLS)

Then, aVar(d) = aVar(bOLS) + aVar(b2SLS) - 2 aCov (bOLS, b2SLS)
= aVar(b2SLS) – aVar(bOLS)

Endogeneity Test (Hausman)

• H = (b2SLS – bOLS)’V-1(b2SLS – bOLS)

where V = V2SLS – VOLS.

• There are different variations of H, depending on which estimator

of V is used. Using V[bOLS] and V[b2SLS] can create problems in
small sample (V may not be pd).

• There are a couple of solutions to this problem, for example,

imposing a common estimate of σ. If we use s2, the OLS estimator,
we have Durbin’s (1954) version of the test.

28
RS – Lecture 8

Endogeneity Test: The Wu Test

• The Hausman test has some computation issues.

• Simplification: The Wu test.
• Consider a regression y = Xβ + ε, an array of proper instruments Z,
and an array of instruments W that includes Z plus other variables
that may be either clean or contaminated.

• Wu test for H0: X is clean. Setup:

(1) Regress X on Z. Keep fitted values X̂ = Z(Z’Z)-1Z’X
(2) Using W as instruments, do a 2SLS regression of y on X, keep
RSS1.
(3) Do a 2SLS regression of y on X and a subset of m columns of X̂
that are linearly independent of X. Keep RSS2.
(4) Do an F-test: F = [(RSS1 – RSS2)/m]/[RSS2/(T-k)].

Endogeneity Test: The Wu Test

• Under H0: X is clean, the F statistic has an approximate Fm,T-k

distribution.

• The test can be interpreted as a test for whether the m auxiliary

variables from X̂ should be omitted from the regression.

• When a subset of X̂ of maximum possible rank is chosen, this

statistic turns out to be asymptotically equivalent to the Hausman test
statistic.

Note: If W contains X, then the 2SLS in the second and third steps
reduces to OLS.

29
RS – Lecture 8

Endogeneity Test: The Wu Test

Note: If W contains X, then the 2SLS in the second and third steps
reduces to OLS.

Davidson and MacKinnon (1993) point out that the DWH test really
tests whether possible endogeneity of the right-hand-side variables
not contained in the instruments makes any difference to the
coefficient estimates.

• These types of exogeneity tests are usually known as DWH (Durbin,

Wu, Hausman) tests.

Endogeneity Test: Augmented DWH Test

• Davidson and MacKinnon (1993) suggest an augmented regression

test (DWH test), by including the residuals of each endogenous right-
hand side variable.

• Model: y = X β + Uγ + , we suspect X is endogenous.

• Steps for augmented regression DWH test:

1. Regress x on IV (Z) and U:
x = Z П + U φ + υ  save residuals vx
2. Do an augmented regression: y = Xβ + Uγ + vx δ + ε
3. Do a t-test of δ. If the estimate of δ, say d, is significantly different
from zero, then OLS is not consistent.

30
RS – Lecture 8

Endogeneity Test: Augmented DWH Test

Intuition: Since each instrument, Z, is uncorrelated with , x is

uncorrelated with  only if vx is uncorrelated with . Then, the DWH
tests becomes
H0: E[vx  ] = 0.

• This is the most popular version of the DWH test.

Implication of DWH: Reject H0  OLS is inconsistent. IV results

should be preferred (the rest of lecture puts some breaks to this
implication!)

Wu Test (Greene)
+----------------------------------------------------+
| Ordinary least squares regression |
| LHS=LWAGE Mean = 6.676346 |
+----------------------------------------------------+
+--------+--------------+----------------+--------+--------+----------+
|Variable| Coefficient | Standard Error |b/St.Er.|P[|Z|>z]| Mean of X|
+--------+--------------+----------------+--------+--------+----------+
|Constant| -6.60400*** .50833 -12.992 .0000 |
|EXP | .01735*** .00057 30.235 .0000 19.8538|
|OCC | -.04375*** .01489 -2.937 .0033 .51116|
|ED | .07840*** .00275 28.489 .0000 12.8454|
|WKS | .00355*** .00114 3.120 .0018 46.8115|
|WKSHAT | .25176*** .01065 23.646 .0000 46.8115|
+--------+------------------------------------------------------------+
| Note: ***, **, * = Significance at 1%, 5%, 10% level. |
+---------------------------------------------------------------------+

--> Calc ; list ; Wutest = b(kreg)^2 / Varb(kreg,kreg) $

+------------------------------------+
| Listed Calculator Results |
+------------------------------------+
WUTEST = 559.119128 (=23.646^2) => OLS is inconsistent!

31
RS – Lecture 8

Measurement Error
• DGP: y* = x* +  -  ~ iid D(0, σε2)
- all of the CLM assumptions apply.
Problem: x*, y* are not observed or measured correctly. x, y are
observed:
x = x* + u u ~ iid D(0, σu2) -no correlation to ,v
y = y* + v v ~ iid D(0, σv2) -no correlation to ,u

• Let’s consider two cases:

- CASE 1 - Only x* is measured with error (y=y*).
- CASE 2 - Only y* is measured with error (x=x*).

Measurement Error
CASE 1 - Only x* is measured with error.
y = y* = x* + 
y = (x – u) +  = x +  – u = x + w
E[x’w] = E[(x* + u)’( – u)] = – σu2 ≠ 0
 OLS biased & inconsistent. We need IV!

• Typical IV solution: Find another noisy measure of x*, say z:

z = x* + η η ~ iid D(0, σw2) -no correlation to , v, u
Check IV conditions:
- Cov(z, ) = Cov(x*+ η, ) = 0
- Cov(z, x) = Cov(x*+ η, x*+u) = Var(x*) 0
Then,
bIV = Cov(z, x)-1 Cov(z, y) = Var(x*)-1 Cov(z, y)
 IV removes the variance in noise.

32
RS – Lecture 8

Measurement Error
• Q: What happens when OLS is used –i.e., we y regress on x?
A: Least squares attenuation:
c o v ( x ,y ) cov(x * u,  x *  )
p lim b = 
v a r(x ) v a r( x *  u )
 v a r( x * )
= < 
v a r( x * )  v a r(u )

• Q: Why is OLS attenuated?

y = x* + 
x = x* + u
y = x + ( – u) = x + v, cov(x,v) = -  var(σu2)

Some of the variation in x is not associated with variation in y. The

effect of variation in x on y is dampened by the measurement error.

Measurement Error
CASE 2 - Only y* is measured with error.
y* = y - v = x* +  = x + 
 y = x +  + v = x + ( + v)
• Q: What happens when y is regressed on x?
A: Nothing! We have our usual OLS problem since  and v are
independent of each other and x*. CLM assumptions are not violated!

• Q: Is measurement error in finance/economics a problem?

A: Yes! In surveys and forms, mistakes are common. Most relevant
problem: often, economic theories deal with unobservables (x*).

Famous unobservables: Market portfolio, innovation, growth

opportunities, potential output, target debt-equity ratio, business
cycles, worker’s skills.

33
RS – Lecture 8

Measurement Error: Proxy Variables

• Often, economic theories deal with unobservables (x*). To test these
theories, practitioners use a proxy (x), instead of x*.

A proxy is a variable that has a “close” relation (usually, linear) with

the unobservable:
x = δ x* + u (typical measurement error problem!)

Example: The CAPM: E[Ri – Rf]= i E[RMP – Rf ]

The market portfolio (MP) is unobservable. According to Roll's
(1977) critique, this makes the CAPM untestable!

In practice, we proxy it by a representative stock market index:

RIndex = δ RMP + u

Measurement Error: Proxy Variables

Example: Testing the CAPM I.
(1) CAPM regression:
Ri – Rf = αi + i (RMP – Rf ) + 
H0: αi=0 (αi is the pricing error. Jensen’s alpha.)
(2) MP unobservable. Proxy: S&P 500 stock market index
RSP500 = η RMP + u  RMP = θ RSP500 + u’
(3) Working CAPM regression
Ri - Rf = αi + i [(θ RSP500 + u’) – Rf] + 
= αi + iθ RSP500 – i Rf + ξ (ξ = i u’ + )
Or,
Ri = αi + δi Rf + γi RSP500 + ξ
where γi = iθ and δi = 1 – i
 αi can be estimated directly, but i cannot be estimated directly!

34
RS – Lecture 8

Measurement Error: Proxy Variables

• Ri = αi + δi Rf + γi RSP500 + ξ (ξ = i u’ + )

(4) Usually, Rf is assumed constant

Ri = αi* + γi RSP500 + ξ
where αi* = αi + δi Rf
Ri = αi* + γi RSP500 + ξ
We can do a regression to estimate αi‘ and γi.
But since γi = iθ  i cannot be estimated!

Note: It is common to just work with “excess returns” directly. In this

case, the proxy would be:
RSP500 – Rf = η (RMP – Rf) + u

Measurement Error: Proxy Variables

Example: Testing the CAPM II. We extend the CAPM (APT style):
(1) CAPM regression with more explanatory variables (W):
Ri – Rf = αi + i (RMP – Rf) + ψi W + 
H0: ψi = 0
(2) MP unobservable. Proxy: S&P 500 stock market index
RSP500 = η RMP + u  RMP = θ RSP500 + u*
(3) Working CAPM extended regression
Ri = αi + δi Rf + γi RSP500 + ψi W + ξ (ξ = i u* + )

OLS estimates αi, δi (=1 – i), γi(= θ i), ψi (but, not i directly!). H0
can be tested. (In general, smearing complicates the estimation.)

Note: Assuming a constant Rf, we get estimates αi*, γi, ψi.

35
RS – Lecture 8

Measurement Error: Smearing Again (Greene)

M u ltip le re g re s s io n : y =  1 x 1 *   2 x 2 *  
x 1 * is m e a s u re d w ith e r ro r; x 1  x 1 *  u
x 2 is m e a s u r e d w ith o u t e rro r .
T h e re g r e s s io n is e s tim a te d b y le a s t s q u a re s
P o p u la r m y th # 1 . b 1 is b ia s e d d o w n w a rd , b 2 c o n s is te n t.
P o p u la r m y th # 2 . A ll c o e ffic ie n ts a re b ia s e d to w a rd z e ro .
R e s u lt fo r th e s im p le s t c a s e . L e t
 ij  c o v ( x i * , x j * ) , i, j  1 , 2 ( 2 x 2 c o v a r ia n c e m a trix )
 ij  ijth e le m e n t o f th e in v e r s e o f th e c o v a ria n c e m a trix
 2  v a r (u )
F o r th e le a s t s q u a re s e s tim a to rs :
 1    2  12 
p lim b 1   1  2 11 
, p lim b 2   2   1  2 11 
1     1    
T h e e ffe c t is c a lle d " s m e a r in g ."

Measurement Error: Twinsville

• Q: Does education affects earnings?

A: We expect two people with similar natural abilities but different
levels of education to be differently paid. To estimate returns-to-
schooling, economists often use a linear regression model relating log
earnings (y) to years of education (x*), with additional control
variables (U). The error term represents the effects of person-to-
person variation that have not been controlled for. The DGP:
y = x* + Uγ + 
• We expect two people with similar natural abilities:
 More education, more earnings. We expect >0.

• Problem: x* is usually self-reported, and often reported with error.

36
RS – Lecture 8

Measurement Error: Twinsville

• Linear model: y = x + Uγ + 
• H0:  = 0.
• We do not observe x*, we observe self-reported x. We need to find
an instrument to estimate the model.

• Famous application from the econ literature: Ashenfelter/Kreuger

(AER,1994): A wage equation for twins that includes two measures of
x: each twin reports their own and their twin’s schooling.

• The data suggests that between 8% and 12% of the measured

variance in schooling levels is error.

• Instrument: Reported schooling by the twin.

Measurement Error: Twinsville

37
RS – Lecture 8

Measurement Error: Twinsville

Omitted Variables: IV Conditions (Again)

• The omitted variables problem is, probably, the most popular IV
application in microeconomics and Corporate Finance.

• Typical omitted variables situation: In the CEO compensation model,

we want to test the causal effect of networking on compensation, but a
CEO's unobserved skills blurs the causality, since Cov(x, ) ≠ 0.

• Recall that IV estimators are consistent if the instruments, Z, used are

both valid and relevant/informative. That is, we look for Z such that
(1) Cov(X, Z) ≠ 0 - relevance condition
(2) Cov(Z, ) = 0 - valid condition (exclusion restriction)

In an omitted variables problem, we can think of (2) as broken into

two parts: (a) Z is uncorrelated to , & (b) only affects y through X.

38
RS – Lecture 8

Omitted Variables: IV Conditions (Again)

• (2) ⇒ Z is not only uncorrelated to , but only affects y through X –
after all, it is excluded from structural equation!
Z→X→y

From the 2nd part: Once I know the effect of Z on X, I can throw Z.

• Historically, the emphasis has been on the valid (exogeneity) condition,

Cov(Z, )=0.

• But, the past 25 years added an additional source of concern: the

Cov(X, Z) may not be high enough. That is, (from the first stage)
may not be very informative about X:
X = ZП + Uδ + V – V ~N(0, σV2I)

Omitted Variables: Finding Good IVs

• Back to CEO compensation model. We need IVs, Z, such that
(1) Explain the variation in networking –i.e., Cov(x, Z) ≠ 0
(2) Do not directly affect CEO compensation –i.e., Cov(Z, ) = 0.

• It is not difficult to find a Z that meets (2), the valid condition. Many
variables are not correlated with , the error term from the CEO
compensation structural equation.

Examples: Potential IVs, uncorrelated with .

Earthquakes in New Zealand; past debt of Denton, TX; asteroids
hitting the Atlantic the year of the CEO’s birth; number of letters on
the name of CEO’s high school.

• We like these potential IVs; they look random or orthogonal to a

CEO compensation model (unrelated to ). They can be safely
excluded from the structural equation. That is, they meet Cov(Z, )=0.

39
RS – Lecture 8

Omitted Variables: Finding Good IVs

• But, it is dubious the effect of these IVs on networking, x. The
relevant condition is likely not met.

Note: Deaton (2010) calls the variables in the examples external, since
they are not determined by the model. They may not be exogenous.

• The key is to find a Z correlated with X –i.e., the relevant condition–

uncorrelated with  (and with the omitted variable, unobserved skills.)

• Starting with Angrist (1990) and Angrist and Krueger (1991) (for us,
A&K), who study the effect on earnings of civilian work experience
and schooling, respectively, there has been an emphasis on using a Z
that can be defined by a natural experiment when the IV problem is
caused by omitted variables.

• Usually, a natural experiment is exogenous to a structural model. Like

the previous external examples, the exclusion condition is met.

Omitted Variables: Finding Good IVs

• The key is to find a natural experiment (defining Z) that is correlated
with X and has no direct effect on y –the impact on y is through X.

Z is an exogenous event  resulting values of X induced by Z may be

considered randomized –a key feature in lab/medical experiments.

• In a lab experiment (the gold standard in experimental sciences), a

researcher randomly assigns a treatment to a group, creating two
groups: treated and not-treated or control. Then, the researcher studies
the effect of the treatment on, say, the group’s health.

The key feature is the randomization of the treatment. The lab

researcher needs to show that the two groups are comparable along all
dimensions relevant for the outcome variable (age, gender, previous
health, etc.) except the one involving the treatment.

40
RS – Lecture 8

Omitted Variables: Finding Good IVs

• Recall that in the CEO compensation model, we want to test the
causal effect of networking on compensation, but an omitted variable –
the CEO's unobserved skills– creates endogeneity.

• A solution to the omitted variables problem is to assign networking

(x) randomly: we have two similar groups of CEOs (with similar skills!)
and randomly we assign them values (say, large network & small
network).

• Of course, this randomized experiment is not possible.

• But, suppose we have a natural event, Z, unrelated to CEO

compensation, which randomly assigns networking, x, to two groups.
Then, we can test causality, without the omitted variable problem.

Omitted Variables: Natural Experiments

• We need to find a natural event, Z, unrelated to y, which randomly
assigns X to two groups.

• We use Z to identify causality. This is why natural experiments are

popular in economics & finance (especially, in Corporate Finance).

• We define natural experiments as historical (exogenous) episodes that

provide observable, quasi-random variation in treatment subject to a
plausible identifying assumption.

• “Natural” point out that us (the researchers) did not design the
episode to be analyzed, but can use it to identify causal relationships.

41
RS – Lecture 8

Omitted Variables: Natural Experiments

• Steps of a natural experiment:
(1) Experiment defines an IV: Zi=1 (i treated), Zi=0 (i control).
(2) Identify two groups:
– treated (all i with Zi=1) with observations: y(1), X(1)
– control (all i with Zi=0) with observations: y(0), X(0)
(3) We analyze differences between (y(1), X(1)) & (y(0), X(0)).

• Remarks: These steps will be treated like a lab experiment if we show

that the treatment is in fact randomly assigned. We need to show that
two groups are comparable except for the treatment.

• This is the key for the experiment to be valid. We need to convince

the audience that the we have a quasi-random treatment.

Finding Good IVs: Natural Experiments

Example: There are significant persistent differences in development
(y) among similar cities. One explanation: Location (x); proximity to
other cities matter. Cities close to another city enjoy externalities (say,
transportation and school networks). We want to test this hypothesis.

I would like to estimate a model: y = x + Uγ +  (but location, x,

is also endogenous. OLS will not work!)

- Ideal experiment: Identify 2 similar cities and remove a city next to it.

- Natural experiment: German division in 1949.

Cities close to the border lost connection to the cities on the other side
of the border. It looks like randomly removing a city! –from Redding
& Sturm (2008, AER).

42
RS – Lecture 8

Finding Good IVs: Natural Experiments

• Now, we need to convince the reader that the German division
provides a legitimate quasi-random treatment:
We need to show that the partition was exogenous (not based on
development, y –i.e., unrelated to the structural model!)

- We need to show that there was no cofounding treatment (something

else happening along with the partition). For example, after the
partition, a city close to the border may be in fear of war.

Finding Good IVs: Natural Experiments

• In the context of Natural Experiments, a good instrument, Z, should
satisfy:
(1) Explain the variation in x –i.e., Cov(x, Z) ≠ 0
(2) Do not directly affect y –i.e., Cov(Z, ) = 0.
(3) As good as randomly assigned.

• Only condition (1) is the only one we can directly check, through the
first-stage regression, where we get . Given that  is unobservable,
the legitimacy of (2) is usually left for theory or common sense. A
researcher should also convince the audience about the validity of (3).

• Finding a Z that meets all requirements is not easy.

43
RS – Lecture 8

Finding an Instrument: A&K (1991)

• Back to the question: Does education, x, affect earnings, y?
We use the same linear model: y = x + Uγ + 
We expect  > 0. But, since Cov(x, ) ≠ 0; we know that OLS b is
inconsistent. We need IV estimation.

Note: In general, U does not capture much of the variation of y.

• Angrist and Krueger’s (1991, QJE) Natural experiment: Find an

exogenous historical event that creates variation in schooling.

Exogenous event: Compulsory schooling laws according to age, not

years of schooling completed.

Finding an Instrument: A&K (1991)

• Years of schooling vary by quarter of birth (QOB=z):
– In the U.S., it is legal to drop out at 16.
– Someone born in Q1 is a little younger and can drop out with
less schooling, than someone born in Q4  Cov(z, x) ≠ 0.
• QOB can be treated as a source of exogeneity in schooling,
unrelated to individual ability  Cov(z, ) = 0.

• That is, Z should affect earnings only through its effect on schooling.

44
RS – Lecture 8

Finding an Instrument: A&K (1991)

• The data for the 1930-39 cohort show that men born earlier in the
year have lower schooling. QOB can be an instrument  there is a
first stage: x = π z + Dγ + v (Z: Dummy variable for QOB)

Source: Angrist and Krueger (1991), Figure I

Finding an Instrument: A&K (1991)

• In a reduced form, we can see a relation between earnings and

QOB: y = Z Г + U φ + ξ

45
RS – Lecture 8

Finding an Instrument: A&K (1991)

• People born in Q1 do obtain less schooling
– But pay close attention to the scale of the y-axis.
– Mean differences between Q1 and Q4 are small (in education,
the difference is only 0.151 and in log earnings 0.014).

• Thus, we need large T since R2X,Z will be very small

– A&K had over 300,000 observations for the 1930-39 cohort

• Final 2SLS model interacted QOB with year of birth (30), state of
birth (150):
– OLS: bOLS = .0628 (s.e. = .0003) (large T  small SE’s).
– 2SLS: b2SLS = .0811 (s.e. = .0109)
– Var[bIV] > Var[bOLS], as expected. (But, maybe too large?)

Finding an Instrument: A&K (1991)

• OLS estimate does not appear to be badly biased. But...

46
RS – Lecture 8

IV with a Dummy Variable

• QOB is a dummy IV. It is the treatment in the natural experiment. In
the simplest model, with one dummy IV, the IV estimator becomes:
Cov ( Z , y ) indirect effect of z on y E[ y | Z  1]  E[ y | Z  0]
bIV   
Cov ( Z , x ) indirect effect of z on x E[ x | Z  1]  E[ x | Z  0]

• This is the Wald (1944) estimator (also, grouping estimator): A ratio of

differences in (group) means (y & x) in treated and control groups.

• To get the above result, recall

Cov[Z,Y] = E[ZY] – E[Y] E[Z]
= E[Y|Z=1] P[Z=1] – E[Y] P[Z=1]
Derivation trick: Y = Y [Z + (1 – Z)]. Then, take expectations on the
last term and some algebra delivers:
Cov[Z,Y] = {E[Y|Z=1] – E[Y|Z=0]} P[Z=1] P[Z=0]

IV with a Dummy Variable

• Similar work for Cov(X, Y) gets the result: bIV = Wald Estimator.

• Interpretation of Wald Estimator, as a ratio of slopes:

- First stage: x = π1 + π2 z + v
- Reduced form: y = γ1 + γ2 z + ξ
Taking conditional expectations on Z above and simple algebra:
π2 = E[x|Z=1] – E[x|Z=0]
γ2 = E[y|Z=1] – E[y|Z=0]

E [ y | Z  1]  E [ y | Z  0]  2
Then, bIV  
E [ x | Z  1]  E [ x | Z  0]  2
• The Wald estimator is known as local IV or local average treatment
effect, LATE (under some assumptions, bIV = E[y(1) – y(0)|compliers]).

47
RS – Lecture 8

IV with a Dummy Variable

• Application to A&K (1991):

bIV = (5.90271 – 5.8916)/(12.7969 – 12.6881) = 0.1021

Interpretation: The Wald estimator measures the effect of an extra
year of schooling on those (dropout) students for whom an earlier
birth –i.e., Z changes from 0 to 1– would have been forced to
complete an extra year of schooling before dropping out.

IV with a Dummy Variable

• The above result can be extended to IV with multiple dummy
instruments. For example, J categories; say, 4 QOB: Q1, Q2, Q3, Q4.

From the structural equation: yi = xi + i

 E[yi |Zi] = E[xi|Zi]

• We replace the expectations (say, E[yi|Zi=j]) with sample analogs

( j). Then, in this case, the IV estimator is the same as the coefficient
from a regression of J group means between Y and X , weighted by
the size of the groups.

48
RS – Lecture 8

Finite Sample Problems

• IV estimators are consistent if the instruments, Z, used are both valid
and relevant/informative, but they may be subject to significant finite
sample biases.

• We now look at two distinct sources of finite sample bias:

– The use of IVs that are only weakly related to the endogenous
variable(s), resulting in “weak identification” of the parameters of
interest. This is the weak instruments problem.

– The use of “too many” instruments relative to the available

sample size. This is the overidentification problem.

Weak Instruments: Definition and Implications

• The explanatory power of Z may not be enough to allow inference on
. In this case, we say Z is a weak instrument.

Definition: Weak Instrument

IVs are weak if the mean component of X that depends on the IVs –
ZП– is small relative to the variability of X, or equivalently, to the
variability of the error V.

• Implications:
– Gleser and Hwang (1987) and Dufour (1997) show that CIs and
tests based on t-tests and F (Wald) tests are not robust to weak IVs.
– The concern is not just theoretical: Numerical studies show that
coverage rates of conventional 2SLS CIs can be very poor when
IVs are weak, even if T is large.

49
RS – Lecture 8

Weak Instruments: Detection (Greene)

• Usual detection of weak IVs: Check if H0 (weak instruments): П = 0:
– Test H0: П = 0 with a standard F-test on Z in the 1st stage
regression.
– Rule of thumb: For a single endogenous regressor, Staiger and
Stock (1997) suggest that 1st stage F < 10 is cause for concern.
– Low partial-R2X,Z –exogenous variable U is partialed out;
see, Shea (1997).
– Large Var[bIV].

Note: There is a theoretical problem when, under a H0, we have

unidentified parameters. Under H0: П = 0,  is not identified.

Weak Instruments: A&K (1991)

• True story: The graduate labor class at the University of Michigan
does replication exercises. Two students, Regina Baker and David
Jaeger replicated the results in Angrist and Krueger (1991).

• They and their professor, John Bound, notice two things:

(1) The results are imprecise and unstable when the controls and
instrument sets change.
(2) The results become precise and stable only when the 1st stage F
tests reject H0: П = 0 –i.e., when instruments are not weak.

Note: Consider the first stage: X = ZП + ξ.

Even if П = 0 in the DGP, as the number of instruments increases
the R2 of the first stage regression in the sample can only increase.

50
RS – Lecture 8

Weak Instruments: A&K (1991)

Note: From BJB (1995, JASA). Different instruments deliver

different 1st stage F-stats. In only (2) there is a significant F-stat!

Weak Instruments: A&K (1991)

Note: As the number of IVs increase, b2SLS gets closer to bOLS.

51
RS – Lecture 8

Weak Instruments: A&K (1991)

• BJB suspected the presence of irrelevant IV. Then, they estimated
the IV coefficient with a randomly assigned Z so that π=0 by
construction. They reproduced the OLS estimate.
 BJB’s suggestion: look at the 1st stage F-stat.

Weak Instruments: A&K (1991)

• QOB looked promising as an IV for education. What went wrong?

• Potential problems with QOB as an IV:

(1) Correlation between QOB and schooling is weak
- Small Cov(X, Z) introduces finite-sample bias, which will be
exacerbated with the inclusion of many IV’s.
(2) QOB may not be completely exogenous
- Recall that even small Cov(Z, ) will cause inconsistency, and
this will be exacerbated when Cov(X, Z) is small.

• QOB qualifies as a weak instrument that may be correlated with

unobserved determinants of wages (e.g., family income).

52
RS – Lecture 8

Weak Instruments: LIML

• There are alternative estimators, which have better small sample
properties than 2SLS with weak instruments. One popular choice is
LIML (limited information maximum likelihood), where we assume joint
normality for the reduced form errors, u’ = (ξ, V)’.

• Limited Information? We estimate only one equation, say, the first

one yj, in a set of simultaneous equations:
yj = Yj j + Zj γj + j j = 1, 2, ..., k (structural equations)
where Yj is a matrix of k1 included endogenous variables Zj is a
matrix of included instruments. There are l2 = l – l1 excluded instruments.

• Suppose we are interested in estimating the first equation y1 (Yj=Xj):

y1 = X1 1 + Z1 γ1 + 1
X1= Z1 Γ11 + Z2 Γ21 + V1

Weak Instruments: LIML

• Define θ = (1, γ1, Γ11, Γ21) and
, , 1 0 γ1 Γ11
Yi = , Zi = , B= 1 , Γ
, , 0 Γ21
• Assume u1’ = (ξ1, V1)’ ~ N(0, Σ). Then, the average log-likelihood:

L(θ, Σ)= log(2π log |Σ|) ∑ B Yi Γ′Zi)′ Σ B Yi Γ′Zi)

Note: The Jacobian of the transformation from Yi to u1 is B whose

determinant is 1.

After a lot of algebra, the log of the concentrated log-likelihood with

respect to (Σ, 1, Γ11, Γ21) is:
L(θ, Σ)= log(2π log κ 1) log | |

53
RS – Lecture 8

Weak Instruments: LIML

• After a lot of algebra, the log of the concentrated log-likelihood with
respect to (Σ, 1, Γ11, Γ21):
L(θ, Σ) = log 2 log κ 1) log | |
where
κ 1) =
where δ 1, ‐1), M1 = IT – PZ1.

• Maximizing L(θ, Σ) is equivalent to minimizing κ 1). Thus,

sometimes LIML estimators are called least variance ratio estimator.

• ̂ κ 1) 1, since span(Z1) ⊂ span(Z) and the numerator of

κ 1) cannot be smaller than the denominator for any δ. For any
equation just identified, ̂ 1.

Weak Instruments: LIML

• Thanks to the special form of B and no exclusion restrictions in the
endogenous variable regression, there is a closed-form solution to the
LIML estimator (see Greene’s textbook for details).

Let Υ′ Z1 X1]. Then,

(γ1, 1)’ = [Υ′ IT – ̂ MZ) Υ]-1 Υ′ IT – ̂ MZ) y1 (*)
where ̂ is the is the smallest characteristic root of W1W-1, with:
W1 = ′ and W= ′ .

• The LIML estimator (*) is a K-class estimator (Theil (1961)). Note:

 2SLS estimator is a K-class estimator with ̂ = 1,
 OLS estimator is a K-class estimator with ̂ = 0.
 LIML = 2SLS when the equation is just identified.

54
RS – Lecture 8

Weak Instruments: LIML

• This estimator is proposed by Anderson and Rubin (1949, 1950). It
is the ML counterpart of the 2SLS.

• Under some assumptions, LIML and 2SLS have the same

asymptotic distribution. But, in finite samples, they can differ.

• It turns out that LIML is a linear combination of the OLS and 2SLS
estimates (with the weights depending on the data), and the weights
happen to be such that they approximately eliminate the 2SLS bias.

Note: In the presence of “many instruments” (using group asymptotics)

2SLS is inconsistent, but LIML still is consistent.

• More in Lecture 16, in the context of SEM.

T. W. Anderson (1918 – 2016, USA)

Weak Instruments: LIML

55
RS – Lecture 8

Weak Instruments: Finance application

Example: The consumption CAPM.
After (many) assumptions, excess returns for a risky asset are a
(linear) function of the covariance of the asset’s returns with
consumption growth:
Et [rt+1] – rf = γσrΔ – σr2/2
where rt+1= ln(1+Rt+1) - Rt+1: return on a risky asset.
2
σr = Var[ln(1+Rt+1)] = Var(rt+1)
σΔ2 = Var[ln(ct+1) – ln(ct)]
σrΔ = Cov[ln(ct+1) – ln(ct), rt+1]
γ = Risk aversion coefficient from a CRRA utility function.

• The C-CAPM is easy to test using linear regressions.

• There is also a non-linear version of the C-CAPM.

Weak Instruments: Finance application

• In both linear and nonlinear versions of the model, IVs are weak --
see Neeley, Roy, and Whiteman (2001), and Yogo (2004).

• In the linear model in Yogo (2004):

X (endogenous variable): consumption growth
Z (the IVs): twice lagged nominal interest rates, inflation,
consumption growth, and log dividend-price ratio.

• But, log consumption is close to a random walk, consumption

growth is difficult to predict. This leads to the IVs being weak.
 Yogo (2004) finds F-statistics for H0: П = 0 in the 1st stage
regression that lie between 0.17 and 3.53 for different countries.

56
RS – Lecture 8

Weak Instruments: Remedies (Greene)

• Symptom: The relevance condition, plim(Z’X/T ) ≠ 0, is close to being
violated.
• Remedy:
– Not much – most of the discussion is about the condition, not
what to do about it.
– Pick your best instrument and report just-identified results.
– Use LIML? Requires a normality assumption. Probably, not too
restrictive.

Weak Instruments: Testing

• Irrelevant IVs –i.e., П=0– and weak IVs bias the IV estimation.
Under Weak IVs, conventional asymptotics fail (see, Staiger and
Stock (1997).

• Small simulation (replications = 2,000).

- Simple case: one endogenous variable, one IV. Parameters
Z, , V ~ N(0, Σ . Set unit variances, but Cov(,V)=ρ
=1; l=1 & 5; T = 100 & 1,000; ρ = .99 & .30
- Compute t =(b2SLS – 1)/SE(b2SLS)
- We determine empirical size of 5% t-test (check|t2SLS|>1.96)
- We study 3 cases:
1) Strong instruments (when l=1, π = 1; when l=5, π’ = [1 1 0 0 0])
2) Weak instruments (when l=1, π = .1; when l=5, π’ = [.1 .1 0 0 0])
3) Irrelevant instruments π 0 (approximated by .0001)

57
RS – Lecture 8

Weak Instruments: Testing

Empirical size of 5% t-test (b2SLS )
T = 100 T = 100
Quality of IV l=1 (ρ=.99) l=1 (ρ=.30) l=5( ρ=.99) l=5 (ρ=.30)
Strong .065 (0.99) .044 (1.00) .088 (1.02) .051 (1.01)
Weak .195 (1.31) .006 (2.26) .852 (1.70) .057 (1.22)
Irrelevant .633 (2.01) .001 (1.40) .995 (1.99) .045 (1.28)

Empirical size of 5% t-test (b2SLS )

T = 1,000 T = 1,000
Quality of IV l=1 (ρ=.99) l=1 (ρ=.30) l=5( ρ=.99) l=5 (ρ=.30)
Strong .051 (1.00) .002 (1.00) .049 (1.00) .050 (1.00)
Weak .093 (0.79) .002 (0.93) .257 (1.13) .059 (1.05)
Irrelevant .631 (2.01) .004 (0.60) .995 (1.99) .043 (1.31)

Weak Instruments: Testing

• In the presence of weak IVs, the usual tests have size problems.
They are also not asymptotically pivotal: the distribution depends on
nuisance parameters (ρ, П) that cannot be consistently estimated.

• Anderson and Rubin (1949) propose a test of H0:  = 0, the AR

stat, an F-test, that has good properties under the usual situations
encountered under IV estimation.

• Intuition of AR test.
- Subtract from model Y0:
y – Y0 = Y  – Y0 +  = Y ( – 0 ) + 
- Substitute 1st stage:
y – Y0 = (ZП + V) ( – 0) + 
= ZП ( – 0 ) + V ( – 0 ) + 
= ZΦ + W

58
RS – Lecture 8

Weak Instruments: Testing

y – Y0 = ZΦ + W
where
Φ = П ( – 0 )
W = V ( – 0 ) + 
Now, we can estimate Φ using OLS, since Z is uncorrelated with W.
Note that testing H0: Φ = 0  H0:  = 0.

• The AR stat is the usual F-stat for testing Φ = 0. Under the usual
assumptions (fixed regressors and normal errors), the AR stat follows
the usual F distribution.

• Under weak instruments –see Staiger and Stock (1997):

l ∙ AR = l ∙ F(Φ=0) → χ2(l).

Weak Instruments: Testing

Note:
- Note that under H0:  = 0, the F-stat does not depend on П.
- The AR test is a joint test. It tests the joint hypothesis  = 0 and Z
is uncorrelated with .

• It turns out that the power of the AR test is not very good when
l>1. The AR test tests whether Z enters the (y – Y0) equation. The
AR test sacrifices power: It ignores the restriction Φ = П ( – 0 ).

• Low power leads to very wide CIs based on such tests. Kleibergen
(2002) and Moreira (2001) propose an LM test whose H0 rate is
robust to weak IVs. (LM test first estimates П under H0:  = 0.)

59
RS – Lecture 8

Weak Instruments: Testing

• There is an interesting literature on constructing CI under garbage
instruments. That is, CI that are robust to the presence of weak and
irrelevant instruments. See Staiger and Stock (1997) and Kleibergen
(2002, 2003).

Example: It is possible to invert the AR stat to get a weak

instrument robust CI interval for :
CIα = {0 : l ∙ AR ≤ χ2α (l)}

Weak Instruments: Pre-testing

• If one uses an F-test to detect weak IVs as a pre-test procedure, then

the usual pre-testing issues arise for subsequent inference –see Hall,
Rudebusch, and Wilcox (1996).

• In practice, researchers do tend to inspect and report the strength

of the first stage. Tests and CI will not have the appropriate nominal
size.

• Chioda and Jansson (2006) propose a similar statistic to the AR-stat

to build a C.I.. It has a non-standard distribution conditional on the 1st
stage F-stat. The C.I. are wider than the ones that do not condition
on the 1st stage.

60
RS – Lecture 8

2SLS Bias with Many Instruments

• Let’s go back to b2SLS = (Y' Pz Y)-1 Y'Pz y1
and look at the bias: E[b2SLS – ] = E[(Y' Pz Y)-1 (П'Z‘ + V') Pz ].
This expectation is hard to evaluate because the expectation operator
does not pass through the inverse (Y' Pz Y)-1, a nonlinear function.

• New Tool: Group asymptotics (or “many instruments asymptotics”). We use

an asymptotic argument but, now, we allow l (number of instruments,
“the group”) to grow at the same rate as the sample size, T.

Group asymptotics assumes a condition:

lim
→
If α = 0, then standard aysmptotics (Lecture 7) applies –i.e., l is fixed.

2SLS Bias with Many Instruments

Intuition:
Lots of instruments  “instruments weak.”

The number of first-stage parameters (the πj’s) grows with l. Many

parameters become “incidental,” creating consistency problems.

• Bekker (1994) and Newey and Smith (2004) show that GMM-type
approaches to estimating structural parameters using instrumental
variables, which include IV and 2SLS, may have substantial bias when
l is not small relative to T.

• While 2SLS estimator is inconsistent in this many instruments

environment, other estimators, including LIML & jackknife IV
(JIVE), remain consistent and asymptotically normal.

61
RS – Lecture 8

2SLS Bias with Many Instruments

• Group asymptotics gives us something like an expectation, but we
can take these expectations through non-linear functions:
E[b2SLS – ] ≈ E[(Y' Pz Y)-1] E[(П'Z' )]+ E[(Y' Pz Y)-1] E[V'Pz ]
Since E[(П'Z' )] = 0, E[b2SLS – ] ≈ E[(Y' Pz Y)-1] E[V'Pz ]

• Substituting in the first stage:

E[b2SLS - ] ≈ E[(П'Z‘ + V')' Pz (ZП + V)]-1 E[V'Pz ]
= E[[П'Z' ZП] + E[V'PzV]]-1 E[V'Pz ]

• Recall properties of trace (tr):

- tr: linear operator, goes through E[.], invariant to cyclic permutations
- V'PzV is a scalar,  V'PzV = tr(V'PzV)
- tr( Pz) = rank(Pz)= l since Pz is is and idempotent matrix.

2SLS Bias with Many Instruments

• Then,
E[V'PzV] ≈ E[tr(V'PzV)] = E[tr(PzVV')] = tr(Pz E[( VV')]
= tr(Pz σVV I) = σVV tr(Pz) = σVV l.

• Similar results applies to E[V'Pz ] = σεV l.

• Then, E[b2SLS – ] ≈ σεV l E[(П'Z' ZП) + σVV l]-1

= σεV /σVV E[(П'Z' ZП)/(σVV l )+1]-1

Note that F = [E[(П'Z' ZП)]/l ]/[σVV ] is the population F-statistic

for H0: П = 0 in the 1st stage regression. Thus,
E[b2SLS – ] ≈ σεV /σVV [1/(F+1)]

62
RS – Lecture 8

2SLS Bias with Many Instruments

• Thus,
E[b2SLS – ] ≈ σεV /σVV [1/(F+1)]

• Suppose the 1st stage coefficients, П, are zero. Then, F = 0

 E[b2SLS – ] ≈ σεV /σVV = σεV /σYY (OLS bias!)

Intuition: If П = 0, then any variation in X̂ in the sample comes from

V. The variation in X̂ is not different from the variation in X.

Note: This bias can affect tests, for example, the Hausman test.

2SLS Bias with Many Instruments

• From above:
F = E[П'Z' ZП]/(σVV l )
E[b2SLS – ] ≈ σεV /σVV [1/(F+1)]
Remarks:
- If П ≠ 0, but F-stat is small, then 2SLS will be biased towards OLS.
- Weak instruments: Instruments with small F-stat.
- The weak instrument bias tends to get worse as we add more (weak)
instruments (by adding IVs with no explanatory power, the only thing
changing in F is l!). These irrelevant IVs are referred as “garbage
instruments.”
- As we add IVs, X̂ gets closer to X. Then, 2SLS becomes OLS.
- If the IVs are very relevant (F ⟶ ∞), the IV bias goes to 0.
“One good instrument is better than 50 garbage instruments.”

63
RS – Lecture 8

Excessive Overidentification
• Situation: l is much larger than k. Possible “overidentification.”

• Extreme case: Suppose l =T. In this case, Z is a square matrix:

b2SLS = [W'Z'X]-1 W'Z'y = [X'Z(Z’Z)-1Z’X ]-1 X'Z(Z’Z)-1Z’y
= [X'Z Z-1 Z'-1 Z’X ]-1 X'Z Z-1 Z'-1 Z’y = [X'X ]-1 X'y = b

Since b is inconsistent and biased when E[ε|X]≠0, then so is b2SLS.

• While nobody will set l=T, a similar finite sample bias occurs in less
extreme cases. In general, as l → T, we see that b2SLS → bOLS.

Note: For the IV asymptotic theory to be a good approximation, T

must be much larger than l (say, T – l > 40 & grow linearly with T.)

Excessive Overidentification
• Angrist and Pischke (2009) report that “just-identified 2SLS is
approximately unbiased.” They report a simulation with weak IVs, using
OLS, just-identified IV and 2SLS with 20 IVs (and LIML too):

64
RS – Lecture 8

Excessive Overidentification
• Angrist and Pischke (2009) also report coverage rates of 95% C.I.
(The coverage rate is the probability that a C.I. includes the true
parameter.) Coverage rates for OLS and 2SLS are poor.

• Using lots of instruments can bias estimation (too many weak

instruments) and cause innacurate asymptotic approximations.

Excessive Overidentification (Greene)

• Obvious symptom: Z has many more columns than X
– 1st stage of 2SLS almost reproduces X
– 2d stage of 2SLS becomes OLS, which is biased.

• Detection:
– Visual – there is no test.
– Check b2SLS and bOLS. It they are similar, check that this is
not a result of “too many IVs.”

• Remedy:
– Fewer instruments? (Several methodological problems with this
idea). Donald and Newey (2001) consider this option.
– Jackknife estimation –see Ackerberg and Devereux (2009).

65
RS – Lecture 8

Instrument Exogeneity: Detection & Remedies

• Symptom: The valid condition, plim(Z’ε/T )=0, is close to being
violated. Since the errors are not observed, very difficult to check.

• The valid condition is an exclusion restriction on the model:

y =Y+Z +
Y =ZП+V
 The exclusion restriction imposes H0: = 0.

• If the exclusion is incorrect –i.e., θ = θ0 0–,  will show an

omitted variables bias problem. In the simple one exogenous
variable & one IV case, it is easy calculate the bias:
bIV = β + θ0/
The smaller , the bigger the bias (the bias is worse with weak IVs).

Instrument Exogeneity: Detection & Remedies

• Detection of instrument exogeneity:
– Endogenous IV’s: Inconsistency of bIV that makes it no better
(and probably worse) than bOLS.
– Durbin-Wu-Hausman test: Endogeneity of the problem
regressor(s). But, DWH tests do not have good properties in the
presence of weak instruments.

• Remedy:
– Some modifications of the DWH have been suggested under
weak instruments, see Hahn and Hausman (2002, 2005).
– Avoid endogenous weak instruments.
– General problem: It is not easy to find good instruments in
theory and in practice. Find natural experiments.

66
RS – Lecture 8

IV: Remarks (Greene)

• Finding good instruments –i.e., meet both conditions- is not easy.

• When only the relevant condition is emphasized, OLS can be better

than IV (“the cure can be worse...”). Even “clever” IVs can have low
correlations with X and create severe finite-sample bias. The bias
tends to be worse when there are many overidentifying restrictions (l
is large relative to k).

• For the simple case of one endogenous variable, the F-stat in the
1st stage can help to identify weak IVs. With many IVs, Stock and
Yogo (2005) provide rules of thumb regarding the weakness of the
IVs based on a statistic due to Cragg and Donald (1993).

• Large T will not help. A&K and Consumption CAPM tests have
very large samples!

IV: Remarks (Greene)

• Just identified IV is approximately unbiased (or less biased) even
with weak instruments (although it is not possible to see this from
the bias formula.)

67
RS – Lecture 8

What should you do in practice? (Pischke)

• Report 1st-stage and think about whether it makes sense. Are the
magnitude and sign as you would expect?
• Report the F-stat on the excluded IVs. The bigger this is, the better.
Fs above 10 to 20 are considered relatively safe, lower Fs put you in
the danger zone.
• Pick your best single IV and report just-identified estimates using
this one only. Just-identified IV is approximately median-unbiased.
• Check over-identified 2SLS estimates with LIML. If the LIML
estimates are very different, or SEs are much bigger, worry.
• Check coefficients, t-stats, and F-stats for excluded IVs in the
reduced-form regression of dependent variables on IVs. The
reduced-form estimates are just OLS, so they are unbiased. If the
relationship you expect is not in the reduced form, it is probably not
there.

IV: Final General Remarks

• Finding good instruments is not easy. A good natural experiment,
which defines the IV, is worthy of a paper.

• Angrist: “Tell a story about why a particular IV is a good

instrument.” In the omitted variable case: “Does the IV, for all
intents and purposes, randomize the endogenous regressor?”

• IV models can be very informative, but it is your job (as author of

the paper) to convince the audience.

• Good IV models are generally interesting in their own right, and

should not be treated as “robustness” checks.

• The emphasis on IV with natural experiments is part of the quasi-

experimental revolution, which shifted the emphasis in applied
economics (and finance) from theory to empirical experiments.

68
RS – Lecture 8

IV: Final General Remarks

• This shift in mainly microeconomics (and, lately in other areas of
economics and finance) from theory to empirical experiments: “From
Mas-Colell to Angrist and Pischke.”

• Criticism to IV estimation using natural experiments:

- No theory. For example, there is no optimizing labor (structural)
model behind A&K (1991). Q: Where does the reduced form
equation come from?
- Difficult to interpret the results. For example, LATE is an average,
which may or may not contain information about the parameter of
interest. It may not be useful (see Heckman and Urzua (2009), for an
example, where LATE is not informative).

IV: Final General Remarks

• Answer from natural experimentalists to the criticism: Big
skepticism about structural models. Thus, no modeling is good!
Angrist and Pischke (2010): “The explosion of IV methods, including
LATE estimation, has led to greater “credibility” in applied econometrics.”

• References:
- “Mostly Harmless Econometrics: An Empiricist's Companion,”
textbook by Angrist and Pischke (2009).
- “Instruments, Randomization, and Learning about Development,”
by Deaton (2010, JEL).
- “The Empirical Economist's Toolkit: From Models to Methods,” by
Panhans and Singleton (2015, Duke Working Paper).

69