You are on page 1of 57

Econometrics

for M.Sc. and doctoral students


Amelie Wuppermann
amelie.wuppermann@econ.lmu.de

Department of Economics
University of Munich
Winter Semester 2014/15

3. Instrumental variables
3.1

Two-stage least squares and IV estimation

3.2

Some applications of IV estimation

3.3

Asymptotic properties of the IV estimator

(W 5.2.15.2.5)

3.4

IV estimation in practice

(W 5.2.6)

3.5

Tests in the IV model

(W 6.3.1, 6.3.2)

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

(W 5.1.1, 5.1.2)

31

3.1 Two-stage least squares and IV estimation


We begin by writing down the population model (notation as
before),
y = 0 + 1 x1 + 2 x2 + + K xK + u .
(1)
Assume cov(xj , u) = 0 for j = 1, . . . , K 1. Thus, the variables
x1, . . . , xK1 are exogenous. Also, we maintain that E(u) = 0.
We now consider the case where cov(xK , u) 6= 0. The source of
this endogeneity could be a left-out variable, for instance.
Without additional information, we cannot estimate the
parameters = (0, 1, . . . , K ) consistently.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

32

3.1 Two-stage least squares and IV estimation


An instrumental variable (an instrument) is a variable, z1, that
is not in the population model and that satisfies two conditions:
1. cov(z1, u) = 0 (i. e., z1 is exogenous in the population), and
2. z1 is partially correlated with xK (i. e., z1 is relevant).
The second condition can be stated formally using a linear
projection,
xK = 0 + 1x1 + + K1xK1 + 1z1 + rK .

(2)

The instrument must satisfy 1 6= 0, i. e., z1 is partially correlated


with xk , after controlling for the other exogenous variables,
x1, . . . , xK1.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

33

3.1 Two-stage least squares and IV estimation


The familiar statement, the instrument must be correlated with
the endogenous variable, holds only in the special case in which
there is only one explanatory variable in the population model
(which then is also the endogenous variable, of course).
Put differently, the second condition, 1 6= 0, is equivalent to
cov(z1, xK ) 6= 0 if K = 1. Why?
Note that the exogenous variables x1, . . . , xK1 all satisfy the
two conditions for instruments and can thus act as their own
instruments.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

34

3.1 Two-stage least squares and IV estimation


Wooldridge calls the linear projection of the endogenous variable,
xK , on all instruments, x1, . . . , xK1, z1 (and a constant), the
reduced form of xK .
Often this is instead called the first stage regression. Well see
why in a second.
Note that the term reduced form is also used differently in
the IV context. It typically refers to the regression of y on
x1, . . . , xK1, z1.
But well stick to Wooldridges terminology for now.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

35

3.1 Two-stage least squares and IV estimation


Heres the intuition for how instrumental variables work. Plugging
the reduced-form equation
xK = 0 + 1x1 + + K1xK1 + 1z1 + rK
into the population model yields
y = 0 + 1x1 + 2x2 + + K1xK1

+K (0 + 1x1 + + K1xK1 + 1z1 + rK ) + u

y = 0 + 1x1 + . . . K1xK1 + 1z1 + v ,

(3)

where v = u + K rk , j = j + K j , 1 = K 1 .
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

36

3.1 Two-stage least squares and IV estimation


Are the OLS assumptions satisfied in (3)?
Yes! We can verify that v = u + K rk is uncorrelated with all
regressors in this equation (x1, . . . , xK1, z1).
Moreover, we can recover the structural parameters from
j = j + K j and 1 = K 1.
The linear projection (2), i. e., the reduced-form of xK , can be
estimated because it contains only observable variables.
All model variables (i. e., y, any element of x, and z) may be
continuous, discrete, or even binary.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

37

3.1 Two-stage least squares and IV estimation


This procedure yields the two-stage least squares estimator
(2SLS or TSLS):
1. Stage: Estimate the reduced-form equation (2). Obtain the
predicted values of the endogenous variable.
2. Stage: Replace the original values of the endogenous variable
with the predicted values in the structural model (1) and
estimate the model.
Well discuss the statistical properties of this procedure in detail
later (section 3.3).
Moreover, well shortly extend it to the case where we have more
than one instrument per endogenous variable.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

38

3.1 Two-stage least squares and IV estimation


Next, we derive the (one-step) instrumental variables estimator.
y = x + u
structural model
x = (1, x2, . . . , xK )
structural explanatory variables
z = (1, x2, . . . , xK1, z1) exogenous variables
E(xu) 6= 0, but by assumption on the instrument z1, E(z u) = 0.
We can now proceed as in the derivation of the OLS estimator:
z y = z (x + u)
E(z y) = E(z x)
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

39

3.1 Two-stage least squares and IV estimation


Under the assumption rank E(z x) = K, we obtain
= [E (z x)]

E(z y)

As before, this equation holds in the population. It tells us that


is identified under our assumptions.
How does this relate to our earlier derivation of 2SLS?

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 10

3.1 Two-stage least squares and IV estimation


Next, we need an estimator of . We apply the analog principle,
i. e., we replace population expectations with their sample analogs
(sample averages) for a random sample i = 1, . . . , N .
Heres the instrumental variables estimator:
!1
!
N
N
X
X
1
1

b =
z i xi
zi yi
N i=1
N i=1

= (Z X)

(Z Y )

(4)

The instrumental variables estimator can be obtained in one step,


but it is equivalent to the two-step 2SLS estimator.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 11

3.1 Two-stage least squares and IV estimation


We have found two equivalent ways to estimate the structural
parameters when one of the structural explanatory variables is
endogenous and we have an instrument for it.
The assumptions are the same in both approaches:
1. cov(z1, u) = 0 (i. e., z1 is exogenous in the population), and
2. z1 is partially correlated with xK (i. e., z1 is relevant).
There is a major difference between these assumptions:
The first assumption is untestable. (Why?)

The second assumption can be tested. (How?)


Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 12

3.1 Two-stage least squares and IV estimation


The next level of generality is to allow for multiple instruments
per endogenous variable.
As before, xK is endogenous.
z1, . . . , zM , that satisfy

We have M instruments,

1. cov(zh, u) = 0 , h = 1, . . . , M (exogeneity), and


2. z1, . . . , zM are partially correlated with xK (relevance).
We could compute M different IV estimates using the procedures
discussed above. But can we use all M instruments at once?
Now we really need the two-step 2SLS approach!
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 13

3.1 Two-stage least squares and IV estimation


Define z = (1, x1, . . . , xK1, z1, . . . , zM ) as the vector of all
exogenous variables. (There are K 1 + M plus the constant.)
Define the linear projection of the endogenous variable on z as
xK = 0 + 1x1 + + K1xK1
+1z1 + + M zM + rk .

(5)

By definition, the systematic component of this projection,


xK = 0 + 1x1 + + K1xK1
+1z1 + + M zM ,

is uncorrelated with u in the structural model. (Why?)


Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 14

3.1 Two-stage least squares and IV estimation


Note that xK satisfies the conditions for a valid instrument.
While xK is an artificial variable defined by a population
projection, we can estimate it by simply estimating (5) by OLS
this is the first-stage (reduced form) regression.
bi = (1, x1i, . . . , xK1,i, x
We can then compute the vectors x
bKi)
bi and yi for i = 1, . . . , N , we
for each observation i. Stacking x
obtain the second stage of the 2SLS estimator (for a proof of the
second equality, see Wooldridge, p. 97):

1 
 
1 

X
Y = X
X

Y .
b = X
X
X

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 15

3.1 Two-stage least squares and IV estimation


Summary of the IV approach
To summarize, the 2SLS procedure consists of these steps:
1. Obtain the predicted value of the endogenous variable, x
bKi,
from the first-stage regression of xK on all exogenous variables.
bi = (1, x1i, . . . , xK1,i, x
2. Regress yi on x
bKi).

It is crucial that the first-stage regression include the exogenous


structural variables as well (and not only the instruments
z1, . . . , zM ) because in a technical sense, x1, . . . , xK1 are
instruments, too!
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 16

3.1 Two-stage least squares and IV estimation


As in the case with one exogenous instrument, the second IV
assumption (relevance) can be tested. In the first-stage regression,
test H0 : 1 = = M = 0 using an F -test.
One instrument would be sufficient to achieve identification of
the structural parameters.
In general, if we have more instruments than endogenous variables
(here, if M > 1 since we have only one endogenous variable), then
we have over-identification (otherwise, its just-identification).
Below, we will see specification tests for IV models that can be
applied when there is over-identification.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 17

3.1 Two-stage least squares and IV estimation


Proxy variables vs. instrumental variables
Both proxies and instruments provide potential solutions to the
left-out variables problem.
Both proxies and instruments must be redundant in the structural
equation.
But there is a major difference:
A proxy should be highly correlated with the left-out variable

(so that it can be used as a stand-in for the left-out variable).


An instrument must be uncorrelated with the left-out variable
(since it must be uncorrelated with the error term).
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 18

3.2 Some applications of IV estimation


In microeconomic applications, it is typically hard to find a suitable
instrument.
In this section, well consider two prominent examples:
A (conceptually) simple example is the reaction of market

demand to exogenous variations in price.

A more difficult example is the estimation of the returns to

schooling.

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 19

3.2 Some applications of IV estimation


Market demand
Suppose we want to estimate the response of market demand to
exogenous changes in prices.
Why is this difficult?
Prices are not exogenously given since they are partially
determined by market demand.
What would be a suitable instrument?
A variable that is correlated with price but does not directly affect
the quantity demanded.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 20

3.2 Some applications of IV estimation


We could use a variable that affects market supply, since this also
affects prices but does not directly affect demand.
For agricultural goods, this could be weather conditions.
This idea and thus the principle of IV estimation was discovered
as early as 1928 by the agricultural economist, Philip G. Wright.
If youre interested, this story is in the textbook by Stock &
Watson (2007, p. 424426).

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 21

3.2 Some applications of IV estimation


Returns to schooling
Why are we interested in the returns to schooling?
The typical approach is to estimate a (log) wage equation,
log(wage) = 0 + 1exper + 2exper2 + 3educ + abil + .
Ability cannot be observed, and we already know that the
estimates of the parameters are inconsistent if ability is
correlated with the other explanatory variables (say education).
There are two potential solutions: using a proxy variable or an
instrumental variable.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 22

3.2 Some applications of IV estimation


The requirements for an instrument in the returns to education
setting can be derived from the IV conditions.
Condition 1: exogeneity
The instrument must be uncorrelated with ability (and with )
Condition 2: relevance
The instrument must be partially correlated with years of
education (i. e., after controlling for all other exogenous variables)
The instrument must also be redundant, i. e., it must not have
a direct influence on the market wage. (Otherwise, it would be
a left-out variable which would effectively result in a violation of
condition 1.)
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 23

3.2 Some applications of IV estimation


Many instruments have been discussed in the literature. Here are
some examples.
Parents education (sometimes only mothers education)
Proximity of a college or a university (Card, 1995)

Month or quarter of birth (Angrist & Krueger, 1991)

For each of these potential instruments, we can discuss whether


conditions 1 and 2 are satisfied. Either of these conditions can be
problematic!
IQ is a proxy for ability, but cannot be used as an instrument!
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 24

3.3 Asymptotic properties of the IV estimator


We now consider the general case. The population model is
y = x + u ,

(6)

where x is a 1 K vector, including a constant.


Several elements of x may be correlated with u.
The 2SLS assumptions are listed on the next page. The common
structure with, but also the differences to, the OLS assumptions
should be obvious.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 25

3.3 Asymptotic properties of the IV estimator


Assumption 2SLS.1: E(z u) = 0.
In words, u has mean zero and is uncorrelated with the 1 L
vector z that contains a constant and all the exogenous variables.
In general, K 6= L.
Assumption 2SLS.2a: rank E(z z) = L.
In words, there is no linear dependence (multicollinearity)
among the exogenous variables. This is a standard assumption in
multivariate regression models.
Assumption 2SLS.2b: rank E(z x) = K.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 26

3.3 Asymptotic properties of the IV estimator


Assumption 2SLS.2b (the so-called rank condition) is whats
substantively new in the IV setting compared to OLS.
A necessary condition for rank E(z x) = K is that L K, i. e.,
we have at least as many instruments including the exogenous
elements of x as there are elements in x (order condition).
In the case with more than one endogenous variable in x, the
rank condition is difficult (but not impossible) to test.
As a remark, if x = z (i. e., if all elements of x are exogenous),
then assumptions 2SLS.1 and 2SLS.2a/b boil down to our earlier
assumptions OLS.1 and OLS.2.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 27

3.3 Asymptotic properties of the IV estimator


Theorem (consistency of the 2SLS estimator)
Under assumptions 2SLS.1 and 2SLS.2a/b, the 2SLS estimator

b =

n
X
i=1

n
X
i=1

xizi

xizi

n
X
i=1

n
X
i=1

zi zi

!1

zi zi

!1

n
X
i=1

n
X
i=1

!1

zi xi
zi yi

(7)

obtained from a random sample is consistent for the parameter


vector of the structural model (6).
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 28

3.3 Asymptotic properties of the IV estimator


The proof is analogous to the consistency proof for the OLS
estimator. (You can do the proof as an exercise, if you want.)
Theres a generic structure of consistency proofs in econometrics:
b = + stuff.
Write

Re-arrange so that stuff involves only the right-hand side

variables (in this case, x, z, u).


Use the assumptions (here, 2SLS.1 and 2SLS.2a/b) to show
that something can be inverted or vanishes.
Apply a law of large numbers and Slutskys theorem.

Note that we once again assume that we have a random sample


of the population so that the observations are i.i.d.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 29

3.3 Asymptotic properties of the IV estimator


Assumption 2SLS.3: E(u2z z) = 2 E(z z), where 2 = E(u2),
i. e., the error terms are homoskedastic.
Theorem (asymptotic normality of the 2SLS estimator)
b
Under assumptions 2SLS.1 through 2SLS.3, the 2SLS estimator, ,
is asymptotically normally distributed. Specifically,


n
o1

a
1
2

b
.
N N 0, E(x z) [E(z z)] E(z x)
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 30

3.3 Asymptotic properties of the IV estimator


The proof is similar to the proof of the asymptotic normality of
the OLS estimator.
The variance matrix stated in the theorem can be estimated using
sample analogs (averages) for the expectations involving x and

z, and

2

P
N
1
b .
the formula
b2 = N K
i=1 yi xi

Important: The 2SLS variance matrix is not the variance matrix


of the OLS estimator in the second step of a two-step procedure
(since u
bi = yi xib is not the OLS residual from the second
step).

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 31

3.3 Asymptotic properties of the IV estimator


Without the homoskedasticity assumption 2SLS.3, the 2SLS
estimator is still asymptotically normal, but with a different (more
complicated) variance matrix.
The heteroskedasticity-robust variance formula is a generalization
of the OLS robust variance formula, see Wooldridge, p. 106,
equation (5.34).
Alternatively, we could use a weighted estimator similar to WLS
if 2SLS.3 fails.

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 32

3.3 Asymptotic properties of the IV estimator


Theorem (relative efficiency of the 2SLS estimator)
Under assumptions 2SLS.1 through 2SLS.3, the 2SLS estimator is
efficient in the class of all instrumental variables estimators using
instruments linear in z.
The proof is omitted (see Wooldridge, p. 103 if you are interested).
If assumption 2SLS.3 does not hold (i. e., in the presence of
heteroskedasticity), more efficient (GMM-type) estimators are
available. (This is similar to using WLS in the OLS set-up with
heteroskedastcity.)
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 33

3.3 Asymptotic properties of the IV estimator


Finite-sample bias of the 2SLS estimator
We have shown that both OLS and 2SLS (IV) are consistent
under the respective assumptions.
However, as shown on the next page, there is a crucial difference:
OLS is unbiased as long as E(u|x) = 0.

2SLS is not unbiased even if E(u|z) = 0.

The expectation of the 2SLS estimator may not even exist.


(Kinal, 1980) showed that the number of moments of the 2SLS
estimator that exist is one less than the number of over-identifying
restrictions.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 34

3.3 Asymptotic properties of the IV estimator


The expectation of the OLS estimator is


E bOLS

= + E (X X) X u
h
i
1
= + E (X X) X E(u|X) = .

This holds because by OLS.1, E(u|x) = 0 in the population.


The expectation of the 2SLS or IV estimator is


E bIV

= + E (Z X) Z u
h
i
1
= + E (Z X) Z E(u|Z, X) 6= !

The problem is that 2SLS.1 ensures that E(u|z) = 0, but says


nothing about E(u|z, x) which generally is nonzero.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 35

3.3 Asymptotic properties of the IV estimator


OK, so the IV estimator is biased under assumption 2SLS.1. But
now we are a little bit confused how can it then be consistent?
The consistency proof works slightly differently:

1 
!
1
1

b
plim IV = plim +
(Z X)
(Z u)
N
N

That is, for consistency we essentially only need to show that


p
1

N (Z u) 0 which is implied by assumption 2SLS.1.


The underlying issue is that the plim and expectation operators
behave differently.

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 36

3.4 IV estimation in practice


Good things to say about instrumental variables and the 2SLS
estimator if assumptions 2SLS.1 through 2SLS.3 hold:
Instrumental variables identify the structural parameters were

interested in even when explanatory variables are endogenous.


2SLS has attractive asymptotic properties.

Not so good things about using instruments and 2SLS in practice:


We already know that its hard to find valid instruments.
2SLS is biased in finite samples.

Weak instruments exacerbate this problem (as well see next).

2SLS estimates are less precise than OLS estimates.


Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 37

3.4 IV estimation in practice


Weak instruments
Even if samples are large so that asymptotic arguments (i. e.,
consistency) are justified, weak instruments may cause trouble.
Heres a simple example. Consider y = 0 + 1x1 + u, with x1
endogenous.
We use z1 as an instrument for x1, with cov(x1, z1) 6= 0.
It follows from the IV consistency proof that in this setting

cov(z1, u)
u corr(z1, u)
b
plim 1,IV = 1 +
= 1 +
.

cov(z1, x1)
x1 corr(z1, x1)

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 38

3.4 IV estimation in practice


Obviously, when cov(z1, u) = 0, the IV estimator is consistent.
But even a small amount of correlation between z1 and u in the
sample causes an inconsistency, and this problem is exacerbated
by a small correlation between z1 and x1.
In this setting, a weak instrument, z1, is thus defined as having
a low correlation with the endogenous variable, x1. This problem
persists in more general settings.
When instruments are weak, it may be better to stick to OLS. To
see this, consider the corresponding formula for OLS:
u
b
plim 1,OLS = 1 +
corr(x1, u) .
x1

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 39

3.4 IV estimation in practice


How to detect weak instruments in practice:

Low R2 in the first-stage (reduced form) regression.

Low

F -statistic for joint significance of the excluded


instruments.
Staiger & Stocks (1997) rule of thumb:
F < 10 is problematic
F < 5 is a disaster
Famous example: Bound, Jaeger, and Bakers (1995) critique of
Angrist and Kruegers (1991) study.
Cruz and Moreira (2005) re-evaluate Bound et al. (1995) and
come to less negative conclusions about Angrist and Kruegers
(1991) results. See also Angrist and Pischke (2009), pp. 2136.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 40

3.4 IV estimation in practice


Angrist and Pischkes (2009, pp. 2123) guidelines for IV studies:
Report the first stage and think about whether it makes sense.
Report the F -statistic on the excluded instruments.
Pick your single best instrument and report just-identified
estimates using this one alone.
Check over-identified 2SLS estimates with LIML.
(This is the Limited Information Maximum Likelihood version
of IV; we do not cover LIML-IV in this course.)
Look at the coefficients, t-statistics, and F -statistics for
excluded instruments in the reduced-form regression of
dependent variables on instruments.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 41

3.4 IV estimation in practice


More helpful reading on weak instruments
Hands-on summaries of the literature:
Hahn and Hausman (2003)
Murray (2006)

Textbook tips:
Brief: Cameron and Trivedi (2005), Section 4.9.1

Extensive: Angrist and Pischke (2009), Section 4.6.4

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 42

3.4 IV estimation in practice


Large standard errors
A separate problem is that 2SLS standard errors are generally
larger than OLS standard errors. (This is intuitive.)
The inflation of standard errors can be substantial, as is shown
for a special case on the next page (Cameron & Trivedi (2005),
Section 4.9.3).
In practice, this often creates a difficult situation:
OLS estimates are inconsistent(, but precise).

2SLS estimates are consistent, but much less precise and often

even insignificant.

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 43

3.4 IV estimation in practice


Lets consider the case with one endogenous variable x and one
instrument z. Assume E(u|z) = 0. In this special case, we have


1
2
1

b
[
= (x z) z z (z x)
Avar IV


[ bOLS
Avar
2

/x x
=
=
,

2
(z x) /[(z z)(x x)]
rxz
where for simplification we use the somewhat unusual notation
x = (x1, . . . , xN ) and z = (z1, . . . , zN ).
1 < rxz < 1 is the sample correlation coefficient between x and
z.
Shea (1997) derives a similar formula for the general case.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 44

3.4 IV estimation in practice


Thus, (asymptotic) IV standard errors are larger than OLS
standard errors.
Why is this bad news?
Importantly, these differences between IV and OLS standard
errors are not related to the question whether we can correctly
estimate the standard errors in small samples (remember: this is
the problem with (cluster)-robust standard errors that we talked
about in chapters 2.3 and 2.9!).

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 45

3.5 Tests in the IV model


Consider the linear regression model y = x + u. Earlier in
this course, we stressed that the crucial independence assumption
E(u|x) = 0 that insures identification of the parameter vector
cannot be tested with data on y and x alone.
In this section, we consider two types of tests in the IV model:

Tests for exogeneity of an explanatory variable in the structural

model if we have a valid instrument: Hausman tests

Tests for exogeneity of the instruments if we have more

instruments than we need: overidentification tests

Important: We also need to make untestable assumptions to


apply these tests!
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 46

3.5 Tests in the IV model


Hausman tests: intuition
Jerry Hausman (1978) developed a general principle for the
construction of tests for the specification of regression models.
We will encounter other applications of this test principle later in
this course, and we will provide a more formal treatment there.
We first provide the basic intuition and then derive a simple
version of a Hausman test that can be used to test the exogeneity
of an explanatory variable.

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 47

3.5 Tests in the IV model


Consider a model with only one explanatory variable, x, that is
(potentially) endogenous. Assume we have a valid instrument, z.
bOLS and bIV should
If x is exogenous in the structural model,

be identical (technically, they have the same probability limit


since both are consistent).
bIV is still consistent but bOLS is not
If x is endogenous,
consistent any more, so the two estimates should be different
(they have different probability limits).
We could construct a test statistic for H0: bIV bOLS . This is
the general version of the Hausman test.

We consider a simpler version that is based on linear regression.

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 48

3.5 Tests in the IV model


Regression version of the Hausman test for endogeneity
Wooldridges notation in this section needs some getting used to.

Consider a structural model with one endogenous variable, y2:


y 1 = z 1 1 + 1 y 2 + u 1

We have at least one instrument for y2. The variable z contains


the exogenous structural variables z1 and all instruments for y2.
The reduced form of y2 (the first stage) is thus
y2 = z2 + v2 ,

where E(z v2) = 0 by construction of the linear projection.


Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 49

3.5 Tests in the IV model


The residuals of the two equations on the previous slide are
correlated, that is E(u1v2) 6= 0, if and only if y2 is endogenous.
Thus, we run an auxiliary regression,
y1 = z11 + 1y2 + 1vb2 + error ,

where vb2 is the residual from the reduced form regression for y2.

Our null hypothesis is 1 = 0. If rejected, then y2 is endogenous.


Practical problem: This test works only if we have a valid
instrument for y2 (which is untestable).
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 50

3.5 Tests in the IV model


Tests for over-identifying restrictions: intuition
Suppose we have two plausible instruments for one endogenous
variable. Then we can compute two IV estimates, each using only
one of the two instruments.
Then, under the null hypothesis that both instruments are valid,
p
b
b
IV,1 IV,2 0 .

We can use this insight to construct a Wald-type test statistic.

What if the null is rejected? Then we know that at least one of


the two instruments is not valid but we dont know which. And
of course, it might also be the case that neither is valid.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 51

3.5 Tests in the IV model


Regression version of the test for over-identifying restrictions
The structural model is slightly more general than before:
y 1 = z 1 1 + y 2 1 + u 1

(8)

The dimensions of the vectors we consider are as follows:


z1 : 1 L1 vector of exogenous variables (in the structural model)
y2 : 1 G1 vector of endogenous variables
z2 : 1 L2 vector of instruments (not in the structural model)
z = (z1, z2) : 1 L vector of all exogenous variables, L = L1 + L2
We assume that the model is over-identified, i. e., L2 > G1.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 52

3.5 Tests in the IV model


The null hypothesis is that all elements of z2 are exogenous.
In the first step, we estimate the structural model (8) by 2SLS.
In an auxiliary regression, we regress the residual obtained from
the first regression, u
b1, on all exogenous variables, z.

Intuition: Under H0, E(z u1) = 0. Thus, the exogenous variables


should not have any explanatory power in this auxiliary regression.
The test statistic is N Ru2 2Q1 , where Ru2 is obtained from the
auxiliary regression, N is the sample size, and Q1 = L2 G1.

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 53

3.5 Tests in the IV model


Theres a major problem: This test works only if we have more
instruments than we need (over-identification) and if we are
convinced that as many as we need for just-identification are
valid. (Again, identification essentially cannot be tested.)
For example, if we have two candidate instrumental variables
whose validity is equally dubious, this test is not helpful.
Tests for over-identifying restrictions are thus viewed with
increasing skepticism. For a formal discussion, see Parente and
Santos Silva (2012).
A good critical discussion of IV approaches to identification is
given by Deaton (2010).
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 54

References in addition to Wooldridges book


Angrist, J. D. and S. Pischke (2009): Mostly Harmless Econometrics.
Princeton: Princeton University Press.
Cameron, A. C. and P. K. Trivedi (2005): Microeconometrics. Cambridge,
UK, and New York: Cambridge University Press.
Cruz, L. and M. J. Moreira (2005): On the validity of econometric
techniques with weak instruments. Journal of Human Resources, 40(2), 393
410.
Deaton, A. (2010): Instruments, randomization, and learning about
development. Journal of Economic Literature, 48, 424455.
Hahn, J. and J. Hausman (2003): Weak instruments: Diagnosis and cures
in empirical econometrics. American Economic Review, Papers & Proceedings,
93(2), 118125.
Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 55

References in addition to Wooldridges book


Parente, P. M. D. C. and J. M. C. Santos Silva (2011): A cautionary
note on tests for overidentifying restrictions. Economics Letters, 115(2), 314
317.
Murray, M. P. (2006): Avoiding invalid instruments and coping with weak
instruments. Journal of Economic Perspectives, 20(4), 111132.
Shea, J. (1997): Instrument relevance in multivariate linear models: A simple
measure. Review of Economics and Statistics, 79(2), 348352.
Stock, J. H. and M. W. Watson (2007): Introduction to Econometrics,
2nd Edition. New York: Addison-Wesley.
Wright, P. G. (1928): The Tariff on Animal and Vegetable Oils. New York:
Macmillan.

Amelie Wuppermann, Econometrics, Winter Semester 2014/15

3 56

You might also like