Econ 140 (Spring 2018) - Section 12

∗
Caroline, Chris, Jimmy, Kaushiki, Leah

April 15, 2018

1 Instrumental Variables Regression

1.1 Intuition and general model
In the last classes we have discussed several problems (or internal threats) that make the error term
correlated with the regressor (Cov(X, u) 6= 0) such as omitted variables, errors in variables, and simul-
taneity causality. Instrumental variables (IV) regression is a nice way to address these issues. To
understand how IV works think of the variation in X as having two parts: one that is correlated with
u (the bad variation), and a second the part that is not correlated with u (the good variation).
Roughly speaking IV regression uses information from another variable Z (called instrument) to isolate
the good variation in X and disregard the bad variation (which biases the OLS estimates). As a
result, the IV estimate of the coecient on X will give us a consistent estimate of the eect of the
good variation part of X on Y.

An important thing to keep in mind is that for this approach to work we need to nd an instrument
Z that is related to X and unrelated to things we don't observe that might also change Y (so that Z
aects Y only through X ). The rst condition is empirically testable, but unfortunately this is not
true for the second condition.

The general IV regression model is

Yi = β0 + β1 Xi + β2 Wi + ui i = 1, ..., n
where:

• Yi is the dependent variable

• ui is the error term (representing measurement error and/or omitted factors)

• Xi is the endogenous regressor (Cov(X, u) 6= 0)

• Wi is the included exogenous regressor (Cov(W, u) = 0)
• Zi is the instrument, which satises:

 instrument relevance: Cov(Z, X) 6= 0

 instrument exogeneity: Cov(Z, u) = 0

The IV regression assumptions are now

1. E (ui |Wi ) = 0.
2. (Xi , Wi , Zi , Yi ) are i.i.d. draws from their joint distribution.

3. Large outliers are unlikely: The X 0 s, W 0 s, Z 0 s, and Y have nonzero nite fourth moments.

4. The two conditions for a valid instrument hold: instrument relevance and instrument exogeneity.
∗
We thank previous GSIs for the great section material they built over time. These section notes are heavily based
on their previous work.

1
1.2 Estimation: Two Stage Least Square (TSLS)
• First stage: run the regression Xi = π0 + π1 Zi + π2 Wi + vi using OLS and compute the predicted

value X
bi = π
b0 + π
b1 Zi + π
b2 Wi .

 Remark 1: Note that the rst stage decomposes X into two: a problematic component (vi )
correlated with ui and a problem-free component (π0 + π 1 Zi + π 2 Wi ) uncorrelated with ui .
 Remark 2: Instrument relevance means π
b1 is statistically signicant.

• Second stage: run the regression Yi = β0 + β1 X

bi + β2 Wi + ui using OLS in order to get βb0T SLS ,
βb1T SLS and βb2T SLS .

 Remark 3: The idea behind TSLS is to use the predicted problem-free component of Xi to
get a consistent estimate of β1 .
1

 Remark 4: When you run the 2-step procedure manually in Stata the standard errors are
incorrect because they do not recognize that it is the 2nd stage of a two-stage process.
Luckily, Stata corrects for this automatically when you use a canned command such as
ivregress.
 Remark 5: As before, the error u might be heteroskedastic, so it is important to use
heteroskedastic-robust standard errors.

1.3 Terminology
In general, there can be multiple endogenous regressors (X 0 s), multiple included regressors
(W s), and multiple instrumental variables (Z s). For IV regression to be possible there must be at
0 0
0 0
least as many Z s as X s. Regarding the number of instruments (m) and the number of endogenous
regressors (k ) we have the following terminology:

• m = k: regression coecients exactly identied

• m > k: regression coecients overidentied

• m < k: regression coecients underidentied

In the last case, we cannot estimate these coecients by IV regression.

1.4 Checking instrument validity (for m instruments)

For instrumental variables regression it is essential to assess whether a given set of instruments is valid.
Otherwise you can end up with meaningless results (and wrong policy prescriptions!).

• Instrument relevance:

Why is this important? Intuitively the more the variation in X is explained by the instruments,
the more info is available for use in IV regression (and the more accurate is the estimator too).

To test this, run the regression of the rst stage

Xi = π0 + π1 Z1i + ... + πm Zmi + πm+1 Wi + vi

and do the following hypothesis test (F-test):

H0 : π1 = π2 = ... = πm = 0 vs H1 : at least one πj 6= 0, j = 1, ..., m

In the case of a single endogenous regressor, if F-statistic < 10 ⇒ weak instruments.
2 TSLS

estimator is biased, and TSLS t-statistics and condence intervals are unreliable.

1
And what about β2 ? In this case β2 is already consistent because by assumption Wi is one of our good regressors,
in the sense that E (ui |Wi ) = 0.
2
Why 10? Take a look at Appendix 12.5.

2
 • Instrument exogeneity:

Intuitively, exogeneity of the instruments means that they are uncorrelated with ui .

 Exactly identied case: We cannot test exogeneity. But you can provide your personal
knowledge on the subject to defend your instrument (this is indeed a very common practice).

 Overidentied case: The overidentifying restrictions test (the J-statistic):

When the model is over-identied, we can use what is called the Sargan test, or the Hansen
test, or the J-test. In essence, the test uses information from k good IVs to test whether
the extra m−k IVs are valid or not. In other words, the test assumes there are at least k
good IVs.

∗ step 1: take the residuals from TSLS estimation (note: use Xi , not X
bi )
 
bTi SLS = Yi − βb0T SLS + βb1T SLS Xi + βb2T SLS Wi
u
∗ step 2: estimate the following regression using OLS (using homoskedastic-only standard
error estimation):
bTi SLS = δ0 + δ1 Z1i + ... + δm Zmi + δm+1 Wi + ei
u
∗ step 3: do the following hypothesis test (F-statistic with homoskedasticity):
-H0 : δ1 = δ2 = ... = δm = 0 vs H1 : at least one δj 6= 0, j = 1, ..., m
- CalculateJ = mF and compare with critical values from χ2m−k (J ∼ χ2m−k , where
m−k is the degree of overidentication, the # of instruments minus the # of endogenous
regressors).
- If you reject the null, at least one (maybe more than one) instrument is not exogenous.
- If you fail to reject the null, then you have evidence that the instruments are approx-
imately uncorrelated with bTi SLS .
u

1.5 IV in practice
• Where do valid instruments come from? If IV is so awesome, why don't we use it all the
time? Because in practice, it is really hard to nd valid instruments. Two approaches are:

 Economic theory. Example: weather aects only supply side in agricultural markets.

 Exogenous source of variation in X arising from random phenomena. Example: earthquake

damage increased average class size in some school districts, but it was not related to student
achievement directly.

• Consistency vs Eciency: While IV estimators are consistent given valid instruments, they
can be much less ecient than the OLS estimator and can behave poorly in nite samples.
These problems are greatly magnied if instruments are weakly correlated with the endogenous
variables.

3
2 Exercises
Problem 1 [SW 12.9]
A researcher is interested in the eect of military service on human capital. He collects data from a
random sample of 4000 workers aged 40 and runs the OLS regression Yi = β0 + β1 Xi + ui , where Yi is
the worker i's annual earnings and Xi is a binary variable that is equal to 1 if the person served in the
military and 0 otherwise.

a) Explain why the OLS estimates are likely to unreliable. (Hint: Which variables are omitted from
the regression? Are they correlated with military service?)

b) During the Vietnam War there was a draft, where priority for the draft was determined by a
national lottery. (The days of the year were randomly reordered 1 through 365. Those with
birthdates ordered rst were drafted before those with birthdates ordered second, and so forth.)
Explain how the lottery might be used as an instrument to estimate the eect of military service
on earnings.

Problem 2 [SW E12.1 (Stata)]

How does fertility aect labor supply? That is, how much does a woman's labor supply fall when she
has an additional child? In this exercise you will estimate this eect using data for married women
aged 21-35 with two or more children from the 1980 US Census.
3 The data are available on the
textbook web site www.pearsonhighered.com/stock_watson in the le Fertility and described in the
le Fertility_Description.

a) Regress weeksm1 (weeks worked by the mom) on the indicator variable morekids using OLS.
On average, do women with more than two children work less than women with two children?
How much less?

b) Explain why the OLS regression estimated in (a) is inappropriate for estimating the causal eect
of fertility (morekids) on labor supply (weeksm1).

c) The dataset contains the variable samesex, which is equal to 1 if the rst two children are of the
same sex (boy-boy or girl-girl) and equal to 0 otherwise. Are couples whose rst two children are
of the same sex more likely to have a third child? Is the eect large? is it statistically signicant?

d) Explain why samesex is a valid instrument for the IV regression of weeksm1 on morekids.

e) Is samesex a weak instrument?

f ) Estimate the regression of weeksm1 on morekids using samesex as an instrument. How large
is the fertility eect on labor supply?

g) Do the results change when you include the variables agem1, black , hispan, and othrace in the
labor supply regression (treating these variables as exogenous)? Explain why or why not.

Problem 3 [Checking exogeneity: J-test]

In an IV regression model with one regressor, Xi , and two instruments, Z1i and Z2i , the value of the
J -statistic is J = 18.2.

a) Does this suggest that E(ui |Z1i , Z2i ) 6= 0? Explain.

b) Does this suggest that E(ui |Z1i ) 6= 0? Explain.

3
These data were used Angrist and Evans (1998) Children and Their Parents' Labor Supply : Evidence from
Exogenous Variation in Family Size, American Economic Review

4
Problem 4 [Weak instruments]
In an IV regression model with one regressor, Xi , and one instrument, Zi , the rst-stage regression of
Xi onto Zi has R2 = 0.05 and N = 100. Is Zi a strong instrument? [Hint: use the R2 to compute the
homoskedastic-only F-stat]. Would your answer change if R
2 = 0.05 and N = 500.

Problem 5 [Returns to schooling]

Kling (2001) analyzed in detail the use of college proximity as an instrument for schooling. This
exercise uses the same data from the NLS young men's cohort on 3,010 males aged 24 to 34 years old
in 1976. The model estimated is

ln wi = α + β1 si + β2 ei + β3 e2i + X0 γ + ui

where si denotes years of schooling, ei denotes years of work experience (calculated as ei = agei −si −6),
e2i denotes experience squared, and X is a vector with 26 control variables (geographic indicators and
parental education). Assume these control variables are exogenous.

a) A classmate tells that this regression suers from OVB and thus you must be careful in inter-
preting the coecients. How many endogenous variables do we have? How many instruments
do you need?

b) The author decides to use 3 instruments. The rst one is col4, an indicator for whether a four-
year college is nearby. The other two are age 2
and age . The author argues that (i) age and
age2 are highly correlated with ei and e2i , and (ii) that they can be omitted from the model
for log-wage since it is work experience that matters. Do you agree with him? Is the model
just-identied or overidentied? Can you test the exogeneity of the instruments?

c) The results for β1 from OLS and IV are reported in the table below. Interpret the eect of
schooling on wages for the OLS regression and IV regression. Comment about the precision of
the estimates.

Table 1: Returns to Schooling: IV estimates

OLS IV

Schooling (si ) 0.073 0.132

(0.004) (0.049)

R2 0.304 0.207
First-stage F-stat for si - 8.07
Obs 3010 3010

Notes : This table only reports the coecient

for schooling and its standard error in paren-

theses. The other variables are part of the re-
gression but the coecients are not included
in the table. The rst-stage F-stat is a weak
instrument diagnostic.

5
3 Solution to the exercises
Problem 1 [SW 12.9]
a) There are other factors that could aect both the choice to serve in the military and annual
earnings. One example could be education, although this could be included in the regression as
a control variable. Another variable is ability which is dicult to measure, and thus dicult to
control for in the regression.

b) The draft was determined by a national lottery so the choice of serving in the military was
random. Because it was randomly selected, the lottery number is uncorrelated with individual
characteristics that may aect earning and hence the instrument is exogenous. Because it aected
the probability of serving in the military, the lottery number is relevant.

Problem 2 [SW E12.1 (Stata)]

a) The coecient is -5.387, which indicates that on average women with more than 2 children work
5.387 fewer weeks per year than women with 2 or fewer children.

b) Both fertility and weeks worked are choice variables. A woman with a positive labor supply
regression error (a woman who works more than average) may also be a woman who is less likely
to have an additional child. This would imply that morekids is positively correlated with the
regression error, so that the OLS estimator of βmorekids is positively biased.

c) The linear regression of morekids on samesex (a linear probability model) yields

\
morekids = 0.346 + 0.068 ∗ samesex
(0.001) (0.002)

so that couples with samesex = 1 are 6.8 percentage points more likely to have an additional
child that couples with samesex = 0. The eect is highly signicant (t-statistic = 35.2).

d) samesex is random and is unrelated to any of the other variables in the model including the
error term in the labor supply equation. Thus, the instrument is exogenous. From (c), the rst
stage F-statistic is large (F = 1238) so the instrument is relevant. Together, these imply that
samesex is a valid instrument.

e) No, see the answer to (d).

f ) The estimated value of

IV
βmorekids = - 6.314.

g) The results do not change in an important way. The reason is that samesex is unrelated to
agem1, black , hispan, othrace, so that there is no omitted variable bias in the IV regression of
(f ).

Problem 3 [Checking exogeneity: J-test]

a) Yes. Under the null hypothesis of instrument exogeneity, the J statistic is distributed as a
χ22−1 random variable, with a 1% critical value of 6.63. Thus we reject the null, the statistic is
signicant, and instrument exogeneity E(ui |Z1i , Z2i ) = 0 is rejected.

b) No. The J test suggests that E(ui |Z1i , Z2i ) 6= 0, but doesn't provide evidence about whether the
problem is with Z1i or Z2i (or both).

6
Problem 4 [Weak instruments]

R2 /k 0.05
N = 100 : Fhomosk−only = 2
= = 5.16
(1 − R )/N − k − 1 0.95/98
R2 /k 0.05
N = 500 : Fhomosk−only = 2
= = 26.2
(1 − R )/N − k − 1 0.95/498

The rst case (with N = 100) is lower than the rule-of-thumb value of 10, so Zi is weak instrument.
The second case (with N = 500) is greater than 10, so Zi is not weak.

Problem 5 [Returns to schooling]

a) There are 3 endogenous variables: schooling si owing to lack of data on ability, and the two
2
work experience variables ei and ei since both are calculated using si which is endogenous (recall
ei = agei − si − 6). Accordingly, at least 3 instruments are needed.

b) The model is just-identied (3 endogenous variables and 3 instruments) so we cannot test exo-
geneity. The use of age and age squared as instruments can be questioned. Although age is clearly
exogenous, some unobservables such as social skills may be correlated with both age and wage.
This illustrates the general point that there can be disagreement on assumptions of instrument
validity.

c) The OLS estimate of β1 is 0.073, so that wages rise by 7.3% on average with each extra year
of schooling holding everything else constant. This estimate is an inconsistent estimate of β1
given omitted ability, as discussed before. The IV estimate (or 2SLS estimate) is 0.132, so that
an extra year of schooling is estimated to lead to a 13.2% increase in wage. The IV estimator
is much less ecient than OLS. The standard error of β̂1,OLS is over 10 times larger than the

standard error of β̂1,IV .