MICRO-ECONOMETRICS

ECO 6175

ABEL BRODEUR

Week 5

1/ 32
Instrumental Variable

Outline:
I (1) Instrumental Variable
I (2) Angrist and Krueger (1991)
I (3) Stata

2/ 32
Objective:

How can we evaluate government policies?

I Understanding Instrumental Variable

Reference:
I Angrist and Krueger, 1991. “Does Compulsory
School Attendance Affect Schooling and Earnings?”
I Angrist, 1990. “Lifetime Earnings and the Vietnam
Era Draft Lottery”

3/ 32
Instrumental Variables: Example
Remember that we wanted to run the long regression

yi = α + ρSi + γAi + i .

One solution was to look for variables which can act as

proxies for ability Ai . The key to regression is that we
need something that captures all the variation in Ai ,
otherwise we are left with OVB

The IV solution is to isolate variation in Si which is

unrelated to Ai . The variable which does the “isolating” is
the instrumental variable. The good thing is that we just
need some of the variation in Ai . But IV comes with other
possible complications

4/ 32
Instrumental Variables

Call the instrumental variable Zi . A valid instrument needs

to satisfy three (two) conditions:
I (1*) Zi is as good as randomly assigned
I (2) Zi satisfies the exclusion restriction, i.e. it does not
appear as a separate regressor in the long
regression we would like to run
I (3) Zi affects the endogenous regressor Si
Of these conditions, only (3) can be tested. This is the
strength of the first stage. Conditions 1 and 2 have to be
argued based on knowledge from outside the data we
have

5/ 32
Instrumental Variables

Three causal effects we can think about:

I (1) The causal effect of Zi on Si
I (2) The causal effect of Zi on yi
I (3) The causal effect of Si on yi
We are interested ultimately in the last one

6/ 32
Instrumental Variables: Notions

The instrumental variables language comes from old style

simultaneous equations models but we can think of it as
related to the three causal effects

Structural equation: the regression of earnings on

schooling is called the structural equation:

yi = α + ρSi + ηi ,

where ηi = γAi + i , i.e. it is a structural error term, not a

regression residual (causal effect 3)

7/ 32
Instrumental Variables: Notions
The instrumental variables language comes from old style
simultaneous equations models but we can think of it as
related to the three causal effects

First stage: the regression of schooling on the instrument

is called the first stage (causal effect 1)

Si = π10 + π11 Zi + ζ1i

Reduced form: the regression of earnings on the

instrument (causal effect 2)

yi = π20 + π21 Zi + ζ2i .

8/ 32
Instrumental Variables

Conditions 1 and 3 on the instrument are enough to get

causal effects 1 and 2. These conditions are sufficient for
the first stage and the reduced forms to have a causal
interpretations. Note that the reduced form coefficient
might be interesting on its own

To get causal effect 3, we also need condition 2, the

exclusion restriction. This condition is usually the most
difficult requirement

9/ 32
Linking the Three Equations
The coefficients in the three equations are linked.
Substitute the first stage into the structural equation:

yi = α + ρSi + ηi
= α + ρ[π10 + π11 Zi + ζ1i ] + ηi
= (α + ρπ10 ) + ρπ11 Zi + (ζ1i + ηi )
= π20 + π21 Zi + ζ2i .

Hence, the reduced form coefficients are:

π20 = α + ρπ10

π21 = ρπ11

10/ 32
Indirect Least Squares

It is straightforward to see that

π21
ρ= ,
π11
i.e. the IV estimate is equal to the ratio of the reduced
form coefficient on the instrument to the first stage
coefficient. This is called indirect least squares

11/ 32
Just Identified vs. Over-Identified
Indirect least squares only works when there is one
endogenous regressor and one instrument. Such a model
is just identified (there is only one single solution to get ρ
from the first stage and reduced form coefficients)

If there are multiple instruments for a single endogenous

regressor the model is over-identified

Si = π10 + π11 Z1i + π12 Z2i + ζ1i

yi = π20 + π21 Z1i + π22 Z2i + ζ2i

There is no unique way to get ρ from π11 , π12 , π21 and π22 .
Two stage least squares (2SLS) is a particular average (it
is optimally weighted GMM estimator for the
homoskedastic model)
12/ 32
Angrist and Krueger (1991)

US compulsory schooling and school entry rules

Compulsory age is 16. You can drop out on your 16th

birthday (even if in the middle of the school year)

School entry is once a year, and cutoffs are based on

birthdays

13/ 32
Angrist and Krueger (1991)

Variation comes from the compulsion side of the

schooling problem. Three conditions:
I (1) Random assignment: are birthdays random with
respect to the counterfactual earnings for different
schooling levels
I (2) Do birthdays satisfy the exclusion restriction, or
could birthdays be correlated with earnings for other
reasons than their effect on schooling?
I (3) Do birthdays affect schooling?

14/ 32
Angrist and Krueger (1991)

Data are from the 1980 Census

359,509 men born 1930 to 1939 (in their 40s when

observed)

For these men we have the year of birth, quarter of birth,

years of schooling, and earnings in 1979

15/ 32
First Stage regressions of Schooling on
Quarter of Birth

Schooling Schooling Schooling Schooling

Regressor (1) (2) (3) (4)
quarter 2 0.057 0.057
(0.017) (0.016)
quarter 3 0.117 0.113
(0.016) (0.016)
quarter 4 0.092 0.151 0.091 0.148
(0.013) (0.016) (0.013) (0.016)
9 year of birth X X
dummies

16/ 32
Reduced Form Regressions of Log Wages
on Quarter of Birth

Log Wages Log Wages Log Wages Log Wages

Regressor (1) (2) (3) (4)
quarter 2 0.0045 0.0046
(0.0034) (0.0034)
quarter 3 0.0149 0.0150
(0.0033) (0.0033)
quarter 4 0.0068 0.0135 0.0068 0.0135
(0.0027) (0.0034) (0.0027) (0.0034)
9 year of birth X X
dummies

17/ 32
IV Regressions of Log Wages on
Schooling
Coefficient on schooling from an OLS regression:
0.071 (0.0004)

Log Wages Log Wages Log Wages Log Wages

Regressor (1) (2) (3) (4)
Schooling 0.074 0.103 0.075 0.105
(0.028) (0.020) (0.028) (0.020)

Instruments Quarter 4 4 quarter Quarter 4 4 quarter

dummies dummies
9 year of birth X X
dummies

18/ 32
IV with a Dummy Instrument: Wald
Estimator

A way to look at IV with a binary instrument (e.g. Q4):

cov(lnyi , Q4i ) E[lnyi |Q4i = 1] − E[lnyi |Q4i = 0]

βIV = =
cov(Si , Q4i ) E[Si |Q4i = 1] − E[Si |Q4i = 0]

This is called the Wald estimator (Wald, 1944)

19/ 32
Constructing the Wald Estimator
The first stage and reduced form are

Si = π11 + π14 Q4i + ζ1i ,

lnyi = π21 + π24 Q4i + ζ2i .
Taking expectations conditionally on Q4i yields

E[lnyi |Q4i = 1] = π21 + π24

E[lnyi |Q4i = 0] = π21
Hence, the reduced form coefficients on Q4i are the
differences in group means with the instrument switched
on and off:

π24 = E[lnyi |Q4i = 1] − E[lnyi |Q4i = 0]

π14 = E[Si |Q4i = 1] − E[Si |Q4i = 0]
The IV estimate is the ratio of the two
20/ 32
Wald Estimate of the Return to Schooling

Born quarter 1, 2, 3 Born quarter 4 Difference

(1) (2) (3)
Log Earnings 5.8993 5.9051 0.0068
(0.027)

Schooling 12.747 12.839 0.092

(0.013)

Wald Estimate 0.074

(0.028)

21/ 32
IV Estimates LATE

If all assumptions are satisfied, IV estimates LATE (Local

Average Treatment Effect)
I LATE is the average effect of X on Y for those whose
treatment status has been changed by the instrument
Z
Angrist (1990)
I Uses the Vietnam draft lottery as in IV for military
service
I Random sequence numbers were assigned to each
birth date in cohorts of 19-year-olds
I Draft did not perfectly determinate military service

22/ 32
IV Estimates LATE

First stage: being eligible for the draft increases veteran

status by about 16 percentage points (the mean of
veteran status is about 27 percent)

Second stage: serving in the army lowers earnings by

between $2,050 and $2,741 per year

LATE: IV estimates the average effect of military service

on earnings for the subpopulation who enrolled in military
service because of the draft but would not have served
otherwise (exclude those exempted)

23/ 32
RCT and Instruments

Using IVs can be useful for randomized control trials

I Participation is often voluntary among those
randomly assigned to treatment
I Positive selection bias (participate if think that will
benefit from treatment)

A solution is to instrument the treatment with whether you

were offered the treatment (estimates LATE)

24/ 32
Measurement Error Problem

Instrumental variables solve the measurement error

problem

Suppose you have an instrument zi , correlated with the

signal x1i and uncorrelated with the error wi

In the bivariate regression you get

cov(yi , zi ) cov(α + βxi + i , zi ) βcov(xi , zi )

β̂IV = = = =β
cov(x̃i , zi ) cov(xi , zi ) cov(xi , zi )

For instance, if your proxy for IQ is not measured well,

take another test as the instrument

25/ 32
Weak Instruments
Bound, Jaeger and Baker (1995) pointed out that the
quarter of birth instruments explain only a tiny proportion
of the variation in schooling. This leads to problems:
I The 2SLS estimator with weak instruments is biased
in small samples
I Any inconsistency from a small violation of the
exclusion restriction gets magnified by weak
instruments

Adding more weak instruments will increase the bias of

the 2SLS. The F-statistic decreases and the bias
increases
I Test whether using 30 instruments (quarter of birth
times year of birth dummies) instead of 3 (quarter of
birth) decreases the F-statistic
26/ 32
Bound, Jaeger and Baker

OLS IV OLS IV
Regressor (1) (2) (3) (4)
Coefficient 0.063 0.142 0.063 0.081
(0.000) (0.033) (0.000) (0.016)

Instruments 4 quarter 4 quarter x

dummies year of birth
9 year of birth X X
dummies
Number of Instruments 3 30

F-Statistic 13.486 4.747

27/ 32
Weak Instruments

Report the first stage and think about whether it makes

sense. Are the magnitude and sign as we expect?

Report the F-statistic on the excluded instruments. The

bigger this is, the better. F s above 10-20 are usually
considered safe

Pick your best single instrument and report just-identified

estimates using this one only. Just-identified IV is
approximately median-unbiased

28/ 32
Weak Instruments

Check over-identified 2SLS estimators with LIML (limited

information maximum likelihood). If the LIML estimates
are very different, or standard errors are much bigger,
worry

The reduced-form estimates are just OLS, so they are

unbiased. If the relationship is not in the reduced-form,
worry

29/ 32
Testing Over-Identifying Restrictions
When we have more instruments than we need to identify
an equation, we can test whether the additional
instruments are valid in the sense that they are
uncorrelated with the error term

See the Sargan test and the Hansen test (the latter is
robust to heteroskedasticity)

The test is useful since rejecting the null hypothesis

means that our logic for choosing the IVs must be
re-examined
I Unfortunately it does not tell us which IVs fail the
exogeneity requirement
I Tells us whether estimates change when we select
different subsets from a set of possible instruments
30/ 32
Stata
Effect of education on wages (Wooldridge’s book)
I Instruments: parents’ education and husband
education
https://stats.idre.ucla.edu/stata/examples/eacspd/econometric-
analysis-of-cross-section-and-panel-data-by-jeffrey-m-
wooldridgechapter-5-instrumental-variables-estimation-of-
single-equation-linear-models/

Effect of institutions on growth (Acemoglu et al. (2001))

I Instrument: settler mortality
http://economics.mit.edu/faculty/acemoglu/data/ajr2001

31/ 32
Recent IV Papers
Hill, A. J., 2015. “The Girl Next Door: The Effect of
Opposite Gender Friends on High School Achievement,”
American Economic Journal: Applied Economics, 7(3):
147-177.

Autor et al., 2016. “The Contribution of the Minimum

Wage to US Wage Inequality over Three Decades: A
Reassessment,” American Economic Journal: Applied
Economics,8(1): 58-99.

Black et al., 2015. “The Impact of the Great Migration on

Mortality of African Americans: Evidence from the Deep
South,” American Economic Review, 105(2): 477-503.

32/ 32