Instrumental Variables

Kosuke Imai

Harvard University

S TAT 186/G OV 2002 C AUSAL I NFERENCE

Fall 2019

Kosuke Imai (Harvard) Instrumental Variables Stat186/Gov2002 Fall 2019 1 / 13

Instrumental Variables

From randomized encouragement design to general instrumental

variabes approach:

U Y

Z T

Instruments in the nature natural experiments

1 random assignment of Z
2 no direct effect of Z on Y

Kosuke Imai (Harvard) Instrumental Variables Stat186/Gov2002 Fall 2019 2 / 13

Classical Instrumental Variables Estimator
Linear model (in matrix notation):

Y = Xβ +  where E() = 0n and X is n × K

Endogeneity:
E(i | X) 6= 0
Instruments Z is n × L
1 Exogeneity: E(i | Z) = 0
2 Exclusion restriction: Zi does not belong to the outcome model
3 Rank condition: Z> X and Z> Z have full rank
Experimental setting:
Xi = the treatment and pre-treatment covariates
Zi = the randomized encouragement and pre-treatment covariates
Identification
1 K = L: just-identified
2 K < L: over-identified
3 K > L: under-identified
Kosuke Imai (Harvard) Instrumental Variables Stat186/Gov2002 Fall 2019 3 / 13
Geometry of Instrumental Variables

Projection matrix (onto S(Z)): PZ = Z(Z> Z)−1 Z>

“Purge” endogeneity: Xb = PZ X
Since PZ = P>
Z and PZ PZ = PZ , we have

b > X)
β̂IV = (X b −1 X
b>Y
= (X> Z(Z> Z)−1 Z> X)−1 X> Z(Z> Z)−1 Z> Y

Two stage least squares:

1 Regress X on Z and obtain the fitted values X
b
2 Regress Y on X
b
We do not assume the linearity of X in Z

Kosuke Imai (Harvard) Instrumental Variables Stat186/Gov2002 Fall 2019 4 / 13

Asymptotic Inference
Estimation error:

β̂IV − β = (X> Z(Z> Z)−1 Z> X)−1 X> Z(Z> Z)−1 Z> 

n
! n
!−1 n
!−1
 1X 1 X 1 X 
= Xi Z>
i Z Z>
i i Z X
i i
>
 n n n 
i=1 i=1 i=1
n
! n
!−1 n
!
1X 1X 1X
× Xi Z>
i Zi Z>
i Zi i
n n n
i=1 i=1 i=1

p
Thus, β̂IV −→ β
Under the homoskedasticity, V( | Z) = σ 2 In :
√ d
n(β̂IV − β) N (0, σ 2 [E(Xi Z> > −1 > −1
i ){E(Zi Zi )} E(Zi Xi )] )

Kosuke Imai (Harvard) Instrumental Variables Stat186/Gov2002 Fall 2019 5 / 13

Residuals and Robust Standard Error

ˆ = Y − Xβ̂IV and not ˆ 6= Y − X

b β̂IV
k2
kˆ p
Under homoskedasticity: σ̂ 2 = n−K −→ σ 2

\
V( β̂IV ) = σ̂ 2 {X> Z(Z> Z)−1 Z> X}−1 = σ̂ 2 (X
b > X)
b −1

Sandwich heteroskedasticity consistent estimator:

bread = {X> Z(Z> Z)−1 Z> X}−1 X> Z(Z> Z)−1

n
!
X
> 2 2 >
meat = Z diag(ˆ
i )Z = ˆi Zi Zi
i=1
> b > X)
b −1 X
b > diag(ˆ b > X)
b −1
bread meat bread = (X 2i )X(
b X

Robust standard errors for clustering, auto-correlation, etc.

Kosuke Imai (Harvard) Instrumental Variables Stat186/Gov2002 Fall 2019 6 / 13

Multi-valued Treatment (Angrist and Imbens. 1995. J. Am. Stat. Assoc)

Two stage least squares regression:

Yi = α + βTi + ηi ,
Ti = δ + γZi + i

Binary encouragement and binary treatment,

\
Cov(Y i ,Zi )
p
β̂ = \
−→ CATE (no covariate)
Cov(T i ,Zi )
p
β̂ −→ CATE (with covariates)
Binary encouragement multi-valued treatment
Monotonicity: Ti (1) ≥ Ti (0)
Exclusion restriction: Yi (1, t) = Yi (0, t) for each t = 0, 1, . . . , K

Kosuke Imai (Harvard) Instrumental Variables Stat186/Gov2002 Fall 2019 7 / 13

Estimator
p Cov(Yi , Zi ) E(Yi (1) − Yi (0))
β̂TSLS −→ =
Cov(Ti , Zi ) E(Ti (1) − Ti (0))
K K  
X X Yi (1) − Yi (0) 
= wjk E  Ti (1) = j, Ti (0) = k
j −k
k =0 j=k +1

where wjk is the weight, which sums up to one, defined as,

(j − k ) Pr(Ti (1) = j, Ti (0) = k )

wjk = PK PK .
(j 0 − k 0 ) Pr(T (1) = j 0 , T (0) = k 0 )
0
k =0 0 0
j =k +1 i i

Easy interpretation under the constant additive effect assumption

for every complier type

Assume encouragement induces at most only one additional dose

Then, wk = Pr(Ti (1) = k , Ti (0) = k − 1)
Kosuke Imai (Harvard) Instrumental Variables Stat186/Gov2002 Fall 2019 8 / 13
Quarter of Birth (Angrist and Krueger. 1991. Q. J. Econ.)
Instrument for educational attainment to address “ability bias”
Outcome: men’s log weekly earnings in 1980
Compulsory education law in US: students must attend school until
they reach age 16
Those born in the third or fourth quarter typically finish tenth grade
before reaching age 16
Instrument at most decreases years of education by one year
Weak instrument: first quarter vs. 2nd to 4th quarter
1920s cohorts: est. = −0.126, s.e. (HC) = 0.016, corr = −0.016
1930s cohorts: est. = −0.109, s.e. (HC) = 0.013, corr = −0.014
Wald estimates:
1920 cohorts: est. = 0.072, s.e. (HC) = 0.022
1930 cohorts: est. = 0.102, s.e. (HC) = 0.024
OLS estimates:
1920 cohorts: est. = 0.080, s.e. (HC) = 0.0004
1930 cohorts: est. = 0.071, s.e. (HC) = 0.0004
Kosuke Imai (Harvard) Instrumental Variables Stat186/Gov2002 Fall 2019 9 / 13
first
quarter.Ifthisconditionis satisfied,
thenwe can get fourth quarter.Foreachlevelofschooling,
j, thisdifference
someidea ofthesizeand characteristicsofthegroupcon- isthefraction ofthepopulation
whoseschooling is switched
CDFs for First and Fourth Quarter of Birth
totheACR through
tributing theACR weighting function. byquarterofbirthfromlessthanj yearsto at leastj years.
The CDF's ofschooling byquarterofbirthformenin Figures3 and4 showdifferences
intheCDF ofschooling
the1970and 1980Censusesaregraphed in Figures1 and byquarterofbirth.In eachfigure,differences between the

CDF
I . n

0.8

0.6

0.4

0.2

0 4 8 12 16 20

YEARS OF SCHOOLING

Figure2. SchoolingCDF by QuarterofBirth(Men Born 1930-1939; Data Fromthe 1980 Census). Quarterof birth: , first;- -, fourth.

Kosuke Imai (Harvard) Instrumental Variables Stat186/Gov2002 Fall 2019 10 / 13

Analysis of Weak Instruments
Recall the Wald estimator:
Cov(Yi , Zi )
β̂IV =
Cov(Ti , Zi )

β̂IV does not exist if the instrument is irrelevant

Consider the following model:
Yi = α + βTi + i ,
Ti = γ Z + ηi , where E(i | Zi ) = E(ηi | Zi ) = 0
|{z} i
≈0

where (i , ηi ) follows a bivariate normal with mean zero. Then,

Pn s
i Zi d V(i )
β̂iv − β ≈ Pni=1 Corr(i , ηi ) + Wi
i=1 ηi Zi V(ηi ) |{z}
Cauchy

Rich literature on asymptotic analysis for weak instruments

Kosuke Imai (Harvard) Instrumental Variables Stat186/Gov2002 Fall 2019 11 / 13
Simulated Instruments (Bound et al. 1995. J. Am. Stat. Assoc)

Simulation exercise:
1 Simulate Zi from Bernoulli with success probability equal to its
empirical estimate
2 Compute the Wald estimate as before

1920s cohorts:
Estimates: min = −694.718, 1st Qu. = −0.093, median = 0.0876,
3rd Qu. = 0.260, max = 36.236
Std. Errors: min = 0.057, 1st Qu. = 0.185, median = 0.393, 3rd
Qu. = 1.467, max = 4865.657

1930s cohorts:
Estimates: min = −36.223, 1st Qu. = −0.117, median = 0.078,
3rd Qu. = 0.284, max = 202.667
Std. Errors: min = 0.064, 1st Qu. = 0.197, median = 0.421, 3rd
Qu. = 1.814, max = 427582

Kosuke Imai (Harvard) Instrumental Variables Stat186/Gov2002 Fall 2019 12 / 13

Summary

Instrumental variables as a general strategy for coping with

selection bias
randomization of instruments
monotonicity
exclusion restriction
Extensions to multi-valued treatment
Weak instruments

Readings:
I MBENS AND RUBIN . Chapter 25
A NGRIST AND P ISCHKE . Chapter 4

Kosuke Imai (Harvard) Instrumental Variables Stat186/Gov2002 Fall 2019 13 / 13