Gov 2002: 10.

Instrumental
Variables
Matthew Blackwell
November 5, 2015

1 / 48
1. IV setup

2. IV with constant treatment effects

3. IV with heterogenous treatment effects

4. IV extensions

2 / 48
1/ IV setup

3 / 48
Where are we? Where are we
going?

• We saw how to identify and estimate effects under no

unmeasured confounding and with repeated measurements
• What if we have neither? Are we doomed?
• Not necessarily if you can identify some exogenous sources of
variation that drives the treatment.
• Instrumental variables allows for unmeasured confounding on
the the treatment-outcome relationship.
• Use the unconfounded variation in the instrument to help
identify treatment effects.

4 / 48
Basic IV setup with DAGs
𝑈

𝑍 𝐷 𝑌

exclusion restriction

• 𝑍 is the instrument, 𝐷 is the treatment, and 𝑈 is the

unmeasured confounder
• Exclusion restriction
▶ no common causes of the instrument and the outcome
▶ no direct or indirect effect of the instrument on the outcome
not through the treatment.

• First-stage relationship: 𝑍 affects 𝐷

5 / 48
An IV is only as good as its
assumptions

𝑍 𝐷 𝑌

exclusion restriction

• Finding a believable instrument is incredibly difficult and some

people never believe any IV setups.
• When effects vary, the IV approach estimates a “local” ATE
that is local to this particular instrument.

6 / 48
IVs in the field
• Angrist (1990): Draft lottery as an IV for military service
(income as outcome)
• Acemoglu et al (2001): settler mortality as an IV for
institutional quality (GDP/capita as outcome)
• Levitt (1997): being an election year as IV for police force size
(crime as outcome)
• Kern & Hainmueller (2009): having West German TV
reception in East Berlin as an instrument for West German TV
watching (outcome is support for the East German regime)
• Nunn & Wantchekon (2011): historical distance of ethnic
group to the coast as a instrument for the slave raiding of
that ethnic group (outcome are trust attitudes today)
• Acharya, Blackwell, Sen (2015): cotton suitability as IV for
proportion slave in 1860 (outcome is white attitudes today)

7 / 48
2/ IV with
constant
treatment effects
8 / 48
IV with constant effects

• Let’s write down a causal model for 𝑌𝑖 with constant effects

and an unmeasured confounder, 𝑈𝑖 :

𝑌𝑖 (𝑑, 𝑢) = 𝛼 + 𝜏𝑑 + 𝛾𝑢 + 𝜂𝑖

• If we connect this with a consistency assumption, we get the

this regression form:

𝑌𝑖 = 𝛼 + 𝜏𝐷𝑖 + 𝛾𝑈𝑖 + 𝜂𝑖

• Here we assume that 𝔼[𝐷𝑖 𝜂𝑖 ] = 0, so if we measured 𝑈𝑖 , then

we would be able to estimate 𝜏.
• But Cov(𝛾𝑈𝑖 + 𝜂𝑖 , 𝐷𝑖 ) ≠ 0 because 𝑈 is a common cause of 𝐷
and 𝑌 .

9 / 48
The role of the instrument

• If we have an instrument, 𝑍𝑖 , that satisfies the exclusions

restriction, then

Cov(𝛾𝑈𝑖 + 𝜂𝑖 , 𝑍𝑖 ) = 0

• It must be independent of 𝑈𝑖 and it has no correlation with 𝜂𝑖

because neither does the treatment.
Cov(𝑌𝑖 , 𝑍𝑖 ) = Cov(𝛼 + 𝜏𝐷𝑖 + 𝛾𝑈𝑖 + 𝜂𝑖 , 𝑍𝑖 )
= Cov(𝛼, 𝑍𝑖 ) + Cov(𝜏𝐷𝑖 , 𝑍𝑖 ) + Cov(𝛾𝑈𝑖 + 𝜂𝑖 , 𝑍𝑖 )
= 0 + 𝜏Cov(𝐷𝑖 , 𝑍𝑖 ) + 0

10 / 48
IV estimator with constant effects

𝑌𝑖 = 𝛼 + 𝜏𝐷𝑖 + 𝛾𝑈𝑖 + 𝜂𝑖

• With this in hand, we can formulate an expression for the

average treatment effect here:

Cov(𝑌𝑖 , 𝑍𝑖 ) Cov(𝑌𝑖 , 𝑍𝑖 )/𝕍[𝑍𝑖 ]

𝜏= =
Cov(𝐷𝑖 , 𝑍𝑖 ) Cov(𝐷𝑖 , 𝑍𝑖 )/𝕍[𝑍𝑖 ]
• Reduced form coefficient: Cov(𝑌𝑖 , 𝑍𝑖 )/𝕍[𝑍𝑖 ]
• First stage coefficient: Cov(𝐷𝑖 , 𝑍𝑖 )/𝕍[𝑍𝑖 ]

11 / 48
Weak instruments
• Natural estimator:

̂ 𝑖 , 𝑍𝑖 )
Cov(𝑌
𝜏
̂𝐼𝑉 =
̂ 𝑖 , 𝑍𝑖 )
Cov(𝐷
• What happens with a weak first stage? Can show that this
estimator converges to:
𝑝 Cov(𝑍𝑖 , 𝑈𝑖 )
𝜏
̂𝐼𝑉 → 𝜏 +
Cov(𝑍𝑖 , 𝐷𝑖 )
• If Cov(𝑍𝑖 , 𝐷𝑖 ) is small, then even very small violations of the
exclusion restriction Cov(𝑍𝑖 , 𝑈𝑖 ) ≠ 0 can lead to large
inconsistencies and finite sample bias.
• Important to convey the strength of the first-stage via 𝑡-test
or 𝐹-test with multiple instruments.

12 / 48
Wald Estimator

• Binary instrument leads to the Wald estimator:

Cov(𝑌𝑖 , 𝑍𝑖 ) 𝔼[𝑌𝑖 |𝑍𝑖 = 1] − 𝔼[𝑌𝑖 |𝑍𝑖 = 0]

𝜏= =
Cov(𝐷𝑖 , 𝑍𝑖 ) 𝔼[𝐷𝑖 |𝑍𝑖 = 1] − 𝔼[𝐷𝑖 |𝑍𝑖 = 0]
• Intuitively:

effect of instrument on outcome

effect of instrument on treatment

13 / 48
What about covariates?

• No covariates up until now. What if we have a set of

covariates 𝑋𝑖 that we are also conditioning on?
• Let’s start with linear models for both the outcome and the
treatment:
𝑌𝑖 = 𝑋𝑖′ 𝛽 + 𝜏𝐷𝑖 + 𝜀𝑖
𝐷𝑖 = 𝑋𝑖′ 𝛼 + 𝛾𝑍𝑖 + 𝜈𝑖
• Now, we assume that 𝑋𝑖 are exogenous along with 𝑍𝑖 :

𝔼[𝑍𝑖 𝜈𝑖 ] = 0 𝔼[𝑍𝑖 𝜀𝑖 ] = 0

𝔼[𝑋𝑖 𝜈𝑖 ] = 0 𝔼[𝑋𝑖 𝜀𝑖 ] = 0
• …but 𝐷𝑖 is endogenous: 𝔼[𝐷𝑖 𝜀𝑖 ] ≠ 0

14 / 48
Getting the reduced form
• We can plug the treatment equation into the outcome
equation:

𝑌𝑖 = 𝑋𝑖′ 𝛽 + 𝜏[𝑋𝑖′ 𝛼 + 𝛾𝑍𝑖 + 𝜈𝑖 ] + 𝜀𝑖

= 𝑋𝑖′ 𝛽 + 𝜏[𝑋𝑖′ 𝛼 + 𝛾𝑍𝑖 ] + [𝜏𝜈𝑖 + 𝜀𝑖 ]
= 𝑋𝑖′ 𝛽 + 𝜏[𝑋𝑖′ 𝛼 + 𝛾𝑍𝑖 ] + 𝜀∗𝑖
= 𝑋𝑖′ 𝛽 + 𝜏𝔼[𝐷𝑖 |𝑋𝑖 , 𝑍𝑖 ] + 𝜀∗𝑖

• Red value in the brackets is the population fitted value of the

treatment, 𝔼[𝐷𝑖 |𝑋𝑖 , 𝑍𝑖 ]
• Because 𝑍𝑖 and 𝑋𝑖 are uncorrelated with 𝜈𝑖 and 𝜀𝑖 , then this
fitted value is also independent of 𝜀∗𝑖 .
• Thus, the population regression coefficient of a 𝑌𝑖 on
[𝑋𝑖′ 𝛼 + 𝛾𝑍𝑖 ] is the average treatment effect, 𝜏.

15 / 48
Two-stage least squares

• Estimate 𝛼
̂ and 𝛾
̂ from OLS and form fitted values:

̂ 𝑖 |𝑋𝑖 , 𝑍𝑖 ] = ̂
𝔼[𝐷 𝐷𝑖 = 𝑋𝑖′ 𝛼
̂+𝛾
̂ 𝑍𝑖 .

• Regress of 𝑌𝑖 on 𝑋𝑖 and ̂
𝐷𝑖 . Add and subtract 𝜏̂
𝐷𝑖 :

𝑌𝑖 = 𝑋𝑖′ 𝛽 + 𝜏̂
𝐷𝑖 + [𝜀𝑖 + 𝜏(𝐷𝑖 − ̂
𝐷𝑖 )]

• Key question: is ̂
𝐷𝑖 uncorrelated with the error?
• ̂
𝐷𝑖 is just a function of 𝑋𝑖 and 𝑍𝑖 so it is uncorrelated with 𝜀𝑖 .
• We also know that ̂
𝐷𝑖 is uncorrelated with (𝐷𝑖 − ̂
𝐷𝑖 )?

16 / 48
Two-stage least squares
• Heuristic procedure:
1. Run regression of treatment on covariates and instrument
2. Construct fitted values of treatment
3. Run regression of outcome on covariates and fitted values
• Note that this isn’t how we actually estimate 2SLS because
the standard errors are all wrong.
• Computer wants to calculate the standard errors based on 𝜀∗𝑖 :

𝜀∗𝑖 = 𝑌𝑖 − 𝑋𝑖′ 𝛽 − 𝜏̂
𝐷𝑖

• but what we really want is the standard errors based on 𝜀𝑖 :

𝜀𝑖 = 𝑌𝑖 − 𝑋𝑖′ 𝛽 − 𝜏𝐷𝑖

17 / 48
Nunn & Wantchekon IV example

18 / 48
General 2SLS

• Notational convenience: combine 𝑋𝑖 and 𝐷𝑖 into one matrix,

𝑋𝑖 , of size 𝑘, where one column contains 𝐷𝑖 .
• The structural model, then is:

𝑌𝑖 = 𝑋𝑖′ 𝛽 + 𝜀𝑖

• 𝑍𝑖 will be a vector of 𝑙 exogenous variables that includes any

exogenous variables in 𝑋𝑖 plus any instruments.
• Key assumption on the instruments:

𝔼[𝑍𝑖 𝜀𝑖 ] = 0

19 / 48
Nasty Matrix Algebra
• Projection matrix projects values from the columns of 𝑍𝑖 to
the columns of 𝑋𝑖 :
Π = (𝔼[𝑍𝑖 𝑍𝑖′ ])−1 𝔼[𝑍𝑖 𝑋𝑖′ ] (projection matrix)
𝑋𝑖̃ = Π′ 𝑍𝑖 (fitted values)
• To derive the 2SLS estimator, take the fitted values, Π′ 𝑍𝑖 and
multiply both sides of the outcome equation by them:
𝑌𝑖 = 𝑋𝑖′ 𝛽 + 𝜀𝑖
Π′ 𝑍𝑖 𝑌𝑖 = Π′ 𝑍𝑖 𝑋𝑖′ 𝛽 + Π′ 𝑍𝑖 𝜀𝑖
𝔼[Π′ 𝑍𝑖 𝑌𝑖 ] = 𝔼[Π′ 𝑍𝑖 𝑋𝑖′ ]𝛽 + 𝔼[Π′ 𝑍𝑖 𝜀𝑖 ]
𝔼[Π′ 𝑍𝑖 𝑌𝑖 ] = 𝔼[Π′ 𝑍𝑖 𝑋𝑖′ ]𝛽 + Π′ 𝔼[𝑍𝑖 𝜀𝑖 ]
𝔼[Π′ 𝑍𝑖 𝑌𝑖 ] = 𝔼[Π′ 𝑍𝑖 𝑋𝑖′ ]𝛽
𝔼[𝑋𝑖̃ 𝑌𝑖 ] = 𝔼[𝑋𝑖̃ 𝑋𝑖′ ]𝛽
𝛽 = (𝔼[𝑋𝑖̃ 𝑋𝑖′ ])−1 𝔼[𝑋𝑖̃ 𝑌𝑖 ]

20 / 48
How to estimate the parameters
• Collect 𝑋𝑖 into a 𝑛 × 𝑘 matrix 𝐗 = (𝑋1′ , … , 𝑋𝑛′ )
• Collect 𝑍𝑖 into a 𝑛 × 𝑙 matrix 𝐙 = (𝑍1′ , … , 𝑍𝑛′ )
̂ = 𝐙(𝐙′ 𝐙)−1 𝐙′ 𝐗 be the matrix of fitted values for 𝐗,
• Let 𝐗
then we have
𝑁 𝑝
• Matrix party trick: 𝐗′ 𝐙/𝑛 = (1/𝑛) ∑𝑖 𝑋𝑖 𝑍𝑖′ → 𝔼[𝑋𝑖 𝑍𝑖′ ].
• Take the population formula for the parameters:

𝛽 = (𝔼[𝑋𝑖̃ 𝑋𝑖′ ])−1 𝔼[𝑋𝑖̃ 𝑌𝑖 ]

• And plug in the sample values (the 𝑛 cancels out):

̂ ′ 𝐗)−1 𝐗
𝛽 ̂ = (𝐗 ̂ ′𝐲

• This is how R/Stata estimates the 2SLS parameters

21 / 48
Asymptotics for 2SLS
̂ ′ 𝐗)−1 𝐗
𝛽 ̂ = (𝐗 ̂ ′𝐲

• We can insert the true model for 𝐲:

̂ ′ 𝐗)−1 𝐗
𝛽 ̂ = (𝐗 ̂ ′ (𝐗𝛽 + 𝜀)

̂ ′𝐗 = 𝐗
• Using the matrix party trick and that 𝐗 ̂ ′ 𝐗,
̂ we have

̂ ′ 𝐗)−1 𝐗
𝛽 ̂ = (𝐗 ̂ ′ 𝐗𝛽 + (𝐗
̂ ′ 𝐗)−1 𝐗
̂ ′𝜀
= 𝛽 + (𝐗̂ ′ 𝐗)
̂ −1 𝐗
̂ ′𝜀
−1
̂𝑖 𝑋
= 𝛽 + [𝑛−1 ∑ 𝑋 ̂𝑖′ ] 𝑛−1 ∑ 𝑋
̂𝑖 𝜀𝑖
𝑖 𝑖

𝑝
̂𝑖 𝜀𝑖 → 𝔼[𝑋
• Consistent because 𝑛−1 ∑𝑖 𝑋 ̂𝑖 𝜀𝑖 ] = 0.

22 / 48
Asymptotic variance for 2SLS
−1
̂𝑖 𝑋
√𝑛(𝛽̂ − 𝛽) = (𝑛−1 ∑ 𝑋 ̂𝑖′ ) (𝑛−1/2 ∑ 𝑋
̂𝑖 𝜀𝑖 )
𝑖 𝑖

̂𝑖 𝜀𝑖 converges in distribution to
• By the CLT, 𝑛−1/2 ∑𝑖 𝑋
𝑁(0, 𝐵), where 𝐵 = 𝔼[𝑋̂𝑖′ 𝜀′𝑖 𝜀𝑖 𝑋
̂𝑖 ].
𝑝
̂𝑖 𝑋
• By the LLN, 𝑛−1 ∑𝑖 𝑋 ̂𝑖′ → 𝔼[𝑋 ̂𝑖 𝑋̂𝑖′ ].
• Thus, we have that √𝑛(𝛽̂ − 𝛽) has asymptotic variance:

̂𝑖 𝑋
(𝔼[𝑋 ̂𝑖′ ])−1 𝔼[𝑋
̂𝑖′ 𝜀′𝑖 𝜀𝑖 𝑋
̂𝑖 ](𝔼[𝑋
̂𝑖 𝑋
̂𝑖′ ])−1

• Replace with the sample quantities to get estimate of the

robust 2SLS variance estimator:
̂ ′ 𝐗)
ar(𝛽)̂ = (𝐗
v̂ ̂ −1 ( ∑ 𝑢𝑖2̂ 𝑋
̂𝑖 𝑋 ̂ ′ 𝐗)
̂𝑖′ )(𝐗 ̂ −1
𝑖

where 𝑢𝑖̂ = 𝑌𝑖 − 𝑋𝑖′ 𝛽̂

23 / 48
Overidentification

• What if we have more instruments than endogenous variables?

• When there are more instruments than causal parameters
(𝑙 > 𝑘), the model is overidentified.
• When there are as many instruments as causal parameters
(𝑙 = 𝑘), the model is just identified.
• With more than one instrument and constant effects, we can
test for the plausibility of the exclusion restriction(s) using an
overidentification test.
• Is it plausible to find more than one instrument?

24 / 48
Overidentification tests
• Sargan-Hausman test:
▶ Under the null of all valid instruments, using all instruments
versus a subset should only differ by sampling variation.
▶ Regress 2SLS residuals, 𝜀𝑖̂ on 𝑋𝑖 and calculate 𝑅𝑢2 from this
regression.
▶ Under the null (and homoskedasticity), 𝑁𝑅𝑢2 ∼ Χ2𝑙−𝑘 .
▶ Degrees of freedom depends on how many overidentifying
restrictions there are.
• If we reject the null hypothesis in these overidentification
tests, then it means that the exclusion restrcitions for our
instruments are probably incorrect.
• Note that it won’t tell us which of them are incorrect, just
that at least one is.
• These overidentification tests depend heavily on the constant
effects assumption

25 / 48
3/ IV with
heterogenous
treatment effects
26 / 48
Instrumental Variables and
Potential Outcomes
• Basic idea of IV:
▶ 𝐷𝑖 not randomized, but 𝑍𝑖 is
▶ 𝑍𝑖 only affects 𝑌𝑖 through 𝐷𝑖

• 𝐷𝑖 now depends on 𝑍𝑖 ⇝ potential treatments:

𝐷𝑖 (1) = 𝐷𝑖 (𝑧 = 1) and 𝐷𝑖 (0).
• Consistency:

𝐷𝑖 = 𝑍𝑖 𝐷𝑖 (1) + (1 − 𝑍𝑖 )𝐷𝑖 (0)

• Outcome can depend on both the treatment and the

instrument: 𝑌𝑖 (𝑑, 𝑧) is the outcome if unit 𝑖 had received
treatment 𝐷𝑖 = 𝑑 and instrument value 𝑍𝑖 = 𝑧.

27 / 48
Key assumptions

1. Randomization
2. Exclusion Restriction
3. First-stage relationship
4. Monotonicity

28 / 48
Randomization

• Need the instrument to be randomized:

[{𝑌𝑖 (𝑑, 𝑧), ∀𝑑, 𝑧}, 𝐷𝑖 (1), 𝐷𝑖 (0)] ⟂⟂ 𝑍𝑖

• We can weaken this to conditional ignorability

• But why believe conditional ignorability for the instrument but
not the treatment?
• Best instruments are truly randomized.
• Identifies the intent-to-treat (ITT) effect:

𝐸[𝑌𝑖 |𝑍𝑖 = 1] − 𝐸[𝑌𝑖 |𝑍𝑖 = 0] = 𝐸[𝑌𝑖 (𝐷𝑖 (1), 1) − 𝑌𝑖 (𝐷𝑖 (0), 0)]

29 / 48
Exclusion Restriction

• The instrument has no direct effect on the outcome, once we

fix the value of the treatment.

𝑌𝑖 (𝑑, 1) = 𝑌𝑖 (𝑑, 0) for 𝑑 = 0, 1

• Given this exclusion restriction, we know that the potential

outcomes for each treatment status only depend on the
treatment, not the instrument:

𝑌𝑖 (1) ≡ 𝑌𝑖 (1, 1) = 𝑌𝑖 (1, 0)

𝑌𝑖 (0) ≡ 𝑌𝑖 (0, 1) = 𝑌𝑖 (0, 0)

• NOT A TESTABLE ASSUMPTION

30 / 48
The linear model with
heterogeneous effects

• As usual, rewrite 𝑌𝑖 using consistency:

𝑌𝑖 = 𝑌𝑖 (0) + (𝑌𝑖 (1) − 𝑌𝑖 (0))𝐷𝑖

= 𝛼 0 + 𝜏 𝑖 𝐷 𝑖 + 𝜂𝑖

• Here, we have 𝛼0 = 𝐸[𝑌𝑖 (0)] and 𝜏𝑖 = 𝑌𝑖 (1) − 𝑌𝑖 (0).

31 / 48
First Stage

• This next assumption is a little mundane, but turns out to be

very important: the instrument must have an effect on the
treatment.
𝐸[𝐷𝑖 (1) − 𝐷𝑖 (0)] ≠ 0
• Otherwise, what would we be doing? The instrument
wouldn’t affect anything.
• Implies that Cov(𝐷𝑖 , 𝑍𝑖 ) ≠ 0

32 / 48
Monotonicity

• Lastly, we need to make another assumption about the

relationship between the instrument and the treatment.
• Monotonicity says that the presence of the instrument never
dissuades someone from taking the treatment:

𝐷𝑖 (1) − 𝐷𝑖 (0) ≥ 0

• Note if this holds in the opposite direction 𝐷𝑖 (1) − 𝐷𝑖 (0) ≤ 0,

we can always rescale 𝐷𝑖 to make the assumption hold.

33 / 48
Monotonicity means no defiers

• This is sometimes called no defiers.

• With a binary treatment and a binary instrument, there are
four groups:
Name 𝐷𝑖 (1) 𝐷𝑖 (0)
Always Takers 1 1
Never Takers 0 0
Compliers 1 0
Defiers 0 1
• These compliance groups are sometimes called principal strata.
• The monotonicity assumption remove the possibility of there
being defiers in the population.
• Anyone with 𝐷𝑖 = 1 when 𝑍𝑖 = 0 must be an always-taker and
anyone with 𝐷𝑖 = 0 when 𝑍𝑖 = 1 must be a never-taker.

34 / 48
Local Average Treatment Effect
(LATE)

• Under these four assumptions, the Wald estimator is equal

what we call Local average treatment effect (LATE) or the
complier average treatment effect (CATE).
• This is is the ATE among the compliers: those that take the
treatment when encouraged to do so.
• That is, the LATE theorem, states that:

𝐸[𝑌𝑖 |𝑍𝑖 = 1] − 𝐸[𝑌𝑖 |𝑍𝑖 = 0]

= 𝐸[𝑌𝑖 (1) − 𝑌𝑖 (0)|𝐷𝑖 (1) > 𝐷𝑖 (0)]
𝐸[𝐷𝑖 |𝑍𝑖 = 1] − 𝐸[𝐷𝑖 |𝑍𝑖 = 0]
• This fact was a massive intellectual jump in our understanding
of IV.

35 / 48
Proof of the LATE theorem
• Under the exclusion restriction and randomization,

𝐸[𝑌𝑖 |𝑍𝑖 = 1] = 𝐸[𝑌𝑖 (0) + (𝑌𝑖 (1) − 𝑌𝑖 (0))𝐷𝑖 |𝑍𝑖 = 1]

= 𝐸[𝑌𝑖 (0) + (𝑌𝑖 (1) − 𝑌𝑖 (0))𝐷𝑖 (1)] (randomization)
• The same applies to when 𝑍𝑖 = 0, so we have

𝐸[𝑌𝑖 |𝑍𝑖 = 0] = 𝐸[𝑌𝑖 (0) + (𝑌𝑖 (1) − 𝑌𝑖 (0))𝐷𝑖 (0)]

• Thus, 𝐸[𝑌𝑖 |𝑍𝑖 = 1] − 𝐸[𝑌𝑖 |𝑍𝑖 = 0] =

𝐸[(𝑌𝑖 (1) − 𝑌𝑖 (0))(𝐷𝑖 (1) − 𝐷𝑖 (0))]

=𝐸[(𝑌𝑖 (1) − 𝑌𝑖 (0))(1)|𝐷𝑖 (1) > 𝐷𝑖 (0)] Pr[𝐷𝑖 (1) > 𝐷𝑖 (0)]
+𝐸[(𝑌𝑖 (1) − 𝑌𝑖 (0))(−1)|𝐷𝑖 (1) < 𝐷𝑖 (0)] Pr[𝐷𝑖 (1) < 𝐷𝑖 (0)]
=𝐸[𝑌𝑖 (1) − 𝑌𝑖 (0)|𝐷𝑖 (1) > 𝐷𝑖 (0)] Pr[𝐷𝑖 (1) > 𝐷𝑖 (0)]
• The third equality comes from monotonicity: with this
assumption, 𝐷𝑖 (1) < 𝐷𝑖 (0) never occurs.
36 / 48
Proof (continued)

𝐸[𝑌𝑖 |𝑍𝑖 = 1]−𝐸[𝑌𝑖 |𝑍𝑖 = 0] = 𝐸[𝑌𝑖 (1)−𝑌𝑖 (0)|𝐷𝑖 (1) > 𝐷𝑖 (0)] Pr[𝐷𝑖 (1) > 𝐷𝑖 (0)]

• We can use the same argument for the denominator:

𝐸[𝐷𝑖 |𝑍𝑖 = 1] − 𝐸[𝐷𝑖 |𝑍𝑖 = 0] = 𝐸[𝐷𝑖 (1) − 𝐷𝑖 (0)]

= Pr[𝐷𝑖 (1) > 𝐷𝑖 (0)]

• Dividing these two expressions through gives the LATE.

37 / 48
Is the LATE useful?

• Once we allow for heterogeneous effects, all we can estimate

with IV is the effect of treatment among compliers.
• This is a unknown subset of the data.
▶ Treated units are a mix of always takers and compliers.
▶ Control units are a mix of never takers and compliers.

• Without further assumptions, 𝜏𝐿𝐴𝑇 𝐸 ≠ 𝜏𝐴𝑇 𝐸 .

• Complier group depends on the instrument ⇝ different IVs
will lead to different estimands.
• 2SLS “cheats” by assuming that the effect is constant, so it is
the same for compliers and non-compliers.

38 / 48
Randomized trials with one-sided
noncompliance
• Will the LATE ever be equal to a usual causal quantity?
• When non-compliance is one-sided, then the LATE is equal to
the ATT.
• Think of a randomized experiment:
▶ Randomized treatment assignment = instrument (𝑍𝑖 )
▶ Non-randomized actual treatment taken = treatment (𝐷𝑖 )
• One-sided noncompliance: only those assigned to treatment
(control) can actually take the treatment (control). Or

𝐷𝑖 (0) = 0∀𝑖 ⇝ Pr[𝐷𝑖 = 1|𝑍𝑖 = 0] = 0

• Maybe this is because only those treated actually get pills or

only they are invited to the job training location.

39 / 48
Benefits of one-sided
noncompliance
• One-sided noncompliance ⇝ no “always-takers” and since
there are no defiers,
▶ Treated units must be compliers.
▶ ATT is the same as the LATE.
• Thus, we know that: 𝐸[𝑌𝑖 |𝑍𝑖 = 1] − 𝐸[𝑌𝑖 |𝑍𝑖 = 0] =
𝔼[𝑌𝑖 (0) + (𝑌𝑖 (1) − 𝑌𝑖 (0))𝐷𝑖 |𝑍𝑖 = 1] − 𝔼[𝑌𝑖 (0)|𝑍𝑖 = 0]
(exclusion restriction + one-sided noncompliance)
=𝔼[𝑌𝑖 (0)|𝑍𝑖 = 1] + 𝐸[(𝑌𝑖 (1) − 𝑌𝑖 (0))𝐷𝑖 |𝑍𝑖 = 1] − 𝔼[𝑌𝑖 (0)|𝑍𝑖 = 0]
=𝔼[𝑌𝑖 (0)] + 𝔼[(𝑌𝑖 (1) − 𝑌𝑖 (0))𝐷𝑖 |𝑍𝑖 = 1] − 𝔼[𝑌𝑖 (0)]
(randomization)
=𝔼[(𝑌𝑖 (1) − 𝑌𝑖 (0))𝐷𝑖 |𝑍𝑖 = 1]
=𝔼[𝑌𝑖 (1) − 𝑌𝑖 (0)|𝐷𝑖 = 1, 𝑍𝑖 = 1] Pr[𝐷𝑖 = 1|𝑍𝑖 = 1]
(law of iterated expectations + binary treatment)
=𝔼[𝑌𝑖 (1) − 𝑌𝑖 (0)|𝐷𝑖 = 1] Pr[𝐷𝑖 = 1|𝑍𝑖 = 1]
(one-sided noncompliance) 40 / 48
• Noting that Pr[𝐷𝑖 = 1|𝑍𝑖 = 0] = 0, then the Wald estimator is
just the ATT:
𝐸[𝑌𝑖 |𝑍𝑖 = 1] − 𝐸[𝑌𝑖 |𝑍𝑖 = 0]
= 𝐸[𝑌𝑖 (1) − 𝑌𝑖 (0)|𝐷𝑖 = 1]
Pr[𝐷𝑖 = 1|𝑍𝑖 = 1]
• Thus, under the additional assumption of one-sided
compliance, we can estimate the ATT using the usual IV
approach

41 / 48
4/ IV extensions

42 / 48
Falsification tests
𝑈

𝑍 𝐷 𝑌

exclusion restriction

• The exclusion restriction cannot be tested directly, but it can

be falsified.
• Falsification test Test the reduced form effect of 𝑍𝑖 on 𝑌𝑖 in
situations where it is impossible or extremely unlikely that 𝑍𝑖
could affect 𝐷𝑖 .
• Because 𝑍𝑖 can’t affect 𝐷𝑖 , then the exclusion restriction
implies that this falsification test should have 0 effect.
• Nunn & Wantchekon (2011): use distance to coast as an
instrument for Africans, use distance to the coast in an Asian
sample as falsification test.
43 / 48
Nunn & Wantchekon falsification
test

44 / 48
Size, characteristics of the
compliers

• While we cannot identify who is a complier and who is not a

complier in general, we can estimate the size of the complier
group:

Pr[𝐷𝑖 (1) > 𝐷𝑖 (0)] = 𝐸[𝐷𝑖 (1)−𝐷𝑖 (0)] = 𝐸[𝐷𝑖 |𝑍𝑖 = 1]−𝐸[𝐷𝑖 |𝑍𝑖 = 0]

• Can extend this to calculate features of the complier group:

▶ Covariate means, variances, etc.
▶ Abadie (2003) shows how to weight the data to estimate these
quantities.

45 / 48
Multiple instruments

• Different instruments ⇝ different LATEs

▶ Instrument 1, 𝑍𝑖1 with LATE 𝜏1
▶ Instrument 2, 𝑍𝑖2 with LATE 𝜏2
• Use both in the first stage:

̂
𝐷𝑖 = 𝜋1 𝑍1𝑖 + 𝜋2 𝑍2𝑖 .

46 / 48
2SLS as weighted average

• MHE shows that the 2SLS estimator using these two

instruments is a weighted sum of the two component LATEs:

𝜌2𝑆𝐿𝑆 = 𝜓𝜏1 + (1 − 𝜓)𝜏2 ,

where the weights are:

𝜋1 Cov(𝐷𝑖 , 𝑍1𝑖 )
𝜓=
𝜋1 Cov(𝐷𝑖 , 𝑍1𝑖 ) + 𝜋2 Cov(𝐷𝑖 , 𝑍2𝑖 )
• Thus, the 2SLS estimate is a weighted average of causal
effects for each instrument, where the weights are related to
the strength of first-stage.

47 / 48
Covariates and heterogeneous
effects
• It might be the case that the above assumptions only hold
conditional on some covariates, 𝑋𝑖 . That is, instead of
randomization, we might have conditional ignorability:

[{𝑌𝑖 (𝑑, 𝑧), ∀𝑑, 𝑧}, 𝐷𝑖 (1), 𝐷𝑖 (0)] ⟂⟂ 𝑍𝑖 |𝑋𝑖

• We would also have exclusion conditional on the covariates:

Pr[𝑌𝑖 (𝑑, 0) = 𝑌𝑖 (𝑑, 1)|𝑋𝑖 ] = 1 for 𝑑 = 1, 0

• Under these assumptions, with fully saturated first and second

stages, then 2SLS estimates a weighted average of the
covariates-specific LATEs (very similar to regression).
• Abadie (2003) shows how to estimate the overall LATE using
a weighting approach based on a “propensity score” for the
instrument.
48 / 48