PMDA

Lecture 7: Instrumental Variables (IV)

A few updates and reminders

Problem sets:
● Problem Set 3: Panel Data & Diff-in-diff due Feb 22th
● Problem Set 4: Lasso Regression due March 15th
Agenda today

● Another method for causal inference in observational data:

○ Instrumental Variables (IV) regression
○ Workshop
 Instrumental Variables (IV) regression

● IV is a general method to eliminate endogeneity, i.e. when you have a

regression model

with cov(X,e)≠0
● The key requirement is that you must find (at least) another variable, an
instrument Z, that satisfies two conditions:

1. (Relevance) it is correlated with X: cov(Z,X)≠0

2. (Exogeneity) it is not correlated with e: cov(Z,e)=0

IV in a causal graph

Z X

● Visualization of the two conditions for instrument validity:

○ Relevance = there is a link between Z and X
○ Exogeneity = there is no direct link between Z and Y

● Q: How is this different from the causal graphs with omitted variables we
have considered before?
● A: An omitted variable by definition is part of the error e, so it must have a
direct link to Y
The two-stage least squares (TSLS) estimator

● Idea: use Z to isolate the variation in X that is not correlated with e (i.e., the
exogenous variation in X) and use it to estimate β_1

● TSLS procedure. Run two regressions:

1. First-stage regression of X on Z

Exogenous part is the predicted value,

2. Second-stage regression of Y on predicted value

Because is uncorrelated with e (why?), the estimator of β_1 is unbiased

 Not to be confused with the Frisch-Waugh theorem

The example we saw of the Frisch-Waugh theorem:

1. First regress X on Z

2. Then regress Y on the residual

Q: What are the differences?

A:
Frisch-Waugh TSLS
● Z is an omitted variable that causes ● Z is exogenous
endogeneity if not included ● Z has a no direct effect on Y
● Z has a direct effect on Y ● It is the predicted value from
● It is the residual from the stage 1 the stage 1 regression that
regression that eliminates the endogeneity eliminates the endogeneity
Example 1. Demand estimation

● β_1 is demand elasticity = % change in quantity for a 1% change in price

(this is the interpretation in a log-log model)
● Data: you observe prices and quantities in different regions or time periods
● We saw that this is an example of simultaneous causality bias: there is a
supply curve in the background so P and Q are determined in equilibrium
Example 1. Demand estimation

● Fitting a regression through the scatterplot gives a biased estimate of

elasticity
● But, what if only the supply curve shifted and the demand stayed constant?

● TSLS estimates the demand curve by isolating shifts in price and quantity
that arise from shifts in supply
● Z is a variable that shifts supply but not demand
 Instruments for demand estimation

Examples of instruments for the regression

1. Suppose the demand is for milk and use Z = rain in dairy-producing regions
○ Q: Relevant?

A: Less rain → less milk produced → higher price

○ Q: Exogenous?
A: Rain doesn’t directly affect how much milk you buy
2. Suppose the demand is for cigarettes and use Z = sales tax in different US
states
○ Q: Relevant?

A: Sales tax affects all prices so also cigarette price

○ Q: Exogenous?
A: Sales tax not determined by cigarette market
Another Example

Does putting criminals in jail reduce crime? (Levitt 1996)

● You have data of crime rates and incarceration rates at the state level
● You want to measure how crime rates are affected by incarceration rates
● You also have data on economic conditions, age, and other demographics

What type of bias can we have?

A: Simultaneous bias:
more crime → more prisoners (if police do their job)
more prisoners → less crime

What instrument would you use?

A: “Lawsuits aimed at reducing prison overcrowding”
● Relevance: They slow the growth of prisoner incarcerations
● Exogeneity: Only if overcrowding litigation is induced by prison
conditions but not by crime rate or its determinants
Instruments can isolate “as if random” variation

● Say we want to understand the effect of class size on students’ performance

test_scores= β_0+β_1 class_size + e
○ Q: Endogeneity problem in this regression?
○ A: OVB, for example due to parental income/involvement
● Suppose you have a natural experiment: an earthquake makes affected
districts to double up classrooms in some schools (others have to close)
● Let Z = distance to the epicenter
○ Q: Is this a valid instrument?
○ A: Relevant because class size is correlated with the distance to the
epicenter: districts close to epicenter are most severely affected
○ Exogenous if distance to epicenter is unrelated to any other factors
affecting student performance such as being english learner or other
disruptive effects of the earthquake on performance
○ Q: Can you explain in words how TSLS gives you a causal estimate?
○ A: The first stage regression isolates the variation in class size that is “as
if randomly assigned”. Using this part in the second stage regression
thus gives a causal estimate
Discussion Question

Suppose you wish to measure the impact of smoking on the weight of

newborns
You are planning to use the following model,

where
● bw is the birth weight,
● male is a dummy variable equal to 1 if the baby is a boy or 0 otherwise,
● order is the birth order of the child,
● y is the log income of the family,
● cig is the amount of cigarettes per day smoked during pregnancy,
● i indexes the observation
● β’s are the unknown parameters.

Q1: What could be the problem in using OLS to estimate the above model?
Q2: Suppose you have data on the average price of cigarettes in the state of
residence. Would this information help you to identify the true parameters of
the model?
Generalizations

● We can have more than one instrument: Z_1,…,Z_m

● We can also add control variables: W_1,…,W_r

1. First-stage regression: regress X on Z_1,…,Z_m and W_1,…,W_r and get

predicted value

2. Second-stage regression: regress Y on and W_1,…,W_r

○ Note: the standard errors from this regression are not the correct
ones because they do not account for estimation uncertainty in
○ Traditional softwares/packages account for this when computing
standard errors
Generalizations

● We can also have more endogenous regressor: X_1,…,X_k

● In this case, we need that the model is exactly identified or over identified:
○ Exactly identified: when m=k, that is, same number of instruments and
endogenous regressors
○ Overidentified: when m>k
○ Underidentified: when m<k

● In the latter case, we cannot estimate all the model parameters

Can you test for instrument validity?

● Requirements for validity when you have multiple instruments Z_1,…,Z_m

1. Relevance: at least one instrument is correlated with X (a bit more
involved when we have multiple endogenous regressors, see Stock & Watson
Chapter 12)
2. Exogeneity: all the instruments are uncorrelated with e

● Can you test these conditions?

○ You can test for relevance by looking at the first-stage regression
○ You can test for exogeneity but only when you have more than one
instrument
 Testing relevance

● Relevance is akin to sample size: the more relevant the instruments, the more
variation in X is explained by the instruments, so that we have more info for the
IV regression => More relevant implies more accurate estimator

● First stage regression:

○ The instruments are said to be weak if all the a_1,…, a_m are zero or nearly
zero
○ Weak instruments explain little of the variation of X:
■ E.g., something that shifts the supply curve but only by little
○ Weak instruments imply that the TSLS estimator is biased towards the OLS
estimator, and usual statistical inference (standard errors, hypothesis tests)
is misleading
 Testing relevance

● You can test for weak instruments by:

○ Consider the F-statistic testing the hypothesis that a_1,…, a_m are all zero
○ Rule of thumb: if first-stage F-statistic < 10, you have weak instruments
○ Note that simply rejecting the null hypothesis that the coefficients are zero
isn’t enough – you need the F-statistic to be large (> 10)
○ There are more formal tests for weak instruments (e.g., built-in in python
and R)
 Testing relevance

What if you have weak instruments?

○ If you have many: a few strong and many weak instruments, drop some of
the weak until you get F-stat >10

○ Your standard error may increase but keep in mind that the original
standard errors were not meaningful

○ If the coefficients are exactly identified and you don’t have strong enough
instruments:
■ Try to find additional stronger instruments
■ Proceed with weak instruments but use other methods (e.g., Limited
information maximum likelihood estimator, LIML)
Testing exogeneity

● You can test for exogeneity only if you have more than one instrument
● The test is called the “J-test of overidentifying restrictions”: it tests that all
instruments are exogenous (H_0)
● Intuition:
○ Suppose there are two instruments → could compute two separate
TSLS estimates → if estimates are very different something must be
wrong: one of the two instruments (or both) must be invalid
○ The J-test makes this comparison in a statistically precise way
Testing exogeneity
More formally:
1. Build:

2. Estimate:

3. Test whether the delta’s are zero by computing the F-statistic for:
delta_1=...=delta_m=0

Why can’t we do the same approach when m=k (same instruments as

endogenous regressors)?

Note: If the J-test rejects, you don’t know which instruments are endogenous.
In this case you need to use your knowledge of the problem to decide what is
the most appropriate next step (see workshop)
IV regression in Python

● The TSLS estimator is computed in Python using the command IV2SLS

from linearmodels.iv: “from linearmodels.iv import IV2SLS”
● Syntax is IV2SLS (Y,[const,W], X, Z ) where X is the endogenous variable, Z
the instrument(s) and W the control(s)
● Diagnostic tests can be run by
○ iv_reg.first_stage: Reports statistics from the first stage include the partial
F-statistic which tests for relevance (weakness of instruments)

○ iv_reg.sargan: This is the J-test of instrument exogeneity. It can only be

used if you have more instruments than endogenous regressors. The null
hypothesis is of exogeneity
IV regression in R

● The TSLS estimator is computed in R using the command ivreg

● Syntax is ivreg (Y ~ X + W | W + Z ) where X is the endogenous variable, Z
the instrument(s) and W the control(s)
● Diagnostic tests can be run by adding ,diagnostics = TRUE inside the
summary command. This reports:
○ Weak instruments: This tests the null hypothesis that we have weak
instruments
○ Wu-Hausman (we didn’t discuss this)
○ Sargan: This is the J-test of instrument exogeneity. It can only be used if
you have more instruments than endogenous regressors. The null
hypothesis is of exogeneity