Endogeneity and

Instrumental
Variables
Presented by Justin Balthrop
September 28, 2015

What Exactly Are Instrumental

Variables?

Take the simple case of ordinary least squares regression with a single
explanatory variable:

A fundamental assumption of estimating _1 is that the correlation between

X and u is zero.

If this is not the case (X is endogenous), then using instrumental variables Z

can essentially detect the part of X which is *not* correlated with the error
term.

An Instrument must satisfy

Relevance and Exclusion

Loosely speaking, the relevance restriction imposes that the instrument Z is

non-trivially related to the endogenous regressor X.

Exclusion, or exogeneity requires that Z not be systematically related to the

error term, u.

Why do we need IV?

Before we worry about external validity and the big picture implications of
our results, we need to satisfy internal validity.

Three main sources of internal validity issues are:

Omitted variables bias
Simultaneity bias
Errors-in-variables bias

Appropriately instrumenting for endogenous regressors can eliminate these

biases

Univariate IV Estimation: Two Stage

Least Squares

Stage 1: Identify the portion of X that is uncorrelated with the error, u

This gives estimates of _0 and _1, which are used to get predicted X values:

Stage 2: Replace X values with estimated X

Underlying Assumptions of 2SLS

Instrument validity

_0 and _1 are well-estimated in the first stage (large samples)

Why it works:

Careful about Standard Errors

Second-stage OLS standard errors are not correct

They need to be adjusted for the fact that the explanatory variables are estimated
See Woolridge for the math, STATA for the code- ivreg, robust

Other considerations:

Heteroskedasticity
Appropriate clustering
Instrument relevance- more relevance lower estimator variance and higher Rsquared in the first stage

IV Regression in a Multivariate Model

Aside from messy algebra, estimation generalizes rather easily

Key identification criterion: at least as many Z as endogenous X

Dont forget to instrument for interactions between endogenous X

Underidentified = too few to estimate _vec (correctly, anyway)
Exactly identified = equal number of Z and endogenous X
Overidentified = too many instruments

Testing for Instrument Relevance

Assume one endogenous X

First stage regress is therefore:

Relevance comes from at least one _i different from zero

If not, the instrument is weak

Why are weak instruments so bad?

Back to the simple model:

With estimator:

Weak instruments leads to a near-zero denominator, and the resulting

sampling distribution cannot be accurately approximated by its asymptotic
distribution

Measuring the Strength of Instruments

The first-stage F-test

Tests the hypothesis that instruments Z_i do not enter the first-stage regression
Small F-stat (less than 10) are the result of weak instruments

If the set of instruments is weak, get better instruments

If that is impossible, consider dropping the weakest to improve the first-stage F

This is somewhat ad-hoc

Too many Instruments = Tests for

Overidentifying Restrictions

Assume we have multiple valid instruments with a single endogenous

regressor.

Intuition: If we perform 2SLS using both instruments separately and arrive

at completely different results, it shouldnt be that both instruments are
valid.

Statistics: J-test

J-Test of Overidentifying Restrictions

Step 1: estimate the conditional expectation function using TSLS and both instruments

Step 2: Compute predicted Y values using the actual Xs

Step 3: Compute residuals

Step 4: Regress residuals against all instruments Z and exogenous regressors X

Step 5: Test the hypothesis that all coefficients on Z_i are zero, with J-statistic J= mF

Here, F is the F-stat from testing coefficients on Z_i

If some instruments are exogenous and others endogenous, J-stat will be large, rejecting the
null that all instruments are exogenous

Sargan Test for Overidentification

1. Estimate the 2SLS IV regression - Extract residuals
2. Regress these residuals on all exogenous variables and extract R2
3. Calculate nR2 which is 2 distributed
4. Compare the value with the critical value in the chi-square table with
degrees of freedom equal to # instruments less #

If the statistic (nR2) exceeds the critical 2 value, conclude the instruments
are invalid.
They are not uncorrelated with the error term and hence has some explanatory
power in the main equation.
Be very careful: The test assumes that one instrument is valid.

If all instruments do not fulfill the criteria Cov(zi,ui) = 0, then the test might
suggest that the instruments are valid, even when they are not

Durbin-Wu-Hausman Test

Balances the consistency of IV against the efficiency of LS

H0: IV and LS both consistent, but LS is efficient
H1: Only IV is consistent

DWH test for a single endogenous regressor:

DWH = (bIV bLS) / (s2bIV s2bLS) ~ N(0,1)

If |DWH| > 1.96, then X is endogenous and IV is the preferred estimator despite its
inefficiency

A roughly equivalent procedure for DWH:

1. Estimate the first-stage model

2. Include the first-stage residual in the structural model along with the endogenous X
3. Test for significance of the coefficient on residual

Note: Coefficient on endogenous X in this model is b IV (standard error is

smaller, though)
First-stage residual is a generated regressor

The

following example is
taken from the University of
Albany Center for Social and
Demographic Analysis
presentation on IV
Estimation

Angrist and Krueger (1991), J.L.E.

Returns to education (Y = wages)

Problem of omitted ability bias

Years of schooling vary by quarter of birth

Compulsory schooling laws, age-at-entry rules

Someone born in Q1 is a little older and will be able to drop out sooner than someone born in Q4

Q.O.B. can be treated as a useful source of exogeneity in schooling

Angrist and Krueger (1991), J.L.E.

People born in Q1 do obtain less

schooling

But pay close attention to the scale of

the y-axis
Mean difference between Q1 and Q4
is only 0.124, or 1.5 months

So...need large N since R2X,Z will

be very small
A&K had over 300k for the 1930-39
cohort

Source: Angrist and Krueger (1991), Figure I

Angrist and Krueger (1991),

J.L.E.

Final 2SLS model interacted QOB with year of birth (30), state of birth (150)
OLS: b = .0628 (s.e. = .0003)
2SLS: b = .0811 (s.e. = .0109)

Least squares estimate does not appear to be badly biased by omitted variables
But...replication effort identified some pitfalls in this analysis that are instructive

Bound, Jaeger, and Baker (1995), J.A.S.A.

Potential problems with QOB as an IV

Correlation between QOB and schooling is weak

Small Cov(X,Z) introduces finite-sample bias, which will be exacerbated with the inclusion of many IVs

QOB may not be completely exogenous

Even small Cov(Z,e) will cause inconsistency, and this will be exacerbated when Cov(X,Z) is small

QOB qualifies as a weak instrument that may be correlated with unobserved determinants of
wages (e.g., family income)

Bound, Jaeger, and Baker (1995), J.A.S.A.

Even if the instrument is good, matters can be made far worse with IV as opposed to LS
Weak correlation between IV and endogenous regressor can pose severe finite-sample bias

Andreally large samples wont help, especially if there is even weak endogeneity between IV and error

First-stage diagnostics provide a sense of how good an IV is in a given setting

F-test and partial-R2 on IVs

Lewbel (2012) Method of Identification

Mostly applicable to models with an unobserved common factor

Identification is achieved by having regressors that are uncorrelated with the product of
heteroskedastic errors

ConsiderY1,Y2 as observed endogenous variables, X a vector of observed exogenous

regressors, and =(1,2) as unobserved error processes.

Consider a structural model of the form:

Y1 = X1+Y21+1 (1)
Y2 = X2+Y12+2 (2)

Higher-moment considerations (restricting correlations of with X);

In the presence of heteroskedasticity related to at least some elements of X,
identification can be achieved.