6/19/2024

Endogeneity: Problem Example: Omitted Variables bias problem

• Y = β0 + β1Y1 + β2X1 + u
• Full model:
• If Cov(Y1,u) ≠ 0
log(wage) = β0 + β1edu + β2abil + β3exper + u (1)
→ Endogeneity: OLS unbiased and inconsistent
• If "abil" (ability) cannot be collected → (1) faces Omitted Variables bias
Y1 Y problem. The model will become:
log(wage) = β0 + β1edu + β3exper + u (2)
Note:
• Y1, Y2, ... Yn : Endogenous variables. • Theoretical perspective: Cov(edu, abil) ≠ 0. Since u contains abil →
• X1, X2, ... Xn : Exogenous variables.
u Cov(edu,u) ≠ 0 → (2) gets biased and inconsistent estimator of β1.

Omitted Variables bias problem: Biased and inconsistent

u contains an omitted variable which correlate with Y1 estimators

Vu Huu Thanh 2 Vu Huu Thanh 3

2 3

Endogeneity Solution stage 2: Structural

Endogeneity Solution stage 1: Choosing Zs
equation
• New instrument variable (IV) z satisfies two condions: • Regress zi on Y1:
Cov(z,Y1) ≠ 0 Y1 = b0 + b1X1 + b2Z2 + b3Z3 + v1 (1)
Cov(z,u) = 0 → z is not correlated with omitted variable contained in u. • Calculate Y1* (fitted value):
Y1 = Y1*+ v1
Z Y1 Y
• Replace Y1 by Y1*+ v1 then estimate the main model
Y = β0 + β1(Y1*+ v1) + β2X1 + u = β0 + β1Y1* + β2X1 + u + β1v1 (2)
x
u
X1, Z2, Z3: Instruments for Y1

z is called as an instrumental variable for Y1 (1) and (2): Structural equation

Vu Huu Thanh 4 Vu Huu Thanh 5

4 5

1
 6/19/2024

Y = β0 + β1Y1* + β2X1 + (u + β1v1) Instrumental Variables: Notation

Composite error

• Y, Y1: endogenous variables (correlated with u)

• Cov(Y1*,u) = 0
• Z, X: exogenous variables (uncorrelated with u)
• E(u + β1v1| Y1*,X1) = 0
• X: included instruments, clean variables (“controls”)
• Cov(u + β1v1, Y1*) = Cov(u + β1v1, Z1) =0
• Z: excluded instruments, IVs (excluded = not included in the structural equation.)

Structural equation ෡𝟎 + 𝜷
෡ 𝟏 𝒀𝟏 + ෍ 𝜽
෡𝒊 𝑿𝒋 + 𝒖
IV estimation – instrument variable estimation or 2SLS – Two stages least 𝒀= 𝜷
square) 𝒏

First stage equation 𝒀𝟏 = 𝝅

ෝ𝟎 + ෍ 𝝅 ෡𝒊 𝑿𝒋 + 𝒗
ෝ 𝒌 𝒁𝒌 + ෍ 𝒃
𝒋=𝟏

Second stage equation ෡𝟎 + 𝜷

𝒀=𝜷 ෡𝟏𝒀
෡𝟏 + ෍ 𝜽 ෡ 𝟏𝒗)
෡𝒊 𝑿𝒋 + (𝒖 + 𝜷

Vu Huu Thanh 6 Vu Huu Thanh 7

6 7

IVs requirement
A regressor can be correlated with the error term when:
• Omitted variables
• Selection bias • Z has a causal effect on Y1
• Measurement error IVs • Z affects Y only through Y1 (Z does not have a direct effect on Y)
• Simultaneous equations
• There is no confounding for the effect of Z on Y.
• Misspecification
• Correlated shocks across linked equations
• Model has a lagged dependent variable and a serially correlated error term

Vu Huu Thanh 8 Vu Huu Thanh 9

8 9

2
 6/19/2024

IVs: Important Notation

1. Endogenous variables are variables correlated with u:
• Test of endogeinity • data: wages2
2. Instrumental variables must be correlated with the endogenous regressors → • describe
They must be not weak:
• Test of weak identification • dependent variable: wages
• Test of underidentification • independent (exogenous) independent: exper, expersq
3. The instruments must be uncorrelated with the structural error term. • instruments: motheduc, fatheduc
• Test of overidentification
Note: If the model is overidentified (the number of excluded instruments exceeds the number of
endogenous regressors), we can perform above test. If the model is just identified, we cannot.

Vu Huu Thanh 10 Vu Huu Thanh 11

10 11

Diagnostic tests
• Ommited variable:
log(wage) = β0 + β1educ + β2exper + β3exper2 + u1
regress lwage educ exper expersq
1. Test of endogeneity
2. Test of overidentifying restrictions
• Employ motheduc and fatheduc as instruments 3. Test of weak identification
Detecting weak instruments
• Reduce form equation for educ – first stage: 4. Test of underidentification
Educ = b0 + b1exper + b2exper2 + b1motherduc + b1fatherduc + v1
• Second stage (replace educ with edu* and v1)
log(wage) = β0 + β1educ* + β2exper + β3exper2 + u1 + β1v1
ivregress 2sls lwage exper expersq (educ = motheduc fatheduc), first
ivreg2 lwage exper expersq (educ = motheduc fatheduc), first

Vu Huu Thanh 12 Vu Huu Thanh 13

12 13

3
 6/19/2024

Test of endogeneity Test of overidentification

• H0: Variables are exogenous
• H1: Variables are endogenous • Sargan’s test, Hansen’s test or Wooldridge’s robust score test
• Reject H0 → Endogeneity • H0: All instruments are uncorrelated with u

ivregress 2sls lwage exper expersq (educ = fatheduc), first → No overidentifying restrictions
estat overid → The model is just identified

ivregress 2sls lwage exper expersq (educ = motheduc fatheduc), first

estat overid

ivreg2 lwage exper expersq (educ = motheduc fatheduc), first endog(educ)

ivregress 2sls lwage exper expersq (educ = motheduc fatheduc), first
estat endogenous

ivreg2 lwage exper expersq (educ = motheduc fatheduc), first endog(educ)

Vu Huu Thanh 14 Vu Huu Thanh 15

14 15

Detection of weak instruments Test of weak identification

• Test whether instrument are weak
• If cov(IVs, E) is weak: 2SLS estimator can be biased. Weak instruments cause instrumental-variables
• Two types of test: Tests of underidentification and Tests of weak
estimators to be biased
identification.
• H0: Instruments are weak
• Goal: determine when instruments are irrelevant (test of underidentification)
• If the test statistic exceeds the critical value, we can conclude that our instruments are not weak
or weak (tests of weak identification)
• F > 10

ivregress 2sls lwage exper expersq (educ = motheduc fatheduc), first

estat firststage

ivreg2 lwage exper expersq (educ = motheduc fatheduc), first endog(educ)

Vu Huu Thanh 16 Vu Huu Thanh 17

16 17

4
 6/19/2024

Test of underidentification

• Test whether instruments are irrelevant (π = 0)

• If we reject underidentification, it can still be the case that our model is only
weakly identified since instruments are weak.

ivreg2 lwage exper expersq (educ = motheduc fatheduc), first endog(educ)

Vu Huu Thanh 18

18