Scribd
Upload
16 views12 pages

IV Regression with feols and ivreg

Uploaded by

sahrish.khan
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
16 views12 pages

IV Regression with feols and ivreg

Uploaded by

sahrish.khan
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

Tutorial 5

Sahrish Aisha Khan

2024-12-01
knitr::opts_chunk$set(echo = TRUE, warning = FALSE, message = FALSE)

library(fixest)
library(ivreg)
library(gmm)
library(flextable)
library(readxl)
library(tidyverse)
library(stargazer)
library(car)
library(lmtest)
library(modelsummary)

mroz <- read_excel("[Link]")


mroz <- [Link](mroz)

A. summary statistics
summary(mroz$wage)

## Min. 1st Qu. Median Mean 3rd Qu. Max.


## 0.000 0.000 1.625 2.375 3.788 25.000

summary(mroz$educ)

## Min. 1st Qu. Median Mean 3rd Qu. Max.


## 5.00 12.00 12.00 12.29 13.00 17.00
summary(mroz$exper)

## Min. 1st Qu. Median Mean 3rd Qu. Max.


## 0.00 4.00 9.00 10.63 15.00 45.00

2. OLS estimates of a model for log(wages) controlling for education, experience and
experience square.
# ols <- lm(log(wage) ~ educ + exper + I(exper^2), data=mroz)
# Problem: wages are zero for some individuals in the sample, and therefore
log(wage) is -infinity
# you need to construct a new dataset selecting only idividuals with positive
wages

mroz2 <-subset(mroz, wage>0)


ols <- lm(log(wage) ~ educ + exper + I(exper^2), data=mroz2)
summary(ols)
##
## Call:
## lm(formula = log(wage) ~ educ + exper + I(exper^2), data = mroz2)
##
## Residuals:
## Min 1Q Median 3Q Max
## -3.08404 -0.30627 0.04952 0.37498 2.37115
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -0.5220406 0.1986321 -2.628 0.00890 **
## educ 0.1074896 0.0141465 7.598 1.94e-13 ***
## exper 0.0415665 0.0131752 3.155 0.00172 **
## I(exper^2) -0.0008112 0.0003932 -2.063 0.03974 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.6664 on 424 degrees of freedom
## Multiple R-squared: 0.1568, Adjusted R-squared: 0.1509
## F-statistic: 26.29 on 3 and 424 DF, p-value: 1.302e-15

C. A test for weak instruments


strong1 <- lm(educ ~ exper + I(exper^2) + mothereduc, data=mroz2)
summary(strong1)

##
## Call:
## lm(formula = educ ~ exper + I(exper^2) + mothereduc, data = mroz2)
##
## Residuals:
## Min 1Q Median 3Q Max
## -7.4423 -1.2963 -0.0837 1.1761 5.9870
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 9.775103 0.423889 23.061 <2e-16 ***
## exper 0.048862 0.041669 1.173 0.242
## I(exper^2) -0.001281 0.001245 -1.029 0.304
## mothereduc 0.267691 0.031130 8.599 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 2.111 on 424 degrees of freedom
## Multiple R-squared: 0.1527, Adjusted R-squared: 0.1467
## F-statistic: 25.47 on 3 and 424 DF, p-value: 3.617e-15

anova(strong1)

## Analysis of Variance Table


##
## Response: educ
## Df Sum Sq Mean Sq F value Pr(>F)
## exper 1 0.52 0.52 0.1157 0.7339
## I(exper^2) 1 10.46 10.46 2.3479 0.1262
## mothereduc 1 329.56 329.56 73.9459 <2e-16 ***
## Residuals 424 1889.66 4.46
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# the F statistic for a test of not significance of the coefficient on the


variable mothereduc is 73.9, well above 10
# instrument is strong

The first step in testing for weak instruments is to estimate the first-stage regression,
where the endogenous variable is regressed on the instrument(s) and any exogenous
control variables. Use summary() to examine the coefficients for the instrument(s).
Significant coefficients (small p-values) suggest that the instrument is relevant. The F-
statistic of the first-stage regression is the key test for weak instruments. If F<10, the
instrument is considered weak. If F≥10, the instrument is considered strong enough.
D. 2SLS (two stage least square) estimator
Obtain the 2SLS (two stage least square) estimator performing all the steps:

- stage 1: regress endogenous variable on exogenous variables and instruments


- obtain predicted values
- stage 2: regress dependent variable on exogenous regressors and predicted
values form stage 1

#2SLS estimator

step1 <- lm(educ ~ exper + I(exper^2) + mothereduc, data=mroz2)


summary(step1)

##
## Call:
## lm(formula = educ ~ exper + I(exper^2) + mothereduc, data = mroz2)
##
## Residuals:
## Min 1Q Median 3Q Max
## -7.4423 -1.2963 -0.0837 1.1761 5.9870
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 9.775103 0.423889 23.061 <2e-16 ***
## exper 0.048862 0.041669 1.173 0.242
## I(exper^2) -0.001281 0.001245 -1.029 0.304
## mothereduc 0.267691 0.031130 8.599 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 2.111 on 424 degrees of freedom
## Multiple R-squared: 0.1527, Adjusted R-squared: 0.1467
## F-statistic: 25.47 on 3 and 424 DF, p-value: 3.617e-15

educhat <- fitted(step1)


step2 <- lm(log(wage) ~ exper + I(exper^2) + educhat, data = mroz2)
summary(step2)

##
## Call:
## lm(formula = log(wage) ~ exper + I(exper^2) + educhat, data = mroz2)
##
## Residuals:
## Min 1Q Median 3Q Max
## -3.15299 -0.34773 0.02906 0.39023 2.35624
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 0.1981861 0.4933427 0.402 0.68809
## exper 0.0448558 0.0141644 3.167 0.00165 **
## I(exper^2) -0.0009221 0.0004240 -2.175 0.03019 *
## educhat 0.0492630 0.0390562 1.261 0.20788
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.709 on 424 degrees of freedom
## Multiple R-squared: 0.04559, Adjusted R-squared: 0.03884
## F-statistic: 6.751 on 3 and 424 DF, p-value: 0.0001861

stage 1: regress endogenous variable on exogenous variables and instruments obtain


predicted values stage 2: regress dependent variable on exogenous regressors and
predicted values form stage 1
E. fixest package
Let’s use now the fixest package; use the following command where x1 and x2 are the
exogenous regressors, x3 is the endogenous regressor and z is the instrument(s):

feols(y ~ x1 + x2 | x3 ~ z, data = dataset)

iv <- feols(log(wage) ~ exper + I(exper^2) | educ ~ mothereduc, mroz2)


summary(iv)

## TSLS estimation - Dep. Var.: log(wage)


## Endo. : educ
## Instr. : mothereduc
## Second stage: Dep. Var.: log(wage)
## Observations: 428
## Standard-errors: IID
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 0.198186 0.472877 0.419107 0.6753503
## fit_educ 0.049263 0.037436 1.315924 0.1889107
## exper 0.044856 0.013577 3.303856 0.0010346 **
## I(exper^2) -0.000922 0.000406 -2.268993 0.0237705 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## RMSE: 0.67642 Adj. R2: 0.116926
## F-test (1st stage), educ: stat = 73.9 , p < 2.2e-16 , on 1 and 424 DoF.
## Wu-Hausman: stat = 2.9683, p = 0.085642, on 1 and 423 DoF.

Here, feols(y ~ x1 + x2 | x3 ~ z, data = dataset) where x1 and x2 are the exogenous


regressors, x3 is the endogenous regressor and z is the instrument(s) The code performs
Instrumental Variable (IV) Regression to estimate the relationship between log(wage)
(dependent variable) and education (educ), while accounting for potential endogeneity in
the education variable by using an instrument (mothereduc).
F. ivreg to obtain 2SLS estimates
Let’s use instead the command ivreg to obtain 2SLS estimates:
(note x1 is the endogenous variable and z1 is the instrument, x2 and x3 are
exogenous variables)

iv <- ivreg(y ~ x1 + x2 + x3 | z1 + x2 + x3, data=dataset)

iv2 <- ivreg(log(wage) ~ educ + exper + I(exper^2) | mothereduc + exper +


I(exper^2), data=mroz2)
summary(iv2)

##
## Call:
## ivreg(formula = log(wage) ~ educ + exper + I(exper^2) | mothereduc +
## exper + I(exper^2), data = mroz2)
##
## Residuals:
## Min 1Q Median 3Q Max
## -3.10804 -0.32633 0.06024 0.36772 2.34351
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 0.1981861 0.4728772 0.419 0.67535
## educ 0.0492630 0.0374360 1.316 0.18891
## exper 0.0448558 0.0135768 3.304 0.00103 **
## I(exper^2) -0.0009221 0.0004064 -2.269 0.02377 *
##
## Diagnostic tests:
## df1 df2 statistic p-value
## Weak instruments 1 424 73.946 <2e-16 ***
## Wu-Hausman 1 423 2.968 0.0856 .
## Sargan 0 NA NA NA
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.6796 on 424 degrees of freedom
## Multiple R-Squared: 0.1231, Adjusted R-squared: 0.1169
## Wald test: 7.348 on 3 and 424 DF, p-value: 8.228e-05

(note x1 is the endogenous variable and z1 is the instrument, x2 and x3 are exogenous
variables)
iv <- ivreg(y ~ x1 + x2 + x3 | z1 + x2 + x3, data=dataset) note the weak instrument test as part
of the output
G. Compare the OLS and IV estimates
m_list <- list(OLS = ols, TWOSLS = iv2, TWOSLSmanually=step2)
msummary(m_list, stars=TRUE)

OLS TWOSLS TWOSLSmanually


(Intercept) -0.522** 0.198 0.198
(0.199) (0.473) (0.493)
educ 0.107*** 0.049
(0.014) (0.037)
exper 0.042** 0.045** 0.045**
(0.013) (0.014) (0.014)
I(exper^2) -0.001* -0.001* -0.001*
(0.000) (0.000) (0.000)
educhat 0.049
(0.039)
[Link]. 428 428 428
R2 0.157 0.123 0.046
R2 Adj. 0.151 0.117 0.039
AIC 1892.0 1908.8 1945.0
BIC 1912.3 1929.1 1965.3
[Link]. -431.599 -458.117
F 26.286 6.751
RMSE 0.66 0.68 0.71
OLS TWOSLS TWOSLSmanually
• p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

note that coefficients of 2SLS and 2SLSmanual are the same but standard errors are
different: when using ivreg command, it computes correct standard errors in second stage,
so always use this command! Compare OLS and IV estimates: the estimated returns to
education is 4.93% with 2SLS, which is lower than the OLS estimate of 10.75%; this is
consistent with a positive omitted variable bias due to positive correlation between the
variable educ and the omitted factor. Note also that the standard errors in 2SLS (the
correct ones) of the instrumeted regressor is quite high and the variable educ is now not
significant. So there is an efficiency loss with the IV regression and at this point we would
like to make sure that we really need IV estimation (i.e. educ is endogenous).
H. GIVE estimates using gmm
mroz2$exper2 <- (mroz2$exper^2)
mm <- cbind(mroz2$mothereduc, mroz2$exper, mroz2$exper2)
gmm <- gmm(log(wage) ~ educ + exper + I(exper^2),
x = mm, data = mroz2)
summary(gmm)

##
## Call:
## gmm(g = log(wage) ~ educ + exper + I(exper^2), x = mm, data = mroz2)
##
##
## Method: twoStep
##
## Kernel: Quadratic Spectral
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 0.19818608 0.56526349 0.35060831 0.72588222
## educ 0.04926295 0.04450138 1.10699836 0.26829464
## exper 0.04485585 0.01453878 3.08525506 0.00203378
## I(exper^2) -0.00092208 0.00040154 -2.29632397 0.02165736
##
## J-Test: degrees of freedom is 0
## J-test P-value
## Test E(g)=0: 1.20773227475269e-24 *******

• create the square of experience variable


• define the moment conditions you are going to use, i.e. the vector of exogenous
variable plus instruments
• estimate the model using those moment conditions gmm <- gmm(y ~ x1 + x2 + x3, x
= mm, data = mroz2)
m_list2 <- list(OLS = ols, TWOSLS = iv2, GMM = gmm)
msummary(m_list2, stars=TRUE)
OLS TWOSLS GMM
(Intercept) -0.522** 0.198 0.198
(0.199) (0.473) (0.565)
educ 0.107*** 0.049 0.049
(0.014) (0.037) (0.045)
exper 0.042** 0.045** 0.045**
(0.013) (0.014) (0.015)
I(exper^2) -0.001* -0.001* -0.001*
(0.000) (0.000) (0.000)
[Link]. 428 428 428
R2 0.157 0.123
R2 Adj. 0.151 0.117
AIC 1892.0 1908.8
BIC 1912.3 1929.1
[Link]. -431.599
F 26.286
RMSE 0.66 0.68
• p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

I. Wu-Hauman test
# First-stage regression
step1 <- lm(educ ~ mothereduc + exper + I(exper^2), data = mroz2)
#here, the endogenous variable (educ) is regressed on the instrumental
variable (mothereduc) and other exogenous variables

# Extract residuals from the first stage


v <- step1$residuals
# Augmented regression
wu_test <- lm(log(wage) ~ educ + exper + I(exper^2) + v, data = mroz2)

# Summary of the model


summary(wu_test)

##
## Call:
## lm(formula = log(wage) ~ educ + exper + I(exper^2) + v, data = mroz2)
##
## Residuals:
## Min 1Q Median 3Q Max
## -3.04565 -0.30313 0.04232 0.39785 2.34435
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 0.1981861 0.4626315 0.428 0.668586
## educ 0.0492630 0.0366249 1.345 0.179324
## exper 0.0448558 0.0132827 3.377 0.000801 ***
## I(exper^2) -0.0009221 0.0003976 -2.319 0.020857 *
## v 0.0683815 0.0396903 1.723 0.085642 .
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.6649 on 423 degrees of freedom
## Multiple R-squared: 0.1627, Adjusted R-squared: 0.1548
## F-statistic: 20.55 on 4 and 423 DF, p-value: 1.733e-15

we reject at 10% but there does not seeem to be a strong rejection of no endogeneity This
p-value is close to 0.1 but greater than typical significance levels (0.05), meaning we fail to
reject the null hypothesis of no endogeneity at the 5% level. However, at a 10%
significance level, the result could be interpreted as weak evidence of endogeneity.
J. two instruments Test for weak instruments
strong2 <- lm(educ ~ exper + exper2 + mothereduc + fathereduc, data=mroz2)
summary(strong2)

##
## Call:
## lm(formula = educ ~ exper + exper2 + mothereduc + fathereduc,
## data = mroz2)
##
## Residuals:
## Min 1Q Median 3Q Max
## -7.8057 -1.0520 -0.0371 1.0258 6.3787
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 9.102640 0.426561 21.340 < 2e-16 ***
## exper 0.045225 0.040251 1.124 0.262
## exper2 -0.001009 0.001203 -0.839 0.402
## mothereduc 0.157597 0.035894 4.391 1.43e-05 ***
## fathereduc 0.189548 0.033756 5.615 3.56e-08 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 2.039 on 423 degrees of freedom
## Multiple R-squared: 0.2115, Adjusted R-squared: 0.204
## F-statistic: 28.36 on 4 and 423 DF, p-value: < 2.2e-16

library(car) #Testing the Joint Validity of Instruments


linearHypothesis(strong2, c("mothereduc=0", "fathereduc=0"))

##
## Linear hypothesis test:
## mothereduc = 0
## fathereduc = 0
##
## Model 1: restricted model
## Model 2: educ ~ exper + exper2 + mothereduc + fathereduc
##
## [Link] RSS Df Sum of Sq F Pr(>F)
## 1 425 2219.2
## 2 423 1758.6 2 460.64 55.4 < 2.2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Both mothereduc and fathereduc are significant predictors of educ individually (based on
their p-values). The joint hypothesis test confirms that they are jointly significant as
predictors of educ. This means they are relevant instruments for use in an instrumental
variables (IV) regression. The F-statistic is 55.4 with a p-value of < 2.2e-16. This means we
reject the null hypothesis that mothereduc and fathereduc are jointly equal to zero.
K. Obtain the 2SLS estimates.
iv3 <- ivreg(log(wage) ~ educ + exper + exper2 | exper + exper2 +
mothereduc + fathereduc , data = mroz2)
summary(iv3)

##
## Call:
## ivreg(formula = log(wage) ~ educ + exper + exper2 | exper + exper2 +
## mothereduc + fathereduc, data = mroz2)
##
## Residuals:
## Min 1Q Median 3Q Max
## -3.0986 -0.3196 0.0551 0.3689 2.3493
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 0.0481003 0.4003281 0.120 0.90442
## educ 0.0613966 0.0314367 1.953 0.05147 .
## exper 0.0441704 0.0134325 3.288 0.00109 **
## exper2 -0.0008990 0.0004017 -2.238 0.02574 *
##
## Diagnostic tests:
## df1 df2 statistic p-value
## Weak instruments 2 423 55.400 <2e-16 ***
## Wu-Hausman 1 423 2.793 0.0954 .
## Sargan 1 NA 0.378 0.5386
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.6747 on 424 degrees of freedom
## Multiple R-Squared: 0.1357, Adjusted R-squared: 0.1296
## Wald test: 8.141 on 3 and 424 DF, p-value: 2.787e-05

L. Wu-Hausman test
# First-stage regression
step5 <- lm(educ ~ mothereduc + fathereduc + exper + I(exper^2), data =
mroz2)
#here, the endogenous variable (educ) is regressed on the instrumental
variable and other exogenous variables

# Extract residuals from the first stage


u <- step5$residuals
# Augmented regression
wu_test1 <- lm(log(wage) ~ educ + exper + I(exper^2) + u, data = mroz2)

# Summary of the model


summary(wu_test1)

##
## Call:
## lm(formula = log(wage) ~ educ + exper + I(exper^2) + u, data = mroz2)
##
## Residuals:
## Min 1Q Median 3Q Max
## -3.03743 -0.30775 0.04191 0.40361 2.33303
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 0.0481003 0.3945753 0.122 0.903033
## educ 0.0613966 0.0309849 1.981 0.048182 *
## exper 0.0441704 0.0132394 3.336 0.000924 ***
## I(exper^2) -0.0008990 0.0003959 -2.271 0.023672 *
## u 0.0581666 0.0348073 1.671 0.095441 .
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.665 on 423 degrees of freedom
## Multiple R-squared: 0.1624, Adjusted R-squared: 0.1544
## F-statistic: 20.5 on 4 and 423 DF, p-value: 1.888e-15

or simply look at results of ivreg again we reject at 10%


M. The Sargan test
res_iv3 <- iv3$residuals
sargan <- lm(res_iv3 ~ exper + exper2 + mothereduc + fathereduc, data=mroz2)
summary(sargan)

##
## Call:
## lm(formula = res_iv3 ~ exper + exper2 + mothereduc + fathereduc,
## data = mroz2)
##
## Residuals:
## Min 1Q Median 3Q Max
## -3.1012 -0.3124 0.0478 0.3602 2.3441
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 1.096e-02 1.413e-01 0.078 0.938
## exper -1.833e-05 1.333e-02 -0.001 0.999
## exper2 7.341e-07 3.985e-04 0.002 0.999
## mothereduc -6.607e-03 1.189e-02 -0.556 0.579
## fathereduc 5.782e-03 1.118e-02 0.517 0.605
##
## Residual standard error: 0.6752 on 423 degrees of freedom
## Multiple R-squared: 0.0008833, Adjusted R-squared: -0.008565
## F-statistic: 0.0935 on 4 and 423 DF, p-value: 0.9845

# compute the test statistic = N*R^2=428*0.0009=0.3852, you cannot reject the


null that instruments are valid.
# again you can check the output of 2SLS.

Common questions

Powered by AI

In the context of measurement for educ, the standard errors of the 2SLS estimates are larger compared to those from OLS, indicating a loss in efficiency. This occurs because 2SLS uses instruments to replace potentially endogenous explanatory variables, which reduces bias but increases variability in the estimates. This trade-off is typical in IV estimation as instruments filter out endogenous variability, often introducing more uncertainty but leading to unbiased estimators compared to biased but more precise OLS estimates .

Using multiple instruments can increase the efficiency of 2SLS estimates by improving the precision of the endogenous variable's predicted values in the first stage. However, it also introduces the complexity of ensuring all instruments meet the validity criteria, i.e., relevance and exogeneity. More instruments increase the identification strength but add the risk of overfitting and may lead to biased estimates if any instrument is weak or invalid. Validity is typically assessed using the Sargan test; if violated, it undermines the 2SLS estimation accuracy .

Endogeneity refers to the situation where an explanatory variable is correlated with the error term in a regression model, often due to omitted variable bias, measurement error, or simultaneity. This correlation biases the results of ordinary least squares (OLS) regression, leading to unreliable estimates. Detecting and addressing endogeneity is crucial because it ensures the validity of the causal interpretations drawn from empirical models. Instrumental variables (IV) approaches, such as 2SLS, are often employed to correct for endogeneity .

The Sargan test evaluates the validity of instruments in IV estimation by testing if the instruments are correlated with the error term, which would violate the IV estimation assumption of instrument exogeneity. The test involves regressing residuals from the IV estimation against the instruments. The test statistic, calculated as the number of observations times the R-squared from this auxiliary regression, follows a chi-square distribution. If the p-value is high, one cannot reject the null hypothesis of instruments being exogenous, thus validating them. A low p-value implies that the instruments are correlated with the error term, indicating invalidity .

A strong instrument in econometric analysis is one that is highly correlated with the endogenous variable but uncorrelated with the error term in the structural equation. The strength of an instrument is usually evaluated using the F-statistic from the first-stage regression, where the endogenous variable is regressed on the instrument(s) and any exogenous control variables. A significant coefficient for the instrument, indicated by a small p-value, suggests relevance. Typically, if the F-statistic is 10 or greater, the instrument is considered strong enough .

The OLS estimate of the returns to education is higher (approximately 10.75%) compared to the 2SLS estimate (approximately 4.93%). This discrepancy suggests the presence of a positive omitted variable bias in the OLS estimate, possibly due to a positive correlation between education and other unobserved factors impacting wages. Consequently, the 2SLS estimate, which accounts for endogeneity by using instruments, likely provides a more unbiased measure, though with larger standard errors, owing to reduced efficiency of the instrumentally estimated equation .

The Generalized Method of Moments (GMM) may be preferred over traditional 2SLS because it can provide more efficient estimates by using a wider set of moment conditions, allowing for heteroskedasticity and autocorrelation consistent (HAC) standard errors. GMM is flexible and can incorporate additional information that allows it to account for the heterogeneity of error structures better than 2SLS, which assumes homoskedastic errors. This adaptability often results in more efficient, albeit computationally more intense, inferences .

The 2SLS (two-stage least squares) estimator is used to manage endogeneity by first estimating the relationship of the endogenous regressor with the instrumental variables (the first stage) and then using the predicted values from this regression as regressors in the second stage where the primary equation of interest is estimated. This two-step process helps in obtaining unbiased and consistent estimates by eliminating the correlation that the endogenous regressor has with the error term, which can otherwise bias the ordinary least squares (OLS) estimates .

The Wu-Hausman test assesses whether an endogenous regressor is actually endogenous by comparing the OLS and IV estimates. In the test, residuals from the first-stage regression (in which the endogenous variable is regressed on instruments and other exogenous variables) are included in an augmented regression model. The significance of the coefficient of these residuals helps determine if there is endogeneity: if the coefficient is significant, it suggests that endogeneity is present and IV estimation is necessary. A non-significant p-value indicates insufficient evidence of endogeneity, suggesting that OLS could be appropriate .

The F-statistic in a linear hypothesis test is used to examine the joint significance of multiple coefficients within a regression model, like testing whether multiple instruments are valid in an IV regression setting. A high and significant F-statistic indicates that together the instruments have a meaningful association with the endogenous variable, confirming their relevance. For instance, the joint hypothesis test in the document indicates that mothereduc and fathereduc are jointly significant with a very high F-statistic of 55.4, suggesting these instruments are relevant for the estimation .

You might also like