Problem Set 5 – ECON 425

Chapter 8 Problems 1 and 3

ANSWERS
i) The estimators are unbiased and consistent when heteroscedasticity is present.
ii) When heteroscedasticity is present, F test is not valid
iii) The OLS estimator are no longer best linear unbiased estimators.
Hence, ii) and iii) are the two consequences of heteroscedasticity.

- It is False because WLS is preferred over OLS when heteroscedasticity is present in the model.

Chapter 8 Computer Exercise 1 and 13 Part (i)

Answers:
i) By the condition in the model
var(u|totwork,educ,age,age2,yngkid)=var(u|male)
lets Assume var(u|male)= A0+ A1 male
This implies that the varience of error term u, in case of female will be A0 and in case
of male it will be A1
ii) U2=189359-28850 male . error term u is higher for women.
iii) Since the reported p value is 0.29 which means only 29% confidence so it is non
significant at 95% confidence or at 5% significane.

1
 CODES
setwd("C:/users/sushi/onedrive/desktop/study materials/Rstudy")
rm(list=ls())
load('sleep75.rdata')
sqage=data$age^2
one.lm=lm("sleep~totwrk+educ+age+sqage+yngkid+male",data=data)
summary(one.lm)
sqreg=resid(one.lm)^2
two.lm=lm(sqreg~male, data)
summary(two.lm)

Call:
lm(formula = " children~ age+ agesq + educ +electric+ urban", data = data)

Residuals:
Min 1Q Median 3Q Max
-5.9012 -0.7136 -0.0039 0.7119 7.4318

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -4.2225162 0.2401888 -17.580 < 2e-16 ***
age 0.3409255 0.0165082 20.652 < 2e-16 ***
agesq -0.0027412 0.0002718 -10.086 < 2e-16 ***
educ -0.0752323 0.0062966 -11.948 < 2e-16 ***
electric -0.3100404 0.0690045 -4.493 7.20e-06 ***
urban -0.2000339 0.0465062 -4.301 1.74e-05 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

coeftest(one.lm, vcov = vcovHC(one.lm, type="HC1"))

t test of coefficients:

2
 Estimate Std. Error t value Pr(>|t|)
(Intercept) -4.22251623 0.24385099 -17.3160 < 2.2e-16 ***
age 0.34092552 0.01917466 17.7800 < 2.2e-16 ***
agesq -0.00274121 0.00035051 -7.8206 6.549e-15 ***
educ -0.07523232 0.00630771 -11.9270 < 2.2e-16 ***
electric -0.31004041 0.06394815 -4.8483 1.289e-06 ***
urban -0.20003386 0.04547093 -4.3992 1.113e-05 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

No, robust standard errors are not always bigger than non-robust errors. It can
even be seen in this above example. Some are less and some are greater.

CODES
setwd("C:/users/sushi/onedrive/desktop/study materials/Rstudy")
rm(list=ls())
load('fertil2.rdata')
library(sandwich)
library (lmtest)
one.lm=lm(" children~age+agesq+educ+electric+urban",data)
summary(one.lm)
coeftest(one.lm, vcov = vcovHC(one.lm, type="HC1"))

Chapter 15 Computer Exercise 1

3
Answers:
i)

one.lm = lm(lwage ~ sibs, data)

summary(one.lm)
##
## Call:
## lm(formula = lwage ~ sibs, data = data)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.97662 -0.25857 0.02503 0.28572 1.22677
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 6.861076 0.022078 310.771 < 2e-16 ***
## sibs -0.027904 0.005908 -4.723 2.68e-06 ***
## ---

4
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.4164 on 933 degrees of freedom
## Multiple R-squared: 0.02335, Adjusted R-squared: 0.0223
## F-statistic: 22.31 on 1 and 933 DF, p-value: 2.68e-06

We get a negative coefficient. Having more siblings is correlated with fewer years of education.
ii)

two.lm = lm(educ ~ brthord, data, na.action=na.exclude)

summary(two.lm)
##
## Call:
## lm(formula = educ ~ brthord, data = data, na.action = na.exclude)
##
## Residuals:
## Min 1Q Median 3Q Max
## -4.8668 -1.5842 -0.7362 2.1332 6.1117
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 14.14945 0.12868 109.962 < 2e-16 ***
## brthord -0.28264 0.04629 -6.106 1.55e-09 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 2.155 on 850 degrees of freedom
## (83 observations deleted due to missingness)
## Multiple R-squared: 0.04202, Adjusted R-squared: 0.04089
## F-statistic: 37.29 on 1 and 850 DF, p-value: 1.551e-09

There is a strong negative correlation.

iii)

three.lm = ivreg(lwage ~ educ | brthord, data = data, na.action=na.exclude)

summary(three.lm)

5
##
## Call:
## ivreg(formula = lwage ~ educ | brthord, data = data, na.action = na.exclud
e)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.8532 -0.2557 0.0435 0.2970 1.3033
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 5.03040 0.43295 11.619 < 2e-16 ***
## educ 0.13064 0.03204 4.078 4.97e-05 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.4215 on 850 degrees of freedom
## Multiple R-Squared: -0.02862, Adjusted R-squared: -0.02983
## Wald test: 16.63 on 1 and 850 DF, p-value: 4.975e-05

We now estimate the return to a year of education as a 13% increase in wages. IF we believe in our
IV, this is a unbiased estimate. Because of missing data on brthord, we are using fewer observations
than in the previous analyses.
iv)
That magical word “identification”!
In the reduced form equation

educ=π0+π1sibs+π2brthord+v,educ=π0+π1sibs+π2brthord+v,
we need π2≠0π2≠0 in order for the β1β1 to be identified. We take the null to
be H0:π2=0H0:π2=0 and look to reject H0H0 at a small significance level.
four.lm = lm(educ ~ sibs + brthord, data, na.action=na.exclude)
summary(four.lm)
##
## Call:
## lm(formula = educ ~ sibs + brthord, data = data, na.action = na.exclude)
##
## Residuals:

6
## Min 1Q Median 3Q Max
## -5.1438 -1.6854 -0.6852 2.0090 5.9950
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 14.29650 0.13329 107.260 < 2e-16 ***
## sibs -0.15287 0.03987 -3.834 0.000135 ***
## brthord -0.15267 0.05708 -2.675 0.007619 **
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 2.137 on 849 degrees of freedom
## (83 observations deleted due to missingness)
## Multiple R-squared: 0.05833, Adjusted R-squared: 0.05611
## F-statistic: 26.29 on 2 and 849 DF, p-value: 8.33e-12

The regression of educ on sibs and brthord (using 852 observations)

yields π^2=0.153π^2=0.153 and se(π^2)=.057se(π^2)=.057. The t statistic is about –2.68, which
rejects H0H0 fairly strongly. Therefore, the identification assumptions appear to hold.
v)

five.lm = ivreg(lwage ~ educ + sibs | sibs + brthord, data = data, na.action=

na.exclude)
summary(five.lm)
##
## Call:
## ivreg(formula = lwage ~ educ + sibs | sibs + brthord, data = data,
## na.action = na.exclude)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.84808 -0.26227 0.03841 0.29901 1.30836
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 4.938527 1.055690 4.678 3.37e-06 ***

7
## educ 0.136994 0.074681 1.834 0.0669 .
## sibs 0.002111 0.017372 0.122 0.9033
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.427 on 849 degrees of freedom
## Multiple R-Squared: -0.05428, Adjusted R-squared: -0.05676
## Wald test: 10.9 on 2 and 849 DF, p-value: 2.124e-05

The standard error on β^educβ^educ is much larger than we obtained in part (iii). The 95% CI
for βeducβeduc is roughly -0.010 to .284, which is very wide and includes the value zero. The
standard error of β^sibsβ^sibs is very large relative to the coefficient estimate,
rendering sibssibs very insignificant.
vi)

data$pred.four <- predict(four.lm)

cor(data$pred.four,data$sib,use="complete.obs")
## [1] -0.9294818

Letting educieduci be the first-stage fitted values, the correlation

between educieduci and sibsisibsi is about -0.930, which is a very strong negative correlation. This
means that, for the purposes of using IV, multicollinearity is a serious problem here and is not
allowing us to estimate βeducβeduc with much precision.