Problem Sets 2 EM2_Solutions

total 40 points

1. (5 points) Suppose that the true model is

Y = 1 + 1:5X1 + U ,

E(X1 ) = 0, E(U jX1 ) = 2X1 ;

V ar(X1 ) = 15:

Obtain E( 2) in estimating Yi = 1 + 2 X1i + Vi , using the fact that E( 2) =

Cov(Y;X1 )
V ar(X1 ) and Cov(Y; X1 ) = E(Y X1 ) E(Y )E(X1 ). (The answers should be numbers,
not symbols).

Cov(X1 ; Y ) E(X1 Y ) E(X1 )E(Y )

E( 2) = =
V ar(X1 ) V ar(X1 )
E(X1 (1 + 1:5X1 + U ))
=
V ar(X1 )
E(X1 ) + 1:5E(X12 ) + E(X1 U )
=
V ar(X1 )
E(X1 U ) EfE(X1 U jX1 )g
= 1:5 + = 1:5 +
V ar(X1 ) V ar(X1 )
EfX1 E(U jX1 )g 2EfX12 g
= 1:5 + = 1:5 +
V ar(X1 ) V ar(X1 )
= 3:5

2. (5 points) Suppose that the true model is

Y = 0:25 + 2X1 + 0:5X12 + U ,

E(U jX1 ) = 0, E(X1 ) = 1; E(X12 ) = 2:5

E(X13 ) = 5:5

1
 Obtain E( 2) in estimating Yi = 1 + 2 X1i + Vi , using the fact that E( 2) =
Cov(Y;X1 )
V ar(X1 ) and Cov(Y; X1 ) = E(Y X1 ) E(Y )E(X1 ). (The answers should be numbers,
not symbols.)

Cov(Y; X1 ) E(Y X1 ) E(Y )E(X1 )

E( 2) = =
V ar(X1 ) V ar(X1 )
Ef( 0:25 + 2X1 + 0:5X12 + U )X1 g 3
=
V ar(X1 )
where E(Y ) = E( 0:25 + 2X1 + 0:5X12 + U ) = 3
0:25E(X1 ) + 2E(X12 ) + 0:5E(X13 ) + E(U X1 ) 3
=
V ar(X1 )
0:25 + 5 + 2:75 3
= = 3.
1:5

3. (8 points) Suppose that a researcher, using data on class size (CS) and average test
scores (T score) from 320 third-grade classes, estimates the OLS regression

T\
score = 475:2 6:13 CS
(21:8) (2:97)

where the standard errors of the OLS estimates are given in parentheses below the OLS
estimates.

(a) (2 points) Construct 95% con…dence interval for the slope coe¢ cient 1.

The 95% con…dence interval for 1 is

CI : [ 6:13 1:96 2:97; 6:13 + 1:96 2:97] [ 11:95; 0:309]

that is the true 1 is between 11:95 and 0:309 by 95% probability.

2
(b) (2 points) Calculate the p-value for the two-side test of the null hypothesis H0 :

1 = 0. Do you reject the null hypothesis at the 5% signi…cance level? At 1%

level?

Calculate the t-statistic

6:13 0
t statistic = 2:064:
2:97
The p-value for the test for H0 : 1 = 0
" #
b 0 6:13 0
Pr j 1 j>j j = 2 ( j 2:064j) 2 ( 2:06)
H0 SE( b 1 ) 2:97
2 0:0197 = 0:0394:

The p-value is less than 0.05, so we can reject the null hypothesis at the 5%
signi…cance level, but not at the 1% signi…cance level because it is larger than
0.01.

(c) (2 points) Calculate the p-value for the two-side test of the null hypothesis H0 :

1 = 5:9. Without doing any additional calculations, determine whether -5.9 is

contained in the 95% con…dence interval for 1.

Calculate the t-statistic

6:13 + 5:9
t statistic = 0:077:
2:97
The p-value for the test for H0 : 1 = 5:9
" #
b 0 6:13 + 5:9
Pr j 1 j>j j = 2 ( j 0:0774j) 2 ( 0:08)
H0 SE( b 1 ) 2:97
2 0:4681 = 0:936:

The p-value is larger than 0.10, so we cannot reject the null hypothesis at the
10%, 5% or 1% signi…cance level. Since CI is based on the same sampling
distribution, it gives the same results as those of the p-value. Because H0 :

1 = 5:9 is not rejected at the 5% level, this value is contained in the 95%
con…dence interval.

3
 (d) (2 points) Construct a 99% con…dence interval for 0.

The 99% con…dence interval for 0 is

CI : [475:2 2:58 21:8; 475:2 + 2:58 21:8] [418:96; 531:44]

that is the true 0 is between 418:96 and 531:44 by 99% probability.

4. (12 points) Suppose we have data on college GPA was used to estimate the following
equation

\
cGP A = 4:60 + 0:0207hsperc 0:00129sat
(2:28) (0:0098) (0:00073)
2
N = 587; R = 0:219

where cGP A is measured on a …ve-point scale (1.0-the highest, 5.0-the lowest), hsperc
is percentile in the high school graduating class (de…ned, for example, hsperc = 5 means
the top 5% of the class), sat is the combined math and reading test scores on the college
entrance exam.

(a) (2 points) Why does it make sense for the coe¢ cient on hsperc to be positive?

hsperc is de…ned so that the smaller it is, the lower the student’s standing in
high school. Everything else equal, the worse the student’s standing in high
school, the lower is his/her expected college evaluation (e.g. low test score),
thus the higher GPA.

(b) (2 points) What is the predicted college GPA when hsperc = 20 and sat = 1050?

4
 Just plug these values into the equation:

\
cGP A = 4:60 + 0:0207 20 0:00129 1050

3:6595

(c) (4 points) Suppose that two high school graduates, A and B, graduate in the same
percentile from high school, but Student A’s SAT score was 140 points higher.
What is the predicted di¤erence in college GPA for two students? Is the di¤erence
statistically signi…cant? Test the following null hypothesis

H0 : E(cGP Ajhsperc = x; sat = z + 140) E(cGP Ajhsperc = x; sat = z) = 0:

The di¤ erence between A and B is simply 140 times the coe¢ cient on sat,
because hsperc is the same for both students. So A is predicted to have a
score -0.00129 140 0.18 lower. And the null hypothesis implies

E(cGP Ajhsperc = x; sat = z + 140) E(cGP Ajhsperc = x; sat = z)

= 0 + 1 x+ 2 (z + 140) ( 0 + 1 x+ 2 z)

= 2 140.

Thus, if 2 is statistically signi…cant, that means 2 is signi…cantly di¤ erent

from zero, then we reject this null hypothesis. The t-statistic is ( 0:00129
0)/ 0:00073 -1.767. Thus, the null hypothesis is rejected at 10 % signi…-
cance level, but not at 5% signi…cant level. It implies that the di¤ erence is
statistically signi…cant at 10% signi…cance level, but not at 5% signi…cance
level.

5
 (d) (4 points) Suppose we assume a di¤erent regression model and obtain

\
cGP A = 4:60 + 0:0206hsperc 0:0085 log(sat)
(2:28) (0:0099) (0:0043)
2
N = 587; R = 0:221

And suppose that two high school graduates, A and B, graduate in the same
percentile from high school, but Student A’s SAT score was 20% higher. What is
the predicted di¤erence in college GPA for there two students? Is the di¤erence
statistically signi…cant?

The di¤ erence between A and B is simply 20 times the coe¢ cient on log(sat),
because hsperc is the same for both students. So A is predicted to have a
1
score 0:0085 100 20 0:0017 lower. And the null hypothesis implies
1
H0 : 2 100 20 = 0. Thus, if 2 is statistically signi…cant, that means

2 is signi…cantly di¤ erent from zero, then we reject this null hypothesis. The
t-statistic is ( 0:0085 0)/ 0:0043 1:977. Thus, the null hypothesis is
rejected at 5 % signi…cance level, but not at 1% signi…cant level. It implies
that the di¤ erence is statistically signi…cant at 5% signi…cance level, but not
at 1% signi…cance level.

5. (10 points) The average starting salary measured in thousand euro for new law school
graduates is determined by

2
salary = 0 + 1 LSAT + 2 LSAT + 3 log(libvol) + 4 rank +u

where LSAT is the average standardized test score for becoming a fully practicing
lawyer, libvol is the number of volumes in the law school library, and rank is a law
school ranking (with rank=1 being the best).

(a) (2 points) Explain why we expect 4 0.

6
 A larger rank for a law school means that the school has less prestige; this
lowers starting salaries. For example, a rank of 100 means there are 99 schools
thought to be better.

(b) (2 points) What signs do you expect for 1 and 2?

The LSAT is measures of the quality of the his/her class. No matter where
better students attend law school, we expect them to earn more, on average.
However, the size of return to the quality of the class would decrease as the
quality increases (diminishing marginal returns). Thus the sign of 2 would
be negative.

(c) (3 points) Using the data, the estimated equation is

\ = 1:63 + 0:048LSAT
salary 0:0002LSAT 2 + 0:042 log(libvol) 0:03rank

Interpret the coe¢ cient on the variable libvol.

This is amount of change in 1000euro salary: a one percent increase in library

volumes implies 0.42 euro increase in predicted average starting salary, other
things equal.

(d) (3 points) What is the predicted ceteris paribus di¤erence in salary for schools
with increase in the average LSAT from 40 to 45? And what is the predicted
ceteris paribus di¤erence in salary for schools with increase in the average LSAT
from 70 to 75? Explain why these two changes are di¤erent.

Holding other factor constant, additional 5 points in the average LSAT from
40 increase the predicted average salary by

E(salaryjLSAT = 45; libvol = x; rank = z) E(salaryjLSAT = 40; libvol = x; rank = z)

= 0:048 45 0:0002 (452 ) 0:048 40 + 0:0002 (402 )

= 0:048 5 0:0002 (452 402 ) 0:155

7
155euro on average. However, additional 5 points in the average LSAT from
70 increase the predicted average salary by

E(salaryjLSAT = 75; libvol = x; rank = z) E(salaryjLSAT = 70; libvol = x; rank = z)

= 0:048 75 0:0002 (752 ) 0:048 70 + 0:0002 (702 )

= 0:048 5 0:0002 (752 702 ) 0:095

95 euro on average. We assume non-linear return to LSAT that allows the

e¤ ect of LSAT on the salary to change depending on the initial level of LSAT,
and it turns out to have diminishing marginal return ( 2 < 0). Thus the
return to LSAT decreases as the level of LSAT increases.