Rodrigo Pinto Economics 103

Department of Economics Introduction to Econometrics

UCLA Spring, 2022

Mid-Term Practice (First Example)

Tuesday, May 16, 2023

Instructions:

ˆ This is a 1 hour and 10 minute exam. You are allowed to use the equations sheet provided
in the site of the course and handwritten notes.

ˆ You are not allowed to use any electronic device. No connection to the internet via WiFi or
any other method is allowed. It is not permitted to use any kind of mobile phone.

ˆ When you are finished with the exam, please turn in the exam questions.

ˆ Cheating of any form will result in a score of 0 (zero) for the exam, in addition to the normal
university disciplinary action.

ˆ Please sign below that you have read, understood, and fulfilled all of the above instructions
and conditions.

Please fill in the following personal information:

First Name

Last Name

UCLA ID #

Signature

Exam Version A

Please start solving the examinations only when you are instructed to do so.
Please stop immediately when instructed to do so.

Good Luck!
 Part I (Questions based on Regression Output):

Questions 1–3 are based on the following regression output

Consider the following linear model based on the Cubic function:

wage = β1 + β2 · educ3 + e, where

1. wage means daily wage measured in dollars.

2. educ means years of education, measured in years of schooling.

3. educ3 means the cubic of education, that is (educ)3 .

The R Output for this linear regression is given below:

Untitled
reg wage educ3

Source | SS df MS Number of obs = 100

‐‐‐‐‐‐‐‐‐‐‐‐‐+‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐ F( 1, 98 ) = 100.00
Model | 10000.00 1 10000.00 Prob > F = 0.0000
Residual | 10000.00 98 100.00 R‐squared = 0.5000
‐‐‐‐‐‐‐‐‐‐‐‐‐+‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐ Adj R‐squared = 0.4950
Total | 20000.00 99 200.00 Root MSE = 10.00

‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐
wage | Coef. Std. Err. t P>|t| [95% Conf. Interval]
‐‐‐‐‐‐‐‐‐‐‐‐‐+‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐
educ3| 0.05 0.0050 10.00 0.000 0.0400 0.0600
_cons | 100.00 10.0000 10.00 0.000 80.000 120.000
‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐

Answer the following questions based on the regression output:

Question 1. What is the expected wage for a person with 10 years of education?

(a) 15

(b) 50

(c) 100

(d) 150

(e) 500

Answer: d.

E(Y |x0 = 10) = b̂1 + b̂2 · x30

= b̂1 + b̂2 · 103
= 100 + 0.05 · 1000
= 100 + 50 = 150

Page 1
 Question 2. What is the marginal effect of another year of education for a person with 10
years of education?

(a) 1.5

(b) 3

(c) 5

(d) 10

(e) 15

Answer: e.

∆E(Y )
δ= = 3 · b̂2 · x20
∆x
= 3 · 0.05 · 102
= 3 · 0.05 · 100
= 3 · 5 = 15

Question 3. What is the estimated elasticity for a person with 10 years of education at his
expected wage?

(a) 0.1

(b) 0.5

(c) 1

(d) 1.5

(e) 2

Answer:c.

∆E(Y ) x0 x0
η= · =δ·
∆x
| {z } ŷ ŷ
δ
10
= 15 ·
150
=1

3
 Questions 4–12 are based on the following regression output
Consider the quadratic model:

wage = β1 + β2 · educ + β3 · educ2 + e.

1. wage means daily wage measured in dollars.

2. educ means the education variable measured in schooling years.

3. educ2 means the squared of education variable (educ)2 .

The regression output for this linear regression is given below:

WageEducEduc2
. reg wage educ educ2

Source | SS df MS Number of obs = 100

‐‐‐‐‐‐‐‐‐‐‐‐‐+‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐ F(2, 97) = 50.00
Model | 10000.00 2 5000.00 Prob > F = 0.0000
Residual | 10000.00 97 100.00 R‐squared = 0.5000
‐‐‐‐‐‐‐‐‐‐‐‐‐+‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐ Adj R‐squared = 0.4900
Total | 20000.00 99 200.00 Root MSE = 10.000

‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐
wage | Coef. Std. Err. t P>|t| [95% Conf. Interval]
‐‐‐‐‐‐‐‐‐‐‐‐‐+‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐
educ | ‐20.000 50.0000 ‐0.40 0.686 ‐120.0000 80.00000
educ2 | 2.000 2.0000 1.00 0.448 ‐2.0000 6.00000
_cons | 200.000 200.0000 1.00 0.422 ‐200.0000 600.0000
‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐

Question 4. What is the expected wage for a person with 10 years of education?

(a) 10

(b) 20

(c) 100

(d) 200

(e) 2000

Answer: d.

E(Y |x0 = 10) = b̂1 − b̂2 · x0 + b̂3 · x20

= b̂1 − b̂2 · 10 + b̂3 · 102
= 200 − 20 · 10 + 2 · 100
= 200

Page 1
 Question 5. What is the marginal effect of another year of education for a person with 10
years of education?

(a) 200

(b) 100

(c) 2

(d) 10

(e) 20

Answer: e.

∆E(Y )
δ= = b̂2 + 2 · b̂3 · x0
∆x
= −20 + 2 · 2 · 10
= −20 + 40 = 20

Question 6. What is the estimated elasticity for a person with 10 years of education at his
expected wage?

(a) 0.1

(b) 0.5

(c) 1

(d) 1.5

(e) 2

Answer: c.

∆E(Y ) x0 x0
η= · =δ·
∆x
| {z } ŷ ŷ
δ
10
= 20 ·
200
=1

5
 Question 7. Parameter λ0 is defined as a linear combination of the model coefficients:
β1 β3
λ0 = − .
200 2
What is the value of the Best Linear Unbiased Predictor (BLUP) for λ0 ?
(a) 20
(b) 1
(c) 10
(d) 2
(e) 0
Answer: e.

b̂1 b̂3
λ̂ = −
200 2
200 2
= −
200 2
=1−1
=0

Question 8. The estimated covariance between estimators b1 , b3 is cov(b

c 1 , b3 ) = 200. What is
the estimated standard error for λ?
(a) -1
(b) 0
(c) 1
(d) 2
(e) 4
Answer: c.

 2  2    
1 1 1 1
var(λ)
ˆ = var(
ˆ b̂1 ) + ˆ b̂3 ) + 2 ·
var( · − · cov(b
ˆ 1 , b3 )
200 2 200 2
1 2
   2    
2 1 2 1 1
= · 200 + ·2 +2· · − · 200
200 2 200 2
=1+1−1
=1
p √
Thus we have that se(λ)
ˆ = var(λ)
ˆ = 1=1

Question 9. What is the test statistic for the null hypothesis H0 : λ0 = 0 against the
̸ 0.
alternative H1 : λ0 =

6
 (a) -1

(b) 0

(c) 1

(d) 2

(e) 10

Answer: b.
λ̂−0 0−0
Test statistic: t̂ = se(λ)
ˆ = 1 =0

Question 10. Using the critical value of tc = 2, the confidence interval for λ is given by:

(a) [−1, 0]

(b) [0, 1]

(c) [0, 2]

(d) [−2, 2]

(e) [−1, 1]

Answer: d.

Confidence Interval: CI = [λ̂ ± tc · se(λ)]

ˆ
= [0 ± 2 · 1]
= [−2, 2]

Question 11. A econometrician spotted a typo in the p-values of the regression output. Let
p1 = 0.422 be the p-value associated with intercept β1 .
Let p2 = 0.686 be the p-value associated with intercept β2 (for educ).
Let p3 = 0.448 be the p-value associated with intercept β3 (for educ2)
Given the estimated coefficients, and its respective standard errors, we should have that:

(a) We should have that p2 < p3

(b) We should have that p1 = p2

(c) We should have that p2 = p3

(d) We should have that p2 < p1

(e) We should have that p1 = p3

Answer: e.
Note that t̂1 = t̂3 ⇒ p̂1 = p̂3
Moreover the absolute value of the t-statistic t̂2 is smaller than the value for the t-statistic t̂2 , t̂3 ,
that is: |t̂2 | = 0.4 < 1 = |t̂1 | = |t̂3 |.
Therefore we must have that: p̂2 > p̂1 = p̂3 .

7
Part II (Multiple Choice Questions that do not use Regression Output):

Question 12. Let X1 , X2 be two random variables and take values in {0, 1} and whose joint
distribution is given by:
Joint Distribution
X1 = 0 X1 = 1 fX2 (x)
X2 = 0 0.2 0.2 0.4
X2 = 1 0.3 0.3 0.6
fX1 (x) 0.5 0.5 1

Mark the choice that is correct:

(a) X1 and X2 are statistically independent.

(b) E(X1 ) = E(X2 )

(c) P (X2 = 1|X1 = 1) = P (X2 = 0|X1 = 0)

(d) var(X1 ) = var(X2 )

(e) Cov(X1 , X2 ) ̸= 0

Answer: a.
X1 , X2 are statistically independent because P (X1 = x1 , X2 = x2 ) = P (X1 = x1 )P (X2 = x2 ) for
all x1 , x2 ∈ {0, 1}.
Independence implies that Cov(X1 , X2 ) = 0. Indeed, we have that:

E(X1 ) = P (X1 = 1) · 1 = 0.5 · 1 = 0.5

E(X2 ) = P (X2 = 1) · 1 = 0.6 · 1 = 0.6
E(X1 · X2 ) = P (X1 = 1, X2 = 1) · 1 = 0.3
Cov(X1 , X2 ) = E(X1 · X2 ) − E(X1 )E(X2 ) = 0.5 · 0.6 − 0.3 = 0

Also:
P (X2 = 1, X1 = 1) 0.3
P (X2 = 1|X1 = 1) = = = 0.6 = P (X2 = 1),
P (X1 = 1) 0.5
and
P (X2 = 0, X1 = 0) 0.2
P (X2 = 0|X1 = 0) = = = 0.4 = P (X2 = 0).
P (X1 = 0) 0.5

8
 Question 13. Let X1 ∼ N (1, 1) and X2 ∼ N (2, 4) be two normally distributed random
variables. Let the correlation between X1 and X2 be ρ. Mark the choice that is correct:

(a) V ar(c1 · X1 + c2 · X2 ) ̸= V ar(c0 + c1 · X1 + c2 · X2 ) whenever c0 ̸= 0.

(b) P ((X1 − 1) > 3) = P ((X2 − 2) > 3) regardless if X1 , X2 correlate.

X1 −1 X2 −2
, then E(Z12 + Z22 ) = 2 regardless if X1 , X2 correlate.



(c) Let Z1 = 1 and Z2 = 2

(d) If X1 and X2 correlate, then their covariance may take any value in the real line [−∞, ∞].


(e) We have that E (X1 − 1) · (X2 − 2) = 0 regardless if X1 , X2 correlate.

Answer: c.
Observe that Z1 , Z2 are standard normal random variables that might be potentially correlated.
Regardless of its correlation, we still have that E(Z12 + Z22 ) = E(Z12 ) + E(Z22 ).
Now observe that Z1 , Z2 are the standardized variables of X1 , X2 . Thus they have means zero and
variance one. Now when a variable X has mean zero, it is always the case that E(X 2 ) = V ar(X).
Therefore we have that:

E(Z12 + Z22 ) = E(Z12 ) + E(Z22 ) = V ar(Z1 ) + V ar(Z2 ) = 1 + 1 = 2.

Let’s compute the mean and variance of these variables for sake of completeness.
Variables Z1 , Z2 have mean zero because:
     
X1 − 1 E(X1 ) − 1 1−1
E(Z1 ) = E = = = 0,
1 1 1
     
X2 − 2 E(X2 ) − 2 2−2
E(Z2 ) = E = = = 0.
2 2 2

Variables Z1 , Z2 have variance one because:

     
X1 − 1 V ar(X1 ) 1
V ar(Z1 ) = V ar = 2
= = 1,
1 1 1
     
X2 − 2 V ar(X2 ) 4
V ar(Z2 ) = V ar = 2
= = 1.
2 2 4

Thus the expectation E(Z12 + Z22 ) is equal to:

E(Z12 + Z22 ) = E(Z12 ) + E(Z22 )

= E((Z1 − 0)2 ) + E((Z2 − 0)2 )
= E((Z1 − E(Z1 ))2 ) + E((Z1 − E(Z2 ))2 )
= V ar(Z1 ) + V ar(Z2 )
=1+1=2

In letter (a), the addition of a constant term does not modify the variance.

In letter (b), the probabilities are different because the random variables defined as (X1 − 1)
and (X2 − 2) have mean zero but different variances.

9
 In letter (d), consider two random variables X1 , X2 such that var(X1 ) > V ar(X2 ), that is, X1
has the bigger variance. Then we have that:

−1 ≤ Corr(X1 , X2 ) ≤ 1
Cov(X1 , X2 )
−1 ≤ p ≤1
V ar(X1 )V ar(X2 )
Cov(X1 , X2 )
−1 ≤ ≤1
se(X1 )se(X2 )
−se(X1 )se(X2 ) ≤ Cov(X1 , X2 ) ≤ se(X1 )se(X2 )

Note that it is clear that the covariance between two variables is bounded by ±se(X1 )se(X2 ),
which falsifies letter d. Now if V ar(X1 ) > V ar(X2 ) then se(X1 )se(X1 ) > se(X1 )se(X2 ), then it
is also true that:

−se(X1 )se(X1 ) ≤ Cov(X1 , X2 ) ≤ se(X1 )se(X1 )

−V ar(X1 ) ≤ Cov(X1 , X2 ) ≤ V ar(X1 )

In summary, it is always the case that the covariance of two variables is bounded by positive and
the negative value of the largest variance.

In letter e, by the definition of the covariance, we have that:

E (X1 − 1) · (X2 − 2) = E (X1 − E(X1 )) · (X2 − E(X2 )) = Cov(X1 , X2 ).

The covariance can also be written in terms of the correlation ρ :

Cov(X1 , X2 ) p
ρ= p ⇒ Cov(X1 , X2 ) = ρ · V ar(X1 )V ar(X2 ).
V ar(X1 )V ar(X2 )

Thus, if ρ ̸= 0 then Cov(X1 , X2 ) ̸= 0, which falsifies letter e.

10
 Question 14. Consider the following regression model:

Yi = β1 + β2 xi + ϵi ,

for i = 1, ..., N . Let ei ∼ 0, σi2 . That is, ei has a distribution whose mean is 0 and its variance

is σ 2 . Let
N N
1 X 1 X
s2x = (xi − x)2 , where x = xi . (1)
N N
i=1 i=1

A data analyst ran a regression of y on x and obtained the following estimates for β1 and β2 :
b̂1 = 4, b̂2 = .5. Define x∗i = 10 × xi . If one were to run a regression of y on x∗ the estimates b̂∗1
and b̂∗2 would be

(a) b̂∗1 = 4, b̂∗2 = 5

(b) b̂∗1 = 4, b̂∗2 = .5

(c) b̂∗1 = 4, b̂∗2 = .05

(d) b̂∗1 = .4, b̂∗2 = .5

(e) b̂∗1 = .4, b̂∗2 = 5

Answer: c.
b̂2
x-transformation x∗ = c · x only changes the estimate for β2 , that is, b̂∗2 = c . the values of the
t-statistics, the inference and the R2 remain the same.

Question 15. Regarding the Simple Regression Model y = β1 +β2 ·x+e, which of the following
is FALSE?

(a) ȳ − b̂1 − b̂2 x̄ = 0, where x̄, ȳ denote sample means.

(b) The LS estimates for the quadratic regression Y = β1 + β2 · x2 + e is not BLUE because the
relation between Y and x is not linear, so the linearity assumption is violated.
cov(x,y)
(c) In the Simple Regression Model, y = β1 + β2 · x + e, b̂2 = var(x) , where cov(x, y) is the
sample covariance and var(x) is the sample variance of x.

(d) Let b̂∗1 , b̂∗2 be estimates for β1 , β2 other than the least squares estimates b̂1 , b̂2 , then it must
be that:
XN N
X
(b̂∗1 + b̂∗2 xi − yi )2 ≥ (b̂1 + b̂2 xi − yi )2 .
i=1 i=1

(e) The sign of the covariance between b̂1 , b̂1 depends only on the sample mean of x.

Answer: b.

11
 Letters a, c display the standard equations that are used to estimate parameters in the simple
regression model.

Letter b is false because the equation is quadratic in x but linear in x2 .

Letter d is correct because the least squares estimators minimize the sum of the square of the
residuals.
This statement deserves some explanation. The estimates b̂1 , b̂2 are defined as the values of pa-
rameters β1 , β2 that minimize the objective equation i (yi − β1 − β2 )2 , that is:
P

X
b̂1 , b̂2 = min (yi − β1 − β2 )2
β1 ,β2
i
PN 2
Thus the sum of the squared residuals, i=1 (b̂1 + b̂2 xi − yi ) is the lowest value that can be
obtained across all possible estimates of β1 and β2 . This means that the sum square of residuals
for any other estimates for β1 and β2 , say b̂∗1 and b̂∗2 , must be bigger than the sum of the square of
residuals computed using the LS estimates b̂1 , b̂2 . In other words, we have that
N
X N
X
(b̂∗1 + b̂∗2 xi − yi )2 > (b̂1 + b̂2 xi − yi )2 ,
i=1 i=1

for any values b̂∗1 , b̂∗2 that differ from b̂1 , b̂2 .
Letter e is correct because the formula for the covariance between estimators b1 , b2 depends on
−x̄ and other terms that are always positive.

Question 16. Let the simple regression model Y = β1 + β2 · X + e. Consider the inference
that tests the null hypothesis H0 : β2 = 0 against H1 : β2 ̸= 0 at significance level α. Which of the
statements is false?

(a) If the standard error of b̂2 decreases, then it is more likely to reject H0 , (everything else
constant).

(b) The higher the absolute value of b̂2 , the more likely it is to reject H0 (everything else con-
stant).

(c) The higher the significance level α, the more likely it is to reject H0 (everything else constant).

(d) Hypothesis H0 : β2 = 0 is not rejected whenever the value 0 belongs to its confidence interval
(with confidence level of 1 − α).

(e) The larger the sample size, the more likely it is to reject H0 , (everything else constant).

(f) The higher the p-value, the more likely you are to reject H0 .

Answer f.

12
 The question requires you to interpret the equations for the estimator standard errors in the
simple regression model:
1 1
V ar(b2 ) = σ 2 · ·
N var(x)
1 var(x) + x̄2
V ar(b1 ) = σ 2 · ·
N var(x)
1 −x̄
Cov(b1 , b2 ) = σ 2 · ·
N var(x)

According to the equations above, we have that:

ˆ The larger the error variance σ 2 , less precise (greater variance) of estimators b1 , b2 .
ˆ The larger the sample size N , more precise (smaller variance) of estimators b1 , b2 .
ˆ The larger the sample variance of the explanatory variable var(x) more precise (smaller
variance) of estimators b1 , b2 .
ˆ Cov(b1 , b2 ) has always the opposite sign of x̄.
The question also requires you to understand the mechanism of single hypothesis testing.
The table below presents a summary of the components of the single hypothesis testing using
p-values:

Hypothesis Testing Using p-Values

Null Alternative p-value Rejection Which
Hypothesis Hypothesis Calculation Rule Tail?
H0 : β = c H1 : β ̸= c p = P (|t(d.f )| > |t̂|) p<α Two-tail test
H0 : β = c H1 : β > c p = P (t(d.f ) > t̂) p<α Right-tail test
H0 : β = c H1 : β < c p = P (t(d.f ) < t̂) p<α Left-tail test
Test statistic Depends on Depends on
depends on c H1 (̸=, >, <) p and α
b̂−c
t̂ = ˆ
Se(b)

For a given p-value, the larger the significance level α, the more likely you are to reject the null
hypothesis.

Question 17. Let the simple regression model Y = β1 + β2 · X + e. Consider the inference
that tests the null hypothesis H0 : βk = 0 against H1 : βk ̸= 0 at significance level α for k = 1, 2.
Which of the statements is correct?

(a) It is less likely to reject the null hypothesis H0 : β1 = 0 if we do the transformation ynew = y·c.
(b) It is less likely to reject the null hypothesis H0 : β2 = 0 if we do the transformation xnew = x·c.
(c) It is less likely to reject the null hypothesis H0 : β2 = 0 if we do the transformation ynew =
y + c.

13
 (d) It is less likely to reject the null hypothesis H0 : β1 = 0 if we standardize both y and x.

(e) It is less likely to reject the null hypothesis H0 : β2 = 0 if we standardize both y and x.

Answer d.
The question explores the properties of linear transformations of the dependent variable y and
the explanatory variable x. First, it is useful to revise what standardize means. The standardized
version x̃ of a variable x is obtained by subtracting x by its sample mean and dividing it by its
standard deviations, namely:
x̃ = − x̄sd(x).
x
As a consequence, x̃ has mean zero, variance one, and standard deviation one. If both x and y are
standardized, then the estimated slope b̂2 of the simple regression model is equal to the correlation
between x and y. The estimate of the intercept b̂1 is zero.
A summary of the properties of the LS estimates under linear transformations of x and y is
listed below:

ˆ The transformation xnew = x·c modifies the estimates b̂2 and se(bˆ 2 ). The estimates b̂1 , se(b)
ˆ 1,
the t-statistics for testing coefficients β1 , β2 and R2 remain the same.

ˆ The transformation ynew = y · c modifies the estimates b̂1 , se(b ˆ 2 ), σ̂ 2 . The inference,
ˆ 1 ), b̂2 , se(b
2
t-statistics for testing coefficients β1 , β2 and R remain the same.

ˆ The transformation xnew = x + c modifies the estimates b̂1 , se(b

ˆ 1 ) and its t-statistic t̂1 . The
ˆ 2 , t̂2 and R2 remain the same.
estimates b̂2 , se(b)

ˆ The transformation ynew = y + c modifies the estimates b̂1 but does not modifies se(b ˆ 1 ), so
it affects its t-statistic. The transformation does not change the estimates b̂2 , se(b)
ˆ 2 , t̂2 and
R2 .

ˆ If x̄ = ȳ = 0 then the estimate b̂1 = 0, t̂1 = 0 and we never reject the null hypothesis
H0 : β1 = 0.

ˆ If we standardize both y and x, then the estimate b̂1 is zero and the estimate b̂2 is equal to
the sample correlation of x, y.

In more general terms, you must be able to read the following table:

14
(a) Note that the sample means of ỹ and x̃ are zero, that is ỹ = 0 and x̃ = 0. Thus the estimated
ˆ
value of b̃1 is given by:
ˆ ˆ
b̃1 = ỹ − b̃2 x̃ = 0.
ˆ ˆ
(b) Moreover the estimated t-statistic associated with b̃1 must be zero as t̃ˆ1 = b̃1 /se(
b b̃) = 0.
(c) Note also that a linear transformation of two random variables X, Y does not change its sample
correlation. But the goodness of fit is simply the square of the correlation between variables.
Thereby the goodness of fit of the original regression and the one that uses the transformed
variables must be the same.
(d) Lastly, note that the sample variance of each of the transformed variables is one. Thus the
covariance of the transformed variables is equal to its correlation.

Question 18. Consider two regressions: Y on X and Y on X. Notationally, let Y = β1y +

β2y X + ϵy , where b̂y1 , b̂y2 denotes the estimates for β1y , β2y , t̂y1 , t̂y2 are its t-statistics and Ry2 denotes
its goodness of fit. Similarly, let X = β1x + β2x Y + ϵx , where b̂x1 , b̂x2 , t̂x1 , t̂x2 , Rx2 denote its respective
estimates. Mark the correct statement regarding these two regressions:

(a) It is always the case that b̂y2 = b̂x2 .

(b) It is always the case that b̂y2 = 1/b̂x2 .

(c) It is always the case that b̂y1 = b̂x1 .

(d) It is always the case that t̂y1 = t̂x1 .

(e) It is always the case that Ry2 = Rx2 .

Answer e.

Letter e is true because the goodness of fit, R2 can be expressed as R2 = (corr(x, y))2 . The
correlation between x and y are the same if we regress Y on X or vice-versa.
cov(x,y) cov(x,y)
Letter a is false because b̂y2 = var(x) while b̂x2 = var(y) . For b̂x2 = b̂y2 to be true, we would need
that var(x) = var(y).
Letter b is false because b̂y2 = cov(x,y) x var(y)
var(x) while 1/b̂2 = cov(x,y) . Note that one could mistakenly
think that the letter is correct by some erroneous rationale. For instance, one could think on
isolating X in the equation: Y = β1 + β2 X + ϵ. This would generated the following equation
X = ββ21 + β12 X − β12 ϵ, which suggests the wrong statement that b̂y2 = 1/b̂x2 .
A simple way to see that letter c is wrong think of a variable x such that x̄ = 0. In this case,
b̂y1 = ȳ − b̂y2 x̄ = ȳ.
On the other hand, b̂x1 = x̄ − b̂x2 ȳ = −ȳ · cov(x,y)
var(y) , which falsifies the letter. In more general terms,

15
for b̂y1 = b̂x1 we need that:

ȳ − b̂y2 x̄ = x̄ − b̂x2 ȳ
⇒ȳ + b̂x2 ȳ = x̄ + b̂y2 x̄
⇒ȳ(1 + b̂x2 ) = x̄(1 + b̂y2 )
ȳ 1 + b̂x2
⇒ =
x̄ 1 + b̂y
|{z}
Depends on Sample Means
| {z 2}
Depends on Sample Covariance

The statement is wrong because the left-hand side depends only on sample means while the
right-hand side of the equation depends on the sample covariances and we can change sample
means without changing the sample covariances by simple adding constants to x or y.
A simple way to check that Letter d is wrong consists in investigating the case where x̄ = 0,
ȳ > 0 and cov(x, y) > 0. Recall that, if x̄ = 0, then b̂y1 = ȳ and b̂x1 = −ȳ · cov(x,y)
var(y) . In this case,
b̂y1 > 0 while b̂x1 < 0 and thereby it cannot be the case that t̂y1 = t̂x1 .

Question 19. Consider the model Y = β1 + β2 X2 + β3 X3 + ϵ. Mark the correct statement

regarding the sample correlation between x2 and x3 :

(a) The lower the absolute value of the sample correlation, the larger the standard errors for
the estimators b2 and b3 .

(b) Any linear transformation of x2 changes the correlation between x2 and x3 and thereby
affects the standard error of estimator b3 .

(c) Re-scaling x2 say x∗2 = 5 · x2 changes the standard error of estimator b2 but does not change
b 2 ) = b̂∗2 /se(b
its t-statistic t̂2 = b̂2 /se(b b ∗2 ) = t̂∗2 .

(d) The estimators b2 , b3 would be most precise if the explanatory variables were equal, that is,
x2 = x3 .

(e) If the sample means x̄2 and x̄1 were zero, that is, x̄2 = x̄1 = 0, then the estimate of the
intercept is also zero, b̂1 = 0.

Answer c.

Letter c states a correct consequence of the linear transformation (see question 17)
Letter e is false because if x̄2 = x̄1 = 0, then b̂1 = ȳ.
The remaining of the items of the question explore the fact that, in multiple regression models,
the higher the correlation between explanatory variables, the higher the estimated standard error
of the leas squares estimates for β2 , β3 .
In is useful to investigate the formula for the variance of estimators b2 and b3 . The formula for
variance of estimators b2 is given by:
1 1 1
V ar(b2 ) = σ 2 · · · 2 )
N var(x2 ) (1 − r2,3

16
The interpretation of this formula is the following :

ˆ The first term (σ 2 ) means that the larger the variance of the error term, the larger the
variance of error term V ar(ei ) = E(e2i ) = σ 2 , the larger the variance of V ar(b2 ).

ˆ The second term (1/N ) means that the larger the sample size N, the smaller the variance
of V ar(b2 ).

ˆ The Third term (1/var(x2 )) means that the larger the sample variance of exploratory vari-
able x2 , the smaller the variance of V ar(b2 ).

ˆ The fourth term (1/(1 − r2,3 2 )) means that the larger the sample correlation of exploratory

variables x2 , x3 , the larger the variance of V ar(b2 ).

We can examine the analogous formula for b3 :

1 1 1
V ar(b3 ) = σ 2 · · · 2 )
N var(x3 ) (1 − r2,3

ˆ The first term (σ 2 ) means that the larger the variance of the error term, the larger the
variance of error term V ar(ei ) = E(e2i ) = σ 2 , the larger the variance of V ar(b2 ).

ˆ The second term (1/N ) means that the larger the sample size N, the smaller the variance
of V ar(b2 ).

ˆ The Third term (1/var(x3 )) means that the larger the sample variance of exploratory vari-
able x3 , the smaller the variance of V ar(b2 ).

ˆ The fourth term (1/(1 − r2,3 2 )) means that the larger the sample correlation of exploratory

variables x2 , x3 , the larger the variance of V ar(b2 ).

Note that the last term is identical for both V ar(b2 ) and V ar(b3 ). The correct answer is (a). The
most important cases when the correlation is mention in mentioned in the course are:

ˆ The increase in the sample correlation among explanatory variables (x2 , x3 ) in the mul-
tiple regression increases the variance of the estimators b2 , b3 .

ˆ The square of the sample correlation between the dependent variable y and exploratory
variable x in the simple regression is the goodness of fit R2 . Thus an increase correlation
between the dependent variable y and exploratory variable x in the simple regression
increases the R2 .

Question 20. Consider the linear Regression model

ln(Yi ) = β1 + β2 xi + ei

where x denotes annual household income (in thousands) and y denotes annual expenses on con-
sumption goods. Let ŷ be the estimated value of Y given the value of the explanatory variable x0 .

17
Consider the following estimates:

γ̂ x = b̂2 x0
γ̂ y = b̂2 ŷ
x0
γ̂ x,y = b̂2
ŷ

Which of the following statements is TRUE?

(a) γ̂ y estimates the percentage increase in consumption expenses associated with an additional
$1,000 in income.
(b) γ̂ x estimates the income elasticity of consumption for the average family.
(c) γ̂ y estimates the increase in consumption expenses for a 1% increase in income.
(d) γ̂ x estimates the average increase in consumption for an addition thousand dollars in income
for the average household.
(e) γ̂ x,y estimates the average % change in food consumption for a 1% increase in income.
Answer b.
The question explores the properties of the Log-Linear Model.
Slope is given by γ̂ y = b̂2 ȳ
Thus a thousand dollars increase leads to γ̂ y increase in consumption expenditure.
Elasticity is given by γ̂ x = b̂2 x̄.
Thus, 1% increase in income leads to γ̂ x percentage increase in consumption expenditure.

A general question on this topic requires you to use the information of the following non-linear
4 . 3 M O D Evariable
transformations of both the explanatory variable x and the dependent L I N G IyS :S U E S 143

Ta b l e 4 . 1 Some Useful Functions, their Derivatives, Elasticities and Other

Interpretation
Name Function Slope ¼ dy/dx Elasticity
x
Linear y ¼ b1 þ b2x b2 b2
y
x
Quadratic y ¼ b1 þ b2x2 2b2x ð2b2 xÞ
y
" #
2 x
Cubic y ¼ b1 þ b2x3 3b2x2 3b2 x
y
y
Log-Log ln(y) ¼ b1 þ b2ln(x) b2 b2
x
Log-Linear ln(y) ¼ b1 þ b2x b2y b2x
or, a 1 unit change in x leads to (approximately) a 100 b2% change in y
1 1
Linear-Log y ¼ b1 þ b2ln(x) b2 b2
x y
or, a 1% change in x leads to (approximately) a b2/100 unit change in y

2.8.3–2.8.4 and will be further discussed in Section 4.5. Note its possible shapes in
Figure 4.5(e). If b2 > 0 the function increases at an increasing rate; its slope is larger
18
for larger values of y. If b2 < 0, the function decreases, but at a decreasing rate.
3. In the linear-log model y ¼ b þ b lnðxÞ the variable x is transformed by the
The first column present a range of models.
The second column presents the function form of the model.
The third column presents the the equation for the slope.
The fourth column presents column presents the the equation for the elasticity.
The interpretation of the slope and elasticity are: The two most important estimates are the slope
and the elasticity.

ˆ Slope = ∆Y /∆X If the estimate value of the slope is δ̂ then:

– One unity change in x leads to (approximately) δ̂ units change in y

– b̂2 in the linear regression Y = β1 + β2 · x + e, estimates the slope of x on y.

ˆ Elasticity = ∆Y
∆X · X
Y : If the estimate value of the slope is η̂ then:

– One percent change in x leads to (approximately) η̂ percent change in y

– b̂2 in the Log-log regression ln(Y ) = β1 + β2 · ln(x) + e, estimates the elasticity of x on
y.

19