Quantitative Methods III - Practice 3

Multiple Linear Regression

Prof. Lorenzo Cavallo: lorenzo.cavallo.480084@uniroma2.eu

Prof. Marianna Brunetti: marianna.brunetti@uniroma2.it

Exercise Given the table:

Regressor (1) (2) (3)

College (X1 ) 0.352 0.373 0.371
(0.021) (0.021) (0.021)
Male (X2 ) 0.458 0.457 0.451
(0.021) (0.020) (0.020)
Age (X3 ) 0.011 0.011
(0.001) (0.001)
North (X4 ) 0.175
(0.037)
South (X5 ) 0.103
(0.033)
East (X6 ) -0.102
(0.043)
Intercept 12.840 12.471 12.390
(0.018) (0.049) (0.057)
Summary and joint tests
F-statistic 21.87
SER 1.026 1.023 1.020
R̄ 2
0.0710 0.0761 0.0814
n 10973 10973 10973

1
The table reports the results on a series of regressions estimated on more than
10,000 individuals who worked full-time for the whole year in a developing country.

The analysis concerned both quantitative and dummy variables1 :

• AHE = logarithm of average hourly earnings (in 2007 units)
• College = dummy variable (1 if graduate, 0 otherwise)
• Male = dummy variable (1 if male, 0 if female)
• Age = age (in years)
• North = dummy variable (1 if Region = North, 0 otherwise)
• East = dummy variable (1 if Region = East, 0 otherwise)
• South = dummy variable (1 if Region = South, 0 otherwise)
• West = dummy variable (1 if Region = West, 0 otherwise)
1. For each of the three regressions, add ∗ (level 5%) and ∗∗
(level 1%) to the
table to indicate the significance of the coefficients.
2. Answer using the regression results reported in column (2)
(a) Is age an important determinant of earnings?
(b) Use an appropriate test statistic or confidence interval to justify your
answer.
(c) Suppose Alvo is a 30-year-old college graduate and Kal is a 40-year-old
college graduate. Construct a 95% confidence interval for the difference
between expected wages.
3. Using the values reported in the table:

(a) Construct the R2 for each of the regressions.

(b) Show how to construct the F statistic for testing the hypothesis that
β4 = β5 = β6 = 0 in the regression shown in column (3).
(c) Is the statistic significant at the 1% level?
(d) Show how to construct the F statistic to test the hypothesis that β4 =
β5 = β6 = 0 in the regression shown in column (3) using the Bonferroni
test.
(e) Construct a 99% confidence interval for β1 in the regression shown in
column (3).
1
A dummy variable is one that takes the values 0 or 1 to indicate the absence or presence of
some categorical effect that may be expected to shift the outcome.

2
Solution

1. Calculating the test statistic:

βi
ti =
S.E.(βi )

and comparing it with the critical value at 5% (t = 1.96) and 1% (t = 2.58),

the following table is obtained (in which the standard errors are indicated in
round brackets and test statistics in square brackets):

Regressor (1) (2) (3)

College (X1 ) 0.352∗∗ 0.373∗∗ 0.371∗∗
(0.021) (0.021) (0.021)
[16.762] [17.762] [17.667]
Male (X2 ) 0.458∗∗ 0.457∗∗ 0.451∗∗
(0.021) (0.020) (0.020)
[21.810] [22.850] [12.550]
Age (X3 ) 0.011∗∗ 0.011∗∗
(0.001) (0.001)
[11.000] [11.000]
North (X4 ) 0.175∗∗
(0.037)
[4.730]
South (X5 ) 0.103∗∗
(0.033)
[3.121]
East (X6 ) -0.102∗
(0.043)
[−2.372]
Intercept 12, 840∗∗ 12, 471∗∗ 12, 390∗∗
(0.018) (0.049) (0.057)
[692.833] [254.510] [217.368]

2. (a) Yes, age is an important determinant of wages.

Using the test statistic, we get:

0.011
t= = 11
0.001

3
 The test statistic is greater than 2.58 and this implies that the age coef-
ficient is statistically significant at the 1% level.

(b) The (1-α)% confidence interval of a coefficient βi is:

(1 − α)%CI(βi ) = [βi ± z1−α/2 × SE(βi )]

In this case the 95% confidence interval of β3 is:

95%CI(β3 ) = [β3 ±z0.025 ×SE(β3 )] = [0.011±1.96×0.001] = [0.009; 0.013]

(c) To construct the confidence interval we need to consider the age diffe-
rence of the two subjects (∆Age = 40 − 30 = 10 years).
The confidence interval will be:

∆Age×[β3 ±z0.025 ×SE(β3 )] = 10×[0.011±1.96×0.001] = [0.0904; 0.1296]

As a percentage it is a difference in expected wages between 9.04% and

12.96%.

3. (a) From the formula of the Adjusted-R2 (R̄2 ), we can get R2 . In particu-
lar, knowing that
n−1 n − 1 RSS
R̄2 = 1 − × (1 − R2 ) = 1 − ×
n−k n − k T SS
n−k
→ R2 = 1 − (1 − R̄2 )
n−1
.

It follows that:

• column (1) = R2 = 1 − 10973−3

10973−1
(1 − 0.0710) = 0.0712
• column (2) = R2 = 1 − 10973−4
10973−1
(1 − 0.0761) = 0.0764
• column (3) = R = 1 −
2 10973−7
10973−1
(1 − 0.0814) = 0.0819

(b) H0 : β4 = β5 = β6 = 0 H1 : βi ̸= 0, for i = 4, 5, 6.

Knowing that:
2 2
Runrestricted − Rrestricted /q
F = 2
,
(1 − Runrestricted )/(n − kunrestricted )

4
 we can test the null hypothesis starting from the R2 of the regressions.
For the non-restricted, the R2 is that of column (3), while for the
restricted we are going to consider the R2 of column (2).
Considering the number of restrictions q = 3 and n = 10973, it follows
that

0.0819 − 0.0764/3
F = = 21.898
(1 − 0.0819)/(10973 − 7)
.

(c) The critical value at 1% F3,∞ = 3.78, is less than the F statistic, so we
reject the null hypothesis at 1% significance level.

(d) The Bonferroni test allows testing joint hypotheses on q coefficients star-
ting from the t statistics relating to individual hypotheses but correcting
the critical value as follows:

c = z1− α/q
2

H0 is rejected if at least one of the individual t-ratios is, in absolute

value, greater than the critical value c, adjusted as above.

In our case, H0 is a joint hypothesis on 3 coefficients, therefore q = 3.

Setting α = 1%, we have:

c = z1− α/q = z1− 0.01/3 = z0.9983 = 2.93

2 2

The t-statistics of the single coefficients subject to the null hypothesis

are t4 = 4.730, t5 = 3.121 and t6 = −2.372. Since |t6 | < c but |t5 | > c
and |t4 | > c, the null hypothesis is rejected.

(e) [β1 ± 2.58 × SE(β1 )] = [0.371 ± 2.58 × 0.021] = [0.3168; 0.4252]