ECO311 Lecture ‘speech’ notes – Chapter 8

University of the Western Cape

Department of Economics
ECO311 Chapter 8
Lecture ‘Speech Notes’

This chapter first conducts the same types of hypothesis testing and confidence interval questions
(as you learned in ECO242) on multivariate regressions, before we conduct t-test on equality of two
parameters, followed by four different types of F-test.

Slide #3
 It is still the same old approach to derive the confidence intervals of variance of error terms
for multivariate regressions, degrees of freedom = n – k, but k is no longer equal to 2 for
multivariate regressions.
 This is a nice example on the confidence intervals of variance of error terms.
 We will proceed to Exercise 2, Question 1.4.

Question 1
(2.6201)2 (2.6201)2
1.4 Pr [(20 − 3) ∙ (30.191)
≤ 𝜎 2 ≤ (20 − 3) ∙ (7.564)
] = 95%
2
Pr[3.8655 ≤ 𝜎 ≤ 15.4288] = 95%
Interpretation: If 100 samples of size 20 are selected and 100 confidence intervals
are constructed, we expect 95 of them to contain σ2. For the above confidence
interval, the variance of error terms is somewhere between 3.8655 and 15.4288 in
the PRF.

Slide #4
 Again, it is the same old approach to derive the confidence intervals of PRF parameters,
even for multivariate regressions. However, one exception is that you need to state the
ceteris paribus condition clearly in the interpretation part.
 Let’s do Exercise 2, Questions 1.1-1.3.

Question 1
1.1 Pr[(55.6148) − (8.6671)(2.898) ≤ 𝛽1 ≤ (55.6148) + (8.6671)(2.898)] = 99%
Pr[30.4975 ≤ 𝛽1 ≤ 80.7321] = 99%

1
 ECO311 Lecture ‘speech’ notes – Chapter 8

Interpretation: If 100 samples of size 20 are selected and 100 confidence intervals
are constructed, we expect 99 of them to contain β1. For the above confidence
interval, if study hours = 0 unit and party hours = 0 unit, on average, final mark is
somewhere between 30.4975 and 80.7321 marks in PRF.
1.2 Pr[(6.1271) − (1.5779)(2.110) ≤ 𝛽2 ≤ (6.1271) + (1.5779)(2.110)] = 95%
Pr[2.7977 ≤ 𝛽2 ≤ 9.4565] = 95%
If 100 samples of size 20 are selected and 100 confidence intervals are constructed,
we expect 95 of them to contain β2. For the above confidence interval, if study hours
increase by 1 hour, providing there is no change in party hours, on average, final
mark will increase by somewhere between 2.7977 and 9.4565 marks in the PRF.
1.3 Pr[(−3.0041) − (1.5214)(1.740) ≤ 𝛽3 ≤ (−3.0041) + (1.5214)(1.740)] = 90%
Pr[−5.6513 ≤ 𝛽3 ≤ −0.3569] = 90%
If 100 samples of size 20 are selected and 100 confidence intervals are constructed,
we expect 90 of them to contain β3. For the above confidence interval, if party
hours increase by 1 hour, providing there is no change in study hours, on average,
expected final mark will decrease by somewhere between 0.3569 and 5.6513 marks
in the PRF.

Slide #5
 Remember there are three types of hypothesis testing: one-sided left-tailed, one-sided right-
tailed, two-sided.
 We still adopt the same old approach to do these hypothesis tests (even if the hypothesized
value is not zero in step 1), except that we once again must refer to the ceteris paribus
condition in the interpretation part.
 Let’s do Exercise 2, Questions 2-3.

Question 2
2.1 Step 1: H0: β1 = 0 H1: β1 ≠ 0
^
1  1* 55.6148  0
Step 2: t comp  ^
  6.4168
se( 1 ) 8.6671
Step 3: 𝑡𝑐𝑟𝑖𝑡 = ±2.110
Step 4: Since computed t-stat (6.4168) is greater than both critical t-stats (+2.110),
H0 is rejected.
[Note: don’t you notice tcomp lies in the rejection area in the graphical sketch below?]
Step 5: The alternative hypothesis would take place, i.e. when study hours = 0 hour
and party hours = 0 hour, on average, final mark will be non-zero in PRF.

α/2 α/2

R R
tcrit = -2.110 tcrit = 2.110
tcomp = 6.4168

2
 ECO311 Lecture ‘speech’ notes – Chapter 8

2.2 Step 1: H0: β2 = 0 H1: β2 > 0

^
 2   2* 6.1271 0
Step 2: t comp  ^
  3.8831
se(  2 ) 1.5779
Step 3: 𝑡𝑐𝑟𝑖𝑡 = 2.567
Step 4: Since computed t-stat (3.8831) is greater than critical t-stat (2.567), H0 is
rejected.
[Note: don’t you notice tcomp lies in the rejection area in the graphical sketch below?]
Step 5: The alternative hypothesis would take place, i.e. when study hour increases,
providing there is no change in party hours, on average, final mark will increase in
the PRF.

R
tcrit = 2.567
tcomp = 3.8831
2.3 Step 1: H0: β3 = 0 H1: β3 < 0
^
 3   3*  3.0041 0
Step 2: t comp  ^
  1.9746
se(  3 ) 1.5214
Step 3: 𝑡𝑐𝑟𝑖𝑡 = −2.567
Step 4: Since computed t-stat (-1.9746) is greater than critical t-stat (-2.567), H0 is
NOT rejected.
[Note: don’t you notice tcomp does NOT lie in the rejection area in the graphical sketch
below?]
Step 5: The null hypothesis would take place, i.e. when party hour increases,
providing there is no change in study hours, on average, final mark will NOT
decrease in the PRF (i.e. party hours is an irrelevant explanatory variable in the PRF).

R
tcrit = -2.567
tcomp = -1.9746

From now on, we won’t show the graphical sketch anymore, but we expect you to do it on your own…

3
 ECO311 Lecture ‘speech’ notes – Chapter 8

Question 3
3.2 Step 1: H0: β2 = 4 H1: β2 > 4
^
 2   2* 5.3325  4
Step 2: t comp  ^
  1.3481
se(  2 ) 0.9884
Step 3: 𝑡𝑐𝑟𝑖𝑡 = 1.310
Step 4: Since computed t-stat (1.3481) > critical t-stat (1.310), H0 is rejected.
Step 5: The alternative hypothesis would take place, i.e. if study hour increases by 1
unit/hour, providing there is no change in party hours, on average, final mark will
increase in the PRF, and such increase is greater than 4 units/marks.
3.3 Step 1: H0: β2 = 5 H1: β2 > 5
^
 2   2* 5.3325  5
Step 2: t comp  ^
  0.3364
se(  2 ) 0.9884
Step 3: 𝑡𝑐𝑟𝑖𝑡 = 2.457
Step 4: Since computed t-stat (0.3364) < critical t-stat (2.457), H0 is NOT rejected.
Step 5: The null hypothesis would take place, i.e. if study hour increases by 1
unit/hour, providing there is no change in party hours, on average, final mark will
increase in the PRF, and such increase is equal to 5 units/marks.
3.4 Step 1: H0: β3 = -1 H1: β3 < -1
^
 3   3*  2.0041 (1)
Step 2: t comp  ^
  1.8008
se(  3 ) 0.5576
Step 3: 𝑡𝑐𝑟𝑖𝑡 = −1.697
Step 4: Since computed t-stat (-1.8008) < critical t-stat (-1.697), H0 is rejected.
Step 5: The alternative hypothesis would take place, i.e. if party hour increases by 1
unit/hour, providing there is no change in study hours, on average, final mark will
decrease in the PRF, and such decrease is greater than 1 unit/mark in absolute
terms.
3.5 Step 1: H0: β3 = -2 H1: β3 < -2
^
 3   3*  2.0041 (2)
Step 2: t comp  ^
  0.0074
se(  3 ) 0.5576
Step 3: 𝑡𝑐𝑟𝑖𝑡 = −1.697
Step 4: Since computed t-stat (-0.0074) > critical t-stat (-1.697), H0 is NOT rejected.
Step 5: The null hypothesis would take place, i.e. if party hour increases by 1
unit/hour, providing there is no change in study hours, on average, final mark will
decrease in the PRF, and such decrease is equal to 2 units/marks.

4
 ECO311 Lecture ‘speech’ notes – Chapter 8

Slides #6-#7
 The hypothesis tests we did so far (t-tests) only tested on one parameter at a time.
 However, now we move on to the trickier case of testing the equality TWO PRF parameters.
 It will be done as a two-sided t-test.
 If you look at the Appendix formulae sheet (available on iKamva), you will notice that the
computed F-statistic equation is (assuming it is about testing 𝛽3 = 𝛽4 ):
̂3 − 𝛽
(𝛽 ̂4 ) − (𝛽3 − 𝛽4 )
𝑡𝑐𝑜𝑚𝑝 =
√𝑣𝑎𝑟(𝛽̂3 ) + 𝑣𝑎𝑟(𝛽 ̂4 ) − 2𝑐𝑜𝑣(𝛽̂3 ; 𝛽
̂4 )

- As we want to test if 𝛽3 = 𝛽4 , it makes sense that (𝛽3 − 𝛽4 ) = 0 in the above equation.

- ̂3 and 𝛽
We substitute the values of 𝛽 ̂4 (from the given information) into the equation.
- The given information will only give you 𝑠𝑒(𝛽 ̂3 ) and 𝑠𝑒(𝛽̂4 ), but don’t you remember
that 𝑣𝑎𝑟(𝛽̂3 ) = square of 𝑠𝑒(𝛽 ̂3 ) and 𝑣𝑎𝑟(𝛽̂4 ) = square of 𝑠𝑒(𝛽 ̂4 )?
- The covariance statistic will definitely provided to you in the question.
 So now let’s do the question on slide #7:
Step 1: H0: β3 = β4 H1: β3 ≠ β4
̂3 −𝛽
(𝛽 ̂4 )−(𝛽3 −𝛽4 ) (−12.9615−0.9396)−(0)
Step 2: 𝑡𝑐𝑜𝑚𝑝 = = = −13.3130
̂3 )+𝑣𝑎𝑟(𝛽
√𝑣𝑎𝑟(𝛽 ̂4 )−2𝑐𝑜𝑣(𝛽
̂3 ;𝛽
̂4 ) √(0.9867)2 +(0.0591)2 −2(−0.0576)

Step 3: 𝑡𝑐𝑟𝑖𝑡 (d.f. = n – k = 10 – 4 = 6; α = 0.05) ±2.447

Step 4: Since computed t-stat is smaller than both the critical t-stats, H0 is rejected.
Step 5: The alternative hypothesis would take place, i.e. β3 is NOT equal to β4 in the PRF.
 We will proceed to do Exercise 2, Questions 4-5.

Question 5
5.1 Step 1: H0: β3 = β4 H1: β3 ≠ β4
̂3 −𝛽
(𝛽 ̂4 )−(𝛽3 −𝛽4)
Step 2: 𝑡𝑐𝑜𝑚𝑝 =
̂3 )+𝑣𝑎𝑟(𝛽
√𝑣𝑎𝑟(𝛽 ̂4 )−2𝑐𝑜𝑣(𝛽
̂3 ;𝛽
̂4 )

(1.8745) − (1.2732)
= = 0.3498
√(1.7334)2 + (0.4444)2 − 2(0.1234)
Step 3: 𝑡𝑐𝑟𝑖𝑡 = ±2.617
Step 4: Since computed t-stat lies between the critical t-stats, H0 is NOT rejected.
Step 5: The null hypothesis would take place, i.e. β3 is equal to β4 in the PRF.

 Lastly, read the question carefully; if the test is no longer about the equality of β3 and β4, but
rather about, for example, the equality of β2 and β4 (i.e. β2 = β4), we simply need to make
minor changes to the computed t-stat equation, and then we conduct the t-test as usual:
̂2 − 𝛽
(𝛽 ̂4 ) − (𝛽2 − 𝛽4 )
𝑡𝑐𝑜𝑚𝑝 =
√𝑣𝑎𝑟(𝛽 ̂2 ) + 𝑣𝑎𝑟(𝛽 ̂4 ) − 2𝑐𝑜𝑣(𝛽̂2 ; 𝛽
̂4 )

 Similarly, if the equality test is about whether β2 = β3, the computed t-stat equation is:
̂2 − 𝛽
(𝛽 ̂3 ) − (𝛽2 − 𝛽3 )
𝑡𝑐𝑜𝑚𝑝 =
√𝑣𝑎𝑟(𝛽 ̂2 ) + 𝑣𝑎𝑟(𝛽 ̂3 ) − 2𝑐𝑜𝑣(𝛽̂2 ; 𝛽
̂3 )

5
 ECO311 Lecture ‘speech’ notes – Chapter 8

Slides #9-#10
 F-test on the overall significance of multivariate regression. This is the easiest type of F-test.
 Null hypothesis: all PRF slope parameters equal to zero, i.e. R2 of PRF = 0
 Alternative hypothesis: at least one PRF slope parameter is non-zero, i.e. R2 of PRF is not 0
 The primary ‘goal’ of this F-test is to reject the null hypothesis, as we always hope R2 is not 0.
𝑅2 /(𝑘−1)
 Computed F-statistic: 𝐹𝑐𝑜𝑚𝑝 = (1−𝑅2 )/(𝑛−𝑘)
 Critical F-statistic can be obtained from the F-table, but we need to be certain of the degrees
of freedom of numerator (k – 1) and degrees of freedom of denominator (n – k).
 We will proceed to do Exercise 2, Questions 6-8.

Question 6
Step 1: H0: β2 = β3 = 0
H1: Not all slope parameters are simultaneously equal to zero
R 2 /(k  1) (0.8765) /(3  1)
Step 2: Fcomp    60.3259
(1  R ) /(n  k ) (1  0.8765) /(20  3)
2

Step 3: Fcrit: Degrees of freedom of numerator =k–1=2

Degrees of freedom of denominator = n – k = 17
∴ 𝐹𝑐𝑟𝑖𝑡 = 3.59

Step 4: Since 𝐹𝑐𝑜𝑚𝑝 > 𝐹𝑐𝑟𝑖𝑡 , H0 is rejected.

Step 5: H1 takes place, i.e. at least one slope parameter is non-zero in the PRF.

Question 8
Step 1: H0: β2 = β3 = β4 = 0
H1: Not all slope parameters are simultaneously equal to zero
R 2 /(k  1) (0.2975) /(4  1)
Step 2: Fcomp    2.8233
(1  R ) /(n  k ) (1  0.2975) /(24  4)
2

Step 3: Fcrit: Degrees of freedom of numerator =k–1=3

Degrees of freedom of denominator = n – k = 20
∴ 𝐹𝑐𝑟𝑖𝑡 = 4.94
Step 4: Since 𝐹𝑐𝑜𝑚𝑝 < 𝐹𝑐𝑟𝑖𝑡 , H0 is NOT rejected.

Step 5: H0 takes place, i.e. slope parameters are simultaneously equal to zero in the PRF.

6
 ECO311 Lecture ‘speech’ notes – Chapter 8

Slides #11-#12
 F-test on marginal contribution of additional explanatory variables aims to test if the
robustness of the regression has improved after adding more explanatory variables
(hopefully they are relevant explanatory variables to help explain variations in the Y values).
 Strictly speaking, if the additional explanatory variables are relevant (e.g. it makes sense to
include them, after using theoretical thinking), the primary ‘goal’ of this F-test is to reject the
null hypothesis in the F-test.
 The original regression – we call it the old regression
 The new regression (with additional explanatory variables) – we call it the new regression
(𝑅2 2
𝑛𝑒𝑤 −𝑅𝑜𝑙𝑑 )


𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑎𝑑𝑑𝑖𝑡𝑖𝑜𝑛𝑎𝑙 𝑒𝑥𝑝𝑙𝑎𝑛𝑎𝑡𝑜𝑟𝑦 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒𝑠
Computed F-statistic: 𝐹𝑐𝑜𝑚𝑝 = (1−𝑅2𝑛𝑒𝑤 )
(𝑛−𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑝𝑎𝑟𝑎𝑚𝑒𝑡𝑒𝑟𝑠 𝑖𝑛 𝑡ℎ𝑒 𝑛𝑒𝑤 𝑟𝑒𝑔𝑟𝑒𝑠𝑠𝑖𝑜𝑛)
 Critical F-statistic can be obtained from the F-table, but we need to be certain of the degrees
of freedom of numerator (number of additional explanatory variables) and degrees of
freedom of denominator (n – number of parameters in the new regression).
 Beware of all the terms in the denominator of the computed F-stat equation: these terms all
refer to the new regression with additional explanatory variables added.
 Let’s do the question on slide #12, at α = 0.05:
Step 1: H0: The two additional explanatory variables do not contribute H1: H0 is not true
(0.4707−0.3175)/2
Step 2: 𝐹𝑐𝑜𝑚𝑝 = (1−0.4707)/(33−5) = 6.1910
Step 3: 𝐹𝑐𝑜𝑚𝑝 = 3.34 (d.f. of numerator = 2, d.f. of denominator = 28)
Step 4: Since the computed F-stat is greater than the critical F-stat, H0 is rejected.
Step 5: The alternative hypothesis would take place, i.e. the two additional explanatory
variables (part-time work dummy, laptop dummy) should be included in the regression.
 We will proceed to do Exercise 2, Questions 12-13.

Question 13
Step 1: H0: The three additional explanatory variables do not contribute H1: H0 is not true
(0.6987−0.5588)/3
Step 2: 𝐹𝑐𝑜𝑚𝑝 = (1−0.6987)/(46−6) = 6.1910
Step 3: 𝐹𝑐𝑜𝑚𝑝 = 2.84 (d.f. of numerator = 3, d.f. of denominator = 40)
Step 4: Since the computed F-stat is greater than the critical F-stat, H0 is rejected.
Step 5: The alternative hypothesis would take place, i.e. the three additional explanatory
variables (lecture attendance, tutorial attendance, laptop dummy) should be included in the
regression.

7
 ECO311 Lecture ‘speech’ notes – Chapter 8

Slides #13-#15
 F-test on restricted least squares aims to test if the robustness of the regression has
improved after imposing restrictions on some of the regression parameters (hopefully they
are relevant to explain variations in the Y values).
 The original regression without any restrictions imposed on the regression parameters – we
call it the unrestricted regression
 The revised regression with restrictions imposed on regression parameters – we call it the
restricted regression
2 2
(𝑅𝑈𝑅 −𝑅𝑅 )/𝑚
 F-test on restricted least squares: 𝐹𝑐𝑜𝑚𝑝 = 2 )/(𝑛−𝑘)
(1−𝑅𝑈𝑅
 One trick is that we need to examine the two given regressions very carefully to find out
how many restrictions (m) are imposed on the parameters of the restricted regression.
 Critical F-statistic can be obtained from the F-table, but we need to be certain of the degrees
of freedom of numerator (number of restrictions imposed) and degrees of freedom of
denominator (n – number of parameters in the unrestricted regression).
 Beware of the terms in the denominator of the computed F-stat equation: these terms all
refer to the unrestricted regression with additional explanatory variables added.
 Let’s do the question on slide #15, at α = 0.05:

Unrestricted regression: Year mark = β1 + β2Study + β3Laptop + β4Lecture + β5Economics

Restricted regression: We assume that Study and Lecture have the same impact on year
mark, i.e. β2 = β4. In addition, we assume that Laptop and Economics have the same impact
on year mark, i.e. β3 = β5. In other words, two restrictions are imposed, or m = 2.

This explains why the restricted regression can be simplified as follows:

Year mark = β1 + β2Study + β3Laptop + β4Lecture + β5Economics
Year mark = β1 + β2Study + β3Laptop + β2Lecture + β3Economics
Year mark = β1 + β2(Study + Lecture) + β3(Laptop + Economics)

Step 1: H0: The restricted regression is correct H1: H0 is not true

2 2
(𝑅𝑈𝑅 −𝑅𝑅 )/𝑚 (0.3707−0.3618)/2
Step 2: 𝐹𝑐𝑜𝑚𝑝 = 2 )/(𝑛−𝑘) = (1−0.3707)/(35−5) = 0.2121
(1−𝑅𝑈𝑅
Step 3: 𝐹𝑐𝑜𝑚𝑝 = 3.32 (d.f. of numerator = 2, d.f. of denominator = 30)
Step 4: Since the computed F-stat is smaller than the critical F-stat, H0 is NOT rejected.
Step 5: The null hypothesis would take place, i.e. the restricted regression is better, and it is
incorrect to impose the restrictions that the slope parameters of Study Hours and Lecture
Attendance variables are the same, and the slope parameters of Laptop dummy and Matric
Economics dummy are the same.

 We will proceed to do Exercise 2, Questions 14-15.

8
 ECO311 Lecture ‘speech’ notes – Chapter 8

Question 15
Step 1: H0: The restricted regression is correct H1: H0 is not true
(0.6700−0.4500)/1
Step 2: 𝐹𝑐𝑜𝑚𝑝 = (1−0.6700)/(123−3)
= 80.0000
Step 3: 𝐹𝑐𝑟𝑖𝑡 = 6.85 (d.f. of numerator = 1, d.f. of denominator = 120)
Step 4: Since the computed F-stat is greater than the critical F-stat, H0 is rejected.
Step 5: The alternative hypothesis would take place, i.e. restricted regression is NOT better,
and it is incorrect to impose the restriction that the slope parameters of LABOUR and
CAPITAL are the same.

Slides #16-#18
 F-test on structural instability – Chow Test. Note that this test focuses on time-series
regressions. Later in the chapter on dummy variables we will look at structural instability on
cross-sectional regressions with piecewise linear regression model.
 The aim of the Chow test is to find out whether it is a correct decision to run the regression
twice for two different sub-periods (if it is suspected that the slope parameters have
changed from a particular time onwards).
 Strictly speaking, if we clearly notice an abrupt change in the slope upon doing a rough XY-
scatter plot, then we hope to prove statistically with the Chow test that it is indeed better to
run two separate regressions, one for each sub-period.
 If we run the regression for the full period by ignoring the possible changes in slope
parameters from a particular time onwards – we call it the restricted regression.
 If we subdivide the full period into two (or more) sub-periods and run the regressions for
both sub-periods, we call them the unrestricted regressions. In this module, we will assume
two sub-periods and two unrestricted regressions.
 We will definitely provide you the residual sum of squares of the restricted regression, i.e.
𝑅𝑆𝑆𝑅 .
 We will also definitely provide you the residual sum of squares of each unrestricted
regression (i.e. 𝑅𝑆𝑆1 and 𝑅𝑆𝑆2 ), so that you derive the overall residual sum of square, i.e.
𝑅𝑆𝑆𝑈𝑅 = 𝑅𝑆𝑆1 + 𝑅𝑆𝑆2.
 Please also take note that k = number of parameters, 𝑛1 = sample size of the first sub-period,
𝑛2 = sample size of the second sub-period (note: 𝑛 = 𝑛1 + 𝑛2 ).
(𝑅𝑆𝑆𝑅 −𝑅𝑆𝑆𝑈𝑅 )/𝑘
 Computed F-statistic: 𝐹𝑐𝑜𝑚𝑝 = 𝑅𝑆𝑆
𝑈𝑅 /(𝑛1 +𝑛2 −2𝑘)
 Critical F-statistic can be obtained from the F-table, but we need to be certain of the degrees
of freedom of numerator (number of parameters = k) and degrees of freedom of
denominator (n1 + n2 – 2k).

9
 ECO311 Lecture ‘speech’ notes – Chapter 8

 Let’s do the question on slide #18, at α = 0.01, for the regression savings = f(income):
1990-2001 2002-2015 1990-2015
Unrestricted Unrestricted Restricted
(sub-period1) (sub-period2) (full period)
β1-hat 1.0161 153.4947 62.4226
β2-hat 0.0803 000.0148 00.0376
R2 0.9021 000.2971 00.7672
RSS 1 785.0320 10 005.2200 23 248.3000
n n1 = 12 n2 = 14 n = 26

RSSUR = RSS1 + RSS2 = 1785.0320 + 10005.2200 = 11790.2520

RSSR = 23248.3000 n1 = 12 n2 = 14 n = 26

Step 1: H0: Parameter of INCOME variable is the same for the whole duration (1990-2015)
H1: H0 is not true
(𝑅𝑆𝑆𝑅 −𝑅𝑆𝑆𝑈𝑅 )/𝑘 (23248.3000−11790.2520)/2
Step 2: 𝐹𝑐𝑜𝑚𝑝 = = = 10.6901
𝑅𝑆𝑆𝑈𝑅 /(𝑛1 +𝑛2 −2𝑘) 11790.2520/(12+14−2∙2)
Step 3: 𝐹𝑐𝑟𝑖𝑡 = 5.72 (d.f. of numerator = 2, d.f. of denominator = 22)
Step 4: Since the computed F-stat is greater than the critical F-stat, H0 is rejected.
Step 5: The alternative hypothesis would take place, i.e. the slope parameter of the INCOME
variable is no longer the same since 2002.

 We will proceed to do Exercise 2, Questions 9-10.

Question 10
Step 1: H0: The parameters of the INCOME and R/US$ variable are the same for the whole
duration (1950-2015)
H1: H0 is not true
(752.44−361.53)/3
Step 2: 𝐹𝑐𝑜𝑚𝑝 = 361.53/(36+30−2∙3) = 21.6253
Step 3: 𝐹𝑐𝑟𝑖𝑡 = 4.13 (d.f. of numerator = 3, d.f. of denominator = 60)
Step 4: Since the computed F-stat is greater than the critical F-stat, H0 is rejected.
Step 5: The alternative hypothesis would take place, i.e. the slope parameters of the
INCOME and R/US$ variables are no longer the same since 1986.

Slides #19-#24
 We will use the multivariate regression Marks = f(Study, Attendance) as an example.
 You must be able to distinguish the following statistics from the E-Views regression output
on slide #19 as an example:
- Dependent variable: Marks
- Method: Ordinary Least Squares
- Sample size (i.e. included observations): 20
- Constant parameter (β1-hat): 42.3837
- Slope parameter of Study (β2-hat): 3.8170
- Slope parameter of Attendance (β3-hat): 0.9859

10
 ECO311 Lecture ‘speech’ notes – Chapter 8

- Standard error of β1-hat: 2.5936

- Standard error of β2-hat: 1.6766
- Standard error of β3-hat: 1.1929
- R-squared: 0.9353
- Adjusted R-squared: 0.9277
- Standard error of regression (σ-hat): 5.1149
- Residual sum of squares (RSS): 444.7531
- Durbin-Watson d-statistic (for Ch. 12): 0.7044
- Akaike info criterion (AIC) (for Ch. 13): 6.2397
- Schwarz info criterion (SIC) (for Ch. 13): 6.3890
- Computed F-statistic (for joint significance): 122.7976

- Prob (F-statistic): it means the lowest α level that can reject the H0 on joint significant.
Think of this regression: d.f. of numerator = k – 1 = 2, d.f. of denominator = n – k = 17
Using the F-table, the critical F-statistics are:
𝛼 = 0.01: 𝐹𝑐𝑟𝑖𝑡 = 6.11
𝛼 = 0.05: 𝐹𝑐𝑟𝑖𝑡 = 3.59
𝛼 = 0.10: 𝐹𝑐𝑟𝑖𝑡 = 2.64
[Note: the lower the α, the higher the critical F-statistic]
The computed F-stat (122.7976) is clearly much greater than all three critical F-stats, and
this is why Prob (F-statistic) = 0.0000, meaning alpha must be as small as 0.00 = 0%,
before there is any chance for computed F-stat to be smaller than the critical F-stat, in
order not to reject H0. In other words, we can also say that the joint statistical
significance of this regression is so strong that we are very certain we can reject H0.

- Computed t-stat in connection with β1-hat: (42.3837 – 0) / 2.5936 = 16.3414

- Computed t-stat in connection with β2-hat: (3.8170 – 0) / 1.6766 = 2.2766
- Computed t-stat in connection with β3-hat: (0.9859 – 0) / 1.1929 = 0.8265
- Prob (β1-hat): 0.0000 – it means the lowest α level that can reject H0 in the two-sided
hypothesis test with hypothesized value being zero (i.e. H0: β1 = 0; H1: β1 ≠ 0)
- Prob (β2-hat): 0.0360 – it means the lowest α level that can reject H0 in the two-sided
hypothesis test with hypothesized value being zero (i.e. H0: β2 = 0; H1: β2 ≠ 0)
- Prob (β3-hat): 0.4200 – it means the lowest α level that can reject H0 in the two-sided
hypothesis test with hypothesized value being zero (i.e. H0: β3 = 0; H1: β3 ≠ 0)

11
 ECO311 Lecture ‘speech’ notes – Chapter 8

The last three bullets above may seem very confusing; in order to fully understand what these
probabilities mean, let’s first find out the critical t-statistics (d.f. = 17) at the three different alphas:
𝛼 = 0.01: 𝑡𝑐𝑟𝑖𝑡 = ±2.898
𝛼 = 0.05: 𝑡𝑐𝑟𝑖𝑡 = ±2.110
𝛼 = 0.10: 𝑡𝑐𝑟𝑖𝑡 = ±1.740
[Note: the lower the α, the higher the critical t-statistics in absolute terms]

Starting with the two-sided hypothesis test on β1, given the computed t-stat is a huge value
(16.3414), it means we are very certain that we can reject H0, and this explains why Prob (β1-hat) =
0.000, as it means even if α is small (zero…) and the critical t-stat is very big, we can still reject H0,
because the computed t-stat is too big.

Moving on with the two-sided hypothesis test on β2, given the computed t-stat is 2.2766, it means
we can only reject H0 when α = 0.05 (critical t-stats = + 2.110) or α = 0.10 (critical t-stats = + 1.740),
but we cannot reject H0 when α = 0.01 (critical t-stats = + 2.898). This explains why Prob (β2-hat) =
0.0360, as it means α must be at least 0.036 or 3.6% before there is any chance we can reject H0.
That is, it is impossible to reject H0 when α = 0.01, because 0.01 is smaller than 0.036 (i.e. failing to
meet the ‘α must be at least 0.036’ criterion).

Lastly, for the two-sided hypothesis test on β3, given the computed t-stat is 0.8265, it means we
cannot reject H0 at all three α values. This explains why Prob (β3-hat) = 0.4200, as it means α must be
at least 0.4200 or 42% before there is any chance we can reject H0. That is, it is impossible to reject
H0 when α = 0.01 / 0.05 / 0.10, because 0.01, 0.05 and 0.10 are all smaller than 0.4200 (i.e. failing to
meet the ‘α must be at least 0.4200’ criterion). To conclude, we can say that the computed t-stat is
so weak that we are certain we cannot reject H0 (β3 = 0), i.e. X3 (lecture attendance) could be an
irrelevant explanatory variable that should not feature in the PRF.

12
 ECO311 Lecture ‘speech’ notes – Chapter 8

 Unofficially, we argue that if computed t-statistic lies between -1.96 and +1.96, we will
regard this computed t-statistic as ‘weak’ or insignificant.
[Hint: check the critical t-stat when n is infinite (∞) and α = 0.05 – you will find +-1.96].
 Let’s end this chapter by doing Exercise 2, Question 11 as a revision question.

Question 11
11.1 People who work in the day time could perform worse
People with a laptop could study more efficiently and can perform better
academically
Hence, according to theory, it is argued that work dummy and laptop dummy should
be included as additional explanatory variables.
(33−1)
11.2 Adjusted R2 of regression #1: 1 − (1 − 0.3175) ∙ (33−3) = 0.2720
(33−1)
Adjusted R2 of regression #2: 1 − (1 − 0.4707) ∙ (33−5) = 0.3951
Hence, the second regression has better goodness of fit.
11.3 Step 1: H0: Work dummy and Laptop dummy do not contribute H1: H0 is not true
(0.4707−0.3175)/2
Step 2: 𝐹𝑐𝑜𝑚𝑝 = (1−0.4707)/(33−5) = 4.0521
Step 3: 𝐹𝑐𝑟𝑖𝑡 = 3.34 (d.f. of numerator = 2, d.f. of denominator = 28)
Step 4: Since the computed F-stat is greater than the critical F-stat, H0 is rejected.
Step 5: The alternative hypothesis would take place, i.e. the two additional
explanatory variables should be included in the regression.

13