5/31/2023

National Economics University

E-PhD, Cohort 6

EPHD 3121: Econometrics

LECTURE 1: MULTIPLE REGRSSSION MODELS

Bach Ngoc Thang

[email protected]

Hanoi, 2023

Instructor’s brief profile

• Education
• Ph.D. in Economics (UQ, 2014); M.A. and B.A. in
Development Economics (NEU, 2005 and 2002).
• Teaching
• Macroeconomics 1 & 2 (NEU, since 2014); Econometrics (E-
PhD, since 2019); Principles of Macroeconomics (BBAE,
since 2020), International Trade (UQ, 2010 – 14);
• Research interests
• Development economics, governance, SMEs.
• Some notable publications in World Development, Journal of
Development Studies, Economics of Transition and
Institutional Change, The Developing Economies, etc.
• Google Scholar:
https://scholar.google.com/citations?hl=en&user=nydtUkcAA
AAJ&view_op=list_works&sortby=pubdate

5/31/2023 2

1
 5/31/2023

Outlines
• Course introduction
• From economic to econometric model
• Estimating the parameters
• Hamburger chain data (andy.dta)
• Sampling properties
• Model specification
• Family income equation (edu_inc.dta)
• Poor data, collinearity, and insignificance
• Cars data (cars.dat)
• Required reading: Chap. 5&6 (Hill et al., 2011)
5/31/2023 3

Course introduction
• Course objectives
• Changes in this semester
• Students’ assessment
• Course syllabus
• Students’ expectations?
• What do you expect to learn from this course?
• What could the instructor do to enhance students’ greater
learning outcomes?

5/31/2023 4

2
 5/31/2023

Economic model
• The interplay between sales and advertising
expenditure:
𝑆𝑎𝑙𝑒𝑠 = 𝛽1 + 𝛽2 𝑃𝑟𝑖𝑐𝑒 + 𝛽3 𝐴𝑑𝑣𝑒𝑟𝑡 (1)
Where, 𝛽1 , 𝛽2 , 𝛽3 are the unknown parameters.
• A quantitative inference:
Marginal analysis: change in Sales when Advert
increase by one unit:
∆𝑆𝑎𝑙𝑒𝑠 𝜕𝑆𝑎𝑙𝑒𝑠
𝛽3 = = (2)
∆𝐴𝑑𝑣𝑒𝑟𝑡 (𝑃𝑟𝑖𝑐𝑒 ℎ𝑒𝑙𝑑 𝑐𝑜𝑛𝑠𝑡𝑎𝑛𝑡) 𝜕𝐴𝑑𝑣𝑒𝑟𝑡

5/31/2023 5

Where does an economic model

come from?
• From economic/management theory:
• The relationship between advertising and sales?
• The responsiveness of sales to advertising? Which
goods/services are more responsive to advertising?
• From empirical investigation:
• A toolkit model: lack of theoretical foundations?
• Data mining?
• From practices or observations:
• Are all the facts/objects observable?
• A combination of all the above?
• A due literature review is needed.

5/31/2023 6

3
 5/31/2023

The econometric model

• Again, the sales-advertising nexus:
𝑆𝑎𝑙𝑒𝑠
= 𝐸 𝑆𝑎𝑙𝑒𝑠 + 𝑒 = 𝛽1 +𝛽2 𝑃𝑟𝑖𝑐𝑒 + 𝛽3 𝐴𝑑𝑣𝑒𝑟𝑡 + 𝑒 (3)
• The general model:
𝑦 = 𝛽1 + 𝛽2 𝑥2 +𝛽3 𝑥3 + ⋯ +𝛽𝐾 𝑥𝐾 + 𝑒 (4)
Where, 𝛽1 , 𝛽2 , … , 𝛽𝐾 are the unknown coefficients to be
estimated.
• The marginal analysis:
∆𝐸(𝑦) 𝜕𝐸(𝑦)
𝛽𝑘 = |𝑜𝑡ℎ𝑒𝑟 𝑥𝑠 ℎ𝑒𝑙𝑑 𝑐𝑜𝑛𝑠𝑡𝑎𝑛𝑡 = (5)
∆𝑥𝑘 𝜕𝑥𝑘

5/31/2023 7

The multiple regression plane

5/31/2023 8

4
 5/31/2023

Monthly sales, price, and advertising

in Big Andy’s Burger Barn

5/31/2023 9

Estimating the unknown

parameters
• Minimizing the sum of squares function:
𝑆 𝛽1 , 𝛽2 , 𝛽3
𝑁 𝑁
2 2
= ෍ 𝑦𝑖 − 𝐸(𝑦𝑖 ) = ෍ 𝑦𝑖 − 𝛽1 − 𝛽2 𝑥𝑖2 −𝛽3 𝑥𝑖3 (5)
𝑖=1 𝑖=1
• The least squares estimators:
𝑏1 , 𝑏2 , 𝑏3 correspond to the unknown
parameters/coefficients 𝛽1 , 𝛽2 , 𝛽3 .
• Are the above OLS estimators BLUE?

5/31/2023 10

10

5
 5/31/2023

Stata practice: Sales equation

• Data source: andy.dat
• To do:
• Variable description and summary statistics
• Conducting OLS estimates of Sales on Price and Advert.
• Interpreting estimated results.

5/31/2023 11

11

OLS estimates for Sales equation

5/31/2023 12

12

6
 5/31/2023

The error variance and standard

error (s.e.)
• The error variance:
2 𝑆𝑆𝐸 σ𝑁
𝑖=1 𝑒𝑖Ƹ
2 σ𝑁 ො 𝑖 )2
𝑖=1(𝑦𝑖 −𝑦
𝜎ො = = = (6)
𝑁−𝐾 𝑁−𝐾 𝑁−𝐾
• The standard errors of estimated coefficient 𝑏2 :
𝜎2
𝑣𝑎𝑟 𝑏2 = 2 σ𝑁 (7)
(1 − 𝑟23 ) 𝑖=1(𝑥𝑖2 − 𝑥ҧ2 )2

σ(𝑥𝑖2 −𝑥ҧ2 )(𝑥𝑖3 −𝑥ҧ3 )

𝑟23 = (8)
σ(𝑥𝑖2 −𝑥ҧ2 )2 (𝑥𝑖3 −𝑥ҧ3 )2

5/31/2023 13

13

The standard error (cont’d)

𝜎2
𝑣𝑎𝑟 𝑏2 = 2 σ𝑁 (7𝑎)
(1 − 𝑟23 ) 𝑖=1(𝑥𝑖2 − 𝑥ҧ2 )2

• Factor affecting the variance of 𝑏2 :

• The variance of estimated (random) error 𝜎ො 2 .
• A larger sample size N.
• More variance in an explanatory variable around its
mean.
• A larger correlation between x2 and x3.

5/31/2023 14

14

7
 5/31/2023

Sampling properties:
Assumptions of the MRM
• MR1: 𝑦𝑖 = 𝛽1 + 𝛽2 𝑥𝑖2 + … + 𝛽𝐾 𝑥𝑖𝐾 + 𝑒𝑖 , 𝑖 = 1, … , 𝑁
• MR2: 𝐸(𝑦𝑖 ) = 𝛽1 + 𝛽2 𝑥𝑖2 + … + 𝛽𝐾 𝑥𝑖𝐾 ↔ 𝐸 𝑒𝑖 = 0
• MR3: 𝑣𝑎𝑟(𝑦𝑖 ) = 𝑣𝑎𝑟 𝑒𝑖 = 𝜎 2
• MR4: 𝑐𝑜𝑣(𝑦𝑖 ; 𝑦𝑗 ) = 𝑐𝑜𝑣(𝑒𝑖 , 𝑒𝑗 ) = 0, 𝑖 ≠ 𝑗
• MR5: The values of each 𝑥𝑖𝑘 are not random and are not
exact linear functions of the other explanatory variables.
• MR6: 𝑦𝑖 ~ 𝑁[(𝛽1 + 𝛽2 𝑥𝑖2 + … + 𝛽𝐾 𝑥𝑖𝐾 ), 𝜎 2 ] ↔ 𝑒𝑖 ~
N(0, 𝜎 2 ).

5/31/2023 15

15

Model specification:
Omitted variables
• Essential features of model choice:
• Choice of functional forms
• Choice of explanatory variables to be included in the model
• Whether the assumptions of MR1 – MR6 hold
• Omitted variables:
• The econometric model of family income regressed on
husband’s and wife’s years of education:
𝑦 = 𝛽1 + 𝛽2 𝑥2 +𝛽3 𝑥3 + 𝑒
• The omitted-variable bias of omitting wife’s year of
education:
෣
𝑐𝑜𝑣(𝑥 2 , 𝑥3 )
𝑏𝑖𝑎𝑠 𝑏2∗ = 𝐸 𝑏2∗ − 𝛽2 = 𝛽3 (9)
෣
𝑣𝑎𝑟(𝑥2 )

5/31/2023 16

16

8
 5/31/2023

Correlation matrix:
Family income data

5/31/2023 17

17

Stata practice:
Family income data
• Data source: edu_inc.dat
• To do:
• Regressing family income (Faminc) on both husband’s
and wife’s years of education (Hedu and Wedu).
• Omitting wife’s years of education in the above
specification.
• Determining upward or downward estimates.
• Adding the number of young children (Kl6) as another
regressor.

5/31/2023 18

18

9
 5/31/2023

Family income data: estimated

models
• On both husband’s and wife’s years of education:

• On husband’s years of education only:

• Adding the number of young children:

5/31/2023 19

19

Model specification:
Irrelevant variables
• Adding two artificially generated variables X5 and
X6:

• The consequences of adding irrelevant explanatory

variables:
• Reducing the precision of the estimated coefficients.

5/31/2023 20

20

10
 5/31/2023

Choosing the model

• The basis of theoretical and general understanding
of the relationship.
• Unobserved heterogeneity: omitted-variable bias,…
• Significance tests: F- and T-statistics test.
• Using model selection criteria: Adjusted R-squared,
Akaike information criterion (AIC), Schwarz
criterion (BIC): see pages 237 - 8.
• The general specification test (RESET): see pages
238 - 9.
• Violation of the MRM assumptions MR1 – 6?

5/31/2023 21

21

Poor data, collinearity and

insignificance
• The survey data issues:
• Most economic data is non-experimental, or
“uncontrolled” data.
• The solution: Randomized Control Trial (RCT), quasi-
experiments, field or laboratory experiments, but the
disadvantages?
• More systemic issues: regressors no longer
exogenous or independent
• Correlated by definition/construction.
• Co-movement or confounding factors.
• Unobserved heterogeneity.
5/31/2023 22

22

11
 5/31/2023

Collinearity
• Consequences:
• The standard errors are large, leading to insignificant
estimates of the coefficients/parameters.
• Sensitive estimators, due to addition or deletion of a few
observations, or variables.
• An example:
• Data source: cars.dat
• To do: (i) regressing energy consumption (miles per
gallon, MPG) on number of cylinders; (ii) adding on
engine displacement (ENG) and vehicle weight (WGT).

5/31/2023 23

23

Collinearity:
Estimated models of car data
• On the number of cylinders:

• Adding on engine displacement and vehicle weight:

5/31/2023 24

24

12
 5/31/2023

Identifying and mitigating

collinearity
• Identifying:
• Excessively high partial correlation or R-squared
obtained from the OLS regression of one explanatory
variable on all the remaining explanatory variables. As a
rule of thumb, say above 0.8.
• Mitigating:
• A better sample.
• Using non-sample information, for example, using priors
to impose some restrictions on the parameters.
• Production technology: CTS, IRT, …

5/31/2023 25

25

Next week
• Lecture 2: Using indicator variables
• Indicator and qualitative factors.
• Application.
• Log-linear and log-log model.
• Treatment effects.
• Required reading: Chap. 4&7 (Hill et al., 2011).

5/31/2023 26

26

13