ECO630B: ECONOMETRICS

Innocent Makuta

University of Malawi
Department of Economics

26th October, 2021

Innocent Makuta (UNIMA) ECO630B: ECONOMETRICS 26th October, 2021 1 / 14

ECONOMETRICS

Contents
Introduction
Methodology of Econometrics
Types/Structures of economic data

Innocent Makuta (UNIMA) ECO630B: ECONOMETRICS 26th October, 2021 2 / 14

INTRODUCTION

Question
How would you...?
Test the law of demand for maize in Malawi(price and quantity
demanded)?
Estimate the consumption function (income and consumption
expenditure)?
Estimate the inflation function for Malawi (money supply and
inflation rate)?
Evaluate the effectiveness of FISP on crop production?
Evaluate the effectiveness of an advertising campaign on sales
revenues?

Innocent Makuta (UNIMA) ECO630B: ECONOMETRICS 26th October, 2021 3 / 14

INTRODUCTION

Question
How would you...?
Test the law of demand for maize in Malawi(price and quantity
demanded)?
Estimate the consumption function (income and consumption
expenditure)?
Estimate the inflation function for Malawi (money supply and
inflation rate)?
Evaluate the effectiveness of FISP on crop production?
Evaluate the effectiveness of an advertising campaign on sales
revenues?

Answer
ECONOMETRICS

Innocent Makuta (UNIMA) ECO630B: ECONOMETRICS 26th October, 2021 3 / 14

INTRODUCTION

What is Econometrics?
Literally, Econometrics means ”economic measurement”
Economentrics involves the use of statistical tools to:
(i) test (validate) economic theories
(ii) estimate economic relationships
(iii) evaluate government and business policies
Econometrics forms the bridge between economic theory and the real
world.

Innocent Makuta (UNIMA) ECO630B: ECONOMETRICS 26th October, 2021 4 / 14

INTRODUCTION

What tools are needed in Econometrics?

Economic theory:
Provides the theoretical relationships between/among economic
variables
Relationships expressed verbally/qualitatively
Mathematical economics:
Relationships among economic variables expressed quantitatively as
functions
Assumes deterministic (nonstochastic) relationships among variables
Mathematical statistics:
Functional forms of relationships expressed stochastically to allow for
empirical estimation and testing
Economic statistics:
Provides the raw data for estimating and testing economic
relationships/ theories
Innocent Makuta (UNIMA) ECO630B: ECONOMETRICS 26th October, 2021 5 / 14
INTRODUCTION

Keynes on the Requirements for an Economist

John Maynard Keynes (1951)
”The master-economist must possess a rare combination of gifts. He must
reach a high standard in several directions and must combine talents not
often found together.
He must be mathematician, historian, statesman, philosopher- in
some degree.
He must understand symbols and speak words.
He must contemplate the particular in terms of the general and touch the
abstract and concrete in the flight of thought”

Innocent Makuta (UNIMA) ECO630B: ECONOMETRICS 26th October, 2021 6 / 14

THE METHODOLOGY OF ECONOMETRICS

Here, we outline the procedure/process of analysing an economic problem

Step 1: Statement of theory or hypothesis
Step 2: Specification of the mathematical model
Step 3: Specification of the econometric model
Step 4: Obtaining data
Step 5: Estimation of the econometric model
Step 6: Hypothesis testing
Step 7: Prediction or Forecasting
Step 8: Use of the Model for Control or policy purposes

Innocent Makuta (UNIMA) ECO630B: ECONOMETRICS 26th October, 2021 7 / 14

THE METHODOLOGY OF ECONOMETRICS

Statement of theory or hypothesis

A theory or hypothesis is qualitative (or verbal) expression of the
relationship between/among variables
quantity demanded of a product is negatively (inversely) related to
the product’s own price (law of demand)
Consumption increases with the level of income of an individual (or
household)
Inflation rate increases with money supply
Advertising expenditure increases sales revenues

Innocent Makuta (UNIMA) ECO630B: ECONOMETRICS 26th October, 2021 8 / 14

THE METHODOLOGY OF ECONOMETRICS

Specification of the mathematical model

A mathematical model expresses the relationship between/among
variables in a functional form.
Example:
Y = α + βX
Y is the dependent variable
X is the independent variable
α and β are parameters or coefficients of the model
The sign on the parameter β tells us the direction of the relationship
between X and Y
The numerical value of the parameter β tells us the magnitude of the
response of Y to a unit change in X

Innocent Makuta (UNIMA) ECO630B: ECONOMETRICS 26th October, 2021 9 / 14

THE METHODOLOGY OF ECONOMETRICS

Specification of the econometric model

The mathematical model assumes an exact (or deterministic)
relationship between variables
For example, if
Y = α + βX = 500 + 0.8X
then given X = 4, 000, Y is predicted to be exactly 3, 700
However, relationships among economic variables are generally inexact
To allow for the inexact relationships the mathematical model is
turned into an econometric model by the introduction of the error
term (disturbance term)

Y = α + βX + u

Innocent Makuta (UNIMA) ECO630B: ECONOMETRICS 26th October, 2021 10 / 14

THE METHODOLOGY OF ECONOMETRICS

Obtaining data
Data in collected in order to allow for estimation of the parameters of
the econometric model
Data is usually collected from a sample (not entire population)
through surveys

Estimation of the econometric model

This stage involves calculating/obtaining numerical estimates of the
parameters of the econometric model
Estimation requires use of statistical methods such as Least Squares
Method or Maximum likelihood method

Innocent Makuta (UNIMA) ECO630B: ECONOMETRICS 26th October, 2021 11 / 14

THE METHODOLOGY OF ECONOMETRICS

Hypothesis testing
This involves confirmation or refutation of the theory or hypothesis in
step 1 on the basis of the estimates obtained in step 5
Hypothesis testing involves use of inferential statistical tools such as
confidence interval method or test-of-significance method

Innocent Makuta (UNIMA) ECO630B: ECONOMETRICS 26th October, 2021 12 / 14

THE METHODOLOGY OF ECONOMETRICS

Prediction or Forecasting
Here, we use the model to
predict the future value(s) of the dependent, or forecast, variable Y
on the basis of known or expected future value(s) of the independent,
or predictor, variable X

Use of the Model for Control or policy purposes

Here, we use the model to
achieve targeted/desired level of the dependent, or target, variable Y
by manipulating value(s) of the independent, or control, variable X

Innocent Makuta (UNIMA) ECO630B: ECONOMETRICS 26th October, 2021 13 / 14

STRUCTURE OF ECONOMIC DATA

Cross-sectional data:
This is data collected on one or variables collected from several
observation units (households, firms, countries) at the same point
in time
Time series data:
This is data collected on one or variables collected from a particular
observation unit at different times
The data is collected at regular time intervals e.g. daily, weekly,
monthly, quarterly, annually etc
Panel (or Longitudinal) data:
This is data collected on one or variables collected from several
observation units (households, firms, countries) over time
Panel has both cross-sectional and time series components

Innocent Makuta (UNIMA) ECO630B: ECONOMETRICS 26th October, 2021 14 / 14

 REGRESSION ANALYSIS: ESTIMATION

Innocent Makuta

University of Malawi
Department of Economics

22nd July, 2021

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 1 / 136
CONTENTS

Simple Regression
Extensions to Simple Regression
Multiple Regression

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 2 / 136
SIMPLE REGRESSION

Contents
Definition of Regression Analysis
Population regression line
Population regression function
Sample regression function
Point Estimation: Least Squares
Numerical Properties Least Squares Estimators
Assumptions
Statistical Properties Least Squares
Gauss-Markov Theorem
Coefficient of Determination
Point Estimation: Maximum Likelihood

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 3 / 136
REGRESSION ANALYSIS

What is Regression Analysis?

Suppose you have two random variables X and Y , such that the
values assumed by Y are dependent on the values assumed by X .

X ⇒ Independent variable
Y ⇒ Dependent variable

Regression analysis is concerned with the estimation and/or

predicting the population mean value of the dependent variable on
the basis of known or fixed values of the independent variables

E (Y |X )

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 4 / 136
ILLUSTRATION

Suppose that one collects the following consumption(Y) and income (X)
from 27 households
Income(X) Consumption(Y)
400 370 375 375 380 380 380 385 385 390
500 440 445 445 450 450 450 455 455 460
600 520 525 525 530 530 530 535 535 540

From the information in the table, the unconditional mean of Y is

E (Y ) = 450

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 5 / 136
ILLUSTRATION

The conditional means of Y is calculated for each level of income. These

conditional means are shown in last column of the table below

Income Consumption E (Y |X )
400 370 375 375 380 380 380 385 385 390 380
500 440 445 445 450 450 450 455 455 460 450
600 520 525 525 530 530 530 535 535 540 530

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 6 / 136
POPULATION REGRESSION LINE (PRL)
The PRL is the line connecting all the conditional means of the dependent
variable E (Y |X )

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 7 / 136
POPULATION REGRESSION FUNCTION (PRF)

As can be seen from the graph, the conditional expected values

E (Y |Xi ) increase with Xi
That is, E (Y |Xi ) are a function of Xi

E (Y |Xi ) = f (Xi )

This function is the Population Regression Function (PRF)

The PRF shows how the mean/ average response of Y varies with X

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 8 / 136
POPULATION REGRESSION FUNCTION (PRF)

The functional form of the PRF E (Y |Xi ) = f (Xi ) is determined

empirically (or from theory)
Assumption: The PRF is linear. That is, E (Y |Xi ) is a linear function
of Xi
The PRF can the be written as

E (Y |Xi ) = β0 + β1 Xi

Linearity in variables vs linearity in parameters

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 9 / 136
POPULATION REGRESSION FUNCTION (PRF)

Actual Y values Yi are spread around the conditional mean E (Y |Xi )

item the deviation of actual Y values Yi around the conditional mean
E (Y |Xi ) is the error term (or disturbance term), denoted ui
That is,
ui = Yi − E (Y |Xi )
Stochastic specification of the PRF

Yi = E (Y |Xi ) + ui = β0 + β1 Xi + ui

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 10 / 136
Importance of the Error Term

Substitute for ommited variables

Vagueness of theory
Unavailability of data
Peripheral variables

Measurement errors (poor proxy variables)

Wrong functional form
Intrinsic randomness of human behaviour

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 11 / 136
SAMPLE REGRESSION FUNCTION

In practice,we collect/have information from a sample an not the

population
Therefore, the PRF is estimated on using sample information
The Sample Regression Function (SRF) may written as:

Ybi = βb0 + βb1 Xi

where:
Ybi is the estimator of E (Y |Xi )
βb0 is the estimator of β0
βb1 is the estimator of β1

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 12 / 136
SAMPLE REGRESSION FUNCTION

The stochastic form of the Sample Regression Function (SRF) may written
as:
Yi = Ybi + ubi = βb0 + βb1 Xi + ubi
where:
ubi is the estimator of E (ui )
ubi is known as the residual

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 13 / 136
TERMINOLOGY

Independent Variable Dependent Variable

Explanatory variable Explained variable
Stimulus variable Response variable
Predictor variable Predicted variable
Exogenous variable Endogenous variable
Regressor Regressand

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 14 / 136
ESTIMATION

Recall: We want to estimate the PRF

Yi = E (Y |Xi ) + ui = β0 + β1 Xi + ui

on the basis of the SRF

Yi = βb0 + βb1 Xi + ubi

Methods of Estimation
1 Least Squares
2 Maximum Likelihood
3 Method of Moments

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 15 / 136
ORDINARY LEAST SQUARES (OLS)

Given a set of data on random variables X and Y , The OLS

procedure fits a sample regression line (SRL) to the data.
Fitting criterion of OLS
OLS fits the SRL to the data such that the sum of the squared
residuals is as small as possible. That is,
X
minimise ubi 2

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 16 / 136
ORDINARY LEAST SQUARES (OLS)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 17 / 136
ORDINARY LEAST SQUARES (OLS)

Using the Least Squares fitting criterion, βb1 is given as:

P P P
n Xi Yi − Xi Yi
β1 =
b P 2 P 2
(a)
n Xi − ( Xi )
P
(Xi − X )(Yi − Y )
= (b)
(Xi − X )2
P
P
xi yi
= P 2 (c)
xi

while βb0 is given as:

βb0 = Y − βb1 X (d)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 18 / 136
Derivation of the expression for βb0 and βb1

Recall: We want to estimate the PRF

Yi = E (Y |Xi ) + ui = β0 + β1 Xi + ui (1)
on the basis of the SRF

Yi = βb0 + βb1 Xi + ubi (2)

Making ubi subject in equation (2), we get:

ubi = Yi − βb0 − βb1 Xi (3)

Squaring and summing on both sides of equation (3), we have:

X X 2
ubi 2 = Yi − βb0 − βb1 Xi (4)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 19 / 136
Derivation of the expression for βb0 and βb1

ubi 2 partially wrt βb0 and βb1 we obtain:

P
Differentiating

ubi 2
P
∂ X  X 
=2 Yi − βb0 − βb1 Xi (−1) = −2 Yi − βb0 − βb1 Xi (5)
∂ βb0

ubi 2
P
∂ X  X 
=2 Yi − βb0 − βb1 Xi (−Xi ) = −2 Yi − βb0 − βb1 Xi (Xi )
∂ βb1
(6)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 20 / 136
Derivation of the expression for βb0 and βb1

Setting the partial derivatives in equations (5) and (6) equal to zero above
yield the following FOCs (also called normal equations)
X 
−2 Yi − βb0 − βb1 Xi = 0
X X
⇒ Yi = nβb0 + βb1 Xi (7)

X 
−2 Yi − βb0 − βb1 Xi (Xi ) = 0
X X X
⇒ Yi Xi = βb0 Xi + βb1 Xi 2 (8)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 21 / 136
Derivation of the expression for βb0 and βb1

Solving equations (7) and (8) simultaneously, we get:

X X X
n × (8) : n Yi Xi = nβb0 Xi + nβb1 Xi 2
X X X X X 2 
− Xi × (7) : − Yi Xi = nβ0
b Xi + β1
b Xi
 
X X X X  X  2
n Yi Xi − Yi Xi = βb1 n Xi 2 − Xi (9)

Solving for βb1 in equation (9) yields:

P P P
n Xi Yi − Xi Yi
βb1 = P 2 P 2
(a)
n Xi − ( Xi )

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 22 / 136
Derivation of the expression for βb0 and βb1
Recall that:
P P
X X Xi Yi
(Xi − X )(Yi − Y ) = Xi Yi −
n
and
( Xi )2
X X P
(Xi − X )2 = Xi 2 −
n

Therefore:
P P
P Xi Yi
P
(Xi − X )(Yi − Y ) Xi Yi −
= n
(Xi − X )2 P 2 ( Xi )2
P P
Xi −
P Pn P
n Xi Yi − Xi Yi
= P 2 P 2
n Xi − ( Xi )
Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 23 / 136
Derivation of the expression for βb0 and βb1
Hence
P P P
n Xi Yi − Xi Yi
βb1 = P 2 P 2
(a)
n Xi − ( Xi )
P
(Xi − X )(Yi − Y )
= (b)
(Xi − X )2
P

Letting xi = Xi − X and yi = Yi − Y then:

P
(Xi − X )(Yi − Y )
βb1 = (b)
(Xi − X )2
P
P
xi yi
= P 2 (c)
xi

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 24 / 136
Derivation of the expression for βb0 and βb1

Since the value of βb1 is now known, we can substitute this value into (7)
X X
Yi = nβb0 + βb1 Xi

and solve for βb0

X X
nβb0 = Yi − βb1 Xi
P P
Yi Xi
βb0 = − βb1
n n
= Y − β1 X
b (c)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 25 / 136
OLS technique: An Example

Calculate βb0 and βb1

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 26 / 136
Derivation of the expression for βb0 and βb1
βb1 can alternatively be expressed as follows:
P
xi yi
β1 = P 2
b
xi
P

xi Yi − Y
= P 2
xi
P P
xi Yi − Y xi
= P 2
xi
P
xi Yi
= P 2
xi

and also as:

X xi
βb1 = wi Yi where wi = P 2
xi
wi is called the least squares weight
Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 27 / 136
NUMERICAL PROPERTIES OF OLS ESTIMATORS

The sample regression function

Yi = βb0 + βb1 Xi + ubi

can be expressed in deviation form as follows

 
Yi = Y − βb1 X + βb1 Xi + ubi


Yi − Y = βb1 Xi − X + ubi
yi = βb1 xi + ubi

Notice that, in deviation form,

ybi = βb1 xi

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 28 / 136
PROPERTIES OF OLS ESTIMATORS

Numerical/algebraic properties
Sample-level properties
Assumption-free

Statistical properties
Population-level properties
Assumption-dependent

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 29 / 136
NUMERICAL PROPERTIES OF OLS ESTIMATORS
Property 1: The regression line passes through the sample means of X
and Y (X , Y )

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 30 / 136
NUMERICAL PROPERTIES OF OLS ESTIMATORS

Property 2: The mean value of the estimated Y is equal to the mean

value of the actual Y
Yb = Y
Proof
 

Ybi = βb0 + βb1 Xi = Y − βb1 X + βb1 Xi = Y + βb1 Xi − X

Summing on both sides yields

X X X

Ybi = Y + βb1 Xi − X = nY

Dividing both sides by n

Yb = Y

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 31 / 136
NUMERICAL PROPERTIES OF OLS ESTIMATORS
Property 3: The mean value of the residuals is zero
X
ubi = ub = 0

Proof
X X 
ubi = Yi − βb0 − βb1 Xi
Xh   i
= Yi − Y − β1 X − β1 Xi
b b
X
X

= Yi − Y + βb1 Xi − X
=0

And so,
P
ubi 0
ub = = =0
n n

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 32 / 136
NUMERICAL PROPERTIES OF OLS ESTIMATORS
Property 4: The residuals are uncorrelated with the predicted Ys
X X
ubi ybi = ubi Ybi = 0

Proof
X X
ubi ybi = (yi − ybi ) ybi
X  
= yi − βb1 xi βb1 xi
X 2X 2
= βb1 xi yi − βb1 xi
xi yi 2 X 2
P X P 
xi yi
= P 2 xi yi − P 2 xi
xi xi
( xi yi )2 ( xi yi )2
P P
= P 2 − P 2
xi xi
=0

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 33 / 136
NUMERICAL PROPERTIES OF OLS ESTIMATORS

Now
X X  
ubi ybi = ubi Ybi − Y
X X
= ubi Ybi − Y ubi
X
= ubi Ybi

Therefore X X
ubi ybi = 0 ⇒ ubi Ybi = 0

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 34 / 136
NUMERICAL PROPERTIES OF OLS ESTIMATORS
Property 5: The residuals are uncorrelated with the explanatory variable
X X
ubi xi = ubi Xi = 0

Proof
X X
ubi xi = (yi − ybi ) xi
X 
= yi − βb1 xi xi
X X
= xi yi − βb1 xi2
P 
X xi yi X 2
= xi yi − P 2 xi
xi
X X
= xi yi − xi yi
=0

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 35 / 136
ASSUMPTIONS

Assumption 1:Linear Regression Model

The regression model is linear in parameters
(The model may or may not be linear in the variables)

Yi = β0 + β1 Xi + ui (1)
Yi = β0 + β1 Xi + β1 Xi2 + ui (2)
lnYi = β0 + β1 lnXi + ui (3)
1
Yi = β0 + Xi + ui (4)
β1
1
Yi = + ui (5)
β0 + β1 Xi

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 36 / 136
ASSUMPTIONS

Assumption 2: Conditional Regression

The values of X are fixed in repeated sampling
X is nonstochastic

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 37 / 136
ASSUMPTIONS

Assumption 3:Zero mean value of the error term

E (ui |Xi ) = 0
Implication:
The average value of the errors for each level of the regressor is zero
Factors not explicitly introduced in the regression model do not
systematically affect the mean value of Y

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 38 / 136
ASSUMPTIONS

Assumption 4: Homoskedasticity
The error has a constant variance for every given value X

var (ui |Xi ) = σ 2

Notation

var (ui |Xi ) = E [ui − E (ui |Xi )|Xi ]2

= E (ui2 |Xi ) − [E (ui |Xi )]2
= E (ui2 |Xi )

Implication:
All Y values corresponding to the different levels of the X values are
equally reliable and therefore equally important

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 39 / 136
ASSUMPTIONS

Figure: Homoskedasticity

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 40 / 136
ASSUMPTIONS

Figure: Heterokedasticity

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 41 / 136
ASSUMPTIONS

Assumption 5: No autocorrelation/ No serial correlation

cov (ui , uj |Xi , Xj ) = 0
Notation

cov (ui , uj |Xi , Xj ) = E {[ui − E (ui |Xi )|Xi ] [uj − E (uj |Xj )|Xj ]}
= E (ui |Xi )(uj |Xj ) since E (ui |Xi ) = E (uj |Xj ) = 0
= E (ui uj )

Implication:
There is no systematic effect of excluded variables (and therefore
subsumed in the error term) on Y as a result of possible associations
among the errors

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 42 / 136
ASSUMPTIONS

Assumption 6: Zero covariance between the the error term and

explanatory variables
cov (ui , Xi ) = E (ui Xi ) = 0
Notation

cov (ui , Xi ) = E {[ui − E (ui |Xi )|Xi ] [Xi − E (Xi |Xi )|Xi ]}
= E {ui [Xi − E (Xi )]}
= E (ui Xi ) − E (Xi )E (ui )
= E (ui Xi )

Implication
The effect of X on Y can be assessed independently of the ommitted
variables and therefore subsumed in the error term)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 43 / 136
ASSUMPTIONS

Assumption 7:No specification error

Specification relates to:
number of variables included in the model
functional form of the model
probabilistic assumption about the error term
There is no error with respect to any of the above

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 44 / 136
ASSUMPTIONS

Assumption 8: No perfect multicollinearity

Given the PRF:
Yi = β0 + β1 X1i + β2 X2i + ui
X1 and X2 are not linearly dependent
Implication: It is possible to assess the separate influence of X1 and X2

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 45 / 136
ASSUMPTIONS

Assumption 9: Normality
The error term is normally distributed. Combined with assumptions 3 and
4, we have that:
ui ∼ N(0, σ 2 )
Adding assumption 5, it means the errors are normally and independently
distributed (NID)
ui ∼ NID(0, σ 2 )
Justification for assuming Normality
Central Limit Theorem (CLT)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 46 / 136
ASSUMPTIONS

Assumption 9: Normality
Consequences of assuming Normality
Yi is normally distributed
Recall:

Yi = β0 + β1 Xi + ui

Since X, β0 and β1 are constant/ fixed, then

Yi = f (ui )

Since ui is normally distributed,then Yi is also normally distributed

Yi ∼ N β0 + β1 Xi , σ 2

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 47 / 136
ASSUMPTIONS

Assumption 9: Normality
Consequences of assuming Normality
βb1 is normally distributed
Recall:
X
βb1 = wi Yi
X
= wi (β0 + β1 Xi + ui )

Since X, β0 and β1 are constant/ fixed, then

βb1 = f (ui )

Since ui is normally distributed,then βb1 is also normally distributed

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 48 / 136
ASSUMPTIONS

Assumption 9: Normality
Consequences of assuming Normality
βb0 is normally distributed
Recall:

βb0 = Y − βb1 X

Since X, Y and X are constant/ fixed, then

βb0 = f (βb1 )

Since βb1 is normally distributed,then βb0 is also normally distributed

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 49 / 136
STATISTICAL PROPERTIES OF ESTIMATORS

A good estimator should possess the following properties properties:

Finite Sample properties
Unbiasedness
Minimum Variance

Asymptotic properties
Consistency
Asymptotic normality

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 50 / 136
Unbiasedness

Suppose we are interested in estimating a parameter θ.

Assume the estimator is denoted by θb
Bias of an estimator is the difference between the mean value of the
estimator and the true value of the parameter

bias(θ) b −θ
b = E (θ)

When the mean value of the estimator is equal to the true value of
the parameter, the estimator is said to be unbiased

Unbiasedness ⇒ E (θ)
b =θ

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 51 / 136
Unbiasedness

θb1 is unbiased
θb2 is biased

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 52 / 136
Minimum Variance

Recall: Variance is a measure of precision or reliability of an estimator

The smaller the variance the more precise an estimator is.
An estimator (say θb1 ) is said to be a minimum variance estimator of a
parameter (say θ) if the variance of θb1 is the least/smallest in a class
of all estimators of the parameter

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 53 / 136
Minimum Variance

θb1 is a minimum variance estimator

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 54 / 136
Consistency

An estimator (say θb1 ) is said to be a consistent estimator of a

parameter (say θ) if it approaches the true value of the parameter as
the sample size increases indefinitely.
Two conditions ensures that mean that the bias of the estimator
decreases as the sample size increases indefinitely.

⇒ the bias of the estimator decreases

⇒ the variance of the estimator decreases

More formally  
plim θb1 = θ

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 55 / 136
Consistency

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 56 / 136
STATISTICAL PROPERTIES OF OLS ESTIMATORS

Under the classical (Gaussian) assumptions, OLS estimator possess the

following properties properties:
Linearity
Unbiasedness
Minimum Variance
Consistency
Asymptotic normality

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 57 / 136
Finite Sample Statistical properties

Property: Linearity
Recall that
X xi
βb1 = wi Yi where wi = P 2
xi

This can be rewritten as:

X
βb1 = wi Yi = w1 Y1 + w2 Y2 + ... + wn Yn

This means that βb1 is linear function of Y

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 58 / 136
Finite Sample Statistical properties

The least squares weight w has the following properties:

P
X xi
wi = P 2 = 0 (1)
x
Pi 2
X x 1
wi2 = P i
2 = P 2 (2)
xi2 xi
  P 2
x x
P i 2 xi = P i2 = 1
X X
wi xi = (3)
xi xi

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 59 / 136
Unbiasedness of OLS estimators

E (βb1 ) = β1

Proof
P P
xi yi xi Yi X
β1 = P 2 = P 2 =
b wi Yi
xi xi
X
= wi (β0 + β1 Xi + ui )
X X X
= β0 wi + β1 wi Xi + wi ui
X
= β1 + wi ui

P
βb1 = β1 + wi ui

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 60 / 136
Unbiasedness of OLS estimators

Taking conditional expectations gives

X
E (βb1 |Xi ) = β1 + E (wi ui |Xi ) = β1

Now by the law of iterative expectation

h i
E (βb1 ) = Ex E (βb1 |Xi ) = Ex [β1 ] = β1

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 61 / 136
VARIANCE OF OLS ESTIMATORS

X σ2
var (βb1 ) = var (βb1 |Xi ) = σ 2 wi2 = P 2
xi
Proof
h i2
var (βb1 |Xi ) = E βb1 − E (βb1 |Xi )|Xi
h i2
= E βb1 − β1 |Xi
hX i2
=E wi ui |Xi
= E [(w1 u1 + w2 u2 + ... + wn un ) (w1 u1 + w2 u2 + ... + wn un ) |Xi ]2
= E w1 2 u1 2 + E w2 2 u2 2 + ... + E wn 2 un 2

+ 2E (w1 w2 u1 u2 ) + 2E (w2 w3 u2 u3 ) + ... + 2E (wn−1 wn un−1 un )

= w1 2 E u1 2 + w2 2 E u2 2 + ... + wn 2 E un 2

+ 2w1 w2 E (u1 u2 ) + 2w2 w3 E (u2 u3 ) + ... + 2wn−1 wn E (un−1 un )

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 62 / 136
VARIANCE OF OLS ESTIMATORS

Using assumptions of homoskedasticity and no serial correlation, we have

that:

var (βb1 |Xi ) = w1 2 σ1 2 + w2 2 σ2 2 + ... + wn 2 σn 2

 
X 1
= σ2 wi2 = σ 2 P 2
xi
σ2
=P 2
xi

Since the Xs are assumed to be non-stochastic, then the conditional

variance is equal to unconditional variance. That is
X σ2
var (βb1 |Xi ) = var (βb1 ) = σ 2 wi2 = P 2
xi

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 63 / 136
ESTIMATOR OF σ 2
σ 2 is unobservable since we don’t have information from the entire
population.
Given the simple regression
Yi = βb0 + βb1 Xi + ubi
unbiased estimator of σ 2 is:
ubi 2
P
2
σ
b =
n−2
Generalisation: Given the regression
Yi = βb0 + βb1 X1i + βb2 X2i + ... + βbk−1 Xk−1,i + ubi
unbiased estimator of σ 2 is:
ubi 2
P
b2 =
σ
n−k
where k is the number of parameters in the regression model
Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 64 / 136
GAUSS-MARKOV (GM) THEOREM

Gauss-Markov Theorem
Given the assumptions of classical linear regression model, the least
squares estimators have the minimum variance in a class of all linear
and unbiased estimators
That is, least squares estimators are BLUE (Best among Linear
Unbiased Estimators).

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 65 / 136
GAUSS-MARKOV (GM) THEOREM

Proof
Recall that the OLS estimator for β1 is
X
βb1 = wi Yi

where wi P
are the least squares weights
˜
Let β1 = ai Yi be another linear estimator where the weight wi is

ai = wi + bi

That is,
X X
β˜1 = ai Yi = ai Yi
X
= ai (β0 + β1 Xi + ui )
X X X
= β0 ai + β1 ai Xi + ai ui

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 66 / 136
GAUSS-MARKOV (GM) THEOREM

For β̃ to be unbiased [ie E (β̃) = β], the conditions are that:

X
ai = 0 (1)
X
ai Xi = 1 (2)

In this case,
X X X X
β˜1 = β0 ai + β1 ai Xi + ai ui = β1 + ai u i

Taking expectations on both sides and using the assumption of

noncorrelation between the regressors and the errors, we have:
 
E β˜1 = β1

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 67 / 136
GAUSS-MARKOV (GM) THEOREM
Now
X X XX
ai = 0 ⇒ (wi + bi ) =
wi + bi = 0
X X
⇒ bi = 0 since wi = 0
and
X X X X
ai Xi = 1 ⇒ (wi + bi )Xi =wi Xi + bi Xi = 0
X X
⇒ bi Xi = 0 since wi Xi = 1
and
X X X X X
ai2 = 0 ⇒ (wi + bi )2 = wi2 + bi2 + 2 wi bi
 
x
P i 2 bi
X X X
⇒ wi2 + bi2 + 2
xi
X X
2 2
⇒ wi + bi

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 68 / 136
GAUSS-MARKOV (GM) THEOREM
Now, the variance of β˜1 is given by
h i2 h i2
var (β˜1 |Xi ) = E β˜1 − E (β˜1 |Xi )|Xi = E β˜1 − β1 |Xi
hX i2
=E ai ui |Xi

Using assumptions of homoskedasticity and no serial correlation, we have

that:
X X X
var (β˜1 |Xi ) = σ 2 ai2 = σ 2 wi2 + σ 2 bi2
X
= var (βb1 |Xi ) + σ 2 bi2

And so
X
var (β˜1 |Xi ) ≥ var (βb1 |Xi ) since σ 2 bi2 ≥ 0

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 69 / 136
Consistency
 
plim βb1 = β1
Proof
X
βb1 = β1 + wi ui
P 
(Xi − X )(ui − u) /n
P
xi ui
= β1 + P 2 = β1 + P 
xi (Xi − X )2 /n

Taking probability limits on both sides we have that:

P 
  plim (Xi − X )(ui − u) /n
plim β1 = plim (β1 ) +
b P 
plim (Xi − X )2 /n
cov (Xi , ui )
= β1 +
var (Xi )
= β1
Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 70 / 136
COEFFICIENT OF DETERMINATION

The coefficient of determination (denoted r 2 ) provides a measure of

goodness of fit.
r 2 is bounded between zero and one (or 100%)
r 2 measures the proportion of the changes in the dependent variable
explained by the regression model
Example: If r 2 = 0.7 means that, on average, the model explains 70%
of the variation in the dependent variable.

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 71 / 136
COEFFICIENT OF DETERMINATION
Recall:

yi = βb1 xi + ubi = ybi + ubi

Squaring and summing on both sides, we obtain

X X X X
yi2 = ybi 2 + ubi 2 + 2 ybi ubi
X X X
yi2 = ybi 2 + ubi 2
TSS = ESS + RSS

where:

TSS is the total sum of squares

ESS is the explained sum of squares
RSS is the residual sum of squares

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 72 / 136
COEFFICIENT OF DETERMINATION
Now,

ESS RSS
r2 = =1− (1)
TSS TSS
P 2
ybi
= P 2 (2)
yi
2P 2 P 2 
βb1 xbi 2 x
= P 2 = βb1 P i2 (3)
yi y
2  i
2 Sx2
P  
2 x i /(n − 1)
= βb1 P 2 = βb1 (4)
yi /(n − 1) Sy2
( xi yi )2
P
= (5)
xi2 yi2
P P

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 73 / 136
COEFFICIENT OF CORRELATION

The coefficient of correlation, denoted by r , is the square root of the

coefficient of determination.
P
xi yi
r = qP P
xi2 yi2

It is a measure of linear association or linear dependence AND not a

cause-and-effect relationship
It is bounded between −1 and 1

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 74 / 136
ANALYSIS OF VARIANCE

Study of the components of TSS.

Recall:

TSS = ESS + RSS

X X X
yi2 = ybi 2 + ubi 2

Each of the components has its degrees of freedom (df)

X X X
yi2 = ybi2 + ubi2
(n − 1) = (k − 1) + (n − k)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 75 / 136
ANALYSIS OF VARIANCE

ANOVA table
Source SS df MSS=SS/df
P 2
βb2 x 2
P
Model ybi 1
P1 2 i
P 2 ubi
Residuals ubi n−2 b2
=σ
P 2 n−2
Total yi n−1

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 76 / 136
Maximum Likelihood Estimation

Given the regression

Yi = β0 + β1 Xi + ui

The error vector is assumed to be normally distributed. That is,

ui ∼ N(0, σ 2 )

One implication of this assumption is that the dependent variable y is

normally distributed

Yi ∼ N(β0 + β1 Xi , σ 2 )

This is because Y is a linear combination of u. And a linear

combination of a normally distributed variable is also normally
distributed

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 77 / 136
Maximum Likelihood Estimation

Fact
E (ui uj |Xi Xj ) = 0 for i 6= j implies that

E (Yi Yj |Xi Xj ) = 0

That is, if the errors are uncorrelated then the y ’s are also
uncorrelated
Therefore, each yi is normally and independently distributed

Yi ∼ NID(β0 + β1 Xi , σ 2 )

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 78 / 136
Maximum Likelihood Estimation

Given
Yi = β0 + β1 Xi + ui where Yi ∼ NID(β0 + β1 Xi , σ 2 ) (1)
then the probability density function (PDF) of Yi is

(Yi − β0 − β1 Xi )2
 
1
f (Yi ) = √ exp − (2)
πσ 2 2σ 2

Because the Yi are independent of each other, the joint PDF can be
written as
n
(Yi − β0 − β1 Xi )2
  P 
2 1
f (Y1 , Y2 , ..., Yn |β0 , β1 , σ ) = √ exp −
πσ 2 2σ 2
(3)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 79 / 136
Maximum Likelihood Estimation

The likelihood function is

!
(Yi − β˜0 − β̃1 Xi )2
P
− n2
L(β˜0 , β˜1 , σ̃ 2 |yi , X ) = (πσ̃ 2 ) exp − (4)
2σ̃ 2

Taking natural log on both sides of equation (4), we obtain the

log-likelihood function:

(Yi − β˜0 − β˜1 Xi )2

P
n n
lnL(β˜0 , β˜1 , σ̃ 2 |yi , X ) = − lnπ − lnσ̃ 2 − (5)
2 2 2σ˜2

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 80 / 136
Maximum Likelihood Estimation

Taking the partial derivative of equation (5) wrt β˜0 , β˜1 and σ˜2

(Yi − β˜0 − β˜1 Xi )

P
∂lnL(.)
= (6)
∂ β˜0 σ̃ 2
(Yi − β˜0 − β˜1 Xi )(Xi )
P
∂lnL(.)
= (7)
∂ β˜1 σ̃ 2
(Yi − β˜0 − β˜1 Xi )2
P
∂lnL(.) n
= − + (8)
∂ σ̃ 2 2σ̃ 2 (σ̃ 2 )2

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 81 / 136
Maximum Likelihood Estimation
Setting the first partial derivatives in (6) and (7) equal to zero yields the
following normal equations

(Yi − β˜0 − β˜1 Xi )

P
=0
σ˜2 X X
⇒ Yi = nβb0 + βb1 Xi (9)
(Yi − β˜0 − β˜1 Xi )(Xi )
P
=0
σ˜2 X X X
⇒ Yi Xi = βb0 Xi + βb1 Xi 2 (10)

Important: Normal equations (9) and (10) are exactly the same as
those obtained under the OLS procedure
Therefore, under normality, ML estimators are identical to OLS
estimators
Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 82 / 136
Maximum Likelihood Estimation

Setting the first partial derivatives in (8) equal to zero yields the following
normal equation

(Yi − β˜0 − β˜1 Xi )

P
n
− 2+ =0
2σ̃ (σ̃ 2 )2
(Yi − β˜0 − β˜1 Xi )2
P P 2
2 ui
⇒ σ̃ = = (11)
n n
Important:
ui2
 P 
Since the LS estimator b2
σ b2
σ = is unbiased, it follows that ML
 P 2 n−k
2 2 ui
estimator σ̃ σ̃ = is biased
n

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 83 / 136
Extensions to the Simple Regression

Contents
Regression through the Origin
Functional forms
Scaling and Units of measurement

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 84 / 136
Regression through the Origin

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 85 / 136
Regression through the Origin

Given
Yi = β0 + β1 Xi + ui (1)
if β0 = 0, then we have a regression through the origin

Yi = β1 Xi + ui (2)

The sample counterpart of equation (2) is

Yi = βb1 Xi + ubi (3)

Making ubi subject in equation (2), we get:

ubi = Yi − βb1 Xi (4)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 86 / 136
Derivation of the expression for βb1
Squaring and summing on both sides of equation (4), we have:
X X 2
ubi 2 = Yi − βb1 Xi (5)
P 2
Differentiating ubi wrt βb1 we obtain:

∂ ubi 2
P X  X 
=2 Yi − βb1 Xi (−Xi ) = −2 Yi − βb1 Xi (Xi ) (6)
∂ βb1
Setting the partial derivative in equation (6) equal to zero above yield the
following FOC (also called normal equation)

X 
−2 Yi − βb1 Xi (Xi ) = 0
X X
⇒ Yi Xi = βb1 Xi 2 (7)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 87 / 136
Derivation of the expression for βb1

Solving for βb1 in equation (8) yields:

P
Xi Yi
β1 = P 2
b (8a)
Xi
X Xi
= di Yi where di = P 2 (8b)
Xi

Important:
In the expression for βb1 , we use the RAW sums of squares and
cross-products

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 88 / 136
Regression through the Origin

Variance of βb1
We know that
X
βb1 = di Yi

Under the classical assumptions, the variance of βb1 is

X
Var (βb1 ) = σ 2 di2
X  Xi 2
2
=σ P 2
X
i
σ2

1
= σ2 P 2 = P 2
Xi Xi

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 89 / 136
Regression through the Origin
Property: the sum of the residuals is not equal to zero
X
ubi 6= 0
Proof (byP
contradiction)
Suppose ubi = 0, then summing on both sides of
Yi = βb1 Xi + ubi
gives
X X X X
Yi = βb1
Xi + ubi = βb1 Xi
P
Yi
∴ βb1 = P
Xi
This contradicts the fact that
P
Xi Yi
β1 = P 2
b
Xi
X
∴ ubi 6= 0
Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 90 / 136
Regression through the Origin
Property: r 2 can be negative
Proof
Recall:
ESS
r2 =
TSS
RSS
=1−
TSS
P 2
ubi
=1− P 2
yi
P 2 b2 P 2
Y − β1 Xi
= 1 − Pi 2
Yi2 − nY

If
X 2 RSS
βb12 Xi2 < nY ⇒ RSS > TSS ⇒ > 1 ⇒ r2 < 0
TSS
Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 91 / 136
Regression through the Origin

Caution:
Always use the conventional model with an intercept
If the intercept is included included in the model but turns out
insignificant(ie equal to zero), we have a regression through the origin
If infact the model has an intercept but we use a regression through
the origin, we would be committing a specification error

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 92 / 136
Functional Form

Types of models
Log-linear model
Semi-log model
Reciprocal model
Logarithmic reciprocal model

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 93 / 136
Log-linear model
Consider the following exponential model

Yi = β0 Xiβ1 e ui

Taking natural logs on both sides, we have:

lnYi = α + β1 lnXi + ui where α = lnβ0

If the assumptions of the CLRM are fulfilled, the parameters can be

estimated using OLS
The slope coefficient of the log-linear model (β1 ) measures the
elasticity of Y wrt X
Recall that elasticity is computed as:

dY /Y dY X
=
dX /X dX Y

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 94 / 136
Log-linear model

Now from the model lnYi = α + β1 lnXi + ui , we have:

dlnYi
β1 =
dlnXi
Recall that
dlnY 1 dY
= ⇒ dlnY =
dY Y Y
dlnX 1 dX
= ⇒ dlnX =
dX X X
Therefore    
dlnYi dY dX dY X
β1 = = / =
dlnXi Y X dX Y
The log-linear model is called a constant elasticity model because it
assumes that the elasticity coefficient is constant throughout

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 95 / 136
Log linear Model

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 96 / 136
Semi-log model

Log-lin model
A log-lin model has the form

lnYi = β0 + β1 Xi + ui

Using differentiation, we get

dlnYi (dY /Y ) dY 1
β1 = = =
dXi dX dX Y

That is, the slope coefficient of the log-lin model (β1 ) measures
relative (proportional) change in Y for a given absolute change in the
value of X
In a log-lin model, β1 must be multiplied by 100 (to get growth rate)
before being interpreted

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 97 / 136
Log-lin Model

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 98 / 136
Semi-log model

Lin-Log model
A lin-log model has the form

Yi = β0 + β1 lnXi + ui

Using differentiation, we get

dYi dY dY
β1 = = = X
dlnXi (dX /X ) dX

That is, the slope coefficient of the lin-log model (β1 ) measures
absolute change in Y for a given relative (proportional) change in the
value of X
In a lin-log model, β1 must be multiplied by 0.01 (divided by 100)
before being interpreted

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 99 / 136
Lin log Model

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 100 / 136
Choice of functional form

Use the underlying theory

Coefficients of the chosen model should satisfy apriori expectations
Choose a model with a high R-squared

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 101 / 136
Multiple Regression

Interpretation of partial regression coefficients

OLS estimation
Frisch-Waugh-Lovell (FWL) Theorem
Variance of OLS estimators
Properities of OLS estimators
Maximum Likelihood Estimation
R-squared and adjusted R-squared

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 102 / 136
Interpretation of partial regression coefficients

Given the regression model

Yi = β0 + β1 X1i + β2 X2i + ui

how are the partial regression coefficients interpreted?

The partial regression coefficients are partial derivatives

∂E (Y |X1 , X2 ) ∂(β0 + β1 X1i + β2 X2i )

= = β1
∂X1 ∂X1
∂E (Y |X1 , X2 ) ∂(β0 + β1 X1i + β2 X2i )
= = β2
∂X2 ∂X2

Since the coefficients are partial derivatives, they are interpreted on a

ceteris paribus basis

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 103 / 136
Derivation of the expression for βb0 , βb1 and βb2

Recall: We want to estimate the PRF

Yi = E (Y |Xi ) + ui = β0 + β1 X1i + β2 X2i + ui (1)

on the basis of the SRF

Yi = Ybi + ui = βb0 + βb1 X1i + βb2 X2i + ubi (2)

Making ubi subject in equation (2), we get:

ubi = Yi − βb0 − βb1 X1i − βb2 X2i (3)

Squaring and summing on both sides of equation (3), we have:

X X 2
ubi 2 = Yi − βb0 − βb1 X1i − βb2 X2i (4)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 104 / 136
Derivation of the expression for βb0 , βb1 and βb2
ubi 2 partially wrt βb0 , βb1 and βb2 we obtain:
P
Differentiating

ubi 2
P
∂ X 
=2 Yi − βb0 − βb1 X1i − βb2 X2i (−1)
∂ βb0
X 
= −2 Yi − βb0 − βb1 X1i − βb2 X2i (5)
∂ ubi 2
P X 
=2 Yi − βb0 − βb1 X1i − βb2 X2i (−X1i )
∂ βb1
X 
= −2 Yi − βb0 − βb1 X1i − βb2 X2i (X1i ) (6)
∂ ubi 2
P X 
=2 Yi − β0 − β1 X1i − β2 X2i (−X2i )
b b b
∂ βb2
X 
= −2 Yi − βb0 − βb1 X1i − βb2 X2i (X2i ) (7)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 105 / 136
Derivation of the expression for βb0 , βb1 and βb2
Setting the partial derivatives in equations (5) and (6) equal to zero above
yield the following FOCs (also called normal equations)
X 
−2 Yi − βb0 − βb1 X1i − βb2 X2i = 0
X X X
⇒ Yi = nβb0 + βb1 X1i + βb2 X2i (8)

X 
−2 Yi − βb0 − βb1 X1i − βb2 X2i (X1i ) = 0
X X X X
⇒ Yi X1i = βb0 X1i + βb1 X1i 2 + βb2 X1i X2i (9)

X 
−2 Yi − βb0 − βb1 X1i − βb2 X2i (X2i ) = 0
X X X X
⇒ Yi X2i = βb0 X2i + βb1 X1i X2i + βb2 X2i2 (10)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 106 / 136
Derivation of the expression for βb0 , βb1 and βb2
Expressing equations (9) and (10) in deviation form, we get:

X X X
yi x1i = βb1 x1i2 + βb2 x1i x2i (11)
X X X
yi x2i = βb1 x1i x2i + βb2 x2i2 (12)

P 2 P
Multiplying equations (11) by x2i and (12) by x1i x2i and subtracting
the resultant equations, we obtain:
X X X X X X
yi x1i x2i2 = βb1 x1i2 x2i2 + βb2 x1i x2i x2i2 (13)
X X X 2 X X 
− yi x2i x1i x2i = βb1 x1i x2i + βb2 x1i x2i x2i2 (14)
X X X X  X X  X 2 
yi x1i x2i2 − yi x2i x1i x2i = βb1 x1i2 x2i2 − x1i x2i
(15)
Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 107 / 136
Derivation of the expression for βb0 , βb1 and βb2

Solving for βb1 in equation (15) yields:

P P 2 P P
yi x1i x2i − yi x2i x1i x2i
βb1 = (16)
x2i − ( x1i x2i )2
P 2P 2 P
x1i

Interchanging for x1i and x2i in equation (16) yields:

P P 2 P P
yi x2i x1i − yi x1i x1i x2i
β2 =
b (17)
x2i − ( x1i x2i )2
P 2P 2 P
x1i

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 108 / 136
Derivation of the expression for βb0 , βb1 and βb2

Substituting the value of βb1 and βb2 from (16) and (17) respectively into
(8) X X X
Yi = nβb0 + βb1 X1i + βb2 X2i

and solving for βb0 gives

βb0 = Y − βb1 X1 − βb2 X2 (18)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 109 / 136
Frisch-Waugh-Lovell (FWL) Theorem

Given the regression

Yi = Ybi + ui = βb0 + βb1 X1i + βb2 X2i + ubi

βb1 measures the effect of regressor X1 on the dependent variable Y

holding constant (partialling/ netting out) the effect of X2
Similarly, βb2 measures the effect of regressor X2 on the dependent
variable Y holding constant (partialling/ netting out) the effect of X1

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 110 / 136
Frisch-Waugh-Lovell (FWL) Theorem

Proof
Run a regression of Y on X2 to have

Yi = α
b0 + α
b1 X2i + vbi (1a)
or in deviation form
yi = α
b1 x2i + vbi (1b)

where
P
yi x2i
α
b1 = P 2 (2)
x2i

Important: The residual vbi is yi without x2 in it [That is,the remainder of

yi after netting (or partialling) out x2 from it]

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 111 / 136
Frisch-Waugh-Lovell (FWL) Theorem

Run a regression of X1 on X2 to have

X1i = δb0 + δb1 X2i + wbi (3a)

or in deviation form
x1i = δb1 x2i + wbi (3b)

where
P
x1i x2i
δ1 = P 2
b (4)
x2i

Important: The residual wbi is x1 without x2 in it [That is, the remainder

of x1 after netting (or partialling) out x2 from it]

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 112 / 136
Frisch-Waugh-Lovell (FWL) Theorem

Run a regression of vbi on wbi to have

vbi = λ
b1 wbi + zbi (5)

where
P
wbi vbi
λ1 = P 2
b (6)
wbi

Important: λ b1 measures the effect x1 on y holding constant (netting

or partialling out) the effect of x2
Goal: To show that
λ
b1 = βb1

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 113 / 136
Frisch-Waugh-Lovell (FWL) Theorem

Making substitutions vbi and wbi from equations (1b) and (3b) respectively
into equation (6), we obtain

P
wbi vbi
λ1 = P 2
b
wbi
P 
x1i − δb1 x2i (yi − α b1 x2i )
= P   2
x1i − δb1 x2i

b1 δb1 x2i2
P P P P
yi x1i − α b1 x1i x2i − δb1 yi x2i + α
= P 2 b2 P 2 P (7)
x1i + δ1 x2i − 2δb1 x1i x2i

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 114 / 136
Frisch-Waugh-Lovell (FWL) Theorem

Making substitutions αb1 and δb1 from equations (2) and (4) respectively
into equation (7), we obtain

y
P
i x1i − α 1
P
x 1i x2i − b1 P yi x2i + α
δ 1
b1 P x 2
δ
b b 2i
λ
b1 = P 2 b2 P 2 P
x +δ x2i − 2δ1 x1i x2i
b
1iP 1  X P 
X yi x2i x1i x2i X
λ1 =
b yi x1i − x1i x2i − yi x2i
x2 x2i2
P P
P   P 2i  X
yi x2i x x
+ P 2 P 1i 2 2i x2i2
x2i x2i
x1i x2i 2 X 2
P  P 
X
2 x1i x2i X
x1i + P 2 x2i − 2 P 2 x1i x2i (8)
x2i x2i

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 115 / 136
Frisch-Waugh-Lovell (FWL) Theorem

x2i2 , we
P
Multiplying the numerator and denominator of equation (8) by
get

X X X X X X
λ
b1 = yi x1i x2i2 − yi x2i x1i x2i − x1i x2i yi x2i
X X
+ yi x2i x1i x2i
X X X 2 X 2
x1i2 x2i2 + x1i x2i − 2 x1i x2i
P P 2 P P
yi x1i x2i − yi x2i x1i x2i
λ
b1 =
x2i − ( x1i x2i )2
P 2P 2 P
x1i
= βb1 (9)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 116 / 136
Variance of βb1

σ2 x2i2 σ2
P
var (βb1 ) = P =
x2i2 − ( x1i x2i )2 x1i2 1 − r12
2
P

x1i2
P P

Proof

βb1 = λ
b1
P
wbi vbi
= P 2
wbi
X wbi
= hi vbi where hi = P 2 (1)
wbi

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 117 / 136
Variance of βb1

hi has the following properties

X X  wbi  1 X 
hi = P 2 =P 2 x1i − δb1 x2i
wbi wbi
1  X X 
=P 2 x1i − δb1 x2i = 0 (i)
wbi
X X  wbi 2 1 1
2
hi = P 2 = P 2 = P 2
wbi wbi x1i − δb1 x2i
P 2
1 x
= 2
= P 2 P 2 2iP (ii)
x2i − ( x1i x2i )2
P
P 2 ( x1i x2i ) x1i
x1i − P 2
x2i

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 118 / 136
Variance of βb1

hi has the following properties

X X  wbi  1 X 
hi x1i = P 2 x1i = P 2 x1i − δb1 x2i x1i
wbi wbi
1 X 2 b X 
=P 2 x1i − δ1 x1i x2i
wbi
P 2 ( x1i x2i )2
P
x1i − P 2
x
= P 2i 2 = 1 (iii)
P 2 ( x1i x2i )
x1i − P 2
x2i

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 119 / 136
Variance of βb1

hi has the following properties

X X  wbi  1 X 
hi x2i = P 2 x 2i = P 2 x1i − δ
b x
1 2i x2i
wbi wbi
1 X X 
=P 2 x1i x2i − δb1 x2i2
wbi
X P  
1 x1i x2i X 2
=P 2 x1i x2i − P 2 x2i
wbi x2i
=0 (iv)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 120 / 136
Variance of βb1
from equation (1), we have that:
X
βb1 = hi vbi
X  
= hi yi − δb1 x2i
X X
= hi yi − δb1 hi x2i
X
= hi y i (2)

If all classical assumption are holding true, then:

  X 1
var βb1 = σ 2 hi2 = σ 2 P 2
wbi
P 2
σ 2 x2i2
P
2 x2i
=σ P 2 P 2 =P 2P 2
x2i − ( x1i x2i )2 x2i − ( x1i x2i )2
P P
x1i x1i
(3)
Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 121 / 136
Variance of βb1

Dividing
P 2 P the numerator and denominator on the RHS of equation (3) by
x1i 2
x2i and simplifying, we get:

σ2
x2 σ2
  P
var βb1 = P 1i = (4)
( x1i x2i )2
" #
( x1i x2i )2
P
1− P 2 P 2
P 2
x1i 1 − P 2 P 2
x1i x2i x1i x2i

( x1i x2i )2
P
Noting the that P 2 P 2 is the coefficient of determination in the
x1i x2i
regression of X1 on X2 , we re-express equation (4) as:
  σ2
var βb1 = P (5)
x1i2 1 − r12
2

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 122 / 136
Numerical Properties of OLS estimators

Given the SRF

Yi = Ybi + ui = βb0 + βb1 X1i + βb2 X2i + ubi

the OLS estimators have the following numerical properties

the regression line passes through the means of Y , X1 and X2
the mean of the estimated Ys is equal to the means of the actual Ys
the sum (and mean) of the residuals is equal to zero
the residuals are uncorrelated with X1 and X2
the errors are uncorrelated with the predicted Ys

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 123 / 136
Statistical Properties of OLS estimators

Given the SRF

Yi = Ybi + ui = βb0 + βb1 X1i + βb2 X2i + ubi

the OLS estimators have the following statistical properties

Linearity
Unbiasedness
Minimum variance
Consistency

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 124 / 136
Linearity of βb1

from equation (1), we have that:

X
βb1 = hi yi
X
= hi Yi

That is, βb1 is a linear estimator

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 125 / 136
Unbiasedness of βb1

E (βb1 ) = β1

X X
βb1 = hi yi = hi Yi
X
= hi (β0 + β1 X1i + β2 X2i + ui )
X X X X
= β0 hi + β1 hi X1i + β2 hi X2i + hi ui
X
= β1 + hi ui

P
βb1 = β1 + hi ui

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 126 / 136
Unbiasedness of OLS estimators
Taking conditional expectations gives
X
E (βb1 |Xi ) = β1 + E (hi ui |Xi )
X  wbi ui  X 1
= β1 + E P 2 |Xi = β1 + P 2 E (wbi ui |Xi )
wbi wbi
X 1   
= β1 + P 2 E x1i − δb1 x2i ui |Xi
wbi
X 1 h i
= β1 + P 2 E (x u |X
1i i i ) − δ
b1 E (x u |X
2i i i )
wbi
= β1

Now by the law of iterative expectation

h i
E (βb1 ) = Ex E (βb1 |Xi ) = Ex [β1 ] = β1

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 127 / 136
GAUSS-MARKOV (GM) THEOREM

Gauss-Markov Theorem
Given the assumptions of classical linear regression model, the least
squares estimators have the minimum variance in a class of all linear
and unbiased estimators
That is, least squares estimators are BLUE (Best among Linear
Unbiased Estimators).

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 128 / 136
GAUSS-MARKOV (GM) THEOREM

Proof

Recall that the OLS estimator for β1 is

X
βb1 = hi Yi

Let β˜1 =
P
ai Yi be another linear estimator where the weight wi is

ai = hi + bi

That is,
X X
β˜1 = ai yi = ai Yi
X
= ai (β0 + β1 X1i + β2 X2i + ui )
X X X X
= β0 ai + β1 ai X1i + β2 ai X2i + ai u i

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 129 / 136
GAUSS-MARKOV (GM) THEOREM
For β̃ to be unbiased [ie E (β̃) = β], the conditions are that:
X
ai = 0 (1)
X
ai X1i = 1 (2)
X
ai X2i = 0 (3)

In this case,
X X X X
β˜1 = β0 ai + β1 ai X1i + β2 ai X2i + ai ui
X
= β1 + ai ui
Taking expectations on both sides and using the assumption of
noncorrelation between the regressors and the errors, we have:
 
E β˜1 = β1

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 130 / 136
GAUSS-MARKOV (GM) THEOREM
Now

X X X
X
ai = 0 ⇒ (hi + bi ) =
hi + bi = 0
X X
⇒ bi = 0 since hi = 0

and
X X X X
ai X1i = 1 ⇒ (hi + bi )X1i = hi X1i + bi X1i = 1
X X
⇒ bi X1i = 0 since hi X1i = 1

and
X X X X
ai X2i = 0 ⇒ (hi + bi )X2i = hi X2i + bi X2i = 0
X X
⇒ bi X2i = 0 since hi X2i = 0

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 131 / 136
GAUSS-MARKOV (GM) THEOREM

and
X X X X X
ai2 = (hi + bi )2 = hi2 + bi2 + 2 hi bi
 
X X X wbi
⇒ hi2 + bi2 + 2 P 2 bi
wbi
P 
X X wbi bi
⇒ hi2 + bi2 + 2 P 2
wbi
P P !
X X x b
1i i − δ
b1 x b
2i i
⇒ hi2 + bi2 + 2 P 2
wbi
X X
⇒ hi2 + bi2

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 132 / 136
GAUSS-MARKOV (GM) THEOREM
Now, the variance of β˜1 is given by
h i2 h i2
var (β˜1 |Xi ) = E β˜1 − E (β˜1 |Xi )|Xi = E β˜1 − β1 |Xi
hX i2
=E ai ui |Xi

Using assumptions of homoskedasticity and no serial correlation, we have

that:
X X X
var (β˜1 |Xi ) = σ 2 ai2 = σ 2 hi2 + σ 2 bi2
X
= var (βb1 |Xi ) + σ 2 bi2

And so
X
var (β˜1 |Xi ) ≥ var (βb1 |Xi ) since σ 2 bi2 ≥ 0

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 133 / 136
Consistency

Exercise

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 134 / 136
Maximum Likelihood estimation

Exercise

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 135 / 136
R-squared vs Adjusted R-squared

Exercise

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: ESTIMATION 22nd July, 2021 136 / 136
 REGRESSION ANALYSIS: HYPOTHESIS TESTING

Innocent Makuta

University of Malawi
Department of Economics(DoE)
official email: [email protected]
personal email: [email protected]

30th July,2021

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 1 / 50

CONTENTS

Meaning of Hypothesis Testing

Errors in Hypothesis Testing
Approaches to Hypothesis Testing
Test of Individual Significance
Analysis of Variance
Test of Overall Significance
Test of Linear Restrictions
Test of Structural change

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 2 / 50

Meaning of Hypothesis Testing

Recall:
The Population regression function (PRF) of interest is:

Yi = β0 + β1 X1i + β2 X2i + ... + βk−1 Xk−1,i + ui

However, we estimate the PRF on the basis of a sample regression

function (SRF)

Yi = βb0 + βb1 X1i + βb2 X2i + ... + βbk−1 Xk−1,i + ubi

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 3 / 50

Meaning of Hypothesis Testing

Further recall:
Step 1 in econometric methodology involves stating hypotheses.
These hypotheses are statements of relationships that (may) exist
between or among variables in the PRF
Technically, hypotheses claim that the parameters βj in the PRF take
on certain values (hypothesized values) or satisfy certain conditions
(restrictions)
One example could be that β1 = 0. That is, variable X1 have no
effect (is not related to) the dependent variable Y
Another example could be that β2 = β3 . That is, the effect of
variable X2 on the dependent variable Y is equal to the effect of
variable X3 on the dependent variable Y

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 4 / 50

Meaning of Hypothesis Testing

Further recall:
The hypotheses stated above are called Null hypotheses
The null hypothesis is denoted H0
Step 4 involves estimating the numerical values of the βbj on the basis
of the sample regression function

The problem of Hypothesis Testing

Is/are the numerical values obtained from estimation (Step 4) compatible
with (or sufficiently close to) the hypothesized values (step 1) so that
we do not reject the null hypothesis?

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 5 / 50

Meaning of Hypothesis Testing

If the null hypothesis is rejected, the estimated value (βbj ) is said to

be statistically significant
That is, the estimated value is not compatible with ( materially or
substantially different from) the hypothesized value
In this case (of rejecting H0 ), we then adopt a counter hypothesis
called the alternative hypothesis
The alternative hypothesis is denoted H1
If we fail to reject the null hypothesis the estimated value (βbj ) is said
to be statistically insignificant
That is, the estimated value is compatible with ( sufficiently close to)
the hypothesized value

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 6 / 50

One-tailed vs Two-tailed Test

The null hypothesis

H0 : β 1 = 0
can have two possible alternative hypotheses, countering this null.
Either, a non-directional alternative (not equal to)

H1 : β1 6= 0

or one of the following directional alternatives

H1 : β 1 < 0

H1 : β 1 > 0

One-tailed tailed: When H1 is directional

Two-tailed tailed: When H1 is non-directional

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 7 / 50

Types of Hypotheses

Given the Population regression function (PRF) of interest:

Yi = β0 + β1 X1i + β2 X2i + ... + βk−1 Xk−1,i + ui

one can test the following hypotheses

Test of individual significance

H0 : βj = 0
H1 : βj 6= 0

Test of overall significance

H0 : β1 = β2 = ... = βk−1 = 0
H1 : Not all βs are equal to zero

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 8 / 50

Types of Hypotheses

Test of linear restrictions

Examples:two coefficients are equal

H0 : β2 = β3 or H0 : β2 − β3 = 0
H1 : β2 6= β3 H1 : β2 − β3 6= 0

or: two coefficients add up to one

H0 : β 1 + β 2 = 1
H1 : β1 + β2 6= 1

Test of parameter stability (structural breaks)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 9 / 50

Errors in Hypothesis Testing

Two types of errors can be committed in conducting hypothesis testing:

Type I error
Type II error

Type I error
This consists of rejecting a true or correct H0
The probability of committing Type I error is called the level of
significance and is denoted by α

Type II error
This consists of accepting a false or incorrect H0
The ability of a test to avoid committing type II error is called power
of a test.
Power = 1 − Pr (Type II error)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 10 / 50

Errors in Hypothesis Testing

Decision
Status of H0
Reject Fail to reject
H0 correct/true Type I error
H0 incorrect/false Type II error

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 11 / 50

Approaches to Hypothesis Testing

Question
How does one determine whether or not to reject a null hypothesis?

Answer
Confidence Interval (CI) approach
Test of Significance approach
p-value approach

Important
To use any of the approaches, one needs to know the probability
distribution followed by estimator.
Estimators must be standardised

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 12 / 50

Probability distribution of OLS estimators
Recall
The error term is normally distributed with a mean of zero and a
constant variance of σ 2
ui ∼ N(0, σ 2 )
OLS estimators (βb1 ,βb1 ,...,βbk−1 ) are normally distributed because they
are linear combinations of the error term
 
βbj ∼ N βj , var (βbj )
!
σ 2 1
2 2
P
where var (βbj ) = σ wi = P 2
xj 1 − Rj2
 2
σ
(n − k) is follows a chi-square χ2 distribution
b
σ2
 2 P 2
σ 2 2 ubi
(n − k) ∼ χn−k where σ
b
b =
σ2 n−k
Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 13 / 50
Probability distribution of OLS estimators

What distributions do the following quantities follow?

P 2 P 2
RSS = ubi = Yi − βb0 − βb1 X1i − βb2 X2i − ... − βbk−1 Xk−1,i
P 2 P b 2
ESS = yb = i β0 + βb1 X1i + βb2 X2i + ... + βbk−1 Xk−1,i
ESS
R2 =
TSS

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 14 / 50

Standardisation of OLS estimators
An OLS estimator βbj can be standardised as follows:

βbj − βj βbj − βj
Z= =v ! ∼ N(0, 1)
se(βbj ) u 2
u σ 1
tP
xj 1 − Rj2
2

However, σ 2 is unobservable. On replacing the unobservable σ 2 with

b2 , we
its unbiased estimator σ

βbj − βj βbj − βj
t= =v ! ∼ tn−k
se(βbj ) u 2
u σ 1
t Pb
xj 1 − Rj2
2

Therefore, we can use the t distribution to establish to test

hypotheses for βj
Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 15 / 50
Confidence Interval Approach

What is a Confidence Interval?

Given two positive numbers δ and α(0 < α < 1), a Confidence interval is
the random interval (βbj − δ, βbj + δ) which contains the true parameter βj
with a probability 1 − α

Pr (βbj − δ ≤ βj ≤ βbj + δ) = 1 − α

where
α is the level of significance
1 − α is the confidence coefficient
βbj − δ is the lower confidence limit
βbj + δ is the upper confidence limit

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 16 / 50

Confidence Interval Approach

Using the standardised variables, the CI for βbj can be expressed as

Pr −tα/2 ≤ t ≤ tα/2 = 1 − α
!
(βbj − βj )
Pr −tα/2 ≤ ≤ tα/2 = 1 − α
se(βbj )

Rearranging, we have the 100(1 − α)% CI as

 
Pr βbj − tα/2 se(βc
j ) ≤ βj ≤ βj + tα/2 se(βj ) = 1 − α
b b

Or, more compactly, as

βbj ± tα/2 se(βbj )

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 17 / 50

Confidence Interval Approach

The confidence interval (CI) is also called a region of acceptance

This is because if the hypothesised value of parameter βj fall in this
region, we accept (fail to reject) the null hypothesis
The region(s) outside the CI is (are) called region(s) of rejection or
critical region(s)
This is because if the hypothesised value of parameter βj fall in this
region, we reject the null hypothesis

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 18 / 50

Confidence Interval Approach

Suppose a researcher is estimating a regression

Yi = β1 + β2 Xi + ui

with the intention of testing H0 : β2 = 0.3 against H1 : β2 6= 0.3

She uses a sample of 10 observations and obtains the following:

βb2 = 0.5091, se(βb2 ) = 0.0357

Task: Construct a 59% confidence interval for β2 [Hint:tα/2 = 2.306]

Question: On the basis of the results, should the null hypothesis be
accepted or be rejected?

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 19 / 50

Confidence Interval Approach

Recall that the area under a PDF is 1

If α = 0.05 it means that the critical region is 5% of the total area
under the PDF
It also means that the confidence interval (or acceptance region)
takes 5% of the total area under the PDF
Under a two tailed test, say,

H0 : βj = 0
H1 : βj 6= 0

the critical region is equally spread on the two tails of the

α
distribution. If α = 0.05,then the area on each tail is = 0.025 [i.e.
2
2.5% in each tail]

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 20 / 50

Confidence Interval Approach

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 21 / 50

Confidence Interval Approach

In one tailed test, the entire critical region is on one side of the
distribution
Under a lower (or left) tailed test, say,

H0 : βj = 0
H1 : βj < 0

the critical region is on the lower (or left) tail of the distribution.
If α = 0.05,then the area on the lower (or left) tail is 5% of the entire
area under the PDF.
The (1 − α)100% confidence interval is given by

[βbj − tα se(βbj ), ∞)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 22 / 50

Confidence Interval Approach

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 23 / 50

Confidence Interval Approach

Under an upper (or right) tailed test, say,

H0 : βj = 0
H1 : βj > 0

the critical region is on the upper (or right) tail of the distribution.
If α = 0.05,then the area on the upper (or right) tail is 5% of the
entire area under the PDF
The (1 − α)100% confidence interval is given by

(−∞, βbj + tα se(βbj )]

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 24 / 50

Confidence Interval Approach

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 25 / 50

Test of Significance Approach

The key idea is to compare a test statistic obtained from estimation

with critical value(s) of a distribution.
The test statistic is also called a computed statistic
The test statistic must be standardised
Decision rule: Reject H0 if

|Statisticcomputed | > Statisticcritical

Critical value(s) of a distribution are boundaries/limits which if

passed/exceed by a test statistic, a null hypothesis should be rejected

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 26 / 50

p-value Approach

p-value is the lowest level of significance at which a null hypothesis

can be rejected.
In other words, it is the actual probability of obtaining a value of the
test statistic equal to or geater the computed value obtained given
the sample at hand
For a given sample, as the value of test statistic increases, the p-value
decreases
Decision rule: Reject H0 if

p-value < α

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 27 / 50

Test of Individual Significance: CI Approach

Procedure
1 Estimate the regression using OLS

Yi = βb0 + βb1 X1i + βb2 X2i + ... + βbk−1 Xk−1,i + ubi

2 Construct a 100(1 − α)% for βj

3 Decision:
Reject H0 if βj falls outside the CI ⇒ βbj is statistically significant
Do not reject H0 if βj falls within the CI⇒ βbj is statistically
insignificant

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 28 / 50

Test of Individual Significance: Test of Significance
Approach

1 Estimate the regression using OLS

Yi = βb0 + βb1 X1i + βb2 X2i + ... + βbk−1 Xk−1,i + ubi

2 Compute the test statistic

(βbj − βj )
tcomputed =
se(βbj )

where βj is the value hypothesised under H0

3 Obtain the critical test value at α level of significance (tα/2,(n−k)
4 Decision:
Reject H0 if |tcomputed | > tα/2,n−k
Do not reject H0 if |tcomputed | ≤ tα/2,n−k

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 29 / 50

Test of Individual Significance: Test of Significance
Approach

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 30 / 50

Analysis of Variance(ANOVA)

Analysis of variance is the decomposition of total variation of the

dependent variable Y

TSS = ESS + RSS

X X X
yi2 = ybi2 + ubi2
where
X X X X
ybi2 = βb1 yi x1i + βb2 yi x2i + ... + βbk−1 yi xk−1,i

Each component has its own degrees of freedom (df)

Component Degrees of freedom
TSS n−1
ESS k −1
RSS n−k

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 31 / 50

Analysis of Variance: ANOVA Table

SS
Source SS df MSS =
P 2 df
ybi
ybi2
P
ESS k −1
k −1
P b2 ubi2
RSS ui n−k
nP− k
yi2
yi2
P
TSS n−1
n−1
where
SS is sum of squares
df is degrees of freedom
MSS is Mean of sum of squares

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 32 / 50

Analysis of Variance

Question:How does one use the information in the ANOVA Table to

test hypotheses?
Answer:
Construct the following ratio
P 2
ESS/df yb /(k − 1)
F = = P i2
RSS/df ubi /(n − k)

Note that with the normality of the error term, the F ratio follows an
F distribution
Now, remember that the expected value of the denominator is
 2 
ubi
E σ2) = σ2
= E (b
n−k

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 33 / 50

Analysis of Variance

It can be shown that the expected value of the numerator is

!
ybi2
P  P P P
βb1 yi x1i + βb2 yi x2i + ... + βbk−1 yi xk−1,i
E =E
k −1 k −1
 P P P 
β1 yi x1i + β2 yi x2i + ... + βk−1 yi xk−1,i
= + σ2
k −1
If we assume that

β1 = β2 = ... = βk−1 = 0

then,

ybi2
P 
E = σ2
k −1

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 34 / 50

Analysis of Variance
Therefore, if all the slope parameters are assumed to be equal to zero
simultaneously,
β1 = β2 = ... = βk−1 = 0
then the ratio
ESS/df
F =
RSS/df
will tend to unity (one)
That is, an F-ratio close to one is evidence in favour of the condition
β1 = β2 = ... = βk−1 = 0
However, if not all β are equal to zero, then the F-ratio will get bigger
and bigger.
That is, an F-ratio substantially greater than one is evidence in
against the condition
β1 = β2 = ... = βk−1 = 0
Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 35 / 50
Test of Overall Significance

Given the PRF:

Yi = β0 + β1 X1i + β2 X2i + ... + βk−1 Xk−1,i + ui

the test of overall significance tests the null hypothesis that all the slope
parameters are equal to zero simultaneously

H0 : β1 = β2 = ... = βk−1 = 0

against the alternative

H1 : Not all βj s are equal zero (j = 1, 2, ..., k − 1)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 36 / 50

Test of Overall Significance: The F-test

Estimate the regression using OLS

Yi = βb0 + βb1 X1i + βb2 X2i + ... + βbk−1 Xk−1,i + ubi

Compute the test statistic (F-statistic)

ESS/df
F =
RSS/df

Obtain the critical test value at α level of significance Fα/2,k−1,n−k )

Decision:
Reject H0 if |Fcomputed | > Fcritical ⇒ F-statistic is statistically
significant
Do not reject H0 if |Fcomputed | ≤ Fcritical ⇒ F-statistic is statistically
insignificant

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 37 / 50

The Relationship between R 2 and F
Recall: We have computed the F ratio in term of ESS and RSS
The F ratio can also be computed in terms of R 2
Recall:
   
ESS/(k − 1) n − k ESS n−k ESS
F = = =
RSS/(n − k) k − 1 RSS k − 1 TSS − ESS
Dividing the numerator and the denominator by TSS, we get
R2 R 2 /(k − 1)
   
n−k ESS/TSS n−k
F = = =
k − 1 1 − ESS/TSS k − 1 1 − R2 (1 − R 2 )/(n − k)
With this expression, it is easy to see that testing the null

H0 : β1 = β2 = ... = βk−1 = 0

is equivalent to testing the null

H0 : R 2 = 0
Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 38 / 50
Test of Linear Restrictions

Given the PRF

Yi = β0 + β1 X1i + β2 X2i + ... + βk−1 Xk−1,i + ui

one may be interested in testing various restrictions

For example, one may want to test the restriction that the sum of β1
and β2 is equal to a certain constant, say, 1

H0 : β 1 + β 2 = 1

Or, one may want to test the restriction that X1 and X2 have no
effects on Y
H0 : β 1 = β 2 = 0

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 39 / 50

Test of Linear Restrictions

To test any restriction, you begin with the original model. This model
is called the Unrestricted model
Next, you incorporate the restriction on the unrestricted model to
form the restricted model

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 40 / 50

Test of Linear Restrictions: Formulating Restricted Models
Given the PRF

lnYi = β0 + β1 lnX1i + β2 lnX2i + ... + βk−1 lnXk−1,i + ui

suppose, you want to test restriction that

H0 : β 1 + β 2 = 1

This restriction can rewritten as

H0 : β 1 = 1 − β 2

Incorporating this restriction onto the unrestricted model, we obtain

the restricted model as follows
     
Yi X2i Xk−1,i
ln = α + β2 ln + ... + βk−1 ln + vi
X1i X1i X1i

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 41 / 50

Test of Linear Restrictions: Formulating Restricted Models

Given the PRF

Yi = β0 + β1 X1i + β2 X2i + β3 X3i + ... + βk−1 Xk−1,i + ui

Suppose you may want to test the restriction that X1 and X2 have no
effects on Y
H0 : β 1 = β 2 = 0
Incorporating this restriction onto the unrestricted model, we obtain
the restricted model as follows

Yi = β0 + β3 X3i + ... + βk−1 Xk−1,i + ui

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 42 / 50

Test of Linear Restrictions

If the restriction is correct, then the R 2 from the unrestricted model

2 will be very different from the R 2 from the restricted model R 2
RUR R
Similarly, if the restriction is correct, then the RSS from the
unrestricted model RSSUR will be very different from the RSS from
the restricted model RSSR
However, if the restriction is incorrect/false, the R 2 or the RSS from
the unrestricted and the restricted model will not differ materially

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 43 / 50

Test of Linear Restrictions: The F-test
2 (or
Estimate the unrestricted model using OLS and obtain the RUR
RSSUR )
Estimate the restricted model using OLS and obtain the RR2 (or
RSSR )
Compute the test statistic (F-statistic)
2 − R 2 )/m
(RUR R
F = 2 /(n − k)
1 − RUR
or
(RSSR − RSSUR )/m
F =
RSSUR /(n − k)
Obtain the critical test value at α level of significance Fα/2,k−1,n−k )
Decision:
Reject H0 if |Fcomputed | > Fcritical ⇒ F-statistic is statistically
significant
Do not reject H0 if |Fcomputed | ≤ Fcritical ⇒ F-statistic is statistically
insignificant
Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 44 / 50
Test of Structural change

Over time ( indeed over space), there may a structural change in the
relationship between the regressand and the regressor(s)
Structural change means that the value of the parameters of the
model do not remain the same (i.e the parameter have not been
stable) over time (or spatially)
To test for a structural change, we use the Chow Test

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 45 / 50

Test of Structural change: Chow Test

Given the regression for the whole period

Yt = β0 + β1 X1i + β2 X2i + ... + βk−1 Xk−1,i + ui

One can partition our sample into two periods, one before the
structural change and the other after the structural change.
Let the model for the period before the structural change be

Yt = α0 + α1 X1i + α2 X2i + ... + αk−1 Xk−1,i + vi

Let the model for the period after the structural change be

Yt = δ0 + δ1 X1i + δ2 X2i + ... + δk−1 Xk−1,i + wi

the first model assumes no structural change while the latter two
model assume structural change

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 46 / 50

Test of Structural change: Chow Test

First model is the restricted model since it imposes the structural

breaks
the last two models together give us the unrestricted model
The null hypothesis of no structural change can be stated as
α0 = δ0 = β0 and α1 = δ1 = β1 ... and αk−1 = δk−1 = βk−1

Assumptions
The errors ui , vi and wi are normally distributed
the two sub-samples are distributed independently. Technically, the
errors for the sub-samples vi and wi are distributed independently

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 47 / 50

Test of Structural change: Chow Test

Estimate the restricted model

Yt = β0 + β1 X1i + β2 X2i + ... + βk−1 Xk−1,i + ui

and obtain RSSR (with df = n1 + n2 − k)

Estimate the model

Yt = α0 + α1 X1i + α2 X2i + ... + αk−1 Xk−1,i + vi

and obtain RSS1 (with df = n1 − k)

Estimate the model

Yt = δ0 + δ1 X1i + δ2 X2i + ... + δk−1 Xk−1,i + wi

and obtain RSS2 (with df = n2 − k)

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 48 / 50

Test of Structural change: Chow Test

Since the two sub-samples are distributed independently, add RSS1

and RSS2 to obtain the unrestricted RSS

RSSUR = RSS1 + RSS2

(with df = n1 + n2 − 2k)
Compute the test statistic

(RSSR − RSSUR )/k

F =
RSSUR /(n1 + n2 − 2k)

Decision:
Reject H0 if |Fcomputed | > Fcritical ⇒ There has been a structural
break
Do not reject H0 if |Fcomputed | ≤ Fcritical ⇒ No structural break

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 49 / 50

Test of Structural change: Chow Test

Limitations of the Chow Test

The assumptions underlying the test must be satisfied
The Chow test does not pin down the source of the structural break
(whether the difference is on the account of the intercept or the
slopes)
Sensitive to the choice of the structural break point

Innocent Makuta (UNIMA) REGRESSION ANALYSIS: HYPOTHESIS TESTING 30th July,2021 50 / 50