See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/269113492

APPLIED ECONOMETRICS With Eviews Applications

Book · November 2013

CITATIONS READS

3 59,890

2 authors, including:

Ali Göksu
University of Eurasia
30 PUBLICATIONS   126 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Working on organization of the International Conference on Economic and Social Studies 2017 (ICESoS 2017) View project

CALL FOR PAPERS: International Journal of Applied Statistics and Econometrics View project

All content following this page was uploaded by Ali Göksu on 10 October 2015.

The user has requested enhancement of the downloaded file.

APPLIED ECONOMETRICS
With Eviews Applications

IBU Publications
2 Applied Econometrics with Eviews Applications

APPLIED ECONOMETRICS
With Eviews Applications

Authors:
Uğur ERGÜN Ali GÖKSU
ugur.ergun@ibu.edu.ba ali.goksu@ibu.edu.ba

Publisher:
International Burch University

Editor in Chief:
Prof.Dr. Mehmet Uzunoğlu

Reviewed by:
Assoc. Prof. Dr. Ercan BALDEMİR, Muğla Sıtkı Koçman University
Assoc. Prof. Dr. Safet KOZAREVIĆ, Faculty of Economics, University of Tuzla

DTP & Design:

Nihad Obralić

DTP and Prepress:

International Burch University

Printed by:

Circulation: 500 copies

Place of Publication: Sarajevo
Copyright: International Burch University, 2013
International Burch University Publication No: 23

Reproduction of this Publication for educational or other non-commercial purposes is authorized

without prior permission from the copyright holder. Reproduction for resale or other commercial
purposes prohibited without prior written permission of the copyright holder.

Disclaimer: While every effort has been made to ensure the accuracy of the information, contained in
this publication, International Burch University will not assume liability for writing and any use made of
the proceedings, and the presentation of the participating organizations concerning the legal status of
any country, territory, or area, or of its authorities, or concerning the delimitation of its frontiers or
boundaries.
-------------------------------------------------
CIP - Katalogizacija u publikaciji
Nacionalna i univerzitetska biblioteka
Bosne i Hercegovine, Sarajevo

330.43(075.8)

ERGUN, Uğur
Applied econometrics : with eviews applications
/ Uğur Ergun, Ali Göksu. - Sarajevo :
International Burch University, 2013. - 272 str. :
graf. prikazi ; 24 cm. - (IBU Publications)

Bibliografija: str. 267-271.

ISBN 978-9958-834-29-5
1. Göksu, Ali
COBISS.BH-ID 20792838
--------------------------------------------------------
 3

Uğur ERGUN Ali GÖKSU

APPLIED ECONOMETRICS
With Eviews Applications

IBU Publications
Sarajevo, 2013
4 Applied Econometrics with Eviews Applications
 5

PREFACE
This book is designed as auxiliary source for the students who are
taking Applied Econometrics course. It is intended to clarify basic
econometrics methods with examples especially for Finance
students.

In this book, two main dimensions have been configured. In the first
chapters some basic information regarding scientific research is given
in order to polish student’s ability to understand following chapters
better. Following chapters are organized to provide more detailed
information about some specific methods and their applications in
finance by examples.

15th November 2013

Uğur ERGÜN Ali GÖKSU

6 Applied Econometrics with Eviews Applications
 7

LIST OF ABBREVIATIONS

2SLS : Two-Stage Least Squares

ACF : Autocorrelation Function
ADF : Augmented Dickey-Fuller
AIC : Akaike Information Criterion
ANOVA : Analysis of Variance
AR : Autoregressive
ARDL : Autoregressive Distributed Lag
ARIMA : Autoregressive Integrated Moving Average
ARMA : Autoregressive Moving Average
ARCH : Autoregressive Conditional Heteroscedasticity
BLUE : Best Linear Unbiased Estimator
BUE : Best Unbiased Estimator
CAPM : Capital Asset Pricing Model
CLRM : Classical Linear Regression Model
CUSUM : Cumulative Sum of Control Chart
DF : Dickey-Fuller
DW : Durbin Watson
ECM : Error Correction Model
EGARCH : Exponential GARCH
EGLS : Estimated Generalized Least Squares
ESS : Explained Sum of Squares
EU : European Union
FDI : Foreign Direct Investment
FDL : Distributed Lag Model
GARCH : Generalozed Autoregressive Conditional
Heteroscedasticty
8 Applied Econometrics with Eviews Applications

GDP : Gross Domestic Product

GLS : Generalized Least Squares
GMM : Generalized Method of Moment
GQ : Goldfeld-Quant
IS-LM : Investment Saving–Liquidity Preference Money
Supply
ISE : Istanbul Stock Exchange
IV : Instrumental Variables
LM : Langrage Multiplier
MA : Moving Average
NCLRM : Normal Classical Linear Regression Model
OLS : Ordinary least Squares
PRF : Population Regression Function
P-P : Probability-Probablity
PP : Philip-Peron
QQ : Quantile-Quantile
RESET : Regression Specification Errror Test
RSS : Residual Sum of Squares
SEM : Structural Equation Modelling
SER : Standard Errror of the Regression
SIC : Schwarz Information Criterion
SSR : Sum of Squared Residuals
TARCH : Threshold ARCH
TSS : Total Sum of Squares
VAR : Vector Autoregressive
VECM : Vector Error Correction Model
VIF : Variance Inflation Factor
WLS : Weighted Least Squares
 9

TABLE OF CONTENTS

PREFACE ......................................................................................................... 5
LIST OF ABBREVIATIONS ................................................................................. 7
TABLE OF CONTENTS ...................................................................................... 9

CHAPTER 1
SCIENTIFIC RESEARCH
DEFINITION ................................................................................................... 21
AIMS OF THE SCIENTIFIC RESEARCH ............................................................ 21
RESEARCH METHODS & METHODOLOGY .................................................... 22
Research Methods.................................................................................... 22
Research Methodology ............................................................................ 22
The Scientific Method .............................................................................. 23
CRITERIA OF A GOOD RESEARCH.................................................................. 23
POSSIBLE USERS OF RESEARCH OUTCOME .................................................. 25
RESEARCH TYPES .......................................................................................... 25
Based on the Research Purpose ............................................................... 25
Descriptive Research ............................................................................ 25
Exploratory Research ............................................................................ 26
10 Applied Econometrics with Eviews Applications

Analytical Explanatory Research .......................................................... 28

Predictive Research .............................................................................. 28
Based on the Research Process ................................................................ 29
Qualitative Research ............................................................................ 30
Quantitative Research .......................................................................... 31
Based on the Research Outcome ............................................................. 31
Applied Research .................................................................................. 31
Basic Research ...................................................................................... 32
Based on the Research Logic .................................................................... 33
Deductive Research .............................................................................. 33
Inductive Research................................................................................ 33
CONCEPTUAL RESEARCH .............................................................................. 34
EMPIRICAL RESEARCH .................................................................................. 34
HISTORICAL RESEARCH ................................................................................. 35

CHAPTER 2
MODELS
ECONOMIC MODEL ...................................................................................... 37
TYPES OF ECONOMIC MODELS..................................................................... 38
Visual Models ........................................................................................... 38
Mathematical Models .............................................................................. 38
Empirical Models ...................................................................................... 38
Static Model ............................................................................................. 39
Dynamic Models ....................................................................................... 39
WHICH THEORY IS APPROPRIATE? ............................................................... 40
ECONOMETRIC MODELS .............................................................................. 40
Attributes of a Good Econometric Model ................................................ 42
Structural Equation Modeling (SEM) ........................................................ 42
 11

The Reduced Form Model ........................................................................ 42

Stochastic Models .................................................................................... 43
Deterministic Models ............................................................................... 43
Analytical Framework ............................................................................... 45
Methodological Framework ..................................................................... 46
Research Process Flow Chart ................................................................... 47
Regression Analysis Flow Chart ............................................................... 48

CHAPTER 3
DATA
QUALITATIVE DATA ...................................................................................... 49
In-Depth Interviews .................................................................................. 49
Direct Observation ................................................................................... 50
Written Documents .................................................................................. 50
Qualitative Methods................................................................................. 50
Participant Observation ........................................................................... 51
Direct Observation ................................................................................... 51
Unstructured Interviewing ....................................................................... 52
Case Studies.............................................................................................. 52
QUANTITATIVE VERSUS QUALITATIVE DATA ........................................... 53
WHY ARE QUANTITATIVE AND QUALITATIVE DATA IMPORTANT? .............. 53
QUANTITATIVE DATA TYPES ......................................................................... 53
Primary Data............................................................................................. 53
Secondary Data ........................................................................................ 54
Experimental Data .................................................................................... 54
Observational data ................................................................................... 54
Time Series ............................................................................................... 55
Panel Data ................................................................................................ 56
12 Applied Econometrics with Eviews Applications

Cross Sectional Data ................................................................................. 57

ARTIFICIAL EXPLANATORY VARIABLES ......................................................... 58
The Chow Breakpoint Test ....................................................................... 58
DATA TRANSFORMATION............................................................................. 59
Log Transformation .................................................................................. 59
When or not to Log?................................................................................. 59
How to Choose? ....................................................................................... 60
Differencing .............................................................................................. 60
Percentage Change................................................................................... 61
Base Year .................................................................................................. 61

CHAPTER 4
REGRESSION
DEFINITION ................................................................................................... 63
SIMPLE REGRESSION MODELS...................................................................... 63
ESTIMATION OF THE LINEAR REGRESSION MODEL ..................................... 65
Ordinary Least Squares Estimation Method (OLS) ................................... 65
Multivariate Regression Model ................................................................ 68
CLASSICAL LINEAR REGRESSION MODEL (CLRM) ASSUMPTIONS ................ 70
STATISTICAL PROPERTIES OF OLS ................................................................. 72
Linearity .................................................................................................... 72
Unbiasedness ........................................................................................... 75
Efficiency .................................................................................................. 75
Normality .................................................................................................. 75
GAUSS-MARKOV THEOREM ......................................................................... 77
COEFFICIENT OF DETERMINATION, GOODNESS OF FIT ............................... 78
 13

CHAPTER 5
CLASSICAL LINEAR REGRESSION MODEL (CLRM)
LINEARITY ..................................................................................................... 83
Modeling Nonlinearities (with OLS) ......................................................... 84
Testing Linearity ....................................................................................... 84
The Ways to Correct Linearity Problem: .................................................. 88
HETEROSCEDASTICITY .................................................................................. 89
Consequences .......................................................................................... 90
Detection .................................................................................................. 90
Which Test to Apply? ............................................................................. 101
The Ways to Correct Heteroscedasticity Problem ................................. 102
AUTOCORRELATION ................................................................................... 102
Consequences of Autocorrelation in the Residuals ............................... 103
First Order Autocorrelation .................................................................... 103
Detection ................................................................................................ 104
The Ways to Correct Autocorrelation Problem ...................................... 111
MULTICOLLINEARITY .................................................................................. 111
Consequences ........................................................................................ 111
Causes of Multicollinearity ..................................................................... 112
Detection ................................................................................................ 113
The Ways to Handle Multicollinearity .................................................... 114
EXOGENEITY ............................................................................................... 116
NORMALITY ................................................................................................ 117
CONSISTENCY ............................................................................................. 121
14 Applied Econometrics with Eviews Applications

CHAPTER 6
HYPOTHESIS TESTING
DEFINITIONS ............................................................................................... 123
F-Test: A JOINT TEST OF SIGNIFICANCE ...................................................... 131

CHAPTER 7
SPECIFICATIONS
CHOOSING INDEPENDENT VARIABLES ....................................................... 135
Omitting a Relevant Variable ................................................................. 136
Detection ............................................................................................ 137
Wald Test (Coefficient Restrictions) ................................................... 137
Omitted Variables Test ....................................................................... 140
Solution ............................................................................................... 142
Including an Irrelevant Variable in the Regression equation ................. 142
How to Choose Correct Variables? ..................................................... 143
Redundant Variables Test................................................................... 143
CHOOSING A FUNCTIONAL FORM .............................................................. 145
Functional Forms: ................................................................................... 145
The Log-log Regression Model (Double Log) ...................................... 145
Cobb-Douglas Production Function .................................................... 145
Lin-Log Model (semi log) .................................................................... 146
Log-ln Model ....................................................................................... 146
Quadratic Forms ................................................................................. 146
Inverse Form ....................................................................................... 146
Intercept Dummy Independent Variable ............................................ 146
Slope Dummy Independent Variable .................................................. 147
Lags..................................................................................................... 147
Mixed Functional Forms ......................................................................... 147
 15

Consequences of Choosing the Wrong Functional Form ....................... 148

Detection and Correction of Functional Form Specification Errors ....... 148
Maintained Hypothesis Methodology ................................................ 148
Theory/Testing Methodology ............................................................. 148
Testing-Down Approach ..................................................................... 149
General Functional Forms ...................................................................... 149
Testing-up Approach .......................................................................... 149
Other Tests for Functional Form ......................................................... 150
SPECIFICATION TESTS ................................................................................. 150
Residual Tests ......................................................................................... 150
Correlograms and Q-Statistics ............................................................ 150
Correlograms of Squared Residuals.................................................... 152
Histogram and Normality test ............................................................ 152
Serial Correlation LM Test .................................................................. 152
ARCH-LM Test ..................................................................................... 153
Whites Heteroscedasticity Test .......................................................... 155

CHAPTER 8
PARAMETER STABILITY
CHOW BREAKPOINT TEST FOR STRUCTURAL BREAK ................................. 158
CHOW FORECAST TEST FOR STRUCTURAL BREAK...................................... 159
CUSUM TEST ............................................................................................... 160
CUSUM OF SQUARES TEST ......................................................................... 161
ONE STEP AHEAD FORECAST TEST ............................................................. 162
N STEP FORECAST TEST .............................................................................. 163
RECURSIVE COEFFICIENT ESTIMATES ......................................................... 164
WALD TEST FOR STRUCTURAL CHANGE ..................................................... 165
16 Applied Econometrics with Eviews Applications

CHAPTER 9
INTERPRETING THE LINEAR REGRESSION MODELS
SIMPLE LINEAR REGRESSION MODEL ......................................................... 167
MULTIPLE LINEAR REGRESSION MODEL .................................................... 168
REGRESSOR SELECTION .............................................................................. 174

CHAPTER 10
ENDOGENEITY
SOURCES OF ENDOGENEITY ....................................................................... 176
Autocorrelation with Lagged Dependent Variable ................................ 176
Omitted Variable Bias............................................................................. 177
Measurement Error ................................................................................ 177
Simultaneity............................................................................................ 178
ENDOGENEITY TEST .................................................................................... 179
SOLUTION ................................................................................................... 181
Ad hoc Approaches................................................................................. 181
Instrumental Variables Estimation ......................................................... 181
Two Stage Least Squares ........................................................................ 185
Weak Instrument.................................................................................... 186
Generalized Method of Moment ........................................................... 186

CHAPTER 11
REGRESSIONS WITH DUMMY VARIABLES
A DUMMY INDEPENDENT VARIABLE .......................................................... 189
INTERPRETATION........................................................................................ 190
QUALITATIVE VARIABLE WITH MORE THAN TWO CATEGORIES
(POLYTOMOUS FACTORS) .......................................................................... 191
CONSTRUCTING INTERACTION REGRESSORS ............................................. 193
 17

DUMMY DEPENDENT VARIABLE................................................................. 195

TYPES OF VARIABLES .................................................................................. 195
Continuous or Quantitative Variables .................................................... 195
Interval - Scale Variables .................................................................... 195
Continuous Ordinal Variables ............................................................. 196
Ratio - Scale Variables ........................................................................ 196
Qualitative or Discrete Variables............................................................ 196
Nominal Variables .............................................................................. 196
Ordinal Variables ................................................................................ 196
Categorical Variables.......................................................................... 197
Preference Variables........................................................................... 197
Multiple Response Variables .............................................................. 198

CHAPTER 12
TIME SERIES
STATIC MODELS .......................................................................................... 199
DYNAMIC MODELS ..................................................................................... 200
TIME SERIES ................................................................................................ 200
Time Series Regression with Lag ............................................................ 200
Lag Selection........................................................................................... 202
Finite Distributed Lag Models ................................................................ 205
Autoregressive Modeling (AR) ............................................................... 205
Stationarity ............................................................................................. 206
Random Walk ......................................................................................... 208
Seasonality ............................................................................................. 210
Steps for Testing in the AR(p) with Deterministic Tren .......................... 211
Lagged Dependent Variables.................................................................. 213
Moving Average ..................................................................................... 214
18 Applied Econometrics with Eviews Applications

Autocorrelation Function ....................................................................... 215

Autoregressive Moving Average (ARMA) ............................................... 217
ARIMA (p,q) ............................................................................................ 218
Predicting with ARMA Models ............................................................... 219
Autoregressive Distributed Lag (ARDL) Model ....................................... 219
If Dependent and Independent Series are Stationary......................... 223
If dependent and independent series are non-stationary (Spurious
regression) .......................................................................................... 225
COINTEGRATION ........................................................................................ 225
Bivariate Cointegration (Engle-Granger Cointegration) ......................... 225
Error Correction Model (ECM) ............................................................ 233
Diagnostic Checking ........................................................................... 237
Dependent and Independent variables have unit roots but are not
cointegrated ....................................................................................... 238
Multivariate Cointegration (Johansen, 1988, 1991) .............................. 240
Vector Error Correction Model ........................................................... 242
Diagnostic Checking ........................................................................... 244
At first, whether there is serial correlation is checked as follows; ..... 244
Potential Pitfalls of the Cointegration Method ...................................... 247
DIAGNOSTIC CHECKING .............................................................................. 247

CHAPTER 13
VOLATILITY
ARCH MODEL .............................................................................................. 249
DIAGNOSTIC CHECKING .............................................................................. 256
VOLATILITY SPILLOVER, GARCH (1,1) MODEL ............................................ 258
 19

CHAPTER 14
GRANGER CAUSALITY
CAUSALITY IN BOTH DIRECTIONS ............................................................... 265
GRANGER CAUSALITY IN COINTEGRATED VARIABLES................................ 265

CHAPTER 15
VECTOR AUTOREGRESSION MODEL (VAR)
DEFINITION ................................................................................................. 267
LAG LENGTH SELECTION ............................................................................. 270
FORECASTING WITH “VAR” ........................................................................ 270
VECTOR AUTOREGRESSIONS WITH COINTEGRATED VARIABLES ............... 272

REFERENCES ............................................................................................... 277

20 Applied Econometrics with Eviews Applications
 Scientific Research 21

CHAPTER 1

SCIENTIFIC RESEARCH

DEFINITION
Scientific research is a systematic, controlled, empirical, amoral,
public and critical investigation of natural phenomena guided by
theory and hypotheses about the presumed relations among such
phenomena1.

AIMS OF THE SCIENTIFIC RESEARCH

The main objective of a research is to find out answers to questions
through the application of scientific procedures. Furthermore, each
research has its own specific purpose.

Research objectives may include but are not limited to following

groupings:

a) To review and synthesize available knowledge

b) To test a hypothesis of a causal relationship between
variables
c) To investigate some existing situation or problem
1
Lee, H.B & Kerlinger, F.N(2000)
22 Applied Econometrics with Eviews Applications

d) To explore possible solutions to a specific issue

e) To analyze general issues
f) To offer a new procedure
g) To explain a new scientific idea
h) To generate new knowledge, etc.

RESEARCH METHODS & METHODOLOGY

Research Methods

Research methods include all the methods or techniques that are

used for performing a research process. Actually, research method is
a part of the research methodology that helps in constructing it.

Research methods can be classified as follows:

 Methods related to the data collection;

 Statistical techniques used to analyze research questions;
 Methods used to evaluate the accuracy of the results
obtained through statistical techniques.

Research Methodology

Research methodology is a way of solving systematically the research

problem. It is actually a science on how a research is done
scientifically. Researchers also need to comprehend related theories
and to understand the assumptions underlying various techniques
which are applicable for the research focus and questions. Certain
techniques and procedures are applicable only to specified problems.
Thus, a researcher should design his/her research based on the
problem which he/she tries to solve by applying appropriate
statistical techniques.

A research methodology should answer the following questions;

 Why a research study has been undertaken? What is the

motivation of the research?
 Scientific Research 23

 What is the problem statement? How is it defined?

 What data have been used and how have the data been
collected?
 What statistical or econometric method has been conducted?
 What is the contribution of the research?
 What is the significance of the research?
 What are the implications?

The Scientific Method

The scientific method relies on empirical evidence, utilizes relevant

concepts, is committed to only objective considerations, presupposes
ethical neutrality, aims at nothing but making only adequate and
correct statements about population objects and results into
probabilistic predictions. Its methodology is available to all concerned
for critical scrutiny and for use in testing the conclusions through
replication, aiming at formulating more general axioms or what can
be termed as scientific theories.

CRITERIA OF A GOOD RESEARCH

One expects scientific research to satisfy the following criteria2:

 The purpose of the research should be clearly defined

and common concepts should be used.
 The research procedure used should be described in
sufficient detail to permit another researcher to repeat
the research for further advancement, keeping the
continuity of what has already been attained.
 The procedural design of the research should be carefully
planned to yield results that are as the objective as
possible.
 The researcher should report with complete frankness,
explain all flaws in the procedural design and estimate
their effects on the findings.

2
Fox, J.H.(1958)
24 Applied Econometrics with Eviews Applications

 The analysis of data should be sufficiently adequate to

reveal its significance and the methods of analysis used
should be appropriate. The validity and reliability of the
data should be checked carefully.
 Conclusions should be confined to those parts justified by
the data of the research and limited to those for which
the data provide an adequate basis.
 Greater confidence in research is warranted if the
researcher is experienced, has a good reputation in doing
research and is a person of integrity.

In a more systematic way, the qualities of a good research can be

stated as follows3:

i- Systematic: It means that a research is structured

through following specified steps that are to be taken
in a specified sequence in accordance with the well-
defined set of rules. Systematic characteristic of the
research does not rule out creative thinking, but it
certainly does reject the use of guessing and intuition
in arriving at conclusions.

ii- Logical: This implies that research is guided by the

rules of logical reasoning and the logical process of
induction and deduction are of great value in carrying
out a research. Induction is the process of reasoning
from a part to the whole whereas deduction is the
process of reasoning from some premise to a
conclusion which follows from that very premise. In
fact, logical reasoning makes research more
meaningful in the context of decision making.

iii- Empirical: A research is related basically to one or

more aspects of a real situation and deals with
concrete data that provides a basis for external validity
to the research results.

3
Bellenger D.N & A. Greenberg, B.A. (1978)
 Scientific Research 25

iv- Replicable: This characteristic allows research results

to be verified by replicating the study and thereby
building a sound basis for making decisions.

POSSIBLE USERS OF RESEARCH OUTCOME

The researches are expected to produce usefull implications for the
policy makers, managers, investors, advisors or academicans as
follows;

 The Government – for making or/and developing their

policies;
 Managers, Investors, Advisors – use for profit maximization;
 Academics – use for conducting further studies;

RESEARCH TYPES4
Based on the Research Purpose

Descriptive Research

Descriptive Research includes surveys and fact-finding enquiries of

different kinds. The major purpose is generally to describe an event,
phenomenon, or a relationship. The term Ex post facto research is
used instead of descriptive research in social sciences. The researcher
only report what has happened or what is happening and has no
control over the variables. The researcher mostly seeks to measure
such items as frequency of shopping, preferences of people, or
similar variables in this research type. This type of research also
includes attempts by researchers to discover causes even if they
cannot control the variables. The methods of research utilized in
descriptive research are all survey methods, including comparative
and correlational methods.

4
Kothari, C.R. & Garg, G.D. (2004)
26 Applied Econometrics with Eviews Applications

Descriptive research is conducted to describe facts as they exist. It is

used to identify or obtain information for a specific issue. Descriptive
research takes further steps in exploring an issue than exploratory
research which is employed to clarify the structural issues. Therefore,
even in a descriptive study, you have to spend time in refining your
research questions and be specific about the phenomena you are
studying. In short, descriptive research may be defined as statistical
summarization of the collected data.

Example 1.1
- What are the feelings of Academics faced with redundancy?
- What kind of ethical behaviors do students expect at the
universities?

Exploratory Research5

The objective of an exploratory research is the development of

hypotheses rather than their testing, whereas formalized research
studies are those with substantial structure and specific hypotheses
to be tested. If there are very scant studies which are used to clarify
existing information, exploratory research is conducted. This kind of
research is also employed to explore patterns, scientific ideas or
hypotheses. In exploratory research, the main idea is to obtain close
insights and knowledge regarding the issue which is studied.

The exploratory research techniques include case studies,

observations and historical analysis. Quantitative and qualitative data
can be used. The researcher assesses the consistency of the results
which are obtained empirically with the existing theories and
concepts or develop them.

Exploratory research provides informative additional for further

research rather than giving answers to problems or issues.
Exploratory or qualitative research is used to obtain deep insights

5
Collis, J & Hussey, R. (2009).
 Scientific Research 27

into the behavior of few consumers, or to gain preliminary

information about the market.

Common exploratory research methods include depth interviews,

projective techniques, and focus groups. Researchers employ
exploratory research when little is known about the topic and
previous theories or ideas do not apply. For example, if you wanted
to study how to get students to use the computer lab in a college
environment, you might first have to do exploratory research to
figure out which students might need the lab and what appeals to
this demographic. Exploratory research clarifies problems, gathers
data and creates initial hypothesis and theories about subjects. The
primary point of exploratory research is to give researchers pertinent
information and help them to form initial hypotheses about the
subject.

Both exploratory and descriptive research can be conducted in the

same research, but exploratory should be conducted before
descriptive research in order to figure out data and hypothesis. Then,
in the second part of the research, defined hypotheses are tested.

Exploratory studies are predominantly used:

i- To satisfy the researcher's curiosity and desire for

better understanding of a particular topic;
ii- To determine the feasibility of undertaking a more
careful study;
iii- To develop research techniques and suggest direction
for future research.

Example 1.2
- Why do some cities have higher traffic accident rate than
others?
- Why are some children becoming piddling?
28 Applied Econometrics with Eviews Applications

Analytical Explanatory Research6

Analytical or explanatory research is an extension of the descriptive

research. The researcher not only describes the issue, but also
analyses and explains the reasons why or how the issue being studied
happened. Analytical research goes further by examining and
measuring causal relationship among the factors in the phenemenon.

The important elements of Analytical explanatory research are;

 Identification
 Controlling
 Exploring causal links among factors
A variable is a factor or characteristic of an issue which are
observable and measurable.

Example 1.3
- Examining the Turkey balance of payment deficit movement
during 2008 global financial crisis.
- Examining trends of real and nominal value of New Turkish Lira
against Euro after 2000.

Predictive Research

Predictive research aims to generalize findings from the analysis

conducted on the basis of hypothesized, general relationships, and
provides explanations for interactions or interchanges in a specific
and general situations.

If the predictive research provides robust, consistent and clear

explanations, its empirical finding can be applicable to the similar
cases, helps to understand more complex phenomenons. Predictive
research provides ‘how’, ‘why’ and ‘where’ answers to current events
and also to similar events that may occur in the future. It is also
helpful in situations where ‘what if’ questions are being asked.

6
Collis, J & Hussey, R. (2009)
 Scientific Research 29

Example 1.4
- What type of marketing strategy can improve the sales?
- How would an increase in interest rates affect inflation rate?
- During financial crises, what would happen to sales of
personnel computer?

Based on the Research Process7

Looking at the approach adopted by the researcher we can also

differentiate among researches. Some researchers prefer to take a
quantitative approach in addressing their research question(s) and
design a study that involves collecting quantitative data (and/or
qualitative data that can be quantified) and analyzing them using
statistical methods. Others prefer to take a qualitative approach to
address their research question(s) and design a study that involves
collecting only qualitative data and analyzing it using interpretative
methods. A comprehensive study should incorporate elements of
both.

Qualitative and quantitavie research can be used together in the

same research such as;
 Collect data through Qualitative techniques and quantify
them by counting the frequency of occurrence of specific
key words or themes in order to employ statistical methods,
 Collect qualitative data and analyse them by non-numerical
methods, or
 Collect numerical data and use statistical methods to analyze
them.

Some researchers avoid taking a quantitative approach because they

are not confident with statistical methods and think that a qualitative
approach is easier. Many researchers consider that it is pretty hard to
start and decide about overall design for a quantitative study, but it is
easier to conduct the analysis and write up the findings since it is well
structured. Qualitative research is realtively more difficult in the
stage of analyzing data and interpreting it although it is easier to
start.
7
Collis, J & Hussey, R. (2009).
30 Applied Econometrics with Eviews Applications

For instance, quantitative data such as absenteeism rates or

productivity levels may be collected in order to analyze the the
impact of night shifts on stress rates. Alternatively, same issue may
be explored by collecting qualitative data consist of night workers
perceptions and analyze it by statistical methods.
In fact, the research scope, philosophical preferences of the
researcher and the data vailabilty are the main factors in defining the
type of research.

Qualitative Research8

The qualitative research is about exploring issues, understanding

facts, and investigating the why and how of the decision making. It
uses induction scientific method and non-statistical techniques for
sample of focus rather than large samples. Qualitative research aims
to gather an in-depth understanding of human behavior and the
reasons that govern such behavior. The subjective data which is
gathered through unstructured or semi-structured techniques (in
depth interviews or group discussions) includes the perception and
interpretation of people. It produces comprehensive information,
and is exploratory or investigative. Findings are not conclusive and
cannot be used to make general conclusions about the population of
interest. It is necessary, to develop an initial understanding and
sound base for further decision making.

This type of research aims at discovering the underlying motives and

desires used in depth interviews for this purpose. Other techniques
used in such research are word association tests, sentence
completion tests, story completion tests and similar projective
techniques. Attitude or opinion research i.e., research designed to
find out how people feel or what they think about a particular subject
or institution is also a type of qualitative research. Qualitative
research is especially important in the behavioral sciences where the
aim is to discover the underlying motives of particular behavior.
Through such research we can analyze the various factors which
motivate people to behave in a particular manner or which make

8
Kothari, C.R. (2004).
 Scientific Research 31

people like or dislike something. It may be concluded that, to apply a

qualitative research method in practice is relatively huge challenge
and therefore, while doing such research, one should seek guidance
from experienced psychologists.

Example 1.5
- What is the role of the experience in participation on public
health education sessions for smokers and drug users?
- How should the nurses handle patients who refuse to follow
instructions?

Quantitative Research

Quantitative research focuses on gathering numerical data and

making general conclusions from a group of people. Its objective is to
test the hypothesis, investigate cause, effect, relationship and to
make sound predictions. Mostly, research samples are consisted of
numbers and statistics selected from population randomly. Research
questions are determined clearly in the beginning of the research.
Confirmatory deduction scientific research method is used to test
specific hypothesis and theories with randomly collected data.
Conclusion is made based on the statistical finding. In data collection,
structured techniques such as online questionnaires, on-street or
telephone interviews are used mainly.

Example 1.6
- To examine the interaction between inflation and balance of
payment.
- To examine the relationship between usage and need of
nursing services in the rural areas compared to urban areas.

Based on the Research Outcome

Applied Research

Applied research is a study that has been conducted in order to apply

its findings to solve a specific, existing problem. It is the application of
32 Applied Econometrics with Eviews Applications

available knowledge to improve management practices and policies.

The research project is likely to be short term (often less than 6
months) and the immediacy of the problem is more important than
amply academic theorizing. For example, you might be investigating
the reorganization of an office layout, the improvement of safety in
the workplace or the reduction of wasting raw materials or energy in
a production process. The output from this type of research is likely
to be a consultant’s report, articles in professional or trade
magazines and presentations to practitioners. Another type of
applied research that is conducted in academic institutions often
goes under the general title of educational scholarship (instructional
research or pedagogic research). This type of study is concerned with
improving the educational activities within the institution and the
output is likely to be case studies, instructional software or
textbooks. It focuses on finding an immediate solution to an existing
problem.

Example 1.7
- The researcher collects information to test the effectiveness of
traffic policies and fines in minimizing traffic causalities.

Basic Research

When the research problem is of a less specific nature and the

research is being conducted primarily to improve our understanding
of some general issues without emphasis on its immediate
application, it is classified as a basic or pure research. For example,
you might be interested in whether personal characteristics influence
people’s professional choices (career, etc.). Basic research is
considered as the most academic type of research, since the primary
aim is to make a contribution to the existing knowledge, usually for
the common good, rather than to solve a specific problem for one
organization. Basic research focuses on problem solving theoretically,
and provides contribution to the existing scientific knowledge which
can be used to find solutions for future problems.
 Scientific Research 33

Basic research is conducted;

- to provide new information,
- to make contribution to the knowledge,
- to improve understanding of new phenemenon, and
- to formulate a theory.

Example 1.8
- The researcher collects information to understand further the
relationship between socioeconomic status and the intention to
obey to traffic rules.

Based on the Research Logic

Deductive Research

Theoretical framework is developed and tested through empirical

observations and particular instances are deducted from general
inferences. For this reason, the deductive method is referred to as
moving from the general to the particular. Deductive research starts
with the general and proceeds to the specific and the steps are as
follows: develop a model, form a hypothesis, gather data to test the
hypothesis, and in the end use the data to conclude whether or not
the model describes reality accurate.

Example 1.9
- To test the theories of motivation in different cultures and
sectors.

Inductive Research

Inductive research represents type of study in which general

inferences are induced from particular instances. Individual
observations are used make general statements and it is referred to
as moving from the specific to the general. Inductive research starts
with the specific and proceeds to the general, thus this type of
research starts with data collection, followed by examining the data
for patterns, forming a hypothesis and then constructing a theory.
34 Applied Econometrics with Eviews Applications

Example 1.10
- To investigate the relationship between production level and
shift time in a factory in order to test the theory stating that
production levels vary with working time length.

A particular research may be applied study, analytical study with

quantitative approach or both qualitative and quantitative
approaches together It can employ deductive or inductive methods,
exploratory or descriptive research and analytical or predictive
research.

CONCEPTUAL RESEARCH
Conceptual research is related to some abstract idea(s) or theory. It is
generally used by philosophers and thinkers to develop new concepts
or to reinterpret existing ones.

EMPIRICAL RESEARCH
Empirical research relies on experience or observation alone, often
without seeking for system and theory. It is data-based research,
resulting in conclusions which are capable of being verified by
observation or experiment, thus it can be named as experimental
type of research.

The researcher starts with a working hypothesis or guesses the

probable results. Then he collects the data and sets up experimental
designs which according to him will manipulate the persons or the
materials concerned so as to bring forth the desired information.
Evidence gathered through experiments or empirical studies is today
considered to be the most powerful support possible for a given
hypothesis.
 Scientific Research 35

HISTORICAL RESEARCH9
Historical research utilizes historical sources such as documents,
remains, etc. in order to study events or ideas from the past,
including the philosophes of persons and groups at any remote point
of time. The purpose is to make people aware of what has happened
in the past offering the opportunity to learn from past failures and
successes, and apply them to present problems, make predictions
based on that and test hypothesis concerning relationships or trends.

Example 1.11
- To investigate the family trees
- To investigate the role of Ottoman Empire in the First
World War 1

9
Kumar, R. (2008).
36 Applied Econometrics with Eviews Applications
 Models 37

CHAPTER 2

MODELS

ECONOMIC MODEL
Economic modeling is one of the most important parts of a research
and the economic theory generally. The researcher organizes and
designs his/her ideas through economic modeling. The model helps
the researcher to rationally and logically support the hypothesis or
answer research questions defined in the beginning of the research.
The researcher follows certain steps in this model and delivers
scientific findings in the end. As a result, he/she produces logical and
beneficial results for the society in general.

Example 2.1
- Aggregate Supply – Aggregate Demand (AS/AD) Model,
- Loanable Funds Model
- HMC Macro Sim- simulation model,
- IS/LM Model
38 Applied Econometrics with Eviews Applications

TYPES OF ECONOMIC MODELS

Visual Models

Visual models involve graphical representation of different economic

issues. It consists of graphs with lines and curves which provides
information and ideas in brief about an issue of interest. They are
generally used in textbooks for teaching in order to clarify the
thoughts or theories. The models help to present the complex
relationships between economic variables.

Example 2.2
Supply-and-Demand model
General Equilibrium model

Mathematical Models

The mathematical models are systems of simultaneous equations

with an equal or greater number of economic variables. They consist
of more complex variables and relationships. These models require a
good previous knowledge of algebra or calculus for its
implementation.

For example, a basic microeconomics model includes a supply

function (shows the behavior of producers), a demand curve (shows
the behavior of purchasers) and an equilibrium equation. The
variables in this model represent a type of economic activity or
information that either determines or is determined by that activity
and are classified as endogenous or exogenous.

Example 2.3
To examine the demand elasticity for luxury cars in low
income countries

Empirical Models

Empirical models are one type of mathematical models designed to

be used along with data. The basic model is mathematical and the
 Models 39

data is gathered for the variables of interest; afterward statistical or

econometric techniques are applied to estimate the values of the
variables which are used previously in the model.

Example 2.4
To investigate the changes in income when investment changes
one percent
2.2.4 Simulation Models

Simulation models are mainly used and created by using different

computer software. The basic features of mathematics are required.
The mathematical complexity varies depending on the research
scope.

Example 2.5
The equations of the model are programmed in a software
programming language

Static Model

Most models which are used in economics are comparative statics

models. These models provide information about what happens over
time. The model estimate generally starts with predefined
equilibrium condition, and then a "shock" to the model is applied
(changing the value of one or more of the variables). In the end, the
new equilibrium is obtained without an exposition of what happened
in the transition from first equilibrium to the second.

Example 2.6
To investigate the impact of class size on the average test
score of the students

Dynamic Models

Dynamic models directly incorporate time into the model. This is

usually done in economic modeling by using differential equations.
Sometimes dynamic models better represent the subtleties of
40 Applied Econometrics with Eviews Applications

business cycles, because time lags in behavioral response and timing

strongly influence the character of one cycle.

Example 2.7
To examine the role of Interest rate on Inflation rate
movements over time

WHICH THEORY IS APPROPRIATE?

While describing the economic activity of consumers in general or a
certain type of business decisions the theory of adaptive expectations
is probably most suiTable. A rational expectations theory is the most
appropriate one when describing an economic activity or en
economic environment where the decision-makers are likely to be
sophisticated and well-informed. The behavior and actions of traders
in the finance markets provide a good example for rational
expectations. Portfolio investors, traders of financial assets or
company shareholders observe the market and seek for new
economic information virtually whole the working day. The market
forces adjust themself instantly to the macroeconomic indicators,
expecting the impact of such policy on interest rates and
consequently share prices. Hence, the choice of the theory of
expectations depends on the context. Models incorporating and
related to both theories would be developed and used.

ECONOMETRIC MODELS
Econometrics is a science and art of using economic theory and
statistical techniques to analyze economic data.10 An econometric
model should be constructed based on economic theory, experience
or critical thinking. Econometrics finds and explains the relationship
between economic variables based on outcomes of statistical
techniques using the available data. It clarifies the interaction and
mediation between data, economic theory and statistical techniques.
10
Stock, J.H & Watson, M.W. (2008).
 Models 41

Econometrics aims to explore relationship between economic

variables and interpret the results obtained through statistical
techniques.

Econometrics models can be classified as follows:

i- The models developed to find out relationships

between past and present.

Example 2.8
The impact of previous days’ stock returns on the current stock
return

ii- The models that examine relationship between

economic variables over time

Example 2.9
To examine the degree of relationship between export and
inflation over time

iii- The models which investigate the relationship

between different variables measured at a given point
of time.

Example 2.10
To investigate the impact of student-teacher ratio on student
exam results

iv- The models which consider the relationship between

different variables for different units over time

Example 2.11
To examine the relations between inflation and
unemployment in different countries
42 Applied Econometrics with Eviews Applications

Attributes of a Good Econometric Model

Features of the good econometric model may be summarized as

follows;

 Parsimonious, explain a lot with minimum specification,

optimum sized model;
 Identifying- unique values exist for the parameter.
 Goodness of fit- it shows the power of the model in explaining
changes in regress.
 Theoretical Consistency- consistent with the related economic
theories
 Predictive power.

Structural Equation Modeling (SEM)

Structural equations are the equations specific for the economic

theory. There are different types of structural equations such as:

i- Behavioral equation, consumption equation

Y= C + I + X – M

ii- Technical relationships, production function

Q = f (L, K)

iii- Identities, Keynesian macro model

The Reduced Form Model

It includes reduced form of equations. Keynesian macro model is a

complete structural form with two equations for the two endogenous
variables is supposed to be exogenous. Solving the
model for gives two reduced from equations:
 Models 43

In different notation;

The parameters are called reduced form

parameters. Reduced form equations indicate that the endogenous
variables are correlated with the exogenous regressors. Thus, each
equation can be estimated by 2SLS using all of the exogenous
variables in the system.

In the reduced form of equations the endogenous variables are

expressed in terms of the exogenous and lagged variables. A special
reduced form model is the model with only exogenous explanatory
variables. These types of models are called the classical regression
model. In this model all of the explanatory variables should be
exogenous.

Stochastic Models

Stochastic modeling is a technique of predicting outcomes and takes

into account a certain degree of randomness, or unpredictability.
Economic relationship is not an exact relationship; a disturbance or
error term, should be added to right hand side of the equation. It
is also called stochastic term in the non systematic part of the
regression equation.

Deterministic Models

Deterministic model is a mathematical model in which outcomes are

determined through known relationships among variables and events
44 Applied Econometrics with Eviews Applications

without random variation. In deterministic models, given input

always produces the same output, such as in a known chemical
reaction. In comparison, stochastic models use ranges of values for
variables in the form of probability distributions.
 Models 45

Analytical Framework

LITERATURE REVIEW

DATA COLLECTION
a- Daily data
- Stock market indices* and Exchange rates*
b- Monthly data
- Stock market indices*, Exchange rates*, Interest rates*,
Export and Import volume**, GDP per capita**, FDI**, M3**
* Turkey, EU member countries (27), non-EU member
European countries (5),the US and Japan
** Turkey

APPLICATIONS
Financial linkages between Turkey and EU &US & Japan
(external linkages)
Interrelationship among ISE, welfare and monetary policy
(internal linkages)

DIAGNOSTIC TESTING

ECONOMETRIC ANALYSIS

REPORT

PRESENTATION AND
PUBLICATION

Figure 2.1 Analytic Framework

46 Applied Econometrics with Eviews Applications

Methodological Framework

Neighborhood Effect How do financial linkages react to

the development? Testing Gosh
(1999) theory
Trade
Stock Markets Effect
Exchange Rates
Interest Rates Foreign Direct
Investment Effect

Development Effect To policy makers,

Long-Term Short-Term To provide information to sustain

Unionization Effect development and eliminate external
Relationship Relationship
and internal shocks

Macroeconomic Variables Monetary Policy

Effect To investors,

To determine the segmented and

Welfare Effect
linked times and points between
countries and Financial &Monetary
variables for profit maximization
FIGURE 2.2 Methodological Framework
 Models 47

Research Process Flow Chart

Define Research Problem

Review related theories

Review previous research

findings

Determine Research
objectives and questions

Research Design

Data Collection

Analyzing data

Interpretation

Figure 2.3 Research Process Flow Chart

48 Applied Econometrics with Eviews Applications

Regression Analysis Flow Chart

Stationary
Test,
ADF and PP

Coint. Test Vector

EngleGranger Autoregressive
and Johansen Analysis

Vector Error Correction Forecast Error Variance

Model Decomposition Analysis

Granger Causality Test Generalized Impulse

Response Analyses

Cointegration with Recursive Cointegration Permanent and

rolling window (Hansen and Johansen) Transitory Decomposition
approach Analysis

The Asymmetric Non-linear Cointegration The Threshold

Vector Error Models Cointegration Technique
Correction Model

Diagnostic testing Stability Test

Stochastic Assumption
Deterministic Assumptions
 Data 49

CHAPTER 3

DATA

Data are the piece of information or knowlege which are used for
reasoning or calculation as measurement.

QUALITATIVE DATA
Qualitative data is extremely varied in nature. It includes virtually any
information that can be captured but is not numerical in nature. Here
are some of the major categories or sources of qualitative data:

In-Depth Interviews

In-depth interviews include both individual interviews (one-on-one)

as well as "group" interviews (including focus groups). The data can
be recorded in a wide variety of ways including audio recording,
video recording or written notes. In-depth interviews differ from
direct observation primarily in the nature of the interaction. In
interviews it is assumed that there is a questioner and one or more
interviewees. The purpose of the interview is to probe the ideas of
the interviewees about the fact of interest.
50 Applied Econometrics with Eviews Applications

Example 3.1
 Which business leaders or role models do you respect?
 How do you handle conflict on a team project?
 How do you see the efficiency of current economic policies?
 What sectors should government support? Why?

Direct Observation

Direct observation differs from interviewing in that the observer does

not actively examine the respondent. It includes field research to
illustrate some aspect of the phenomenon. The data can be obtained
through interviews (stenography, audio, and video) or pictures,
photos or drawings (courtroom drawings of witnesses).

Example 3.2
 To analyze the student’s reaction to the male and female
teachers
 To analyze the impact of family income on their children’s
behavior in the public schools

Written Documents

Written documents refer to existing and available documents. It

includes newspapers, magazines, books, websites, memos,
transcripts of conversations, annual reports, and so on.

Example 3.3
To analyze the ethical sensitiveness in the novels published
between two world wars

Qualitative Methods

Generally, the qualitative methods are limited by the imagination of

the researcher. There are a wide variety of methods that are
common in qualitative measurements. Some of them are mentioned
below.
 Data 51

Participant Observation

Participant observation is one of the most common and demanding

methods for qualitative data collection. The researcher should be a
participant in the environment or context being observed. How to
enter the context, the role of the researcher as a participant, the
collection and storage of field notes, and the analysis of field data are
discussed in the literature on participant observation. Participant
observation is usually long term, involving several months or years of
intensive work because the researcher is required to be natural part
of the environment in order to assure that the observations are
natural.

Direct Observation

Direct observation is different from participant observation in the

following ways:
 A direct observer doesn't necessary become a participant in
the environment. Instead, direct observer tries to be as
retiring as possible so as not to spoil or bias the observations.
 Direct observation suggests a more detached perspective. The
researcher is watching rather than participating. Technology
can be a useful part of direct observation. For instance, one
can videotape the phenomenon or observe from far or behind
one-way mirrors.
 Third, direct observation tends to be more focused than
participant observation. The researcher is observing certain
sampled situations or people rather than trying to become
immersed in the entire context.
 Finally, direct observation tends not to take as long as
participant observation. For instance, one might observe
child-mother interactions under specific circumstances in a
laboratory setting from behind a one-way mirror, looking
especially for the nonverbal cues being used.

Structured Interviewing

It is associated with quantitative research and the main goal is to

collect data from large samples ensuring consistency of response. In
52 Applied Econometrics with Eviews Applications

this interview, the questions are clear and fully understood by the
respondents and, ach respondent answer same set of question in
order to obtain an objective result minimizing interviwer bias.

The questions are selected based on research objectives. It includes

detailed instructions and information regarding interview questions.

Unstructured Interviewing

Unstructured interviewing requires direct interaction between the

researcher and a respondent or group. It differs from traditional
structured interviewing in the following:
- There is no formal structured instrument, questions or
protocol except some limited guiding questions. But, the
researcher may have some initial guiding questions or core
concepts to ask about.
- The researcher or interviewer can move the conversation in
any direction of interest that may come up.
- Unstructured interviewing is particularly useful for exploring a
topic broadly.
- Each interview is unique with no predefined questions to all
respondents.
- It is more difficult to analyze unstructured interview data.

Case Studies11

A case study is an intensive study of a specific individual or specific

entity. For instance, Freud developed case studies of several
individuals as the basis for the theory of psychoanalysis and Piaget
did case studies of children to study developmental phases. There is
no single way to conduct a case study, and a combination of methods
(e.g., unstructured interviewing, direct observation) can be used.

11
http://www.socialresearchmethods.net
 Data 53

QUANTITATIVE VERSUS QUALITATIVE DATA

Quantitative methods focus only on numbers and frequencies rather
than on meaning and experience; they are associated with the
scientific and experimental approach without in depth description.

Qualitative methods are ways of collecting data which are concerned

with describing, rather than drawing statistical inferences and
provide a more in-depth and rich description analysis.

In modern research, most psychologists tend to adopt a combination

of qualitative and quantitative approaches, which allow statistically
reliable information obtained from numerical measurement to be
backed up and enriched by the information gained from the research
participants' explanations.

WHY ARE QUANTITATIVE AND QUALITATIVE DATA

IMPORTANT?
Quantitative and qualitative method provide different outcomes, and
are often used together to get a full picture of a population. For
example, if data are collected on annual income (quantitative),
occupation data (qualitative) could also be gathered to get more in
detail the information on the average annual income for each type of
occupation. Quantitative and qualitative data can be obtained from
the same data unit depending on whether the variable of interest is
numerical or categorical.

QUANTITATIVE DATA TYPES

Primary Data

Primary data are those that the researcher collects himself. The
researcher is expected to have accessed primary data while using
54 Applied Econometrics with Eviews Applications

quantitative methods, and primary data collection is necessary when

a researcher cannot find the data needed in secondary sources.

Secondary Data

A secondary data research uses existing data. It can be obtained from

very wide range of sources such as: literature, industry surveys,
reports, databases and information systems.

Experimental Data12

Experimental data is obtained through experiments in order to

examine economic policy or treatment. It is much more expensive to
conduct compared to observational data. It also has administration
difficulties and ethical consideration.

In sciences, experimental data is data produced by a measurement,

test method, experimental design or quasi-experimental design. In
clinical research any data produced as a result of clinical trial is
experimental data. It can be qualitative or quantitative, each being
appropriate for different investigations.

Observational data

Most of the data in the economic research is obtained through

observations. It includes surveys (telephone surveys, on street
surveys, mail surveys, and interviews), company records, government
records, past financial data records, etc. In economics, the
researchers usually do analysis with observational data in different
circumstances which are not under the control of researchers.

Researchers who use observational data can obtain data from lab
scientists, make reliable conclusions and recognize the limitations of
data that are gathered in more natural settings.

12
http://en.wikipedia.org/wiki/Experimental_data
 Data 55

Time Series13

A time series is a collection of observations of variables obtained

through repeated measurements over time. A times series allows the
researcher to identify presumed changes within a population over
time. It can also show the impact of cyclical, seasonal and irregular
events on the data item being measured. Time series can be
classified into two different types: stock and flow.

A stock series is a measure of certain attributes at a point in time. A

flow series is a series which is a measure of activity over a given
period. An original time series shows the actual movements in the
data over time and includes any movements due to cyclical, seasonal
and irregular events.

A cyclical effect is any regular fluctuation or changes in daily, weekly,

monthly or annual data. For example, the number of people using
public transportation has regular peaks and troughs during each day
of the week, depending on the point of time during a day. A seasonal
effect is any variation dependent on a particular season of year. For
example, fruit and vegeTable prices can vary depending on whether
or not they are 'in-season'. An irregular effect is any movement that
occurred at a specific point of time, but is unrelated to a season or
cycle, for example; a natural disaster, the introduction of legislation,
or a one-off major cultural or sporting event.

A seasonally adjusted series involves estimating and removing the

cyclical and seasonal effects from the original data. Seasonally
adjusting a time series is useful if you wish to understand the
underlying patterns of change or movement in a population, without
the impact of the seasonal or cyclical effects. For example,
employment and unemployment are often seasonally adjusted so
that the actual change in employment and unemployment levels can
be seen, without the impact of periods of peak employment such as
Christmas/New Year when a large number of casual workers are
temporarily employed.

13
http://www.abs.gov.au
56 Applied Econometrics with Eviews Applications

A trend series is a seasonally adjusted series that has been further

adjusted to remove irregular effects and 'smooth' out the series to
show the overall 'trend' of the data over time. For example, the trend
is often used when analyzing economic indicators such as
employment and unemployment levels.

Example 3.4

Table 3.1 Macroeconomic Indicators

Obs. Number Year Inflation Export Interest rate
1 1980 2000 1250 20
2 1981 2000 560 25
3 1982 2000 780 10
- ---- --- --- --
- ---- --- --- --
- ---- --- --- --
34 2012 2012 8000 100
35 2013 2012 9000 250

Panel Data14

Panel data is consisted of multiple entities or variables observed at

multiple time periods. It is also known as longitudinal or cross-
sectional time-series data. These entities could be states, companies,
individuals, countries, interest rates, exchange rates, etc. Panel
analysis uses panel data to examine changes in variables over time
but also differences in variables between several subjects
simultaneously.

Panel data allows the researcher to control variables which are not
possible to be observed or measured such as cultural factors or
differences in business practices across companies, variables that
change over time but not across entities, and other. New variables
can be included at different levels of analysis (i.e. students, schools,
districts, states). It is suitable for multilevel or hierarchical modeling.

14
Baltagi, B.H. (2008).
 Data 57

Example 3.5

Table 3.2 University Graduates by University and Year

Obs. Number University Year Bachelor Grad. Master Grad.
1 Sivas University 2000 1250 20
2 Tampin University 2000 560 25
3 Bangi University 2000 780 10
- ---- --- --- --
- ---- --- --- --
- ---- --- --- --
34 Sivas University 2012 8000 100
35 Tampin University 2012 9000 250
36 Bangi University 2012 5400 80

Cross Sectional Data

Cross Sectional Data consists of multiple entities or variables which

are observed at a point of time. Analysis of cross-sectional data
usually consists of comparing the differences among the entities.

For example, the researcher wants to measure current education

level in a society. A sample of 2,000 people randomly can be selected
from that population (cross section of that population), and
examination of the percentage of educated people in the sample can
be conducted. This cross-sectional sample provides us with a
snapshot of that population, at that one point of time. We can only
describe the current proportion of educated people in the society not
changes over time, decreasing or increasing, through cross sectional
data analysis.

Example 3.6

Table 3.3 Observing Math Test Scores in Bosna Sema Schools

Obs. Number Test Score Father Income ($) Father Education Gender
1 68 3,000 Primary Male
2 98 2,000 Primary Male
3 40 3,500 High School Female
- ---- --- --- --
- ---- --- --- --
- ---- --- --- --
57 25 10,000 Master Male
58 Applied Econometrics with Eviews Applications

ARTIFICIAL EXPLANATORY VARIABLES

Qualitative explanatory variables are unobserved variables. The
constant term, dummy variables and deterministic trends are
qualitative explanatory variables. If dependent variable shows a
linear trend and not one of the explanatory variables has such
pattern, then estimation result will show a lot of residual
autocorrelation. A deterministic linear trend variable may be
included in the equation to solve this problem.

The Chow Breakpoint Test

If the data exhibits different regimes whether they should have been
modeled with different specification or not, it should be tested
through the Chow breakpoint test as follows:

∑
∑

Where, m : number of sub-samples

SR : RSS for the whole sample, (restricted sum of squares)
Si : RSS for the sub- sample,
K: number of estimated parameters

F-distribution is valid if the error terms are independently and

identically normally distributed. The null hypothesis of no structural
breaks in the sample period is tested with the chow breakpoint test.
For a break at one date the test resembles the Chow forecast test,
but different null hypothesis is tested with a different test statistic.
More dates can be specified for the breakpoint test.15

15
Vogelvang, B. (2005).
 Data 59

Example 3.7

Table 3.4 Eviews Output for Chow Breakpoint Test with One
Breakpoint
Chow Breakpoint Test: 1970Q1
Null Hypothesis: No breaks at specified breakpoints
Varying regressors: All equation variables
Equation Sample: 1954Q1 1994Q4

F-statistic 74.39175 Prob. F(2,160) 0.0000

Log likelihood ratio 107.8245 Prob. Chi-Square(2) 0.0000
Wald Statistic 148.7835 Prob. Chi-Square(2) 0.0000

Table 3.4 Eviews Output for Chow Breakpoint Test with Multiple
Breakpoint
Chow Breakpoint Test: 1970Q1 1982Q2 1988Q3
Null Hypothesis: No breaks at specified breakpoints
Varying regressors: All equation variables
Equation Sample: 1954Q1 1994Q4

F-statistic 40.60198 Prob. F(6,156) 0.0000

Log likelihood ratio 154.2646 Prob. Chi-Square(6) 0.0000
Wald Statistic 243.6119 Prob. Chi-Square(6) 0.0000

DATA TRANSFORMATION
Log Transformation

The log transformation yields appealing interpretation of coefficients

and model. The interpretation is good for small changes only. The log
transformation makes coefficients invariant to rescaling. For
example, lnY may look more normal than Y. But also, lnY has a
narrower range than Y.

When or not to Log?

 Do not log zeroes or negative values

 Do not log dummies
60 Applied Econometrics with Eviews Applications

 Potentially large monetary variables are often logged:

revenue, income, wages;
 Large integer values are often logged: population, number of
employees, number of students;
 Small integer values are usually not: age, education, number
of children;
 Percentages can be logged or not

For example, an increase in unemployment rate from 4% to 4.5% is

half a percentage point change. However it implies a 12.5%
percentage change in the population of unemployed. The difference
in logs is log 4.5 − log 4 = 11.8%. This is an approximation of the
percentage change, good for small changes.

How to Choose?

Sometimes it is unclear which form to choose, and thus we have to:

i- Rely on economic theory and previous studies.

ii- Think about what is implied by particular functional forms for
the relevant range of the variables (even if your data does not
include the whole relevant range).
iii- Do not compare if the dependent variable is
different.
iv- Even in the case of the same dependent variable, e.g. linear
and log-linear, beware of selecting the functional form on the
sole basis of . Selecting the functional form on
the basis of fit only gives you an equation that works well for
your particular sample.

Differencing

Stationarity is an issue for time series data and is a pre-condition to

perform regression analysis. Nonstationary time series data should
be converted into stationary data by taking differences. The time
series data which is stationary after first differencing is called
stationary in order one (I(1)).
 Data 61

Percentage Change

It is calculated by subtracting old one from new one and dividing the
difference by old one. It is used for the give emphasis return.

Base Year

A base year is used for comparison of the two or more time series
data levels. The arbitrary level of 100 is selected so that percentage
changes (either rising or falling) over year can be easily depicted. Any
year can be chosen as base year, but generally recent years are
chosen and the other observations are adjusted based on the base
year.
62 Applied Econometrics with Eviews Applications
 Regression 63

CHAPTER 4

REGRESSION

DEFINITION
Regression is an econometric technique for estimating the
relationships among variables. It helps to analyze how the typical
value of the dependent variable changes when any one of the
independent variables changes, while the other are held constant
(ceteris paribus). Linear regression estimates how much y changes
when x changes one unit. The main purpose of linear regression
analysis is to assess associations between dependent and
independent variables.

SIMPLE REGRESSION MODELS

Linear Regression with one Regressor,

(4.1)

Where;

i denotes observations, i = 1,2,…..,N

64 Applied Econometrics with Eviews Applications

is called the dependant variable, regressand, response variable,

measured variable, explained variable, outcome variable,
experimental variable, or output variable.

is called independent variable, regressor, explanatory variable,

predictor variable, controlled variable, input variable or exogenous
variable.

is called population regression line or function

is constant term or intercept term

is regression coefficient, slope of the regression line

is the deviation of the actual value of an observation from the true

regression line which is called the error term, stochastic term,
residual term or disturbance term. It consists of omitted independent
variables and measurement error.

Regression equation consists of two components:

i- Deterministic component,
ii- Nonsystematic or stochastic components,
is called intercept, are the slope coefficients. They all are
called regression coefficients or regression parameters.

Example 4.1
(4.2)
implies degree and direction of the relationship between Income
and Consumption.
y
𝛽
𝛽 𝑥𝑖
𝑥 𝑦

e1
e2

𝑥 𝑦 x

Figure 4.1 Scatter Plot of Simple Regression

 Regression 65

is the population regression line as shown in Figure 4.1

are error terms for the first and second observation

ESTIMATION OF THE LINEAR REGRESSION MODEL

Ordinary Least Squares Estimation Method (OLS)

OLS minimizes the squared difference between observed and

predicted parameters from the model. Differences are called
residuals or error terms.

(4.3)

̂ ̂ ̂ ̂ (4.4)

Predicted values of the regression ̂ ̂ ̂ are OLS

estimators.

And the residuals are computed as follows:

̂ (4.5)

the residuals differ from disturbances. The residuals are observed,

the disturbances are not. OLS is trying to get a best model in order to
obtain residuals as small as possible. Main objective is to minimize
the sum of the squares of the residuals (ESS) as follows:

(̂ ̂ ̂ ) ∑ (4.6)

In order to determine the minimum f with respect to predicted

parameters of the regression or to minimize ESS, we can find out the
values of regression coefficients which make ESS as small as possible
as indicated in Eq.4.5. ESS is a function of regression coefficients.
66 Applied Econometrics with Eviews Applications

∑ ̅ ̅
̂ (4.7)
∑ ̅

̂ ̅ ̅ (4.8)

Example 4.2

Regression equation in Eq.4.2 is estimated using OLS and

found following results
8.4

8.2

8.0

7.8
INCOME

7.6

7.4

7.2

7.0

6.8
6.8 7.0 7.2 7.4 7.6 7.8 8.0 8.2 8.4

CONSUMPTION
Figure 4.2 Eviews Output for Scatter Plot between Income and
Consumption

Using the data income.xls OLS estimation result is presented below:

 Regression 67

Table 4.1 Eviews Output for Simple Regression Model Estimates

Dependent Variable: CONSUMPTION
Method: Least Squares
Sample: 1954Q1 1994Q4
Included observations: 164

Variable Coefficient Std. Error t-Statistic Prob.

INCOME 1.001918 0.003254 307.9252 0.0000

C -0.110026 0.025047 -4.392831 0.0000

R-squared 0.998294 Mean dependent var 7.592739

Adjusted R-squared 0.998284 S.D. dependent var 0.388767
S.E. of regression 0.016105 Akaike info criterion -5.407225
Sum squared resid 0.042019 Schwarz criterion -5.369422
Log likelihood 445.3925 Hannan-Quinn criter. -5.391879
F-statistic 94817.94 Durbin-Watson stat 0.284313
Prob(F-statistic) 0.000000

Example 4.3:

18,000

16,000

14,000

12,000
LOTSIZE

10,000

8,000

6,000

4,000

2,000

0
0 40,000 80,000 120,000 160,000 200,000

PRICE
Figure 4.3 Eviews Output for Regression Line
68 Applied Econometrics with Eviews Applications

Table 4.2 Eviews Output for Simple Regression Model Estimates

Dependent Variable: PRICE
Method: Least Squares
Sample: 1 546
Included observations: 546

Variable Coefficient Std. Error t-Statistic Prob.

LOTSIZE 6.598768 0.445847 14.80053 0.0000

C 34136.19 2491.064 13.70346 0.0000

R-squared 0.287077 Mean dependent var 68121.60

Adjusted R-squared 0.285766 S.D. dependent var 26702.67
S.E. of regression 22567.05 Akaike info criterion 22.89003
Sum squared resid 2.77E+11 Schwarz criterion 22.90579
Log likelihood -6246.977 Hannan-Quinn criter. 22.89619
F-statistic 219.0558 Durbin-Watson stat 1.089325
Prob(F-statistic) 0.000000

Multivariate Regression Model

(4.9)

Multivariate regression model includes more than one regressor.

In regression analysis, main objective is to observe how dependent

variable responds to changes in the values of independent variables.
The causal relationship which is detected through regression analysis
should be based on the relevant economic theory. Each slope
coefficient measures the rate of change in the mean value of
dependent variable for a unit change in the value of an independent
variable, holding the values of the other independent variables
constant. How many independent variables should be included in the
regression equation depends on the theory, research objective, and
assumptions.
 Regression 69

It is assumed that the independent variables are not random

variable. Their values are fixed in repeating samples. In other words,
regression analysis is conditional on the given values of independent
variables.

Error term, ei, includes all the independent variables which are not
defined in the regression because of the data unavailability or
measurement error. We assume that effect of error term on
dependent variable is marginal.

Example 4.4

(4.10)

The multivariate regression equation above shows the impact of Lot

Size, Number of bedrooms, Number of Bathrooms, and Number of
Stories on the determination of House Price

Table 4.3 Eviews Output for Multiple Regression Model Estimates

Dependent Variable: PRICE
Method: Least Squares
Sample: 1 546
Included observations: 546

Variable Coefficient Std. Error t-Statistic Prob.

LOTSIZE 5.429174 0.369250 14.70325 0.0000

BEDROOMS 2824.614 1214.808 2.325153 0.0204
BATHRMS 17105.17 1734.434 9.862107 0.0000
STORIES 7634.897 1007.974 7.574494 0.0000
C -4009.550 3603.109 -1.112803 0.2663

R-squared 0.535547 Mean dependent var 68121.60

Adjusted R-squared 0.532113 S.D. dependent var 26702.67
S.E. of regression 18265.23 Akaike info criterion 22.47250
Sum squared resid 1.80E+11 Schwarz criterion 22.51190
Log likelihood -6129.993 Hannan-Quinn criter. 22.48790
F-statistic 155.9529 Durbin-Watson stat 1.482942
Prob(F-statistic) 0.000000
70 Applied Econometrics with Eviews Applications

= 7634.897

CLASSICAL LINEAR REGRESSION MODEL (CLRM)

ASSUMPTIONS
i- The regression model is linear in the parameters (can
be nonlinear or linear in the dependent and
independent variables). The Regression model should
correct variables without omitted variable bias and
have correct functional form.

ii- The regressors are assumed to be fixed or non-

stochastic in repeated sampling. This assumption may
not be appropriate for all economic data; if
independent variables and error term are
independently distributed, the results based on the
classical assumption hold true provided the analysis is
conditional on the particular values of independent
variables drawn in the sample. However, if
independent variable and error term are uncorrelated,
the classical results hold true asymptotically (i.e. in
large samples).

iii- Given the values of the (independent) variables, the

expected, or mean, value of the error term is zero.
|

The conditional expectation of the error term, given

the values of the independent variables, is zero. Since
the error term represents the influence of other
factors that are not included in the regression
equation and may be essentially random, it is very
logical to assume that their mean or average value is
zero.
 Regression 71

As a result of this critical assumption, we can write the

regression equation as:

| | (4.11)

This can be interpreted as the model for mean or

average value of yi conditional on the xi values. This is
the population (mean) regression function (PRF). In
regression analysis, the main objective is to estimate
this function. If there is only one independent variable,
it can be visualized as the (population) regression line.
If there is more than one independent variable, it
would be a curve in a multi-dimensional graph. The
estimated PRF is denoted by ̂ .

Suppose that | , then | , we

subtract 3 from the error term and add 3 to the
intercept (constant term) as follows;

, can be written as,

, where and

then | . Assumption is satisfied.

iv- The variance of each , given the values of x, is

constant, or homoscedastic

| (4.12)

v- There is no correlation between two error terms. That

is, there is no autocorrelation.

( | ) (4.13)

If there is autocorrelation, an increase in the error

term in one period affect the error term in the next.
72 Applied Econometrics with Eviews Applications

vi- There are no perfect linear relationships among the

independent variables. This is the assumption of no
perfect multicollinearity.

vii- Error term is not correlated with the independent

variables.

| (4.14)

viii- All independent variables should be exogenous which

are defined outside of the model.

ix- The regression model should be correctly specified.

Alternatively, there is no specification bias or
specification error in the model used in empirical
analysis. It is implicitly assumed that the number of
observations, N, is greater than the number of
parameters (K) estimated.

x- Although it is not a part of the CLRM, it is assumed that

the error term follows the normal distribution with
zero mean and (constant) variance. It is necessary for
the hypothesis testing.

(4.15)

STATISTICAL PROPERTIES OF OLS

Linearity

The estimators are linear, that is, they are linear functions of the
dependent variable, . Linear estimators are easy to understand and
deal with compared to nonlinear estimators. Linearity assumption
can be expressed as follows:

(4.16)
 Regression 73

Linear models can be expressed in a form that is linear in the

parameters by a transformation of the variables. Nonlinear models,
on the other hand, cannot be transformed to the linear form. The
non-linearity of interest here is the one which cannot be
accommodated into a linear conditional mean after transformation.

Reset-type test (Ramsey, 1969) is the most common test for testing
the linearity assumption. This testing procedure involves the
estimation of the following (auxiliary) regression:

̂ ̂ (4.17)

,
.

We can now test the statistical significance of using t-test, F-

statistic or LM test as follows:

̂ (4.18)

∑ ̂ ∑ ̂
∑ ̂
(4.19)

(4.20)

when linearity does not hold, the OLS estimators are biased and
inconsistent. In other words, estimation and testing results are
invalid and the model should be restructured.

Nonlinearity is seen very clear in a plot of the observed versus

predicted values or a plot of residuals versus predicted values, which
are a part of standard regression output. The points in the first plot
should be symmetrically distributed around a diagonal line or a
horizontal line in the second plot. The evidence of a "bowed" pattern
should be examined carefully. It indicates that the model makes
systematic errors whenever it is making unusually large or small
predictions.
74 Applied Econometrics with Eviews Applications

If the transformation seems to be appropriate, a nonlinear

transformation may be applied to the dependent and/or
independent variables, as follows:

- A log transformation may be feasible if the data are strictly

positive,

- Adding another regressor which is a nonlinear function of one

of the other variables. For example, if you have regressed
on , and the graph of residuals versus predicted suggests a
parabolic curve, then it may make sense to regress on both
and (i.e., -squared). The latter transformation is
possible even when and/or have negative values, except
logging.

Example 4.5

200,000

160,000

120,000
PRICE

80,000

40,000

0
0 40,000 80,000 120,000 160,000 200,000

PRICEF

Figure 4.4 Eviews Output for Scatter Graph of Observed and

Predicted Values of House Prices Based on The Regression
Equation above
 Regression 75

Unbiasedness

The estimators ( ̂ ) are unbiased, that is, in repeated

applications of the method, on average, the estimators
approach their true values.

(̂) (4.21)

Efficiency

In the class of linear estimators, OLS estimators have

minimum variance. As a result, the true parameter values can
be estimated with least possible uncertainty; an unbiased
estimator with the least variance is called an efficient
estimator.

Normality

The assumption of normality can be expressed as follows:

⁄ ́ (4.22)

If the assumption of normality does not hold, then the OLS

estimator ( ̂ ) remains the Best Linear Unbiased Estimator
(BLUE), i.e. it has the minimum variance among all linear
unbiased estimators. However, without normality one cannot
use the standard for the t and F distributions to perform
statistical tests.

The following null hypothesis should be specified before

normality test.

The null is that the skewness16 and kurtosis17

coefficients of the conditional distribution of (or,
equivalently, of the distribution of ) are 0 and 3, respectively:

16
a measure for the degree of symmetry in the variable distribution
17
a measure for the degree of peakedness/flatness in the variable distribution.
76 Applied Econometrics with Eviews Applications

The above assumptions can be tested jointly using the Jarque-

Bera test (JB, 1981) which follows asymptotically a chi-square
distribution:

[ ̂ ̂ ] (4.23)

̂ [ ∑ ̂ ∑ ̂ ] (4.24)

̂ [ ∑ ̂ ∑ ̂ ] (4.25)

Example 4.6
80
Series: Residuals
70 Sample 1 546
Observations 546
60
Mean -9.39e-12
50 Median -805.5036
Maximum 85311.59
40 Minimum -52757.51
Std. Dev. 18198.08
30 Skewness 0.893173
Kurtosis 5.649682
20
Jarque-Bera 232.3196
10 Probability 0.000000
0
-40000 -20000 0 20000 40000 60000 80000

Figure 4.5 Eviews Output for Scatter Plot

Scatter plot shows whether the series are distributed normally or not.
It skewed the left side and shows leptokurtic behavior.
 Regression 77

14
Series: Residuals
12 Sample 1954Q1 1994Q4
Observations 164
10
Mean -1.77e-15
Median 0.000850
8 Maximum 0.032516
Minimum -0.046830
6 Std. Dev. 0.016056
Skewness -0.224505
4 Kurtosis 2.604619

Jarque-Bera 2.445893
2
Probability 0.294362
0
-0.0375 -0.0250 -0.0125 0.0000 0.0125 0.0250

Figure 4.6 Eviews Output for Scatter Plot

Result of skewness and kurtosis indicates that income series are

normally distributed around the mean.

If the residuals do not follow a normal pattern, omitted variables,

model specification, functional forms and linearity should be
checked. Normally, normality does not represent much of a problem
in big samples.

GAUSS-MARKOV THEOREM
Under the assumed above mentioned conditions, OLS estimators are
BLUE (best linear unbiased estimators). This is the essence of the
well-known Gauss–Markov theorem which provides a theoretical
justification for the method of least squares. With the added
assumption of normality, the OLS estimators are best unbiased
estimators (BUE) in the entire class of unbiased estimators, whether
linear or not. With normality assumption, CLRM is known as the
normal classical linear regression model (NCLRM).

For the linear regression model, an estimate of the variance of the

error term is:

∑
̂ (4.26)
78 Applied Econometrics with Eviews Applications

that is, the residual sum of squares (RSS) divided by (N-K), which is
called the degrees of freedom (df), N being the sample size and K
being the number of regression parameters estimated, an intercept
and (K – 1) slope coefficients. ̂ is called the standard error of the
regression (SER) or root mean square.

Example 4.7

Table 4.4 Eviews Output for Simple Regression Model Estimates

Dependent Variable: CONSUMPTION
Method: Least Squares
Included observations: 164

Variable Coefficient Std. Error t-Statistic Prob.

INCOME 1.001918 0.003254 307.9252 0.0000

C -0.110026 0.025047 -4.392831 0.0000

R-squared 0.998294 Mean dependent var 7.592739

Adjusted R-squared 0.998284 S.D. dependent var 0.388767
S.E. of regression 0.016105 Akaike info criterion -5.407225
Sum squared resid 0.042019 Schwarz criterion -5.369422
Log likelihood 445.3925 Hannan-Quinn criter. -5.391879
F-statistic 94817.94 Durbin-Watson stat 0.284313
Prob(F-statistic) 0.000000

∑
√̂ √

COEFFICIENT OF DETERMINATION, GOODNESS OF

FIT
It measures the proportion of the variance in the dependent variable
which is explained by variations in all independent variables.

After estimation of a particular linear model, a question comes up

such as: how well does the estimated regression line fit the
 Regression 79

observation? A popular measure for the goodness of fit is called R2 (R

squared) and is defined as follows:

Total Sum of Squares (TSS): ∑ ∑ ̅ (4.27)

Explained Sum of Squares (ESS): ∑ ̂ ̅ (4.28)

Residual Sum of Squares (RSS): ∑ (4.29)

TSS=ESS+RSS

∑ ̂ ̂
∑ ̂
(4.30)

The R2 indicates which proportion of the sample variation in is

explained by the model.

. The closer to one the better is the fit,

while closer to zero fit is becoming worse.
If ESS = 0, (RSS = TSS) model does not explain anything.

If RSS = 0, (ESS = TSS) meaning that the model

fit perfect.

Some source of variation in exploratory variable is much harder to

explain and can be incomparable. For example, models which explain
consumption, changes in consumption or consumption growth are
not comparable in terms of their R2.

R2 can be interpreted as a measure of quality of the model and its

linear approximation. It measures mainly linear approximation. The
value of R2 never decrease when the number of regressor increase
even if the added variables have no explanatory power on
regressand. A common way to deal with this problem is to correct the
80 Applied Econometrics with Eviews Applications

variance estimates for the degrees of freedom. This gives the value of
adjusted R2, or ̅ as follows:

∑ ̂ ̂ ∑
̅
̂
∑
∑ ̂

[ ] (4.31)

This new measure of goodness of fit punishes the inclusion of

additional explanatory or independent variables in the model. Even,
it may decline when a variable is added. ̅ is always smaller than R2.

As the number of independent variable in the regression model

increases, the ̅ become significantly smaller than R2. The impact of
adding more independent variables in the value of R2 is eliminated
through ̅ The R2 is always positive, but the values of ̅ can be
negative. Additionaly:

i. An increase in R2 or ̅ does not mean that an added

variable is statistically significant.
ii. A high R2 or ̅ does not mean that the regressors are the
true cause of the dependent variable. (still need a causal
model.)
iii. A high R2 or ̅ does not mean there is no omitted
variable bias (nor does low R2 or ̅ mean that there is
omitted variable bias).
iv. A high R2 or ̅ not necessarily mean that the regressors
are appropriate nor does a low R2 or ̅ mean the
regressors are inappropriate.

Example 4.8

̅ [ ]

It is possible to test whether the increase in R2 is statistically

significant. It is same as testing whether the coefficient of new added
 Regression 81

regressor is equal to zero. The appropriate F-statistics can be written

as follows:

(4.32)

Where, in the model with and without new

added regressor. J is the number of newly added regressors. It
measures the dispersion of the dependent variables estimates around
its mean.

Example 4.9

Table 4.5 Eviews output for multiplee regression model estimates

Dependent Variable: PRICE
Method: Least Squares
Sample: 1 546
Included observations: 546

Variable Coefficient Std. Error t-Statistic Prob.

LOTSIZE 4.871811 0.355481 13.70484 0.0000

BATHRMS 16285.39 1641.569 9.920629 0.0000
BEDROOMS 2802.653 1147.600 2.442186 0.0149
STORIES 5711.979 981.0879 5.822087 0.0000
AIRCO 13851.00 1702.051 8.137832 0.0000
C -932.5029 3424.701 -0.272287 0.7855

R-squared 0.586284 Mean dependent var 68121.60

Adjusted R-squared 0.582454 S.D. dependent var 26702.67
S.E. of regression 17254.69 Akaike info criterion 22.36048
Sum squared resid 1.61E+11 Schwarz criterion 22.40776
Log likelihood -6098.412 Hannan-Quinn criter. 22.37897
F-statistic 153.0489 Durbin-Watson stat 1.546388
Prob(F-statistic) 0.000000
82 Applied Econometrics with Eviews Applications
 Classıcal Lınear Regressıon Model (CLRM) 83

CHAPTER 5

CLASSICAL LINEAR REGRESSION

MODEL (CLRM)

LINEARITY
Linearity is the situation in which the relationship between the
dependent variable and the independent variables is linear. It is the
main assumption of the OLS estimation method. If linearity exists, the
effect of an exogenous variable on the endogenous variable is the
same for all values of the other exogenous variables in the model
(that is, the slope of the population regression function is constant,
so that the effect on y of a unit change in x does not depend on the
value of one or more exogenous variable). In this sense, non-linearity
has important consequences for the interpretation of the estimation
results.

The regression function can be nonlinear in two different cases:

i- The effect on endogenous variable of a change in

exogenous variable might be greater or smaller for
different values of the exogenous variable.
84 Applied Econometrics with Eviews Applications

ii- The effect on endogenous variable of a change in

exogenous variable depends on the value of another
exogenous variable. For example, the effect on test
scores of reducing the class durations may be greater
or lower in schools where the parents of children have
a higher level of education.

Nonlinear regression model may be:

 The regression functions of a nonlinear function of the

exogenous variables but is a linear function of the
unknown parameters.
 The regression functions of a nonlinear function of the
unknown parameters and may or may not be a
nonlinear function of the exogenous variables (logit,
probit, etc.).

Modeling Nonlinearities (with OLS)

 Based on the theory identify a nonlinear relationship

 Specify a nonlinear function (including a quadratic or an
interaction term) and estimate by OLS
 Compare two models and investigate whether or not the
nonlinear model improves upon a linear model
 Plot the nonlinear regression function
 Estimate the effect on endogenous variable of a unit change
in exogenous variable.
 Interpret the estimation results

Testing Linearity

The linearity of the regression can be examined:

i- Visually- by plots of the residuals against any of the

independent variables, or against the predicted values.
The points should be symmetrically distributed around
a diagonal line in the former plot or a horizontal line in
the latter plot. Look carefully for evidence of a
 Classıcal Lınear Regressıon Model (CLRM) 85

"bowed" pattern, indicating that the model makes

systematic errors whenever it is making unusually
large or small predictions.

Example 5.1
18,000

16,000

14,000

12,000
LOTSIZE

10,000

8,000

6,000

4,000

2,000

0
-60,000 -20,000 0 20,000 60,000 100,000

RESID01
Figure 5.1 Eviews Output for Lot Size Versus Predicted Residual

18,000

16,000

14,000

12,000
LOTSIZE

10,000

8,000

6,000

4,000

2,000

0
-60,000 -20,000 0 20,000 60,000 100,000

RESID01
Figure 5.2 Eviews Output for Kernel Density, Lot Size Versus
Predicted Residual
86 Applied Econometrics with Eviews Applications

Example 5.2

.48

.44

.40
CONSF

.36

.32

.28

.24
-.08 -.06 -.04 -.02 .00 .02 .04 .06 .08

RESID01
Figure 5.3 Eviews Output for Ice-Cream Consumption Predicted
Versus Residual

ii- Ramsey RESET test (Ramsey, 1969) may be

constructed by adding powers of fitted values to the
regression model. It is a test of linear specification
against a nonlinear specification. It tests the
hypothesis stating that the values of added parameters
are zero, such as:

̂ ̂ ̂ ̂ ̂ ̂ ̂ (5.1)

includes square fitted values.

The null hypothesis of linearity is

=0 (5.2)
 Classıcal Lınear Regressıon Model (CLRM) 87

which is tested by F-statistics. If F-statistics is larger

than its critical value, conclude that there is enough
evidence to reject the null hypothesis of linearity.

F-statistics is formed as follows;

̂ ̂
(5.3)
̂

SSR : sum of squared residuals

M : the number of restrictions
N : the number of observations
K : the number of explanatory variables

Example 5.3

Table 5.1 Eviews Output for Ramsey Reset Test

Ramsey RESET Test
Equation: UNTITLED
Specification: PRICE LOTSIZE BEDROOMS BATHRMS STORIES C
Omitted Variables: Powers of fitted values from 2 to 4

Value df Probability
F-statistic 0.339738 (3, 538) 0.7966
Likelihood ratio 1.033391 3 0.7932

F-test summary:
Sum of Sq. df Mean Squares
Test SSR 3.41E+08 3 1.14E+08
Restricted SSR 1.80E+11 541 3.34E+08
Unrestricted SSR 1.80E+11 538 3.35E+08
Unrestricted SSR 1.80E+11 538 3.35E+08

LR test summary:
Value df
Restricted LogL -6129.993 541
Unrestricted LogL -6129.476 538
88 Applied Econometrics with Eviews Applications

Unrestricted Test Equation:

Dependent Variable: PRICE
Method: Least Squares
Sample: 1 546
Included observations: 546

Variable Coefficient Std. Error t-Statistic Prob.

LOTSIZE 1.531674 19.47670 0.078641 0.9373

BEDROOMS 556.3758 10084.57 0.055171 0.9560
BATHRMS 5127.654 61702.43 0.083103 0.9338
STORIES 2241.466 27288.20 0.082140 0.9346
C 8644.588 85554.31 0.101042 0.9196
FITTED^2 1.64E-05 6.41E-05 0.256362 0.7978
FITTED^3 -1.42E-10 4.81E-10 -0.294963 0.7681
FITTED^4 3.96E-16 1.28E-15 0.310148 0.7566

R-squared 0.536426 Mean dependent var 68121.60

Adjusted R-squared 0.530394 S.D. dependent var 26702.67
S.E. of regression 18298.76 Akaike info criterion 22.48160
Sum squared resid 1.80E+11 Schwarz criterion 22.54464
Log likelihood -6129.476 Hannan-Quinn criter. 22.50624
F-statistic 88.93529 Durbin-Watson stat 1.484699
Prob(F-statistic) 0.000000

The reset F statistic is equal to 0.339 and the corresponding p-value is 0.7966. There
is no evidence to reject the null hypothesis of linearity at the 5% significance level.

The Ways to Correct Linearity Problem:

i- Using a nonlinear transformation to the dependent

and/or independent variables may be used if the
transformation is appropriate. For example, if the data
are strictly positive, a log transformation may be
feasible.

ii- Add another regressor which is a nonlinear function of

one of the other variables. For example, regress on
, and the graph of residuals versus predicted suggests
a parabolic curve, then it may make sense to regress
on both and (i.e., -squared). The latter
 Classıcal Lınear Regressıon Model (CLRM) 89

transformation is possible even when and/or have

negative values, whereas logging may not be.

HETEROSCEDASTICITY
for all observations. That is, the variance of the error
term is constant (Homoscedasticity) over the sample period. If the
error terms do not have constant variance, they are said to be
heteroscedastic.

Errors may increase as the value of an independent variable

increases. Annual family expenditures for education differ among rich
and poor families or educated and uneducated families. A research
may include two income group of families may face
heteroscedasticity problem. Measurement error can be also occurred
if some of respondant gives more accurate responses.

Heteroscedasticity is occurred, If;

- There are subpopulation differences or other interaction
effects. For instance, the effect of education in employment
differs for villagers and urban parts.
- There are model misspecifications. For instance, instead of
using , using the log of or instead of using , using .
- There are omitted variables. Omitting the important variables
from the model may cause bias. In the correctly specified
model, the patterns of heteroscedasticity are expected
disappear.

If the plot of residuals shows some uneven envelope of residuals, so

that the width of the envelope is considerably larger for some values
of than for others, a more formal test for heteroscedasticity should
be conducted.

H0 : the variance of the error term is constant (homoscedastic)

H1 : the variance of the error term is heteroscedastic
90 Applied Econometrics with Eviews Applications

Consequences

Heteroscedasticity does not result in unbiased parameter estimates.

OLS estimates are no longer BLUE. That is, among all of the unbiased
estimators, OLS does not provide the estimate with the smallest
variance. Depending on the nature of the heteroscedasticity,
significance tests can be too high or too low. The standard errors are
biased when heteroscedasticity is present. This in turn leads to bias in
test statistics and confidence interval.

Detection

i- It can be inspected through Visual Inspection. Do a

visual inspection of residuals plotted against fitted
values; or, plot the independent variables suspected to
be correlated with the variance of the error term.

Example 5.4
60,000

50,000

40,000

30,000

20,000

10,000

-10,000
71 72 73 74 75 76 77 78 79 80 81 82 83 84 85

CONSUMF RESID01
Figure 5.4 Eviews Output for Residuals Plotted Against Fitted Values
 Classıcal Lınear Regressıon Model (CLRM) 91

40

30

20

10

-10

-20
250 500 750 1000 1250

WAGEF RESID01

Figure 5.5 Eviews Output for Residuals Plotted Against Fitted Values

ii- White Test for Heteroscedasticity (White, 1980)

It is actually a special case of Breusch-Pagan test. It

involves an auxiliary regression of squared residuals,
but excludes any higher order terms. It is very general.
If the number of the observation is small, power of the
white test becomes weak. It can be performed by
obtaining least squares residuals and modeling the
square residuals as a multiple regression which
includes independent variables and their squares and
second degree products (interaction term) as follows:

(5.4)

White test for heteroscedasticity is performed through

the null hypothesis of no heteroscedasticity. Instead of
F statistics, the statistics NR2 (sample size multiplied by
the coefficient of determination) is in use.

(5.5)
92 Applied Econometrics with Eviews Applications

Example 5.5

Table 5.2 Eviews Output for White Test for Heteroscedasticity

Heteroskedasticity Test: White

F-statistic 9.564407 Prob. F(2,161) 0.0001

Obs*R-squared 17.41601 Prob. Chi-Square(2) 0.0002
Scaled explained SS 13.63430 Prob. Chi-Square(2) 0.0011

Dependent Variable: RESID^2

Method: Least Squares
Sample: 1954Q1 1994Q4
Included observations: 164

Variable Coefficient Std. Error t-Statistic Prob.

C 0.009461 0.011185 0.845830 0.3989

INCOME -0.002690 0.002937 -0.915986 0.3610
INCOME^2 0.000194 0.000192 1.006889 0.3155

R-squared 0.106195 Mean dependent var 0.000256

Adjusted R-squared 0.095092 S.D. dependent var 0.000326
S.E. of regression 0.000310 Akaike info criterion -13.30390
Sum squared resid 1.54E-05 Schwarz criterion -13.24720
Log likelihood 1093.920 Hannan-Quinn criter. -13.28088
F-statistic 9.564407 Durbin-Watson stat 1.050351
Prob(F-statistic) 0.000119

iii- The Breusch-Pagan Test

Proposed by Breusch and Pagan (1980). It is a Lagrange

multiplier test for heteroscedasticity and designed to
detect any linear form of heteroscedasticity.

It tests the null hypothesis that the error variances are

all equal versus the alternative that the error variances
are a multiplicative function of one or more variables.
 Classıcal Lınear Regressıon Model (CLRM) 93

Example 5.6

Table 5.3 Eviews Output for the Breush Pagan

Heteroskedasticity Test: Breusch-Pagan-Godfrey
F-statistic 29.82872 Prob. F(3,1468) 0.0000
Obs*R-squared 84.57450 Prob. Chi-Square(3) 0.0000
Scaled explained SS 450.4796 Prob. Chi-Square(3) 0.0000

Test Equation:
Dependent Variable: RESID^2
Method: Least Squares
Sample: 1 1472
Included observations: 1472

Variable Coefficient Std. Error t-Statistic Prob.

C -27.03619 4.356212 -6.206352 0.0000

MALE 4.711490 2.170099 2.171094 0.0301
EXPER 0.716098 0.107900 6.636675 0.0000
EDUC 7.224143 0.907955 7.956499 0.0000

R-squared 0.057456 Mean dependent var 12.55743

Adjusted R-squared 0.055529 S.D. dependent var 41.11146
S.E. of regression 39.95372 Akaike info criterion 10.21603
Sum squared resid 2343368. Schwarz criterion 10.23042
Log likelihood -7515.001 Hannan-Quinn criter. 10.22140
F-statistic 29.82872 Durbin-Watson stat 1.991852
Prob(F-statistic) 0.000000

iv- Ramsey’s Test for Heteroscedasticity

RESET stands for Regression Specification Error Test

and was proposed by Ramsey (1969). The classical
normal linear regression model is specified as:

(5.6)

Where disturbance vector e is presumed to follow the

multivariate normal distribution
Specification error is a comprehensive term which
94 Applied Econometrics with Eviews Applications

covers departures the maintained model assumptions.

Serial correlation, heteroscedasticity, and non-
normality violate error term assumptions which
are .

RESET specification test is a general test for the

following types of specification errors:
a. Omitted variable bias (if set of independent
variables do not include all relevant variables).

b. Wrong functional form (if some or all of the

dependent and independent variables should
be transformed to logs, powers, reciprocals,
or in some other way, but not).

c. Correlation between independent variable

and error term (may be caused measurement
error in simultaneity).

d. The existence of lagged dependent variable

values.

e. Serially correlated error terms.

Under those specification errors above, OLS estimators

is biased and inconsistent. Conventional inference
procedures are not valid. Ramsey (1969) indicates that
any or all of these specification errors above produce a
non-zero mean vector.

The null and alternative hypotheses of the RESET test

are:

The test is based on following augmented regression

equation:

(5.7)
 Classıcal Lınear Regressıon Model (CLRM) 95

Specification error test examines the restriction of

. The main question in constructing the RESET
test is to identify the variables which should be in the Z
matrix. The Z matrix can include the variables which
are not in the original specification, so that the test
of is called as the omitted variables test.

Example 5.7

Table 5.4 Eviews Output for the Ramsey´s Test for Heteroscedasticity
Ramsey RESET Test
Equation: UNTITLED
Specification: CONS INCOME PRICE TEMP TIME C
Omitted Variables: Powers of fitted values from 2 to 3

Value df Probability
F-statistic 5.331333 (2, 23) 0.0125
Likelihood ratio 11.42686 2 0.0033

F-test summary:
Mean
Sum of Sq. df Squares
Test SSR 0.010957 2 0.005478
Restricted SSR 0.034590 25 0.001384
Unrestricted SSR 0.023634 23 0.001028
Unrestricted SSR 0.023634 23 0.001028

LR test summary:
Value df
Restricted LogL 58.91245 25
Unrestricted LogL 64.62588 23

Unrestricted Test Equation:

Dependent Variable: CONS
Method: Least Squares
Sample: 1 30
Included observations: 30
96 Applied Econometrics with Eviews Applications

Variable Coefficient Std. Error t-Statistic Prob.

INCOME 0.036697 0.021613 1.697907 0.1030

PRICE -19.02496 12.82023 -1.483980 0.1514
TEMP 0.060590 0.038732 1.564338 0.1314
TIME 0.017399 0.012889 1.349847 0.1902
C 3.614740 2.410354 1.499672 0.1473
FITTED^2 -53.50632 32.01791 -1.671137 0.1082
FITTED^3 54.47286 29.16295 1.867879 0.0746

R-squared 0.811717 Mean dependent var 0.359433

Adjusted R-squared 0.762599 S.D. dependent var 0.065791
S.E. of regression 0.032056 Akaike info criterion -3.841725
Sum squared resid 0.023634 Schwarz criterion -3.514779
Log likelihood 64.62588 Hannan-Quinn criter. -3.737132
F-statistic 16.52607 Durbin-Watson stat 1.063651
Prob(F-statistic) 0.000000

v- Test for ARCH Effect

It is common for financial variables. Engel (1982)

detected that large and small forecast errors tend to
occur in clusters so that the conditional variance of
error term is the autoregressive function of the past
errors. Ignoring ARCH effect can result in inefficiency
of the estimation. ARCH(q) effect can be written as
follows:

(5.8)

This is a test for ARCH (q) versus ARCH (0).

 Classıcal Lınear Regressıon Model (CLRM) 97

Example 5.8

Table 5.5 Eviews Output for the ARCH Effect

Heteroskedasticity Test: ARCH
F-statistic 0.373376 Prob. F(1,27) 0.5463
Obs*R-squared 0.395563 Prob. Chi-Square(1) 0.5294

Test Equation:
Dependent Variable: RESID^2
Method: Least Squares
Sample (adjusted): 2 30
Included observations: 29 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

C 0.000879 0.000326 2.700516 0.0118

RESID^2(-1) 0.111842 0.183033 0.611045 0.5463

R-squared 0.013640 Mean dependent var 0.000993

Adjusted R-squared -0.022892 S.D. dependent var 0.001426
S.E. of regression 0.001442 Akaike info criterion -10.17897
Sum squared resid 5.62E-05 Schwarz criterion -10.08467
Log likelihood 149.5950 Hannan-Quinn criter. -10.14943
F-statistic 0.373376 Durbin-Watson stat 1.456000
Prob(F-statistic) 0.546280

vi- Goldfeld-Quant (GQ) Test

The GQ test (GQ, 1965) can be used when it is believed

that the variance of the error term increases
consistently or decreases consistently as increases.
This test is commonly used because it is easy to apply
when one of the regressors (or another r.v.) is
considered to possess proportionality factor of
heteroscedasticity. The test has two limits: its difficult
to reject the null hypothesis of homoscedasticity and
the fact that it do not allow to verify other forms of
heteroscedasticity. This test is based on the hypothesis
that the error variance is related to a regressor .
98 Applied Econometrics with Eviews Applications

The test procedure is the following:

a) The observations on and are sorted

following the ascending order of the regressor
.

b) The sample observations are divided into

three subsamples omitting the central one.

c) The regression models are estimated through

OLS on the first and third subsample (then on
(n-c)/2 observations; the number of
observations considered has to be sufficiently
large.

d) The relative RSS is calculated, denoted as RSS1

and RSS2 and derived the Goldfeld-Quandt
test as follows:

GQ = R = RSS2/RSS1 (5.9)

If the sample value of the F statistics is greater than

the critical value at the chosen significance level, the
null hypothesis of homoscedasticity is rejected. If R is
large, then RSS2 is greater that RSS1, which means that
residuals increase with the regressor.

The power of this test depends on the number of

omitted observations (usually n/3 observations have to
be omitted). If too many observations are excluded,
the RSS2 and RSS1 will take too low degrees of
freedom, if too few observations are excluded, the test
power is low because the comparison between RSS2
and RSS1 becomes less effective.

vii- Park Test

This test procedure involves three steps:

 Classıcal Lınear Regressıon Model (CLRM) 99

a) Modeling OLS estimation to derive the OLS

residuals,
, (5.10)

OLS estimates produce following OLS residuals;

̂ ̂ ̂ ̂ (5.11)

b) The derivation of the which are

considered as dependent variable in the
regression;

(5.12)

where the only regressor is the log of the real

value considered as proportionality factor.

c) The estimation results of the model are used to

verify the presence of heteroscedastic errors.

viii- Glesjer LM Test

The Glejser (1969) test is similar to the Breusch-Pagan-

Godfrey test. This test tests against an alternative
hypothesis of heteroscedasticity. The auxiliary
regression that Glejser proposes regresses the
absolute value of the residuals from the original
equation. As with the previous tests, this statistic is
distributed from a chi-squared distribution with
degrees of freedom equal to the number of variables.

ix- Harvey-Godfrey LM Test

The Harvey (1976) test for heteroscedasticity is similar

to the Breusch-Pagan-Godfrey test. Harvey tests a null
hypothesis of no heteroscedasticity against
heteroscedasticity of the form of
100 Applied Econometrics with Eviews Applications

, (5.13)

where, again is a vector of independent variables.

Example 5.9

Table 5.6 Eviews Output for the Glesjer LM Test

Heteroskedasticity Test: Glejser
F-statistic 1.415749 Prob. F(4,25) 0.2578
Obs*R-squared 5.540550 Prob. Chi-Square(4) 0.2362
Scaled explained SS 5.568536 Prob. Chi-Square(4) 0.2338

Test Equation:
Dependent Variable: ARESID
Method: Least Squares
Sample: 1 30
Included observations: 30

Variable Coefficient Std. Error t-Statistic Prob.

C 0.237778 0.194819 1.220507 0.2337

INCOME -0.000224 0.001399 -0.160103 0.8741
PRICE -0.737679 0.506247 -1.457151 0.1575
TEMP 0.000317 0.000287 1.105262 0.2796
TIME -0.000379 0.000936 -0.404613 0.6892

R-squared 0.184685 Mean dependent var 0.025450

Adjusted R-squared 0.054235 S.D. dependent var 0.022864
S.E. of regression 0.022235 Akaike info criterion -4.623290
Sum squared resid 0.012360 Schwarz criterion -4.389757
Log likelihood 74.34934 Hannan-Quinn criter. -4.548580
F-statistic 1.415749 Durbin-Watson stat 1.380946
Prob(F-statistic) 0.257755
 Classıcal Lınear Regressıon Model (CLRM) 101

Example 5.10

Table 5.7 Eviews Output for the Harvey-Godfrey Test

Heteroskedasticity Test: Harvey
F-statistic 1.026380 Prob. F(4,25) 0.4130
Obs*R-squared 4.231692 Prob. Chi-Square(4) 0.3756
Scaled explained SS 6.189628 Prob. Chi-Square(4) 0.1854

Test Equation:
Dependent Variable: LRESID2
Method: Least Squares
Sample: 1 30
Included observations: 30

Variable Coefficient Std. Error t-Statistic Prob.

C 0.490030 23.89891 0.020504 0.9838

INCOME -0.103229 0.171591 -0.601596 0.5529
PRICE -4.340469 62.10253 -0.069892 0.9448
TEMP 0.025815 0.035195 0.733499 0.4701
TIME -0.023174 0.114794 -0.201871 0.8417

R-squared 0.141056 Mean dependent var -8.529700

Adjusted R-squared 0.003625 S.D. dependent var 2.732573
S.E. of regression 2.727615 Akaike info criterion 4.995744
Sum squared resid 185.9971 Schwarz criterion 5.229277
Log likelihood -69.93616 Hannan-Quinn criter. 5.070453
F-statistic 1.026380 Durbin-Watson stat 1.543329
Prob(F-statistic) 0.413020

Which Test to Apply?

To choose the wrong test may not detect presence of

heteroscedasticty. The most general test is the White test. But, it has
limited power against large number of alternatives. In most cases
visual inspections of the residuals can be the best method to detect
heteroscedasticty correctly.
102 Applied Econometrics with Eviews Applications

The Ways to Correct Heteroscedasticity Problem

A common way to alleviate this problem is to use logarithms of all

variables rather than their levels. Consequently, our first step in
handling the heteroscedasticity problem is to consider a log linear
model.

As noted before, sometimes heteroscedasticity results from improper

model specification. There may be subgroup differences. Effects of
variables may not be linear. Perhaps some important variables have
been left out of the model. If these represent a problem then deal
with them first.

Using Weighted Least Squares is a more difficult option but superior

when you can make it work right. It uses weighted least squares.
Generalized Least Squares (GLS) is a technique that will always yield
estimators that are BLUE when either heteroscedasticity or serial
correlation is present.

AUTOCORRELATION
Autocorrelation can only occur in the models that include time series
data and it means that either the model is specified with an
insufficient number of lagged variables or not all the relevant
explanatory variables are specified in the model.

The error term catch the influence of the not included variables
affecting dependent variable. Persistence effect of the excluded
variables causes positive autocorrelation. If those excluded variables
are observable and includable in the model, autocorrelation test
result is an indication of a misspecification model.

Autocorrelation test is also regarded as misspecification test.

Incorrect functional forms, omitted variables and an inadequate
dynamic specification of the model can cause autocorrelation.
 Classıcal Lınear Regressıon Model (CLRM) 103

Consequences of Autocorrelation in the Residuals

i- The standard errors are underestimated, so t-values

are overestimated.

ii- High values for the t-statistics and R2 are observed in

the estimation output. It means that the result is false
if the output is not correctly interpreted

iii- OLS estimates remain unbiased, but it becomes

inefficient.

First Order Autocorrelation

The most popular form of the autocorrelation is the first-order

autoregressive process;

̀ (5.14)

Error term is assumed to depend on its previous value as follows;

(5.15)

where, is an error term with mean zero and constant variance ,

which exhibits no serial correlation. It means that the value of the
error term in any observation is equal to times its value in the
previous observation plus a new error term. If then
there is no residual autocorrelation and autocorrelation condition of
OLS (Gauss-Markov) is satisfied.

If | | then the first autoregressive process is stationary. It

means that the mean, variance and covariance of do not change
over time.

satisfies the Gauss-Markov conditions. The transformation like

will generate homoscedastic non autocorrelated errors.
104 Applied Econometrics with Eviews Applications

That is all observations should be transformed as

.18

If no autocorrelation is present and OLS is BLUE. If OLS

estimator will be misleading because standard errors will be based on
the wrong formula.

There are various autocorrelation tests. All tests have same null
hypothesis of absence of autocorrelation in the disturbance term.
The tests have different alternative hypothesis because of differences
of the order of the autocorrelation.

The existence of autocorrelation may be an indication of

misspecification. A possible way to eliminate autocorrelation
problem is to change model specification.

For example;

H0: no autocorrelation in the disturbance term

(5.16)

(the disturbance term has autocorrelation of the pth order)

With

Detection

i- Graphical method

Plotting error term to detect autocorrelation

18
Verbeek, M. (2004).
 Classıcal Lınear Regressıon Model (CLRM) 105

RESID01
2,000

1,500

1,000

500

-500

-1,000

-1,500
50 100 150 200 250 300 350 400 450 500 550
Figure 5.6 Eviews Output for Residual Plot

RESIDUAL
100

80

60

40

20

-20

-40

-60

-80
1960 1965 1970 1975 1980 1985 1990 1995

Figure 5.7 Eviews Output for Residual Plot

106 Applied Econometrics with Eviews Applications

ii- The Breusch-Godfrey LM-test

Breusch(1978) and Godfrey (1978) developed this

autocorrelation test. The idea behind the Breusch-
Godfrey test is as follows:

Suppose that the following linear model is estimated

with OLS.

(5.17)

Next residuals are computed and following equation is

estimated with OLS.

(5.18)

Null hypothesis (no

autocorrelation)

In Eviews BG test is called the serial correlation LM

test. Eviews computes and F-statistics to test that all
the are zero.
 Classıcal Lınear Regressıon Model (CLRM) 107

Example 5.11

Table 5.8 Eviews Output the Breusch-Godfrey LM-Test .

Breusch-Godfrey Serial Correlation LM Test:
F-statistic 1.814271 Prob. F(3,22) 0.1740
Obs*R-squared 5.949988 Prob. Chi-Square(3) 0.1141

Test Equation:
Dependent Variable: RESID
Method: Least Squares
Included observations: 30
Presample missing value lagged residuals set to zero.

Variable Coefficient Std. Error t-Statistic Prob.

PRICE -0.328497 0.848957 -0.386942 0.7025

TEMP -0.000742 0.000829 -0.894623 0.3807
TIME 0.001529 0.001950 0.784184 0.4413
INCOME -0.002193 0.002959 -0.741184 0.4664
C 0.290480 0.381957 0.760505 0.4550
RESID(-1) 0.516004 0.239559 2.153970 0.0425
RESID(-2) 0.064238 0.325189 0.197540 0.8452
RESID(-3) 0.176768 0.310595 0.569127 0.5750

R-squared 0.198333 Mean dependent var 4.67E-17

Adjusted R-squared -0.056743 S.D. dependent var 0.034537
S.E. of regression 0.035503 Akaike info criterion -3.615225
Sum squared resid 0.027730 Schwarz criterion -3.241572
Log likelihood 62.22838 Hannan-Quinn criter. -3.495690
F-statistic 0.777545 Durbin-Watson stat 1.571966
Prob(F-statistic) 0.612620

iii- The Box-Pierce and the Ljung-Box tests (Q test) (1970)

The Box-Pierce and the Ljung-Box test have asymptotic

distribution, with p degrees of freedom under the null
hypothesis of no autocorrelation. This test uses
autocorrelation of the residuals. The estimated
autocorrelation coefficients are defined as ̂ .
̂ (5.19)
√ √
108 Applied Econometrics with Eviews Applications

The theoretical autocorrelation coefficients ̂ are zero

under the null hypothesis. The Q-test does not look at the
individual autocorrelation coefficients. It considers the sum
of a number of squared autocorrelation coefficients as
follows:

∑ ̂ (5.20)

but this test has low power of detecting autocorrelation.

The main difference between Q-test and BG test is that
one specific order of the autoregressive process specified
should be chosen under the alternative hypothesis.

Example 5.1219

Table 5.9 Eviews Output for the Box-Pierce and the Ljung-Box Tests (Q test)
Included observations: 30

Autocorrelation Partial Correlation AC PAC Q-Stat Prob

. |*** | . |*** | 1 0.367 0.367 4.4564 0.035

. |* . | . | . | 2 0.104 -0.036 4.8264 0.090
. | . | . | . | 3 0.013 -0.015 4.8326 0.184
. *| . | . *| . | 4 -0.134 -0.150 5.4918 0.240
***| . | .**| . | 5 -0.370 -0.318 10.759 0.056
.**| . | . | . | 6 -0.241 -0.003 13.074 0.042
.**| . | .**| . | 7 -0.281 -0.226 16.366 0.022
. *| . | . | . | 8 -0.171 -0.021 17.635 0.024
. *| . | . *| . | 9 -0.094 -0.115 18.043 0.035
. *| . | .**| . | 10 -0.171 -0.331 19.452 0.035
. | . | . | . | 11 -0.051 -0.038 19.585 0.051
. |* . | . | . | 12 0.206 0.059 21.845 0.039
. |*** | . |**. | 13 0.365 0.255 29.365 0.006
. |* . | . *| . | 14 0.199 -0.133 31.745 0.004
. |* . | . *| . | 15 0.105 -0.204 32.455 0.006
. |* . | . |* . | 16 0.129 0.093 33.601 0.006

19
Vogelvang,B. (2005)
 Classıcal Lınear Regressıon Model (CLRM) 109

iv- The Durbin Watson test (1950)

This is the most known autocorrelation test. This test is used for first
order autocorrelation.

DW test assumes that error term is stationary and normally

distributed with zero mean. It tests the null hypothesis H0 that the
errors are uncorrelated.

Test the H0: no autocorrelation versus H1:first order residual

autocorrelation.

This test can be used if the explanatory variables are exogenous and
a constant term has been included in the model. DW –statistics is not
used, if lagged dependent variables are present as explanatory
variables in the model. It can be employed if the explanatory
variables are exogenous and the model includes intercept. DW-
statistics should be used if all the conditions are satisfied. Otherwise
it is more informative to use Breusch-Godfrey test in the research
paper.

DW –statistic has been defined as:

∑
∑
(5.21)

Properties of the DW-statistics;

DW-statistics cannot be tabulated. But it is possible to drive

distributions of a lower (dL) and an upper (dU) bound. These two
distributions depend on the number of the observations (n) and the
number of the explanatory variables (K). Therefore DW–statistics is
tabulated as follows:
110 Applied Econometrics with Eviews Applications

Table 5.10 Lower and Upper Bounds for 5% Critical Values of the DW Test20
Number of regressors including intercept
Number of K=3 K=5 K=7 K=9
Observations dL dU dL dU dL dU dL dU
25 1.206 1.550 1.038 1.767 0.868 2.012 0.702 2.280
50 1.462 1.628 1.378 1.721 1.291 1.822 1.201 1.930
75 1.571 1.680 1.515 1.739 1.458 1.801 1.399 1.867
100 1.634 1.715 1.592 1.758 1.550 1.803 1.506 1.850
200 1.748 1.789 1.728 1.810 1.707 1.831 1.686 1.852

If there is autocorrelation. OLS is no longer BLUE, then EGLS

(Cochrane-Orcutt) method can be used.

DW test for autocorrelation has some shortcomings such as;

a) The form of model (explanatory variables) should be known
b) The test result is sometimes inconclusive

Example 5.13 Ice-cream

Table 5.11 Eviews Output for the Durbin Watson Test

Dependent Variable: CONS
Method: Least Squares
Sample: 1 30
Included observations: 30

Variable Coefficient Std. Error t-Statistic Prob.

INCOME 0.001890 0.002340 0.807659 0.4269

PRICE -1.104229 0.846905 -1.303841 0.2042
TEMP 0.003341 0.000480 6.961297 0.0000
TIME 0.001099 0.001565 0.702250 0.4890
C 0.322449 0.325914 0.989367 0.3320

R-squared 0.724430 Mean dependent var 0.359433

Adjusted R-squared 0.680339 S.D. dependent var 0.065791
S.E. of regression 0.037197 Akaike info criterion -3.594163
Sum squared resid 0.034590 Schwarz criterion -3.360630
Log likelihood 58.91245 Hannan-Quinn criter. -3.519454
F-statistic 16.43025 Durbin-Watson stat 0.947014
Prob(F-statistic) 0.000001

20
Savin, N.E. & White, K.J. (1977)
 Classıcal Lınear Regressıon Model (CLRM) 111

The Ways to Correct Autocorrelation Problem

An autocorrelation is an indicator of model misspecification. If the

model misspecification occurs then change the model not the
estimation method (from OLS to EGLS). There are three types of
model misspecifications such as:

 Omitted variable bias- excluding relevant independent

variables can cause autocorrelation. Find out the relevant
omitted variable and include in the model
 Functional form; to use log transformation can eliminate
autocorrelation problem
 Dynamic misspecification; whether the model is static or
dynamic should be decided. Inclusion of the lagged
endogenous and exogenous variables eliminate the
autocorrelation problem.

MULTICOLLINEARITY
Multicollinearity is a data problem. Collinearity between variables is
always present. It becomes a problem and violation of the classical
assumptions if the correlations among the independent variables are
very strong. It can affect accuracy of the parameter estimates.
Multicollinearity misleadingly increases the standard errors. Thus, it
makes some variables statistically insignificant while they should be
otherwise significant. It is like two or more people singing loudly at
the same time. One cannot discern who is singing how. The influence
of the independents offset each other. In multicollinearity, the
parameter estimates become inaccurate.

To use too many dummy variables may also cause multicollinearity.

Consequences

The precision of estimation falls because of very big errors that may
be highly correlated errors and very large sampling variances of the
coefficients. Large standard errors can be caused by things besides
multicollinearity. If two independent variables are highly and
112 Applied Econometrics with Eviews Applications

positively correlated, then their slope coefficient estimators tend to

be highly and negatively correlated. If the effect of one parameter is
overestimated, then the effect of the other parameter is
underestimated. Hence, coefficient estimates tend to vary from one
sample to the next.

When high multicollinearity is present, confidence intervals for

coefficients tend to be very wide and t-statistics tend to be very
small. Coefficients have to be larger in order to be statistically
significant. It is very difficult to reject the null hypothesis when
multicollinearity is present.

If the objective is only to predict endogenous variable from a set of

exogenous variables, then multicollinearity is not a problem. The
predictions can be accurate, and the overall R2 (or adjusted R2)
quantifies how well the model predicts the y values.

If the objective is to understand how the various x variables impact y

over observations, then multicollinearity can be a big problem. One
problem is that the individual p values can be misleading (a p value
can be high, even though the variable is important). The second
problem is that the confidence intervals on the regression
coefficients can be very wide. The confidence intervals may even
include zero, which means you can’t even be confident whether an
increase in the x value is associated with an increase, or a decrease,
in y. Because the confidence intervals are so wide, excluding a
subject (or adding a new one) can change the coefficients and may
even change their signs.

Causes of Multicollinearity

i- Improper use of dummy variables

ii- Including a variable computed from other variables in

the equation (e.g. family income = husband’s income +
wife’s income, and the regression includes all three
income measures)
 Classıcal Lınear Regressıon Model (CLRM) 113

iii- Including the same or almost the same variable twice

(height in feet and height in inches; or, more
commonly, two different operationalization of the
same identical concept)

Detection

 Variance inflation factors (VIF) is used to measure how much

the variance of the estimated coefficients is increased over
the case of no correlation among the independent variables.

 If VIF = 0, No multicollinearity
 If VIF ≥ 0, There is multicollinearity

 Regression coefficients change drastically while adding or

deleting an x variable.
 A regression coefficient is negative when theoretically
endogenous variable increase with increasing values of that
exogenous variable, or the regression coefficient is positive
when theoretically endogenous variable decrease with
increasing values of that exogenous variable.
 Collinearity diagnostics less than 0.2 indicates high
collinearity.
 None of the individual coefficients has a significant t-statistic,
but the overall F-test is significant.
 A regression coefficient has a non-significant t-statistic, even
though on theoretical grounds that exogenous variable should
provide substantial information about endogenous variable.
 High R2 (larger than .75) but only a few significant t-values.
 High correlation coefficients between estimated values in the
correlation matrix indicate possible multicollinnearity (≥ .80).
 Large standard errors and low t-statistic values
 Wrong signs of parameter estimates
 Whether the coefficients are sTable should be checked when
different samples are used. Sample period may be divided
into subsamples randomly. If the coefficients differ
dramatically, multicollinearity may be a problem.
 Using the same data do a slightly different specification of a
model. Observe the changes.
114 Applied Econometrics with Eviews Applications

 In particular, as variables are added, changes in the signs of

effects should be considered (e.g. switches from positive to
negative) which seems theoretically questionable. If there are
suppressor effects, such changes may be logical, otherwise
they may indicate multicollinearity.

The Ways to Handle Multicollinearity

 The sample size can be increased since when sample

size is increased, standard error decreases (when all
other things are equal). This mitigates the high
multicollinearity problem which causes high standard
errors of coefficients.
 The highly inter-correlated variable(s) may be omitted
from analysis (generally if larger than 0.80). This
method is misguided if the variables were there due to
the theory of the model, and they should have been.
2
Remove each variable individually and observe R . The
2
model with higher R should be chosen.
 Combine variables into a composite variable through
building indexes (creating an index theoretical and
empirical reason to justify this action).
 Transform the offending independent variables by
subtracting the mean from each case. The resulting
centered data may well display considerably lower
multicollinearity. Interpretations of b and beta must
be changed accordingly.
 Information from prior researches can be used
 The model may be re specified. Intercorrelated
variables may be omitted from analysis but substitute
their cross product as an interaction term or in some
other way combine the intercorrelated variables. If a
correlated variable is a dummy variable, other
dummies in that set should also be included in the
combined variable in order to keep the set of dummies
conceptually together. To do this, compute the mean
of each independent variable, and then replace each
value with the difference between it and the mean.
For example, if the variable is weight and the mean is
 Classıcal Lınear Regressıon Model (CLRM) 115

72, then enter "6" for a weight of 78 and "-3" for a

weight of 69.
 Assign the common variance to each of the covariates
by some probably arbitrary procedure. Treat the
common variance as a separate variable and
decontaminate each covariate by regressing them on
the others and using the residuals. That is, analyze the
common variance as a separate variable.

Example 5.14

Table 5.12 SPSS Outputs for Variance Inflation Factor

Coefficientsa
Model Collinearity
Statistics
Tolerance VIF
1 Nbedroom .768 1.301
Nbath .825 1.212
Nstories .799 1.251
a. Dependent Variable: lot size

Coefficientsa
Model Collinearity
Statistics
Tolerance VIF
1 Nbedroom .854 1.171
Nbath .841 1.189
lot size .955 1.047
a. Dependent Variable: Nstories
Coefficientsa
Model Collinearity
Statistics
Tolerance VIF
1 Nbedroom .820 1.220
lot size .976 1.024
Nstories .833 1.200
a. Dependent Variable: Nbath
116 Applied Econometrics with Eviews Applications

Coefficientsa
Model Collinearity
Statistics
Tolerance VIF
1 lot size .962 1.040
Nstories .895 1.118
Nbath .867 1.153
a. Dependent Variable: Nbedroom

EXOGENEITY
There are two types of variables in macro-econometric models:
endogenous and exogenous. Endogenous variables are explained by
the equations, either the stochastic or the identities. Exogenous
variables are determined outside of the model not explained within
the model. They are taken as given.

The concept of exogeneity has been analyzed and elaborated by

Engle, Hendry and Richard (1983). All explanatory variables should be
uncorrelated with the error term. It means that explanatory variables
have to be determined outside of the model (they are exogenous).

| (5.22)

Suppose we have the following model, where are

positively correlated.

(5.23)

If we are trying to find out the relationship between the price of a

sish kebab and the quantity sold across a wide variety of Turkish
restaurants.

, (5.24)
 Classıcal Lınear Regressıon Model (CLRM) 117

Together with the price, it is obvious that the quality has also
significant impact on sales quantity. Also, there is very strong positive
relationship between price of sish kebab and the quality. Therefore
are highly correlated.

NORMALITY
Normality test was introduced by Fisher (1948), but, a standard
Jarque-Bera (JB, 1980) test is being used widely. It is expressed in
terms of the third and fourth moments of the disturbances, as
follows:

The null hypothesis is H0 and the residuals are normally distributed.

( ) (5.25)
S: measure of skewness

(5.26)

K: measure of Kurtosis

(5.27)

The values for K and S when the variable is normally distributed:

K=3 and S=0

Example 5.15
600
Series: Residuals
Sample 1 1472
500 Observations 1472

400 Mean -5.11e-15

Median -0.312358
Maximum 30.70154
300 Minimum -13.52944
Std. Dev. 3.544851
200
Skewness 1.615750
Kurtosis 11.71098

100 Jarque-Bera 5294.525

Probability 0.000000
0
-15 -10 -5 0 5 10 15 20 25 30

Figure 5.8 Eviews Output for Normality

118 Applied Econometrics with Eviews Applications

8
Series: Residuals
7 Sample 1 30
Observations 30
6
Mean 2.04e-17
5 Median 0.003040
Maximum 0.076192
4 Minimum -0.060326
Std. Dev. 0.034537
3 Skewness 0.065254
Kurtosis 2.988689
2
Jarque-Bera 0.021450
1 Probability 0.989332

0
-0.06 -0.04 -0.02 0.00 0.02 0.04 0.06 0.08

Figure 5.9 Eviews Output for Normality

ii- QQ and PP plots

Plotting the empirical distribution of residuals against

the normal counterparts can reveal normality
problem.

Example 5.16
WAGE MALE
30 2.4

2.0
20
Quantiles of Normal

Quantiles of Normal

1.6
10
1.2

0
0.8

-10 0.4
0 10 20 30 40 50 0 1 2

Quantiles of WAGE Quantiles of MALE

EXPER EDUC
60 8

7
40
Quantiles of Normal

Quantiles of Normal

5
20
4

3
0
2

-20 1
0 10 20 30 40 50 0 1 2 3 4 5 6

Quantiles of EXPER Quantiles of EDUC

Figure 5.10 Eviews Output for QQ and PP Plots

 Classıcal Lınear Regressıon Model (CLRM) 119

100

80
Empirical Quantiles

60

40

20

0
.24 .28 .32 .36 .40 .44 .48 .52 .56

Quantiles of CONS

INCOME PRICE
TEMP TIME
Figure 5.11 Eviews Output for QQ and PP Plots

If normality is rejected, the test gives nothing about how to correct

this deviation from normality. The best way is to look at the plot of
the residuals. It can be caused by outliers in the residuals. These
outliers can be eliminated through using a dummy variable in the
regression equation.

What to do if data is not normal:

i- Increase the sample size

ii- Transform data (log values). Misspecification and

functional form can spoil normality assumptions.
120 Applied Econometrics with Eviews Applications

Example 5.17
LNWAGE LNEXPER
1.6 1.2
Distance to points above median

Distance to points above median

1.0
1.2
0.8

0.8 0.6

0.4
0.4
0.2

0.0 0.0
0.0 0.4 0.8 1.2 1.6 0.0 0.5 1.0 1.5 2.0 2.5 3.0

Distance to points below median Distance to points below median

LNEDUC MALE
.6 1
Distance to points above median

Distance to points above median

.5

.4

.3 0

.2

.1

.0 -1
0.0 0.2 0.4 0.6 0.8 1.0 1.2 0 1 2

Distance to points below median Distance to points below median

Figure 5.12 Eviews Output for QQ and PP Plots with Logged Data

240
Series: Residuals
Sample 1 1472
200 Observations 1472

160 Mean -4.87e-15

Median 0.004848
Maximum 1.118146
120 Minimum -1.704765
Std. Dev. 0.286360
80
Skewness -0.390370
Kurtosis 5.574270

40 Jarque-Bera 443.8339
Probability 0.000000
0
-1.5 -1.0 -0.5 0.0 0.5 1.0

Figure 5.13 Eviews Output for Normality with Logged Data

 Classıcal Lınear Regressıon Model (CLRM) 121

CONSISTENCY
Consistency is a large sample property. If there are more
observations, the probability that the estimator is some positive
number far from the true value becomes smaller. A minimum
requirement for an estimator to be useful appears to be that it is
consistent. OLS estimator is consistent if it holds that:

∑ ̀ (5.28)

The variance of b decreases as the sample size increases.

{ } (5.29)

It says that the error term is mean zero and uncorrelated with any of
the explanatory, independent variables. The consistency property
implies that the estimator ̂ itself does not have an asymptotic
distribution. Sufficient condition for a consistent estimator is that the
bias and the variance should both tend to be zero as the sample size
increases. The specification of the linear regression model is as
follows:

(5.29)

Deterministic assumptions:

i- is endogenous, dependent variable

ii- is explained by the K linearly independent exogenous
variables
iii- The first exogenous variable equals to one. This is constant
term included in the model.
iv- The difference between the number of observations (N) and
the number of explanatory variables (K) gives degrees of
freedom (N-K) of the model.
v- Exogenous variables (explanatory) are linearly independent
and form the systematic part of the model.
122 Applied Econometrics with Eviews Applications
 Hypothesıs Testıng 123

CHAPTER 6

HYPOTHESIS TESTING

Its objective is to determine whether or not sample data support a

belief or hypothesis about the population(s) from the sample(s)
which is drawn. For example, one may want to know if a proposed
policy is equally supported by men and women.

DEFINITIONS
 Statistical significance: Is a result that is not likely to occur
randomly, but rather is likely to be attribuTable to a specific
cause.

 Null hypothesis: denoted by H0 and always asserts that no

difference exists between a population parameter and a
single value.

 Alternative hypothesis: is denoted by H1 and assert that there

is difference between a population parameter and a single
value.

 Test Statistic: a tool employed by the researcher in order to

make decision (reject or not reject) on the null hypothesis.
124 Applied Econometrics with Eviews Applications

 Degrees of freedom: refer to how free one is to set the value

of a variable once an observation has been drawn. Drawing
one observation removes one degree of freedom (to set a
value, N - 1, where N is the sample size). N – 1 is the number
of degrees of freedom for the t-test.21

 Confidence level: the percentage of times one would expect

that the sample represents the population; it is commonly set
at 0.90, 0.95, or 0.99.

 Significance level: the probability of rejecting a null

hypothesis when it is true. It is denoted by the ˝α˝ and is
commonly set at 0.10, 0.05, or 0.01.22 Before running any
statistical test, significance level should be defined first.

 Confidence interval: refers to the interval of values around

the test statistic in which the null hypothesis is not rejected. It
is the probability of making wrong decision. Confidence
intervals are generally notated as follows:

[ ], 95% confidence interval (6.1)

Confidence interval

tcritical 0 tcritical
Figure 6.1 Confidence Interval two Sided

21
the z-test does not involve degrees of freedom since it assumes that the population
variance is known
22
significance level + confidence level = 1
 Hypothesıs Testıng 125

One-sided hypothesis test puts all of the significance level "at one
end of a distribution.

One-sided test: ,

critical region
acceptance region p = α, reject null
1- α, accept null

0 tcritical
Figure 6.2 Confidence Interval one Sided

Two-sided hypothesis test: Suppose that we have null hypothesis,

, where is the estimated value and is the valuei
chosen by researcher . If null hypothesis is not true following
alternative hypothesis holds:

critical region
p = α/2, reject null

acceptance region
1- α/2, accept null

-tcritical tcritical
Figure 6.3 Acceptance Region Two Sided

The p-value is the smallest value that would lead to rejection of the
null hypothesis. It is also called the significance probability.
Generally, the smaller the p-value, the more likely is it that one can
reject the null hypothesis. A p−value of 0.003 means that there is
only a 3 in a thousand chance of reaching the wrong conclusion. The
126 Applied Econometrics with Eviews Applications

smaller the p- value, the stronger the evidence against the null
hypothesis.

| | (6.2)

Generally, following conventions are applied:

If p-value > .10 → the observed difference is “not significant”

if p-value ≤ .10 → the observed difference is “marginally significant”
if p-value ≤ .05 → the observed difference is “significant”
if p-value ≤ .01 → the observed difference is “highly significant”

However, for hypothesis testing to work properly, data should be

relatively normally distributed. That is, the distribution of data
should be:
- mound-shaped
- fairly symmetrical
- free of significant skew and kurtosis
The less these conditions hold, the less reliable are tests of
hypothesis and, thereby, the conclusions become less reliable. The
ordered steps of hypothesis testing for one sample are as follow:

i- Calculate the mean and standard error (Se) of the

sample

ii- Define the hypothesized value to be tested

iii- Define the significance level

iv- Construct the null and alternative hypothesis

v- Calculate the standard error of the mean

(6.3)
√

vi- Calculate the t statistics

(6.4)
 Hypothesıs Testıng 127

vii- With N-1 degree of freedom, specify the critical value

form the t distribution Table

viii- Reject null hypothesis

if | | or
if probability value, p < significance level

ix- Interpret the result based on the research questions

Ultimately, only two possible conclusions can be drawn for a

hypothesis test: Reject the null hypothesis, in which case one
concludes that there is sufficient evidence to conclude that the
alternative hypothesis is true.

Do not reject the null hypothesis, in which case one concludes that
there is insufficient evidence to indicate that the alternative
hypothesis is true.

Test statistic is used to determine whether the researcher should

reject or not reject the null hypothesis.23

(6.5)

Where is the standard error of .

It follows t distribution with N-K degrees of freedom. To be precise

we can reject the null hypothesis if t-test result is larger than critical
values which are summarized in Table 6.1.

Example 6.1

We aim to measure relative impact of a diploma on the beginning

salary. Our sample includes graduates from the different universities
with a bachelor degree having a mean annual starting salary of
$10,450. Faculty of Economics graduates from eighty universities, are
randomly selected. Their starting salaries have a mean of $12,050. If
23
The term accept should not be used in the context of hypothesis testing.
128 Applied Econometrics with Eviews Applications

the standard deviation is $950, use a 0.01 level of significance to test

the claim that University graduates with an economics degree have a
mean starting salary that is greater than the mean for graduates with
a bachelor degree.

Information given in the question is following:

µ=10,450
n=80
=12,050
S=950
Α=0.01

Step one: H0: µ=10,450

H1: µ>10,450

Step two: parametric one group of samples  σ unknown

but s is known

Therefore, we use t test with the formula

Calculate the test statistics t value:

√
√

Step three: Identify the critical value or p-value.

We find the critical t value 2.66 by

Or, we find p-value is less than 0.0001 by using the

test statistics t=15.068,
one-tailed, in Standard Normal Distribution Table.

Step four: Draw a graph included the test statistics value, the
critical value and the critical region(s) or compare the
P-value with the significance level α. And then make
a conclusion of the hypothesis.
 Hypothesıs Testıng 129

2.66 15.068
Figure 6.4 Critical Region

Step five: the value of t-statistics is highly significant. It means

that there is enough evidence to reject the null
hypothesis and the mean of our sample equals to
the population mean.

If probability of obtaining t value is less than chosen probability value

the significance levels (10%, 5%, or 1%) we can reject null hypothesis.
We conclude that estimated t value is statistically significant or
significantly different from zero.

Table 6.1 Critical Values for One Sided and Two Sided t-Test
(for infinite sample)
Significance Levels (%) Critical values
One-sided two-sided
1 2.58 2.58
5 1.96 1.96
10 1.64 1.64

Example 6.2

If the impact of income level on the demand for ice-cream is

examined:

First step: define a regression equation based on the

related economic theories, such as:
130 Applied Econometrics with Eviews Applications

Second step: regress consumption on Price, Income and

Temperature using OLS, WLS or other
estimation methods

Third step: Define the significance level ( 10%, 5% or 1%)

Let’s define the significance level as 5%.

Fourth step: compute t-statistics

(0.8343) (0.0011) (0.0004)

t-values -1.25 2.82 7.76

Fifth step: Interpret the results

Reported t-values above indicate that there is

enough evidence to conclude that income and
temperature have positive significant impact
on the ice-cream consumption at the 5%
significance level. Low t- value of the
estimates of the price coefficient (-1.25)
indicates that there is not enough evidence to
reject null hypothesis that no relationship
between price and consumption exists.
 Hypothesıs Testıng 131

F-Test: A JOINT TEST OF SIGNIFICANCE

It is possible to test more than one restriction simultaneously.

(6.6)

Alternative hypothesis is that H0 is not true (at least one of

coefficients is not equal to zero).

(6.7)

It implies that there is no linear relationship between dependent and

independent (explanatory) variables.

(6.8)

S1: residual sum of squares of the full model,

S0: residual sum of squares of the restricted model
N: Sample size,
K: number of variables

Large values for the test statistics imply rejection of null hypothesis.
Critical values for the F-test are attached.

If we use the definition of R2, f statistic can be written as:

(6.9)

Where, are the goodness-of-fit measures for the

unrestricted and the restricted model respectively.

Example 6.3

At first we regress price of coffee on price on cocoa and report the

results in the Table 6.2
132 Applied Econometrics with Eviews Applications

Table 6.2 Eviews Output

Dependent Variable: PCOFFEE
Included observations: 462

Variable Coefficient Std. Error t-Statistic Prob.

PCOCOA 1.183739 0.030164 39.24323 0.0000

C 12.33664 2.307020 5.347436 0.0000

R-squared 0.770004 Mean dependent var 89.77775

Adjusted R-squared 0.769504 S.D. dependent var 53.50328
S.E. of regression 25.68695 Akaike info criterion 9.334163
Sum squared resid 303516.8 Schwarz criterion 9.352065
Log likelihood -2154.192 Hannan-Quinn criter. 9.341211
F-statistic 1540.031 Durbin-Watson stat 0.166605
Prob(F-statistic) 0.000000

In the second step we add price of tea as a regressor and report the
results in Table 6.3

Table 6.3 Eviews Output

Dependent Variable: PCOFFEE
Included observations: 462

Variable Coefficient Std. Error t-Statistic Prob.

PCOCOA 1.022619 0.039617 25.81265 0.0000

PTEA 0.016525 0.002759 5.989058 0.0000
C -5.431432 3.707959 -1.464803 0.1437

R-squared 0.786674 Mean dependent var 89.77775

Adjusted R-squared 0.785745 S.D. dependent var 53.50328
S.E. of regression 24.76546 Akaike info criterion 9.263249
Sum squared resid 281517.5 Schwarz criterion 9.290103
Log likelihood -2136.811 Hannan-Quinn criter. 9.273822
F-statistic 846.3184 Durbin-Watson stat 0.161762
Prob(F-statistic) 0.000000

In the third step calculate F-value as follows:

 Hypothesıs Testıng 133

In the fourth step we interpret the results.

The significant F-value reported above indicates that adding the

regressors has significant contribution to the model.

Restricted model has of zero by construction; its test statistic can

also be written using the relationship between F and R2:

(6.10)

If the test statistics based on F does not reject the null hypothesis, we
can conclude that the model performs rather poorly, a model with
only one intercept term would not be statistically worse. If the test
statistics reject the null hypothesis, we cannot conclude that the
model is good.

If the test statistic is enough to reject null hypothesis, we can

conclude that the model is good. It is possible to observe
contradiction between t-test and F-test results. While t-test do not
reject null hypothesis, F-test can reject or vice versa.

Example 6.424:

The result indicates that reported R2 is significantly different than

zero.

24
Verbeek, M. (2004)
134 Applied Econometrics with Eviews Applications
 Specıfıcatıons 135

CHAPTER 7

SPECIFICATIONS

Before estimating a regression model, model specification should be

checked. Specification errors in the regression model estimation may
result series problem with the robustness and reliability of the
estimation. Specification checking is constructed under three
subtitles which are choosing independent variables, functional form,
and stochastic term.

CHOOSING INDEPENDENT VARIABLES

The researchers define the independent variables based on the
economic theory, experience, common sense, critical thinking and
research objectives. The independent variables which are chosen
should satisfy the estimation assumptions. Any departure from those
assumptions may cause unreliable and wrong findings.

If the economic theory requires including one or more independent

variables in the regression equation, then they should be definitely
included in the model. Leaving a related variable outside of the
model can cause bias in the estimation, and including irrelevant
variables in the regression equation leads to higher variances of the
estimation. It is mostly difficult to decide whether to include or
exclude an independent variable in the regression equation, if it is
not included in the economic theory.
136 Applied Econometrics with Eviews Applications

Omitting a Relevant Variable

Data unavailability and researchers ignorance may cause a regression

equation without relevant independent variable. Omitted variables
bias is the bias resulting from leaving a relevant independent variable
out of the regression equation. In a multiple regression equation the
coefficient , represents the amount of changes in the dependent
variable when independent variable change one unit holding
constant the other independent variables in the equation. If a
relevant variable is excluded, it is not held constant for the
calculation and interpretation of coefficient . Because of this,
omission of a relevant independent variable may cause bias pushing
the expected value of the estimated coefficient out from the true
value of the population coefficient.

Assume that the independent variable is omitted from the

regression equation.

(7.1)

The error term includes the omitted independent variable as follows:

, (7.2)

then Assumption 2 (unbiasedness) does not hold.

(7.3)

Also, if and are correlated, Assumption 3 (efficiency) does not

hold.

(7.4)

Gauss- Markov theorem is not valid, and OLS is not BLUE, estimate of
regression coefficient is biased. OLS overestimate the coefficient of
independent variable .

̂ (7.5)
 Specıfıcatıons 137

Omission of a closely related independent variable implies that OLS is

no longer BLUE.

If a relevant variable is omitted from the regression equation:

i- Estimate of the coefficient of that variable is absent

ii- The coefficients of the included independent variables

are likely to be biased

If omitted variable is not correlated with the included independent

variable and the coefficient of omitted independent variable is zero,
then OLS is BLUE.

Detection

i- Check the economic theory to identify relevant

independent variable

ii- Check the compatibility of the sign of estimated

coefficients with the related economic theory

Wald Test (Coefficient Restrictions)

The Wald test (Wald, 1943) statistics is the estimation of unrestricted

regression without imposing the coefficient restrictions specified by
the null hypothesis. It measures how close the unrestricted
estimation satisfies the restrictions under the null hypothesis. The
unrestricted estimation satisfies the restrictions if the imposed
restrictions are true.

The Wald test statistic is computed as follows:

which is linear model

Linear restrictions are as follows:

(7.6)
138 Applied Econometrics with Eviews Applications

R : q x k matrix,
r : a q vector

Wald statistics:

, (7.7)

If we assume that errors are independent and identically normally

distributed, F-statistic can be computed as follows:

̅ ̅
(7.8)

̅ : Residual vector of the restricted regression. F-statistic does

compare the residual sum of squares computed with and without the
imposed restrictions. F-statistic is small if the restrictions are true
and there is little difference in two residual sums of squares.

Example 7.1

Estimation of Cobb-Douglas production function (Cobb, Douglas,

1928) using the data provided the following results:

Table 7.1 Eviews Output

Dependent Variable: LOG(OUTPUT)
Method: Least Squares
Sample: 1 569
Included observations: 569
Variable Coefficient Std. Error t-Statistic Prob.
C -1.711459 0.096711 -17.69672 0.0000
LOG(CAPITAL) 0.207570 0.017188 12.07677 0.0000
LOG(LABOR) 0.714847 0.023142 30.89002 0.0000
R-squared 0.837830 Mean dependent var 1.674907
Adjusted R-squared 0.837257 S.D. dependent var 1.185050
S.E. of regression 0.478067 Akaike info criterion 1.367125
Sum squared resid 129.3580 Schwarz criterion 1.390028
Log likelihood -385.9472 Hannan-Quinn criter. 1.376062
F-statistic 1462.078 Durbin-Watson stat 1.963300
Prob(F-statistic) 0.000000
 Specıfıcatıons 139

LOG(OUTPUT) = C(1) + C(2)*LOG(CAPITAL) + C(3)*LOG(LABOR)

LOG(OUTPUT) = -1.71145941868 + 0.207570296302*LOG(CAPITAL) +

0.714846786693*LOG(LABOR)

In order to test hypothesis of constant returns to scale, type the

following restriction in the dialog box:

c(2)+c(3) = 1

Eviews reports following output:

There is only one restriction. Two test statistics are identical with the
p-values of both statistics indicates that the null hypothesis of
constant returns to scale can be rejected.

Table 7.2 Eviews Output for Wald Restriction Test

Wald Test:
Equation: Untitled

Test Statistic Value df Probability

F-statistic 20.33514 (1, 566) 0.0000

Chi-square 20.33514 1 0.0000

Null Hypothesis Summary:

Normalized Restriction (= 0) Value Std. Err.

-1 + C(2) + C(3) -0.077583 0.017205

Restrictions are linear in coefficients.

To test more than one restriction, separate the restrictions by

commas.

If the hypothesis is the elasticity of output with respect to labor is 1/4

and with respect to capital 3/4, type the following restrictions:

c(2) = 1/4, c(3) = 3/4

140 Applied Econometrics with Eviews Applications

Eviews reports following result:

Significant F-statistics and a Chi-square indicate that there is enough
evidence to reject null hypotheses defined as restrictions.

Table 7.3 Eviews Output for Wald Restriction Test

Wald Test:
Equation: Untitled

Test Statistic Value df Probability

F-statistic 12.27250 (2, 566) 0.0000

Chi-square 24.54501 2 0.0000

Null Hypothesis Summary:

Normalized Restriction (= 0) Value Std. Err.

-1/4 + C(2) -0.042430 0.017188

-3/4 + C(3) -0.035153 0.023142

Restrictions are linear in coefficients.

Omitted Variables Test

This test gives options to add one or more variables to an existing

equation and to test whether the set makes a significant contribution
in explaining the variation in the dependent variable. The null
hypothesis in this case is:

the additional variable or variables are not significant

The output from F-statistics, Likelihood ratio statistics, p-values and

the estimation results of the unrestricted model is reported.

The F-statistic is based on the difference between the residual sums

of squares of the restricted and unrestricted regressions.

Number of observation of original and test equation should be equal

in order to employ the omitted variables test. It also can be
employed if the equation is specified by listing regressors not by
 Specıfıcatıons 141

formula. It is available for LS, TSLS, ARCH, binary, ordered, censored,

truncated and count methods.

Example 7.2

Table 7.4 Eviews output for Omitted Variables Test Results

Dependent Variable: CONS
Method: Least Squares
Included observations: 30
Variable Coefficient Std. Error t-Statistic Prob.
INCOME 0.000505 0.001988 0.253942 0.8014
C 0.316715 0.168665 1.877771 0.0709
R-squared 0.002298 Mean dependent var 0.359433
Adjusted R-squared -0.033334 S.D. dependent var 0.065791
S.E. of regression 0.066878 Akaike info criterion -2.507551
Sum squared resid 0.125235 Schwarz criterion -2.414138
Log likelihood 39.61326 Hannan-Quinn criter. -2.477667
F-statistic 0.064487 Durbin-Watson stat 0.392752
Prob(F-statistic) 0.801396
Omitted Variables: PRICE TEMP
F-statistic 33.15603 Prob. F(2,26) 0.0000
Log likelihood ratio 38.01235 Prob. Chi-Square(2) 0.0000
Test Equation:
Dependent Variable: CONS
Method: Least Squares
Included observations: 30
Variable Coefficient Std. Error t-Statistic Prob.
INCOME 0.003308 0.001171 2.823722 0.0090
C 0.197315 0.270216 0.730212 0.4718
PRICE -1.044414 0.834357 -1.251759 0.2218
TEMP 0.003458 0.000446 7.762213 0.0000
R-squared 0.718994 Mean dependent var 0.359433
Adjusted R-squared 0.686570 S.D. dependent var 0.065791
S.E. of regression 0.036833 Akaike info criterion -3.641296
Sum squared resid 0.035273 Schwarz criterion -3.454469
Log likelihood 58.61944 Hannan-Quinn criter. -3.581528
F-statistic 22.17489 Durbin-Watson stat 1.021170
Prob(F-statistic) 0.000000
142 Applied Econometrics with Eviews Applications

Solution

i- Expected bias analysis may be applied. Omitting a

variable may have caused it in the estimated
coefficient of one of the independent variable in the
model. Expected bias can be estimated as follows:

(7.9)

ii- denotes the correlation coefficient between

included and omitted variable in the regression. Sign of
expected bias should be checked with the sign of the
unexpected result. If they are same, then the variable
could be the reason of bias. This analysis should be
used if there is obviously a bias.

iii- If the omitted variable is obvious and available, it

should be included in the model.

iv- If the omitted variable is not available, a proxy variable

should be found which is closely related to the omitted
variable and included in the model.

Including an Irrelevant Variable in the Regression equation

Including an irrelevant variable in the regression equation may cause

an increase in the variances of the estimated coefficients of included
independent variables. It may reduce the precision of the estimation
and does not cause bias.

Suppose that an independent variable is not related or it is obvious

that it has no relationship with the dependent variable such as:

(7.10)

(7.11)

If is not related, = 0. Hence Assumption 2 and 3 holds, OLS

estimators are not biased and remain BLUE.
 Specıfıcatıons 143

Standard errors on the estimated coefficients are larger in the model

which includes irrelevant dependent variable than the optimum
model. These results are higher than t-ratio.

When it is detected that it is irrelevant, it should be omitted from the

regression equation.

How to Choose Correct Variables?

i- Following the Economic theory

ii- Check whether it is significant with correct sign

iii- Has ̅

iv- Check the other coefficients sign, after the variable is

included

If all the conditions above are satisfied, then variable belongs to the
regression equation. Following techniques are strongly
recommended to be employed for specification bias problem:

i- Scanning to develop a tesTable theory: it is about

analyzing a data set for the purpose of developing a
tesTable theory or hypothesis. An economic theory or
hypothesis should have been tested on a different data
set before giving reference to theory or hypothesis.
ii- Sensitivity analysis: it refers to employing different
alternative specifications to determine whether the
estimation result is robust. How sensitive an
estimation result is to a change in different
specifications should be examined.

Redundant Variables Test

This test allows testing for statistical significance of a subset of the

included variables. It tests whether the coefficients of the variables in
a regression equation are zero. If they are equal to zero, they should
be omitted from the equation. It is available for LS, TSLS, ARCH,
144 Applied Econometrics with Eviews Applications

binary, ordered, censored, truncated and count methods and can be

employed if the equation is specified by listing regressors not by
formula.

Example 7.3

Table 7.5 Eviews Result for Redundant Variable Test Results

Dependent Variable: PRICE
Variable Coefficient Std. Error t-Statistic Prob.
LOTSIZE 4.040758 0.350757 11.52011 0.0000
BEDROOMS 2353.935 1084.623 2.170280 0.0304
BATHRMS 15480.96 1538.040 10.06538 0.0000
DRIVEWAY 8052.076 2107.907 3.819939 0.0001
FULLBASE 5697.870 1649.906 3.453451 0.0006
PREFAREA 8873.391 1731.846 5.123660 0.0000
RECROOM 4069.431 1972.623 2.062954 0.0396
AIRCO 12353.98 1592.301 7.758570 0.0000
STORIES 6316.716 958.5377 6.589950 0.0000
C -6619.304 3510.132 -1.885770 0.0599
R-squared 0.645637 Mean dependent var 68121.60
Adjusted R-squared 0.639687 S.D. dependent var 26702.67
S.E. of regression 16028.57 Akaike info criterion 22.22028
Sum squared resid 1.38E+11 Schwarz criterion 22.29908
Log likelihood -6056.136 Hannan-Quinn criter. 22.25108
F-statistic 108.5081 Durbin-Watson stat 1.562506
Prob(F-statistic) 0.000000
Redundant Variables: DRIVEWAY
F-statistic 14.59193 Prob. F(1,536) 0.0001
Log likelihood ratio 14.66544 Prob. Chi-Square(1) 0.0001
Dependent Variable: PRICE
Method: Least Squares Included observations:
546 Variable Coefficient Std. Error t-Statistic Prob.
LOTSIZE 4.380574 0.343556 12.75070 0.0000
BEDROOMS 1893.405 1091.458 1.734749 0.0834
BATHRMS 15307.73 1556.706 9.833411 0.0000
FULLBASE 5890.468 1669.875 3.527489 0.0005
PREFAREA 9731.681 1738.805 5.596763 0.0000
RECROOM 4275.759 1996.683 2.141431 0.0327
AIRCO 12392.28 1612.295 7.686113 0.0000
STORIES 6824.015 961.2320 7.099238 0.0000
C -1099.044 3239.084 -0.339307 0.7345
R-squared 0.635990 Mean dependent var 68121.60
Adjusted R-squared 0.630567 S.D. dependent var 26702.67
S.E. of regression 16230.15 Akaike info criterion 22.24348
Sum squared resid 1.41E+11 Schwarz criterion 22.31440
Log likelihood -6063.469 Hannan-Quinn criter. 22.27120
F-statistic 117.2792 Durbin-Watson stat 1.583598
Prob(F-statistic) 0.000000
 Specıfıcatıons 145

CHOOSING A FUNCTIONAL FORM

To specify the classical linear regression model, a specific functional
form should be chosen. Any functional form that is linear in
parameters can be chosen. If the incorrect functional form is chosen,
then the model is misspecified. If the model is misspecified then it
may not be a reasonable approximation of the true data generation
process. We make a functional form specification error when we
choose the wrong functional form.

Constant term: constant term should be included in the regression

model unless there is some strong reason for opposite such as the
data is in the close neighborhood. Not including a constant term
causes inflated t-ratio.

Functional Forms:

The Log-log Regression Model (Double Log)

(7.12)

When

Consider the following exponential regression model:

, (7.13)

It can be expressed in logs:

, (7.14)

It is called linear in logs and can be estimated by OLS on the condition

that classical assumptions are satisfied.

Cobb-Douglas Production Function

(7.15)
146 Applied Econometrics with Eviews Applications

When L changes 1%, Q changes by %

Lin-Log Model (semi log)

(7.16)

holding other variables constant.

Theimpact of a variation in on y decreases as gets larger.

Log-ln Model

(7.17)

when changes one unit yi changes 100 % holding the other

regressors constant. The impact of a
variation .

Quadratic Forms

, i=1,…..,n (7.18)

Cost curve is U-shaped and cost is quadratic in output with

Inverse Form

(7.19)
is large.

Intercept Dummy Independent Variable

The intercept dummy independent variable changes the intercept

but the slope remains constant. One dummy variable is used for two
categories. If there are more than two categories more than one
dummy independent variable can be set in the regression equation.
 Specıfıcatıons 147

To include two same dummy variable cause perfect multicollinearity

and violates Assumption 6.

Slope Dummy Independent Variable

(7.20)

There are two equations as follows:

(7.21)

(7.22)

Each equation can be estimated separately.

Interaction term should be included if there is reason to believe that

the slopes are different across categories.

Lags

(7.23)

The length of time between cause and effect is called a lag.

. measures the impact of

previous observation on the current observation. If lag structure take
place over more than one time period, it is called distributed lags.

Mixed Functional Forms

It is possible to mix functional forms as follows:

(7.24)

In this case, yi are a semi-log function of xi, a quadratic function of φi,

and a linear function of zi. The marginal effect and elasticity for each
of these variables is given by the formulas above.
148 Applied Econometrics with Eviews Applications

How to evaluate functional form?

In the residual plot, whether there is a systematic pattern between

and , should be checked. Different functional forms should be
used.

Consequences of Choosing the Wrong Functional Form

The OLS estimator will be biased and not valid.

Detection and Correction of Functional Form Specification Errors

Following two alternative methodologies are used to choose a

specific functional form for a model:

i. Maintained hypothesis methodology

ii. Theory/testing methodology

Maintained Hypothesis Methodology

This methodology uses theory and/or tractability to choose a specific

functional form. Once a specific functional form is chosen, it is
treated as a maintained hypothesis and not tested using the sample
data. Choosing functional form based on tractability is not good
practice. Relying on theory alone may not be a good practice. This is
because there are many situations when theory has nothing to say
about the appropriate functional form. If the wrong functional form
is chosen, then the parameter estimates will be biased and all tests of
hypothesis, strictly speaking, will be incorrect.

Theory/Testing Methodology

This methodology involves the following steps:

i- A set of specific functional forms that are consistent

with theory is identified. If a specific functional form is
inconsistent with the theory being used to guide the
 Specıfıcatıons 149

specification of the statistical model, then it should not

be considered.

ii- Statistical tests should be conducted to determine

which specific functional form should be chosen.

iii- One of the following two approaches may be used in

the third step;
1) Testing-down approach; 2) Testing-up approach.

Testing-Down Approach

When using the testing-down approach, we begin with a general

model and test-down to a more specific model. To test for nonlinear
terms and interaction terms, begin with a general model that
includes one or more of these terms. For example, the general
model might include nonlinear terms such as X2 and/or lnX, or
interaction terms such as X*A. A t-test and/or an F-test can be
employed to test whether these terms belong in the model.

General Functional Forms

A more systematic approach is to begin with a general functional

form. This general functional form has one or more specific
functional forms as a special case. An F-test or a t-test is used to test
whether a specific functional form is the appropriate functional form.

Testing-up Approach

When using the testing-up approach, begin with a specific model and
test-up to a more general model. To test for nonlinear terms and
interaction terms, a specific model that does not include one or more
of these terms may be appropriate. For example, the specific model
might not include nonlinear terms such as X2 and/or lnX, or
interaction terms such as X*A. We then use a Lagrange multiplier
test to test whether these terms should be added to the model.
150 Applied Econometrics with Eviews Applications

Other Tests for Functional Form

A number of other criteria and statistical tests are also used to test
for functional form. Some of these are the following:

i- Adjusted R2

ii- Ramsey’s Reset Test

iii- Recursive Residual Test

SPECIFICATION TESTS
Residual Tests

Correlograms and Q-Statistics

This test displays the autocorrelation and partial autocorrelations of

the squared residuals up to specified number of lags. It is available
for LS, TSLS, nonlinear LS, binary, ordered, censored, and count
methods.
 Specıfıcatıons 151

Example 7.4

Table 7.6 Eviews Result for Correlograms and Q-Statistics

Sample: 1954Q1 1994Q4
Included observations: 164

Autocorrelation Partial Correlation AC PAC Q-Stat Prob

.|******| .|******| 1 0.869 0.869 126.03 0.000

.|***** | *|. | 2 0.723 -0.130 213.81 0.000
.|***** | .|* | 3 0.644 0.196 283.88 0.000
.|**** | *|. | 4 0.543 -0.185 333.97 0.000
.|*** | .|. | 5 0.432 -0.016 365.96 0.000
.|** | **|. | 6 0.301 -0.229 381.58 0.000
.|* | .|. | 7 0.179 -0.014 387.12 0.000
.|* | .|* | 8 0.131 0.162 390.13 0.000
.|* | *|. | 9 0.077 -0.101 391.17 0.000
.|. | .|. | 10 -0.006 -0.047 391.18 0.000
.|. | .|. | 11 -0.058 -0.004 391.78 0.000
*|. | .|. | 12 -0.080 0.027 392.93 0.000
*|. | .|. | 13 -0.089 0.000 394.37 0.000
*|. | .|. | 14 -0.089 0.032 395.81 0.000
*|. | .|. | 15 -0.097 -0.002 397.52 0.000
*|. | .|. | 16 -0.095 -0.012 399.18 0.000
.|. | .|* | 17 -0.053 0.095 399.71 0.000
.|. | .|. | 18 -0.009 0.038 399.72 0.000
.|. | .|. | 19 0.006 -0.049 399.73 0.000
.|. | .|. | 20 0.027 0.041 399.87 0.000
.|. | .|. | 21 0.060 0.008 400.56 0.000
.|* | .|. | 22 0.080 -0.035 401.77 0.000
.|* | .|. | 23 0.080 -0.052 403.02 0.000
.|. | .|. | 24 0.061 -0.024 403.74 0.000
.|. | .|* | 25 0.059 0.077 404.42 0.000
.|* | .|. | 26 0.081 0.034 405.73 0.000
.|* | .|. | 27 0.081 -0.002 407.03 0.000
.|* | .|. | 28 0.078 0.070 408.24 0.000
.|* | .|. | 29 0.087 0.001 409.78 0.000
.|* | .|. | 30 0.085 -0.051 411.23 0.000
.|* | .|. | 31 0.075 -0.028 412.38 0.000
.|. | .|. | 32 0.050 -0.055 412.90 0.000
.|. | .|. | 33 0.014 -0.017 412.94 0.000
.|. | .|. | 34 -0.010 -0.028 412.96 0.000
.|. | .|. | 35 -0.030 0.010 413.15 0.000
.|. | .|* | 36 -0.030 0.126 413.34 0.000
152 Applied Econometrics with Eviews Applications

Correlograms of Squared Residuals

This test displays the autocorrelation and partial autocorrelations of

the squared residuals up to specified number of lags defined. It can
be used to check ARCH in the residuals. If autocorrelation and partial
autocorrelation is equal to zero, there is no ARCH effect in the
residuals.

Histogram and Normality test

This test provides you a histogram and descriptive statistics of the

residuals, including Jarque-Bera statistics (JB, 1981) for testing
normality. If the residuals are normally distributed, the histogram
should be bell-shaped and Jarque-Bera test statistics should not be
significant.

Example 7.5
16
Series: Residuals
14 Sample 1954Q1 1994Q4
Observations 164
12
Mean 4.12e-14
10 Median 2.84e-13
Maximum 1.73e-11
8 Minimum -1.23e-11
Std. Dev. 6.60e-12
6
Skewness 0.355963
Kurtosis 2.721785
4

Jarque-Bera 3.992331
2
Probability 0.135855
0
-1.0e-11 -2.8e-27 1.0e-11

Figure 7.1 Eviews Result for Normality Test

Serial Correlation LM Test

This test is an alternative to the Q-statistics for testing serial

correlation. Unlike Durbin Watson test (AR (1)), LM test can be used
to test higher order ARMA errors.

The null hypothesis of the LM test is there is no serial correlation up

to chosen lag order. Eviews reports two test statistics. The F-statistics
 Specıfıcatıons 153

is an omitted variable test for the joint significance of all lagged

residuals. Omitted variables are residuals not independent variables.
R2 statistic is the Breusch-Godfrey LM test statistics.

Example 7.6

Table 7.7 Eviews Result for Serial Correlation LM Test

Breusch-Godfrey Serial Correlation LM Test:

F-statistic 2322039. Prob. F(2,156) 0.0000

Obs*R-squared 163.9945 Prob. Chi-Square(2) 0.0000

Test Equation:
Dependent Variable: RESID
Method: Least Squares
Sample: 1954Q1 1994Q4
Included observations: 164
Presample missing value lagged residuals set to zero.

Variable Coefficient Std. Error t-Statistic Prob.

TBR -3.87E-12 1.03E-14 -375.3065 0.0000

Y 5.33E-11 5.35E-14 994.7556 0.0000
MF -1.00E-11 1.23E-13 -81.65197 0.0000
M -1.09E-10 1.09E-13 -997.9981 0.0000
CPR 1.68E-14 9.17E-15 1.830093 0.0691
C 3.58E-10 8.06E-13 444.1100 0.0000
RESID(-1) -0.001390 0.001407 -0.988141 0.3246
RESID(-2) -0.000686 0.001019 -0.673236 0.5018

R-squared 0.999966 Mean dependent var 4.12E-14

Adjusted R-squared 0.999965 S.D. dependent var 6.60E-12
S.E. of regression 3.91E-14 Sum squared resid 2.38E-25
F-statistic 663413.6 Durbin-Watson stat 1.993998
Prob(F-statistic) 0.000000

ARCH-LM Test

This is LM test for ARCH (Engle, 1982) in the residuals. Ignoring ARCH
effect can cause inefficiency in estimation. The null hypothesis is that
154 Applied Econometrics with Eviews Applications

there is no ARCH effect in the residuals. It is computed from the

residual test regression as follows:

, (7.25)

e is the residual.

This is a regression of the squared residuals on constant and lagged

squared residuals up to order q. The F-statistic is an omitted variable
test for the joint significance of all lagged squared residuals. Engels
LM test statistic is equal to number of observations times R2 statistic.

Example 7.7

Table 7.8 Eviews Result for ARCH-LM Test

Heteroscedasticity Test: ARCH

F-statistic 122.9920 Prob. F(4,155) 0.0000

Obs*R-squared 121.6674 Prob. Chi-Square(4) 0.0000

Test Equation:
Dependent Variable: RESID^2
Method: Least Squares
Sample (adjusted): 1955Q1 1994Q4
Included observations: 160 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

C 6.23E-24 2.91E-24 2.141565 0.0338

RESID^2(-1) 0.997611 0.080486 12.39489 0.0000
RESID^2(-2) -0.245270 0.111848 -2.192884 0.0298
RESID^2(-3) 0.229981 0.111806 2.056967 0.0414
RESID^2(-4) -0.120375 0.080923 -1.487518 0.1389

R-squared 0.760421 Mean dependent var 4.23E-23

Adjusted R-squared 0.754238 S.D. dependent var 5.75E-23
S.E. of regression 2.85E-23 Sum squared resid 1.26E-43
F-statistic 122.9920 Durbin-Watson stat 1.983015
Prob(F-statistic) 0.000000
 Specıfıcatıons 155

Whites Heteroscedasticity Test

It is a test for heteroscedasticity in the residuals from a LS regression

(White, 1980). In the presence of heteroscedasticity standard errors
are no longer valid but OLS estimates are still consistent. In order to
correct heteroscedasticity, weighted least squares estimation
method can be employed or chosen the robust standard error option
to correct the standard errors.

White test is test of null hypothesis of that there is no

heteroscedasticity.

For example we estimate following regression,

, (7.26)

the test statistics based on auxiliary regression.

(7.27)

The White test statistic is computed the number of observations

times R2 from the regression.

Example 7.8

Table 7.9 Eviews Result for Whites Heteroscedasticity Test

Heteroscedasticity Test: White

F-statistic 120.7133 Prob. F(18,145) 0.0000

Obs*R-squared 153.7404 Prob. Chi-Square(18) 0.0000
Scaled explained SS 123.4851 Prob. Chi-Square(18) 0.0000

Test Equation:
Dependent Variable: RESID^2
Method: Least Squares
Sample: 1954Q1 1994Q4
Included observations: 164
Collinear test regressors dropped from specification
156 Applied Econometrics with Eviews Applications

Variable Coefficient Std. Error t-Statistic Prob.

C -2.02E-19 1.57E-20 -12.81748 0.0000

TBR 1.60E-22 9.94E-22 0.160650 0.8726
TBR^2 9.78E-24 7.02E-24 1.392979 0.1658
TBR*Y -4.17E-22 4.82E-23 -8.644330 0.0000
TBR*MF -3.64E-22 1.37E-22 -2.658945 0.0087
TBR*M 8.36E-22 1.03E-22 8.125272 0.0000
TBR*CPR 9.93E-24 1.29E-23 0.767886 0.4438
Y 1.35E-20 2.00E-21 6.763355 0.0000
Y^2 2.95E-21 9.05E-23 32.57495 0.0000
Y*MF 2.77E-21 3.16E-22 8.783545 0.0000
Y*M -1.19E-20 3.70E-22 -32.14765 0.0000
Y*CPR -1.57E-24 3.91E-23 -0.040015 0.9681
MF 4.74E-20 2.68E-21 17.66015 0.0000
MF*M -1.08E-20 3.93E-22 -27.41762 0.0000
MF*CPR 1.52E-22 1.33E-22 1.146168 0.2536
M^2 1.25E-20 3.99E-22 31.25539 0.0000
M*CPR 1.64E-23 8.29E-23 0.197514 0.8437
CPR -1.07E-21 9.46E-22 -1.130305 0.2602
CPR^2 -4.91E-24 6.30E-24 -0.780155 0.4366

R-squared 0.937442 Mean dependent var 4.33E-23

Adjusted R-squared 0.929676 S.D. dependent var 5.71E-23
S.E. of regression 1.52E-23 Sum squared resid 3.33E-44
F-statistic 120.7133 Durbin-Watson stat 1.313970
Prob(F-statistic) 0.000000
 Parameter Stabılıty 157

CHAPTER 8

PARAMETER STABILITY

Aim of the structural break and parameter stability test is to check

whether the parameters of the model are constant or stable over the
subsamples of the data. The stability of the parameter estimates in
the sample period can be obtained by computing recursive
coefficient estimates and looking at their plots. It provides
information about the stability of the estimates. Main idea is
repeatedly adding one observation and re-estimating the
parameters. From plots we can conclude that the estimates are
stable.

One empirical technique is to divide the whole samples into two

subsamples, and use first subsample for estimation and second
subsample for testing. If there is obvious data points where structural
break have taken place, the subsamples might be determined based
on these points. If there is no clear structural break, a rule-of-thumb
which is 80% to 90% of the observations for estimation and
remainder for testing is followed.
158 Applied Econometrics with Eviews Applications

CHOW BREAKPOINT TEST FOR STRUCTURAL BREAK

Possible structural breaks points are defined. This test fits the
equation separately for each subsample and sees whether there are
significant differences in the estimated equations. If significant
difference is detected, then it is concluded that there is structural
change in the relationship. This test can be used with least squares
and two-stage least squares estimation method. F-statistic, with a
single breakpoint is computed as follows:

̅ ̅
(8.1)

̅ ̅ : the restricted sum of squared residuals,

: the sum of squared residuals from subsample 1,
: the sum of squared residuals from subsample 2,
T : the total number of observations
K : the number of parameters in the equation

This formula can be generalized for more than one breakpoint.

The Log likelihood ratio test result is also reported.

Major drawback of the Chow breakpoint test (Chow, 1960) is that

each subsample requires the number of observation as many as the
number of parameters.
 Parameter Stabılıty 159

Example 8.1

Table 8.1 Eviews Result for Chow Breakpoint Test

Dependent Variable: INFL
Method: Least Squares
Sample: 1954Q1 1994Q4
Included observations: 164
Variable Coefficient Std. Error t-Statistic Prob.
TBR -0.430756 0.406092 -1.060735 0.2904
M -2.864898 1.129140 -2.537240 0.0121
CPR 0.946117 0.388299 2.436569 0.0159
C 19.14048 7.279982 2.629194 0.0094
R-squared 0.453612 Mean dependent var 4.280384
Adjusted R-squared 0.443367 S.D. dependent var 2.566663
S.E. of regression 1.914931 Akaike info criterion 4.161328
Sum squared resid 586.7136 Schwarz criterion 4.236934
Log likelihood -337.2289 Hannan-Quinn criter. 4.192021
F-statistic 44.27745 Durbin-Watson stat 0.848191
Prob(F-statistic) 0.000000
Chow Breakpoint Test: 1970Q2
Null Hypothesis: No breaks at specified breakpoints
Varying regressors: All equation variables
Equation Sample: 1954Q1 1994Q4
F-statistic 15.66846 Prob. F(4,156) 0.0000
Log likelihood ratio 55.38695 Prob. Chi-Square(4) 0.0000
Wald Statistic 62.67385 Prob. Chi-Square(4) 0.0000

CHOW FORECAST TEST FOR STRUCTURAL BREAK

If there is only a few observations or the number of observations are
less than the number of parameters in the equation, then Chow
Forecast test should be used. This test can be used with least squares
or two stage least squares estimation method. F statistic is computed
as follows:

̅ ̅
(8.2)

̅ ̅ : the residuals sum of squares when the equation is fitted to all T

observations
: the residuals sum of squares when the equation is fitted to T 1
observations
160 Applied Econometrics with Eviews Applications

The Log likelihood ratio test result is also reported.

Example 8.2

Table 8.2 Eviews Result for Chow Forecast Test

Chow Forecast Test: Forecast from 1980Q3 to 1994Q4
F-statistic 1.498472 Prob. F(58,102) 0.0374
Log likelihood ratio 101.0741 Prob. Chi-Square(58) 0.0004
Test Equation:
Dependent Variable: INFL
Method: Least Squares
Sample: 1954Q1 1980Q2
Included observations: 106
Variable Coefficient Std. Error t-Statistic Prob.
TBR 0.756093 0.494190 1.529963 0.1291
M 12.49727 4.787969 2.610141 0.0104
CPR 0.052019 0.457198 0.113778 0.9096
C -78.75231 30.25633 -2.602838 0.0106
R-squared 0.606890 Mean dependent var 4.421247
Adjusted R-squared 0.595328 S.D. dependent var 2.770335
S.E. of regression 1.762317 Akaike info criterion 4.008141
Sum squared resid 316.7876 Schwarz criterion 4.108648
Log likelihood -208.4315 Hannan-Quinn criter. 4.048877
F-statistic 52.48973 Durbin-Watson stat 1.321713
Prob(F-statistic) 0.000000

CUSUM TEST
The CUSUM test statistic (Brown, Durbin, and Evans, 1975) is based
on cumulative sums of scaled recursive residuals and is plotting the
cumulative sum together with the 5% critical lines against time
(Vogelvang, 133). If the cumulative sum goes outside of the 5%
critical lines, then the test shows parameter instability. The CUSUM
test is based on the following recursive residual test statistics defined
as follows:

∑ (8.3)
 Parameter Stabılıty 161

w : recursive residual
s : the standard error of the regression fitted to all T observations.

The expectations of the CUSUM statistics are zero under the null
hypothesis of constant parameters.

Example 8.3

40

30

20

10

-10

-20

-30

-40
1955 1960 1965 1970 1975 1980 1985 1990

CUSUM 5% Significance

Figure 8.1 Ewievs Result for CUSUM Test

CUSUM OF SQUARES TEST

The CUSUM of squares statistics is a cumulative sum of squared
residuals. The expectations of the CUSUM of squares statistics run
from zero at the first observation until the value of one at the end of
the sample period, under the null hypothesis of constant coefficients
and variance. The CUSUM test statistic (Brown, Durbin, and Evans,
1975) is defined as follows:

∑ ∑ (8.4)
162 Applied Econometrics with Eviews Applications

The expected value of S under the null hypothesis of parameter

constancy or stability is,

[ ] (8.5)

If t=k, then it goes zero

If t=T, then it goes to unity
St is plotted together with the 5% critical lines against time.
Movement of the critical lines shows parameter instability.

Example 8.4 1.2

1.0

0.8

0.6

0.4

0.2

0.0

-0.2
1955 1960 1965 1970 1975 1980 1985 1990

CUSUM of Squares 5% Significance

Figure 8.2 Ewievs Result for CUSUM of Squares Test

ONE STEP AHEAD FORECAST TEST

This test produces a plot of the residuals and standard errors of the
sample points whose probability value is equal or less than 15%. The
upper portion of the plot repeats the recursive residuals and
standard errors displayed by the recursive residuals option, and the
lower portion of the plot indicates the values for those sample points
where the hypothesis of parameter stability would be rejected at the
5%, 10%, or 15% significance levels. If p-value is less than 0.05, it
 Parameter Stabılıty 163

means that the recursive residual go outside the two standard error
bounds at this point of observation.

Example 8.5

.000 -4

.025
-8
.050

.075

.100

.125

.150
1955 1960 1965 1970 1975 1980 1985 1990

One-Step Probability
Recursive Residuals

Figure 8.3 Ewievs Result for One Step Ahead Forecast Test

N STEP FORECAST TEST

This test employs the recursive calculations. This test does not
require specifying a forecast period; it calculates all feasible cases,
starting with the smallest possible sample size in estimating and
forecasting equation through adding one observation at a time.
Upper portion of the plot indicates recursive residuals and lower
portion shows the corresponded p-values.
164 Applied Econometrics with Eviews Applications

Example 8.6

.000 -4

.025
-8
.050

.075

.100

.125

.150
1955 1960 1965 1970 1975 1980 1985 1990

N-Step Probability
Recursive Residuals

Figure 8.4 Ewievs Result for N Step Forecast Test

RECURSIVE COEFFICIENT ESTIMATES

This test shows the evolution of estimates for all coefficients as more
and more of the data are used in the estimation. It gives a plot of
selected coefficients in the equation for all feasible recursive
estimations including two standard error bands around the estimated
coefficients.

If the coefficient shows significant variations as more observation is

added to the estimation equation, then there is enough evidence for
instability.
 Parameter Stabılıty 165

Example 8.7
8 160

6
120

4
80
2
40
0

0
-2

-4 -40
60 65 70 75 80 85 90 60 65 70 75 80 85 90

Recursive C(1) Estimates Recursive C(2) Estimates

± 2 S.E. ± 2 S.E.

4 400

2
0
0

-2 -400

-4
-800
-6

-8 -1,200
60 65 70 75 80 85 90 60 65 70 75 80 85 90

Recursive C(3) Estimates Recursive C(4) Estimates

± 2 S.E. ± 2 S.E.

Figure 8.5 Eviews Result for Recursive Coefficient Estimates

WALD TEST FOR STRUCTURAL CHANGE

Chow test for parameter stability is used on the condition that the
errors are independent, normally distributed and equal residual
variance. Wald test is used in the case of unequal residual variance in
different sub-samples.
A Wald statistic for the null hypothesis of no structural change in
independent samples is constructed as follows:

(8.6)

which has asymptotic distribution.

166 Applied Econometrics with Eviews Applications

Degrees of freedom are equal to the number of estimated

parameters in the estimated b vector.

Example 8.8

Table 8.3 Eviews Test Result for Wald Test

Ramsey RESET Test:

F-statistic 0.382882 Prob. F(1,159) 0.5370

Log likelihood ratio 0.394447 Prob. Chi-Square(1) 0.5300

Test Equation:
Dependent Variable: INFL
Method: Least Squares
Sample: 1954Q1 1994Q4
Included observations: 164

Variable Coefficient Std. Error t-Statistic Prob.

TBR -0.521658 0.432586 -1.205906 0.2296

M -3.500625 1.528213 -2.290665 0.0233
CPR 1.160862 0.521347 2.226659 0.0274
C 22.87393 9.466138 2.416395 0.0168
FITTED^2 -0.023118 0.037361 -0.618775 0.5370

R-squared 0.454925 Mean dependent var 4.280384

Adjusted R-squared 0.441212 S.D. dependent var 2.566663
S.E. of regression 1.918635 Akaike info criterion 4.171118
Sum squared resid 585.3042 Schwarz criterion 4.265626
Log likelihood -337.0317 Hannan-Quinn criter. 4.209485
F-statistic 33.17572 Durbin-Watson stat 0.847180
Prob(F-statistic) 0.000000
 Interpretıng The Lınear Regressıon Models 167

CHAPTER 9

INTERPRETING THE LINEAR

REGRESSION MODELS

SIMPLE LINEAR REGRESSION MODEL

(9.1)

We assume that CLRM assumptions are satisfied.

measures the expected change in if the changes one unit.

{ | }
(9.2)
168 Applied Econometrics with Eviews Applications

Example 9.1

Table 9.1 Eviews Output for Simple Regression Model Estimates

Dependent Variable: CONS
Method: Least Squares
Included observations: 30
Variable Coefficient Std. Error t-Statistic Prob.
TEMP 0.003107 0.000478 6.502305 0.0000
C 0.206862 0.024700 8.374902 0.0000
R-squared 0.601593 Mean dependent var 0.359433
Adjusted R-squared 0.587365 S.D. dependent var 0.065791
S.E. of regression 0.042262 Akaike info criterion -3.425533
Sum squared resid 0.050009 Schwarz criterion -3.332120
Log likelihood 53.38299 Hannan-Quinn criter. -3.395649
F-statistic 42.27997 Durbin-Watson stat 0.623564
Prob(F-statistic) 0.000000

Regression result based on the impact of temperature on the ice-

cream consumption is reported in Table 9.1. ,
t values are 8.374902 and 6.502305 respectively. It is
obvious that T- statistics are statistically significant. It means that
there is enough evidence to reject the null hypothesis of there is no
impact.

The coefficient 0.003107 means that one unit increase in the

temperature causes 0.003107 unit change of ice cream consumption.

MULTIPLE LINEAR REGRESSION MODEL

(9.3)

We assume that CLRM assumptions are hold.

The regressor measures the expected change in if the

changes one unit when we hold other regressors constant. This is
called ceteris paribus condition. The interpretation of multiple linear
regression estimates is valid under ceteris paribus condition. It is not
possible to interpret one single regressor estimates without knowing
 Interpretıng The Lınear Regressıon Models 169

the other regressors of the model. If the research focus is to

investigate only the relationship between the regressor and
regressand or the impact of any change in regressor on the
regressand then the other regressors in the regression equation are
called control variables. They should be included in the model to
make estimation robust.

Example 9.2

Table 9.2 Eviews Output for Multiple Regression Model Estimates

Dependent Variable: CONS
Method: Least Squares
Included observations: 30

Variable Coefficient Std. Error t-Statistic Prob.

TEMP 0.003458 0.000446 7.762213 0.0000

PRICE -1.044414 0.834357 -1.251759 0.2218
INCOME 0.003308 0.001171 2.823722 0.0090
C 0.197315 0.270216 0.730212 0.4718

R-squared 0.718994 Mean dependent var 0.359433

Adjusted R-squared 0.686570 S.D. dependent var 0.065791
S.E. of regression 0.036833 Akaike info criterion -3.641296
Sum squared resid 0.035273 Schwarz criterion -3.454469
Log likelihood 58.61944 Hannan-Quinn criter. -3.581528
F-statistic 22.17489 Durbin-Watson stat 1.021170
Prob(F-statistic) 0.000000

Table 9.2 shows the OLS estimation result for regression and includes
three independent variables.
In multivariate regression model,
estimates are interpreted under ceteris paribus condition. It means
that one unit change in temperature causes 0.003458 unit increase in
ice-cream consumption when the other variables price and income
hold constant.

Sometimes, it is very difficult to maintain the other regressors

constant due to possible collinearity between them. For example, if
the regression model includes age and experience together as
170 Applied Econometrics with Eviews Applications

regressors, it is impossible to control experience and keep constant,

while age varies or vice versa. For example:

(9.4)

In the regression model above, it is impossible to conclude that

coefficient measures the effect of age given that age-squared as
constant. In this situation, we should see the following derivative
(write eq. number) which can be interpreted as the marginal effect of
a change in age on the regressand when the other variables
(excluding in the regression model are held constant.

{ | }
(9.5)

Example 9.3

Table 9.3 Eviews Output for Multiple Regression Model Estimates

Dependent Variable: WAGE
Method: Least Squares
Sample: 1 3294
Included observations: 3294

Variable Coefficient Std. Error t-Statistic Prob.

SCHOOL 0.599229 0.033402 17.94010 0.0000

EXPER 0.157261 0.024170 6.506507 0.0000
C -2.476683 0.469997 -5.269573 0.0000

R-squared 0.091489 Mean dependent var 5.757585

Adjusted R-squared 0.090937 S.D. dependent var 3.269186
S.E. of regression 3.116999 Akaike info criterion 5.112529
Sum squared resid 31974.31 Schwarz criterion 5.118084
Log likelihood -8417.335 Hannan-Quinn criter. 5.114517
F-statistic 165.7051 Durbin-Watson stat 1.820012
Prob(F-statistic) 0.000000

If we want to analyze whether the effect of age is different between

men and women, we can include a dummy variable for men as
follows:
 Interpretıng The Lınear Regressıon Models 171

(9.6)

The effect of changing age on income between male and female can
be interpreted as follows:

{ | }
(9.7)

The effect of changing age is for females and for males.

Example 9.4

Table 9.4 Eviews Output for Multiple Regression Model Estimates

Dependent Variable: WAGE
Method: Least Squares
Sample: 1 3294
Included observations: 3294

Variable Coefficient Std. Error t-Statistic Prob.

SCHOOL 0.630602 0.034653 18.19771 0.0000

EXPER 0.203125 0.109467 1.855590 0.0636
EXPER^2 -0.004833 0.006596 -0.732760 0.4638
MALE 1.344831 0.107685 12.48853 0.0000
C -3.576700 0.536916 -6.661567 0.0000

R-squared 0.132729 Mean dependent var 5.757585

Adjusted R-squared 0.131675 S.D. dependent var 3.269186
S.E. of regression 3.046358 Akaike info criterion 5.067287
Sum squared resid 30522.89 Schwarz criterion 5.076546
Log likelihood -8340.822 Hannan-Quinn criter. 5.070602
F-statistic 125.8392 Durbin-Watson stat 1.905606
Prob(F-statistic) 0.000000

Table 9.4 shows the estimation result for the regression equation
specified in Eq.9.7. Now it is possible to interpret marginal effect of a
change in experience on wage when the other independent variables,
schooling and sex are held constant.
172 Applied Econometrics with Eviews Applications

Elasticity measures the relative change in the regressand due to any

relative changes in the regressors. Elasticity is estimated in the linear
regression model includes logarithms of regressors. For example the
coefficient in the following regression model measures the
relative change in regressand due to relative change in the
regressor . Explaining , rather than can be helpful to reduce
heteroscedasticity.

(9.8)

(9.9)

Example 9.5

Table 9.5 Eviews Output for Multiple Regression Model Estimates

Dependent Variable: LOG(WAGE)
Method: Least Squares
Sample: 1 3294
Included observations: 3294

Variable Coefficient Std. Error t-Statistic Prob.

SCHOOL 0.120493 0.006517 18.48886 0.0000

LOG(EXPER) 0.249870 0.071636 3.488034 0.0005
EXPER^2 9.63E-05 0.000612 0.157181 0.8751
MALE 0.242488 0.020440 11.86326 0.0000
C -0.457195 0.121089 -3.775700 0.0002

R-squared 0.138814 Mean dependent var 1.587268

Adjusted R-squared 0.137766 S.D. dependent var 0.622725
S.E. of regression 0.578240 Akaike info criterion 1.743863
Sum squared resid 1099.717 Schwarz criterion 1.753122
Log likelihood -2867.142 Hannan-Quinn criter. 1.747178
F-statistic 132.5375 Durbin-Watson stat 1.879750
Prob(F-statistic) 0.000000

Result in the Table 9.5 indicates that there is a relative change in

wage because of absolute change in school, relative change in
experience, marginal effect of experience and absolute change in
 Interpretıng The Lınear Regressıon Models 173

male. One unit relative change in expertise results 0.071636 unit

relative changes in wage.

If is dummy variable, or any other variable which may take

negative values, then we cannot use log transformation. Regression
model can include some variables in log and some in levels. We
should include in the model as follows:

(9.10)

In this case measures the relative change in due to an absolute

change of one unit in the regressor . If it is a dummy variable for
males, it can be interpreted as relative wage differential between
men and women.

How our interpretations are effected due to the misspecifying the

regressors? Assume that we have following two regression models to
estimate.

(9.11)

(9.12)

If we estimate the first model while, in fact, the second model is the
correct one, then it will increase the variance of the estimators for
the relevant regressors. Thus we will get higher variance and less
reliable estimates.

If the second model is estimated instead of first model which is

correct, then it will cause omitted variable bias. In this case the error
term includes the omitted regressor. Then Assumption 2
(unbiasedness, Assumption 3
(efficiency ) do not hold Gauss- Markov theorem
does not apply, and OLS is not BLUE; estimates of regression
coefficient is biased. OLS overestimate the coefficient of independent
variable .
174 Applied Econometrics with Eviews Applications

REGRESSOR SELECTION
To overcome the problem of including irrelevant regressors and
omitting relevant regressors in the regression model following
methods can be used together or separately:

 The potentially relevant variables should be determined

based on the related economic theory
 The relevant regressor should be defined by checking the
literature and using the ‘common sense’
 The regressors should be determined based on research
questions and research focus
 ‘From specific to general’ approach may be used: the
approach suggests to form regression model as simple as
possible and add regressors until the adequate model
specification is obtained
 ‘General to specific modeling’ (Charemza and Deadman,
1999) approach may also be used: it starts with very
general specification which is called general unrestricted
model and reduces size and number of regressors in each
step by testing restrictions until the model with adequate
specification is obtained. Besides, including more variables
and starting with the general model can cause
multicollinearity problem which misguides the
researchers.

If two regressors are omitted from the model, joint test should be
used rather than two tests separately.

Best option is to start with the optimum model and test, whether the
restrictions imposed by the model are correct, and whether the
restrictions not imposed by the model should be imposed.

In the first category misspecification test for omitted variable bias,

autocorrelation and heteroscedasticity test can be employed.
Parameter restriction test, such as, whether one or more regressors
have zero coefficients can be employed.
 Endogeneıty 175

CHAPTER 10

ENDOGENEITY

In a statistical model, a parameter or variable is said to be

endogenous when there is a correlation between the independent
variable and the error term.

Endogeneity can arise as a result of the measurement error, auto

regression with auto correlated errors, simultaneity and omitted
variables. Broadly, a loop of, bivariate causality between the
independent and dependent variables of a model leads to
endogeneity.

Tariffs are an example of endogeneity problem.

(10.1)

(10.2)

The current value of depends on the current value of , is

influenced by current shocks to as follows:

(10.3)

Thus, it causes endogeneity problem.

This is the violation of OLS assumption.
176 Applied Econometrics with Eviews Applications

In the presence of endogeneity, the linear model does not

correspond to a conditional expectation and OLS cannot produce
unbiased and consistent parameter estimates. Ceteris paribus is not
valid and OLS estimator is biased and inconsistent and does not
produce best linear approximation. Hypotheses tests are misleading.
It can cause rejecting a hypothesis that is true (Type I Error) and fail
to reject a hypothesis that in fact is false (Type II Error). For example,
inflation is independent of all other factors within a given period, but
influenced by the previous years’ export and interest rate. Then we
can say that inflation is exogenous within the period, but endogenous
over time.

OLS is no longer BLUE if endogeneity exists.

SOURCES OF ENDOGENEITY
Autocorrelation with Lagged Dependent Variable

(10.4)

The OLS estimator provided that { } { } and

the other assumptions are met. However, if is depended on the
first order autocorrelation as follows:

(10.5)

we can rewrite the above model as:

(10.6)

and we also can write:

(10.7)

Indicating that the error term is correlated with . Thus, if

, OLS method does not produce consistent estimators.
 Endogeneıty 177

A possible solution is the use of maximum likelihood or instrumental

variables techniques. Durbin–Watson test is not valid to test
autocorrelation in the model (Eq.10.7), because the condition that
the explanatory variables should be treated as deterministic is
violated. An alternative test, Breusch–Godfrey Lagrange Multiplier
test for autocorrelation can be applied. This test statistic is computed
as T times the R2 of a regression of the least squares residuals et on
et−1 and all included explanatory variables (including the relevant
lagged values of yt ). Under H0, the test statistic asymptotically has a
Chi-squared distribution with 1 degree of freedom.

OLS is inconsistent whenever the model which is being estimated

does not correspond to a conditional expectation. For example, a
lagged dependent variable, combined with autocorrelation of the
error term.

Omitted Variable Bias

Let’s assume an independent variable (z) is correlated with another

independent variable (x) and error term in the model.

(10.8)

Omit from the model and run the following regression:

, (10.9)

(10.10)

In this new regression model is correlated with error term

(because there is correlation between and the omitted variable i.
Multicollinearity problem occurs.

Measurement Error

If true value of one independent variable is not available, instead of

following is defined:

, (10.11)
178 Applied Econometrics with Eviews Applications

is the measurement error, and the following regression model is

estimated as:

, (10.12)

Actually, following regression is estimated mistakenly:

(10.13)

(10.14)

(10.15)

and ui depend on , they are correlated and OLS estimator

is inconsistent. That is, { } and one of the necessary
conditions for consistency of OLS estimator is violated.

Simultaneity

Simultaneity mostly occurs in the dynamic models.

Suppose that we have two structural equations as follows:

, (Keynesian consumption function) (10.16)

Where denotes a country`s real per capita consumption and is

real capita income. has a causal interpretation reflecting the
impact of income on consumption ( ). It measures how
much individuals increase their consumption if their income increases
by one unit.

is also defined as:

(10.17)

Where is real per capita investment? This equation says that the
total consumption plus total investment should equal total income in
a closed economy without government intervention.
 Endogeneıty 179

Assume that Assumption 11 holds, which says that is i.i.d. over

time with zero mean, variance and are independent
( { } ). This assumption says that investment is
exogeneous and determined independently of the error term. In
contrast, both and are endogeneous variables, which are
jointly (simultaneously) determined in the model. The above model
in Eq. 10.16, and Eq. 10.17, is a simple simultaneous model in
structural form (structural model).

is endogenous in the consumption function Eq.10.16. Because

influences through Eq.10.17. We can no longer argue that and
are uncorrelated. Consequently, the OLS estimator for will be
inconsistent.

To overcome this problem, the reduced form of this model may be

considered, in which the endogeneous variables and are
expressed as a function of the exogeneous variable and the error
term as follows.

(10.18)

(10.19)

The usual assumptions of error term cannot be applied to above

reduced form model.

ENDOGENEITY TEST
Is there evidence that correlation between the potentially
endogenous variables and the error term is strong enough to result in
substantively biased estimates?

Instrument relevance test: Are the instruments sufficiently strongly

correlated with the potentially endogenous variables?

Exogeneity/excludability of instruments: Are the instruments

genuinely uncorrelated with the main equation residuals?
180 Applied Econometrics with Eviews Applications

In the presence of weak instruments, the TSLS estimator can actually

produce worse results than OLS. First step is the identify instruments
which are strongly correlated with the endogeneous variables.

Hausman (1978) test can be used for endogeneity of independent

variables.

i- Run the first stage regressions using OLS and save the
residuals. This is called reduced form equation

ii- Assume that the initial model is

(10.20)

iii- Include the residuals obtained from first stage regression as

additional regressors in the main equation and estimate by
using OLS.

iv- Test the joint significance of the first stage residuals.

Instruments should also be exogeneous with regard to the

dependent variable in the main equation and excludable from the
main regression equation.

The test will analyze whether or not there is a correlation between

the independent, and that part of the suspect variables variation that
is not explained by exogeneous factors.
 Endogeneıty 181

SOLUTION
Ad hoc Approaches

- A proxy can be used that does not suffer from the

endogeneity problem.

- The suspected variable can be lagged by one or more periods.

This class of solution has some advantages. It is easy to implement

and does not require additional variables or data.

But it has also following disadvantages:

- Interpretation of proxy variable is difficult.

- It is difficult to figure out the level of endogeneity concern

and adequacy of the solution.

Instrumental Variables Estimation

The best way to deal with endogeneity problem is through

instrumental variables (IV) techniques and the most common IV
estimator is the Two Stage Least Squares (TSLS) estimates.

An exogenous variable (instrument) that is strongly correlated with

the potentially endogenous regressor should be found. The
instrument only influences the dependent variable through the
potentially endogenous independent variable. Selection of
appropriate instruments is a crucial issue.

When the model is interpreted as a conditional expectation, the

ceteris paribus condition only refers to the included variables, while
for a causal interpretation it also includes the unobservables (omitted
variables) in the error term.

An instrumental variable z2i, is assumed to be uncorrelated with the

model error εi but correlated with the endogenous regressor x2 . The
main problem is to find appropriate instrument. Another problem is
182 Applied Econometrics with Eviews Applications

that standard errors of instrumental variables estimators are typically

quite high compared to those of the OLS estimator. The most
important reason for this is that the instrument and regressor have a
low correlation.

Advantages:
 Rigor and transparency;

 Amenability to empirical testing, appropriateness of the

instruments;

(10.21)

contains some endogeneous variables

(10.22)

Z is the matrix of all endogenous variables in the model, and the

number of instruments should be equal to the number of
endogenous variables.

The IV estimate is:

∑ ̅ ̅
̂ (10.23)
∑ ̅ ̅

The R2 in IV is not useful since it can also be negative. Interpretation

of it is not valid.

Example 10.1

Consider a sub-sample of subjects for whom we wish to understand

whether college degree have an effect on wages. Unfortunately, to
understand the effect of a college degree, we need to have a proxy
for the subjects’ intrinsic ability. Intrinsic ability may influence the
likelihood of obtaining a college degree.
 Endogeneıty 183

The intrinsic ability may also influence the wages you obtain. As a
result a positive estimate on return on college degree may be
attributed.

We wish to understand the impact of education on wages. We are

unable to measure individuals non-education based capabilities,
which not only influence wages but also the choice and ability to
complete a degree. Here these capabilities are the unobserved
heterogeneity causing bias in the estimated effect of education on
wage. In this case probably a positive impact since they are boosting
the estimated effect of education.

The problem using an OLS in cases which suffers from endogeneity is

that the error term and the explanatory variables become correlated.

(10.24)

This is caused by the unobserved element (omitted variable) since it

is hidden in the error term.

Instruments (zi) are variables used to explain a variable we suspect of

being endogenous and which are exogenous with respect to the main
equation;

(20.25)

(10.26)

For example25; Suppose we explain an individual’s log wage by a

number of personal characteristics, x1i , as well as the number of
hours person i is working (x2i) by means of a linear model as follows:

(10.27)

In the Eq. 10.27, εi includes all unobservable factors that affect a

person’s wage, including things like ‘ability’ or ‘intelligence’. Typically,
it is obvious that the number of hours a person is working partly also

25
Verbeek, M. (2004)
184 Applied Econometrics with Eviews Applications

depends on these unobserved characteristics. If this is the case, OLS

is consistently estimating the conditional expected value of a
person’s wage given, among other things, how many hours he or she
is working, but not consistently estimating the causal effect of
working hours. That is, the OLS estimate for 2 would reflect the
difference in the expected wage of two arbitrary persons with the
same observed characteristics in x1i , but working x2 and x2 + 1 hours,
respectively.

In the above example, variable that is correlated with working hours

is x2i but not correlated with the unobserved ‘ability’ factors that are
included in εi. Variables relating to the composition of one’s family
may serve as instrumental variables.

The assumptions captured in the moment conditions are identifying.

So that, they cannot be tested statistically. Besides, over identifying
restrictions can be tested when there are more conditions than
actually needed for identification.

If the instrument z2i is valid, then the endogeneity of x2i can be

tested. Housman (1978) proposes to compare the OLS and IV
estimators for . Assuming E{εizi} = 0, the IV estimator is consistent.
If, in addition, E{εi,x2i} = 0, the OLS estimator is also consistent and
should differ from the IV one by sampling error only.

In the first step, estimate a regression explaining x2i from x1i and z2i,
and save the residuals, save . This is the reduced form equation.
Next, add the residuals to the model of interest and estimate the
following regression by OLS.

(10.28)

If , is exogenous. We can easily test the endogeneity of

by performing a standard t-test on in the above regression.
This test requires the assumption that the instrument is valid and
therefore does not help to determine which identifying moment
condition, { } { } is appropriate.
 Endogeneıty 185

To define the variables correlated with the endogenous variable but

uncorrelated with the part of the error term that is due to the
unobserved heterogeneity is a big challenge. A good instrument
should be correlated with the key independent variable, but not with
the main equation dependent variable. If there are four endogenous
regressors, there should be at least four different instruments.

In a simultaneous equations context, sufficient instruments should be

available in the system (order condition for identification). If there
are five exogenous variables in the system that is not included in the
equation of interest, there can be up to five endogenous regressors.
If there is only one endogenous regressor, there are five different
instruments to choose from. It is also possible and advisable to
estimate more efficiently by using all the available instruments
simultaneously26.

Two Stage Least Squares

Whether the instruments chosen are valid and IV estimation is

necessary can be checked by the help of TSLS:

- In the first step the reduced form is estimated by OLS (that is:
a regression of the endogenous regressors upon all
instruments).

- In the second step the original structural equations are

estimated by OLS, while replacing all endogenous variables on
the right hand side with their predicted values from the
reduced form.

Endogeneity test: Is there evidence that correlation between the

potentially endogenous variables and the error term is strong enough
to result in substantively biased estimates?

Instrument relevance test: Are the instruments sufficiently strongly

correlated with the potentially endogenous variables?

26
Verbeek, M. (2004)
186 Applied Econometrics with Eviews Applications

Exogeneity/excludability of instruments: Are the instruments

genuinely uncorrelated with the main equation residuals?

Are the instruments individually statistically significant?

Are their signs and magnitudes sensible?

Are the instruments jointly statistically significant? Look for a high F-

statistic.

Weak Instrument

Simple OLS can produce better results than TSLS estimator if there is
weak instrument issue. Whether the instruments are strongly
correlated enough with the potentially endogenous variables should
be checked in the first step.

To test for instrument relevance, make sure to run the first stage
regressions of the potentially endogenous variables on all of the
exogenous variables. The properties of the IV estimator can be very
poor, and the estimator can be severely biased, if the instruments
exhibit only weak correlation with the endogenous regressor(s). Even
if the sample size is large the normal distribution may provide a very
poor approximation to the true distribution of the IV estimator.

Generalized Method of Moment

This approach estimates the model parameters directly from the

moment conditions imposed by the model. These conditions can be
linear or nonlinear in the parameters and the number of moment
conditions should be at least as many as the number of unknown
parameters for identification.

The advantages of the generalized method of moments are:

i- It does not require normality assumption.

ii- It allows for heteroscedasticity of unknown form.

 Endogeneıty 187

iii- It can estimate parameters of variables even if the

model cannot be solved analytically from the first
order conditions.

The GMM concept is commonly used to estimate and test asset

pricing models. An asset pricing model, for example the CAPM should
explain the variation in expected returns for different risky
investments. Because some investments are more risky than others,
investors may require compensation for bearing this risk by means of
a risk premium. This leads to variation in expected returns across
different assets27.

27
Cochrane (2001).
188 Applied Econometrics with Eviews Applications
 Regressıons wıth Dummy Varıables 189

CHAPTER 11

REGRESSIONS WITH DUMMY

VARIABLES

It is also called a binary variable. A dummy variable takes only two

values, 0 or 1.

A DUMMY INDEPENDENT VARIABLE

A dichotomous factor can be entered into a regression equation by
formulating a dummy regressor, giving 1 for one category of the
factor and 0 for the other category. Dummy variables also are a way
of turning qualitative variables into quantitative variables. Once the
variables are quantitative, then the correlation and regression
techniques can be used.

Regression with dummy explanatory or independent variables is

extremely common and the interpretation of coefficient estimates is
different from the other variables. Regression with dummy
explanatory variables is closely related to Analysis of Variance (or
ANOVA for short). ANOVA is rarely used in economics, but it is an
extremely common tool in other social and physical sciences such as
sociology, education, medical statistics, and epidemiology.
190 Applied Econometrics with Eviews Applications

Regression with dummy variables is a more general and more

powerful tool than ANOVA. If you know how to use and understand
regression, then you have no need to learn about ANOVA.

Let’s assume following regression model includes a continues

independent variable and a dummy independent variable:

(11.1)

It can be interpreted as an intercept shift as follows:

(11.2)

(11.3)

𝐷 𝑦𝑖 𝛽 𝛽 𝛽 𝑥𝑖
y
𝐷 𝑦𝑖 𝛽 𝛽 𝑥𝑖

x
Figure 11.1 Dummy Independent Variable Regression Line

INTERPRETATION
We can say that β is a measure of how much Y tends to change when
X is changed by one unit. But, with the present dummy explanatory
 Regressıons wıth Dummy Varıables 191

variable a “one unit” change implies a change from “No air

conditioner” to “Having an air conditioner”. That is, we can say
“houses with an air conditioner tend to be worth $25,996 more than
houses without an air conditioner.”

QUALITATIVE VARIABLE WITH MORE THAN TWO

CATEGORIES (POLYTOMOUS FACTORS)
Dummy variables can be used as a control in multiple categories.

(11.4)

Example 11.1 Schooling28

Table 11.1 Eviews Output for Regression With Dummy Variable

Dependent Variable: LWAGE76
Method: Least Squares
Included observations: 3010

Variable Coefficient Std. Error t-Statistic Prob.

ENROLL76 -0.070539 0.026723 -2.639635 0.0083

EXP76 0.053647 0.007278 7.371461 0.0000
EXP762 -0.002453 0.000355 -6.905773 0.0000
BLACK -0.326035 0.018211 -17.90325 0.0000
C 6.103856 0.034063 179.1914 0.0000

R-squared 0.111505 Mean dependent var 6.261832

Adjusted R-squared 0.110323 S.D. dependent var 0.443798
S.E. of regression 0.418602 Akaike info criterion 1.097867
Sum squared resid 526.5590 Schwarz criterion 1.107850
Log likelihood -1647.290 Hannan-Quinn criter. 1.101458
F-statistic 94.28118 Durbin-Watson stat 1.715153
Prob(F-statistic) 0.000000

We can, however, say that = -0,326 is a measure of the impact of

the race on the wage. In other words, if we compare two workers

28
Verbeek, M. (2004)
192 Applied Econometrics with Eviews Applications

with the same experience and number of schooling, black people will
always get $0,326 lower wage than the white people with same
personal characteristics.

Example 11.2

Suppose everyone from your data is a University dropout, bachelor

graduates or master graduates. In order to make comparisons
between bachelor and master graduates with university dropouts,
you need to include two dummy variables, as follows:

UNGRAD= 1 if university graduates only,

MGRAD=1 if master graduates.

If regression equation includes more dummy variables:

(11.5)

Example 11.3

Duncan data: Regressing occupational prestige on income and

education produces the following regression equation:

(11.6)

Dependent variable:

Quantitative independent variables:

Qualitative independent variable:

type (bc, prof, wc)

 Regressıons wıth Dummy Varıables 193

The three-category occupational-type factor can be represented in

the regression equation by introducing two dummy regressors,
employing the following coding scheme:

Type D1 D2
Blue collar (bc) 0 0
Professional (prof) 1 0
White collar (wc) 0 1

Note: If there are p categories, we should use p − 1 dummy

regressors. Blue collar implicitly serves as the baseline category to
which the other occupational-type categories are compared.

For Blue collar:

(11.7)

For Professional:

(11.8)

(11.9)

For white collar:

(11.10)

(11.11)

CONSTRUCTING INTERACTION REGRESSORS

Two explanatory variables may interact in determining a response
variable when the partial effect of one depends on the value of the
other. If these regressions are not parallel, then the factor interacts
with one or more of the quantitative explanatory variables. The
dummy-regression model may be constructed to reflect these
interactions.
194 Applied Econometrics with Eviews Applications

Together with the quantitative regressor x and the dummy regressor

D, an interaction regressor xD may be included into the regression
equation. The interaction regressor is the product of the other two
regressors; although xD is a function of x and D, not linear, and
perfect collinearity is avoided.

(11.12)

It causes a change in the slope of the estimated regression line.

Example 11.4

Table 11.2 Eviews Output for Regression with Interaction Variable

Dependent Variable: PRICE
Method: Least Squares
Sample: 1 546
Included observations: 546

Variable Coefficient Std. Error t-Statistic Prob.

LOTSIZE 5.336944 0.404851 13.18248 0.0000

BEDROOMS 5829.629 1195.020 4.878270 0.0000
BATHRMS 18959.39 1764.674 10.74385 0.0000
RECROOM 16466.72 7353.490 2.239306 0.0255
C -2713.233 3831.495 -0.708140 0.4792
LOTSIZE*RECROOM -1.202176 1.230103 -0.977298 0.3289

R-squared 0.505477 Mean dependent var 68121.60

Adjusted R-squared 0.500898 S.D. dependent var 26702.67
S.E. of regression 18864.68 Akaike info criterion 22.53890
Sum squared resid 1.92E+11 Schwarz criterion 22.58618
Log likelihood -6147.119 Hannan-Quinn criter. 22.55738
F-statistic 110.3921 Durbin-Watson stat 1.411099
Prob(F-statistic) 0.000000

Other type of dummy variables is a seasonal dummy, structural

breaks, or shocks.
 Regressıons wıth Dummy Varıables 195

Why is a qualitative independent variable needed?

i- If the research aims to find out the effect of a qualitative

independent variable (For example: Do men earn more than
women?)

ii- To better predict/describe the dependent variable. The errors

can be made smaller by including variables like gender, race,
etc.

iii- Qualitative variables may be confounding factors. Omitting

them can cause biased estimates of other coefficients.

DUMMY DEPENDENT VARIABLE

In the case of dummy dependent variable, there are better
estimation methods than OLS. The two main alternatives are termed
Logit and Probit.

TYPES OF VARIABLES29
Continuous or Quantitative Variables

Interval - Scale Variables

Interval scale variables take on positive or negative values. The

intervals which are equally ordered keep the same importance
throughout the whole scale. They can be used for quantification and
comparision the magnitudes of of differences. For instance, 40°C is
higher than 30°C, and an increase from 30°C to 40°C is twice as much
as the increase from 30°C to 35°C. Counts are also interval scale
measurements (number of publications or citations, years of
education, etc).

29
http://www.unesco.org
196 Applied Econometrics with Eviews Applications

Continuous Ordinal Variables

When a scale is transformed by an exponential, logarithmic or any

other nonlinear transformation, it loses its interval - scale property.
then, the observations should be ordered by their ranks.
Ratio - Scale Variables

Ratio - Scale variables are continuous positive measurements on a

nonlinear scale. For instance, the growth of bacterial population (say,
with a growth function AeBt). In this model, equal time intervals
multiply the population by the same ratio.

Qualitative or Discrete Variables

Discrete variables which are also called categorical variables take a

finite number of numerical values, categories or codes and classified
as follows;

- Nominal variables
- Ordinal variables
- Dummy variables
- Preference variables
- Multiple response variables

Nominal Variables

Nominal variables which are used for qualitative classification has no

order and measured only in terms of certain category. For instance;

Gender:
1. Male 2.Female

Marital Status:
1. Single 2. Married 3. Divorcee 4. Widower

Ordinal Variables

An ordinal variable is a kind of nominal variable. Three-, five-, or

seven- point scales answers are used for evaluation of relative
 Regressıons wıth Dummy Varıables 197

magnitude of quality, importance or relevance in social and

behavioral research.

For instance;
1. The economic status of families in the society might be 'upper
lower' is lower than 'middle', but 'how much higher' is not
known.
2. A question in a questionnaire related with the time
involvement of employees in the social activities to measure
commitment and satisfaction. The respondents show their
involvement by selecting one of the following answers:

1 = Very low
2 = Low
3 = Medium
4 = High
5 = Very high

The variable Involvement is an ordinal variable with 5 points scale.

Ordinal variables can be treated as nominal variables or scale

variables.

Categorical Variables

A categorical variable can be obtained from quantitative variables by

recoding them. For instance, the quantitative variable Income can be
classified into four intervals such as;

[Up to 500 KM] 1

[500 KM, 1000 KM] 2
[1000 KM, 2000 KM] 3
[2000 KM, 4000 KM] 4
[Above 4000 KM] 5

Preference Variables

Preference variables are the values either in a decreasing or

increasing order. For instance, in a survey, a question may be asked
198 Applied Econometrics with Eviews Applications

to respondent to grade the importance of the predetermined

difficulties of doing research by using the number codes from 1 to 5
for the most important difficulty to the least important difficulty.

Multiple Response Variables

Multiple response variables can take more than one value. For
instance, a survey question regarding the purpose of using computers
in research. The respondents could score more than one category.
 Tıme Serıes 199

CHAPTER 12

TIME SERIES

STATIC MODELS

Suppose that we have time series data for two variables dated
contemporaneously. A static model can be written as follows:

(12.1)

Static models are used to model contemporaneous relationship

between two variables when a change in x at time t is believed to
have an immediate effect on y. Static regression models are also used
when exploring the tradeoff between two variables.

Example 12.1

Static Phillips curve, defined by:

(12.2)

Where, Inft is the inflation rate, and Unempt is the unemployment

rate. This form of the Phillips curve assumes a constant natural rate
of unemployment and constant inflationary expectations, and it can
be used to study the contemporaneous tradeoff between them.
[Mankiw (1994, Section 11.2).]. There can be several explanatory
variables in a static regression model.
200 Applied Econometrics with Eviews Applications

DYNAMIC MODELS
One way to model the dynamic relationships is to include lagged
values of regressors on the right hand side of the regression
equation; this is the basis of the distributed- lag model. The
distributed-lag model takes the form:

(12.3)

Given the model above, immediate impact on y is given as , ,

the impact on y after one period, while is the impact on y after
two periods. The final impact on y is that occurs after k periods. It
takes k periods for the full effects of the impulse to be completed.
The sequence of coefficients constitutes the impulse response
function.

TIME SERIES
Time Series Regression with Lag

The researcher which is using time series data faces two problems
which do not exist in the cross sectional data:

i- One time series variable can influence another with a

time lag; and

ii- If the variables are non-stationary, spurious regression

may arise.

Non-stationary time series variables should be transformed into

stationary before running a regression using unit root tests for every
variable in the regression equation.

The value of the dependent variable at a given point of time can

depend not only on the value of the explanatory variable at that time
period, but also on values of the explanatory variable in the past. The
 Tıme Serıes 201

simplest model to incorporate dynamic effects is known as the

distributed lag model.

A regression model helps in measuring the effect of one or more

independent variables on the dependent variable. In the case of time
series data, the effect of some explanatory variables on the
dependent variable may take time.

Example 12.230

Table 12.1 Eviews Output for Time Series Regression with Lags
Dependent Variable: Y
Method: Least Squares
Sample (adjusted): 2005M07 2010M02
Included observations: 56 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

X -131.9943 47.43609 -2.782571 0.0076

X__1_ -449.8597 47.55659 -9.459460 0.0000
X__2_ -422.5183 46.77785 -9.032445 0.0000
X__3_ -187.1041 47.64089 -3.927384 0.0003
X__4_ -27.77104 47.66190 -0.582668 0.5627
C 91173.32 1949.850 46.75914 0.0000

R-squared 0.759855 Mean dependent var 74067.00

Adjusted R-squared 0.735840 S.D. dependent var 10468.82
S.E. of regression 5380.606 Akaike info criterion 20.11995
Sum squared resid 1.45E+09 Schwarz criterion 20.33695
Log likelihood -557.3585 Hannan-Quinn criter. 20.20408
F-statistic 31.64143 Durbin-Watson stat 2.240889
Prob(F-statistic) 0.000000

The result for time series analysis with lag is reported in Table 12.1.
Inflation in Turkey is effected by its own past values up to three lags.
The inflation four months ago does not have any impact on this
month´s inflation level.

30
Koop, G. (2005).
202 Applied Econometrics with Eviews Applications

Lag Selection

If the lag length is not clear, there are many different approaches to
lag length selection in the econometrics literature.

Step1. Choose the maximum possible lag length, qmax, that seems
reasonable to you.

Step2. Estimate the distributed lag model

(12.4)

If the p-value for testing is less than the

significance level you choose (e.g. 0.05) then go no further.
Use qmax as lag length. Otherwise go on to the next step,

Step3. Estimate the distributed lag model

(12.5)

If the p-value for testing is less than the

significance level you choose (e.g. 0.05) then do not go
further. Use qmax -1 as lag length. Otherwise go on to the
next steps

Example 12.3

We estimate that five months is the maximum time period to expect

that training may impact on losses due to accidents. We need to start
with estimating a distributed lag model with lag length equal to 5
(qmax=5). Results are given in Table 12.2.
 Tıme Serıes 203

Table 12.2 Eviews Output for Distributed Lag Model

Dependent Variable: Y
Method: Least Squares
Included observations: 55 after adjustments
Variable Coefficient Std. Error t-Statistic Prob.
X -145.3045 48.18577 -3.015506 0.0041
X__1_ -461.8515 48.21403 -9.579192 0.0000
X__2_ -423.3032 47.84485 -8.847413 0.0000
X__3_ -199.4649 48.25000 -4.133988 0.0001
X__4_ -36.26030 48.27840 -0.751067 0.4563
X__5_ 5.332677 47.95305 0.111206 0.9119
C 91903.04 2207.578 41.63072 0.0000
R-squared 0.770395 Mean dependent var 74023.99
Adjusted R-squared 0.741695 S.D. dependent var 10560.31
S.E. of regression 5367.145 Akaike info criterion 20.13239
Sum squared resid 1.38E+09 Schwarz criterion 20.38787
Log likelihood -546.6408 Hannan-Quinn criter. 20.23119
F-statistic 26.84252 Durbin-Watson stat 2.235865
Prob(F-statistic) 0.000000

Since the p-value corresponding to the explanatory variable xt-5 is

greater than 0.05 we cannot reject null hypothesis that at the
5% significance level. Accordingly we drop this variable from the
model and re-estimate with lag length set equal to 4; providing the
result in Table 12.3.

Table 12.3 Eviews Output for Distributed Lag Model

Dependent Variable: Y
Method: Least Squares
Sample (adjusted): 5 60
Included observations: 56 after adjustments
Variable Coefficient Std. Error t-Statistic Prob.
X -131.9943 47.43609 -2.782571 0.0076
X__1_ -449.8597 47.55659 -9.459460 0.0000
X__2_ -422.5183 46.77785 -9.032445 0.0000
X__3_ -187.1041 47.64089 -3.927384 0.0003
X__4_ -27.77104 47.66190 -0.582668 0.5627
C 91173.32 1949.850 46.75914 0.0000
R-squared 0.759855 Mean dependent var 74067.00
Adjusted R-squared 0.735840 S.D. dependent var 10468.82
S.E. of regression 5380.606 Akaike info criterion 20.11995
Sum squared resid 1.45E+09 Schwarz criterion 20.33695
Log likelihood -557.3585 Hannan-Quinn criter. 20.20408
F-statistic 31.64143 Durbin-Watson stat 2.240889
Prob(F-statistic) 0.000000
204 Applied Econometrics with Eviews Applications

We also should drop the explanatory variable xt-5 from the model and
re-estimate with lag length equal to 3, giving the result in Table 12.4.

Table 12.4 Eviews Output for Distributed Lag Model

Dependent Variable: Y
Method: Least Squares
Sample (adjusted): 4 60
Included observations: 57 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

X -125.9000 46.24049 -2.722722 0.0088

X__1_ -443.4918 45.88164 -9.665999 0.0000
X__2_ -417.6089 45.73324 -9.131409 0.0000
X__3_ -179.9043 46.25205 -3.889650 0.0003
C 90402.22 1643.183 55.01653 0.0000

R-squared 0.757447 Mean dependent var 74153.74

Adjusted R-squared 0.738789 S.D. dependent var 10395.57
S.E. of regression 5313.050 Akaike info criterion 20.07735
Sum squared resid 1.47E+09 Schwarz criterion 20.25657
Log likelihood -567.2045 Hannan-Quinn criter. 20.14700
F-statistic 40.59659 Durbin-Watson stat 2.234934
Prob(F-statistic) 0.000000

The p-value for testing is much less than 0.05. We therefore

conclude that the lag three is belonging to the model. Hence q=3 is
the lag length we choose for this model reported in Table 12.4.

Distributed lag models have the dependent variable depending on an

explanatory variable and time lags of the explanatory variable. If the
variables in the distributed lag model are stationary, then OLS
estimates are reliable and the statistical techniques of multivariate
regressions (e.g. looking at p-values or confidence intervals) can be
used in the interpretation of the estimation results. The lag length in
a distributed lag model can be selected by sequentially using t-tests
beginning with a reasonable large lag length.
 Tıme Serıes 205

Finite Distributed Lag Models

In a finite distributed lag (FDL) model, we allow one or more variables

to affect y with a lag. For example, for annual observations, consider
the model below:

(12.6)

Where gfrt is the general fertility rate and pet is the real dollar value
of the personnel tax exemption. The objective is to see whether, in
aggregate, the decision to have children is linked to tax value of
having a child. The decision would not be based on the immediate
impact from the changes in the personnel exemption.

All of the functional forms can be used in time series regressions.

When using time series data in a regression, the relationship
between y and x may be concurrent or x may serve as a leading
indicator. Past values of x appear as a predictor, either with or
without the current value of x.

Example 12.5

The effect of advertising on sales takes time to be analyzed and it is

cumulative. It can be modeled as follows:

(12.7)

Where, xt is advertising in the current month and the lagged variables

xt-1, xt-2 and xt-3 represents advertising in the two previous months.

Autoregressive Modeling (AR)

It is a regression model where the independent variables are lags of

the dependent variable (an autoregression is a regression of a
variable on lags of itself).

Let’s start with an autoregressive model with the explanatory

variable which is one period lag of dependent variable. This is called
the AR(1) model:
206 Applied Econometrics with Eviews Applications

(12.8)

Implies the type of trend behavior is non-stationary. The

other values of imply stationary behavior. This allows us to provide
a formal definition of the concepts of stationarity and
nonstationarity, at least for the AR(1) model: For the AR(1) model, we
can say that Y is stationary if | φ | < 1 and is non-stationary if φ = 1.
The other possibility, | φ | > 1, is formally, “non-stationary” merely
means “anything that is not stationary. At this stage it is useful to
think of a unit root as implying φ = 1 in the AR(1) model.

Stationarity

Following are different ways to test whether a time series variable, y,

is stationary or has a unit root:

i- In the AR(1) model, if φ = 1, then y has a unit root. If |φ| < 1

then y is stationary.

ii- If y has a unit root, then its autocorrelations with past values
is close to one and does not change as lag length increases.

iii- Non-stationary time series have a long memory, bur

stationary time series do not have long memory.

iv- Non-stationary time series have trend behavior.

v- If y has a unit root, then the first difference of y, Δy, is

mostly stationary. Non-stationary time series are often
referred to as difference stationary series.

vi- If we subtract yt-1 from both sides of the equation in the

AR(1) model, as follows:
if φ = 1, then ρ = 0 and the previous equation is written in
terms of Δyt. It means that the first difference, Δyt fluctuates
randomly around α. Whether a series has a unit root can be
tested for ρ= 0. Furthermore, a time series is stationary if -1
< φ < 1 which is equivalent to -2 <ρ < 0. This is called as the
stationarity condition.
 Tıme Serıes 207

Example 12.6

Table 12.6 Eview Outputs for Unit Root Test

Dependent Variable: CROBEX
Method: Least Squares
Sample (adjusted): 1997M10 2012M12
Included observations: 183 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

C 29.85194 26.16873 1.140748 0.2555

CROBEX(-1) 0.984848 0.012520 78.65950 0.0000

R-squared 0.971578 Mean dependent var 1791.031

Adjusted R-squared 0.971421 S.D. dependent var 1083.960
S.E. of regression 183.2469 Akaike info criterion 13.27041
Sum squared resid 6077875. Schwarz criterion 13.30549
Log likelihood -1212.243 Hannan-Quinn criter. 13.28463
F-statistic 6187.316 Durbin-Watson stat 1.723101
Prob(F-statistic) 0.000000

Null Hypothesis: CROBEX has a unit root

Exogenous: Constant
Lag Length: 0 (Automatic based on SIC, MAXLAG=13)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -1.210194 0.6699

Test critical values: 1% level -3.466176
5% level -2.877186
10% level -2.575189

*MacKinnon (1996) one-sided p-values.

Augmented Dickey-Fuller Test Equation
Dependent Variable: D(CROBEX)
Method: Least Squares
Included observations: 183 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

CROBEX(-1) -0.015152 0.012520 -1.210194 0.2278

C 29.85194 26.16873 1.140748 0.2555

R-squared 0.008027 Mean dependent var 2.755792

208 Applied Econometrics with Eviews Applications

Adjusted R-squared 0.002546 S.D. dependent var 183.4806

S.E. of regression 183.2469 Akaike info criterion 13.27041
Sum squared resid 6077875. Schwarz criterion 13.30549
Log likelihood -1212.243 Hannan-Quinn criter. 13.28463
F-statistic 1.464570 Durbin-Watson stat 1.723101
Prob(F-statistic) 0.227783

Null Hypothesis: D(CROBEX) has a unit root

Exogenous: Constant
Lag Length: 0 (Automatic based on SIC, MAXLAG=13)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -11.74614 0.0000

Test critical values: 1% level -3.466377
5% level -2.877274
10% level -2.575236

Augmented Dickey-Fuller Test Equation

Dependent Variable: D(CROBEX,2)
Method: Least Squares
Included observations: 182 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

D(CROBEX(-1)) -0.868009 0.073897 -11.74614 0.0000

C 2.558105 13.55653 0.188699 0.8505

R-squared 0.433912 Mean dependent var -0.124615

Adjusted R-squared 0.430767 S.D. dependent var 242.3693
S.E. of regression 182.8616 Akaike info criterion 13.26626
Sum squared resid 6018904. Schwarz criterion 13.30147
Log likelihood -1205.230 Hannan-Quinn criter. 13.28054
F-statistic 137.9719 Durbin-Watson stat 2.018993
Prob(F-statistic) 0.000000

Unit root test result given in Table 12.6 indicate that Croatian stock
market index is stationary in first difference (I(1)).

Random Walk

Consider the case where φ = 1 (or, equivalently, ρ= 0) and α = 0. In

this case the AR (1) model can be written as:
 Tıme Serıes 209

(12.9)

This is referred to as the random walk model. Since φ = 1, y has a

unit root and is non-stationary. Random walk model is commonly
thought to hold for stock prices and foreign exchange rates. The price
of a stock today is the price of a stock yesterday plus an
(unpredicTable) error term. If stock prices do not follow a random
walk, then the change in stock price becomes predicTable and
investors have arbitrage possibilities.

The non-stationary time series variables contain a unit root. These

series contain a stochastic trend. If these time series are differenced,
stationary time series are obtained. For this reason, they are also
called difference stationary.

It is used mostly for forecasting. The current value of a dependent

variable is defined only by its own past values. pth order
Autoregressive model is thus:

(12.10)

Let us subtract yt-1 from both sides of the previous equation. We

obtain following equation after rearranging:

(12.11)

ρ = 0 implies that the AR(p) time series y contains a unit root; if -2 < ρ
< 0, then the series is stationary. In Eq. 12.11, ρ = 0 clarifies the logic
of unit root series. If ρ = 0 then the term yt-1 will drop out of the
equation and only terms involving Δy or its lags will be present in the
regression. “If a unit root is present, then the data can be differenced
to obtain stationarity”.

There is also another regression model which yields trend behavior.

The term δt in the Eq.(12.10) is called a deterministic trend since it is
an exact (deterministic) function of time. In contrast, unit root series
contain a so-called stochastic trend.
210 Applied Econometrics with Eviews Applications

(12.12)

These series can exhibit trend behavior through the incorporation of

a deterministic trend and they are called as trend stationary.
Stationary models with a deterministic trend can yield time series
plots that closely resemble those from non-stationary models having
a stochastic trend. Looking at time series plots alone is not enough to
tell whether a series has a unit root or not.

Example 12.7

Table 12.7 Eviews Output for Wald Coefficient Restriction Test

Wald Test:
Equation: Untitled

Test Statistic Value df Probability

F-statistic 1.464570 (1, 181) 0.2278

Chi-square 1.464570 1 0.2262

Null Hypothesis Summary:

Normalized Restriction (= 0) Value Std. Err.

-1 + C(2) -0.015152 0.012520
Restrictions are linear in coefficients.

Wald coefficient restriction test reported in Table 12.7 indicate that

coefficient of first lagged variable is not different than 1 meaning that
our model follows the random walk.

Seasonality

Seasonal dummies can also be used as explanatory variables. These

dummy variables may be included as additional explanatory variables
in the AR (p) with deterministic trend model.

For instance, with quarterly data, you can create the dummy
variables:
(1) D1 = 1 if an observation is from the first quarter (= 0 otherwise);
 Tıme Serıes 211

(2) D2 = 1 if an observation is from the second quarter (= 0

otherwise); and
(3) D3 = 1 if an observation is from the third quarter (= 0 otherwise).

If the AR(p) with deterministic trend model includes seasonal

dummies, OLS still provides good estimates of all coefficients and
related statistical tests can be used.

Example 12.831

Table 12.8 Eviews Output for Seasonality Analysis

Dependent Variable: DCROBEX
Method: Least Squares
Sample (adjusted): 1997M11 2012M12
Included observations: 182 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

C -0.010373 14.24460 -0.000728 0.9994

DCROBEX(-1) 0.138657 0.074865 1.852084 0.0657
D01 28.98210 48.49478 0.597633 0.5508

R-squared 0.019372 Mean dependent var 2.966044

Adjusted R-squared 0.008415 S.D. dependent var 183.9647
S.E. of regression 183.1890 Akaike info criterion 13.27526
Sum squared resid 6006918. Schwarz criterion 13.32807
Log likelihood -1205.049 Hannan-Quinn criter. 13.29667
F-statistic 1.768029 Durbin-Watson stat 2.031438
Prob(F-statistic) 0.173636

The coefficient of the dummy variable defined for possible

seasonality in the model is not significant as reported in Table 12.8. It
means that the model does not include end of the year effect.

Steps for Testing in the AR(p) with Deterministic Tren

Step 1. Choose the maximum lag length, (pmax) that seems

reasonable to you;
Step 2. Estimate using OLS the AR(pmax) with deterministic
trend model as follows:
31
End of the year effect on Crobex index
212 Applied Econometrics with Eviews Applications

(12.13)

If the p-value for testing is less than the

chosen significance level (e.g. 0.10, 0.05, 0.001) then go
to Step 5, using pmax as lag length. Otherwise go on to
the next step.
Step 3. Estimate the following AR(pmax - 1) model:

(12.14)

If the p-value for testing is less than the

predetermined significance level (e.g. 0.10, 0.05 or 0.001)
then go to Step 5, using pmax - 1 as lag length. Otherwise
go on to the next step.
Step 4. Repeatedly estimate lower order AR models until you find
an AR(p) model where is statistically significant (or
run out of lags).
Step 5. Now test for whether the deterministic trend should be
omitted. If the p-value for testing δ = 0 is greater than the
chosen significance level then drop the deterministic
trend variable.

Example 12.9

Table 12.9 Eviews Output for Testing in The AR(P) with Deterministic Trend Model
Dependent Variable: CROBEX Method: Least Squares
Sample (adjusted): 1998M06 2012M12 Included observations: 175 after adjustments
Variable Coefficient Std. Error t-Statistic Prob.
C -9496.055 7984.226 -1.189352 0.2360
DATUM 0.013075 0.010929 1.196377 0.2333
CROBEX(-1) 0.958370 0.016081 59.59740 0.0000
DCROBEX(-1) 0.159776 0.075370 2.119878 0.0355
DCROBEX(-2) 0.087136 0.076015 1.146303 0.2533
DCROBEX(-3) 0.118046 0.074944 1.575127 0.1172
DCROBEX(-4) -0.066389 0.074777 -0.887828 0.3759
DCROBEX(-5) 0.183187 0.074444 2.460729 0.0149
DCROBEX(-6) -0.141791 0.075692 -1.873271 0.0628
DCROBEX(-7) 0.146991 0.075739 1.940763 0.0540
DCROBEX(-8) 0.210142 0.076206 2.757566 0.0065
R-squared 0.976308 Mean dependent var 1828.422
Adjusted R-squared 0.974864 S.D. dependent var 1093.873
S.E. of regression 173.4277 Akaike info criterion 13.21019
Sum squared resid 4932654. Schwarz criterion 13.40912
Log likelihood -1144.892 Hannan-Quinn criter. 13.29088
F-statistic 675.8239 Durbin-Watson stat 2.008453
Prob(F-statistic) 0.000000
 Tıme Serıes 213

Result reported in Table 12.9 shows that the model does not include
deterministic trend.

Lagged Dependent Variables

Sometimes a past value of y is used as a predictor as well. A

relationship of this type might be:

(12.15)

For instance, this month's sales are defined by four months of

advertising expense plus last month's sales.

Example 12.10

Table 12.10 Eviews Output for Lagged Dependent Variables in The

Model
Dependent Variable: LCPI
Method: Least Squares
Sample (adjusted): 3 192
Included observations: 190 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

C 0.120554 0.020397 5.910510 0.0000

LCPI(-1) 1.503330 0.063013 23.85761 0.0000
LCPI(-2) -0.506877 0.062599 -8.097168 0.0000
LIMPORT -0.005756 0.002572 -2.238411 0.0264

R-squared 0.999901 Mean dependent var 16.68963

Adjusted R-squared 0.999899 S.D. dependent var 1.250662
S.E. of regression 0.012542 Akaike info criterion -5.898567
Sum squared resid 0.029260 Schwarz criterion -5.830208
Log likelihood 564.3638 Hannan-Quinn criter. -5.870876
F-statistic 626343.4 Durbin-Watson stat 1.918657
Prob(F-statistic) 0.000000

LCPI = 0.120553908648 + 1.50333021473*LCPI(-1) - 0.506877156909*LCPI(-2) -

0.00575616664906*LIMPORT
214 Applied Econometrics with Eviews Applications

Moving Average

The MA(1) (first order moving average) process is given by the

following equation:

(12.16)

It means that current value of dependent variable, y1 is a weighted

average of current and previous values of error term, e1 and e0 , and
y2 is a weighted average of e2 and e1 .The value of error term in time
t, et is defined by the white noise process et.

The simple moving average model says that observations that deviate
two or more periods are not correlated.

∑ (12.17)

This can be interpreted as the moving average representation of the

autoregressive process.

Autoregressive process (AR) is written as an infinite order moving

average processes if |θ|<1.

Moving average representation might be more convenient than an

autoregressive in some instances.
 Tıme Serıes 215

Example 12.11
Table 12.10 Eviews Output for Moving Average Method
Dependent Variable: STATEUR
Method: Least Squares
Sample (adjusted): 3 4877
Included observations: 4875 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

C 7.511632 1.09E-05 688602.0 0.0000

RES 0.999995 4.36E-06 229152.5 0.0000
RES(-1) -0.017451 4.36E-06 -3999.289 0.0000

R-squared 1.000000 Mean dependent var 7.511036

Adjusted R-squared 1.000000 S.D. dependent var 2.500353
S.E. of regression 0.000762 Akaike info criterion -11.52156
Sum squared resid 0.002826 Schwarz criterion -11.51757
Log likelihood 28086.81 Hannan-Quinn criter. -11.52016
F-statistic 2.63E+10 Durbin-Watson stat 2.034619
Prob(F-statistic) 0.000000

STATEUR = 7.51163178608 + 0.999994663914*RES - 0.0174508542575*RES(-1)

United States unemployment rate shows moving average

characteristics since the coefficients of the error term and first lag of
the error term are statistically significant.

Autocorrelation Function

Defining autocorrelation ρk as:

{ }
{ }
(12.18)

That autocorrelation is a function of k is referred as autocorrelation

function (ACF) or correlogram of the series yt. The ACF is employed to
model the dependencies between observations by describing the
evolution of yt over time.

The ACF provides following information;

- The value of the process and its correlation with previous
observations.
216 Applied Econometrics with Eviews Applications

- The length and power of the process memory.

- How long and how strongly a shock in the process et impact
the values of yt.

(12.19)

(12.20)

(12.21)

(12.22)

In sum, a shock in an MA(1) process impacts yt in two periods only. A

shock in the AR(1) process impacts all future observations with a
decreasing affect.32

Example 12.12
Table 12.12 Eviews Output for Autocorrelation Function
Sample: 1997M09 2012M12
Included observations: 184

Autocorrelation Partial Correlation AC PAC Q-Stat Prob

.|******* .|******* 1 0.985 0.985 181.37 0.000

.|******* *|. | 2 0.966 -0.135 356.75 0.000
.|******* *|. | 3 0.942 -0.141 524.65 0.000
.|******* *|. | 4 0.915 -0.112 683.84 0.000
.|******| .|. | 5 0.889 0.069 834.91 0.000
.|******| *|. | 6 0.858 -0.178 976.31 0.000
.|******| .|* | 7 0.828 0.097 1109.0 0.000
.|******| *|. | 8 0.796 -0.145 1232.2 0.000
.|***** | **|. | 9 0.757 -0.207 1344.1 0.000
.|***** | .|. | 10 0.717 -0.010 1445.1 0.000
.|***** | .|* | 11 0.679 0.164 1536.3 0.000
.|***** | .|. | 12 0.643 0.023 1618.6 0.000
.|**** | *|. | 13 0.607 -0.084 1692.2 0.000
.|**** | *|. | 14 0.568 -0.116 1757.1 0.000
.|**** | .|* | 15 0.533 0.156 1814.7 0.000
.|**** | *|. | 16 0.497 -0.070 1865.1 0.000

32
Verbeek, M. (2004)
 Tıme Serıes 217

.|*** | .|. | 17 0.462 0.017 1908.8 0.000

.|*** | .|* | 18 0.430 0.113 1946.9 0.000
.|*** | .|. | 19 0.401 0.036 1980.3 0.000
.|*** | .|. | 20 0.378 -0.003 2010.2 0.000
.|*** | .|. | 21 0.354 -0.015 2036.4 0.000
.|** | .|. | 22 0.330 -0.009 2059.4 0.000
.|** | .|. | 23 0.310 0.045 2079.8 0.000
.|** | .|. | 24 0.292 0.053 2098.1 0.000
.|** | *|. | 25 0.274 -0.108 2114.3 0.000
.|** | .|. | 26 0.257 -0.052 2128.6 0.000
.|** | .|. | 27 0.241 -0.012 2141.3 0.000
.|** | .|. | 28 0.226 -0.019 2152.5 0.000
.|* | .|* | 29 0.213 0.099 2162.5 0.000
.|* | .|. | 30 0.198 -0.060 2171.2 0.000
.|* | .|. | 31 0.186 -0.039 2178.9 0.000
.|* | .|. | 32 0.174 -0.012 2185.8 0.000
.|* | .|. | 33 0.162 -0.009 2191.7 0.000
.|* | .|. | 34 0.149 0.002 2196.8 0.000
.|* | .|. | 35 0.135 -0.050 2201.0 0.000
.|* | .|. | 36 0.122 -0.046 2204.4 0.000

Autoregressive Moving Average (ARMA)

MA(q) : (12.23)

AR(p): (12.24)

ARMA(p,q) :
(12.25)

For more parsimonious representation, we may want to work with an

ARMA model that contains both an AR and MA part. The general
ARMA model can be also written as:

(12.26)

(12.27)

(12.28)
218 Applied Econometrics with Eviews Applications

Example 12.13

Table 12.13 Eviews Output for ARMA Model

Dependent Variable: DCROBEX
Method: Least Squares
Sample (adjusted): 1998M03 2012M12
Included observations: 178 after adjustments
Convergence achieved after 96 iterations
MA Backcast: 1997M12 1998M02

Variable Coefficient Std. Error t-Statistic Prob.

C 0.367155 19.34648 0.018978 0.9849

AR(1) -0.911170 0.096376 -9.454333 0.0000
AR(2) -0.868000 0.103492 -8.387151 0.0000
AR(3) -0.421543 0.137626 -3.062956 0.0026
AR(4) 0.218220 0.102726 2.124295 0.0351
AR(5) 0.282092 0.078334 3.601121 0.0004
MA(1) 1.103650 0.074023 14.90948 0.0000
MA(2) 1.185603 0.039084 30.33477 0.0000
MA(3) 0.866090 0.073222 11.82832 0.0000

R-squared 0.174800 Mean dependent var 3.877303

Adjusted R-squared 0.135737 S.D. dependent var 184.6922
S.E. of regression 171.7005 Akaike info criterion 13.17862
Sum squared resid 4982298. Schwarz criterion 13.33950
Log likelihood -1163.897 Hannan-Quinn criter. 13.24386
F-statistic 4.474844 Durbin-Watson stat 2.015767
Prob(F-statistic) 0.000061

Inverted AR Roots .55 -.07-.97i -.07+.97i -.66-.32i

-.66+.32i
Inverted MA Roots -.11-.99i -.11+.99i -.88

ARIMA (p,q)
Stationarity of a stochastic process requires that the variances and
autocovariances are finite and independent of time.33

33
Verbeek, M. (2004)
 Tıme Serıes 219

A series which becomes stationary after first differencing is said to be

integrated of order one, denoted I(1). If Δyt is described by a
stationary ARMA(p,q) model, we say that yt is described by an
autoregressive integrated moving average (ARIMA) model of order p,
1, q or in short an ARIMA (p,1,q) model.

First differencing quite often can transform a non-stationary series

into a stationary one. If a series must be differenced twice before it
becomes stationary, then it is said to be integrated of order two,
denoted I(2) and it must have two unit roots.

Predicting with ARMA Models

One of main goal of building a time series model is to predict the

future path of the economic variables using past observations. the
expected quadratic prediction error is minimized to obtain accurate
prediction. The accuracy of the prediction decreases as time horizon
of the prediction increase.

In the prediction of one period ahead, the MA(1) model provides

more accurate prediction result. In further ahead predictions for the
longer time horizon ARMA type models give better result. Mostly,
the autoregressive representation model is most convenient for the
computation of the predictor.

The informational value or short and long run memory contained in

an AR (1) process decreases over time. The forecast error variance
increases as the forecast horizon increase.

In fact the parameters in ARMA models are unknown. Hence, the

estimated values of the parameters are used which produce
additional uncertainty in predictors. However, this uncertainty is
ignored in practise.

Autoregressive Distributed Lag (ARDL) Model

In time-series econometric modeling a dynamic regression will

usually include both lagged dependent and independent variables as
regressors. The dependent variable might be correlated with its lags.
220 Applied Econometrics with Eviews Applications

It means that lags of the dependent variable should be included in

the regression model. In this model the dependent variable depends
on the lags of itself and the explanatory variables as well as lags of
the explanatory variables as follows:

(12.29)

The model described above is called the autoregressive distributed-

lag model, abbreviated as ARDL(p; k). This model also include
deterministic trend (t). Since the model includes p lags of y and q lags
of x, we can write it as ARDL(p,q). In order to perform this model,
series should have same stationarity properties, either they both are
stationary or both have a unit root.

The values of p and k (lags numbers of y and x used) are chosen:

i- On the basis of the statistical significance of the lagged

variables, and

ii- So that the resulting model is well specified (e.g. it

does not suffer from serial correlation).

The ARDL(1,1), or alternatively the first order Dynamic Linear

Regression Model, takes the following form:

(12.30)

Note that y is sTable (it will converge to its equilibrium)

. If the above stability condition is satisfied, the long-run solution (or
steady state) of Eq. (12.30) is given by:

( ) (12.31)

How can we interpret the coefficients in the ARDL model?

The most common way is through the concept of a multiplier. It is

common to focus on the long run or total multiplier.
 Tıme Serıes 221

Let’s assume that x and y is in an equilibrium or steady state (not

changing over time). When x changes one unit, it is affecting y, which
starts to change, eventually settling down in the long run to a new
equilibrium value. The difference between the old and new
equilibrium values for y can be interpreted as the long run effect of x
on y and is the long run multiplier. This multiplier is often of great
interest for policymakers who want to know the eventual effects of
their policy changes in various areas.

The long run multiplier measures the effect of a permanent change in

x. x changes permanently to a new level one unit higher than the
original value. The long run multiplier measures the effect of this sort
of change.

The long run multiplier does not measure the effect of this type of
change. The “marginal effect” interpretation of regression
coefficients can be used for such temporary changes. Previously, we
were interested in the effect of increasing safety training in one
particular month on accident losses. But we did not discuss the effect
of increasing safety training permanently. The long run multiplier for
the ARDL( p, q) model is:

(12.32)

Here, we are assuming that x and y are stationary. How = 0 in the

AR(p) model implied the existence of a unit root has been studied
above. The ARDL model is not the same as the AR model. If = 0
then the long run multiplier is infinite. In fact, it can be shown that
for the model to be sTable, and then we must have > 0.4. Actually,
if x and y are stationary, this condition is satisfied.34

34
For instance, the effect of computer purchases on sales
222 Applied Econometrics with Eviews Applications

Example 12.14

Table 12.14 Microfit output for ARDL model

A. Autoregressive Distributed Lag Estimates

ARDL(2,0,0,1,0) selected based on Schwarz Bayesian Criterion
*******************************************************************************
Dependent variable is X1
132 observations used for estimation from 2001M2 to 2012M1
*******************************************************************************
Regressor Coefficient Standard Error T-Ratio[Prob]
X1(-1) 1.3772 .071186 19.3462[.000]
X1(-2) -.40579 .071801 -5.6515[.000]
C 11.1558 17.2889 .64526[.520]
X2 -.0063771 .018970 -.33617[.737]
X3 .57778 .12130 4.7634[.000]
X3(-1) -.54644 .082772 -6.6017[.000]
X4 .0021061 .0035208 .59819[.551]
*******************************************************************************
R-Squared .98672 R-Bar-Squared .98608
S.E. of Regression 50.7255 F-stat. F( 6, 125) 1547.6[.000]
Mean of Dependent Variable 590.7834 S.D. of Dependent Variable 429.9262
Residual Sum of Squares 321634.5 Equation Log-likelihood -701.9923
Akaike Info. Criterion -708.9923 Schwarz Bayesian Criterion -719.0821
DW-statistic 2.1428
*******************************************************************************

Diagnostic Tests
*******************************************************************************
* Test Statistics * LM Version * F Version *
*******************************************************************************
* * * *
* A:Serial Correlation*CHSQ( 12)= 27.9926[.006]*F( 12, 113)= 2.5344[.005]*
* * * *
* B:Functional Form *CHSQ( 1)= .039857[.842]*F( 1, 124)= .037452[.847]*
* * * *
* C:Normality *CHSQ( 2)= 12.4107[.002]* Not applicable *
* * * *
* D:Heteroscedasticity*CHSQ( 1)= 42.0608[.000]*F( 1, 130)= 60.7955[.000]*
*******************************************************************************
A:Lagrange multiplier test of residual serial correlation
B:Ramsey's RESET test using the square of the fitted values
C:Based on a test of skewness and kurtosis of residuals
D:Based on the regression of squared residuals on squared fitted values
 Tıme Serıes 223

B. Estimated Long Run Coefficients using the ARDL Approach

ARDL(2,0,0,1,0) selected based on Schwarz Bayesian Criterion
*******************************************************************************
Dependent variable is X1
132 observations used for estimation from 2001M2 to 2012M1
*******************************************************************************
Regressor Coefficient Standard Error T-Ratio[Prob]
C 389.9973 750.0502 .51996[.604]
X2 -.22294 .61821 -.36062[.719]
X3 1.0957 2.9052 .37715[.707]
X4 .073627 .10250 .71830[.474]
*******************************************************************************

C. Error Correction Representation for the Selected ARDL Model

ARDL(2,0,0,1,0) selected based on Schwarz Bayesian Criterion
*******************************************************************************
Dependent variable is dX1
132 observations used for estimation from 2001M2 to 2012M1
*******************************************************************************
Regressor Coefficient Standard Error T-Ratio[Prob]
dX11 .40579 .071801 5.6515[.000]
dC 11.1558 17.2889 .64526[.520]
dX2 -.0063771 .018970 -.33617[.737]
dX3 .57778 .12130 4.7634[.000]
dX4 .0021061 .0035208 .59819[.551]
ecm(-1) -.028605 .016488 -1.7349[.085]
*******************************************************************************
List of additional temporary variables created:
dX1 = X1-X1(-1)
dX11 = X1(-1)-X1(-2)
dC = C-C(-1)
dX2 = X2-X2(-1)
dX3 = X3-X3(-1)
dX4 = X4-X4(-1)
ecm = X1 -389.9973*C + .22294*X2 -1.0957*X3 -.073627*X4
*******************************************************************************
R-Squared .43277 R-Bar-Squared .40554
S.E. of Regression 50.7255 F-stat. F( 5, 126) 19.0738[.000]
Mean of Dependent Variable 1.5935 S.D. of Dependent Variable 65.7909
Residual Sum of Squares 321634.5 Equation Log-likelihood -701.9923
Akaike Info. Criterion -708.9923 Schwarz Bayesian Criterion -719.0821
DW-statistic 2.1428
*******************************************************************************
R-Squared and R-Bar-Squared measures refer to the dependent variable
dX1 and in cases where the error correction model is highly
restricted, these measures could become negative.

If Dependent and Independent Series are Stationary

OLS estimation of ARDL (p,q) regression model can be carried out in

the standard way. Testing of the significance of the parameters can
be done using the t-stats and p-values. However, interpretation of
the results is different from the interpretation of standard case.
224 Applied Econometrics with Eviews Applications

Macroeconomic time series are often highly correlated with their

lags. This implies that original form of ARDL face multicollinearity
problems. With the rewritten form taking first difference, we will not
encounter this problem:

(12.33)

The OLS estimation results interpretation under ceteris paribus

condition can still be used in ARDL. Another interpretation concept
which is called multiplier is commonly used in interpretation of ARDL
regression results focusing on the long run or total multiplier.

Suppose that are in an equilibrium or steady state (not

changing over time). When ,
which starts to change, eventually settling down in the long run to a
new equilibrium value. Then the difference between the old and new
equilibrium values for y can be interpreted as the long run effect of x
on y and is the long run multiplier.

The long run multiplier measures the effect of a permanent change in

x. For temporary changes standard interpretation method of
marginal effect can be used.

The long run multiplier for the ARDL (p,q) model is:

Only the coefficients on are important for the long run

behavior.

For the model to be sTable, then we must have in practice, if

are stationary, this condition will be satisfied.
 Tıme Serıes 225

Example 1535

If dependent and independent series are non-stationary (Spurious

regression)

We assume that have unit roots. In practice, we should test

stationarity using ADF, DF, PP etc. We start with the case of
regression model without lags, then proceed to ARDL (p,q) model.

Suppose that we want to estimate following regression:

(12.34)

If have unit roots then all the usual regression results might
be misleading and incorrect. This is called Spurious regression
problem. You should never run a regression of if the variables
have unit roots with the exception of cointegration method.

COINTEGRATION
Bivariate Cointegration (Engle-Granger Cointegration)

The concept of cointegration was first introduced by Granger (1981)

and elaborated further by Engle and Granger (1987), Engle and Yoo
(1987, 1991), Phillips and Ouliaris (1990), Stock and Watson (1988),
Phillips (1991) and Johansen (1988, 1991, 1994).

(12.35)

Written in this way, it is clear that the errors are a linear combination
of y and x. However, x and y both exhibit non-stationary unit root
behavior such that you would expect the error to also exhibit non-
stationary behavior. The error usually have a unit root. Statistically, it
is this unit root in the error term that causes the spurious regression
problem. However, it is possible that the unit roots in y and x “cancel
each other out” and that the resulting error is stationary. In this

35
KOOP, G. (2005)
226 Applied Econometrics with Eviews Applications

special case, called cointegration, the spurious regression problem

disappears and it is valid to run a regression of y and x.

In sum, if y and x have unit roots, but some linear combination of

them is stationary, then we can say that y and x are cointegrated.

 If y and x have unit roots then they have stochastic trends.

However, if they are cointegrated, the error does not have
such a trend. In this case, the error will not get too large and y
and x will not diverge from one another; y and x, in other
words, will trend together. This fact motivates other jargon
used to refer to cointegrated time series. You may hear them
referred as either having common trends or co-trending.

 If we are talking about an economic model involving an

equilibrium concept, e is the equilibrium error. If y and x are
cointegrated then the equilibrium error remains small.
However, if y and x are not cointegrated then the equilibrium
error will have a trend and departures from equilibrium and
will become increasingly large over time. If such departures
from equilibrium occur, then many would hesitate to say that
the equilibrium is a meaningful one.

 If y and x are cointegrated then there is an equilibrium

relationship between them. If they are not, then no
equilibrium relationship exists.

 Departures from equilibrium should not be too large and

there should always be a tendency to return to equilibrium
after a shock occurs. Hence, if an economic model which
implies an equilibrium relationship exists between y and x is
correct, then we should observe y and x as being
cointegrated.

 If y and x are cointegrated then their trends will cancel each

other out. If cointegration is present, then not only do we
avoid the spurious regression problem, but we also have
important economic information.
 Tıme Serıes 227

If y and x are cointegrated, no need to worry about spurious

regression problem.

Regression of y on x is called cointegrating regression and is

represented as follows:

(12.36)

If the linear combination of two non-stationary series is stationary,

then cointegration exists between these two variables. The unit roots
in y and x cancel each other out and the resulting error becomes
stationary.

Bivariate Cointegration (Engle-Granger) test has following steps:

i- Run the regression of y on an intercept and x and save the

residuals.

ii- Carry out a Dickey-Fuller test (Dickey, Fuller, 1979) on the

residuals (without including a deterministic trend36).

iii- If the unit root hypothesis is rejected then conclude that y

and x are cointegrated. However, if the unit root is
accepted then conclude that cointegration does not occur.

System always returns to equilibrium and, hence, that errors never

grow too big. The null hypothesis in the Engle Granger test is no
cointegration and it is based on Dickey Fuller unit root test. If the
structural breaks occur in the data, Engle Granger test has lower
power and can be misleading.

Engle-Granger cointegration test tells whether cointegration is

present or not. It doesn’t provide any information about how many
cointegrating relationships are exist.

36
If it were included it could mean the errors could be growing steadily over time. This
would violate the idea of cointegration.
228 Applied Econometrics with Eviews Applications

i- The null hypothesis in the Engle–Granger test is “no

cointegration” and we conclude “cointegration is
present” if we reject this hypothesis.

ii- The Engle–Granger test has low power and can be

misleading if there are structural breaks in the data.

Since there is cointegration, no need to worry about the spurious

regressions problem. And, we can proceed to an interpretation of our
coefficients without worrying that the OLS estimates are
meaningless.

Example 12.14

Table 12.14 Eviews Output for Unit Root Tests

A. France
Null Hypothesis: FRANCE has a unit root
Exogenous: Constant, Linear Trend
Lag Length: 0 (Automatic based on SIC, MAXLAG=30)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -1.451148 0.8458

Test critical values: 1% level -3.960237
5% level -3.410881
10% level -3.127243

Augmented Dickey-Fuller Test Equation

Dependent Variable: D(FRANCE)
Method: Least Squares
Included observations: 4175 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

FRANCE(-1) -0.001085 0.000748 -1.451148 0.1468

C 1.007433 1.146994 0.878324 0.3798
TREND(12/30/1988) 0.000810 0.000569 1.422733 0.1549

R-squared 0.000557 Mean dependent var 0.704256

Adjusted R-squared 0.000078 S.D. dependent var 27.85584
S.E. of regression 27.85476 Akaike info criterion 9.492603
 Tıme Serıes 229

Sum squared resid 3237004. Schwarz criterion 9.497157

Log likelihood -19812.81 Hannan-Quinn criter. 9.494214
F-statistic 1.162196 Durbin-Watson stat 1.938805
Prob(F-statistic) 0.312900

Null Hypothesis: D(FRANCE) has a unit root

Exogenous: Constant, Linear Trend
Lag Length: 0 (Automatic based on SIC, MAXLAG=30)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -62.67179 0.0000

Test critical values: 1% level -3.960237
5% level -3.410882
10% level -3.127243

*MacKinnon (1996) one-sided p-values.

Augmented Dickey-Fuller Test Equation

Dependent Variable: D(FRANCE,2)
Method: Least Squares
Sample (adjusted): 1/01/1991 12/29/2006
Included observations: 4174 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

D(FRANCE(-1)) -0.969971 0.015477 -62.67179 0.0000

C 0.260374 1.028061 0.253267 0.8001
TREND(12/30/1988) 0.000162 0.000358 0.453170 0.6504

R-squared 0.484982 Mean dependent var 0.001782

Adjusted R-squared 0.484735 S.D. dependent var 38.80164
S.E. of regression 27.85256 Akaike info criterion 9.492446
Sum squared resid 3235717. Schwarz criterion 9.497000
Log likelihood -19807.73 Hannan-Quinn criter. 9.494056
F-statistic 1963.877 Durbin-Watson stat 1.998964
Prob(F-statistic) 0.000000

We want to explore long run relationship between France and

Greece stock market indices. After checking the unit root condition of
the series, we proceed to find out whether the linear combinations of
this two series are stationary or not. Result reported in Table 12.14
230 Applied Econometrics with Eviews Applications

shows that there exists long run relationship between France and
Greece stock market indices. Furthermore, the long run multiplier is
1.299. This indicates that, in the long run, an increase in the France
stock market index by one unit would cause an increase in the
Greece stock market index 1.299 units.

B. GREECE
Null Hypothesis: GREECE has a unit root
Exogenous: Constant, Linear Trend
Lag Length: 1 (Automatic based on SIC, MAXLAG=31)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -1.758819 0.7246

Test critical values: 1% level -3.959993
5% level -3.410762
10% level -3.127172

*MacKinnon (1996) one-sided p-values.

Augmented Dickey-Fuller Test Equation
Dependent Variable: D(GREECE)
Method: Least Squares
Sample (adjusted): 1/03/1989 12/29/2006
Included observations: 4694 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

GREECE(-1) -0.001150 0.000654 -1.758819 0.0787

D(GREECE(-1)) 0.169347 0.014392 11.76662 0.0000
C 0.785073 1.230216 0.638159 0.5234
TREND(12/30/1988) 0.000958 0.000639 1.499781 0.1337

R-squared 0.029146 Mean dependent var 0.873513

Adjusted R-squared 0.028525 S.D. dependent var 41.93447
S.E. of regression 41.33206 Akaike info criterion 10.28201
Sum squared resid 8012112. Schwarz criterion 10.28751
Log likelihood -24127.87 Hannan-Quinn criter. 10.28394
F-statistic 46.93212 Durbin-Watson stat 1.988464
Prob(F-statistic) 0.000000
 Tıme Serıes 231

Null Hypothesis: D(GREECE) has a unit root

Exogenous: Constant, Linear Trend
Lag Length: 0 (Automatic based on SIC, MAXLAG=31)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -57.76232 0.0000

Test critical values: 1% level -3.959993
5% level -3.410762
10% level -3.127172

*MacKinnon (1996) one-sided p-values.

Augmented Dickey-Fuller Test Equation

Dependent Variable: D(GREECE,2)
Method: Least Squares
Date: 06/05/13 Time: 18:17
Sample (adjusted): 1/03/1989 12/29/2006
Included observations: 4694 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

D(GREECE(-1)) -0.831266 0.014391 -57.76232 0.0000

C 0.368006 1.207416 0.304789 0.7605
TREND(12/30/1988) 0.000153 0.000445 0.342473 0.7320

R-squared 0.415633 Mean dependent var 0.000300

Adjusted R-squared 0.415384 S.D. dependent var 54.06902
S.E. of regression 41.34128 Akaike info criterion 10.28224
Sum squared resid 8017396. Schwarz criterion 10.28636
Log likelihood -24129.42 Hannan-Quinn criter. 10.28369
F-statistic 1668.243 Durbin-Watson stat 1.988270
Prob(F-statistic) 0.000000

Exogenous: Constant
Lag Length: 0 (Automatic based on SIC, MAXLAG=30)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -2.993250 0.0356

Test critical values: 1% level -3.431732
5% level -2.862036
10% level -2.567077
232 Applied Econometrics with Eviews Applications

*MacKinnon (1996) one-sided p-values.

C. Unit Root test for the residual series

Augmented Dickey-Fuller Test Equation

Dependent Variable: D(RESID08)
Method: Least Squares
Date: 06/05/13 Time: 18:10
Included observations: 4175 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

RESID08(-1) -0.004221 0.001410 -2.993250 0.0028

C -0.093792 0.756853 -0.123924 0.9014

R-squared 0.002142 Mean dependent var -0.093794

Adjusted R-squared 0.001903 S.D. dependent var 48.95007
S.E. of regression 48.90346 Akaike info criterion 10.61805
Sum squared resid 9979933. Schwarz criterion 10.62109
Log likelihood -22163.18 Hannan-Quinn criter. 10.61913
F-statistic 8.959543 Durbin-Watson stat 1.915883
Prob(F-statistic) 0.002776

Dependent Variable: GREECE

Method: Least Squares
Date: 06/05/13 Time: 18:19
Sample (adjusted): 12/28/1990 12/29/2006
Included observations: 4176 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

C -724.2877 21.79627 -33.22990 0.0000

FRANCE 1.299339 0.009047 143.6155 0.0000

R-squared 0.831690 Mean dependent var 2169.725

Adjusted R-squared 0.831649 S.D. dependent var 1308.362
S.E. of regression 536.8283 Akaike info criterion 15.40971
Sum squared resid 1.20E+09 Schwarz criterion 15.41275
Log likelihood -32173.48 Hannan-Quinn criter. 15.41079
F-statistic 20625.41 Durbin-Watson stat 0.008315
Prob(F-statistic) 0.000000

GREECE = -724.287661619 + 1.29933920604*FRANCE

 Tıme Serıes 233

The residual of the model is found stationary at the 5% significance

level meaning that the model is not spurious and we can accept the
model. Greece and France do have long run relationship. The
variables in the model are cointegrated or have long run equilibrium
relationship.

Error Correction Model (ECM)

If the research aims to explore short run behavior, it is not possible to

use the regression of y on x. In these cases an error correction model
(or ECM for short) should be employed.

The Granger Representation Theorem says that if y and x are

cointegrated, then the relationship between them can be expressed
as an ECM. In this section, we assume y and x are cointegrated.

In order to understand the properties of ECMs let us begin with the

following simple version:

(12.37)

where et-1 is the error obtained from the regression model with y and
x,

(12.38)

and et is the error in the ECM model.

We assume that λ < 0

Note that, if we know et-1, then the ECM is just a regression model.
That is, Δyt is the dependent variable and et-1 and Δxt are explanatory
variables.

The regression model attempts to use explanatory variables to

explain the dependent variable. The ECM says that Δyt depends on
Δxt meaning that changes in x cause y to change. In addition, Δyt
depends on et-1. This latter aspect is unique to the ECM and gives it
its name.
234 Applied Econometrics with Eviews Applications

In sum, the ECM has both long run and short run properties, such as:

- The former properties are embedded in the et-1 term (which is

still the long run multiplier and the errors are from the
regression involving y and x).

- The short run behavior is partially captured by the equilibrium

error term, which says that, if y is out of equilibrium, it is
pulled towards it in the next period.

- Further aspects of short run behavior are captured by the

inclusion of Δxt as an explanatory variable. This term implies
that, if x changes, the equilibrium value of y also change and
that y will also change accordingly.

In ECM the spurious regression problem does not exist. OLS

estimation and t-statistics and p-values can be used for
interpretation, because of:

- y and x both have unit roots; therefore Δy and Δx are

stationary.

- Since y and x are cointegrated, the equilibrium error is

stationary.

- The dependent variable and all explanatory variables in the

ECM are stationary.

The inclusion of et-1 as an explanatory variable brings some new

statistical issues:

- The errors in the model are not directly observed. How they
can be used as an explanatory variable in a regression?

- To replace the unknown errors by the residuals from the

regression of y on x (replace et-1 by ut-1). A simple technique
based on two OLS regressions proceeds as follows:
Step 1. Run a regression of y on x and save the
residuals.
 Tıme Serıes 235

Step 2. Run a regression of Δy on Δ and the residuals

from Step 1 lagged one period.

Note that y and x should have unit roots and be cointegrated before
proceeding to ECM. So far we have discussed the simplest error
correction model. The ECM may also have lags and deterministic
trend like ARDL (p, q) model. Incorporating these features into the
ECM yields:

(12.39)

This expression is still in the form of a regression model and can be

estimated using the two-step procedure described above. The
adjustment to equilibrium intuition also holds for this model.

As the variables Greece and France stock market indices are

cointegrated in the Example 12.14, we can proceed to run the
following ECM.

(12.40)

D(Greece) and D(France) are the first differenced variables;

is intercept;

is short run coefficient. It tells us at what rate it corrects the

previous period disequilibrium of the system. It validates that there
exist a long run equilibrium relationship among the variables.

v is error term

is the one period lag residual of the model. It is also known as an

equilibrium error term. It guides the variables of the system to
restore back to equilibrium. Sign of the error correction term should
be negative and it should be significant.

After estimating the model reported in the Table 12.15, short run
coefficient has been found 0.41 and the coefficient of error term
236 Applied Econometrics with Eviews Applications

has been found 0.15 percent meaning that system corrects its
previous period disequilibrium at a speed of 0.15 percent monthly.
The sign of is found negative and significant indicating the non-
existence of long run equilibrium relationship between Greece and
France stock market indices.

Example 12.15

Table 12.15 Eviews Result for ECM Estimates

Dependent Variable: D(GREECE)
Method: Least Squares
Sample (adjusted): 12/31/1990 12/29/2006
Included observations: 4175 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

C 0.530286 0.654009 0.810824 0.4175

D(FRANCE) 0.413186 0.023514 17.57155 0.0000
V(-1) -0.001512 0.001220 -1.238985 0.2154

R-squared 0.068917 Mean dependent var 0.821274

Adjusted R-squared 0.068470 S.D. dependent var 43.76977
S.E. of regression 42.24474 Akaike info criterion 10.32556
Sum squared resid 7445427. Schwarz criterion 10.33011
Log likelihood -21551.60 Hannan-Quinn criter. 10.32717
F-statistic 154.4006 Durbin-Watson stat 1.731312
Prob(F-statistic) 0.000000
 Tıme Serıes 237

Diagnostic Checking

At first, serial correlation is checked as follows;

Table 12.16 Eviews Output for Serial Correlation Test

Breusch-Godfrey Serial Correlation LM Test:

F-statistic 40.63616 Prob. F(2,4170) 0.0000

Obs*R-squared 79.81422 Prob. Chi-Square(2) 0.0000

Test Equation:
Dependent Variable: RESID
Method: Least Squares
Date: 06/05/13 Time: 18:58
Sample: 12/31/1990 12/29/2006
Included observations: 4175
Presample missing value lagged residuals set to zero.

Variable Coefficient Std. Error t-Statistic Prob.

C 0.000996 0.647883 0.001537 0.9988

D(FRANCE) -0.001451 0.023295 -0.062292 0.9503
V(-1) -0.000639 0.001216 -0.525342 0.5994
RESID(-1) 0.139198 0.015513 8.972856 0.0000
RESID(-2) -0.031297 0.015526 -2.015786 0.0439

R-squared 0.019117 Mean dependent var -6.18E-17

Adjusted R-squared 0.018176 S.D. dependent var 42.23462
S.E. of regression 41.84902 Akaike info criterion 10.30721
Sum squared resid 7303092. Schwarz criterion 10.31480
Log likelihood -21511.30 Hannan-Quinn criter. 10.30990
F-statistic 20.31808 Durbin-Watson stat 1.999187
Prob(F-statistic) 0.000000

Since it is significant we can reject the null hypothesis of there is no

serial correlation, and conclude that our model is serially correlated.

In the second step normality assumption is checked as follows;

Check whether the residuals are normally distributed.

238 Applied Econometrics with Eviews Applications

2,000
Series: Residuals
Sample 12/31/1990 12/29/2006
1,600 Observations 4175

Mean -6.18e-17
1,200 Median -0.506729
Maximum 350.3563
Minimum -413.0776
800 Std. Dev. 42.23462
Skewness -0.111649
Kurtosis 19.47798
400
Jarque-Bera 47242.53
Probability 0.000000
0
-400 -300 -200 -100 0 100 200 300

Figure 12.1 Eviews Output for Normality

JB statistics is significant. We reject the null hypothesis of residuals

are normally distributed. Residuals of the model are not normally
distributed.

Dependent and Independent variables have unit roots but are not
cointegrated

Even though the time series have unit root, the Engle–Granger test
may indicate no cointegration. In such cases, the series may not be
trending together and may not have an equilibrium relationship.
Hence, a regression of y on x should not be run due to the spurious
regression problem. The presence of such characteristics suggests
that basic model should be respecified and other explanatory
variables included. Instead of working with y and x themselves,
difference series can be used.

In this case, the following ARDL model can be employed but with
changes in the variables:

(12.41)

For most time series variables, this specification should not suffer
from multicollinearity problems. Or, second variant of the ARDL
model based on the differenced data can be estimated.
 Tıme Serıes 239

But if you are working with the differences of your time series and
then use the variant of the ARDL that involves differencing the data
you end up with second differenced data:

(12.42)

Where:

(12.43)

OLS estimation and testing can be done in either of these models in a

straightforward way. Whatever model is chosen, it is important to
emphasize that the interpretation of regression results will likewise
change.

More specifically, suppose that:

y: exchange rates and x : interest rates

If y and x are cointegrated, or if both are stationary, we can obtain an

estimate of the long run effect of a small change in interest rates on
exchange rates. If y and x are neither stationary nor cointegrated and
we estimate either of the two preceding equations, we can obtain an
estimate of the long run effect of a small change in the change of
interest rates on the change in exchange rates.

This may or may not be a sensible thing to measure depending on the

particular empirical exercise, such as:

 If all variables are stationary, then an ARDL(p,q) model can be

estimated using OLS. Statistical techniques are all standard
and valid.
 A variant on the ARDL model is often used to avoid potential
multicollinearity problems and provide a straightforward
estimate of the long run multiplier.
 If all variables are nonstationary, the spurious regression
problem can occur.
240 Applied Econometrics with Eviews Applications

 If cointegration is present, the spurious regression problem

does not occur.
 If all variables are nonstationary but the regression error is
stationary, then cointegration occurs.
 Cointegration implies whether an equilibrium relationship
exists.
 Cointegration can be tested using the Engle–Granger test.
This test is a Dickey–Fuller test on the residuals from the
cointegrating regression.
 If the variables are cointegrated, then an error correction
model can be employed. This model captures short run
behavior in a way that the cointegrating regression cannot.
 If the variables have unit roots but are not cointegrated, take
difference and estimate an ARDL model using the differenced
variables. The interpretation of these models can be
awkward.

Multivariate Cointegration

The Johansen test (Johansen, 1988, 1991) is a test for cointegration

that allows more than one cointegrating relationship, unlike the
Engle–Granger method.

Example 12.16
Table 12.16 Eviews Output for Johansen Cointegration Test
Sample (adjusted): 1/06/1989 12/29/2006
Included observations: 4691 after adjustments
Trend assumption: Linear deterministic trend
Series: GERMANY SPAIN ITALY
Lags interval (in first differences): 1 to 4
Unrestricted Cointegration Rank Test (Trace)
Hypothesized Trace 0.05
No. of CE(s) Eigenvalue Statistic Critical Value Prob.**
None * 0.003696 31.25751 29.79707 0.0337
At most 1 0.002596 13.88696 15.49471 0.0861
At most 2 0.000361 1.692354 3.841466 0.1933
Trace test indicates 1 cointegrating eqn(s) at the 0.05 level
* denotes rejection of the hypothesis at the 0.05 level
**MacKinnon-Haug-Michelis (1999) p-values
Unrestricted Cointegration Rank Test (Maximum Eigenvalue)
Hypothesized Max-Eigen 0.05
No. of CE(s) Eigenvalue Statistic Critical Value Prob.**
None 0.003696 17.37056 21.13162 0.1552
 Tıme Serıes 241

At most 1 0.002596 12.19460 14.26460 0.1035

At most 2 0.000361 1.692354 3.841466 0.1933
Max-eigenvalue test indicates no cointegration at the 0.05 level
* denotes rejection of the hypothesis at the 0.05 level
**MacKinnon-Haug-Michelis (1999) p-values
Unrestricted Cointegrating Coefficients (normalized by b'*S11*b=I):
GERMANY SPAIN ITALY
-0.002378 -0.008764 0.004923
0.002253 -0.003252 -0.001580
0.000603 0.000243 0.000169
Unrestricted Adjustment Coefficients (alpha):
D(GERMANY) -1.096077 -0.998472 -0.635771
D(SPAIN) -0.318655 -0.286985 -0.020752
D(ITALY) -1.959926 -0.175790 -0.230935
1 Cointegrating Equation(s): Log likelihood -59785.64
Normalized cointegrating coefficients (standard error in parentheses)
GERMANY SPAIN ITALY
1.000000 3.684964 -2.070029
(0.84823) (0.24787)
Adjustment coefficients (standard error in parentheses)
D(GERMANY) 0.002607
(0.00149)
D(SPAIN) 0.000758
(0.00027)
D(ITALY) 0.004661
(0.00120)
2 Cointegrating Equation(s): Log likelihood -59779.54
Normalized cointegrating coefficients (standard error in parentheses)
GERMANY SPAIN ITALY
1.000000 0.000000 -1.086768
(0.08357)
0.000000 1.000000 -0.266830
(0.02765)
Adjustment coefficients (standard error in parentheses)
D(GERMANY) 0.000358 0.012853
(0.00205) (0.00584)
D(SPAIN) 0.000111 0.003726
(0.00037) (0.00106)
D(ITALY) 0.004265 0.017748
(0.00166) (0.00473)

Johansen cointegration test result reported in Table 12.16 indicate

that there is one cointegrating relationships between variables
meaning that there is significant long run relationship between
variables.
242 Applied Econometrics with Eviews Applications

Vector Error Correction Model

Vector Error Correction Model is used to explore dynamic (short run)

cointegration relationship among more then three variables.

Example 12.17

Table 12.17 Eviews Output for VECM Test

Vector Error Correction Estimates
Sample (adjusted): 1/04/1989 12/29/2006
Included observations: 4693 after adjustments
Standard errors in ( ) & t-statistics in [ ]

Cointegrating Eq: CointEq1

GERMANY(-1) 1.000000

SPAIN(-1) 3.364462
(0.75675)
[ 4.44593]

ITALY(-1) -1.974751
(0.22148)
[-8.91605]

C -139.3244

Error Correction: D(GERMNY) D(SPAIN) D(ITALY)

CointEq1 0.002324 0.000753 0.005015

(0.00158) (0.00029) (0.00128)
[ 1.47398] [ 2.62824] [ 3.93176]

D(GERMANY(-1)) -0.104841 0.007890 0.026180

(0.02289) (0.00416) (0.01851)
[-4.58118] [ 1.89830] [ 1.41404]

D(GERMANY(-2)) -0.042952 -0.010781 -0.016748

(0.02285) (0.00415) (0.01849)
[-1.87952] [-2.59759] [-0.90585]

D(SPAIN(-1)) 0.414903 0.007206 0.015514

(0.13343) (0.02423) (0.10795)
 Tıme Serıes 243

[ 3.10955] [ 0.29736] [ 0.14372]

D(SPAIN(-2)) 0.039475 -0.013438 -0.207321

(0.13332) (0.02421) (0.10786)
[ 0.29608] [-0.55501] [-1.92210]

D(ITALY(-1)) 0.075161 -0.002533 -0.037561

(0.02933) (0.00533) (0.02373)
[ 2.56277] [-0.47560] [-1.58306]

D(ITALY(-2)) 0.023652 0.009558 0.070995

(0.02933) (0.00533) (0.02373)
[ 0.80634] [ 1.79421] [ 2.99176]

C 0.575387 0.271770 0.650120

(0.62632) (0.11374) (0.50670)
[ 0.91868] [ 2.38932] [ 1.28305]

R-squared 0.007436 0.005513 0.006062

Adj. R-squared 0.005953 0.004027 0.004577
Sum sq. resids 8598952. 283602.3 5628036.
S.E. equation 42.84183 7.780367 34.65961
F-statistic 5.014384 3.709923 4.081816
Log likelihood -24289.09 -16283.25 -23294.46
Akaike AIC 10.35461 6.942787 9.930730
Schwarz SC 10.36561 6.953789 9.941732
Mean dependent 0.662387 0.272514 0.624831
S.D. dependent 42.96993 7.796079 34.73920

Determinant resid covariance (dofadj.) 23839045

Determinant resid covariance 23717340
Log likelihood -59824.83
Akaike information criterion 25.50685
Schwarz criterion 25.54399

D(GERMANY) = C(1)*( GERMANY(-1) + 3.36446151586*SPAIN(-1) -

1.97475095482*ITALY(-1) - 139.324408384 ) + C(2)*D (GERMANY(-1)) +
C(3)*D(GERMANY(-2)) + C(4)*D(SPAIN(-1)) + C(5)*D(SPAIN(-2)) + C(6)*D(ITALY(-1))
+ C(7)*D(ITALY(-2)) + C(8)

Hence, estimate the above model by OLS.

244 Applied Econometrics with Eviews Applications

Table 12.18 Eviews Output for Johansen Cointegration Test

dependent Variable: D(GERMANY)
Method: Least Squares
Date: 06/05/13 Time: 19:11
Sample (adjusted): 1/04/1989 12/29/2006
Included observations: 4693 after adjustments
D(GERMANY) = C(1)*( GERMANY(-1) + 3.36446151586*SPAIN(-1) -
1.97475095482*ITALY(-1) - 139.324408384 ) + C(2)*D(GERMANY(-1))
+ C(3)*D(GERMANY(-2)) + C(4)*D(SPAIN(-1)) + C(5)*D(SPAIN(-2)) +
C(6)*D(ITALY(-1)) + C(7)*D(ITALY(-2)) + C(8)

Coefficient Std. Error t-Statistic Prob.

C(1) 0.002324 0.001577 1.473983 0.1406

C(2) -0.104841 0.022885 -4.581177 0.0000
C(3) -0.042952 0.022853 -1.879518 0.0602
C(4) 0.414903 0.133429 3.109552 0.0019
C(5) 0.039475 0.133325 0.296080 0.7672
C(6) 0.075161 0.029328 2.562772 0.0104
C(7) 0.023652 0.029332 0.806340 0.4201
C(8) 0.575387 0.626318 0.918681 0.3583

R-squared 0.007436 Mean dependent var 0.662387

Adjusted R-squared 0.005953 S.D. dependent var 42.96993
S.E. of regression 42.84183 Akaike info criterion 10.35461
Sum squared resid 8598952. Schwarz criterion 10.36561
Log likelihood -24289.09 Hannan-Quinn criter. 10.35848
F-statistic 5.014384 Durbin-Watson stat 2.001250
Prob(F-statistic) 0.000011

C(1) is the coefficient of VECM. It should be negative and significant.

In our model it is positive and not significant. It is unable to validate
Johansen cointegration test results reported in the Table 12.18. The
other coefficients are short run coefficients.

Diagnostic Checking

At first, whether there is serial correlation is checked as follows;

 Tıme Serıes 245

Table 12.19 Eviews Output for Serial Correlation

Breusch-Godfrey Serial Correlation LM Test:

F-statistic 1.150315 Prob. F(2,4683) 0.3166

Obs*R-squared 2.304411 Prob. Chi-Square(2) 0.3159

Test Equation:
Dependent Variable: RESID
Method: Least Squares
Sample: 1/04/1989 12/29/2006
Presample missing value lagged residuals set to zero.

Variable Coefficient Std. Error t-Statistic Prob.

C(1) -0.001607 0.001902 -0.844888 0.3982

C(2) 0.845628 0.576122 1.467794 0.1422
C(3) -0.012230 0.180551 -0.067738 0.9460
C(4) -0.002758 0.133439 -0.020668 0.9835
C(5) -0.336260 0.266225 -1.263068 0.2066
C(6) -0.002110 0.029361 -0.071868 0.9427
C(7) -0.062456 0.052016 -1.200713 0.2299
C(8) -0.422910 0.687281 -0.615338 0.5384
RESID(-1) -0.844129 0.574666 -1.468905 0.1419
RESID(-2) 0.097797 0.206616 0.473326 0.6360

R-squared 0.000491 Mean dependent var 1.22E-15

Adjusted R-squared -0.001430 S.D. dependent var 42.80986
S.E. of regression 42.84045 Akaike info criterion 10.35497
Sum squared resid 8594730. Schwarz criterion 10.36872
Log likelihood -24287.94 Hannan-Quinn criter. 10.35981
F-statistic 0.255626 Durbin-Watson stat 1.999579
Prob(F-statistic) 0.985748
246 Applied Econometrics with Eviews Applications

1,600
Series: Residuals
1,400 Sample 1/04/1989 12/29/2006
Observations 4693
1,200
Mean 1.22e-15
1,000 Median 0.532233
Maximum 240.5610
800 Minimum -301.2473
Std. Dev. 42.80986
600
Skewness -0.298449
Kurtosis 8.059792
400

Jarque-Bera 5075.829
200
Probability 0.000000
0
-300 -200 -100 0 100 200

Figure 12.2 Eviews Output for Normality Test

Table 12.20 Eviews Output for Heteroscedasticity

Heteroskedasticity Test: Breusch-Pagan-Godfrey
F-statistic 84.61243 Prob. F(9,4683) 0.0000
Obs*R-squared 656.3996 Prob. Chi-Square(9) 0.0000
Scaled explained SS 2309.130 Prob. Chi-Square(9) 0.0000
Test Equation:
Dependent Variable: RESID^2
Method: Least Squares
Sample: 1/04/1989 12/29/2006
Included observations: 4693
Variable Coefficient Std. Error t-Statistic Prob.
C -1652.975 163.4808 -10.11113 0.0000
GERMANY(-1) -8.504171 2.413443 -3.523667 0.0004
SPAIN(-1) -21.73891 14.07776 -1.544203 0.1226
ITALY(-1) 1.605723 3.094562 0.518885 0.6039
GERMANY(-2) 0.924901 3.167677 0.291981 0.7703
GERMANY(-3) 8.457493 2.413339 3.504478 0.0005
SPAIN(-2) 21.95260 19.60785 1.119582 0.2629
SPAIN(-3) -6.050174 14.09019 -0.429389 0.6677
ITALY(-2) -1.505254 4.240540 -0.354968 0.7226
ITALY(-3) 1.872900 3.094881 0.605161 0.5451
R-squared 0.139868 Mean dependent var 1832.293
Adjusted R-squared 0.138215 S.D. dependent var 4868.971
S.E. of regression 4519.983 Akaike info criterion 19.67253
Sum squared resid 9.57E+10 Schwarz criterion 19.68628
Log likelihood -46151.60 Hannan-Quinn criter. 19.67737
F-statistic 84.61243 Durbin-Watson stat 1.848660
Prob(F-statistic) 0.000000
 Tıme Serıes 247

Potential Pitfalls of the Cointegration Method

i- If several lags of the same variable are used, it can

cause multicollinearity; if xt is highly correlated with its
own past values, then autocorrelation occurs.

ii- One variable can influence another with a time lag.

iii- If the data are non-stationary, a problem known as

spurious regression may arise.

iv- Interpretation depends on whether, x and y, are

stationary or not.

DIAGNOSTIC CHECKING
The model adequacy should be investigated in the last step of the
model building cycle. Following residual analysis may be performed;

- If ARMA(p,q) model is chosen on the basis of the sample ACF

and PACF, an ARMA(p+1,q) and an ARMA(p,q+1) should be
estimated in order to test the significance of the additional
parameters.

- The residuals of an well specified model should be white

noise.

- A plot of the residuals can be used for checking outliers.

- The significance of residual autocorrelation may be examined.

The autocorrelations are zero for a white noise series. The
Ljung-Box portmanteau test statistics is used in order to test
the residual autocorrelation by comparing with approximate
two standard error bounds:

∑ (12.44)
248 Applied Econometrics with Eviews Applications

rk is the estimated autocorrelation coefficients of the

residuals
k is a number chosen by the researcher.

The structure of the model should be revised, if the model is rejected

at this stage.
 Volatılıty 249

CHAPTER 13

VOLATILITY

ARCH MODEL
This model is closely related to the Autoregressive Conditional
Heteroscedasticity (ARCH), (Engle, 1982).

Suppose the following random walk model:

, (13.1)

, (13.2)

Naturally, stock prices are quite fragile and influenced by expected

and unexpected factors, reflected in the error ( ). This result for
investors is unable to predict future stock price changes. If the future
stock price movements are predicTable, then there would be
arbitrage opportunities. Arbitrage opportunities give advantage to
the smart investors and they are instantly eliminated.

A more realistic model is:

, (13.3)
250 Applied Econometrics with Eviews Applications

This model implies that stock prices, on average, increase by per

period, but are otherwise unpredicTable. This is known as the
random walk with drift model; it is a form of the random walk
model with intercept. It allows stock prices to “drift” upwards over
time.

If this is the case, the behavior of the series is not possible to predict.
Instead, the volatility of financial time series can be explored. In fact,
volatility changes over time are examined.

Mostly, to know the level of macroeconomic indicators is not

important, but their variance. For instance, if the level of inflation is
high, agents or individuals can plan for the future with a high degree
of confidence if the variance of inflation is low. However, if there is a
high variance it is hard to predict what the inflation rate might be
next period. The low level of variance reflects high level of reliable
prediction is possible which also reflect low level of risk. For example,
if the exchange rate exhibits low volatility it becomes easy to plan
whereas high volatility makes it more difficult. The negative effect of
uncertainty partially accounts for the growing number of financial
derivatives (futures and options) that firms can use to hedge against
the risk posed by currency fluctuations.

The volatility is a measure of the riskiness of a stock. Nonetheless,

the riskiness of a portfolio of stocks depends not only on the volatility
of the individual stocks, but also on the correlation between the
stocks in the portfolio (CAPM).

Let’s assume that either asset price follows a pure random walk
model or a random walk with drift:

̅, (13.4)

Taking deviations from mean implies that there is no intercept in the

model.

is an estimator of volatility at time t. High volatility is associated

with big changes. This measure of volatility will be small in sTable
times and large in times of crises.
 Volatılıty 251

Autoregressive models are normally used to model “clustering in

volatility”, which exists in financial time series data. Consider, for
instance, an AR (1) model that uses volatility as the time series
variable of interest:

, (13.5)

This model has volatility in a period depending on volatility in a

previous period. If, then the volatility was high in last period
(e.g. was very large), it will also tend to be high in this period.
Alternatively, if volatility was low last period (e.g. was near
zero) then this period’s volatility will also tend to be low.

Low volatility is followed by low volatility and high volatility is

followed by high. The level of volatility tends to be similar over time.
But, in general, this model implies that intervals or clusters are
observed in time where volatility is low and intervals where it is high.
In empirical studies of asset prices, such a pattern is very common.37

All the series in the model should be stationary. If not, there should
be converted into stationary before proceed to estimate ARCH and
GARCH (Bollerslev, 1986) model. model.

Example 13.1

Let’s start with the model that includes Greece stock index.
GREECE C
Estimate the equation and check the residual through graph to
identify whether the periods of high volatility is followed by period of
high volatility and vice versa.

37
Koop, G. (2005)
252 Applied Econometrics with Eviews Applications

Table 13.1 Eviews Output for Estimation

Dependent Variable: D(GREECE)
Method: Least Squares
Date: 06/06/13 Time: 19:06
Sample (adjusted): 1/02/1989 12/29/2006
Included observations: 4695 after adjustments
Variable Coefficient Std. Error t-Statistic Prob.
C 0.874198 0.611938 1.428573 0.1532
R-squared 0.000000 Mean dependent var 0.874198
Adjusted R-squared 0.000000 S.D. dependent var 41.93003
S.E. of regression 41.93003 Akaike info criterion 10.31009
Sum squared resid 8252650. Schwarz criterion 10.31147
Log likelihood -24201.95 Hannan-Quinn criter. 10.31058
Durbin-Watson stat 1.662471
Go to ARCH test under heteroscedasticity test and run. Check R2 and
p-values. If p-value is less than 5%, we can reject null hypothesis of
there is no ARCH effect.

After rejection of the null hypothesis we can proceed to ARCH

models.

Table 13.2 Eviews Output for Heteroscedasticity Test

Heteroskedasticity Test: ARCH
F-statistic 365.0401 Prob. F(1,4692) 0.0000
Obs*R-squared 338.8343 Prob. Chi-Square(1) 0.0000
Test Equation:
Dependent Variable: RESID^2
Method: Least Squares
Sample (adjusted): 1/03/1989 12/29/2006
Included observations: 4694 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

C 1285.767 108.3720 11.86438 0.0000

RESID^2(-1) 0.268672 0.014062 19.10602 0.0000

R-squared 0.072185 Mean dependent var 1758.125

Adjusted R-squared 0.071987 S.D. dependent var 7504.223
S.E. of regression 7229.077 Akaike info criterion 20.61004
Sum squared resid 2.45E+11 Schwarz criterion 20.61279
Log likelihood -48369.76 Hannan-Quinn criter. 20.61100
F-statistic 365.0401 Durbin-Watson stat 2.203790
Prob(F-statistic) 0.000000
 Volatılıty 253

Go to estimate and choose ARCH model from the list. Choose 1 for
ARCH, 0 for GARCH, and normal distribution, then run the model
which is GARCH(1,0).
Table 13.3 Eviews Output for ARCH Test
Dependent Variable: D(GREECE)
Method: ML - ARCH (Marquardt) - Normal distribution
Sample (adjusted): 1/02/1989 12/29/2006
Included observations: 4695 after adjustments
Convergence achieved after 20 iterations
Presample variance: backcast (parameter = 0.7)
GARCH = C(2) + C(3)*RESID(-1)^2
Variable Coefficient Std. Error z-Statistic Prob.
C 1.113585 0.278842 3.993613 0.0001
Variance Equation
C 774.5477 6.553909 118.1810 0.0000
RESID(-1)^2 0.890166 0.021965 40.52626 0.0000
R-squared -0.000033 Mean dependent var 0.874198
Adjusted R-squared -0.000459 S.D. dependent var 41.93003
S.E. of regression 41.93965 Akaike info criterion 9.985032
Sum squared resid 8252919. Schwarz criterion 9.989156
Log likelihood -23436.86 Hannan-Quinn criter. 9.986482
Durbin-Watson stat 1.662417

Now estimate GARCH(1,1) model

Table 13.4 Eviews output for GARCH(1,1) test
Dependent Variable: D(GREECE)
Method: ML - ARCH (Marquardt) - Normal distribution
Date: 06/06/13 Time: 19:10
Sample (adjusted): 1/02/1989 12/29/2006
Included observations: 4695 after adjustments
Convergence achieved after 15 iterations
Presample variance: backcast (parameter = 0.7)
GARCH = C(2) + C(3)*RESID(-1)^2 + C(4)*GARCH(-1)
Variable Coefficient Std. Error z-Statistic Prob.
C 0.229580 0.175920 1.305023 0.1919
Variance Equation
C 0.697177 0.099060 7.037933 0.0000
RESID(-1)^2 0.119684 0.005399 22.16970 0.0000
GARCH(-1) 0.896867 0.003844 233.3240 0.0000
R-squared -0.000236 Mean dependent var 0.874198
Adjusted R-squared -0.000876 S.D. dependent var 41.93003
S.E. of regression 41.94839 Akaike info criterion 9.101550
Sum squared resid 8254601. Schwarz criterion 9.107049
Log likelihood -21361.89 Hannan-Quinn criter. 9.103484
Durbin-Watson stat 1.662078
254 Applied Econometrics with Eviews Applications

TARCH model

In the Threshold ARCH (TARCH) model innovations or shocks are

divided into intervals and approximated a piecewise linear function
for the conditional standard deviation.38

Two intervals mean that the division is normally at zero and the
impact of positive and negative shocks on the volatility is
differentiated.

Table 13.5 Eviews Output for TARCH Test

Dependent Variable: D(GREECE)
Method: ML - ARCH (Marquardt) - Normal distribution
Sample (adjusted): 1/02/1989 12/29/2006
Included observations: 4695 after adjustments
Convergence achieved after 37 iterations
Presample variance: backcast (parameter = 0.7)
GARCH = C(2) + C(3)*RESID(-1)^2 + C(4)*RESID(-1)^2*(RESID(-1)<0) +
C(5)*GARCH(-1)

Variable Coefficient Std. Error z-Statistic Prob.

C 0.266773 0.188900 1.412241 0.1579

Variance Equation

C 0.617096 0.091341 6.755947 0.0000

RESID(-1)^2 0.122972 0.006082 20.21843 0.0000
RESID(-1)^2*(RESID(-
1)<0) -0.016922 0.006908 -2.449444 0.0143
GARCH(-1) 0.900832 0.003799 237.1060 0.0000

R-squared -0.000210 Mean dependent var 0.874198

Adjusted R-squared -0.001063 S.D. dependent var 41.93003
S.E. of regression 41.95231 Akaike info criterion 9.101390
Sum squared resid 8254382. Schwarz criterion 9.108264
Log likelihood -21360.51 Hannan-Quinn criter. 9.103807
Durbin-Watson stat 1.662122

38
See Glosten et al. (1993) for more information.
 Volatılıty 255

EGARCH model

Standard GARCH models are based on the same (symmetric) effect of

negative and positive shoks on the volatility which is mostly violated
in practice. Normaly, stock returns are more responsive to the
negative shoks than positive shocks. This is called Leverage Effect
(Black, 1976).

The exponential GARCH (EGARCH) model is a GARCH family model

provided by Nelson (1990). This model includes modelling logarithm
and leverage term to detect asymmetry in volatility clustering as
follows;

| |
(13.6)

The model is asymmetric as long as If then positive

shocks generate less volatility then negative shocks.39

Table 13.6 Eviews Output for EGARCH Test

Dependent Variable: D(GREECE)
Method: ML - ARCH (Marquardt) - Normal distribution
Sample (adjusted): 1/02/1989 12/29/2006
Included observations: 4695 after adjustments
Convergence achieved after 18 iterations
Presample variance: backcast (parameter = 0.7)
LOG(GARCH) = C(2) + C(3)*ABS(RESID(-1)/@SQRT(GARCH(-1))) + C(4)
*LOG(GARCH(-1))
Variable Coefficient Std. Error z-Statistic Prob.
C 0.222382 0.115135 1.931490 0.0534
Variance Equation
C(2) -0.098750 0.005062 -19.50920 0.0000
C(3) 0.208688 0.007636 27.32960 0.0000
C(4) 0.992058 0.000647 1532.764 0.0000
R-squared -0.000242 Mean dependent var 0.874198
Adjusted R-squared -0.000881 S.D. dependent var 41.93003
S.E. of regression 41.94850 Akaike info criterion 9.103164
Sum squared resid 8254645. Schwarz criterion 9.108663
Log likelihood -21365.68 Hannan-Quinn criter. 9.105098
Durbin-Watson stat 1.662069

39
See Verbeek (A guide to modern econometrics, 2004) for more information.
256 Applied Econometrics with Eviews Applications

We have estimated four models from ARCH family models. Among

them GARCH(1,1) is the best because it has lowest AIC and SIC
values.

DIAGNOSTIC CHECKING
1- serial correlation

Table 13.7 Eviews Output for Serial Correlation Test

Sample: 1/02/1989 12/29/2006 Included observations: 4695
Autocorrelation Partial Correlation AC PAC Q-Stat Prob
| | | | 1 0.014 0.014 0.9418 0.332
| | | | 2 0.006 0.006 1.1159 0.572
| | | | 3 0.023 0.023 3.6682 0.300
| | | | 4 -0.002 -0.002 3.6835 0.451
| | | | 5 -0.004 -0.005 3.7701 0.583
| | | | 6 -0.015 -0.016 4.8983 0.557
| | | | 7 -0.018 -0.017 6.4174 0.492
| | | | 8 -0.017 -0.016 7.7068 0.463
| | | | 9 -0.024 -0.023 10.378 0.321
| | | | 10 0.014 0.016 11.292 0.335
| | | | 11 -0.010 -0.009 11.745 0.383
| | | | 12 -0.019 -0.019 13.528 0.332
| | | | 13 -0.006 -0.007 13.695 0.396
| | | | 14 -0.031 -0.032 18.368 0.191
| | | | 15 -0.005 -0.005 18.503 0.237
| | | | 16 -0.032 -0.032 23.267 0.107
| | | | 17 -0.021 -0.019 25.321 0.088
| | | | 18 -0.021 -0.021 27.357 0.073
| | | | 19 -0.031 -0.030 31.839 0.033
| | | | 20 -0.006 -0.007 32.008 0.043
| | | | 21 -0.010 -0.011 32.443 0.053
| | | | 22 -0.011 -0.012 33.040 0.061
| | | | 23 -0.010 -0.013 33.482 0.073
| | | | 24 -0.016 -0.018 34.710 0.073
| | | | 25 0.002 -0.002 34.724 0.093
| | | | 26 0.009 0.006 35.096 0.110
| | | | 27 0.007 0.004 35.297 0.131
| | | | 28 -0.011 -0.016 35.830 0.147
| | | | 29 -0.001 -0.004 35.835 0.178
| | | | 30 0.004 -0.001 35.930 0.210
| | | | 31 -0.017 -0.021 37.261 0.203
| | | | 32 -0.034 -0.038 42.780 0.097
| | | | 33 0.002 -0.002 42.798 0.118
| | | | 34 0.011 0.009 43.375 0.130
| | | | 35 0.012 0.008 44.023 0.141
| | | | 36 -0.019 -0.025 45.818 0.127
 Volatılıty 257

Since all the correspondent p-values are more than 5% then we do

not have any evidence to reject null hypothesis of there is no serial
correlation.

2- whether our model has ARCH effect or not

Table 13.8 Eviews Output for Heteroscedasticity Test

Heteroskedasticity Test: ARCH

F-statistic 0.940887 Prob. F(1,4692) 0.3321

Obs*R-squared 0.941100 Prob. Chi-Square(1) 0.3320

Test Equation:
Dependent Variable: WGT_RESID^2
Method: Least Squares
Sample (adjusted): 1/03/1989 12/29/2006
Included observations: 4694 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

C 0.985305 0.039104 25.19718 0.0000

WGT_RESID^2(-1) 0.014159 0.014597 0.969994 0.3321

R-squared 0.000200 Mean dependent var 0.999463

Adjusted R-squared -0.000013 S.D. dependent var 2.485467
S.E. of regression 2.485483 Akaike info criterion 4.659237
Sum squared resid 28985.42 Schwarz criterion 4.661987
Log likelihood -10933.23 Hannan-Quinn criter. 4.660204
F-statistic 0.940887 Durbin-Watson stat 2.000046
Prob(F-statistic) 0.332100

P value is bigger than 5% it means that we do not have enough

evidence to reject null hypothesis of there is no ARCH effect.
258 Applied Econometrics with Eviews Applications

3- Whether the residuals are normally distributed

1,400
Series: Standardized Residuals
1,200 Sample 1/02/1989 12/29/2006
Observations 4695
1,000
Mean 0.029548
Median -0.009699
800
Maximum 7.878602
Minimum -7.410734
600 Std. Dev. 0.999510
Skewness 0.265750
400 Kurtosis 7.154967

200 Jarque-Bera 3432.483

Probability 0.000000
0
-6 -4 -2 0 2 4 6 8

Figure 13.1 Eviews Output for Normality Test

Correspondent p-value is less than 5%. It means that we have enough

evidence to reject null hypothesis of residuals are normally
distributed.

This model does not have serial correlation, ARCH effect but it is not
normally distributed. This model still can be accepted although
residuals are not normally distributed.

VOLATILITY SPILLOVER, GARCH (1,1) MODEL

Mean Equation

(13.7)

Variance Equation

(13.8)

First plot the residuals derived from the mean equation.

 Volatılıty 259

8,000

6,000

4,000
4,000

3,000 2,000

2,000
0
1,000

-1,000

-2,000
1992 1994 1996 1998 2000 2002 2004 2006

Residual Actual Fitted

Figure 13.2 Eviews Output for Residuals Derived from Mean

Equation

: variance of the residual derived from mean equation. It is

known as current days variance or volatility of GREECE.

: previous days residual variance or volatility. It is known as

GARCH term.

: previous periods squared residual derived from mean equation.

It is known as previous days bond return information about
volatility. It is ARCH term.

TURKEY, GERMANY and ENGLAND are known as variance regressors

as they can also contribute in the volatility of GREECE ( ) in the
variance equation.

This model is GARCH(1,1) model which refers to first order ARCH

term and first order GARCH term.
260 Applied Econometrics with Eviews Applications

Here we estimate mean equation and variance equation

simultaneously.

Table 13.9 Eviews Output for Volatility Spillover Test

Dependent Variable: GREECE
Method: ML - ARCH (Marquardt) - Normal distribution
Sample (adjusted): 12/28/1990 12/29/2006
Included observations: 4176 after adjustments
Convergence achieved after 56 iterations
Presample variance: backcast (parameter = 0.7)
GARCH = C(3) + C(4)*RESID(-1)^2 + C(5)*GARCH(-1) + C(6)*TURKEY +
C(7)*GERMANY + C(8)*ENGLAND

Variable Coefficient Std. Error z-Statistic Prob.

FRANCE 1.263438 0.005808 217.5493 0.0000

C -712.8147 21.51183 -33.13594 0.0000

Variance Equation

C 287684.6 9989.890 28.79758 0.0000

RESID(-1)^2 0.793084 0.084408 9.395841 0.0000
GARCH(-1) -0.188672 0.016241 -11.61692 0.0000
TURKEY -1.204154 0.107242 -11.22839 0.0000
GERMANY -4.213912 2.776550 -1.517679 0.1291
ENGLAND -35.63601 1.747883 -20.38809 0.0000

R-squared 0.828314 Mean dependent var 2169.725

Adjusted R-squared 0.828025 S.D. dependent var 1308.362
S.E. of regression 542.5753 Akaike info criterion 14.15328
Sum squared resid 1.23E+09 Schwarz criterion 14.16542
Log likelihood -29544.05 Hannan-Quinn criter. 14.15758
F-statistic 2872.692 Durbin-Watson stat 0.007986
Prob(F-statistic) 0.000000

ARCH is significant. It means that previous days Greece stock market

return information (ARCH) can influence today’s Greece stock market
volatility.
 Volatılıty 261

GARCH is also significant. It means that previous days Greece stock

market index volatility (GARCH) can influence today’s Greece stock
market volatility.

It means that Greece stock market volatility is influenced by its own

ARCH and GARCH factors or own shocks.

Turkey stock market and England stock market are also significant
meaning that Turkey and England stock market volatility influence
the volatility of Greece stock market. In In summary, volatility
spillover from Turkey and England stock markets to Greece stock
market is detected. Greece stock market volatility is mainly defined
by its own shocks such as ARCH and GARCH. It is also influenced by
the Turkey and England stock market volatility. Volatility of the
Germany stock market does not contribute in the volatility of Greece
stock market.
262 Applied Econometrics with Eviews Applications
 Granger Causalıty 263

CHAPTER 14

GRANGER CAUSALITY

If event A happens before event B, then it is possible that A is causing

B. In other words, events in the past can cause events to happen
today. Future events cannot cause past events.

Causality concern is examined by Granger causality (Granger, 1969).

The basic idea is that a variable x Granger causes y if past values of x
can help explain y. Of course, if Granger causality holds this does not
guarantee that x causes y. Nevertheless, if past values of x have
explanatory power for current values of y, it at least suggests that x
might be causing y.

Granger causality is only relevant with time series variables.

Let’s consider Granger causality between two variables (x and y)

which are both stationary. A nonstationary case, where x and y have
unit roots but are cointegrated, will be mentioned below.

Granger Causality in simple ARDL model is:

, (14.1)
264 Applied Econometrics with Eviews Applications

This model implies that last period’s value of x has explanatory power
for the current value of y. The coefficient measures the level of
the influence of on . If , then past values of x have no
effect on y and x does not Granger cause y. An alternative way of
expressing this concept is to say that if then past values of x
have no explanatory power for y over sample period”. Since we know
how to estimate the ARDL and carry out hypothesis tests, it is simple
to test for Granger causality.

OLS estimation and correspondent t-statistics and p-values can be

conducted for the coefficients. If is statistically significant (p-value
is higher than chosen significance level) then we conclude that x
Granger causes y.

The null hypothesis being tested here is H0: which is a

hypothesis that Granger causality does not occur. The test of
may be referred as a test of Granger non-causality.

Granger Causality in ARDL (p,q) model:

,
(14.2)

Here x Granger causes y if any or all of are statistically

significant. Since we are assuming x and y does not contain unit
roots, OLS regression analysis can be used to estimate this model and
correspondent p-values of the individual coefficients can be used to
determine whether Granger causality is present. If any of the p-
values for the coefficients were less significance level,
then Granger causality is present. If none of the p-values is less than
significance level, then Granger causality is not present.

If any or all of the coefficients are significant using t-

statistics or the p-values of individual coefficients, then x Granger
causes y. If none of these coefficients is significant, then x does not
Granger cause y.
 Granger Causalıty 265

CAUSALITY IN BOTH DIRECTIONS

Mostly, it is not clear where the direction of the causality is. There
can be also bidirectional causality. For instance, should past interest
rate changes cause inflation or should the reverse hold? If there is
possibility of causality in either direction, it should be checked
running a regression of x on lags of itself and lags of y.

Bidirectional causality is also equally reasonable. For example,

exchange rates may also affect future interest rate policy.

GRANGER CAUSALITY IN COINTEGRATED

VARIABLES
Testing for Granger causality among cointegrated variables is very
similar to the method outlined above. It is common to work with a
variant of the error correction model (ECM):

, (14.3)

This is an ARDL model except the term .

an estimate of which can be obtained by

running a regression of y on x and saving the residuals.

x Granger causes y if past values of x have explanatory power for

current values of y. Applying this interpretation to the ECM, the past
values of x appear in the terms . This
implies that x does not Granger cause y if
Thus, t-statistics and p-values can be used to test for Granger
causality. Also, the F-tests can be used to carry out a formal test of:
.

Testing whether y Granger causes x can be done by reversing the

roles of x and y in the ECM. If x and y are cointegrated then Granger
266 Applied Econometrics with Eviews Applications

causality must occur (Granger Representation Theorem). That is,

either x must Granger cause y or y must Granger cause x (or both).

Example 14.1

Whether Greece stock market causes France stock market or France

stock market causes Greece stock market?

(14.4)
(14.5)

The variables in the model should be stationary. If not, they should

be converted into stationary variables by taking first difference.

We assume that and are not correlated.

Null Hypothesis is Greece stock market does not granger cause

France stock market. F- statistic can be used to test null hypothesis.

Table 14.1 Eviews Outputs for Granger Causality Test

Pairwise Granger Causality Tests
Sample: 12/30/1988 12/29/2006
Lags: 2

Null Hypothesis: Obs F-Statistic Prob.

DFRANCE does not Granger Cause DGREECE 4173 25.3000 1.E-11

DGREECE does not Granger Cause DFRANCE 0.32394 0.7233

Pairwise Granger Causality Tests

Sample: 12/30/1988 12/29/2006
Lags: 4

Null Hypothesis: Obs F-Statistic Prob.

DFRANCE does not Granger Cause DGREECE 4171 16.7342 1.E-13

DGREECE does not Granger Cause DFRANCE 2.15275 0.0718
 Vector Autoregressıon Model (VAR) 267

CHAPTER 15

VECTOR AUTOREGRESSION
MODEL (VAR)

DEFINITION
In estimating Granger causality between x and y, restricted version of
an ARDL (p, q) model is used with y as the dependent variable to
investigate whether x Granger cause y. Causality in the other
direction can also be examined switching the roles of x and y in the
ARDL. The two equations can be written as follows:

(15.1)

(15.2)

The first one is used to test whether x Granger causes y and the
second one, whether y Granger causes x.

These two equations comprise a VAR. A VAR is the extension of the

autoregressive (AR) model to the case in which there is more than
one variable under study.
268 Applied Econometrics with Eviews Applications

The AR model involves one dependent variable, which depended only

on lags of itself (and possibly a deterministic trend). A VAR has more
than one dependent variable and, has more than one equation. Each
equation uses its explanatory variables lags of all the variables under
study (and possibly a deterministic trend).

The Eq. (15.1) and Eq. (15.2) form a VAR model with two variables.

In the first equation y depends on p lags of itself and on q lags of x.

The lag lengths are p and q. They can be selected using the sequential
testing methods. However, especially if the VAR has more than two
variables, many different lag lengths need to be selected (i.e. one for
each variable in each equation).

It is common to set p = q and use the same lag length for every
variable in each equation. The resulting model is known as a VAR( p)
model.

For example, the following VAR (p) has three variables, y, x and z;

(15.3)

(15.4)

(15.5)

In addition to an intercept and deterministic trend, each equation

contains p lags of all variables in study.

It is assumed that all the variables in the VAR (p) are stationary.
Estimation and testing can be carried out by using OLS. P-values or t-
statistics help to examine whether individual coefficients are
significant.
 Vector Autoregressıon Model (VAR) 269

The reasons to use VARs are:

- It is easy to use.

- It provides a framework for testing for Granger causality

between each set of variables.

Economic theories or common sense help to interpret the results

obtained through VAR model. For instance, x (exam score) caused y
(class size) or x (lot size) influenced y (house price). In both cases it is
not plausible for us to say that y influenced or caused x.

However, there are many circumstances in which neither economic

theory nor common sense can provide sensible interpretation
reflecting causality. For instance, does y (interest rate) cause x
(inflation) or vice versa? Economic theory and common sense tells us
that either can happen and that Granger causality tests helps to
answer these questions.

If cointegration is present one should be careful when interpreting

regression results as reflecting causality.

In the VAR model, the explanatory variables might influence the

dependent variable, but there is no possibility that the dependent
variable influences the explanatory variable.

One of the drawback in VARs is they are not theoretical which are not
based on economic theory. There is theory in selecting the variables
for the VAR.

For example, the interaction between interest rates, the price level,
money supply and real GDP. There have been created many
sophisticated models for this relationship (The IS–LM model).
However, a VAR model says: Interest rates, price level, money supply
and real GDP are related. This relationship can be modeled as each
variable depends on lags of itself and all other variables. There is no
need to establish any link between the empirical VAR and a
theoretical macroeconomic model.
270 Applied Econometrics with Eviews Applications

The VAR has better forecasting performance than sophisticated

macroeconomic models. The regression-based methods can
outperform complicated macroeconomic models. This is a strong
motivation for using VARs.

LAG LENGTH SELECTION

The following sequential testing strategy can be used in lag selection:

Step1. Choose the maximum possible lag length, pmax, that

seems reasonable to you.

Step2. Estimate a VAR( pmax). If any of the variables lagged

pmax periods are significant, use the VAR( pmax),
otherwise proceed to the next step.

Step3. Estimate a VAR( pmax - 1). If any of the variables lagged

pmax - 1 periods are significant, use the VAR( pmax -
1), otherwise proceed to the next step.

Step4. Estimate a VAR( pmax - 2). If any of the variables lagged

pmax - 2 periods are significant, use the VAR( pmax -
2), otherwise proceed to the next step and try pmax -
3, etc.

FORECASTING WITH “VAR”

Forecasting in econometrics is done using time series variables. The
idea is that using observed future is predicted.

To provide some notions for how forecasting is done, consider a

VAR(1) involving two variables, y and x:

(15.6)

(15.7)
 Vector Autoregressıon Model (VAR) 271

is not observable but it can be predicted. Using the Eq. (15.6)

and setting t = T + 1, following equation is obtained for :

(15.8)

This equation cannot be directly used to obtain since we don’t

know what is. It is not clear that what unpredicTable shock or
surprise will hit the economy next period. Furthermore, the
coefficients are not known. But, if the error term (which cannot be
predicted) is ignored and the coefficients are replaced by OLS
estimates a forecast can be obtained which is denoted as ̂ :

̂ ̂ ̂ ̂ ̂ (15.9)

The strategy for how to forecast one period into the future can be
used for two periods on the condition that one is an extension. In the
one period, and are used to create ̂ and ̂ . In the two
period, ̂ and ̂ subject to and . Since the data runs
until period T, what is and are not known. Consequently
and are replaced by ̂ and ̂ . That is, use the
relevant equation from the VAR, is used ignoring the error, replacing
the coefficients by their OLS estimates and replacing past values of
the variables that are not observed by forecast.

In a formula:

̂ ̂ ̂ ̂ ̂ ̂ ̂ (15.10)

̂ can also be calculated using the following formula:

̂ ̂ ̂ ̂ ̂ ̂ ̂ (15.11)

The general strategy which are ignoring the error, replacing

coefficients by OLS estimates and replacing lagged values of variables
that are unobserved by forecasts, can be used in order to obtain
forecasts for any number of future periods for any VAR(p).
272 Applied Econometrics with Eviews Applications

OLS provides estimates only for coefficients which cannot be

precisely corrected, it is also recommended to report confidence
intervals.

VECTOR AUTOREGRESSIONS WITH COINTEGRATED

VARIABLES
So far, we have assumed that all variables are stationary. If some of
the original variables are non-stationary and are not cointegrated,
then non-stationary variables should be differenced and used in the
VAR. This is valid except if the variables are non-stationary in levels
and are cointegrated.

In Granger causality, it is recommended to work with an ECM.

Instead of working with a vector autoregression (VAR), a vector error
correction model (VECM) should be employed. Like the VAR, the
VECM has one equation for each variable in the model.

If there are two variables, y and x, the VECM is:

(15.12)

(15.13)

. (15.14)

The VECM is the same as a VAR with differenced variables, except for
the term . This error correction variable can be estimated by
running an OLS regression of y on x and saving the residuals. Then
OLS can be used to estimate ECMs, and p-values and correspondent
confidence intervals can be obtained. Lag length selection and
forecasting also can be done in a similar way to the VAR, with the
forecasts of the error correction term, must be calculated.
 Vector Autoregressıon Model (VAR) 273

In summary, this is simple using OLS estimates of and and

replacing the error, , by the residual .

Note the following issues:

i- If x and y are stationary, standard statistical methods

based on an ARDL model can be used to test for
Granger causality.

ii- If x and y are non-stationary and are cointegrated,

statistical methods based on an ECM can be used to
test for Granger causality.

iii- Vector autoregressions have one equation for each

variable. Each equation uses one variable as the
dependent variable. The explanatory variables are lags
of all the variables under study.

iv- VARs are useful for forecasting, and Granger causality

test is useful for understanding the relationships
between several series.

v- If all the variables in the VAR are stationary, OLS can

be used to estimate each equation and standard
statistical methods can be employed (p-values and t-
statistics can be used to test for significance of
variables).

vi- If the variables under study non-stationary and are

cointegrated, a variant on the VAR called the Vector
Error Correction Model, or VECM, can be used.
274 Applied Econometrics with Eviews Applications

Example 15.1
Table 15.1 Eviews Output for VAR Analysis
Vector Autoregression Estimates
Sample (adjusted): 1/01/1991 12/29/2006
Included observations: 4174 after adjustments
Standard errors in ( ) & t-statistics in [ ]

GREECE FRANCE TURKEY

GREECE(-1) 1.142527 0.007954 -0.240347

(0.01585) (0.01028) (0.11145)
[ 72.0657] [ 0.77401] [-2.15652]

GREECE(-2) -0.143405 -0.004824 0.262410

(0.01587) (0.01029) (0.11160)
[-9.03349] [-0.46878] [ 2.35140]

FRANCE(-1) 0.190865 1.024415 0.679475

(0.02523) (0.01635) (0.17736)
[ 7.56520] [ 62.6386] [ 3.83107]

FRANCE(-2) -0.190844 -0.029875 -0.709881

(0.02518) (0.01632) (0.17704)
[-7.57802] [-1.83003] [-4.00973]

TURKEY(-1) -0.007488 -0.000914 1.016284

(0.00228) (0.00148) (0.01602)
[-3.28514] [-0.61831] [ 63.4212]

TURKEY(-2) 0.007555 0.001036 -0.015816

(0.00228) (0.00148) (0.01604)
[ 3.31170] [ 0.70028] [-0.98615]
C 1.872509 4.919013 24.41604
(2.19474) (1.42269) (15.4287)
[ 0.85318] [ 3.45755] [ 1.58251]

R-squared 0.998927 0.999085 0.999321

Adj. R-squared 0.998926 0.999083 0.999320
Sum sq. resids 7663596. 3220236. 3.79E+08
S.E. equation 42.88491 27.79919 301.4754
F-statistic 646709.7 757954.6 1021740.
Log likelihood -21607.21 -19797.73 -29747.21
Akaike AIC 10.35659 9.489566 14.25693
Schwarz SC 10.36722 9.500193 14.26756
Mean dependent 2170.302 2227.884 9341.112
S.D. dependent 1308.410 918.1297 11559.02
 Vector Autoregressıon Model (VAR) 275

The corresponding equations are;

GREECE = C(1)*GREECE(-1) + C(2)*GREECE(-2) + C(3)*FRANCE(-1) +

C(4)*FRANCE(-2) + C(5)*TURKEY(-1) + C(6)*TURKEY(-2) + C(7)
(15.15)

FRANCE = C(8)*GREECE(-1) + C(9)*GREECE(-2) + C(10)*FRANCE(-1) +

C(11)*FRANCE(-2) + C(12)*TURKEY(-1) + C(13)*TURKEY(-2) + C(14)
(15.16)

TURKEY = C(15)*GREECE(-1) + C(16)*GREECE(-2) + C(17)*FRANCE(-1) +

C(18)*FRANCE(-2) + C(19)*TURKEY(-1) + C(20)*TURKEY(-2) + C(21)
(15.17)

Table 15.2 Eviews Output for VAR Analysis

Estimation Method: Least Squares
Sample: 1/01/1991 12/29/2006
Included observations: 4174
Total system (balanced) observations 12522

Coefficient Std. Error t-Statistic Prob.

C(1) 1.142527 0.015854 72.06569 0.0000

C(2) -0.143405 0.015875 -9.033488 0.0000
C(3) 0.190865 0.025229 7.565201 0.0000
C(4) -0.190844 0.025184 -7.578016 0.0000
C(5) -0.007488 0.002279 -3.285137 0.0010
C(6) 0.007555 0.002281 3.311698 0.0009
C(7) 1.872509 2.194735 0.853182 0.3936
C(8) 0.007954 0.010277 0.774005 0.4389
C(9) -0.004824 0.010290 -0.468784 0.6392
C(10) 1.024415 0.016354 62.63860 0.0000
C(11) -0.029875 0.016325 -1.830032 0.0673
C(12) -0.000914 0.001478 -0.618314 0.5364
C(13) 0.001036 0.001479 0.700279 0.4838
C(14) 4.919013 1.422688 3.457548 0.0005
C(15) -0.240347 0.111451 -2.156518 0.0311
C(16) 0.262410 0.111598 2.351398 0.0187
C(17) 0.679475 0.177359 3.831065 0.0001
C(18) -0.709881 0.177040 -4.009731 0.0001
C(19) 1.016284 0.016024 63.42118 0.0000
C(20) -0.015816 0.016038 -0.986153 0.3241
C(21) 24.41604 15.42870 1.582508 0.1136

Determinant residual covariance 1.11E+11

276 Applied Econometrics with Eviews Applications

Equation: GREECE = C(1)*GREECE(-1) + C(2)*GREECE(-2) + C(3)

*FRANCE(-1) + C(4)*FRANCE(-2) + C(5)*TURKEY(-1) + C(6)*TURKEY(
-2) + C(7)
Observations: 4174
R-squared 0.998927 Mean dependent var 2170.302
Adjusted R-squared 0.998926 S.D. dependent var 1308.410
S.E. of regression 42.88491 Sum squared resid 7663596.
Durbin-Watson stat 1.995159

Equation: FRANCE = C(8)*GREECE(-1) + C(9)*GREECE(-2) + C(10)

*FRANCE(-1) + C(11)*FRANCE(-2) + C(12)*TURKEY(-1) + C(13)
*TURKEY(-2) + C(14)
Observations: 4174
R-squared 0.999085 Mean dependent var 2227.884
Adjusted R-squared 0.999083 S.D. dependent var 918.1297
S.E. of regression 27.79919 Sum squared resid 3220236.
Durbin-Watson stat 1.998785

Equation: TURKEY = C(15)*GREECE(-1) + C(16)*GREECE(-2) + C(17)

*FRANCE(-1) + C(18)*FRANCE(-2) + C(19)*TURKEY(-1) + C(20)
*TURKEY(-2) + C(21)
Observations: 4174
R-squared 0.999321 Mean dependent var 9341.112
Adjusted R-squared 0.999320 S.D. dependent var 11559.02
S.E. of regression 301.4754 Sum squared resid 3.79E+08
Durbin-Watson stat 2.000858
 References 277

REFERENCES
Baltagi, B.H. (2008).Econometric Analysis of Panel Data. Fourth
Edition. Great Briatin: John Wiley & Sons

Fischer, B. (1976). The pricing of commodity contracts, Journal of

Financial Economics 3.p.167-179.

Bollerslev, T. (1986). Generalized Autoregressive Conditional

Heteroskedasticity. Journal of Econometrics 31.p.307-327.

Box,G.E.P. & Pierce, D.A. (1970). Distribution of the Autocorrelations

in Auotoregressive moving Average Time Series Models.
Journal of American Statistical Association 65. p. 1509-1526.

Breusch, T.S. (1979). Testing for Autocorrelation in Dynamic Linear

Models. Australian Economic Papers 17.p. 334–355.

Breusch, T.S.&Pagan, A.R.(1980). The Lagrange Multiplier Test and Its

Applications to Model Specification in Econometrics. Review
of Economic Studies 47(1). p. 239-253.

Brown, R.L.; Durbin, J. & Evans, J.M. (1975). Techniques for Testing
the Constancy of Regression Relationships over Time. Journal
of the Royal Statistical Society. Ser.B. 37.p.149 – 192.

Charemza, W.W & Deadman, D.F.(1997). New Directions in

Econometric Practice. Cheltenham: Edward Elgar.
278 Applied Econometrics with Eviews Applications

Chow, G. C. (1960). Tests of Equality Between Sets of Coefficients in

Two Linear Regressions. Econometrica 28 (3).p. 591–605.

Cobb, C. W.& Douglas, P. H. (1928). A Theory of Production. American

Economic Review 18.p. 139–165.

Cochrane, J.H.(2001). Asset Pricing. USA: Princeton University Press.

Collis, J & Hussey, R. (2009). Business Research: A Practical Guide for

Undergraduate and Postgraduate Students. Third Edition.
China. Palgrave Macmillan

Dickey, D. A. & Fuller, W. A. (1979). Distribution of the Estimators for

Autoregressive Time Series with a Unit Root. Journal of the
American Statistical Association 74 (366).p. 427–431.

Durbin, J. & Watson, G.S. (1950). Testing for Serial Correlation in

Least-Squares Regression Biometrika 37. p.409-428.

Engle, R.F. (1982). Autoregressive Conditional Heteroscedasticity with

Estimates of the Variance of United Kingdom Inflation ,
Econometrica. 50(4).p. 987-1007.

Engle, R.F. (1982). Autoregressive Conditional Heteroskedasticity with

Estimates of the Variance of U.K. Inflation. Econometrica
50.p. 987-1008.

Engle, R.F. & C.W.J. Granger (1987). Cointegration and Error

Correction: Representation, Estimation and Testing.
Econometrica 55.p.251-76.

Engle, R.F. & B.S. Yoo (1991). Cointegrated Economic Time Series: An
Overview with New Results” in R.F. Engle and C.W.J. Granger
(eds.), Long-run Economic Relationships: Readings in
Cointegration, Oxford: Oxford University Press.

Fisher, R. A. (1948). Conclusions fiduciaires. Ann. Inst. H. Poincaré 10.

p. 191–213.
 279

Fox, J.H.(1958). Criteria of Good Research. Phi Delta Kappan. Vol. 39.
P. 285–86.

Goldfeld, S. M. & Quandt, R. E. (1965). Some Tests for the

Homoscedasticity. Journal of the American Statistical
Association 60. p. 539-547.

Godfrey, L.G. (1978). Testing for Higher Order Serial Correlation in

Regression Equations when the Regressors Include Lagged
Dependent Variables. Econometrica 46. p.1303-1310.

Glejser, H. (1969). A New Test for Heteroscedasticity. Journal of

American Statistical Association. 64. p. 316-323.

Glosten, L. R., Jagannathan, R., and Runkle, D. E. (1993). On the

Relation between the Expected Value and the Volatility of the
Nominal Excess Return on Stocks. Journal of Finance 48.p.
1779-1801.

Granger, C. W. J. (1969). Investigating Causal Relations by

Econometric Models and Cross-spectral
Methods. Econometrica 37 (3).p. 424–438.

Granger, C.W.J. (1981). Some Properties of Time Series Data and

Their Use in Econometric Model Specification. Journal of
Econometrics 16.p. 121-30.

Harvey, A.C. (1976). Estimating Regression Models with Multiplicative

Heteroscedasticity. Econometrica 44(3).p. 461

Hausman, J.A. (1978). Specification Tests in Econometrics.

Econometrica 46 (6).p. 1251-1271.

Jarque, C. M. & Bera, A. K. (1981). Efficient Tests for Normality,

Homoscedasticity and Serial Independence of Regression
Residuals: Monte Carlo Evidence. Economics Letters 7 (4).p.
313–318.
280 Applied Econometrics with Eviews Applications

Johansen, S. (1988). Statistical Analysis of Cointegration Vectors.

Journal of Economic Dynamics and Control 12.p.231-254.

Johansen, S. (1991). Estimation and Hypothesis Testing of

Cointegration Vectors in Gaussian Vector Autoregressive
Models. Econometrica 55.p. 1551-1580.

Johansen, S.& Juselius, K. (1994). Identification of the Long-Run and

the Short Run Structure. An Application to the ISLM Model. J.
Econom 63.p. 7-36.

Koop, G. (2005). Analysis of Financial Data. Second Edition. England:

John Wiley &Sons.

Kothari, C.R.(2004). Research Methodology: Methods and Techniques.

Second Edition. New Age International

Kumar, R. (2011). Research Methodology: A Step-by-Step Guide for

Beginners. Third Edition. Great Britain. SAGE.

Lee, H.B & Kerlinger, F.N. (2000) Foundations of Behavioral Research.

Third Edition. USA, Earl McPeek.

Ljung, G.M. & Box, G.E.P.(1978). On a measure of Lack of Fit in Time

Series. Biometrica 65. p. 297-303.

Nelson, D.B. (1990). ARCH models as diffusion approximations,

Journal of Econometrics 45.p. 7 - 38.

Phillips, P. C. B. & Ouliaris, S. (1990). Asymptotic Properties of

Residual Based Tests for Cointegration. Econometrica 58,
165–193.

Phillips, P.C.B. (1991). Optimal Inference in Cointegrated Systems.

Econometrica 59.p. 283-306.

Ramsey, J.B. (1969). Tests for Specification Errors in Classical Linear

Least Squares Regression Analysis. Journal of the Royal
Statistical Society, Series B., 31(2), 350–371.
 281

Savin, N.E. & White, K.J. (1977). The Durbin-Watson Test for Serial
Correlation with Extreme Sample Sizes or Many Regressors.
Econometrica 45.p. 1989-1996.

Stock, J.H. & Watson, M.W. (1988). Testing for Common Trends.
Journal of the American Statistical Association 83.p. 1097-
1107.

Stock, J.H & Watson, M.W. (2008). Introduction to Econometrics. Brief

Edition. USA: Greg Tobin

Verbeek, M.(2004). A Guide to Modern Econometrics. Second

Edition.England: John Wiley & Sons.

Vogelvang, B. (2005). Econometrics Theory and Applications with

Eviews. England: Pearson

Wald, A. (1943). Tests of Statistical Hypotheses Concerning Several

Parameters when the Number of Observations is Large.
Transactions of the American Mathematical Society 54. p.
426-482.

White,H. (1980). A Heteroskedasticity-Consistent Covariance Matrix

Estimator and a Direct Test for Heteroskedasticity.
Econometrica 48 (4).p.817-838.

White, H. (1980). A Heteroscedasticity-Consistent Covariance Matrix

Estimator and a Direct Test for Heteroscedasticity.
Econometrica. 48.p.817-838.
282 Applied Econometrics with Eviews Applications

Standard Normal Curve Probability Distribution

The table is based on the upper right 1/2 of the Normal Distribution; total area shown is 0.5
The Z-score values are represented by the column value + row value, up to two decimal places
The probabilities up to the Z-score are in the cells

Z 0,00 0,01 0,02 0,03 0,04 0,05 0,06 0,07 0,08 0,09
0,0 0,0000 0,0040 0,0080 0,0120 0,0160 0,0199 0,0239 0,0279 0,0319 0,0359
0,1 0,0398 0,0438 0,0478 0,0517 0,0557 0,0596 0,0636 0,0675 0,0714 0,0753
0,2 0,0793 0,0832 0,0871 0,0910 0,0948 0,0987 0,1026 0,1064 0,1103 0,1141
0,3 0,1179 0,1217 0,1255 0,1293 0,1331 0,1368 0,1406 0,1443 0,1480 0,1517
0,4 0,1554 0,1591 0,1628 0,1664 0,1700 0,1736 0,1772 0,1808 0,1844 0,1879
0,5 0,1915 0,1950 0,1985 0,2019 0,2054 0,2088 0,2123 0,2157 0,2190 0,2224
0,6 0,2257 0,2291 0,2324 0,2357 0,2389 0,2422 0,2454 0,2486 0,2517 0,2549
0,7 0,2580 0,2611 0,2642 0,2673 0,2704 0,2734 0,2764 0,2794 0,2823 0,2852
0,8 0,2881 0,2910 0,2939 0,2967 0,2995 0,3023 0,3051 0,3078 0,3106 0,3133
0,9 0,3159 0,3186 0,3212 0,3238 0,3264 0,3289 0,3315 0,3340 0,3365 0,3389
1,0 0,3413 0,3438 0,3461 0,3485 0,3508 0,3531 0,3554 0,3577 0,3599 0,3621
1,1 0,3643 0,3665 0,3686 0,3708 0,3729 0,3749 0,3770 0,3790 0,3810 0,3830
1,2 0,3849 0,3869 0,3888 0,3907 0,3925 0,3944 0,3962 0,3980 0,3997 0,4015
1,3 0,4032 0,4049 0,4066 0,4082 0,4099 0,4115 0,4131 0,4147 0,4162 0,4177
1,4 0,4192 0,4207 0,4222 0,4236 0,4251 0,4265 0,4279 0,4292 0,4306 0,4319
1,5 0,4332 0,4345 0,4357 0,4370 0,4382 0,4394 0,4406 0,4418 0,4429 0,4441
1,6 0,4452 0,4463 0,4474 0,4484 0,4495 0,4505 0,4515 0,4525 0,4535 0,4545
1,7 0,4554 0,4564 0,4573 0,4582 0,4591 0,4599 0,4608 0,4616 0,4625 0,4633
1,8 0,4641 0,4649 0,4656 0,4664 0,4671 0,4678 0,4686 0,4693 0,4699 0,4706
1,9 0,4713 0,4719 0,4726 0,4732 0,4738 0,4744 0,4750 0,4756 0,4761 0,4767
2,0 0,4772 0,4778 0,4783 0,4788 0,4793 0,4798 0,4803 0,4808 0,4812 0,4817
2,1 0,4821 0,4826 0,4830 0,4834 0,4838 0,4842 0,4846 0,4850 0,4854 0,4857
2,2 0,4861 0,4864 0,4868 0,4871 0,4875 0,4878 0,4881 0,4884 0,4887 0,4890
2,3 0,4893 0,4896 0,4898 0,4901 0,4904 0,4906 0,4909 0,4911 0,4913 0,4916
2,4 0,4918 0,4920 0,4922 0,4925 0,4927 0,4929 0,4931 0,4932 0,4934 0,4936
2,5 0,4938 0,4940 0,4941 0,4943 0,4945 0,4946 0,4948 0,4949 0,4951 0,4952
2,6 0,4953 0,4955 0,4956 0,4957 0,4959 0,4960 0,4961 0,4962 0,4963 0,4964
2,7 0,4965 0,4966 0,4967 0,4968 0,4969 0,4970 0,4971 0,4972 0,4973 0,4974
2,8 0,4974 0,4975 0,4976 0,4977 0,4977 0,4978 0,4979 0,4979 0,4980 0,4981
2,9 0,4981 0,4982 0,4982 0,4983 0,4984 0,4984 0,4985 0,4985 0,4986 0,4986
3,0 0,4987 0,4987 0,4987 0,4988 0,4988 0,4989 0,4989 0,4989 0,4990 0,4990
 283

Critical Values of t-Distribution

The table shows the critical t-values for a given alpha level (one-tailed) and degrees of
freedom
The degrees of freedom are the rows (denoted by df)
Note: The probability levels represent the whole of alpha (you must divide alpha by 2
if you want the t-value for a two-tailed test)

df 0,1000 0,0500 0,0250 0,0100 0,0050 0,0010 0,0005

1 3,078 6,314 12,706 31,821 63,657 318,309 636,619
2 1,886 2,920 4,303 6,965 9,925 22,327 31,599
3 1,638 2,353 3,182 4,541 5,841 10,215 12,924
4 1,533 2,132 2,776 3,747 4,604 7,173 8,610
5 1,476 2,015 2,571 3,365 4,032 5,893 6,869
6 1,440 1,943 2,447 3,143 3,707 5,208 5,959
7 1,415 1,895 2,365 2,998 3,499 4,785 5,408
8 1,397 1,860 2,306 2,896 3,355 4,501 5,041
9 1,383 1,833 2,262 2,821 3,250 4,297 4,781
10 1,372 1,812 2,228 2,764 3,169 4,144 4,587
11 1,363 1,796 2,201 2,718 3,106 4,025 4,437
12 1,356 1,782 2,179 2,681 3,055 3,930 4,318
13 1,350 1,771 2,160 2,650 3,012 3,852 4,221
14 1,345 1,761 2,145 2,624 2,977 3,787 4,140
15 1,341 1,753 2,131 2,602 2,947 3,733 4,073
16 1,337 1,746 2,120 2,583 2,921 3,686 4,015
17 1,333 1,740 2,110 2,567 2,898 3,646 3,965
18 1,330 1,734 2,101 2,552 2,878 3,610 3,922
19 1,328 1,729 2,093 2,539 2,861 3,579 3,883
20 1,325 1,725 2,086 2,528 2,845 3,552 3,850
21 1,323 1,721 2,080 2,518 2,831 3,527 3,819
22 1,321 1,717 2,074 2,508 2,819 3,505 3,792
23 1,319 1,714 2,069 2,500 2,807 3,485 3,768
24 1,318 1,711 2,064 2,492 2,797 3,467 3,745
25 1,316 1,708 2,060 2,485 2,787 3,450 3,725
26 1,315 1,706 2,056 2,479 2,779 3,435 3,707
27 1,314 1,703 2,052 2,473 2,771 3,421 3,690
28 1,313 1,701 2,048 2,467 2,763 3,408 3,674
29 1,311 1,699 2,045 2,462 2,756 3,396 3,659
30 1,310 1,697 2,042 2,457 2,750 3,385 3,646
40 1,303 1,684 2,021 2,423 2,704 3,307 3,551
50 1,299 1,676 2,009 2,403 2,678 3,261 3,496
60 1,296 1,671 2,000 2,390 2,660 3,232 3,460
100 1,290 1,660 1,984 2,364 2,626 3,174 3,390
120 1,289 1,658 1,980 2,358 2,617 3,160 3,373
Infinity 1,282 1,645 1,960 2,326 2,576 3,090 3,291
284 Applied Econometrics with Eviews Applications

F Table for Alpha=0.5

alpha=.05
df upper 1 2 3 4 5 6 7 8 9 10 12 15 20 24 30 40 60 120 inf
df lower
1 161,4 199,5 215,7 224,6 230,2 234,0 236,8 238,9 240,5 241,9 243,9 245,9 248,0 249,1 250,1 251,1 252,2 253,3 254,3
2 18,51 19,00 19,16 19,25 19,30 19,33 19,35 19,37 19,38 19,40 19,41 19,43 19,45 19,45 19,46 19,47 19,48 19,49 19,50
3 10,13 9,55 9,28 9,12 9,01 8,94 8,89 8,85 8,81 8,79 8,74 8,70 8,66 8,64 8,62 8,59 8,57 8,55 8,53
4 7,71 6,94 6,59 6,39 6,26 6,16 6,09 6,04 6,00 5,96 5,91 5,86 5,80 5,77 5,75 5,72 5,69 5,66 5,63

5 6,61 5,79 5,41 5,19 5,05 4,95 4,88 4,82 4,77 4,74 4,68 4,62 4,56 4,53 4,50 4,46 4,43 4,40 4,36
6 5,99 5,14 4,76 4,53 4,39 4,28 4,21 4,15 4,10 4,06 4,00 3,94 3,87 3,84 3,81 3,77 3,74 3,70 3,67
7 5,59 4,74 4,35 4,12 3,97 3,87 3,79 3,73 3,68 3,64 3,57 3,51 3,44 3,41 3,38 3,34 3,30 3,27 3,23
8 5,32 4,46 4,07 3,84 3,69 3,58 3,50 3,44 3,39 3,35 3,28 3,22 3,15 3,12 3,08 3,04 3,01 2,97 2,93
9 5,12 4,26 3,86 3,63 3,48 3,37 3,29 3,23 3,18 3,14 3,07 3,01 2,94 2,90 2,86 2,83 2,79 2,75 2,71

10 4,96 4,10 3,71 3,48 3,33 3,22 3,14 3,07 3,02 2,98 2,91 2,85 2,77 2,74 2,70 2,66 2,62 2,58 2,54
11 4,84 3,98 3,59 3,36 3,20 3,09 3,01 2,95 2,90 2,85 2,79 2,72 2,65 2,61 2,57 2,53 2,49 2,45 2,40
12 4,75 3,89 3,49 3,26 3,11 3,00 2,91 2,85 2,80 2,75 2,69 2,62 2,54 2,51 2,47 2,43 2,38 2,34 2,30
13 4,67 3,81 3,41 3,18 3,03 2,92 2,83 2,77 2,71 2,67 2,60 2,53 2,46 2,42 2,38 2,34 2,30 2,25 2,21
14 4,60 3,74 3,34 3,11 2,96 2,85 2,76 2,70 2,65 2,60 2,53 2,46 2,39 2,35 2,31 2,27 2,22 2,18 2,13

15 4,54 3,68 3,29 3,06 2,90 2,79 2,71 2,64 2,59 2,54 2,48 2,40 2,33 2,29 2,25 2,20 2,16 2,11 2,07
16 4,49 3,63 3,24 3,01 2,85 2,74 2,66 2,59 2,54 2,49 2,42 2,35 2,28 2,24 2,19 2,15 2,11 2,06 2,01
17 4,45 3,59 3,20 2,96 2,81 2,70 2,61 2,55 2,49 2,45 2,38 2,31 2,23 2,19 2,15 2,10 2,06 2,01 1,96
18 4,41 3,55 3,16 2,93 2,77 2,66 2,58 2,51 2,46 2,41 2,34 2,27 2,19 2,15 2,11 2,06 2,02 1,97 1,92
19 4,38 3,52 3,13 2,90 2,74 2,63 2,54 2,48 2,42 2,38 2,31 2,23 2,16 2,11 2,07 2,03 1,98 1,93 1,88

20 4,35 3,49 3,10 2,87 2,71 2,60 2,51 2,45 2,39 2,35 2,28 2,20 2,12 2,08 2,04 1,99 1,95 1,90 1,84
21 4,32 3,47 3,07 2,84 2,68 2,57 2,49 2,42 2,37 2,32 2,25 2,18 2,10 2,05 2,01 1,96 1,92 1,87 1,81
22 4,30 3,44 3,05 2,82 2,66 2,55 2,46 2,40 2,34 2,30 2,23 2,15 2,07 2,03 1,98 1,94 1,89 1,84 1,78
23 4,28 3,42 3,03 2,80 2,64 2,53 2,44 2,37 2,32 2,27 2,20 2,13 2,05 2,01 1,96 1,91 1,86 1,81 1,76
24 4,26 3,40 3,01 2,78 2,62 2,51 2,42 2,36 2,30 2,25 2,18 2,11 2,03 1,98 1,94 1,89 1,84 1,79 1,73

25 4,24 3,39 2,99 2,76 2,60 2,49 2,40 2,34 2,28 2,24 2,16 2,09 2,01 1,96 1,92 1,87 1,82 1,77 1,71
26 4,23 3,37 2,98 2,74 2,59 2,47 2,39 2,32 2,27 2,22 2,15 2,07 1,99 1,95 1,90 1,85 1,80 1,75 1,69
27 4,21 3,35 2,96 2,73 2,57 2,46 2,37 2,31 2,25 2,20 2,13 2,06 1,97 1,93 1,88 1,84 1,79 1,73 1,67
28 4,20 3,34 2,95 2,71 2,56 2,45 2,36 2,29 2,24 2,19 2,12 2,04 1,96 1,91 1,87 1,82 1,77 1,71 1,65
29 4,18 3,33 2,93 2,70 2,55 2,43 2,35 2,28 2,22 2,18 2,10 2,03 1,94 1,90 1,85 1,81 1,75 1,70 1,64

30 4,17 3,32 2,92 2,69 2,53 2,42 2,33 2,27 2,21 2,16 2,09 2,01 1,93 1,89 1,84 1,79 1,74 1,68 1,62
40 4,08 3,23 2,84 2,61 2,45 2,34 2,25 2,18 2,12 2,08 2,00 1,92 1,84 1,79 1,74 1,69 1,64 1,58 1,51
60 4,00 3,15 2,76 2,53 2,37 2,25 2,17 2,10 2,04 1,99 1,92 1,84 1,75 1,70 1,65 1,59 1,53 1,47 1,39
120 3,92 3,07 2,68 2,45 2,29 2,17 2,09 2,02 1,96 1,91 1,83 1,75 1,66 1,61 1,55 1,50 1,43 1,35 1,25
inf 3,84 3,00 2,60 2,37 2,21 2,10 2,01 1,94 1,88 1,83 1,75 1,67 1,57 1,52 1,46 1,39 1,32 1,22 1,00
 285

……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………
……………………………………………………………………………………………………………

View publication stats