JPLy. r
2 ...u

. unit

4% \:

3.3.. .
7 .

gm.

“HIV 3

.N. .91 ‘fi

i‘t‘fléufl
cm?

13‘
g

I

"Haw...

,1} I
inn“ Mr

:4.
2

r1}?
"I'M-:3
R3 vii:

x

,.r

: Il.’ :32.

‘ly.~\..t\: . .1.

‘1

r6

.2”

”WIS

 

 

 

THESIS
“MET

,"\.\_’J

This is to certify that the
dissertation entitled

THREE ESSAYS IN APPLED ECONOMETRICS WITH
APPLICATIONS TO INTERNATIONAL TRADE AND
FINANCE

 

presented by

 

PATRICE WHITELY

LmRARY
Michigan State
University

[_

 

has been accepted towards fulfillment
of the requirements for the

Doctoral degree in Economics

 

 

 

@4115, “'1 mimic/g,

Major Professor's Signature

 

Adj-Vt; T ’lOfifloo 7

Date

MSU is an afiinnatr‘ve-action, equal-opportunity employer

.__._.-._.--.—--.-u-o-a-.-.-o-.-.—.-p---.—

o-v-u_--v-o-o---.—.-o—----—-.--v--o—---p-p-.-o—-—.—.—-—~———

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE

DATE DUE

DATE DUE

 

 

mm 1331“
J

E
LT

“

 

 

 

 

 

 

 

 

 

 

 

 

6107 p;/ClRC/DateDue.indd-p.1

 

TIREE ESSAYS IN APPLIED ECONOMETRICS WITH
APPLICATIONS TO INTERNATIONAL TRADE AND FINANCE

By

Patrice Whitely

A DISSERTATION

Submitted to
Michigan State University
in partial fulfillment of the requirements
For the degree of

DOCTOR OF PHILOSOPHY
Department of Economics

2007

ABSTRACT
THREE ESSAYS IN APPLIED ECONOMETRICS WITH
APPLICATIONS To INTERNATIONAL TRADE AND FINANCE
By

Patrice Whitely

My dissertation consists of three essays in applied Econometrics with applications
in the fields of Trade and International Finance. Chapters 1 and 2 investigate the
relationship between openness and growth. This is an area which has received a great
deal of attention from economists. The usual approach to investigating this relationship
involves the specification of a structural model for gross domestic product (GDP) per
capita with some measure of openness and other variables which are assumed to be
exogenous. The weakness of this approach is the fact that openness may be endogenous.
This may be as a result of simultaneity. Therefore the estimated coefficient of openness
in such a model would be invalid. As such the goal of my paper is to identify the
direction of Granger causality between openness and growth. Granger causality is a
specific, testable definition of causality which does not involve the specification of a
structural model. It is a statement about forecastibility and predictability. The concept of
Granger causality is usually applied to time series data. However chapter 1 shows how
this concept can be extended for use with panel data I find that the direction of Granger
causality for low and high income countries is from openness to growth.

Chapter 2 uses the popular ‘Barro’ regression to investigate the relationship

between openness and growth. This involves running regressions of the growth rate of

income over a period of time on initial income as well as other control variables. This
approach is used in order to address the issue of simultaneity between growth and
openness. Openness at the beginning of the period cannot be caused income growth over
the entire period. The ‘Barro’ methodblogy also allows me to measure the impact of
openness on a country’s growth rate as opposed to investigating the impact of openness
on a country’s GDP per capita as is usually done. I find that initial income, investment
rates and initial openness all have an important impact on the rate at which a country
grows.

Chapter 3 determines whether or not the popular monetary model of exchange
rate determination describes exchange rate movements in Latin American countries. In
the monetary model of exchange rate determination, the exchange rate is viewed as the
value of one country’s money supply against another. Initially there was some support for
this model. However, subsequently these results have been invalidated by the discovery
of the existence of a unit root in the exchange rate sequence. This led to studies trying to
discover whether or not there exists a cointegrating or long run relationship between the
exchange rate and the monetary fundamentals money supply and income. This paper
investigates the applicability of the model to exchange rates in Latin America. Latin
American countries are usually characterized by exchange rate volatility and high
inflation. I find that despite the instability and volatility of exchange rates in these
countries, there is still a long run relationship between the exchange rate and the

fundamentals.

Dedicated to my mother, my sister and the memory of my father

iv

ACKNOWLEDGEMENTS

I would like to thank my advisor Professor Wooldridge for his constant help and
guidance throughout the dissertation writing process. Professor Wooldridge is an
excellent teacher who has not only improved my research skills but who has also inspired
my teaching style. I hope to one day be as gifted as he is in taking complex concepts and
explaining them in a simple, straightforward, systematic manner. I would also like to
thank my other committee members Professors Steven Matusz, Susan Chun Zhu and
Tamer Cavusgil for their help and advice.

I would like to thank Professor Thomas J eitschko for his help, support and
guidance over my time at Michigan State University. Professor Jeitschko, through his
role as Director of the graduate program, gave me the opportunity to independently teach
courses at MSU starting fiom my third year in the program. He also gave me the
opportunity to teach the Math Review Session. His support and confidence in me have
been priceless. I am also grateful to him for helping to make my graduation experience
especially memorable.

I would also like to thank Professor Paul Menchik who served as graduate
Director of graduate studies during my last two years in the program. As a result of his
support and the support of the department I was able to be fully funded during my final
semesters at MSU.

I owe a debt of gratitude to the staff of the Economics department especially
Jennifer Carducci, Stephanie Smith, Margaret Lynch, Linda Wirick and Belen Feight

whose patience in answering all my questions over the years has been greatly

appreciated. Their constant, patient and generous assistance has definitely made the
process a lot less painful than it would otherwise have been.

I would like to thank all the members of GLOE — Jenny, Jiyoung, Peggy,
Deborah, Mary and Nicole for their'advice and support over the years. I would also like
to thank Nathan and Chris for their friendship and support. I would not have made it
through the program without the support of the above mentioned friends. I am truly
grateful.

I would also like to thank Kevin, Valrie, Neesia and Kevin Jr. for their love and
support. They have been my family and support system during my time at MSU. I am
blessed to know people as generous and kind as the Brooks family. They have provided
an excellent example for me of how to be a Christian.

I would like to thank Linda Bailey for all of her advice and support over the years.
I don’t know what I would have done without Linda’s help. It was comforting to have
another J arnaican already in the program when I arrived. Cedric Taylor is yet another
Jamaican whose friendship has helped me through the program. I would like to thank him
for listening to me and for being good company.

I would also like to thank my Canterbury family especially Sean, Kiezha, Nicole,
Mike, Sarah, Stephanie, Emily and Jenni for their support. Canterbury always provided a
warm, safe, inviting place where I could forget about Economics for a while.

I am especially grateful to Canterbury because that is where I met T.P. Words
cannot express how much his love, friendship and support mean to me. He has been a
constant shoulder to lean on and a constant ear to listen to me vent. His laid back,

relaxed, ‘fiee-style’ approach to life has seen me through many a crisis.

vi

I would also like to thank my friends Nicholas and Peter who I met while
pursuing my Masters degree in Jamaica. Nicholas’ perpetual fi‘iendship even when I
became stressed and moody is greatly appreciated. I am thankful to Peter for his faith in
me, his advice both technical and practical and for being there when I needed support and
encouragement.

Lastly and most importantly I would like to thank God for his love, for his gifts,

and for blessing me with the amazing people in my life who have helped me to achieve

this goal.

vii

TABLE OF CONTENTS

LIST OF TABLES _ ix
LIST OF FIGURES xiii
1 What is the Direction of Causality Between Openness and Growth? A

Panel Data Granger Causality Study . 1
1.1 Introduction .................................................................................. 2
1.2 Literature Review .............................................................................. 3
1.3 Granger Causality Testing ................................................................ 11
1.4 Panel Data Granger Causality Testing .................................................. 17
1.5 Description of Data ........................................................................ 21
1 .6 Methodology ............................................................................... 28
1.7 Results ....................................................................................... 29
1.8 Discussion of Results ..................................................................... 37
1 .9 Conclusion ................................................................................. 40
2 What is the Impact of Openness on Growth? A Panel Data Approach 60
2.1 Introduction ................................................................................ 61
2.2 Literature Review ......................................................................... 62
2.3 Methodology ............................................................................... 71
2.4 Description of Data ........................................................................ 77
2.5 Results ....................................................................................... 78
2.6 Discussion and Conclusions ............................................................. 85
3 Testing the Monetary Model of Exchange Rate Determination: Evidence

from Latin America 97
3.] Introduction ................................................................................. 98
3.2 The Monetary Model .................................................................... 100
3 .3 Literature Review ........................................................................ 102
3.4 Methodology 106
3.5 Description of Sample ................................................................... 111
3.6 Results ..................................................................................... 1 15
3.7 Discussion and Conclusions ............................................................ 123
APPENDIX 142

BIBLIOGRAPHY 151

viii

1.1

1.2

1.3

1.4

1.5

1.6

1.7

1.8

1.9

1.10

1.11

1.12

1.13

1.14

1.15

1.16

LIST OF TABLES

Summary Statistics for Income and Trade ............................................. 43
Correlations ................................................................................. 43
Typical Gravity Equation Coefficients from Bergstrand (1985) .................... 44
Estimated Coefficients from F eenstra, Markusen and Rose (2001) ................ 44
Correlations ........................ 44
Classification of Countries According to Income Levels ............................ 45
Fixed Effects Estimation .................................................................. 46
Fixed Effects Estimation with Interactions ............................................. 47
Pooled Granger Causality Test Results ................................................. 47
Granger Causality Test — Unbalanced Panel - one lag ............................... 48

Granger Causality Test - Balanced Panel — one lag - Causality from Openness to
Growth ...................................................................................... 48

Granger Causality Test - Balanced Panel — one lag - Causality from Growth to
Openness .................................................................................... 48
Fixed Effects Estimation — Openness Based on Trade in Differentiated Products
................................................................................................ 49
Fixed Effects Estimation with Interactions - Openness based on Trade in
Differentiated Products ................................................................... 49
Fixed Effects Estimation — Openness based on Trade in Homogeneous Products

................................................................................................ 50

Fixed Effects Estimation with Interactions - Openness based on Trade in
Homogeneous Products .................................................................. 50

ix

1.17

2.1

2.2

2.3

2.4

2.5

2.6

2.7

2.8

2.9

2.10

2.11

2.12

2.13

2.14

2.15

2.16

2.17

3.1

3.2

3.3

3.4

3.5

List of Country Codes and ID’s ...................................................... 57-59

Summary Statistics ........................................................................ 88
Summary Statistics ........................................................................ 88
Correlations .................... ' ............................................................. 89
Estimation Results (T = 4) ............................................................... 89
IV Estimation Results (T = 3) — using instruments for y only ..................... 89
IV Estimation — using instruments for y, I and G .................................... 90
Estimation Results - using different reduced forms for each time period ......... 90

Estimation Results with Initial Values and using instruments for y only .....91

Estimation of equation 1 with interaction term ........................................ 91
IV Estimation using instruments for y only with interaction ........................ 92
IV Estimation using instruments for y, I and G with interaction .................. 92
Estimation of equation 1 using alternative measure of openness ................... 93
Estimation of equation 1 using alternative measure of openness ................... 93
Results using Alternative Openness Measures and instruments for y only ....... 94

Results using Alternative Openness Measures and instruments for y, I & G ....94

Estimation Results — using different reduced forms for each time period ......... 95
Effect on Average Growth Rate of a 10 Percentage point Increase in Openness .95
Coefficient of Variation of Exchange Rates Over Different Time Periods ...... 136

Results of Individual OLS Estimation of the Monetary Model Using Panel 1 ..137

Individual Unit Root Test Results for Exchange Rates ............................. 138
Individual Unit Root Test Results for Relative Money Supply .................... 139
Individual Unit Root Test Results for Relative GDP ................................ 139

3.6

3.7

3.8

3.9

3.10

3.11

Al.l

A12

A13

A1.4

A1.5

A1.6

A1.7

A1.8

A1.9

A1.10

A21

A22

A23

A2.4

A25

A26

A27

Fixed Effects Estimation of the Panel Version of the Monetary Model .......... 140

Panel Engle and Granger Results ...................................................... 140
Individual Cointegration Test Results ................................................ 140

Panel Cointegration Test Results ...................................................... 141

Error Correction Coefficient Estimates for Panel 1 Countries ..................... 141

Table 3.11 — Error Correction Coefficient Estimates ............................... 141

Unbalanced Panel — Causality fiom Trade to Growth .............................. 143
Unbalanced Panel — Causality fi'om Growth to Trade .............................. 143
Balanced Panel - Three Lags ........................................................... 143
Balanced Panel -— Two Lags ............................................................ 143
Unbalanced Panel — Causality From Trade to Growth .............................. 144
Unbalanced Panel — Causality From Growth to Trade .............................. 144
Balanced Panel — Three Lags ........................................................... 144
Balanced Panel — Two Lags ............................................................ 144
Balanced Panel — one lag - Causality from Openness to Growth .................. 145
Balanced Panel — one lag - Causality from Growth to Openness .................. 145
Estimation Results with Initial Values and using instruments for y only ........ 146
Estimation Results with Initial Values and using instruments for y only ........ 146
Estimation of equation 1 using alternative measure of openness .................. 147
Estimation of equation 1 using alternative measure of openness .................. 148

Results using Alternative Openness Measures and instruments for y only 148
Results using Alternative Openness Measures and instruments for y, I & G ..149

Estimation Results — using different reduced forms for each time period ....... 149

xi

A2.8 Estimation Results with Initial Values and using instruments for y only ........ 150

A29 Estimation Results with Initial Values and using instruments for y only ........ 150

xii

1.1

1.2

1.3

1.4

1.5

1.6

1.7

1.8

1.9

1.10

1.12

2.1

2.2

3.1

3.2

3.3

3.4

3.5

3.6

3.7

LIST OF FIGURES

Relationship Between Income and Trade in 1980 .................................... 51
Relationship Between Income and Trade in 1996 .................................... 51
Trade Percentage in 1980 ................................................................. 52
Trade Percentage in 1996 ................................................................. 52
Scatter Diagram of Values Used in Granger Causality Tests ........................ 53
Scatter Diagram of Values Used in Granger Causality Tests ........................ 53
Long Run Impact of Trade on Growth ................................................. 54
Impact of Growth on Trade ............................................................... 54

Long Run Impact of Openness on Growth — Trade in Differentiated Products ...55
Impact of Growth on Openness - Trade in Differentiated Products ............... 55

Long Run Impact of Openness on Growth — Trade in Homogeneous Products ...56

Impact of Growth on Openness — Trade in Homogeneous Products ................. 56
Relationship Between Initial Trade and Growth ...................................... 96
Relationship Between Initial Income and Growth .................................... 96
The Exchange Rate in Argentina ...................................................... 128
The Exchange Rate in Bolivia ......................................................... 128
The Exchange Rate in Brazil ........................................................... 129
The Exchange Rate in Chile ............................................................ 129
The Exchange Rate in Colombia ...................................................... 130
The Exchange Rate in Costa Rica ..................................................... 130
The ExChange Rate in Ecuador ........................................................ 131

xiii

3.8

3.9

3.10

3.11

3.12

3.13

3.14

3.15

The Exchange Rate in Guyana ......................................................... 131

The Exchange Rate in Jamaica ......................................................... 132
The Exchange Rate in Mexico ......................................................... 132
The Exchange Rate in Nicaragua ...................................................... 133
The Exchange Rate in Peru ............................................................. 133
The Exchange Rate in Trinidad & Tobago ........................................... 134
The Exchange Rate in Uruguay ........................................................ 134
The Exchange Rate in Venezuela ...................................................... 135

xiv

Chapter 1

What is the direction of
causality between openness
and growth? A panel data

Granger causality Study

1.1 Introduction

This paper investigates the relationship between openness and growth. What is the
direction of causality between openness and growth? The theory of comparative
advantage tells us that countries benefit by Specializing in and exporting the good that
they can produce at a lower relative cost. However, in the real world does openness cause
an increase in output? Does trade cause an increase in income? Or is it true that countries
with higher incomes trade more? Does openness indicate future growth or does growth
indicate future openness or are both of these true?

Statistics Show that there is a correlation between trade and growth. According to
a report1 from the World Bank, in the last two decades, twenty-four less developed
countries, (including China, India and Mexico), have become more integrated into the
global economy. During that time period, these countries have doubled their ratio of trade
to national income and in the 19903 their per capita Gross Domestic Product increased by
an annual average of 5%. In less integrated countries, the ratio of trade to output has
decreased, and the number of people below the poverty line has increased. Does this
correlation between trade and growth imply causality? This is a complex question which
economists have used various methods to try to answer.

The traditional approach to investigating the relationship between trade and
growth involves the use of static structural models. This paper takes a different approach
by applying the concept of Granger causality. I use this approach because the goal of this
paper is to identify the direction of causality as defined in Granger (1969) between

openness and growth and not to measure the impact of openness on growth. As I will

 

' “Globalization, Growth and Poverty”, World Bank policy research report. December 2001

discuss in the literature review, this approach has been used in various fields of
economics in order to address the issue of direction of causality. Granger causality tests
are usually conducted using time series data. However, recent developments in
econometric techniques now allow Granger causality tests to be applied to panel data
sets. This study makes use of these techniques in order to identify the direction of
Granger causality between openness and growth.

The paper is organized as follows: section 2 gives a review of the relevant
literature. Section 3 defines the concept of Granger causality in its traditional time series
setting. Section 4 discusses the extension of the concept to panel data. Section 5 describes
the sample used. Section 6 outlines the methodology which was followed. Section 7
presents the results, section 8 discusses the results and section 9 summarizes the main

findings of the paper in the form of a conclusion.

1.2 Literature Review

Does trade cause growth? Grossman and Helpman (1990) use an endogenous
growth model to put forward the answer “it depends.” It depends on whether or not trade
forces the economy to direct resources to activities that generate growth (such as research
and development, increasing product quality, and so on) or if it in fact diverts the
economy from such activities.

In the traditional neoclassical growth model, the rate of grth of income per
capita is unexplained. That is, growth is assumed to be exogenous to the model. As the
name implies, the endogenous growth model endogenizes or explains growth. Growth is

treated as a variable to be determined. Investment in physical and human capital has

spillover effects which allow capital to exhibit non-decreasing returns to scale. Therefore
sustained growth becomes a possibility.

Grossman and Helpman (1990) consider a two-sector, two-factor economy. They
assume that the two factors, land and labour are available in fixed supplies. Output in

each sector 1' is described by the following constant returns to scale production function:
X‘=KF"(T",L") i=1,2 (2)
where

Ti = amount of land used in sector 1'

L' = amount of labour used in sector 1'
K = stock of knowledge capital (a public input)

They also assume that:

k = bX1 (3)

where

K = growth rate of K
Using the Rybczynski Theorem they argue that an increase in the supply of the factor
used intensively in sector 1 the knowledge generating sector will lead to an increase in
growth while an increase in the supply of the factor used intensively in sector 2 will
decrease growth. They then go on to discuss how trade and trade policy can change the
relative supplies of the factors. Grossman and Helpman therefore conclude that the
potential for less developed countries to benefit from international trade does exist but
such trade gains are not automatic. They depend on whether or not trade leads to

knowledge and technology transfers.

In this paper Grossman and Helpman also briefly mention product cycles or the
‘North-South’ theory of product development. This theory was first introduced in Vernon
(1966) and was formally modeled in Krugrnan (1979). Grossman and Helpman (1991a)
builds on earlier work to construct a product cycle model which features endogenous
innovation and endogenous technology transfer. The product cycle theory states that
product innovation takes place in the North while imitation takes place in the South. That
is, the invention and initial manufacturing of a product occurs in the North. Over time, as
the production methods become more standardized, they are copied by firms in the South.
It is argued that innovation occurs in the North because these countries possess the
research and development resources and technology necessary for the introduction of
new products. The existence of lower wages in the South means that, after the technology
has been successfully copied, the bulk of manufacturing and production takes place there.
Of course as the South successfully copies the goods invented in the North and is able to
produce and sell them at a lower price, this forces the North to introduce new products as
well as improve the quality of existing products. Examples of goods that have followed
this cycle include personal computers and consumer electronics.

Both Grossman and Helpman (1990) and (1991 a) show that the potential exists
for countries to benefit from trade. However the authors use the North-South theory of
product development to suggest that less developed countries potentially stand to benefit
the most from international trade. This is because trade allows them to have access to the
technologies invented in the North. This implies that trade may have a greater impact on

growth in low and middle income countries than it does in high income countries.

Grossman and Helpman (1991b), Feenstra (1990) and Matsuyama (1992), among
others, have shown how trade can lead to a reduction in the long run growth rate of
developing countries. They do so by using formal models of the ‘infant industry’
argument. These models illustrate that once real world market imperfections are taken
into consideration, there is no reason to believe that trade will necessarily lead to growth.

Many empirical studies have been undertaken in order to quantify the impact of
trade on growth. These studies usually proceed by running a cross-sectional regression of
per capita income on some measure of trade or openness (such as the ratio of exports
and/or imports to gross domestic product (GDP)), and other variables. These regressions
usually identify a weakly positive relationship.

The main weakness of studies of this type is the fact that trade may be
endogenous. As Elhanan Helpman (1988), Colin Bradford, Jr. and Naomi Chakwin
(1993), Rodrik (1995 a), and several others point out, countries with high incomes (for
reasons other than trade) may trade more. Therefore a positive coefficient on the trade
variable may be reflecting the fact that growth causes trade and not necessarily the other
way around. Furthermore, if there are endogeneity issues, the least squares estimator is
biased and inconsistent. Frankel and Romer (1999) address the endogeneity issue by
using an instrumental variables (IV) approach. A valid instrument should be uncorrelated
with the unobservables that affect growth but correlated with the trade share. As such,
Frankel and Romer use countries’ geographic characteristics as instruments for trade.
They argue that “it is difficult to think of reasons that a country’s geographic
characteristics could have important effects on its income except thrbugh their impact on

trade.” Frankel and Romer estimate the model:

InYi=a+b7}-+cllnN,-+czlnA,-+/ri (1)

where

i = country

N = population
A = area

T = trade share
Y = income

Using geographic characteristics as instruments for 7} they conclude that trade

increases income. They believe that by using IV they have uncovered a causal effect of
trade on income. Their paper suggests that there is no reason to believe that the positive
correlation between trade and income comes about because countries whose incomes are
high for other reasons, trade more. Therefore, Frankel and Romer would lead one to
conclude that trade causes growth.

It has also been argued that another source of endogeneity is omitted variables. If
the relationship between trade and growth is driven by the existence of omitted variables,
then trade may be an endogenous variable. For example, countries with good institutions
may grow faster. However, it could be that one of the reasons that this growth occurs is
because of the impact of these institutions on trade. This would imply that one needs to
isolate the effect of quality of institutions on growth by including it as an independent
variable. Rodrik, Subramanian and Trebbi (2002) attempt to estimate the partial effects of
trade and institutions using the same instrument for trade advocated in Frankel and
Romer (1999). They conclude that the effect of institutions on growth is robust to the

inclusion of trade as an explanatory variable, but the effect of trade on growth is not

robust to the inclusion of institutions. This casts a new light on the findings of Frankel
and Romer. It may be that the results of Frankel and Romer suffer from omitted variable
bias. Rodrik et al. find that trade has a positive impact on institutions. It is therefore
possible that the positive effect of trade on growth found by Frankel and Romer, is
actually reflecting the impact of institutions on growth. Trade may have an indirect
impact on growth through its impact on institutions. Rodrik et al (2002) highlights the
unresolved nature of the trade and growth debate. It also illuminates the difficulty in
specifying a model with truly exogenous explanatory variables. As is well known, if the
explanatory variables in a model are not exogenous, the coefficient estimates are invalid.
This is one reason why some papers have addressed the trade-growth debate by using the
concept of Granger causality. The rationale behind using this approach will be discussed
in detail in the next section.

As the title suggests, the goal of this paper is to identify the direction of
‘causality’ between openness and growth. Granger causality has been used to address the
issue of direction of causality between several variables — money and output [for
example, Hayo ( 1999)] United States (U .S.) and foreign equity market yields [for
example, Cochran and Mansur (1991)] international trade and political
conflict/cooperation [for example, Reuveny and Kang (1996)], among others. It is a
concept that has been applied in fields from Macroeconomics to Finance to Trade.
Several studies, for example, — Ahmad and Hamhirun (1996), Leichenko (2000), and Tao
and Zestos (2002), use Granger causality tests to explore the relationship between trade

and growth.

Tao and Zestos use annual time series data from 1948-1996 to analyze the causal
relations between trade and GDP growth in the US. and Canada. They use the Granger
causality test introduced in Engle and Granger (1987). This causality test requires
cointegration. Inforrnally, two variables are cointegrated if they move together over a
long period of time. If two variables, X and Y, are cointegrated, causality from X to Y
can be established not only from the joint significance of the coefficients of the lagged
values of the X variable but also from the joint significance of the coefficient of the one-
period lagged error term of the cointegrating equation of the two variables. As such, Tao
and Zestos estimate the following vector error correction (VEC) model:

r s

ALGDP, = a1 +aLGDPu,_1 +2.21,- ALGDPH- + th- ALexport,_,- +

i=1 i=1
k

2 CU A Limport,_,- + "It (2)

i=1

r S
A L exp art, = b2 + bL exportut—I + Z 02,- A LGDP,_,- + 2 b2,- A L exp ortt_,- +

i—l i=1

k
262,‘ ALimport,_,- +v2, (3)

i=1
r S
A Limport, = C3 + cLimportut—l + 2 03,- A LGDP,_,- + 2 b3,- A L exp 0rt,_,- +
i=1 I=I
k

2%. A Limport,_,- + V3, (4)

i=1
where:

A LGDP = growth rate of GDP

A L exp ort = growth rate of exports
A Limport = growth rate of imports
u,_1 = one-period lagged error term of the cointegrating vector

Tao and Zestos conclude that for Canada causality exists in every possible
direction. They argue that export growth generated the necessary foreign exchange to pay
for imported goods, which led to domestic economic growth. For the US. causality was
found from exports to GDP. They argue that since the US. dollar is widely accepted for
international payments, there was no causality from imports to exports.

Ahmad and Hamhirun (1996) explore the causal relationship between exports and
economic growth for the countries of the Association of South East Asian Nations
(ASEAN) — Indonesia, Malaysia, the Philippines, Singapore and Thailand. They use
annual data on exports and Gross National Product fi'om 1966-1988. They specify a VEC
model but since the variables in their study are not cointegrated, the error correction term
is excluded from their causality tests. If the error correction term is excluded the Granger
causality test becomes the standard Granger (1969) test. Ahmad and Hamhirun find that
exports do not Granger cause GDP growth in any of the countries. However, they find
Granger causality from economic growth to exports in all five countries.

Leichenko (2000) uses the standard Granger (1969) test to investigate causality
between U. S. foreign exports and economic growth at the regional level. He finds
general support for bidirectional causality between exports and economic growth but
notes that there is variation among regions.

Rodriquez and Rodrik (1999) argue that the measure of “openness” used in many

empirical studies does not accurately reflect the existence or lack thereof of trade barriers.

10

The authors conduct a review of four papers — Dollar (1992), Ben-David (1993), Sachs
and Warner (1995) and Edwards (1998). Dollar (1992) uses two indices of ‘outward
orientation’ — exchange rate distortion and real exchange rate variability. Sachs and
Warner (1995) use a binary variable for their openness indicator. This variable takes on
the value of 0 if the economy is closed according to any of five criteria and takes on the
value of 1 otherwise. Rodriquez and Rodrik show that these methods do not accurately
measure openness. They also argue that there is no evidence that these more complex
measures improve on a simple measure such as trade-weighted tariff averages. They posit
that this is a more direct indicator of trade restrictions. They also suggest that it might be
more fruitful to look for more contingent relationships between trade policy and growth.
That is, they propose asking if trade restrictions operate differently in low versus high
income countries. If the implications of product cycle theories are correct then openness
should have a larger impact on growth in developing countries than it does in developed

countries.

1.3 Granger Causality Mg
(i) Discussion of Approach

What is causality? This paper uses an explicit, testable definition of causality that
is introduced in Granger (1969). This is different from the standard meaning of causality.
The usual way of testing for the standard cause and effect relationship, and measuring the
impact of one variable on another involves the specification of a structural model.
However, in order for the estimates from a structural model to be valid (that is, unbiased

or consistent), the explanatory variables in the model must be exogenous. As discussed in

11

the literature review, one of the main difficulties in investigating the relationship between
openness and growth is the issue of endogeneity. This may arise either because of the
simultaneous bias inherent in the relationship between the two or because of omitted
variable bias.

When one fits a regression line, the existence of a relationship is assumed on the
basis of economic theory. The objective is to measure the ceteris paribus impact of one
variable on another. However this involves introducing the untestable concept of
exogeneity. As was discussed in the literature review, specifying a structural model for
income per capita makes the existence of exogeneity of the explanatory variables highly
questionable. It is therefore not possible to conduct a ‘controlled experiment’ of the
impact of trade on growth.

Granger (1969) introduces an alternative, explicit and testable definition of the
word causality. Some authors including Leichenko (2000) and Ahmad and Hamhirun
(1996) define causality as Granger causality. However, as Harvey (1981) points out,
Granger’s definition of causality is purely statistical and does not correspond to the cause
and effect definition of causality in the philosophical sense. Granger’s definition is
limited in that it is a statement about predictability and forecastibility. However, it’s
strength is in the fact that it is explicitly and easily testable.

It is important to note that Granger causality is a weaker condition than the

condition for exogeneity. Enders (1995) shows that it is possible for a series yr to not
Granger cause the series x, , and yet x, is not exogenous to y, . This leads to the question,

when does Granger causality imply standard causality? The answer is, when we have

exogeneity. Hamilton (1994) argues that if {(x,, y,) : t = 1,2,...} is a bivariate time

12

series: if x, Granger causes y, and if nothing else Granger causes x, , then it is possible
to conclude that x, causes y,. As stated above, exogeneity is a stronger condition than
Granger causality so if x, Granger causes y, , and if x, is exogenous then x, causes yr.
Similarly Enders (1995) shows that if I x, does not Granger cause y, and y, is exogenous
then x, has no impact on y,. Therefore, in the final analysis, to answer the question of

standard causality we will always need exogeneity. It is important to emphasize that
Granger causality is not another way of answering the question of the existence of
standard causality. It is asking a different question. It proposes a different definition of

the concept of causality. It asks the question, ‘does the forecast of y, improve after

taking account of past values of x, , while controlling for past values of y, ?’ It is also able

to give a definitive answer to the question it asks. AS such that is the definition of
causality utilized in this paper.

In this paper, I use a rich panel data set in order to see what allong variation,
both across countries and over time, will illuminate about the relationship between
openness and growth. Specifying a structural model for income per capita would require
the use of instruments for both trade and institutions. Finding suitable instruments is
always a challenge. The currently agreed upon instruments used in Rodrik, Subramanian
and Trebbi (2002), for both trade and institutions, do not vary over time. Therefore it
would be impossible to estimate their coefficients using fixed effects estimation. It
would be difficult to find appropriate time-varying instruments that are available for all
the countries in the sample, over all the years included in the sample. Therefore, as was

previously discussed, the exogeneity of the explanatory variables would be highly

13

questionable. Looking at Granger causality as opposed to standard causality overcomes
these issues. It will be interesting to see what this method reveals about the relationship
between trade and growth. Can past values of openness predict future values of growth,
after controlling for past values of growth or do past values of growth serve as an
indicator of future values of openness after controlling for past values of openness? Or
are both of these true?

This paper extends Tao and Zestos (2002) and other papers that have used
Granger causality to identify the direction of causality between trade and growth by using
a rich panel data set with more recent data. The fact that using panel data increases
estimation efficiency is well known. The use of panel data also overcomes the problem of
lack of sufficient data for some countries. The version of the Granger causality test used
by Tao and Zestos can only be used if the variables are cointegrated and is therefore
somewhat restrictive. I use the standard F -test version of the Granger causality test
extended for use with panel data. This is described in more detail in the following
section.

In this paper, I also address the critique of Rodriquez and Rodrik (1999). I do so
by using more direct measures of openness, which will be described in more detail in
section 5. Rodriquez and Rodrik suggest using trade-weighted tariff averages. However
due to the existence of non-tariff trade barriers, I think that trade percentage would be a
more reliable, direct measure of openness. I also address the critique of Rodriquez and
Rodrik by dividing the sample into groups according to income levels to see if the
relationship between openness and growth is different in low versus middle versus high

income countries.

14

The goal of this paper is not to measure the impact of openness on growth. The
goal is to try to discern whether observed correlations between openness and growth are
driven by the fact that openness Granger causes growth or by the fact that growth
Granger causes openness or both. Looking at whether or not past values of openness help
to predict future values of growth or vice versa will help in understanding the relationship

between openness and growth.

(ii) Definition of Granger Causality
If {( x,, y,) : t = 1,2,. . .} is a bivariate time series, then x, Granger causes y, if
E(yt I Yr—Iaxt—IsJ’t—zaxt—za-«) ¢ 50’: I J’r—i»J’r—2a---) (5)
In other words, if, after we control for past information of y , past values of x help to
predict y , then we say that x, Granger causes y, 2.. As the timing of the information sets

implies, Granger causality is not a statement about contemporaneous causality. Its
strength lies in the fact that it is fairly easy to test without additional assumptions.
Usually, a linear model is put forward for the more general

expectation, E(y, | y,_1,x,_1, y,_2,x,_2,...) , and then a joint test of significance is
carried out for the lags of x. With p lags of y and q lags of x , we have
EO’t Iy,-1,xt_1,y,_2,x¢_2,...)=
a+yly,_1 +....+}'py,_p +fl1x,_1+...flqxt_q (6)
Assuming that {(x,, y,) : t = 1,2,...} is a weakly dependent series - that is, satisfies the

central limit theorem (and therefore the law of large numbers), we can use a Wald test of

 

2 Causality from y to x is defined by reversing the two variables in the previous definition

15

H0 :fll = 0,..., A, = 0 . The standard F test (which is just a rescaled Wald statistic) or the
usual Lagrange multiplier test, maintain homoskedasticity under the null:

Var(y, |y,_1,xt_1,y,_2,xt_2,...) = 0'2,t = 1,2,... (7)

This, rules out structural change in the variance as well as autoregressive forms of
heteroskedasticity such as autoregressive conditional heteroskedasticity (ARCH) and
generalized ARCH (GARCH). The ARCH model was introduced by Engle (1982) while
GARCH was introduced in Bollerslev (1986). In the ARCH model the conditional
variance of the error term, rather than being constant, changes over time as an
autoregressive function of the squared residuals. In the GARCH model the conditional
variance of the error term has both autoregressive and moving average components. That
is, the variance is a function of both the past squared residuals and past variances.
Heteroskedasticity-robust versions of the F statistic are easy to obtain.

Sometimes it is usefirl to write the conditional expectation in (6) in error form:

y, = a + 71y,_1+...+ ypyt_p + fllx,_1+...flqx,_q + u, (8)
where, necessarily, the error term satisfies

E(ut I yt-l vxt—lsyt—2axt—29m) = 0 (9)
The fact that past errors are a function of the conditioning set in (9) implies that the errors

are appropriately serially uncorrelated. Therefore Granger causality tests do not have to

be made robust to serial correlation.

16

1.4 Panel Data Granger Causality Testing

Consider the following model for x, and y, :
K K
k k
yit = “i + 2 7i )J’it-k + Z .5,-( )xit—k +812 (10)
k=l k=l
for each individual 1': 1,...,N and t =1,....,T with K CN‘ and a,- =(/3§1),...,6§"))'

With the usual assumptions, (mentioned above), it is possible to conduct a test for
Granger causality by estimating equation (10) pooled across 1' and t. The null hypothesis

would be:
H0 :flU‘) = 0 Vk =1...K

While the alternative hypothesis would be:
H1:,B(k) ¢ 0 Vk =1...K

In this case the standard F test is asymptotically valid. However, pooling in this manner
does not allow heterogeneity acrossi . It assumes that the lagged coefficients are the same
for each individual in the sample. It is possible to conduct panel Granger causality testing

that overcomes this limitation. We start by making the following assumptions:
A (1) {(xi,,y,~,) :t =1,2...} are weakly dependent for all i.
(2) The errors are homoskedastic. That implies:
Var(y,-t I xi,_1 , yit_1,...) = 0'2 (a constant)
B {(xi1,...x,-T,yi1....y,~T)} are independent across i .

It is then possible to test a Homogenous Non-Causality (I-IN C) Hypothesis where

the null hypothesis is:

17

H0 :fli =0 Vi=1...N
While the alternative hypothesis is:

H1 :,6,- =0 Vi=l....N1

,Bi at 0 Vi =N1 +1,N1 +2,...N

If the HNC hypothesis is not rejected, we can conclude that there is no causality
for all individuals of the sample. If N1 = 0 , then 161' at 0 for all i and there is causality for
all i. If 0 < N1< N, then there are N1 individuals with no causality from x to y and
N - N1 individuals with causality from x to y. If N1 = N , we are back to the case of

failing to reject the null hypothesis.

Consider the test statistic:

N

i =1
where
WiT= individual Wald statistic for H0 : .Bi = 0.
Under the null hypothesis WiT converges asymptotically to a chi-squared distribution,
with K degrees of freedom, as T —> 00, Vi = 1...N with:
E (WIT) = K and WWm = 2K
Therefore by the Lindberg-Levy central limit theorem, under the null hypothesis:

ZHNC =JN/2K *(WHNC—K) (12)

converges asymptotically to a standard normal distribution as T -> 00 first and then

N—>oo.

l8

The test involves taking the Wald statistic for individual Granger causality tests
for individual groups and averaging them to form a statistic that is appropriate for panel

data Granger causality testing. As such, we need two sets of assumptions. We need
assumptions that guarantee that each WiT converges asymptotically to a chi-squared
distribution with K degrees of freedom as T—> 00. We also need an assumption that allows

us to apply the central limit theorem to the averaged statistic in order to guarantee that it

converges asymptotically to the standard normal distribution as N —> 00.

The assumptions in group A imply that under H0, WiT converges asymptotically

to a chi-squared distribution with K degrees of freedom as T —> 00. Adding assumption B
implies that under H0, Z converges asymptotically to a standard normal distribution as T
—> oo firstandthenN —> oo.

Hurlin (2004) argues that this approach can be extended to the case of fixed T by
making adjustments to the first and second moments of WIT . He does this by assuming
normality in order to apply the Magnus (1986) theorem. However, the Magnus theorem
requires strict exogeneity of the regressors, which clearly cannot hold in the model
presented in equation (10). Nevertheless, in his paper, Hurlin conducts Monte Carlo
experiments generating the model presented in equation (10) and computing the mean
and variance of the Wald statistic. He compares these with the calculated moments based
on the Magnus theorem and shows that the differences are negligible. As such I will
make use of the statistic he proposes in his paper, which is outlined below.

For the case of fixed T, Hurlin asserts that by the Magnus Theorem, the second

order moments of WIT exist if and only if:

T>5+2K

19

He then uses the Lyapunov central limit theorem to posit that:

— N

2F _
— N
JN ‘Z,=,Var(W.~r)

 

(13)

converges asymptotically to a standard normal distribution as N ——> oo.
Applying the Magnus theorem he gets:

*(T—ZK—l)

E(WI'T)= K (T—2K—3)

 

(T—2K—I)2*(T—K—3)
(T—2K—3)2*(T—2K—5)

 

Var(W,-T) = 2K *

ifandonlyifT>5+2K
Using these values for E(W,-T)and Var(W,-T) Hurlin puts forward the following

approximated standardized statistic:

 

 

 

ZFHNC= [N*(T—2K-5) * [(T’ZK-3)WHIVC_K] (14)
V(2*K)*(T—K—3) (T-ZK—l)

which converges asymptotically to a standard normal distribution as N —> 00.

In the case of both N and T fixed Hurlin suggests that one can use the

approximated standardized statistic zFHNC and compute an approximation of the

appropriate critical values [cN1(a)] for a fixed N. Based on this he gets:

 

emu) = Za JN‘lr/ar(W,-T) +E(W,-T) (15)

20

1.5 Descgg' tion at Data

(i) Initial Measures

The sample is an unbalanced panel data set of 123 countries with a time span of
20-51 years (depending on data availability). I use annual data on openness and PPP
adjusted constant gross domestic product (GDP) per capita obtained from the Penn World
Tables. The measure of openness used by the Penn World Tables is total trade as a
percentage of GDP. This is a standard measure of openness and is also used by
organizations such as the International Monetary Fund. However in the next section, I
give a detailed description of an alternative method of openness which was also used.
Logs were taken of both series before unit root tests were done. Taking logs allows me to
calculate the elasticity of trade with respect to openness and vice versa, in preliminary
regressions.

When dividing the sample into groups according to income level, I use a balanced
panel of 107 countries with 20 years of data fiom 1981-2000. Table 1.1 gives summary
statistics for the percentage trade and GDP per capita variables in levels. All values are in
1996 dollars. As expected, the average income per capita increases steadily from the low
income group to the high income group in both 1981 and 2000. In 2000, the average
income for the low income group is $1,482.54, for the lower middle income group is
$5,195.70, for the upper middle income group is $10,494.13, and for the high income
group is $25, 083.01.

In 1981, the high income countries have the lowest average trade percentage,
while over the low income to the upper middle income range, the average trade

percentage increases. In 2000, there is a positive relationship between average trade

21

percentage and income group, as the low income group has an average trade percentage
of 84.2, the lower middle income group has an average trade percentage of 87.2, the
upper middle income group has an average trade percentage of 87.9, and the high income
group has an average trade percentage of 101.1. There is therefore some evidence that
larger countries trade more.

Figure 1.1 is a diagram showing the relationship between income and trade in
1980 (the earliest year for which data is available for all 123 countries). It shows a
slightly positive relationship between trade and income. Figure 1.2 illustrates that this
positive relationship increases in 1996 (the latest year for which data was available for all
123 countries). This weakly positive relationship is reflected in the correlation between
trade and income for all countries, which as table 1.2 illustrates, is 6%.

Figures 1.1 and 1.2 indicate that some of the trade percentage values are greater
than 300. Figures 1.3 and 1.4 take a closer look at the values for trade percentage for each
country in 1980 and 1996, respectively. As figure 1.3 shows, in 1980, one country has a
trade percentage greater than 300. This country is Romania, with a trade percentage of
325. Four countries, Romania, Sao Tome and Principe, Singapore, and Luxembourg,
have trade percentages greater than 200. As figure 1.4 illustrates, in 1996 Singapore is the
only country with a trade percentage greater than 300. While Hong Kong, Equatorial
Guinea, Guyana, and Luxembourg are the other countries with trade percentages greater
than 200. Hong Kong, Singapore and Luxembourg are all classified as high income
countries.

When the panel is divided into groups according to income levels, the correlation

is the strongest in the high income group — 29%. The second largest correlation, 17%, is

22

among the upper middle income countries. The low income countries have a correlation
of 15%, while the lower middle income countries show the smallest degree of correlation
between trade and income. The correlation for the middle income countries as a group is
6%.

As table 1.2 illustrates, there is also a weakly positive relationship between trade
and growth ((incomet — incomet_1)/ income“). This correlation is strongest among the

high income countries — 11.4%, and weakest among the low income countries — 0.46%.
So there is also evidence that the relationship between trade and growth may differ across

income groups.

(ii) Alternative Measure of Openness

The initial measure of openness used may not accurately reflect how open to trade
an economy’s borders are. As such, I also use the difference between the volume of trade
predicted by the gravity model and the volume of trade that actually occurred, to
determine how ‘open’ a country’s economy is. The gravity model relates bilateral trade
between countries to GDP, distance, and other factors that might affect trade barriers. The
theory states that after controlling for size, trade between two regions is inversely related
to their bilateral trade barrier relative to the average barrier of the two regions to trade
with all their partners.

I use the following equation for the gravity model:

lnXij =k0+kllnYi+k21an -k3IITDy'+k4CON7}j +k5LANGlj-I-k6FTAU +141]

(16)

where:

23

X ,1 = value of exports from country i to country j

Y,- = real GDP of country i

Dij = distance between i and j

CONT,-j = dummy variable which is 1 if i and j are contiguous

LANGij = dummy variable which is 1 if i and j share a common language

FTAij3 = dummy variable which is 1 if iand j share a common fi'ee trade

agreement

This implies:

Yik Y;-
X,- =

 

exp(k0 + k4C0NT,-- + k5LANG,-I + k6FTAy-) (17)
U

We know that:

Tij ’total trade between i and j= X,~J-+Xj,-

 

Therefore:
YI" Y" + ij' Y."-
Tij_ — Dk; CXpUtO + k4CONTy- + ksLANGiJ + k6FTAl-j) (18)

y
T,- = total trade by country i = 21.7}!-
Therefore:
kl k2 klk 2
_ "’7 Yi Yr ”'1 1;.

TI-Z k

j—_-1 Dij

 

exp(k0 +k4C0NT, +k5LANG, +k6FTAiJ-) (19)

3

 

3 A binary variable for common continent was used as a proxy for this variable

24

= iP
= trade predicted by the gravity model for country i

In order to calculate the predicted trade for each country, I therefore need
estimates of k1...k6 , which can be obtained by estimating equation (16). However
estimating this equation requires data on bilateral trade for each country pair in the
sample. I have data on total trade for each country. This data cannot be used to estimate a
gravity model. Therefore I obtain the estimated coefficients by doing a survey of papers
that have estimated gravity equations.

A number of empirical papers exploring the gravity model have been written.
Several of them have focused on the impact of national borders, for example McCallum
(1995) and Anderson and Van Wincoop (2001). Others have focused on using the gravity
model to distinguish among different theoretical models, for example, Evenett and Keller
(1998), Head and Ries (2000) and F eenstra, Markusen and Rose (2001).

The goal of Evenett and Keller (1998) is to use the gravity model to differentiate
between the Heckscher-Ohlin theory and the Increasing Returns trade theory. As such,
the version of the gravity model that they use focuses on the effect of own country GDP
and trading partner GDP. They do not include variables such as common language and
distance which, as shown in Feenstra, Markusen and Rose (2001), have an impact on
trade flows. Furthermore, Evenett and Keller use a data set of 58 countries which consists
of nearly all industrialized countries but relatively few less developed countries. Head
and Ries focus on the US. and Canada I use a balanced panel of 107 countries therefore

I need to use estimates that were obtained using a larger data set.

25

Bergstrand (1985) presents typical gravity equation coefficient estimates for trade
flows. The coefficients for 1965 and 1976 are presented in table 1.3. All variables except
the dummies are in logarithms. The model used does not include common language,
which is shown to be significant in Feenstra, Markusen and Rose (2001).

Therefore, the estimated coefficients used to construct the predicted trade measure
are taken from Feenstra, Markusen and Rose (2001). This paper presents a
comprehensive version of the gravity model including variables such as common
language and distance which are not in the model used in Evenett and Keller (1998).
Feenstra, Markusen and Rose also use a much wider data set — over 110 countries. The
results of F eenstra, Markusen and Rose (2001) are generally consistent with the findings
in Evenett and Keller (1998) and Head and Ries (2000).

F eenstra, Markusen and Rose get different coefficient estimates depending on
whether trade is in differentiated or homogeneous products. I use the estimated
coefficients for 1980 for both cases. These are reported in table 1.4. Since all the
variables are in logs, the estimates are the elasticities of exports with respect to each
independent variable. So, for example, the elasticity of exports of differentiated goods of
a particular country with respect to its trading partner’s GDP is 0.65. The estimated
coefficients are similar to the typical gravity equation coefficient estimates presented in
Bergstrand (1985). For example, the estimate on partner GDP presented in Bergstrand
(1985) is 0.65 in 1965 and 0.69 in 1976.

The measure of openness I use is:

0,, = lnTitA ‘ ln TitP (20)

where 7},p = trade predicted for country i in year t

26

TM = actual trade for country i in year t

For every country in the sample, predicted trade is greater than actual trade. This
means that the expected value of the error term in equation (19) cannot be zero, implying
that equation (19) may not be a very good model for trade. However this problem may
arise because the coefficients used to construct predicted trade were not estimated with
the sample data used in this paper. As was previously mentioned, because of data
limitations the estimated coefficients were obtained fi'om Feenstra, Markusen and Rose
(2001)

The above measure is constructed for 107 countries for the years 1982-2000. As
table 1.5 illustrates the correlation between actual and predicted trade is very low (0.1,
whether I use the estimates for trade in differentiated or homogeneous goods). However
once both measures of openness (trade percentage and the measure based on the gravity
model) are standardized and first differenced to achieve stationarity, the correlation
between the measures of openness actually used in the Granger causality tests is quite
high. When trade in differentiated products is assumed, the correlation is 0.9858. When
the estimated coefficients for trade in homogeneous products are used, the correlation is
0.9807. These high correlations are reflected in figures 1.5 and 1.6 which are scatter

diagrams of the different measures.

L6 Methodology

First of all, a test for weak dependence must be conducted. As such, I use Choi’s

test statistic for panel unit roots. This statistic allows a great deal of flexibility. For

27

example, it allows the number of time series to be different across groups. Choi (2001)

introduces the inverse normal test. The test statistic is:

=‘/—i— Z‘Vl (Pi) (21)

where (D is the standard normal cumulative distribution firnction and

p,- is the p-value for a unit root test statistic for the i "’ country

For all i, pi has a uniform distribution over the interval [0,1]. Therefore Z

converges to a standard normal distribution as T —> 00 first and then N —> 00. As Choi
(2001) shows, we can reject the null hypothesis that all the time series are unit root non-
stationary when the test statistic is less than the critical value of the lower tail of the
standard normal distribution. Choi’s unit root test was conducted on the log of income
and the log of trade. The unit root test statistic used for constructing Z was the augmented
Dickey-Fuller test.

I then conduct a test for Granger causality by estimating equation (10) pooled
across individuals and over time. This equation is estimated by using both openness and
growth alternately as the dependent variables in the model. Next, I allow heterogeneity
across i by testing the HNC hypothesis using both openness and growth as dependent
variables in the model. I use both alternately because I am interested in discovering

whether openness Granger causes growth or growth Granger causes openness or both.
First I use the entire panel and therefore employ both the ZHNC and the ZFHNC statistic.

Then I divide the data set into groups according to income levels. I classify the countries

as high, upper middle, lower middle or low income based on the system used by the

28

World Bank4 for 1987 (the earliest available year). I then test the HNC hypothesis for
each group by computing the ZFHNC statistic and using the appropriate critical values for

the case of fixed N and T.

1.7 Results

(a) Using Trade Percentage to Measure Openness

(i) Unit Root Testing

For the income variable, I am unable to reject the null hypothesis that for every
individual in the sample the series contains a unit root. I am able to reject this null
hypothesis for the percentage trade variable. However, the p-values for the individual
unit root tests are less than 0.05 for 12% of the countries in the sample. Since the
alternative hypothesis in the Choi test is that at least one of the time series is
stationary, the test is conducted using the difference of both variables. This time the
null hypothesis is rejected for both variables and the p-values for the individual unit
root tests are less than 0.05 for all the countries in the sample for both income and
trade. Therefore the HNC hypothesis is tested using the difference in the log of both
variables. Weak dependence is one of the assumptions that is necessary for the test
statistic to converge to a standard normal distribution, and I now have evidence that

the difference of both series is stationary, and therefore weakly dependent.

 

’ For complete list of classifications see table 1.6

29

(ii) Preliminary Rgggessions

Table 1.7 illustrates a fixed effects regression of the differenced income
variable (referred to as growth) on lags of both growth and the differenced trade
variable. The differences are used because as the above section indicates they are
stationary. The lagged impacts of trade on growth are all positive. The elasticity of
growth with respect to trade after one year is 0.074. This decreases to 0.028 after three
years, increases to 0.037 after four years and then falls to 0.026 after 5 years.
Therefore the long run elasticity of growth with respect to trade is 0.21. A number of
specifications with different numbers of lags were estimated. The results for five lags
of the trade variable are shown because all five lags are significant at the 1% level.
This implies that any impact that trade may have on growth is likely to be positive and
that this impact declines over time. The lagged impact of growth on trade is positive.
There is therefore fiuther evidence that larger countries trade more.

The models shown in table 1.7 were estimated using ordinary least squares for
each country in the sample individually. These results are illustrated in figures 1.7 and
1.8. Figure 1.7 illustrates the long run elasticity of growth with respect to trade for
each country in the sample individually. For 94 of the 123 countries in the sample, the
impact is positive. St. Lucia has the largest elasticity (1.169) while the smallest
elasticity (in absolute value) belongs to St. Kitts and Nevis. Both countries are
classified as middle income. The average long run elasticity is 0.167. Of the 29
countries for which the elasticity is negative, 41% of them were lower middle income.

However the elasticity of growth with respect to trade is positive for most (71% of)

30

lower middle income countries, including St. Lucia, the country with the highest
elasticity.

When the elasticity of trade with respect to growth is examined for the
individual countries, it is positive for 72 countries in the sample. As illustrated in
figure 1.8, the country with the largest magnitude of elasticity is Burkina Faso, with an
elasticity of -1.667. The country with the largest positive elasticity is Indonesia with
an elasticity of 1.587. The country with the smallest elasticity (in absolute value) is
Thailand with an elasticity of -0.0018. The average elasticity is 0.07. Once again, of
the four income groups, the lower middle group has the largest percentage (31%) of
the countries for which the elasticity is negative. However, yet again, for the majority
of lower middle income countries (62%) the impact of growth on trade is positive.
Overall, for the majority of countries in all income groups, the relationship between
trade and growth is positive. However, the lower middle income group had the largest
percentage of countries for which there was a negative relationship between growth
and trade.

Table 1.8 shows the results of fixed effects regressions where the impacts of
trade and growth are allowed to vary by income group. The results indicate that there
is variation in these impacts across income groups. For example, the impact of trade
on growth after one year is 0.048% lower for the middle income group than it is for
the low income group. The impaCt of growth on trade after one year is 0.155% lower
for upper middle income countries, than it is for low income countries. This is further

evidence that the relationship between trade and growth may differ across income

groups.

31

(iii) Panel Data Granger Causality Tests
I start by conducting an F test for the null of non-causality on the pooled data. I

estimate the models of table 1.7 using fixed effects. As expected the null hypothesis of
non-causality is rejected both from growth to trade and flour trade to growth. As table
1.9 illustrates, going from trade to growth the test statistic is 33.05 and the p-value is
0. In testing the hypothesis from growth to trade, the statistic is 11.44 and the p-value
is 0.007. Table 1.7 reports the models for trade and growth that fit the best (in terms
of significance). As table 1.7 implies, an AR (1) model is most appropriate for both
trade and growth.

Table 1.10 reports the results of the panel data Granger causality tests that
allow heterogeneity across individuals. Using one lag, the HNC hypothesis can be
rejected from trade to growth as well as fiom growth to trade, at the 5% level. In all
cases the test statistic is greater, in absolute value, than the critical value. The same is
true when the data set is restricted to a balanced panel with T = 20 for 107 countries.
Therefore, it is necessary to attempt to identify a sub-group of countries for which
there is no causality for all i. In other words, I attempt to identify N], In order to do so
the panel of countries is divided into 4 groups according to income levels — low, lower
middle, upper middle and high.

Table 1.11 shows that for all sub-groups, it is possible to reject the hypothesis
from trade to growth. Going from growth to trade the HNC hypothesis cannot be
rejected for the middle income group. This result is confirmed when the middle

income group is divided into upper middle and lower middle income sub-groups. The

32

HNC hypothesis fi'om growth to trade cannot be rejected for both groups. This implies
that for middle income countries past values of growth do not help to predict future
values of trade. This is not surprising when one recalls that the middle income group
has the lowest correlation between income and trade. The non-rejection of the
hypothesis implies that for middle income countries the direction of Granger causality
is from trade to growth.

The tests are repeated dropping assumption A2 and making the statistics
robust to heteroskedasticity. The tests are also repeated for K = 2 and K = 3. These
results are reported in the appendix. An AR( 1) model is most appropriate for both

trade and growth.

(b) Alternative Measure of Openness

1!) Trade in Differentiated Products
(i) Unit Root Testing

The null hypothesis that every individual in the sample contains a unit root is
rejected for the first difference of the openness variable. The p-values for the
individual unit root tests on the first difference are less than 0.05 for 82% of the
sample. Therefore the HNC hypothesis is tested using the first difference of the

variable.

33

(ii) Preliminggy Rmions

Table 1.13 illustrates the fixed effects regression of the growth variable on lags
of both growth and the differenced openness variable. As was the case with the
percentage trade measure, the coefficients on the lagged values of openness suggest
that openness has a positive impact on growth. The elasticity of grth with respect to
openness after one year is 0.087. With trade as the measure of openness, this elasticity
is 0.074. Once again the elasticity decreases as the lag length increases. Table 1.13
also shows evidence that larger countries are more open. So once again the result is
similar to that which is obtained with trade as the measure of openness.

The models shown in table 1.13 were estimated via OLS for each country in
the sample individually. These results are illustrated in figures 1.9 and 1.10. Figure 1.9
shows the long run elasticity of growth with respect to openness. This impact is
positive for 64% of the countries in the sample. Most of the countries for which the
impact is negative are low income countries. The largest elasticity in absolute value
belongs to Denmark with a value of —0.905. Argentina is the country with the largest
positive long run effect. As figure 1.9 illustrates this value is 0.793. The average long
run elasticity is 0.078 which is lower than the value obtained using trade (0.167). The
country with the smallest elasticity in absolute value is the Netherlands, with an
elasticity of -0.0019.

The average impact of growth on openness is 0.051. However the elasticity is
negative for 57% of the countries in the sample. Of the four income groups the lower
middle group has the largest percentage (38%) of the countries for which the elasticity

is negative. This is also true when trade is used as the measure of openness.

34

Table 1.14 shows evidence that the impact of growth on openness varies across
income groups. For example, the impact of growth on openness after one year is

0.45% higher for upper middle income countries than it is for low income countries.

(iii) Panel Data Granger Causality Tests

Table 1.9 shows the results of conducting an F test for the null of non-causality
on the pooled data. I estimate the models of table 1.13 using fixed effects. The null
hypothesis of non-causality from openness to grth is rejected even at the 1% level.
The test statistic is 29.65 and the p—value is 0. Similarly the null hypothesis fiom
growth to openness is also rejected. The F statistic is 26.05 and the p-value is 0. We
see that the AR(l) model is most appropriate for the openness variable.

Table 1.11 shows that once again using one lag, the HNC hypothesis from
openness to growth as well as fiom growth to openness can be rejected for the entire
sample, at the 5% level. Going from openness to growth, the HNC hypothesis cannot
be rejected for the low and high income groups, while the hypothesis from growth to
openness cannot be rejected for the middle income group. This suggests that for the
low and high income countries the direction of Granger causality is from growth to
openness, while for the middle income countries the direction of Granger causality is

from openness to growth.

35

1!!) Trade in Homogeneous Products
(i) Unit Root Testing

The null hypothesis that every individual in the sample contains a unit root is
rejected using the first difference of i the openness measure. The p-values for the

individual unit root tests are less than 0.05 for 83% of the sample. The HNC

hypothesis is tested using the first difference of the variable.

(ii) Preliminary Rgggessions

Once again, the estimated coefficients on the lagged values of openness, suggest
that a reduction in trade barriers leads to growth. The elasticity of growth with respect to
openness, after one year, is 0.09. This decreases to 0.063 after two years. As table 1.15
shows, the coefficient of lagged growth implies that larger countries are more open.
Table 1.16 illustrates evidence that the impact of growth on openness varies across
income groups.

Figures 1.1] and 1.12 illustrate the results of estimating the models fi'om table
1.15 for each country individually. The average long run elasticity of growth with respect
to openness is 0.073. This elasticity is positive for 64% of the countries in the sample.
Most of the countries for which the impact is negative are classified as low income. As is
the case with trade in differentiated products the country with the largest elasticity is
Denmark, while the country with the largest positive elasticity is Argentina.

The average elasticity of openness with respect to growth is positive (0.079).
However, the impact is negative for approximately half (53%) of the countries in the

sample. As the previous section indicated this is also the case with trade in differentiated

36

products. Once again, the lower middle income countries account for the largest

percentage (39%) of the countries for which the elasticity is negative.

(iii) Panel Data Granger Causality Tests
I conduct an F test for the null of non-causality on the pooled data. The null

hypothesis of non-causality is rejected both from growth to openness and from openness
to growth. In both case the p-value is 0. Once again an AR (1) model is most suitable for
the openness variable.

Using one lag, the HNC hypothesis fiom openness to growth, as well as fiom
growth to openness, can be rejected for the entire sample, at the 5% level. As table 1.11
shows, the hypothesis from openness to growth cannot be rejected for low, high, and
upper middle income countries. The HNC hypothesis from growth to openness cannot be
rejected for the high and middle income countries.

The results for the statistics that are robust to heteroskedasticity are reported in

the appendix.

1 .8 Discussion ot Results

The results using trade percentage as the measure of openness indicate that the
HN C hypothesis from openness to growth is rejected for all income groups. However
using the alternative measures of openness the results are different. When trade in
differentiated products is assumed, the HNC hypothesis from openness to growth cannot
be rejected for the low and high income groups. This suggests that the alternative

measure of openness may capture more complexity in the relationship between openness

37

and growth than the simple measure does. With a sample of 107 countries, the
assumption of trade in differentiated products is more plausible than the assumption of
trade in homogeneous products. Moreover, as Feenstra, Markusen and Rose (2001) points
out product differentiation is usually assumed when deriving the gravity equation.

Interestingly, the results using trade percentage as the measure of openness are the
same as the results using the alternative measure of openness and assuming trade in
differentiated goods when testing the HNC hypothesis in the opposite direction. In both
cases the HNC hypothesis from growth to openness cannot be rejected for middle income
countries only.

If the alternative measure of openness is accepted as being more accurate and if
trade in differentiated products is assumed, the results suggest that the direction of
Granger causality for low and high income groups is from growth to openness while the
direction of Granger causality for middle income countries is from openness to growth. It
may be that for high income countries, there are factors (other than openness) that are
more important to predicting growth. Whereas low income countries may not have the
resources needed to take advantage of the opportunities provided by openness. This
would leave middle income countries as the only group for which the path of openness is
usefirl for forecasting the path of growth. It may be that for these countries, the policies
that helped them to achieve middle class status work together with openness to generate
increases in the growth rate of GDP.

This is consistent with the North-South theory of product development. It may be
that for the developed, high income countries of the North factors such as the research

and development and technology, which they used to innovate, are more important for

38

predicting future values of growth. Middle income countries have the resources and
income needed to be able to imitate the products developed in the North. Openness
provides the opportunity for this technology diffusion to occur and therefore leads to
growth for these countries. On the other hand, low income countries may be too ‘poor’ to
take advantage of the opportunities provided by openness. These countries may not
possess the resources or levels of income required to undertake the costs of imitation.
Therefore technology transfers may not occur. Grossman and Helpman (1990) argues that
technology diffusion is not automatic. If these knowledge transfers do not occur,
openness will not be important for predicting growth in low income countries.

The above findings are consistent with the findings of Ahmad and Hamhirun
(1996) for Indonesia and Singapore. Indonesia is classified as a low income country
while Singapore is classified as a high income country. The above results therefore
suggest that for these two, openness does not Granger cause growth however causality in
the opposite direction is statistically supported. However Ahmad and Hamhirun also
found this result for Malaysia, the Philippines and Thailand which are all classified as
lower middle income countries. The results of this paper suggest that for these countries
openness may Granger cause growth. The difference in results may be attributed to the
fact that Ahmad and Hamhirun investigate the relationship between exports and growth
while I examine the relationship between openness and growth.

The results of this paper also differ fi'om the findings of Tao and Zestos (2002).
Tao and Zestos found that imports Granger cause growth in Canada, and that exports
Granger cause GDP growth in Canada and the US. Both of these countries are classified

as high income. countries. Therefore the results of this paper using the alternative measure

39

of openness imply that openness does not Granger cause growth for these countries.
However the above findings of Tao and Zestos are consistent with the results of this
paper using trade percentage to measure openness. When trade percentage is used to
measure openness the HN C hypothesis from openness to growth is rejected for the high
income group. This similarity in results makes sense since Tao and Zestos explore the

relationship between trade and GDP growth.

1 .9 Conclusion

This paper uses a novel approach to examine the relationship between trade
and growth. It shows how the concept of Granger Causality can be extended to apply
to panel data. A new test statistic allowing for variation across countries as well as
across time is introduced. Hopefully, future work will put forward variants of this
statistic that will allow the relaxation of some of the assumptions made.

Globalization is a powerful force that isn’t going anywhere any time soon.
Economies are becoming more and more open. The results of this paper imply that using
a more complex measure of openness may reveal subtleties about the causal relationship
between openness and growth that may not be obvious. It may be that trade predicts
growth but openness may only predict growth for middle income countries. If the impact
of trade is isolated, there is evidence that trade may Granger cause growth. However just
opening up one’s economy to trade by reducing trade restrictions etc. may not be a
surefire way to achieve growth. Any open economy policy may need to be accompanied
by other growth inducing policies if the benefits of openness are to be reaped. It is

possible that low income countries may not be in a position to reap these benefits.

40

The findings of this paper are not inconsistent with the endogenous growth theory or
with endogenous product cycle theory. It appears that whether or not openness ‘causes’
growth depends. Middle income countries seem to be the ones most likely to take
advantage of the opportunities provided by openness. It may be that for these countries
openness leads the economy to direct resources to activities such as research and
development, increasing product quality etc. that stimulate growth. Middle income
countries can benefit from openness by having access to a larger market for their goods,
more advanced technology, increased variety, etc. Low income countries may not possess
enough resources to take advantage of these opportunities and to use openness as an
engine for growth.

The results of this paper are also not inconsistent with ‘infant industry’ theories. It
may be that in low income countries, openness leads to the destruction of local industries
that are too ‘young’ to survive against the competition from imported goods. In this case
openness may lead to unemployment and dependence on foreign countries for goods.

This paper also highlights the fact that, as was suggested in Rodrik and Rodriquez
(1999), the relationship between openness and growth differs according to income levels.
This is important for development policy. The findings imply that emphasis on openness
as a stimulus for growth is most relevant for middle income countries, such as South
Africa, Korea, and Jamaica, for which the HNC hypothesis from openness to growth is
rejected.

A fruitful area for future study is the causal relationships between innovation, trade

and growth implied by the endogenous growth and endogenous product cycle theories. It

41

would be interesting to see what panel data Granger causality testing reveals about the

linkages among them.

42

Table 1.1 — Summary Statistics for Income and Trade

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Sample Split Year Variable No. of Mean Standard
Observations Deviation
Full 1981 Income 107 6197.973 5778.212
Trade 67.488 46.330
2000 Income 9100.228 9334.541
Trade 84.222 50.787
Low Income 1981 Income 31 1210.275 484.718
Trade 60.777 38.884
2000 Income 1482.541 975.360
Trade 66.693 45.707
Lower 1981 Income 36 3891.873 1563.458
Middle Trade 70.873 41.010
2000 Income 5195.696 2783.269
Trade 87.191 37.458
Upper 1981 Income 18 7231.266 2374.847
Middle Trade 83.075 66.795
2000 Income 10494.130 4218.728
Trade 87.896 47.139
High 1981 Income 22 16154.290 2991.545
Income Trade 58.650 43.402
2000 Income 25083.010 5483.062
Trade 101.059 71.555
Table 1 .2 — Correlations
Sample Split Income and Trade Growth and Trade
Full 0.0612 0.0291
Low Income 0.1528 0.0046
Lower Middle 0.0008 0.0586
Upper Middle 0.1704 0.0049
Middle Income 0.0605 0.0349
High Income 0.2909 0.1 138

 

43

 

 

 

 

 

 

 

 

 

 

Table 1.3 — Typical Gravity Equation Coefficients from Ben strand (1985)
Independent Variables 1965 Values 1976 Values
Own GDP 0.80 0.84

Partner GDP 0.65 0.69
Distance -0.72 -0.72
Common Border 0.61 0.74

EEC Dummy5 0.35 0.30

EFTA Dummy" 0.69 0.69

 

 

 

Table 1.4 - Estimated Coefficients from Feenstra, Markusen and Rose (2001)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Independent Variables Differentiated Goods Homogeneous Goods

Own GDP 1.06 0.54

Partner GDP 0.65 0.82

Distance -1 .04 -0.73

Common Border 0.06 0.05

Common Language 0.71 0.50

F TA 1.53 1.12

Table 1.5 — Correlations

Type of Good Actual Trade and Predicted Stationary Versions of
Trade Trade Percentage and

‘Openness’
Differentiated Products 0.139 0.9858
Homogeneous Products 0.132 0.9807

 

 

 

 

5 BBC dummy = dummy variable equal to one if both countries are members of the European Economic

Community, and zero otherwise

6 EFTA dummy = dummy variable equal to one if both countries are members of the European Free Trade

Area, and zero otherwise

44

 

 

 

Table 1.6 - Classification of Countries Accordin

to Income Levels

 

mg}.

 

 

 

 

 

 

Low Lower Middle Upper Middle
Bangladesh Angola Algeria Australia
Benin Belize Antigua Austria
Burkina Faso Bolivia Argentina Belgium
Burundi Botswana Barbados Canada
Central African Rep.1 Cameroon Brazil Denmark
Chad Cape Verde Cyprus Finland
China Chile Gabon France
Comoros Colombia Greece Hong Kong
1 Congo, Republic of Hungary Iceland
godgfzngleguifii. Costa Rica Iran Ireland
q . . Cote d‘lvoire Korea Rep 1 of Israel
Ethiopia . . I -
Gambia, The . . .
Ghana Dominican Republic Portugal Japan
Guinea Ecuador Romania Luxembourg
Guinea-Bissau Egypt Seychelles Netherlands
El Salvador St Kitts & Nevis New Zealand
Guyana -
Haiti 2” d Trinidad &Tobago :0"me
'"dia aiiizmiia 3mm”, 3:3:er
. enezue a
32:":8'3 Honduras Sweden
Lesdtho Jamaica Swrtz' erland
Mada ascar Jordan Taiwan
Mama Malaysia United Kingdom
Mali Mauritius United States
. . Mexico
Mauritania
. Morocco
Mozambique . .
Namibia
Nepal .
. Nicaragua
Niger .
. . Papua New Gurnea
Nigeria Paraguay
Pakistan
Peru
Rwama Phili ines
Sao Tome and Principe pp
Sierra Leone Poland
Sri Lanka Senegal
. South Africa
Tanzania .
St. Lucia
T090 . 2
Uganda St.Vincent
Zambia Syria
Thailand
Tunisia
Turkey
Zimbabwe
Notes:

1 Rep. = Republic

2 St. Vincent = St. Vincent & the Grenadines

45

 

Table 1.7 — Fixed Effects Estimation

 

 

 

 

 

 

 

 

Independent Variables Growth ATrade
growth_l "0.039 ***0.096
(0.016) (0.028)
Atrade_l ***0.074 ***-0.057
(0.008) (0.015)
Atrade_2 ***0.047
(0.008)
Atrade_3 ***0.028
(0.007)
Atrade_4 ***0.037
(0.007)
Atrade_5 ***0.026
(0.007)

 

 

 

Notes: standard errors are in parentheses
** denotes significance at the 5% level
"*denotes significance at the 1% level

46

 

Table 1.8 — Fixed Effects
V
growth_1

growth_1

1

errorsare

with Interactions

1

-0.015
1
0.022

" denotes significance at the 5% level
"*denotes significance at the 1% level

Table 1.9 - Pooled Granger Causality Test Results

 

 

 

 

 

 

 

 

 

 

 

Openness to Growth Growth to Openness
Measure of Openness F-statistic P-Value F-statistic P-value
Trade Percentage 33 .05 0.0000 11.44 0.0007
Gravity — Differentiated 29.65 0.0000 26.05 0.0000
Products
Gravity — Homogeneous 30.58 0.0000 30.41 0.0000
Products

 

47

 

Table 1.10 - Granger Causality Test - Unbalanced Panel — one 1

 

 

 

 

Statistic Trade to Growth Growth to Trade Critical Value
ZI 111 “L 7.908 4.310 1.650
z”NC 9.112 5.159 1.650

 

 

 

 

Table 1.11 — Granger Causality Test - Balanced Panel - one lag - Causality from

 

 

 

 

Openness to Growth

Measure of All Low Lower Upper Middle High
Openness Countries Income Middle Middle Income Income
Trade 6.551 2.887 4.429 3.596 5.692 2.101
Percentage

Gravity — 2.184 *0.468 2.407 3.246 2.661 *0.092
Differentiated

Goods

Gravity — 2.390 *0.564 2.455 *1.227 2.713 *0.357
Homogeneous

Goods

 

 

 

 

 

 

 

 

Note * denotes non-rejection at the 5% level

Table 1.12 — Granger Causality Test - Balanced Panel - one lag - Causality from Growth

 

 

 

 

to Openness

Measure of All Low Lower Upper Middle High
Openness Countries Income Middle Middle Income Income
Trade 3.811 4.597 *1.l67 *-0.550 *0.635 1.954
Percentgge

Gravity — 3.602 3.331 *1.512 *-0.260 * 1.084 2.290
Differentiated

Goods

Gravity — 3.251 4.195 1.750 *-0.028 * 1.413 *-0.023
Homogeneous

Goods

 

 

 

 

 

 

 

 

Note * denotes non-rejection at the 5% level

48

 

 

Table 1.13 - Fixed Effects Estimation — Openness based on Trade in Difierenfiated
Products

1 -0.037

openness_l

openness_

    

errors are
" denotes significance at the 5% level
*"denotes significance at the 1% level

Table 1.14 — Fixed Effects Estimation with Interactions - Openness based on Trade in
Differentiated Products
Growth
1 -0.037

openness_
openness_

openness_]

openness_]

openness_]
1

 

errors are
" denotes significance at the 5% level
“*denotes significance at the 1% level

49

Table 1.15 — Fixed Effects Estimation — Openness based on Trade in Homogeneous

 

 

 

 

 

Products
Independent Variables Growth Openness
growth_l -0.048 "0.246
(0.026) (0.045)
openness_l ***0.090 ***-0.058
(0.014) (0.024)
openness_2 ***0.063
(0.01 3)

 

 

 

Notes: standard errors are in parentheses
** denotes significance at the 5% level
""denotes significance at the 1% level

Table 1.16 — Fixed Effects Estimation with Interactions - Openness based on Trade in

 

 

 

 

 

 

 

 

 

 

 

Homogeneous Products
Independent Variables Growth Openness
growth_1 -0.050 ***0.389
(0.026) (0.059)
growth_1"'lm "0.248
(0.1 13)
growth_1*um "*0418
(0.1 17)
growth_1 *h 0.360
(0.232)
openness_l ***0.088 l”-0.055
(0.019) (0.024)
openness_2 * * *0.063
(0.013)
openness_] *lm 0.021
(0.033)
openness_] *um -0.024
(0.035)
openness_l *h -0.087
(0.102)

 

 

 

Notes: standard errors are in parentheses
** denotes significance at the 5% level
"*denotes significance at the 1% level

50

 

 

Figure 1.1 — Relationship Between Income and Trade in 1980

1

15000 20000 25000
1
o 8
O
O

 

 

 

 

 

 

(I)
g .00 . . . '
E
> _
o
E} . .
Ito ° ° 0
a) o
E§q 0 ° 0 . o
O‘—
2 . . O __.___....
'- .. .. . . . . C
o
g4 ‘. $ 0 :0 . . .0 O
..Oa . .O .
o o o. o o
I. O Q 0 ‘0 . O
o_ $335. 0
I I I I T I
0 50 100 150 200 250
trade
[- incorre Fitted valuefl

 

Figure 1.2 — Relationship Between Income and Trade in 1996

 

 

 

 

 

O
o —4
8
v
I
O
o -
a .3. '
2 c
9 . . .; °
O
E g L O o 0'. . .
LL
\ N
o
E
o .4
I I I l I
0 100 200 300 400
trade
[- incorre Fitted values

 

51

Figure 1.3 - Trade Percentage in 1980

 

35—520.“. 3!...

For a complete list of country abbreviations and IDs see Table 17

Note

Figure 1.4 - Trade Percentage in 1996

 

35:023. 25:.

52

Figure 1.5 — Scatter Diagram of Values Used in Granger Causality Tests

Trade In Differentiated Products - 2000 values

 

Figure 1.6 — Scatter Diagram of Values Used in Granger Causality Tests

Trade In Homogeneous Products - 2000 values

 

53

Figure 1.7 - LR Impact of trade on growth

 

Flgure 1.8 - Impact of Growth on Trade

 

54

Elasticity

Elasticity

Figure 1.9 - LR Impact of Openness on Growth - Trade In Differentiated
Products

 

Figure 1.10 - Impact of growth on Openness - Trade In Diferentiated Products

 

55

Elasticity

Elasticity

 

Figure 1.11 - LR impact of Openness on Growth - Trade in
Homogeneous Products

 

0 ARG
. COM’ GNQ , ,
O . O . g
Q Q
o . ° . ° 0.
Q . O O . O .
O O .
Q 0 O 0 0
e e ..e e .0. e e e . . d. . O
Q . .0 . O .
° ° ooCPv

.p
O 0 O O o
. . .BRA ’DOMO OGIN . . .e ° e °I3HL OZMB
.° .006 o.

.KOR .ZWE

' BEI- . FRA . ITA . KNA ° . VEN

0 DNK

 

 

 

Figure 1.12 - Impact of growth on Openness - Trade in Homogeneous
Goods

 

0 PRY

' "3N . MDG

OCHN . ‘MUS ,
sBDI OGRD
O

’ e 0 0 e
Q ..FRA O ' Q Q ..0
.. . ’BRA . . .. . 'LUX . : .e .OUSA
O
O

0

e
0‘. e e ' " e
0.. O . .g 9 O O
. ' OFIN . ‘ ° . .. - “3’1”?
e

0 DZA °
. BFA 0 CRI 0 KEN

e BGD ’ IRN

 

 

56

Table 1.17 - List of Country Codes and II)’

S

 

 

 

Country Code ID
Algeria DZA 1
Angola AGO 2
Antigua ATG 3
Argentina ARG 4
Australia AUS 5
Austria AUT 6
Bangladesh 360 7
Barbados BR8 8
Belgium BEL 9
Belize BLZ 10
Benin BEN 1 1
Bolivia BOL 12
Botswana BWA 13
Brazil BRA 14
Burkina Faso BFA 15
Burundi BDI 16
Cameroon CMR 17
Canada CAN 18
Cape Verde CPV 19
Central African
Republic CAF 20
Chad T00 21
Chile CHL 22
China CHN 23
Colombia COL 24
Comoros COM 25
Congo, Dem. Rep. ZAR 26
Congo, Republic of COG 27
Costa Rica CRI 28
Cote d‘lvoire CIV 29
Cyprus CYP 30
Denmark DNK 31
Dominica DMA 32
Dominican Republic DOM 33
Ecuador ECU 34
Egypt EGY 35
El Salvador SLV 36
Equatorial Guinea GNQ 37
Ethiopia ETH 38
Fiji FJI 39
Finland FIN 40
France FRA 41
Gabon GAB 42
GMB 43

Gambia, The

 

57

 

Table 1.17 — List of Country Codes and ID’s - Continued

 

 

 

Country Code ID
Ghana GHA 44
Greece GRC 45
Grenada GRD 46
Guatemala GTM 47
Guinea GIN 48
Guinea-Bissau GNB 49
Guyana GUY 50
Haiti HTI 51
Honduras HND 52
Hong Kong HKG 53
Hungary HUN 54
Iceland ISL 55
India IND 56
Indonesia IDN 57
Iran IRN 58
Ireland IRL 59
Israel ISR 60
Italy ITA 61
Jamaica JAM 62
Japan JPN 63
Jordan JOR 64
Kenya KEN 65
Korea, Republic of KOR 66
Lesotho LSO 67
Luxembourg LUX 68
Madagascar M06 69
Malawi MWI 70
Malaysia MYS 71
Mali MLI 72
Mauritania MRT 73
Mauritius MUS 74
Mexico MEX 75
Morocco MAR 76
Mozambique M02 77
Namibia NAM 78
Nepal NPL 79
Netherlands NLD 80
New Zealand NZL 81
Nicaragua N10 82
Niger NER 83
Nigeria NGA 84
Nomay NOR 85
Pakistan PAK 86
Panama PAN 87

 

58

 

Table 1.17 — List of Country Codes and ID’s - Continued

 

 

 

Country Code ID
Papua New Guinea PNG 88
Paraguay PRY 89
Peru PER 90
Philippines - PHL 91
Poland POL 92
Portugal PRT 93
Romania ROM 94
Rwanda RWA 95
Sao Tome and Principe STP 96
Senegal SEN 97
Seychelles SYC 98
Sierra Leone SLE 99
Singapore SGP 100
South Africa ZAF 101
Spain ESP 102
Sri Lanka LKA 103
St. Kitts 8 Nevis KNA 104
St. Lucia LCA 105
St.Vincent & Grenadines VCT 106
Sweden SWE 107
Switzerland CHE 108
Syria SYR 109
Taiwan TWN 1 10
Tanzania TZA 1 1 1
Thailand THA 112
Togo TGO 1 13
Trinidad &Tobago TTO 114
Tunisia TUN 1 15
Turkey TUR 1 16
Uganda UGA 1 17
United Kingdom GBR 118
United States USA 1 19
Uruguay URY 120
Venezuela VEN 121
Zambia ZMB 122
Zimbabwe ZWE 123

 

 

59

What is the impact of
Openness on Growth? A

Panel Data Approach

60

2.1 Introduction

What is the impact of openness on growth? The benefits of openness have been
touted since Adam Smith introduced the concept of absolute advantage. David Ricardo
built on this concept to formalize the theory of comparative advantage. This theory states
that countries can gain from trade by specializing in and exporting the good that they can
produce at a lower relative cost. Openness provides opportunities for technology
transfers, and increased domestic competition and efficiency resulting from international
competition. However, in practice, do countries benefit from openness? Does openness
lead to an increase in a country’s growth rate? Does openness cause countries to grow
faster?

The usual approach to investigating the relationship between openness and growth
involves the specification of a cross sectional model for gross domestic product (GDP)
per capita with some measure of openness and other variables on the right hand side of
the equation. The argument is that if openness has a positive impact on GDP per capita
then an increase in openness will lead to an increase in GDP per capita. Therefore these
studies do not investigate the impact of openness on a country’s growth rate. The
weakness of this approach is that openness may be endogenous. This may be as a result
of simultaneity. It is possible that countries with high incomes (for reasons other than
trade) may trade more. To overcome the endogeneity issue I use an approach that is
sometimes referred to as the ‘Barro’ regression, because of Barro’s seminal paper. Barro
(1991) investigates the concept of “conditional convergence” by running regressions of
the growth rate of income over a period of time on initial income as well as other control

variables. In this paper I use the Barro growth regression approach but I add initial

61

openness as one of my explanatory variables. This overcomes the issue of simultaneity
between growth and openness. Openness at the beginning of the period cannot be caused
by the growth rate of income over the entire period.

Barro (2000) extends the traditional cross-sectional growth regression,
popularized in Barro (1991) and Mankiw, Romer and Weil (1992), to panel data. This
paper also uses panel data to take advantage of the information provided by considering
variation both across countries and over time. The estimation techniques employed
include pooled ordinary least squares (POLS), fixed effects (FE), and first differencing
(FD), and instrumental variables versions of these. These techniques will be discussed in
detail in section 3. The paper is organized as follows: section 2 reviews the relevant
literature, section 3 discusses the methodology, section 4 describes the data, section 5
presents the results, and section 6 discusses the main findings of the paper in the

conclusion.

2.2 Literature Review

The general framework for analysis used in Barro regressions is the following

model:
git =k0’ +k1yiat—j +k2piJ—j +"3in +k4wi +a,- +1117 , i=1,...,N (1)
where:
. y. _y. _ .
g it = annual average growth rate of Income = ————u j” J

yi,,_j = log of initial period income

17”-]- = primary variable of interest

62

xi, = vector of time varying explanatory variables

w,- = vector of time constant explanatory variables

a,- = time constant unobserved effect

“it = idiosyncratic error term
N refers to the number of individuals in the sample and L is the length of the entire
sample period. Some of the variables in x are measured as period averages while others
are measured as initial period values.

In the earlier papers using this approach, for example, Barro (1991) and Mankiw,
Romer and Weil (1992), the primary variable of interest is the lagged dependent variable.
In more recent papers such as Barro (2000) and Dejong and Ripoll (2006), the primary
variables of interest are inequality and tariffs respectively. Another feature of earlier
papers, is the fact that they take a cross-sectional approach to estimating the model. In
doing so they assume that the x ' s and w's control for all cross-country heterogeneity.
Any unobserved heterogeneity is captured in the idiosyncratic error term so there is no
need for a time constant fixed effect. These more recent papers take a panel approach to
estimating (1) and therefore acknowledge the presence of the time constant unobserved
effect. The different approaches to estimating equation (1) and the assumptions involved
will be examined in more detail in section 3. This section discusses the motivations and
findings of previous papers that have used the Barro regression.

Barro regressions have traditionally been used to explore the concept of
convergence. Barro (1991) asks the question “do poor countries grow faster than rich
countries?” He argues that neoclassical growth models, such as Solow (1956), Cass

(1965), and Koopmans (1965), imply that a country’s per capita growth rate is negatively

63

related to its initial level of income per person. Barro runs cross-sectional regressions of
equation (1). He uses the annual average growth rate of per capita real GDP from 1960-
1985, and fiom 1970-1985 as dependent variables. His primary variable of interest is
lagged income. His x '3 include g0vernment spending, and measures of human capital
and political instability. His measures of human capital include secondary and primary
school enrollment rates, the student-teacher ratio in primary and secondary schools and
the adult literacy rate. He uses the period average for government spending and initial
values for the enrolhnent rates. His sample includes 98 countries.

Barro frnds support for “conditional convergence”. He finds that given the human
capital variables growth is negatively related to the initial level of per capita GDP. Poor
countries grow faster than rich countries, but only for a given quantity of human capital.

Mankiw, Romer and Weil (1992) also test the predictions of the Solow growth
model. They, too, include human capital as an explanatory variable. They argue that the
augmented Solow model (which includes human capital) predicts that countries generally
arrive at different steady state income levels. Therefore income per capita in any country
converges to that country’s steady state value. Thus, the Solow model predicts
“conditional convergence”. That is, convergence occurs only after one controls for the
determinants of the steady state.

Mankiw, Romer and Weil (MRW) are primarily interested in the impact of initial
period income on growth. Their x ' .9 include investment, population growth, and school
enrollment. They estimate the cross-sectional version of equation (1) over the period
1960 to 1985. Their three samples consist of 98, 75, and 22 countries. The coeflicient of

initial income is negative and statistically significant. They therefore argue that once one

64

controls for differences in investment, population growth rates and human capital,
convergence occurs. Thus using a slightly different set of control variables than the one
used in Barro (1991) MRW also find evidence of conditional convergence.

Sachs and Warner (1995) also use cross-sectional growth regressions to explore
the issue of convergence. They too find evidence of conditional convergence, but unlike
Barro (1991) and MRW (1992) they argue that convergence is conditional on good policy
choices. Specifically, they argue that these good policy choices include protecting private
property rights and maintaining economic openness. A country is classified as having
failed the property rights test if it is characterized by at least one of the following
conditions:

(1) A socialist economic structure

(2) Extreme domestic unrest

(3) Extreme deprivation of civil or political rights
A country is classified as having failed the economic openness test if it possesses any of
the following:

(1) A very large percentage of imports covered by quota restrictions

(2) A high percentage of exports covered by state export monopolies and state-set

prices

(3) A socialist economic structure

(4) A black market premium on the official exchange rate of 20% or more

Their data set includes 117 countries over the years 1970-89. Sachs and Warner
replicate the basic Barro (1991) regression of cross-country growth for their sample and

time period including drunmy variables for political and openness qualifications (1 for

65

countries that fail and 0 for countries that pass). Therefore their primary variables of
interest are these dummy variables. They found that political and openness non-qualifiers
grow more slowly than the countries that passed the tests. They argue that this is evidence
that the growth rate over this period was determined by policy choices.

Rodriquez and Rodrik (1999) question the measure of openness used in Sachs and
Warner (1995). By dividing the binary openness variable into different variables based on
each individual characteristic, they show that the significance of the openness dummy is
being driven by the state monopoly on exports and the black market premium criteria.
They then point out that the index for state monopoly on exports was taken from a World
Bank study that only covered 29 African countries. Therefore non-African countries with
restrictive export policies are not captured at all. They also argue that black market
premia reflect a wide range of policy failures such as inflation pressures, high and
growing levels of external debt, and no enforcement of rule of law, among others. They
therefore conclude that the measure of openness used by Sachs and Warner does not
reflect trade policy. On the contrary they argue that there is no evidence that complex
binary measures such as the one used in Sachs and Warner (1995) do a better job of
measuring openness than more direct measures.

As such this paper uses 2 direct measures of openness. The first one is the
standard measure of openness used by institutions such as the international monetary
fund (IMF) and the World Bank — trade percentage ((exports + imports)/GDP). The
second is the difference between the amount of trade predicted by the gravity model and

actual trade.

66

Levine and Renelt (1992) explore the issue of different authors using different
control variables in growth regressions. They state that over 50 variables have been found
to be significantly correlated with growth in at least one regression. They point out that
economists who examine the relationship between fiscal policy measures and growth
ofien ignore trade policy measures while those who explore the relationship between
trade and growth ofien exclude fiscal policy variables. They examine whether the
findings of such studies are robust to changes in the conditioning information set.

The methodology they use is extreme-bounds analysis (EBA). They test the
robustness of coefficient estimates to changes in the control variables. They estimate the
cross-sectional version of equation (1). They distinguish between ‘1’ variables, ‘M’
variables and ‘Z’ variables. ‘1’ variables are the set of variables always included in Barro
regressions. The ‘M’ variable is the variable of interest and the ‘2’ variables are a subset
of variables chosen from a pool of variables identified as significant in past studies.

The data set covers the period 1960-1989 and includes 119 countries. The ‘1’
variables consist of the investment share of GDP, the initial level of real GDP per capita,
the initial secondary-school enrollment rate and the average annual rate of population
growth. They argued that of the 41 growth studies surveyed few included all of the above
but most of them controlled for some subset.

Their main findings are that firstly no reliable, independent statistical relationship
between a wide variety of macroeconomic indicators and growth exists. Secondly there is
a positive and robust correlation between growth and the share of investment to GDP and
thirdly the ratio of trade to output is robustly, positively correlated with the investment

share.

67

Levine and Renelt highlight the difficulty in choosing which variables to include
as explanatory variables. As was mentioned above they found that over 50 variables have
been found to be correlated with growth. Since including 50 variables in a regression
would result in a substantial loss in degrees of fi'eedom and since it would be extremely
difficult, if not impossible, to find data on all 50 variables for a large number of
countries, one has to exercise discretion in choosing which variables to include in the
information set. This decision of course will be guided firstly by the question being
asked.

The findings of Levine and Renelt also provide guidance in this issue. They imply
that the investment share should definitely be included in the information set. I would
also argue that the complete set of ‘1’ variables should be included where possible. The
‘1’ variables were the variables that were included in most growth regressions. This
would imply that they have been ‘agreed upon’ as being important for determining
growth.

A more recent paper that has explored the issue of growth and convergence is
Higgins, Levy, and Young (2006). They study growth and convergence using county
level data from across the United States. Their data set includes 3,058 county level
observations and covers 1970-1998. They estimate the cross-sectional version of equation
(1). They include 41 conditioning variables such as area, percentage of Hispanics, school
enrollment, percentage of the p0pulation below the poverty line, percentage of the
population employed in finance, construction, and entertainment, among others. They are
able to include so many independent variables because of the large number of

observations and because they only include United States data.

68

The estimation techniques that they use are OLS and a procedure which they refer
to as 3SLS-IV which was introduced in Evans (1997). They argue that the OLS
procedures usually used require strong, unrealistic assumptions. However the technique
introduced in Evans (1997) also requires strong unrealistic assumptions. This procedure
and the assumptions will be discussed in detail in section 3. Higgins, Levy and Young
find evidence of convergence and conclude that convergence rates are variable. For
example, the Southern counties converge more than two and a half times faster than the
counties in the Northeast.

Barro (2000) extends the traditional cross-sectional growth regression to panel
data. Using a panel data set over three decades, 1965 to 1975, 1975 to 1985, and 1985 to
1995, Barro explores the relationship between inequality and growth. As his fiamework
for analysis Barro uses the model in equation (1). Therefore his entire sample period is
from 1965 to 1995. So in this case, L= 30, and T= 3. His sample size is 146
observations. His dependent variable is the growth rate of real per capita GDP and his
independent variables include the period averages of the ratios of government spending
and investment to GDP, the initial period value of years of schooling, and initial income,
among others. Since he is interested in the relationship between inequality and growth, he
also includes a measure of inequality, the Gini coefficient, as his primary variable of
interest. The estimation technique that Barro uses is three stage least squares (3SLS) with
mainly lagged values of the regressors as instruments. He finds that, ceteris paribus,
changes in Gini coefficients for income inequality have no significant impact on
economic growth. He concludes that overall there is little relation between income

inequality and growth.

69

Dejong and Ripoll (2006) have also used a panel version of the growth regression.
Using a panel data set of 60 countries from 1975-2000 they explore the contingent
relationships between tariffs and growth. They split the data into 5 non-overlapping half
decades. So, in their model, L = 25, l = 5, and T = 5. However, the dependent variable
in their sample is logged per capita real GDP measured in 1980, 1985, ..., 2000. Their
independent variables include lagged income, (for example when GDP in 1980 is the
dependent variable, GDP in 1975 is an explanatory variable), proxies for human capital,
ad valorem tariffs, investment as a percentage of real GDP and government expenditure
as a percentage of GDP. They distinguish between stock variables measured at the
beginning of the period and flow variables which are period averages. They classify
lagged income and the human capital proxies as stock variables and tariffs, investment
share, and government expenditure as flow variables. To capture contingencies in the
relationship between tariffs and growth, they include as an explanatory variable the
product of lagged income and tariffs. Altemately, they also include the product of tariffs
and the 1975 income rankings. They also capture contingencies by dividing the data set
into two sub-samples — one with high and upper middle income countries and the other
one with lower middle and low income countries.

The authors use 4 different estimation techniques — generalized method of
moments (GMM) systems, GMM difference, panel seemingly unrelated regression
(SUR) and cross-sectional OLS. These techniques will be discussed in detail in section 3.
Dejong and Ripoll conclude that tariffs have a significant and negative impact on growth
among the world’s rich countries but a positive and sometimes insignificant (depending

on the estimation technique) impact on growth among the world’s poor countries.

70

As a study of the relationship between openness and growth Dejong and Ripoll
(2006) is limited. This is because of the existence of non-tariff trade barriers. I am
interested in looking at the impact of openness, and not just one particular type of trade
barrier, on growth. Therefore I use more comprehensive measures. Another limitation of
Dejong and Ripoll (2006) is the fact that population growth is excluded as a control
variable. As noted in MRW (1992) population growth is an important explanatory
variable in the Solow growth model. Also as Levine and Renelt (1992) highlights,
population growth is one of the variables that is included in most regressions and is
therefore agreed upon by most economists as being an important determinant of GDP
growth.

Dejong and Ripoll (2006) can also be criticized on the basis of one of their
proxies for human capital. Their proxies for human capital include average years of
secondary school for males and females over the age of 15 and log of life expectancy. It
can be argued that life expectancy is not a very good measure of human capital. It is a
better measure of the state of a country’s health sector. Furthermore, as Levine and
Renelt (1992) point out, the standard measure of human capital is school enrolhnent

rates.

2.3 Methodology

Different estimation techniques can and have been used to estimate equation (1).
This section provides a unified framework for the analysis of these techniques. The first
approach that was employed was to estimate the equation over the entire time period

L and therefore to use cross-sectional ordinary least squares (OLS). Consistency of these

71

estimates requires the assumption that the independent variables control for all
observable cross-country heterogeneity. Any unobservable factors are included in the
idiosyncratic error term, which is assumed to have a conditional expectation of zero. In
this case j = L and initial income is measured at year 0 only. The model in equation (1)
therefore becomes:
git = k0, + kIJ’i,t-j + kzpi,,_j + k3xit + k4w,- + "it , i =1,..., N (2)
More recently papers such as Barro (2000) and Dijong and Ripoll (2006) have

taken the approach of dividing the sample period L into sub-periods of length I < L . In

.1:

1 time periods.

this case j =1 and the result is a panel data set with N countries, and T =

Using a panel data approach allows us to look at dynamic relationships. If it is assumed
that there is no time constant unobserved heterogeneity, that is, ai =0, then pooled OLS
(POLS) can be used to estimate (2). Sequential exogeneity is required for consistent

estimation of the coefficients of the explanatory variables. This assumption can be

written as:
Eon I qr.) = 0. t =11....,T (3)
Where:
qit = (yiJ—j’PiJ—j’xit’wi)
That is, we need to assume that the idiosyncratic error term is uncorrelated with the
explanatory variables in the same time period.

If we acknowledge the existence of the time constant unobserved effect a, then

POLS leads to biased and inconsistent estimates. Fixed effects and FD are methods of

dealing with the unobserved effect which do not require one to assume that it is

72

uncorrelated with the explanatory variables. Both methods involve eliminating the fixed
effect. However, for consistency of the estimates, they both require the assumption of
strict exogeneity of the idiosyncratic error. This assumption is stronger than sequential
exogeneity and can be written as: ‘
E(u,-, |q,-1,q,-2,...,q,-T) = 0, t =1,...,T (4)

Strict exogeneity means that the error term is uncorrelated with the explanatory variables
in all time periods. The regressors in equation (1) cannot be strictly exogenous. This is
because of the presence of the lagged dependent variable. Therefore if FE or FD are used

as estimation techniques we need to be concerned about the endogeneity of yi, ,_ j . One

way to solve that problem is to use instrumental variables.

There may also be concern about the endogeneity of the xi, variables. If, for

example, the growth rate of GDP has an impact on investment or government spending,
the estimates of these variables would be biased and inconsistent. One approach to
solving this problem is to use initial values for these variables. Another possible solution
is to use instrumental variables. Possible instruments for the endogenous variables
include lagged values of the regressors. Any time constant explanatory variables can also
be used as instruments for the lagged dependent variable.

As discussed in the literature review, Higgins, Levy, and Young (2006) estimate a
cross-sectional version of equation (1) using data fiom 1970-1998 on initial output and
other conditioning variables. However the estimation technique they use is not OLS.
Instead, they use a methodology introduced in Evans (1997). Evans argues that the
assumption, that the conditioning variables control for all cross-country heterogeneity is

too strong.

73

The technique introduced in Evans (1997) starts with the cross-sectional variant

of the model in equation (1). This is then differenced to give the following:
A git = Co + k1 Ayr',t— j + Vi: (5)
It should be noted that the above equation is only true if all of the conditioning variables

are constant over time. Instrumental variables is then used to estimate equation (5).

Lagged values of the x's and the p's as well as the w's can be used as instruments. The

estimates from that regression are used to construct:
a:
"it = git — k1 J’i,t— j (6)
Then r,-, is regressed on xit, pi,,_ j and was follows:
’it = d1 + k2Pi,t- j + k3xi, + k4Wi + eit (7)
Evans (1997) shows that OLS gives consistent estimators of k2,k3 , and k4 . The
weakness of the above method however is the highly unrealistic assumption that all of the

conditioning variables are constant over time. While the w's are constant over time, the

x's and the p's vary. Therefore, 1 will not use this method.

DeJong and Ripoll (2006) use a panel data growth regression approach. The
estimation techniques they use are generalized method of moments (GMM) systems,
GMM difference, panel seemingly unrelated regression (SUR) and cross-sectional OLS.
GM is a method of applying instrumental variables to a system of equations. DeJong
and Ripoll use the GM dynamic panel (systems) estimator developed by Arrellano and
Bover (1995) and Blundell and Bond (1998). They estimate the following version of
equation (1):

yit = k0: + (1 + k1 )yi,t— j + k2pit + k3xiz + “i + ”it (8)

74

They use lagged differences of the variables on the right hand side of the above equation
as instruments. Of course the consistency of the resulting estimates requires the
assumption that there is no correlation between the differences of the regressors and the
unobserved country specific effect. Their primary variable of interest is ad valorem
tariffs. Their x's include three proxies for human capital measured at the beginning of the
period and investment share and government share measured as period averages.

The GM difference estimator they use was developed in Arellano and Bond
(1991). This involves estimating the first difference of equation (8) and using lagged
levels of the explanatory variables as instruments. The rationale behind this technique is
to take account of the unobservable fixed effect and any possible endogeneity. The third
estimation technique used by Dejong and Ripoll is panel SUR. They specify separate
regressions for each half decade and the covariance structure of the error term is allowed
to vary across equations. As they point out, this is closely related to the random effects
estimator. Of course this requires the assumption that the unobserved effect is
uncorrelated with the explanatory variables. The final technique that DeJong and Ripoll
use is cross-sectional OLS. This is the most popular estimation technique used in growth
regressions.

I begin my analysis by estimating the cross-sectional variant of equation (1) using
data from 1981-2000. That is, I start by assuming that L = j =20, and estimate the

equation using OLS. I then divide the entire sample period L into T = 4 sub-periods each
of length I: 5. The result is a panel data set with N =107 countries and T =4 time
periods. The first estimation technique I use for the panel data set is POLS. This

technique ignores the presence of the time constant unobserved effect “1' , but only

75

requires sequential exogeneity of the explanatory variables for consistency. I then use
both FE and FD. Since these techniques require strict exogeneity, the presence of the

lagged dependent variable is a problem. This can be solved using values of y prior to

1985 or any time constant explanatory variables as instruments for M, t_ j .

The primary variable of interest in this paper is openness. The xi, variables

include investment share, government spending and population growth. The time
constant variable in the model is area. There may be concern about feedback from growth
to the investment share and government spending. As such when using FD and FE I use
instruments for these variables. As instruments I use lagged values. That is, I use values
of the variables prior to 1986. For example, when T = 2, I use the 1982 values as
instruments, and when T = 3, I use the 1983 values as instruments.

Another possible estimation technique involves running separate regressions for
each time period using instruments for initial income, investment and government
spending, and using the fitted values as instruments for initial income. The rationale
behind running separate regressions to get the fitted values is to take advantage of the fact
that as more recent data is used more instruments become available. For example when T
= 2, instruments for initial income, investment and government spending include values
of these variables prior to 1986, while when T = 4 the instruments include values of these
variables prior to 1996. This approach is similar to the one put forward in Arellano and

Bond (1991).

76

2.4 Description of Data

The sample is a balanced panel data set of 107 countries with data fiom 1981-
2000. 1 use data on PPP adjusted constant GDP per capita, investment share of GDP,
government share of GDP, openness and population obtained from the Penn World
Tables (PWT). The measure of openness used by the PWT is total trade as a percentage
of GDP. The alternative measure of openness used was constructed as the difference
between the amount of trade predicted by the gravity model and actual trade. For details
see Whitely (2006). Data on area was obtained from the Centre D’Etudes Prospective et
D’lnforrnations Intemationales database.

Table 2.1 gives summary statistics for GDP per capita in 1981, and the growth
rate of GDP from 1981 to 2000 ((GDP in 2000 — GDP in 1981)/GDP in 1981). Not
surprisingly the average value for initial income increases steadily from 1,210 for low
income countries to 16,154 for high income countries. All values are expressed in terms
of 1996 dollars. The value for the growth rate over the period also increases from low
income to lower middle income to higher middle income to high income countries. This
provides preliminary evidence against convergence. Richer countries appear to have
grown faster over the period. I do a more formal test for convergence when I conduct
regression analysis.

Table 2.2 shows the 1981 values for trade share, investment share of GDP, and
the share of government expenditure in GDP. The trade share is highest among the upper
middle income countries (83%) and lowest among the high income countries (59%). The
investment share is highest among the high income countries (23%) and lowest among

the low income countries (10%). Interestingly, the low income countries have the highest

77

value for the share of government expenditure in GDP (24.6%). The upper middle
income countries rank a close second with 24.2% and the high income countries have the
lowest value (15.5%). It seems that the low income countries have the highest level of
government involvement in the economy. This can be viewed as evidence that less
government intervention and a ‘laissez-faire’ attitude are better for the economy than
more government intervention.

Table 2.3 shows the correlation between trade in 1981 and growth over the entire
sample period. For the fill] sample this correlation is 0.14. The correlation between initial
trade and growth is positive for all except the low income countries. The highest
correlation between trade and growth is among the high income countries (0.653). The
positive correlation between trade and growth is reflected in figure 2.1. Table 2.3 also
shows the correlation between initial income and growth. There is evidence of
convergence among the high and upper middle income countries only. Overall the
correlation between initial income and growth is positive. This overall relationship is

illustrated in figure 2.2.

2.5 Results

Table 2.4 shows the results of estimating equation 1 by cross—sectional OLS,
POLS, FE and FD. Once a panel data approach is used initial openness becomes
statistically significant at the 1% level. This is true no matter which of the three
estimation techniques is used. The coefficient on initial openness is 0.02 using POLS,
0.03 using FE and 0.04 using FD. This would imply that a 1 percentage point increase in

openness would increase growth by approximately 0.03 percentage points. The

78

coefficient on initial income is consistently negative and statistically significant at the 1%
level in all three cases. In the cross-sectional approach it is significant at the 5% level.
Using both FE and FD the elasticity of growth with respect to initial income is
approximately -0.1. This is similar to the coefficient of initial income in Barro (2000).
The coefficient using the cross-sectional version of the model is similar to the estimates
on initial income found in Barro (1991). The coefficient on investment is consistently
positive and statistically significant in all cases. They imply that a 1 percentage point
increase in investment would result in a 0.19 percentage point increase in growth.

Table 2.5 illustrates the results when values of income prior to 1985 are used as
instruments for lagged income. In this case the number of time periods is 3. Using both
FE-IV and FD-IV initial openness, initial income, and investment are all significant.
Once again the coefficient on initial openness is approximately 0.03, and the coefficient
on investment is 0.2. The coefficient on initial income is -0.1 using FD-IV and 0 using
FE-IV. When the coefficients are standardized, the results indicate that a 1 standard
deviation increase in openness increases the growth rate by 0.658 standard deviations
using FD-IV and by 0.384 standard deviations using FE-IV. A 1 standard deviation
increase in income decreases the growth rate by approximately 4.046 standard deviations
while a 1 standard deviation increase in investment increases the growth rate by 0.564
standard deviations using FD-IV and 0.624 standard deviations using FE-IV. However, as
table 2.6 illustrates once lagged values are used as instruments for initial income,
investment share and government spending, all the variables become insignificant using

both FE-IV and F D-IV.

79

Table 2.7 shows the results of running separate regressions for each time period
using instruments for initial income, investment and government spending, and using the
fitted values as instruments for initial income. The results are almost identical to the ones
obtained using FE and FD without taking the endogeneity of lagged income into
consideration. The coefficient on initial openness is 0.03 using FE and 0.04 using FD.
Initial openness, initial income, and investment are all significant at the 1% level. The
coefficient on initial income is approximately -0.2 while the coefficient on investment is
0.15. In this case the standardized coefficients show that a 1 standard deviation increase
in openness increases growth by approximately 0.492 standard deviations using F E-IV
and 0.68 standard deviations using FD—IV. A 1 standard deviation increase in initial
income decreases growth by around 7.95 standard deviations while a 1 standard deviation
increase in investment increases growth by around 0.4 standard deviations. So once
again, the relative impacts of openness and investment are similar.

Table 2.8 shows the results of re-rrmning the regressions using initial values for
all variables and income prior to 1985 as instruments for lagged income when FE and FD
are used. All the panel data estimation techniques show that openness is positive and
statistically significant at the 1% level. Once the fixed effect is taken into consideration
the coefficient is 0.04. The coefficient on initial income is only significant using FD-IV.
The elasticity in this case is -0.1. Using FE-IV all other explanatory variables are
insignificant.

The regressions in tables 4-6 are re-run adding the interaction between openness
and initial income. The results are generally consistent with the previous findings.

Overall, openness, initial income and investment are significant at the 1% level. The

80

coefficient on the interaction term showed that the impact of openness on growth does in
fact depend on the level of income. Tables 2.9 and 2.10 illustrate that as the level of
income increases by 1 percentage point the impact of growth on openness decreases by
0.02 percentage points. Therefore the effect of openness on growth decreases as income
increases. The elasticity of growth with respect to income is still approximately -0.1%
and the coefficient on investment is still in the range 0.14 — 0.17. As table 2.11 illustrates
once again when instruments are used for income, investment, and government spending
nothing is significant at the 5% level.

The regressions in tables 4-8 are also re-run using the alternative measure of
openness described in section 4. In this case openness is the difference between the
amount of trade predicted by the gravity model and the amount of trade that actually
occurs. Two different scenarios are investigated. In the first case, trade is assumed to be
in differentiated products and in the second case trade is assumed to be in homogeneous
products. The subscript 1 refers to trade in differentiated products while the subscript 2
refers to trade in homogeneous products. Generally the results for the two different cases
are similar to each other and they are also similar to the results using openness measured
as trade percentage.

As tables 2.12 and 2.13 illustrate once again initial openness, initial income, and
investment are significant at the 1% level once FE and FD are used and without
controlling for the endogeneity of any of the variables. Once again the elasticity of
growth with respect to income is approximately -0.1, and the coefficient of openness is
approximately 0.03, while the coefficient on investment is consistently 0.2. As before

when only the endogeneity of income is taken into consideration, the three ‘usual

81

suspects’ are all significant at the 1% level using FD. As table 2.14 shows, FD implies
that a 1 percentage point increase in openness will lead to a 0.03 percentage point
increase in growth while a 1 percentage point increase in investment will lead to a 0.2
percentage point increase in growth. Yet again the elasticity of growth with respect to
income is -0. 1% . Not surprisingly, the beta coefficients imply that investment and
openness have a similar relative impact. As is the case when trade percentage is used to
measure openness, when instruments are used for income, investment and government
spending, nothing is significant.

Table 2.16 shows the results of using fitted values as instruments for income. In
this case both FE and FD show that initial income and initial openness are significant (at
the 1 % level in most cases). However the magnitudes of the coefficients of both
variables are slightly higher than they are in the previous cases. The coefficient on initial
openness is around 0.04 and the coefficient on initial income ranges from -0.3 to -0.2.
Investment is only significant at the 10% level and its coefficient is slightly lower (0.1).
The same general pattern of results is obtained when initial values are used for
investment and government spending and lagged values of income are used as
instruments for initial income. These results are in the appendix. The appendix also
shows the results of including the interaction between initial openness and initial income
using the alternative measure of openness. Once again the same pattern of results holds.
Overall the results are remarkably robust to estimation technique, measure of openness,
and the inclusion of the interaction term.

Table 2.17 shows the impact on the annual average growth rate of a 10 percentage

point mcmase in openness. To get the figures I use the values of the coefficients for

82

openness and the interaction term from table 10 as well as the values for initial income
from table 1. It implies that a 10 percentage point increase in openness will increase the
annual average growth rate by 0.32 percentage points for the entire sample using FD-IV.
Using FE-IV the effect on the full sample is insignificant. Both estimation techniques
indicate that as the income level decreases, the impact of openness on growth increases.
Using F E-IV the effect of openness on growth becomes significant only once the income
grouping becomes lower middle class while using FD—IV openness has a significant
impact on growth even for the upper middle income countries. In both cases the
magnitude of the impact increases as the income level decreases. The results imply that a
10 percentage point increase in openness will increase the annual average growth rate by
approximately 0.65 percentage points among low income countries.

These findings appear to contradict the findings of DeJong and Ripoll (2006).
They find that the marginal impact of tariffs on growth is decreasing in initial income.
Specifically, using the results of the GM Difference estimator DeJong and Ripoll argue
that a 10 percentage point increase in tariff rates increases average annual growth by
0.181% among low income countries, decreases growth by 0.353% among lower middle
income countries, by 0.839% among upper middle income countries and by 0.123%
among high income countries. Using the GM Systems estimator they argue that a 10
percentage point increase in tariff rates increases average annual growth by 1.309%
among low income countries, by 0.366% among lower middle income countries,
decreases growth by 0.492% among upper middle income countries and by 1.182%

among high income countries.

83

However it should be noted that of the results mentioned above only the impact of
tariffs on growth for low income countries using the GM Systems estimator is
significant at the 5% level or lower. For all other effects the p-value is > 0.05. In some
cases the p-values are as high as 0.8 or 0.9. When DeJong and Ripoll run the regression
with the interaction term between tariffs and initial income the coefficients on both the
tariff rate and the interaction term are insignificant at the 5% level using all 4 estimation
techniques. Only using GMM Systems estimation are these coefficients significant at the
10% level. This might lead one to argue that DeJong and Ripoll actually find that tariffs
have no effect on openness. The strength of their conclusions is therefore questionable.

Mankiw, Romer and Weil (1992) show that the annualized rate of convergence A

solves:
k1=—[1—exp<-xj)1 (9)
Using the value for k1 (the coefficient of lagged income) of -0.1, which is obtained in the

vast majority of regressions, the implied 1. in this paper is 0.02. Mankiw, Romer and Weil
(1992), Barro and Sala-i-Martin (1995) and Barro and Sala-i-Martin (2003) also find
convergence rates of 0.02. This means that only 2% of the gap between a country’s initial
income level and its long run or steady state value is eliminated each year. Mankiw,
Romer and Weil argue that a convergence rate of 2% means that the economy reaches
halfway to steady state in around 35 years. Higgins, Levy and Young (2006) find
convergence rates, for all counties within the United States, of 0.024% using OLS and
0.066% using 3SLS. DeJong and Ripoll (2006) find convergence rates varying from
0.99% to 2% using OLS, panel SUR, and GMM Systems estimation. However, using

GMM Difference estimation they find a much higher convergence rate of 9%.

84

This paper finds support for conditional convergence. Once differences in
investment, government spending, population growth, and openness are taken into
consideration, lower income countries appear to grow faster than higher income
countries. Moreover, the rate of convergence found in this paper is similar to the values

found in other well-known papers investigating the issue of convergence.

2.6 Discussion and Conclusions

This paper provides evidence that initial income, investment rates and initial
openness have an important effect on the rate at which countries grow. The significance
and sign of the estimate of initial income is supportive of the theory of conditional
convergence which is based on the augmented Solow model. An approximate annual
convergence rate of 2% is found. Economic theory has long recognized investment as an
engine of growth. Table 2.2 illustrates that the high income countries have the highest
investment rates while the low income countries have the lowest. This implies correlation
between investment and growth. The results of the regression analysis suggest a causal
relationship between investment and growth.

The goal of this paper is to examine the relationship between openness and
growth. Openness is found to have a positive impact on growth especially for low income
countries. The results suggest that for these countries a 10 percentage point increase in
openness will increase the annual average growth rate by 0.65 percentage points. This
increase while plausible does appear to be a bit large. This may be because it includes the
impact of human capital, which is not explicitly controlled for. Human capital is excluded

as an explanatory variable because one way in which openness may influence growth is

85

through its impact on human capital. It may be the case that openness increases human
capital through the transfer of knowledge and technology, and human capital increases
growth. The impact of openness on growth found in this paper is not independent of the
impact of human capital investment.

The findings of this paper are important from a policy perspective. I apply
relatively advanced panel data estimation techniques to the traditional ‘Barro’ regression
and not only do I still find evidence of convergence, and evidence that openness has an
impact on growth, I also find evidence that the impact of openness varies across income
groups. It seems that openness is not important as an engine of growth for high income
countries. Openness may provide opportunities for middle and low income countries to
grow at a faster rate through knowledge transfers, investment in research and
development, increase in product quality caused by competition, among others. However
it appears that for high income countries emphasis on openness as a growth policy tool is
ill-advised.

This is consistent with life cycle theories of product development. These theories
state that innovation occurs in the high income developed countries of the North while
imitation takes place in the low-wage developing countries of the South. This implies that
middle and low income countries have the potential to benefit the most from openness.
This is because openness provides these countries with access to the technologies
developed in high income countries. Given that transfers of technologies, ideas, and
processes occur, openness leads to growth for low and middle income countries. High
income countries may be advised to focus on the development of technology and cutting

edge products as a means for achieving growth.

86

The results of this paper are remarkably robust to estimation technique and
measure of openness used. This gives a certain level of confidence in the findings. It
seems that once investment rates, population growth, initial income levels, and
government spending are controlled for, openness can increase growth among low
income countries. However, the results do not imply that openness is a ‘magic pill’ that
will automatically confer developed country status on low income countries. Instead they
imply that an open economy policy combined with an emphasis on investment growth,
and stable economic policies can result in an increase in growth. They imply that trade
restricting policies may not be beneficial to low income countries in the long run. It may
be better for these countries to focus on taking advantage of the opportunities provided by
trade by investing in research and development and improvement of product quality so
that they will be more competitive on the international market. Trade provides an
opportunity for less developed economies to learn from more developed countries
through technology transfers. Trade helps to make the processes, ideas and human capital
in rich countries available to lower income countries. While I do find evidence of a
positive impact of openness on growth, I also find evidence of a positive impact of
investment on growth. This implies that any open economy policy should be
accompanied by an emphasis on investment in order to assist countries in achieving a

faster growth rate.

87

Table 2.1 — Summary Statistics

 

 

 

 

 

 

 

Sample Split Variable Observations Mean Std. Dev.
Full Initial Income 107 6198 5778
Growth 0.374 0.513
High Income Initial Income 22 16154 2992
Growth 0.578 0.346
Upper Middle Initial Income 18 7231 2375
Growth 0.506 0.585
Lower Middle Initial Income 36 3892 1563
Growth 0.331 0.480
Low Initial Income 31 1210 485
Growth 0.202 0.557

 

 

 

 

 

Notes: Initial income is measured as PPP adjusted constant GDP per capita in 1981 using 1996 as the base
year. Growth is the annual average growth rate of income from 1981-2000.

Table 2.2 — Summary Statistics

 

 

 

 

 

 

 

Sample Split Variable Observations Mean Std. Dev.
Full Trade 107 0.675 0.463
Investment 0.169 0.081
Government 0.225 0.123
High Income Trade 22 0.587 0.434
Investment 0.230 0.045
Government 0. 1 55 0.092
Upper Middle Trade 18 0.831 0.668
Investment 0.216 0.084
Government 0.242 0. l 36
Lower Middle Trade 36 0.709 0.410
Investment 0. 165 0.066
Government 0.240 0. 1 29
Low Trade 31 0.608 0.389
Investment 0. 1 01 0.062
Government 0.246 0.1 13

 

 

 

 

 

 

 

Notes: Trade is measured as (exports + imports)/GDP. Investment and government spending are also
measured as ratios with respect to GDP.

88

Table 2.3 - Correlations

 

 

 

 

 

 

 

Sample Split Trade and Growth Initial Income and Growth
Full 0.136 0.199

High Income 0.653 -0.412

Upper Middle 0.228 -0.299

Lower Middle 0.282 0.019

Low -0.366 0.106

 

 

 

 

Notes: Correlations are between trade percentage and the annual average growth rate of income from 1981-
2000 and the GDP per capita in 1981 and the average annual growth rate from 1981-2000

4-

 

Growth

log (Area)

Investment

Government -0.001 83

1

error

‘denotes significance at the 10% level
" denotes significance at the 5% level
""denotes significance at the 1% level

-0.01336

  

-0.03885

Table 2.5 — IV Estimation Results (T = 3) — using instruments for y only

 

 

 

 

 

 

 

Regressor FE-IV F D-IV
Initial Openness "' *0.02429 * "' *0.04164
(0.00988) (0.00729)
log (Initial Income) -0.0089 * * * -0. 1006
(0.0551) (0.0272)
Population Growth 0.2782 0.3429
(0.3259) (0.2253)
Investment ***0.2187 ***0.19778
(0.0533) (0.0364)
Government -0.04439 *-0.05 831
(0.05128) (0.03158)

 

 

 

Notes: standard errors are in parentheses
‘denotes significance at the 10% level
"“' denotes significance at the 5% level
""denotes significance at the 1% level

89

 

Table 2.6 — IV Estimation — using instruments for y, I and G

 

 

 

 

 

 

 

Regressor FE-IV F D-IV
Initial Openness 0.07292 0.09354
(0.05523) (0.21 103)
log (Initial Income) -0.0528 -0.1350
(0.0904) (0.2252)
Population Growth A 0.2148 0.3147
(0.4830) (1 .999)
Investment -0.16597 -0.32372
(0.52329) (2.72143)
Government -0.24416 -0.27837
(0.22855) (0.54198)

 

 

 

Notes: standard errors are in parentheses
‘denotes significance at the 10% level
** denotes significance at the 5% level
*"denotes significance at the 1% level

Table 2.7 — Estimation Results - using different reduced forms for each time period

 

 

 

 

 

 

 

Regressor FE-IV FD-IV
Initial Openness ***0.031 12 ***0.04299
(0.00933) (0.01135)
log (Initial Income) ** *-0. 1534 * * *-0.2418
(0.0402) (0.0653)
Population Growth 0.4144 0.4782
(0.3108) (0.3063)
Investment * * *0.1465 ** *0.14765
(0.04817) (0.05685)
Government 0.04138 0.03538
(0.04474) (0.05601)

 

 

 

Notes: standard errors are in parentheses
*denotes significance at the 10% level
*"‘ denotes significance at the 5% level
"*denotes significance at the 1% level

90

 

 

Table 2.8 — Estimation Results with Initial Values and using instruments for y only

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Regressor Cross-Sectional POLS FE-IV FD-IV
Openness 0.00107 ***0.01606 ***0.03643 ***0.04036
(0.0052) (0.00552) (0.01065) (0.0096)
log (Income) 0.0028 -0.0012 -0.0488 ***-0.1375
(0.0019) (0.0020) (0.0674) (0.0332)
log ***0.0050 l***0.0041 -0.0048 "-0.1 101
(Population) (0.0017) (0.0012) (0.0698) (0.0477)
log (Area) ***-0.0051 ***-0.0028
(0.0010) (0.0009)
Investment *0.04894 ***0.12206 0.07927 ***0.10926
(0.02714) (0.03372) (0.05122) (0.03531)
Government -0.001 15 -0.00367 -0.02826 **-0.05839
(0.01622) (0.01343) (0.03253) (0.02391)
Notes: the Huber/White/ Sandwich standard error estimates are in parentheses
*denotes significance at the 10% level
" denotes significance at the 5% level
"‘denotes significance at the 1% level
Table 2.9 — Estimation of equation 1 with interaction term
Regressor Cross- POLS Fixed Effects First-
Sectional Difference
Initial Openness ***-0.07835 0.03629 ***0.18603 II“"0.16325
(0.02443) (0.03374) (0.03764) (0.07266)
Log (Initial Income) ***-0.01 164 -0.00501 ***-0.05082 ***-0.1 1563
(0.00306) (0.00273) (0.00805) (0.01044)
Openness*log(Income) ***0.00996 -0.00223 ***-0.01814 *-0.0148
(0.00284) (0.00359) (0.00441) (0.00849)
Population Growth * **-0.85799 “-0.663 77 *0.43595 *0.40966
(0.18169 (0.27754) (0.23569) (0.23642)
Log (Area) 00003]
(0.00084)
Investment ***0.13111 ***0.15806 ***0.15158 ***0.17328
(0.02838) (0.02697) (0.0003 1) (0.0407)
Government 0.00214 -0.01543 *-0.0452 *-0.05697
(0.01491) (0.01336) (0.02502) (0.03403)

 

 

 

 

 

Notes: the Huber/White/ Sandwich standard error estimates are in parentheses
*denotes significance at the 10% level

** denotes significance at the 5% level

*"denotes significance at the 1% level

91

 

 

Table 2.10 — IV Estimation using instruments for y only with interaction

 

 

 

 

 

 

 

 

 

Regressor FE-IV FD-IV
Intial Openness * "0.29947 * "0.19443
(0.08321) (0.05256)
Log (Initial Income) 0.04280 *"-0.08014
(0.06844) (0.02955)
Openness*log (Income) ***-0.03262 ***-0.0186
(0.00999) (0.00626)
Population Growth 0.50062 *0.39117
(0.34084) (0.22473)
Investment ***0.14248 ***0.16965
(0.0532) (0.03727)
Government -0.06654 *-0.0602
(0.05659) (0.0316)

 

 

Notes: standard errors are in parentheses
‘denotes significance at the 10% level
" denotes significance at the 5% level

"*denotes significance at the 1% level

Table 2.11 — IV Estimation using instruments for y, I and G with interaction

 

 

 

 

 

 

 

 

 

 

Regressor FE-IV FD-IV
Initial Openness *0.5679 0.45997
(0.3065) (1.26222)
Log (Initial Income) -0.02933 -0.10658
(0.09736) (0.20474)
Openness*log (Income) *-0.05612 -0.04178
(0.02835) (0.11418)
Population Growth 0.80493 0.47086
(0.70375) (2.89942)
Investment -0.53513 -0.63 825
(0.76411) (4.23981)
Government -0. 15075 -0.34075
(0.26704) (0.53806)

 

Notes: standard errors are in parentheses
*denotes significance at the 10% level
*"' denotes significance at the 5% level

*"denotes significance at the 1% level

92

 

 

Table 2.12 — Estimation of equation 1 using alternative measure of openness

 

 

 

 

 

 

 

 

 

 

 

 

Regressor Cross-Sectional POLS Fixed Effects First-Difference
Grav Open] a 0.00038 0.00252 ***0.02879 ***0.03 565
- (0.00197) (0.00177) (0.00528) (0.00882)
Log (Initial * *-0.00548 * *-0.00565 "‘ ”-0.07278 * * *-0. 12483
Income) (0.00228) (0.00232) (0.00701) (0.01091)
Population ** *-0.84594 ' *-0.523 77 0.29978 0.37872
Growth (0.21342) (0.28253) (0.23961) (0.24383)
Log (Area) *-0.00132 ** *-0.00187
(0.00072) (0.00051)
Investment ***0.12255 ***0.18164 ***0.20189 ***0.21092
(0.02986) (0.03575) (0.02908) (0.05122)
Government -0.00151 -0.00606 -0.03 729 -0.05064
(0.01548) (0.01363) (0.02564) (0.03327)

 

Notes: aGrav_0penl represents the difference between trade predicted by the gravity model and actual
trade assuming that trade is in differentiated products.
the Huber/White/ Sandwich standard error estimates are in parentheses

I"denotes significance at the 10% level
** denotes significance at the 5% level

*"denotes significance at the 1% level

Table 2.13 - Estimation of equation 1 using alternative measure of openness

 

 

 

 

 

 

 

 

 

 

 

 

Regressor Cross-Sectional POLS Fixed Effects F irst-Difference
Grav OpenZa 0.00296 ** *0.00706 * * *0.02931 * M0.03573
'— (0.00246) (0.0023881) (0.00530) (0.00882)
Log (Initial * *-0.00567 * **-0.00679 * * *-0.07703 * **-0. 12949
Income) (0.00223 (0.00239) (0.00713) (0.01097)
Population * * *-0.88232 * *-0.55055 0.32988 *0.39659
Growth (0.21802) (0.27486) (0.23912) (0.24041)
Log (Area) * *-0.00155 * * *-0.00248
(0.00073) (0.00055)
Investment ***0.11804 ***0.17175 ***0.20099 ***0.21165
(0.03052) (0.03647) (0.02905) (0.05131)
Government 0.00159 -0.00272 -0.03 78 -0.05226
(0.01568) (0.01355) (0.02562) (0.03352)

 

Notes: aGrav_Open2 represents the difference between trade predicted by the gravity model and actual

trade assuming that trade is in homogeneous products.

the Huber/White/Sandwich standard error estimates are in parentheses

*denotes significance at the 10% level

** denotes significance at the 5% level
"*denotes significance at the 1% level

93

 

 

Table 2.14 — Results using Alternative Openness Measures and instruments for y only

 

 

 

 

 

 

 

 

 

 

 

Regressor FE_IV13 FE-IVZb FD-IVl FD-IV2
Initial M0.02174 0.01488 ***0.03565 ***0.03468
Openness (0.00869) (0.01152) (0.00624) (0.00642)
Log (Initial -0.01 105 -0.00938 *"'*-0.10926 ***-0.11246
Income) (0.05579) . (0.05984) (0.02606) (0.02579)
Population 0.31218 0.32892 0.34503 0.36077
Growth (0.32245) (0.32762) (0.22468) (0.22423)
Investment ***0.22729 ***0.24454 ***0.21823 ***0.22206
(0.05271) (0.06006) (0.03684) (0.03794)
Government -0.0466 -0.04384 *-0.05216 0.05325
(0.05024) (0.04985) (0.03136) (0.03136)

 

Notes: a1 represents trade in differentiated products.
b2 represents trade in homogeneous products

standard errors are in parentheses
*denotes significance at the 10% level
*"‘ denotes significance at the 5% level

"*denotes significance at the 1% level

Table 2.15 — Results using Alternative Openness Measures and instruments for y, I & G

 

 

 

 

 

 

 

 

 

 

 

Regressor lemma FE-IVZb FD-IVl FD-IV2
Initial 0.05407 0.05207 -0.10517 -0.16540
Openness (0.03660) (0.04051) (0.58634) (1.12136)
Log (Initial -0.04688 -0.05843 0.08066 0.15203
Income) (0.08480) (0.09269) (0.67593) (1.33141)
Population 0.29935 0.36554 -0.90295 -1 .1 1286
Growth (0.50614) (0.52650) (2.54671) (4.56906)
Investment -0.09105 -o.10491 2.18672 2.83551
(0.45774) (0.46748) (7.37938) (13.57584)
Government -0.25847 -0.2456 -0.26283 .0.1255
(0.22142) (0.22365) (1.39469) (2.52184)

 

Notes: 8‘1 represents trade in differentiated products.

b .
2 represents trade In homogeneous products

standard errors are in parentheses
*denotes significance at the 10% level
** denotes significance at the 5% level

I""“"denotes significance at the 1% level

94

 

 

Table 2.16 — Estimation Results — using different reduced forms for each time period

 

 

 

 

 

 

 

 

 

 

 

Regressor FE-IV 1 a FE-IVZb FD-IV 1 FD-IVZ
Initial ***0.03627 ***0.04167 ***0.04479 ** *0.05328
Openness (0.01039) (0.01 I 70) (0.01273) (0.01336)
Log (Initial I"""'-0.20754 I""""-0.l9593 1""‘-0.32881 ***-0.28581
Income) (0.07841) (0.05922) (0.13972) (0.09047)
Population 0.50124 0.50689 0.52247 0.54241
Growth (0.36607) (0.34234) (0.39365) (0.33780)
lnvestrnent I"0.11867 *0.11414 0.12248 *0.11942
(0.06482) (0.06122) (0.08225) (0.07162)
Government 0.06417 0.04933 0.07074 0.04206
(0.0625) (0.051 17) (0.08665) (0.06356)

 

a . . .
Notes: 1 represents trade 1n differentlated products.

b2 represents trade in homogeneous products
standard errors are in parentheses

*denotes significance at the 10% level

" denotes significance at the 5% level
"*denotes significance at the 1% level

Table 2.17 — Effect on Average Growth Rate of a 10 Percentage point Increase in
Openness

 

 

 

 

 

 

 

 

 

Sample Split FE FD
Full Sample 0.14614 ***0.31997
(0.1 l 190) (0.07701)
High Income -0. 16634 0.14179
(0.17473) (0.1 1234)
Upper Middle 0.09591 ***0.291323
(0.] 1909) (0.08076)
Lower Middle * * *0.29784 * "' *0.40647
(0.10187) (0.07264)
Low ***0.67919 ***0.62393
(0.15506) (0.10539)

 

 

 

Notes: Using the coefficient estimates for openness and openness-income interaction from table 10, this
table shows the impact on growth (measured in annual percentage terms) of a lO-percentage-point increase
in trade. The values for the log of income were taken from table 1.

"*denotes significance at the 1% level

95

Figure 2.1 - Relationship Between Initial Trade and Growth

Initial Trade

 

Initial Income

96

Testing the Monetary Model
of Exchange Rate
Determination: Evidence

from Latin America

97

3.1 Introduction

Since 1980 the monetary model of exchange rate determination has been the
dominant one in the field of international finance. The appeal of the model lies in the
convincing intuition on which it is based. Within this model, the exchange rate is viewed
as the value of one country’s money supply against another. Since its introduction several
papers have been written in an attempt to determine its validity.

Initially, there was some support for this model. However, subsequent papers
were able to identify weaknesses in the evidence supporting the model. One of the most
damaging criticisms of the initial findings was the discovery of the existence of a unit
root in the exchange rate sequence. This led to studies trying to discover whether or not a
cointegrating relationship exists between the exchange rate and the monetary
fundamentals. At first, the consensus was that the exchange rate and the fundamentals in
the model are not cointegrated. However more recent papers, using different techniques
to increase the power of the cointegration tests, suggest that the model may be
appropriate as a long run phenomenon. That is, they imply that there may in fact be a
cointegrating relationship between the exchange rate and the monetary fundamentals.

The papers that have previously been written testing the validity of the monetary
model of exchange rate determination have focused on its applicability to industrialized
first world countries. The goal of this paper is to determine whether or not this model is
appropriate for describing exchange rate movements for Latin American countries.
Countries in Latin America are usually characterized by exchange rate instability.

Exchange rate regimes are usually classified as fixed or flexible. However, in the

real world there are a variety of different exchange rate systems which involve varying

98

degrees of government participation in foreign exchange markets. According to a 2004
USAID Fiscal Reform Project Report exchange rate regimes can be classified as follows:
“fixed—rate regimes”, “intermediate regimes” and “flexible regimes”. “Fixed-rate
regimes” include currency unions, dollarized regimes, currency boards and fixed pegs.
“Intermediate regimes” include horizontal bands, crawling pegs and crawling bands,
while “flexible regimes” include managed and independent floats.

Over the sample period, the 15 countries analyzed have experienced a variety of
different exchange rate systems from fixed exchange rates to crawling bands to managed
floating to free floating. Some of these countries have experienced periods of fiequent
devaluations and currency crises. For example, Mexico experienced currency crises in
1976, 1982 and 1986. Eight out of the fifteen countries experienced a change in the actual
currency used by the economy over the sample period. For example, in September 1975
in Chile the escudo was replaced by the Chilean peso at a rate of 1 peso to 1000 escudos.
These changes in currency were usually driven by high inflation.

The issue being investigated in this paper is whether or not, in these countries
characterized by exchange rate volatility, the value of the exchange rate can be described
by relative money supplies and relative output levels. For these countries that are usually
plagued by high inflation and whose currencies are not heavily traded on the world
market, is there a cointegrating relationship between the exchange rate and the monetary
fundamentals?

Economic theory tells us that despite the volatility in the exchange rate
movements, there should be a long run relationship between the exchange rate between

two countries and their relative money supplies and relative output levels. The monetary

99

approach implies that countries with relatively restrictive monetary policies and/or
relatively rapid increases in growth will experience an appreciation in their exchange
rate. The monetary model is founded on the quantity theory of money which states that if
a country increases its money supply faster than its growth rate of gross domestic product
(GDP), then its prices will rise. Higher prices will then lead to a depreciation of the
currency as the country’s goods become less competitive in international markets.
Despite the fact that Latin American countries usually experience high exchange rate
volatility and high inflation rates and in spite of the fact that Latin American currencies
are not widely traded in international foreign exchange markets, the above analysis
should still hold.

The paper will proceed as follows: Section 2 provides an introduction to the
monetary model. Section 3 presents the literature review. Section 4 outlines the
methodology which will be followed in this paper. Section 5 provides a description of the
data. Section 6 summarizes the results and section 7 highlights the main contributions of

this paper in the form of a conclusion.

3.2 The Monetam Model

The monetary model of exchange rate determination is based on three
relationships. These are the uncovered interest rate parity (U IRP) condition, purchasing
power parity (PPP), and the money demand function for real balances.

The UIRP condition describes the relationship between domestic and foreign
interest rates and the current and expected future spot exchange rate. It can be expressed

as:

100

e 0*
’t"t = EtASt+l (1)
where it = the domestic interest rate

3, = log of the nominal spot exchange rate (domestic cost of foreign

currency)
* denotes the foreign country
This equation implies that the interest rate differential is equal to the expected rate of
change of the currency. The country with the higher interest rate has the currency that is
expected to depreciate.

PPP is an extension of the law of one price to a basket of goods. The law of one
price states that the price of a good in one country should be equal to the price of the
same good in another country once the two prices are expressed in terms of the same
currency. The extension to a basket of goods is usually done using the consumer price

index or wholesale price index. PPP can be expressed as:
4=m-d (a
where p, = log of the price level in the domestic country
The money demand function can be expressed as:
mt ‘ Pt = 01}? — 02'} (3)
Where m, = log of money supply
y, = log of real output
a1 and a2 are both assumed to be positive. A similar equation can be written for the

foreign country:

101

* * *

mt ‘Pt = 01y: "’2": (4)
Note that the parameters of the money demand function are the same for both home and
foreign. This assumption is not a necessary one but it simplifies the analysis.

Combining equations (1) — (4) we get the monetary model. The first step is to
solve equations (3) and (4) for p, and p: , respectively, and then substitute the resulting
expressions into equation (2). This gives:

st=04—mI)+a2(il-iI)—a1(yt—y}'> (5)
Equation (1) is then substituted into the above expression to give:

s,=bo+bl(m,—mf)+bz(y,-yf)+u, t=l,....,T (6)
The error term comes from the uncovered interest rate parity condition which states that

the currency with the higher interest is expected to have the currency depreciation. The

expectation in equation (1) gives rise to the error term.

Weexpect b1=1 and b2 <0.

3.3 Literature Review

One of the earliest papers testing the validity of the monetary model is Frankel

(1979). Frankel estimates a model of the form:
It * . .III *
s, = c0 +(mt —m, )+c2(yt —yt)+C3(r,—1,)+C4(7rt - flit ) + at (7)
where wt = rate of inflation in domestic country

He anticipates c2 < 0, C3 < 0, and C4> 0. Note that this equation is slightly different from

the model presented in equation (6). This is because Frankel derives and estimates the

102

real interest rate differential version of the model. Frankel, rather than accept the UIRP,
tests the validity of it. He concludes that it holds as a long run phenomenon.

He estimates the model using OLS and the Cochrane-Orcutt (C.O.) procedure. He
concludes that when the above equation is estimated for the mark/dollar rate from July
l974-February 1978, the evidence is supportive of the monetary model. The signs of the
coefficients were as he had hypothesized.

One of the main objections to Frankel’s conclusion is the fact that when he used

the CO. procedure, his estimate of p was very close to one (0.99). This raises the

question of non-stationarity of st and casts doubt on his findings.

Subsequent papers were therefore focused on determining whether or not there
exists a long run relationship between the nominal exchange rate and its fundamentals,
that is, whether or not the exchange rate and the monetary fundamentals are cointegrated.
The findings have been inconclusive. MacDonald and Taylor (1993) find evidence of
cointegration between the exchange rate and its fundamentals in a forward looking
monetary exchange rate model for the United States (U .S.) $-DM over the period 1976-
1990. However, Sarantis (1994) uses the same approach to conclude the opposite with
data for the pound-sterling exchange rates of the US, Germany, Japan and France for the
years 1973-1990. Other studies that have found little evidence of a long-run relationship
using data from the post-Bretton Woods float include Meese (1986), Baillie and Selover
(1987), McNown and Wallace (1989), Baillie and Pecchenino (1991) and Sarantis
(1994)

Groen (2000) argues that previous papers have failed to find evidence of

cointegration because of the availability of a short time span of data for the post Bretton

103

Woods float. Engle and Granger (1987) show that the power of cointegration tests
increases with the time span of data used. Two approaches have been taken to increase
the power of cointegration tests. The first approach involves using panel data, while the
second approach involves using a long time span of data.

The first approach was utilized by Groen in his 2000 paper. He uses a panel
version of the Engle and Granger cointegration test with quarterly data over the period
l973:1-1994:4 for monetary fundamentals and bilateral exchange rates with respect to
both the US. $ and the German DM. The sample includes the following 14
industrialized, first world countries: Australia, Austria, Canada, Finland, France,
Germany, Italy, Japan, the Netherlands, Norway, Spain, Sweden, Switzerland and the
United Kingdom.

Groen starts by estimating the following model using OLS on the pooled time
series data:

4

* *
Sit = b0,i + Z xisdis + b1 (mit - mil ) + b2 (yit - yit) + Hit (8)
s=2

where i=1,...,N
1 =1,..., T
dis = seasonal dummies
The asterisks in the above equation refer to both the US. and Germany since he uses
them both alternately as numeraire currencies. He anticipates bl = l, b2< 0 and pit
stationary with mean zero.

In his second stage he tests the estimated residuals of (8) using Levin and Lin’s

(1992) panel unit root tests. That is, he applies a panel augmented Dickey-Fuller

104

regression to the residuals and conducts a one-sided t-test for the coefficient of the lagged
residuals. The results for the 14 country panel were supportive of the monetary model for
both nurneraire countries. Groen was able to reject the null hypothesis at the 5% level or
lower. The cointegrating vector on the US. $ panel was (b1 ,b2) = (0.66, -1.34) and for
the DM panel it was (b1 ,b2) = (0.92, -1.82). The residuals were found to be stationary.
Groen concludes that on a pooled time series level there is cointegration between the
exchange rate and the fundamentals of the monetary model.

Rapach and Wohar (2002) use the second approach to increase the power of the
cointegration test. They test the long run monetary model using annual data for 14
industrialized countries spanning the late 19th or early 20th century to the late 20th
century. The countries included in their sample are Australia, Belgium, Canada,
Denmark, Finland, France, Italy, the Netherlands, Norway, Portugal, Spain, Sweden,
Switzerland, and the United Kingdom. They find support for a simple form of the long
mm monetary model in over half of the countries considered. For those countries, they
estimate a vector error-correction model to uncover the short run dynamics.

They start by examining the integration properties of the three series. They then

estimate equation (6). They also estimate a simpler form of the model by imposing the

restrictions bl = 1 and b2 = -1. To discover how long run equilibrium is restored between

the nominal exchange rates and the monetary frmdamentals, they estimate a bivariate

vector error-correction model in st and ft where ft = (m: — m, ) — (y; — yt).

Rapach and Wohar point out that the panel cointegration test in Groen (2000)

requires one to accept that the monetary model is appropriate for each member of the

105

entire panel. However, in their paper they find that the monetary model is inappropriate
for some of the countries that were included in Groen’s panel. They therefbre conclude
that while there is some support for the long-run monetary model, this support may have
been overstated by Groen.

Another approach that authors have used to test the validity of the monetary
model involves evaluating its forecasting performance. This work tests the hypothesis
that the foreign exchange market is not efficient and is therefore predictable. Meese and
Rogoff, in their seminal 1983 paper, show that out-of-sample forecasts based on the
monetary model do not beat the forecasts generated by a random walk for the US. dollar
rates of Germany, Japan and the United Kingdom for the period 1976—2001. However
Mark (1995) and Chinn and Meese (1995) find evidence that error correction terms,
based on monetary models, do accurately forecast US. $ exchange rates for countries
including Germany, Canada, and Japan within a time span of 1973-1991. On the other
hand, Groen (1999), in doing further analysis of the results of Mark and Chinn and
Meese, observes a break down in forecasting performance when he uses data for a longer

time period (1973-1994).

3.4 Methodology

The goal of this paper is to determine the relevance of the monetary model for
Latin American countries. I start by using OLS to estimate the model described in
equation (6). I then test each of the series to determine their integration properties. The

next step is to determine whether or not there exists a long run relationship between the

exchange rates and their fundamentals.

106

I conduct Engle and Granger cointegration tests for each country within the
sample. Then, following Groen (2000) I conduct a panel version of the Engle and
Granger cointegration test. I use fixed effects on the pooled time series data to estimate

the following model:

1' III
Sit =b10+b1(mu-mu)+b2(y“—yu)+ult , i=1,....,N, t=1,....,T (9)
Then I apply newly developed panel cointegration techniques to the residuals. The results
from the two approaches are compared to determine the applicability of the monetary

model to exchange rates in Latin America.

Panel Data Cointegration Testing

In recent years, the area of panel data cointegration has received a great deal of
attention fi'om economists. Several articles have explored the concept of extending
cointegration testing to panel data7. The main approach that has been taken is to apply
panel data unit root testing to the estimated residuals from a first stage regression.
According to Pedroni (1999), in panels with homogeneous cointegrating vectors, if the
regressors are assumed to be strictly exogenous, raw panel data unit root tests can be
applied to the estimated residuals. Pedroni (1999) states “an interesting special case holds
such that residual based tests for the null of no cointegration have distributions that are
asymptotically equivalent to raw panel unit root tests if and only if the regressors are
exogenous”. Therefore following Groen (2000), I apply Levin and Lin’s unit root test to
the estimated residuals of (6). However, while Groen uses the 1992 version of Levin and

Lin’s test, I use the revised version in Levin, Lin and Chu (2002). The tests suggested in

 

7 For a comprehensive summary see Banerjee (1999)

107

Levin et a1 (2002) are more relevant for panels of moderate size and are therefore more
appropriate for me.

Levin et al (2002) test the null hypothesis that each individual time series contains
a unit root against the alternative hypothesis that each time series is stationary. As both N
and T approach infinity, the test statistic has a limiting normal distribution. However,
Monte Carlo simulations showed that the normal distribution provides a good
approximation to the empirical distribution of the test statistic in small samples. Levin et
al also showed that panel cointegration tests have dramatically more power than separate
unit root tests on each individual time series.

The test statistic introduced in Levin et a1 (2002) is:

 

A
d
’61 = A (10)
STD(d)
where d is the estimate from the regression:
eit =dVit—1+uit (11)
and
~ A A
_ _ 8', ~ v- _1
811 —‘—l/\_ Vi, = “—16;— (12)
Sui Sui
A
where S“,- is the regression standard error in:
Pi
Ayit = ciyit—l + Z kiL A yit—L + “it (13)

i=1

108

A A

and ei, and vi,_1 are the residuals from the regressions of A ya and yi,_1 on A yt_ L (L =
1,..., pi), respectively. { yi, } is the stochastic process observed for the panel of
individuals, i =1,...,N and each contains t =1,...,T time series observations.

Pedroni (1999) introduces panel cointegration tests that allow the regressors to be
endogenous. This is particularly useful in this case because as discussed in the next
section, the sample period includes fixed exchange rate regimes. For the monetary model
to hold over that time span, one or both of the fundamentals must have adjusted to
maintain the fixed rate, implying that the explanatory variables are not exogenous. He

computes the asymptotic distributions and critical values for residual based tests of the

null of no cointegration in panel data allowing multiple regressors and including specific

fixed effects and time trends. In order to test the null hypothesis H0: “all of the

individuals of the panel are not cointegrated” against the alternative hypothesis H1: “all

of the individuals are cointegrated”, Pedroni starts with the model:
yi,t = “i + dz" + blixli,t + b2ix2i,t + + bMixMi,t + em (14)
for t: 1,...,T; i= 1,...,N; m =1,...,M
His panel t-statistic which he argues is analogous to the Levin and Lin panel unit root
statistic applied to the estimated residuals of a cointegrated regression is:
*2 A*2 _1 NT —2
ZtNT= (SN,TZZl{A12iei,—t 1) ZZZL ei,—-t lAei,t (15)

i=1t=1 i: It 1 11;

Where

109

l
SN,T=—ZS (16)
N . ,
l=l 1
A*2 1 T A*2
si=—Zv on
T .
i=1 1]
and
$
A
v it is the residual from:
A A A Ki A A A
em = hi ei,t_1+ Z I’ll-,k Aei,,_k +u,-,t (18)
k=1
and

A
e 1"; is the residual from (13)

A 2 1 T 3 2 ki s T A A
Ln.- =72cl,,+;2(1-m)2 one”... (19)
l
t=l s=l t=s+l
A
Where cm is the residual from:

M A A

Ayn = Z bmi Atom-1+ en (20)
m=1

Pedroni shows that the asymptotic distribution for the panel t-statistic can be expressed in

the form:

2fo dis/N

J;

—> N(0,1) (21)

 

110

E and v are firnctions of the moments of the underlying Brownian motion functions.
Pedroni calculates the values for E and v depending on the number of regressors and
whether or not constants or trends have been included in the regression.

I use the Pedroni test because, unlike the Levin and Lin test, it allows
endogeneity. This test is also appropriate for this case because it allows the number of
regressors to be greater than one. Pedroni assumes that the researcher has in mind a
particular relationship among the variables and is only interested in knowing whether or
not the variables are cointegrated. In my case the ‘particular relationship’ of course is the

monetary model.

3.5 Description of Saw

The sample consists of three different panels. The first panel includes 15 Latin
American countries from 1967 — 2000. Figures 3.1 — 3.15 show the movements in the
exchange rates for each of these countries in alphabetical order.“ The second panel
contains the same 15 Latin American countries over the period 1981 — 2000. The 15
Latin American countries included in the sample are Argentina, Bolivia, Brazil, Chile,
Colombia, Costa Rica, Ecuador, Guyana, Jamaica, Mexico, Nicaragua, Peru, Trinidad
and Tobago, Uruguay, and Venezuela. The third panel consists of 7 countries over a
longer time period, 1957 — 2000. These 7 countries — Brazil, Colombia, Costa Rica,
Ecuador, Nicaragua, Peru, and Venezuela were chosen based on data availability. It will
be interesting to see if using a longer time span but a smaller group of countries, will

change the results.

111

In the early 1980’s most Latin American countries started moving from
“intermediate regimes” such as crawling pegs to “flexible regimes” such as managed or
dirty floating. For example, figure 3.10 illustrates movements in the Mexican peso from
1967 - 2000. In the early 1980’s the peso starts to exhibit more movement than it had in
previous years. In fact, in 1976, the peso moved to a managed floating regime. As figure
3.4 illustrates, after the mid 1970’s, the Chilean peso started experiencing more volatility.

In 1975, the Chilean peso experienced a series of depreciations. In 1977, a policy
of pre-announced devaluation was adopted. The exchange rate was fixed in 1979.
However by 1982, the system had changed to a dirty float. In 1983 there was a sharp
depreciation (which coincided with a fall in real GDP), so a crawling peg was established
until 1985. From 1985-1997 the exchange rate was left to float within two predetermined
bands. The bands were widened to 12.5% in 1997. In that same year, Chile’s exchange
rate system was classified as managed floating. In 1999 the classification was changed to
independent floating. Chile’s experience from 1967 - 2000 is a good example of the
exchange rate regime volatility experienced by most Latin American countries over the
same period.

This volatility in the exchange rate movements may be difficult to see in the
graphs because of the changes to the currencies used over the sample period. Table 3.1
shows the coefficient of variation (cv) for all 15 countries over different time periods. It
illustrates that for all countries except Chile and Jamaica the cv over the last 20 years of
the sample is greater than the cv over the first 14 years of the sample. The most extreme
case is Uruguay which moves from a cv of 0.051 from 1967-1980 to a cv of 4364204

from 1981-2000. For most of the countries the cv over the last 14 years of the sample is

112

greater than the cv over the first 14 years of the sample. It should be noted that for
Jamaica the cv over the first 11 years of the sample is 0.129 implying that more volatility
started occurring around 1978. For Chile the cv over the first 6 years is 0.0298. The
exchange rate for that country started becoming more volatile around 1973.

Of the 15 countries in the sample, Argentina and Ecuador are the only ones that
by the end of the sample period had moved to “fixed-rate regimes”. In the 1990’s
Argentina fixed its peso to the US. dollar. However in 2002 Argentina unpegged the
peso from the US. dollar. In 2000 Ecuador made the conversion to dollarization. As
such, the US. dollar is currently the official currency of Ecuador.

The sample of Latin American countries includes three Caribbean territories —
Guyana, Jamaica and Trinidad and Tobago. As figures 3.8, 3.9, and 3.13 and table 3.1
illustrate they share the characteristic of experiencing greater movement in their
exchange rates from the 1980’s onwards. However over the period 1967 — 2000 these
countries experienced less volatility in their exchange rate regimes than the other Latin
American countries. All three countries currently have a floating exchange rate regime.
However the Trinidad and Tobago dollar (TTD) was not floated until 1993. In 1972, the
TTD was pegged to the pound sterling. In 1976, it was pegged to the US. dollar (U SD) at
a rate of 2.4 TTD to 1 USD. By 1985 the new peg was 3.6 TTD to 1 USD. In 1988, the
rate was devalued to 4.25 TTD to the USD.

As was mentioned in the introduction, for eight countries within the sample —
Argentina, Bolivia, Brazil, Chile, Mexico, Nicaragua, Peru, and Uruguay — there were
changes in the currency used over the sample period. As figures 3.1 to 3.4, 3.10 to 3.12,

and 3.14 illustrate these countries appear to have a currency valued at $0 to 1 US$ at the

113

beginning of the sample period. While these values are very small, none of them is
actually zero. In these countries high inflation rendered the currency useless at some
point or points in time, so the currency was changed. All exchange rates are expressed in
terms of the current currency of the country.

For example, in Argentina, the currency is currently the peso, (formally called the
peso convertible). However in 1967 the currency was the peso moneda nacional. In 1970
this was replaced by the peso ley at a rate of 1 peso ley to 100 pesos moneda nacional. In
1983 the peso argentino replaced the peso ley at a rate of 1 to 10,000. In 1985 the austral
replaced the peso argentine at a rate of l austral for 1000 pesos. In 1991 the peso replaced
the austral at a rate of 1 to 10,000. In the final analysis one peso would be worth
10,000,000,000,000 pesos moneda nacional today.

Colombia, Costa Rica, Ecuador, Guyana, Jamaica, Trinidad and Tobago and
Venezuela experienced no change in their currency over the sample period. For example,
the sucre was the cmrency of Ecuador between 1884 and 2000. Ecuador is currently a
dollarized economy. It will be interesting to see if the individual results will be different
for this group of countries. A priori I would not expect them to be. The relationship
between the exchange rate and the monetary fundamentals should still hold despite
changes in currency.

The data on exchange rates and money supply was obtained from the International
Monetary Fund’s International Financial Statistics website. All exchange rates are
expressed in terms of current currency per USS. The money supply data is the seasonally

adjusted sum of currency outside banks plus demand deposits. It is the measure of money

114

usually referred to as M1. The data on income was obtained from the Penn World Tables.

Income is measured as PPP adjusted constant GDP.

W
(a) Preliminary Analysis

The first step is to estimate the model using OLS. The results for panel 1 are
illustrated in table 3.2. Both the heteroskedasticity robust and the Newey-West standard
errors are reported. The two sets of standard errors are very similar implying that there
may not be much serial correlation present in the model. Using the first panel, for all 15

countries except Colombia b1 is statistically significant at the 1% level. In fact in all
cases the p-value is zero. For 13 countries b1 is approximately 1. However, for Trinidad
and Tobago bj is 0.46 and for Ecuador b] is -4.4. Formal statistical testing shows that the
hypothesis that bl is equal to 1 cannot be rejected at the 5% level for Chile, Costa Rica,

Jamaica, Mexico, and Venezuela using the first panel.
The coefficient of relative GDP is significant at the 1% level for 10 countries. In

some cases the p-value is zero. For Jamaica, b2 is significant at the 5% level. For
Mexico, b2 is significant at the 10% level, whereas b2 is not significant for Guyana,
Venezuela and Brazil. In 8 of the 10 countries for which b2 is significant, it is also
negative. However, for Colombia, Nicaragua and Ecuador b2 is positive.

Overall the results of the OLS estimation are quite supportive of the monetary

model. The general pattern is bl = l and b2 < 0. Table 3.2 also reports the coefficient

from an AR(l) regression of the residuals of each country. The evidence here is mixed.

115

The root ranges from as low as 0.14 for Chile to as high as 0.97 for Ecuador. Interestingly
the results for Chile are quite supportive of the monetary model while the results for
Ecuador are not. Colombia is another country for which the root is close to one and the
results are not supportive of the model. Casual inspection of the roots implies that there is
evidence of cointegration for some countries and no cointegration for others. More
formal tests of cointegration are conducted later.

The results for the second panel of countries show the same general pattern. The
coefficient of relative money supply is significant at the 1% level for every country
except Trinidad and Tobago and Colombia. In fact for each of these 14 countries except

Ecuador, the p-value is zero. For all 14 countries except Chile and Ecuador [)1 is
approximately 1. For Chile b1 is 1.6 and for Ecuador b1 is -1.5. Formal statistical testing

shows that the hypothesis that b1 is equal to 1 cannot be rejected at the 5% level only for

Jamaica and Venezuela. The coefficient of GDP is significant at the 1% level for 8
countries. In some cases the p-value is zero. For Brazil and Guyana it is significant at the
5% level. However for Costa Rica, Jamaica, Nicaragua, Trinidad and Tobago, and

Venezuela, bzis not significant (even at the 10% level). In all cases where bzis

significant it is also negative.

Using the third panel for all 7 countries the p-value on b1 is zero. For all countries
except Ecuador b1 is approximately 1. For Ecuador b1 is -4.17. Formal statistical testing
shows that the hypothesis that b1 is equal to 1 cannot be rejected at the 5% level for Costa

Rica and Venezuela Only for Nicaragua and Venezuela is b2 not significant at the 1%

116

level. For Nicaragua it is significant at the 10% level while for Venezuela it is not
significant at all.

Dickey-Fuller (DF) and augmented Dickey-Fuller (ADF) tests are conducted for
each series. A trend term is included where necessary for the relative money supply and
GDP series based on both visual inspection and formal statistical testing. The results of
the formal unit root tests for the first panel are illustrated in tables 3.3-3.5. They indicate
that the exchange rate series is 1(1) for every country except Brazil. For Brazil the test
results imply that the series is 1(2). The unit root tests on relative GDP imply that for
every country except Brazil the series are 1(1). For Brazil the results indicate that relative
GDP is stationary. The results for money supply are more mixed. For 9 countries the
series is 1(1). For 5 countries — Bolivia, Brazil, Nicaragua, Peru, and Uruguay, the series
is 1(2) and for Colombia the series is 1(0).

Using the third panel the unit root tests indicate that for all 7 countries except
Brazil and Nicaragua all three series are 1(1). Once again for Brazil the exchange rate and
relative money supply series are both 1(2). However in this case relative GDP is 1(1). For
Nicaragua as is the case with the first panel, both the exchange rate and relative GDP are
1(1). A comparison of the results for the two panels indicates that of the 7 countries in
panel three the results differ for Colombia and Peru. Using a longer time span all three
series are I(l) for Colombia while using the shorter time span, the relative money supply
sequence is stationary and relative GDP and the exchange rate are 1(1). For Peru using the
longer time span all three series are 1(1) while panel 1 indicates that only the exchange

rate series and relative GDP are 1(1).

117

The results of the D-F test show the orders of integration based on a mechanical
performance of the test. However in time series analysis most macroeconomic variables
are [(1). An inspection of the root of relative GDP series for Brazil shows that the value
of 0.84 is similar to the value for other countries that are found to be 1(1). Furthermore it
is higher than the root for countries such as Costa Rica, Guyana, and Venezuela for
which the D-F test implies that the series is 1(1). Therefore I treat all three series as 1(1)
for all countries and conduct cointegration tests on the residuals from the monetary model
for all countries. It is possible that the DP test concludes that some of the series are not
1(1) because of the small sample size or because of type 1 or type 2 errors.

Fixed effects is used on the pooled time series data to estimate equation (9). As
table 3.6 indicates, for all three panels the results are generally supportive of the
monetary model. For panels 1 and 2 the p-values for both relative money and relative
GDP are zero. Using the third panel the p-value for relative money supply is zero while

relative GDP is significant at the 10% level. In all cases bl is approximately 1 and b; is
negative. Using panels 1 and 3 the null hypothesis that bl is equal to 1 cannot be rejected

at the 5 % level.

I apply a panel version of the Engle and Granger procedure to the residuals from
the fixed effects regression. As table 3.7 illustrates for panel 1 the t-statistic is -8.82 and a
regression of the residuals on their lagged values shows that the root is 0.71. The critical
values for the Engle and Granger test cannot be used in this case because the procedure is
being applied to panel and not time series data. However the high t-statistic and the size
of the root imply that the residuals may be stationary. There is therefore preliminary

evidence that the panel data implies that there is a long run relationship between the

118

exchange rate and the fundamentals. The results using panel 2 show even more evidence
of co-integration. In this case the t-statistic is -10.21 and the root is 0.48. Panel 3 shows
the weakest support for co-integration. However it should be noted that when Colombia,
Ecuador and Venezuela, the three countries for which the model fits most poorly, are
excluded from this panel the root falls from 0.976 to 0.72.

To find out the order of integration of the panel data I use Choi’s test for panel

unit roots. Choi (2001) introduces the inverse normal test. The test statistic is:

l N -1
Z=—Z<I> (pi) (22)

W 1:]
where (I) = standard normal cumulative distribution function and

p,- = p-value for a unit root test statistic for country i

The null hypothesis is that all the time series are unit root non-stationary while the
alternative hypothesis is that at least one of the time series is stationary. Using the first
panel the null hypothesis cannot be rejected at the 5% level for all three series. When the
test is conducted on the first difference of each series the null hypothesis is rejected and
the p-value for the individual unit root tests is less than 0.05 for the majority of countries
in the sample. This implies that each series is 1(1).

Using panel 3 the null hypothesis cannot be rejected for the money supply and
GDP series. The null hypothesis is rejected for the first difference of all three variables
and the p-values are less than 0.05 for most of the countries in the sample. Once again

this is evidence that each series is 1(1).

119

(b) Formal Cointegration Testing

I use the two-step Engle and Granger procedure to conduct cointegration tests on
the individual countries. Since there are two explanatory variables in the model I obtain
the critical values from Engle and Y00 ( 1987). As table 3.8 shows the null hypothesis of
no cointegration is rejected for Argentina, Bolivia, Chile, Nicaragua, Trinidad and
Tobago, and Uruguay using the first panel. Using the second panel the hypothesis is
rejected for all of the above countries as well as Guyana and Mexico. For panel 3 the
hypothesis can only be rejected for Costa Rica. Additionally table 3.8 shows that for
Brazil, Jamaica and Peru, even though the Engle and Granger test formally shows no
cointegration, the t-statistic for these countries is not as low as it is for Colombia,
Ecuador and Venezuela. So there is evidence of a long run relationship between the
exchange rate and the fundamentals. A comparison with table 3.2 shows that the
countries for which cointegration is found all have residuals with roots of 0.41 or lower.
For the countries for which the hypothesis of no cointegration could not be rejected the
roots are all above 0.41. A comparison with table 3.2 also shows that the results of the
OLS estimation for the countries for which a long run relationship exists are supportive
of the monetary model. For countries such as Colombia and Ecuador where the results
are unsupportive of the model the null hpothesis of no cointegration cannot be rejected.

The first panel data cointegration test that I apply is the Levin et al test. It should
be noted that this test is only valid if the regressors are exogenous. I apply the test to the
residuals from the fixed effects regression. The results of this test indicate that the null
hypothesis of no cointegration can be rejected. As table 3.9 shows, for panel 1 the test

statistic is -3.95 implying that the null hypothesis that the residuals are non-stationary for

120

every individual in the panel can be rejected. For panels 2 and 3 the test statistics are -
4.62 and -1.92 respectively. So this test implies that there is a long run relationship
between the exchange rate and the fundamentals. As implied by the individual test results
the rejection of the null hypothesis is strongest for panel 2 and weakest for panel 3. Panel
2 has the largest number of countries for which the hypothesis of no cointegration is
formally rejected. Recall that for panel 3 formal evidence of cointegration is only found
for Costa Rica. It is therefore not surprising that the Levin et al test shows that the null
hypothesis can only be rejected at the 5% level. For panels 1 and 2, it is rejected at the
1% level.

Table 3.9 also illustrates the results of the Pedroni test. Using all three panels the
null hypothesis of no cointegration for each member of the panel can be rejected even at
the 1% level in favour of the alternative hypothesis that for each country there exists a
cointegrating vector. Once again the rejection of the null hypothesis is strongest for panel
2 and weakest for panel 3. So the Pedroni test also implies that there is in fact a long run
relationship between the exchange rate and the fundamentals. This along with the results
of the fixed effects regression implies that the monetary model holds as a long run

phenomenon.

(c) Error Correction Model

If a set of variables is cointegrated, their short term movements must be
influenced by the deviation from the long run relationship. This leads to the concept of an
error correction model. Error correction models describe how the variables respond to

short term deviations in order to restore equilibrium. The following trivariate vector error

121

correction model in s, , m, —m: , and y, — y: is estimated for each country for which
evidence of cointegration is found:
t '3
AS: = k1 + ksut—l + kll A SH + k12 A (mt—1 - "it—1) + 1‘13 A (yt—l - yt—l) + est (23)

* *
A0": 'mt ) = k2 +kmut—1+k21ASt-1+ k22 A(mz—l -mt—1)+

#
k23 A(yt—1-yt—1)+emt (24)

* *
A04 -yz ) = 1‘3 +kyut—1+k31ASt—l+k32 A(mt—l “'mt-l)+

1‘33 AUG-I ‘yi—1)+eyt (25)

A maximum lag length of 1 is used because additional lags are insignificant. The
above model is estimated for Argentina, Bolivia, Chile, Nicaragua, Trinidad and Tobago
and Uruguay for panel 1, all of the above countries as well as Guyana and Mexico for
panel 2 and Costa Rica for panel 3. The results for panel 1 are illustrated in table 3.10. On
the basis of the results of the Pedroni test, the panel version of the above model is also
estimated. These results are illustrated in table 3.11.

Table 3.10 illustrates that for Argentina, Nicaragua, and Trinidad and Tobago, it
is relative GDP that adjusts to restore the long run equilibrium. For Bolivia, the opposite
result holds. That is, for that country it is the exchange rate and relative money supply
that adjust to restore the long mm equilibrium. The results suggest that when there is a
gap between the exchange rate and the fundamentals both the exchange rate and the
money supply adjust to reduce the exchange rate, that is, appreciate the currency and
restore equilibrium. For Chile it is the relative money supply only that does the adjusting.
For Uruguay, none of the adjustment parameters are significant. This implies that the

variables may not in fact be cointegrated. However, this may just be due to a statistical

122

error. The results for Uruguay are quite supportive of the model and the root of the
residuals is 0.38. Table 3.11 shows that panel 2 implies that money supply does the

adjusting.

3.7 Discussion and Conclusions

In the vein of papers like Groen (2000) and Rapach and Wohar (2002) this paper
continues to find support for the monetary model. Like those authors I show that the
existence of a long run relationship between the monetary fundamentals and the exchange
rate may not be as statistically unsupported as it once seemed. However unlike those
authors I focus on the exchange rate movements of Latin American countries rather than
industrialized first world countries. The contribution of this paper is to show that the
current consensus about the validity of the monetary model of exchange rates for high
income developed countries holds for the countries of Latin America. That is, like
Rapach and Wohar (2002) I show that panel data methods may overstate the support for
the model because they force one to accept that it holds (or doesn’t hold) for each
individual unit in the panel.

The Pedroni test shows that the hypothesis that there is no cointegration for each
member of the panel can be rejected. The cointegrating vectors for panels 1, 2, and 3 are
(1.01, -1.32), (0.95, -4.57), and (1.02, -0.44) respectively. The coefficients are correctly
signed and reasonably sized. This would lead one to accept the monetary model as an
explanation for exchange rate movements among Latin American countries. The

individual country analysis on the other hand shows strong support for the model for

123

 

Argentina, Bolivia, Chile, Costa Rica, Guyana, Mexico, Nicaragua, Trinidad and Tobago
and Uruguay. There is weaker support for the model for Brazil, Jamaica and Peru.
However there are other countries such as Colombia and Ecuador for which not only can
the null hypothesis of no cointegration not be rejected but for which the coefficient
estimates are not supportive of the model.

This leads to the question of why the model fits so well for countries like
Argentina and Bolivia and yet so poorly for others like Colombia and Ecuador. Rapach
and Wohar (2002) find support for the model for countries such as France and Italy but
also find that it does not hold for countries like Denmark and Sweden. They argue that
the failure of the model is caused by instability in the long run relationship between
relative price levels and fundamentals for those countries. However as Diaz (2004)
shows, that is not a likely explanation. Diaz (2004) examines how well the monetary
model describes exchange rates in Mexico from 1945-2002. Diaz tests the monetary
model as well as two of its building blocks, the money demand function and PPP over the
entire sample as well as over two non-overlapping sub-samples. He shows that even
though the relationships upon which the model is based changed over time the model still
described exchange rate movements over the entire sample. He therefore argues that
instabilities in the fundamentals do not provide an explanation for the failure of the
model.

This paper shows that the model holds for countries such as Chile which as
described in section 5 experienced a great deal of volatility in exchange rate movements
over the sample period. The model is also a good fit for countries such as Argentina

which experienced several changes in the actual currency used over the sample period.

124

Similarly the model describes exchange rate movements in countries such as Mexico and
Uruguay quite well, despite the fact that they have the largest coefficients of variation
over the sample periods 1967-2000 and 1981-2000 respectively. Therefore instability and
volatility do not provide explanationsfor the failure of the model in some countries.

It is interesting to note that in Colombia where there is no evidence of a long run
relationship between the exchange rate and the fundamentals there is some question about
the accuracy of the official money supply statistics. This is because a number of
transactions are conducted in US. dollars. This is mainly caused by the presence of the
illegal drug trade. Cocaine accounts for about 25% of that country’s foreign exchange
earnings. US. dollars are used by foreign exchange earners. Ecuador is another country
for which the model fits poorly. Currently Ecuador is a dollarized economy. Dollarization
officially occurred in 2000. However even before the local currency was officially
demonetized, the country was already heavily dollarized. There is therefore once again
the concern that the official money supply data does not reflect that which is actually
occurring in the economy.

It is therefore possible that for these countries there is a long run relationship
between the exchange rate and the fundamentals but the model appears to fit poorly
because of inaccurate data. If the data does not reflect what is happening in the economy
we will not find evidence of this relationship even if it does exist. Another possible
explanation is that for countries such as Colombia and Ecuador, the monetary model does
hold as a long run phenomenon. However the speed of adjustment for these countries is
so slow that it would take a much longer time span of data to find evidence of this

relationship. I

125

It should also be noted that Ecuador and Colombia are both members of the
Andean Community (CAN) and have been since 1969. Venezuela was a member of the
CAN from 1973-2006. The CAN and Mercorsur are the two main trading blocs of South
and Central America. Peru and Bolivia are also members of the CAN. However it should
also be noted that Venezuela and Ecuador are among Colombia’s top 3 export partners
while Columbia and Venezuela are among Ecuador’s top 4 import partners. It seems that
Colombia, Ecuador and Venezuela do a great deal of trading amongst themselves. The
results imply that the model fits poorly for these three countries. This may be due to
factors arising from the large volume of trade that occurs among these countries that are
having an impact on exchange rate movements in these countries.

Whether or not any of the above explanations hold true, this paper shows that the
monetary model describes exchange rate movements in Latin American countries just as
well as it describes exchange rate movements in first world industrialized economies.
This paper shows that there is evidence that the simple monetary model is appropriate
even for countries characterized by exchange rate volatility, that are plagued by high
inflation and whose currencies are not heavily traded on the world market. In spite of
these features it is still possible to find a long run relationship between the exchange rate
and the fundamentals.

Future work may benefit from finding a way to test the accuracy of official data
for less developed countries that possess a prominent black market for foreign currency
or that regularly conduct transactions in foreign currency. Research on less developed
countries generally speaking is handicapped by data availability. Another area of interest

is the relationship between exchange rates and growth. If there is a long run relationship

126

between the exchange rates and the monetary filndamentals and if these variables respond
to short run deviations from the equilibrium, insight on the relationship between
exchange rates and growth can have important policy implications. It would also be
interesting to investigate the impact of trade integration on exchange rates. Some
research, for example, Hooper and Kohlhagen (1978) and Rose, Lockwood and Quah
(2000), has looked at the impact of exchange rates on trade volumes but it would be
interesting to see how the volume of trade between countries influences their exchange

rate movements.

127

Argentine Peso per US $

Figure 3.1 - The Exchange Rate in Argentina

12 --..-.-. .. .---H--..___--_-_.._ -_-_____._----___-, n

 

08 ,.. -.L._fi_. - _ __. _,___. ._.n._.._.._.._r_.__t..

 

 

 

0.4 wm.

02 ..t..-.. -. . .n-,_.,.-_.-n-__.____..l... .

 

0

«eeextese
«tatax
(05°.(05°.(°l°\o .9 .9 (9.9.9.1

 

 

 

 

Boliviano per US 3

lVlal‘I

Bol

 

 

 

 

63%99
99999
9939419

128

Brazilian Real per US 8

Chilean Peeo per US 3

.43‘od’o’n “at M“,

119334951 ’0 6’4

K

Figure 3.3 - The Exchange Rate in Brazil

 

4‘3’5’nfie‘,’ are: $333,459 4’4’.5I’.4‘ 4"4’4’4’4’4 4’4’3‘

Year

Figure 3.4 - The exchange rate In Chlle

 

o’ers‘ets’e‘es’ets‘o’este‘spe‘ffato‘wtt‘t’s’e’

Year

129

Colombian Peso per US 5

Costa Rican Colon per US 5

2500

8

‘8”

1000

L8"

350

300

Figure 3.5 - The exchange rate in Colombia

...,—-.-, .. .,_. ,A‘,. ~— — —. -~-.._-...- ......_.____._ --..-. <w—_~V»—--. —-~—-———-.—-~n—-—v "-..—"4.” —~-‘

 

 

 

 

 

 

 

 

 

 

 

 

3%QQNW55%63%QQNW5 @6«%90\W5 95%«9390
e «aaaaaataaeeeee’eeeeeqqe o’eeeeee
NC” \°§°\°?\9 '9 '3“ v9 '9 '3’ '9 '3 31'9"" \°-’ #939 ’3‘ 39391399 993%.?) \9’ 3.9.8) ’3’")?

Year

Figure 3.6 - The Exchange Rate In Costa Rica

TT'TMTT“T 2...--. - ._ — . ...,.-a...-_ _.-._...-_.- ..-_.,.._ ...,- -.-,“-.. -_..- - ..- -.. -..,-.-“ --....-____-..._.___ _.. . .... ---._._ ...__.V.._.-~_f.-._ .---... --._a_ "...e -- ain‘. ‘ , » 7.7 - -. —._--_-—.__.—_=

 

 

 

 

 

N
01
O

200

150

..—_—.‘_._ t.. e. , L ._ ..

 

100

50

 

 

 

 

 

 

3%QQKW'BPQQ3«%QQNW5?‘JB’\$99NW’5&66’\%QQ
©b««’\«’\’\’\’\««%%%% %%%%%99 0.1 9999036
«6093’@3333933933’39’9K@\9\9\9\9\9\9\&\9\99319319Q

’L
Year

130

Figure 3.7 - The Exchange Rate in Ecuador

 

 

 

 

 

 

 

.
_
_
n

1

m
a m: .8 seem 3.83m

20000 .___-,_ —~

5000

oooN
mam?
mam?
hmmw
ommw
mam?
vmmw
mmmw
Nam?
wmmr
oom—
mam?

_ m8,

moor
wwmr
mwor
vom—
”war
«war
roar
Omar
mum?
who?
who?
char
mum?
thaw
numw
«nap
wnap
onmp

mom—

mom?

Year

Flgure 3.8 - The Exchange Rate in Guyana

 

 

 

 

 

200

180 i———-—————‘77774~~ ,, "w

80 t____
o

W
.m.
D .on a oeo:e>:0

120 ewe-~-

20

 

Year

131

Jamaican 3 per US 3

Mexican Peso per US 5

50

45

8

09
0|

8

N
00

N
0

_e
(1.

_a
0

Figure 3.9 - The Exchange Rate in Jamaica

 

 

 

 

 

 

 

p-_..._~_..__.__ _.____ _

 

 

 

 

«eeexmeueeaeeoxrte (05’\%90\W"J
e «'\'\‘\’\'\‘\'\‘\’\eeeee’eeeeeeeee
.913?d°~ (99.33.339.9399(9333399999(9.3

f v ‘I

(55 63‘50 ’\ % Q 0
(£3 Q
N ’\ NgNggNgN egé’fi‘?

Year

Figure 3.10 - The Exchange Rate in Mexico

 

 

 

 

«eooxrtsseeaeooxwe eeaeeoxrtegseeacoee
’oi'o ««««««««««eeee§e’eeeeeeeqe oooeee
\e\o\<1l°\e\o,9.9\q\e\e 939999.55».8.9433389093qu399$

132

Nicaraguan Cordoba per US 5

Peruvian Nuevo $01 per US 3

14

12

10

Figure 3.11 - The Exchange Rate in Nicaragua

 

F__" - . “..., -._ fin.-- ._.--._..._______ -__._ _ __..

 

 

 

 

 

 

 

 

 

 

~06} see’x999 99 9 «999 99 9 «99
9P f9" .fl90P Mei "(59‘9“ 9‘ .9A .93‘ .9" .9P 9P ”9" P .99? .9P.9P 6H9P 9P .9P .9" MP P 9" 9P9P 9Pb9P .9" 9P¢9P°

9°
01

(a)

N
or

N

d
(I!

u.

.0
0!

«999x99 9,( «999k 99t999~99 bb‘\990
90(69««««'\“CD« '\'\'\99 99’999999999999999
(.9 83'8”“ W9999999W9Pb9 9M9999.9.9999.9.9.9.9

Year

Figure 3.12 - The Exchange Rate in Peru

 

 

 

F'—' V__-..._... . , ,4____,.z.. _.

 

 

 

'\

Year

133

Trinidad and Tobago 3 per US $

Uruguayan Peso per US 5

Figure 3.13 - The Exchange Rate In Trinidad and Tobago

 

 

 

 

 

 

 

., ~74___LL__.4

 

 

 

 

 

 

 

 

 

0 - . r
«$9QNW95653%90\ '5 bbh$9°~$5 963%90
~96 '96 \éo '59 \é ’93 '5" \‘P\ \‘P\ \‘g '3’} r33 \6\ 9% ’93) 9‘93? \P%h@% \an >9” 3% '93, '39 '39 '99 >90" @9535? ~43 *9“ ’39 99 119°
Year
Figure 3.14 - The Exchange Rate in Uruguay
14 -4 -..,.-- +9.-- .. . -e- .-._._.-.. --.e--- __ -.- ---....__. h4~f 7“ __. 4 A- -W -M v“...
12 L— —~7~7V 7" - _ efi—fe— ——. /

 

 

 

 

 

 

 

 

 

 

 

 

 

 

0 1 ..9

 

f

QQN'L'bb‘ogb
9 99999
9.9P.9 .9 .9 9.9.9

«999N995994999N99 9949
be ’\’\’\'\’\’\’\’\’\’\%%%%%P<b<b‘b%
9.9.93999.9.9.9.99.99999999999N .\

K

Year

134

Venezuelan Bolivar per US $

800

700

600

300

200

100

0

’\ 9: 9 9 x
Q) ’\
9°.9°.9 9‘9

-,—.—.--_.. .-.,"...

Figure 3.15 - The Exchange Rate in Venezuela

 

__._ _~— .-- __._.__.~....,~.-......_.-_~._ .....-...-- _., . .5- .. fl... . ... .. . . . — . .....,~-...,._ M...

 

 

 

 

 

 

 

 

 

 

 

 

 

 

P4_L .flmnga.-.“ ...M L“. .

 

 

 

r

«PAPA‘tPaPa‘«P499P9‘9P9P9’9P9P9‘9P9P9P9"
NQNQ’NQNQ’NQNQ’NO"ngg'xg’g’g'qug'g 9'9 9’9'9

Year

135

Table 3.1 — Coefficient of Variation of Exchange Rates Over Different Time Periods

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Country 1967-1980 1981-2000 1967-2000
Argentina 0.015 1.665 0.716
Bolivia 0.0007 58.869 1.558
Brazil 0.0012 0.921 0.434
Chile 88.79 0.0203 1.48
Columbia 0.014 0.034 0.1 12
Costa Rica 0.0043 0.024 0.162
Ecuador 0.002 0.098 0.204
Guyana 0.018 0.24 0.605
Jamaica 2952.23 0.202 1.068
Mexico 0.0047 9.98 3.801
Nicaragua 0.00004 2.488 0.727
Peru 0.0024 1.91 0.654
Trin. & Tob. 0.015 0.062 0.143
Uruguay 0.05 1 4364204 1 .977
Venezuela 0.0001 0.22 0.408

 

 

 

 

136

 

Table 3.2 — Results of Individual OLS Estimation of the Monetary Model Using Panel 1

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Country mt _ m: (b1) y, _ y: ( b2) Root of Residuals

Argentina 0.95 -5.47 0.40
(0.014) (1.15)
[0.014] [1.04]

Bolivia 0.95 -4.43 0.18
(0.02) ‘ (0.96)
[0.02] [0.99]

Brazil 1.02 -0.204 0.73
(0.009) (0.35)
[0.012] (0.45]

Chile 0.96 -2.15 0.14
(0.02) (0.62)
[0.03] [0.62]

Colombia 0.17 10.733 0.87
(0.15) (2.26)
[0.16] [2.499]

Costa Rica 1.06 -3.25 0.47
(0.04) (0.62)
[0.04] [0.696]

Ecuador -4.39 8.63 0.97
(0.503) (1.06)
[0.66] [1.34]

Guyana 1 . l 7 -0.296 0.78
(0.07) (0.47)
[0.09] [0.59]

Jamaica 0.91 -1.04 0.70
(0.08) (0.53)
[0.10] [0.66]

Mexico 0.997 -1.25 0.80
(0.02) (0.68)
40.03] [0.798]

Nicaragua 1.24 4.36 0.41
(0.05) (1.34)
[0.07] [1.64]

Peru 0.94 -4.504 0.53
(0.01) (0.67)
[0.02] [0.697]

Trinidad & Tobago 0.46 -2.15 0.28
(0.04) (0.22)
[0.04] [0.23]

Uruguay 0.96 -3.54 0.38
(0.01) (0.35)
[0.02] [0.398]

Venezuela 1.25 1.08 0.86
(0.14) (1.47)
[0.16] [1.79]

 

 

 

Notes: The explained variable is the log exchange rate
Heretoskedasticity Robust standard errors are in parentheses
Newey-West standard errors are in square brackets

137

 

Table 3.3 — Individual Unit Root Test Results for Exchange Rates

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Country Root Statistic Critical Value Outcome
Argentina 0.99 -0.829 -2.980 1(1)
Bolivia 0.97 -0.863 -2.980 1(1)
Brazil 0.99 -0.808 -2.980 1(2)
(0.81) (-1.728) (-2.980)
Chile 0.93 -2.775 -2.978 1(1)
Colombia 1.01 0.495 -2.980 1(1)
Costa Rica 0.999 -0.027 -2.978 1(1)
Ecuador 1.03 1.506 -2.980 1(1)
Guyana 1 .01 0.427 -2.978 1(1)
Jamaica 1.01 0.288 -2.978 [(1)
Mexico 0.99 -0.410 -2.980 [(1)
Nicaragua 0.99 -0.433 -2.980 1(1)
Peru 0.99 -0.477 -2.980 [(1)
Trin. & Tob. 0.99 -0.159 -2.978 1(1)
Uruguay 0.99 -0.687 -2.978 [(1)
Venezuela 1 .04 1 .746 -2.978 [(1)

 

 

 

 

 

Notes: DF and ADF tests conducted
Critical values differ based on the number of lags

Unit root test results on the first difference of the variable are in parentheses

138

 

Table 3.4 — Individual Unit Root Test Results for Relative Money Supply

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Country Root Statistic Critical Value Outcome

Argentina 0.87 -2.185 -3.572 1(1)

Bolivia 0.86 -2.21 1 -3.572 1(2)
(0.58) (-2.83 l) (-2.98)

Brazil 0.93 -2.275 -3.572 1(2)
(0.76) (-1 .992) (-2.980)

Chile 0.898 -2.786 -2.980 1(1)

Colombia -0.034 -5.659 -3.568 1(0)

Costa Rica 0.673 -2.422 -3.568 1(1)

Ecuador 0.90 -1.559 -3.568 1(1)

Guyana 0.87 -2.074 -3.572 1(1)

Jamaica 0.899 -1 .341 -3.568 1(1)

Mexico 0.88 -2. 148 -3.572 1(1)

Nicaragua 0.91 -2.542 -3.572 1(2)
(0.76) (-2.036) (-2.980)

Peru 0.89 -2. 196 -3.5 72 1(2)
(0.62) (-2.644) (-2.980)

Trin. & Tob. 0.93 -1.767 -2.980 1(1)

Uruguay 0.99 -1 .064 -2.980 1(2)
(0.56) (-2.683) (-2.980)

Venezuela 1.05 1.859 -2.978 1(1)

 

Notes: DF and ADF tests conducted
Critical values differ based on the presence or absence of time trends and the number of lags
Unit root test results on the first difference of the variable are in parentheses

Table 3.5 — Individual Unit Root Test Results for Relative GDP

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Country Root Statistic Critical Value Outcome
Argentina 0.94 -0.972 -2.978 1(1)
Bolivia 0.86 ~2.426 -3.572 1(1)
Brazil 0.84 -4.206 -2.978 1(0)
Chile 0.98 -0.370 -2.978 [(1)
Colombia 0.81 -1.822 -3.572 1(1)
Costa Rica 0.79 -2.836 -2.980 1(1)
Ecuador 0.97 -0.568 -3.568 1(1)
Guyana 0.75 -2.137 -3.572 1(1)
Jamaica 0.75 -2.245 -3.568 1(1)
Mexico 0.85 -2.400 -2.980 1(1)
Nicaragua 0.75 -2.435 -3.568 1(1)
Peru 0.95 -0.726 -2.983 1(1)
Trin. & Tob. 0.86 -1.474 -2.978 [(1)
Uruguay 0.898 -1.622 -2.983 1(1)
Venezuela 0.566 -2.979 -3.572 1(1)

 

 

 

 

 

 

Notes: DF and ADF tests conducted
Critical values differ based on the presence or absence of time trends and the number of lags

139

Table 3.6 — Fixed Effects Estimation of the Panel Version of the Monetary Model

 

 

 

 

 

 

 

Sample mt -m: (bl) y: -y: (b2)
Panell ***1.01 ***-1.32
(0.009) (0.28)
Panel 2 ***0.95 ***-4.57
(0.015) (0.49)
Panel 3 ***1.02 *-0.44
(0.01) (0.246)

 

Notes: The explained variable is the log exchange rate
Standard errors are in parentheses
*, ", and "" denote significance at the 10, 5, and 1% level respectively

Table 3.7 — Panel Engle and Granger Results

 

 

 

 

 

Sample T- Statistic Root
Panel 1 -8.82 0.71
Panel 2 -10.21 0.48
Panel 3 -1.05 0.976

 

 

 

 

 

Notes: The Engle and Granger procedure is applied to the residuals from the fixed effects regression of
equation (9) (the panel version of the monetary model)

Table 3.8 - Individual Cointegration Test Results

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Country Panel 1 Panel 2 Panel 3
Argentina *-3.655 * * *-5 .451

Bolivia * * *-5.18 * * *-4.706

Brazil -2.209 -2.865 -2.604
Chile ***-4.929 "-3.783

Colombia -1 .21 1‘il -1 .3093 -2.460
Costa Rica -3 .279 -1.508 * *-3 .970
Ecuador -0.407 -1.317 -0.170
Guyana -2.127 *-3.474

Jamaica -2.433 -2.840

Mexico -2.462 * * *-6.577

Nicaragua **-3.701 “-3.676 -3.139
Peru -2.974 -3 .094 -3.175
Trinidad & Tobago "-4.238 * * *-4.874

Uruguay ***-4.189 ***-4.234

Venezuela -1 .652 -1.479 -1 .593

 

 

 

 

Notes: The critical values are fiom Engle and Y00 (1987) and are 4.32, 3.67, and 3.28 at the 1,5, and

10% levels respectively

The critical values for ' are from Engle and Granger (1987) and are 4.07, 3.37, and 3.03 at the 1,5,
and 10% levels respectively (3.73, 3.17, and 2.91 with lags)

*, ", and *" denote rejection at the 10, 5, and 1% levels respectively

140

 

Table 3.9 - Panel Cointegation Test Results

 

 

 

 

 

 

Sample Levin et al Pedroni

test statistic critical value test statistic critical value
Panel 1 ***-3.95 -1.65 *‘*-l9.08 -1.65
Panel 2 ***-4.62 -1.65 ***-40.94 -1.65
Panel 3 “-1.92 -1.65 ***-13.71 -1.65

 

 

 

 

 

Notes: The tests are applied to the residuals from the fixed effects regression equation (9) (the panel
version of the monetary model)

‘, ", and "‘ denote rejection at the 10, 5, and 1% levels respectively

5 % critical values are displayed in the table

Table 3.10 — Error Correction Coefficient Estimates for Panel 1 Countries

 

 

 

 

 

 

 

 

 

 

 

Country k k... ky
Argentina -0.1 1 0.22 "-0.05
(0.59) (0.29) (0.02)
Bolivia ***-2.38 "-1.19 0.02
(0.75) (0.54) (0.02)
Chile -0.19 "0.66 0.02
(0.18) (0.28) (0.02)
Nicaragua -0.92 -0.41 ***0.05
(0.57) (0.27) (0.01)
Trinidad & Tobago -0.02 -0.02 ***-0.37
(0.17) (0.21) (0.11)
Uruguay -0.52 0.04 -0.02
(0.35) (0.13) (0.04)

 

Notes: Heteroskedasticity Robust standard errors are in parentheses
*, ", and "‘ denote significance at the 10, 5, and 1% levels respectively

Coefficients are the speed of adjustment parameters for exchange rates, relative money supply, and

relative income (based on equations 23-25)

Table 3.11 — Error Correction Coefficient Estimates

 

 

 

 

 

 

 

 

Sample ks km Icy
Panel 1 -0.075 0.135 -0.0006
(0.083) (0.091) (0.003)
Panel 2 -0.151 "0.285 -0.001
(0. 1 3) (0.146) (0.005)
Panel 3 -0.055 -0.035 0.0005
(0.074) (0.034) (0.003)

 

Notes: Heteroskedasticity Robust standard errors are in parentheses
*, ", and "* denote significance at the 10, 5, and 1% levels respectively

Coefficients are the speed of adjustment parameters for exchange rates, relative money supply, and

relative income (based on equations 23-25)

141

 

 

 

APPENDIX

142

Table A1.1 — Unbalanced Panel — Causality from Trade to Growth

 

 

 

Statistic K = 3 K = 2 Critical
. Value

2F“NL 5.199 7.993 1.650

ZHNC 6.803 6.974 1.650

 

 

 

 

 

Table A1 .2 -— Unbalanced Panel — Causality from Growth to Trade

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Statistic K = 3 K = 2 Critical
Value
zFHNC 4.108 7.230 1.650
ZHNC 5.549 6.369 1.650
Table A1.3 — Balanced Panel — Three Lags
Sample Split Trade to Growth Growth to Trade Critical Value
All Countries 5.086 4.1 19 1.650
Low Income * 1 . 199 *0.800 4.615
Lower Middle 5.642 *4.454 4.538
Upper Middle * 1.882 *-0.666 4.950
Middle Income 5.693 *3.252 4.356
High Income *0.873 *3.038 4.815
Note * denotes non-rejection at the 5% level
Table A1.4 — Balanced Panel — Two Lags
Sample Split Trade to Growth Growth to Trade Critical Value
All Countries 2.310 3.493 1.650
Low Income *0.803 * l .321 3.106
Lower Middle * 1.186 *1.648 3.049
Upper Middle *1.223 *0.793 3.356
Middle Income * 1.821 *1.960 2.913
High Income ”“1518 3.311 3.256

 

 

 

 

Note * denotes non-rejection at the 5% level

143

 

 

Robust Results

Table A1.5 — Unbalanced Panel —- Causality From Trade to Growth

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Statistic K = 3 K = 2 K = 1- Critical Value
ZI 111 'C 8.734 9.341 6.325 1.650
ZHNC 10.861 8.043 7.373 1.650
Table A1.6 — Unbalanced Panel - Causality From Growth to Trade

Statistic K = 3 K = 2 K = 1 Critical Value
ZFHNC 7.628 10.662 4.842 1.650
ZHNC 9.591 9.091 5.743 1.650
Table A1.7 — Balanced Panel — Three Lags

Sample Split Trade to Growth Growth to Trade Critical Value
All Countries 17.006 14.756 1.650

Low Income *3.381 *4.279 4.615

Lower Middle 18.332 11.694 4.538

Upper Middle 6.505 *3.121 4.950

Middle Income 18.724 1 1.350 4.356

High Income *4.156 9.681 4.815

Note * denotes non-rejection at the 5% level

Table A1.8 — Balanced Panel — Two Lags

Sample Split Trade to Growth Growth to Trade Critical Value
All Countries 8.091 1 1.916 1.650

Low Income 4.535 7.323 3.106

Lower Middle 5.262 7.170 3.049

Upper Middle *2.381 *2.883 3.356

Middle Income 6.165 8.173 2.913

High Income 3.575 5.806 3.256

 

 

 

 

 

Note "‘ denotes non-rejection at the 5% level

144

 

 

 

 

Table A1.9 — Balanced Panel — oneLag - Causality from Openness to Growth

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Measure of All Low Lower Upper Middle High
Openness Countries Income Middle Middle Income Income
Trade 7.812 2.821 8.008 *0.958 7.092 2.769
Percentage

Gravity — 5.003 2.240 5.548 *0.883 5.040 *0.478
Differentiated ’

Goods

Gravity — 5.039 2.324 5.174 *0.954 4.776 *0.872
Homogeneous

Goods

Note " denotes non-rejection at the 5% level

Table A1.10 - Balanced Panel — one lag - Causality from Growth to Openness
Measure of All Low Lower Upper Middle High
Openness Countries Income Middle Middle Income Income
Trade 8.559 10.218 2.086 *0.986 2.272 3.187
Percentage

Gravity - 9.551 9.598 3.080 2.361 3.878 3.596
Differentiated

Goods

Gravity — 9.888 12.236 3.449 1.898 3.912 *1.153
Homogeneous

Goods

 

 

 

 

 

 

 

Note "‘ denotes non-rejection at the 5% level

145

 

 

Table A2.1 — Estimation Results with Initial Values and using instruments for y only

 

 

 

 

 

 

 

 

 

 

 

 

Regressor Cross-Sectional POLS FE-IV FD-IV
Grav_Open1a -0.00283 0.00091 * * *0.03135 * * *0.03424
(0.00230) (0.00171) (0.00792) (0.00712)
Log (Income) 0.001681 -0.00151 -0.04268 * * *-0. 14389
(0.00192) (0.00224) (0.06928) (0.0311)
Log * * *0.00469 ** *0.00301 0.01660 **-0.09921
(Population) (0.00150) (0.00114) (0.07274) (0.04687)
Log (Area) * * *-0.00490 * * *-0.00393
(0.00099) (0.00084)
Investment ***0.05873 ***0.14472 “0.09357 ***0.13425
(0.02267) (0.03965) (0.0458) (0.02999)
Government 0.00019 -0.00014 -0.03 199 -0.05438
(0.01536) (0.01373) (0.03292) (0.02404)

 

Notes: ‘Grav_Open1 represents the difference between trade predicted by the gravity model and actual
trade assuming that trade is in differentiated products.
the Huber/White/Sandwich standard error estimates are in parentheses

*denotes significance at the 10% level

** denotes significance at the 5% level
"*denotes significance at the 1% level

Table A2.2 — Estimation Results with Initial Values and using instruments for y only

 

 

 

 

 

 

 

 

 

 

 

 

Regressor Cross-Sectional POLS FE-IV F D-IV
Grav_Open2al -0.00239 *0.00421 ***0.02581 ***0.03463
(0.00309) (0.00248) (0.00804) (0.00661)
Log (Income) 0.00259 -0.00213 -0.04975 ***-0. 14826
(0.00166) (0.00223) (0.07092) (0.02993)
Log ***0.00519 “0.00236 0.00491 **-0.10580
(Population) (0.00161) (0.00116) (0.07119) (0.04456)
Log (Area) ***-0.00512 ***-0.00395
(0.00097) (0.00083)
Investment “0.05799 ***0.14000 ***0.10431 ***0.13231
(0.02334) (0.0396) (0.03949) (0.02923)
Government 0.00065 -0.00024 -0.03021 **-0.05498
(0.01596) (0.01372) (0.03241) (0.02403)

 

Notes: ‘Grav_Open2 represents the difference between trade predicted by the gravity model and actual
trade assuming that trade is in differentiated products.
the Huber/White/Sandwich standard error estimates are in parentheses

*denotes significance at the 10% level

" denotes significance at the 5% level
"*denotes significance at the 1% level

146

 

 

 

Estimation Results Using Alternative Measure of Ogenness With Interaction Term

Table A2.3 — Estimation of equation 1 using alternative measure of openness

 

 

 

 

 

 

 

 

 

 

 

Regressor Cross- POLS Fixed Effects F irst-
Sectional Difference
Grav_Openl 8 "-0.03882 -0.00241 "0.08161 0.04589
(0.01598) ‘ (0.01571) (0.03503) (0.07390)
Log (Initial * *0.02980 -0.00128 * * *-0. 12439 * * *-0. 12507
Income) (0.01373) (0.01341) (0.03456) 40.01080)
Openness*log * *0.00473 0.00058 -0.00668 -0.00130
(Income) (0.00188) (0.00178) (0.00438) (0.00893)
Population ***-0.93 838 *-0.5311 1 0.30573 0.37994
Grth (0.21634) (0.29091) (0.23913) (0.24439)
Log (Area) -0.001 13 ***-0.00185
(0.00073) (0.00052)
Investment ***0.11794 ***0.1818 ***0.19745 ***0.2103
(0.02801) (0.03577) (0.02917) (0.04937)
Government -0.00013 -0.00575 -0.0373 —0.05058
(0.01fl) (0.01355) (0.02559) (0.03344)

 

 

 

Notes: ‘Grav_Openl represents the difference between trade predicted by the gravity model and actual
trade assuming that trade is in differentiated products.
the Huber/White/Sandwich standard error estimates are in parentheses

‘denotes significance at the 10% level
" denotes significance at the 5% level
"*denotes significance at the 1% level

147

 

Table A2.4 — Estimation of equation 1 using alternative measure of openness

 

 

 

 

 

 

 

 

 

 

 

 

 

Regressor Cross- POLS Fixed Effects First-
Sectional Difference
Grav_0pen2 -0.00518 *0.03 756 * * *0.08912 0.03933
(0.02269) (0.01951) (0.03419) (0.07703)
Log (Initial -0.00261 * *-0.01 821 * * *-0.09927 -0. 12950
Income) 40.0082) (0.00754) (0.01443) (0.01099)
Openness*log 0.00101 -0.003 69 *-0.00743 -0.00045
(Income) (0.00267) (0.0022 8) (0.00420) (0.00924)
Population ** *-0.89836 *-0.51082 0.33377 *0.39722
Growth (0.22062) (0.27049) (0.23832) (0.24144)
Log (Area) * *-0.00154 * * *-0.00251
(0.00073) (0.00054)
Investment ***0.11635 ***0.17548 ***0.l9765 ***0.21144
(0.03102) (0.03661) (0.02901) (0.04946)
Government 0.00246 -0.00674 -0.04076 -0.05228
(0.01576) (0.01359) (0.02559) (0.0336)

 

Notes: ‘Grav_0pen2 represents the difference between trade predicted by the gravity model and actual
trade assuming that trade is in homogeneous products.
the Huber/White/Sandwich standard error estimates are in parentheses

*denotes significance at the 10% level
** denotes significance at the 5% level
”*denotes significance at the 1% level

Table A2.5 — Results using Alternative Openness Measures and instruments for y only

 

 

 

 

 

 

 

I Regressor FE-IVI‘ FE-Ivz" FD-IVl FD-IV2
Initial Openness *0.10799 M0.13959 0.04653 0.04876
(0.06084) (0.05679) (0.05153) (0.04965)
Log (Initial -0.10445 -0.05895 **-0.12016 ***-0.11788
Income) (0.08829) (0.05591) (0.06075) (0.03109)
Openness*log -0.01068 "-0.01530 -0.00139 -0.00178
(Income) (0.00741) (0.00714) (0.00651) (0.00623)
Population 0.34926 0.40895 0.34638 0.36106
Growth (0.31027) (0.31941) (0.22525) (0.22450)
Investment ***O.20482 ***O.21692 ***0.21684 ***0.22085
(0.05375) (0.05723) (0.03777) (0.0381 1)
Government -0.04272 -0.05016 *-0.05198 *-0.05325
(0.04836) (0.04918) (0.03142) (0.0314)

 

 

 

 

 

 

Notes: 'l represents trade in differentiated products.
I’2 represents trade in homogeneous products
standard errors are in parentheses

*denotes significance at the 10% level
** denotes significance at the 5% level
*"denotes significance at the 1% level

148

 

 

Table A2.6 — Results using Alternative Openness Measures and instruments for y, I & G

 

 

 

 

 

 

 

 

 

Regressor FE-IVl“ Fla-Ivz1S FD-IVl FD-1V2
Initial Openness 0.20502 0.02128 -0.l6214 -O.23611
(0.15283) (0.24598) (0.47810) (0.46613)
Log (Initial -017200 —0.05457 0.11742 0.03213
Income) (0.17212) (0.11950) (0.48134) (0.23328)
Openness*log .0.02021 0.00413 0.01639 0.030010
(Income) (0.01592) (0.02814) (0.03981) (0.04757)
Population 0.18026 0.37677 -O.43278 -O.19664
Growth (0.47593) (0.48897) (0.77140) @56746)
Investment 0.04189 -O.13063 1.24093 0.75354
(0.37703) (0.33366) (2.05612) (1.11497)
Government -O.35308 —0.2292 025952 0.27744
(0.22412) (0.19996) (0.81786) (0.59057)

 

 

 

 

 

Notes: '1 represents trade in differentiated products.
b2 represents trade in homogeneous products
standard errors are in parentheses
‘denotes significance at the 10% level

** denotes significance at the 5% level
"*denotes significance at the 1% level

Table A2.7 — Estimation Results — usingfiifferent reduced forms for each time period

 

 

 

 

 

 

 

 

Regressor FE-IV1° FE-IV2b FD-IVl FD-IV2
Initial Openness * * *0.35358 * * *0.24401 * * *0.44016 ** *0.26607
(0.12512) (0.07306) (0.16583) (0.09867)
Log (Initial * * *-0.44949 * * *-0.22658 * * *-0.60517 * * *-0.31842
Income) (0.15000) (0.05784) (0.2065 3) (0.08634)
Openness*log ***0.35358 ***-0.02575 *"‘-0.05017 M-0.02710
(Income) (0.12512) (0.00878) (0.02070) (0.01208)
Population *0.53480 *0.59124 "0.61483 ”0.64029
Growth (0.31477) (0.30577) (0.30455) (0.29923)
Investment 0.09501 *0.10454 0.10449 *0.1 1095
(0.06047) (0.05515) (0.06665) (0.06376)
Government 0.01648 0.01232 0.01893 0.01385
(0.04606) (0.04356) (0.05799) (0.05401)

 

 

 

 

 

 

Notes: ‘1 represents trade in differentiated products.
I’2 represents trade in homogeneous products
standard errors are in parentheses
*denotes significance at the 10% level

" denotes significance at the 5% level
*"denotes significance at the 1% level

149

Table A2.8 — Estimation Results with Initial Values and using instruments for y only

 

 

 

 

 

 

 

 

 

 

 

 

Regressor Cross- POLS FE—IV FD-IV
Sectional
Grav_Openla **-0.04278 -0.00205 ***0.15867 0.06391
$01693) (0.01588) (0.05414) (0_.05167)
Log (Initial "0.03506 -0.00065 ***-0.l7729 "-0.14790
Income) (0.01422) (0.01356) (0.06583) (0.05952)
Openness*log ***0.00515 0.00050 **-0.01616 -0.00327
(Income) (0.00199) (0.00181) (0.00675) (0.00657)
Population ***-1.12059 **-0.59562 0.46637 *0.43529
Growth (0.26149) (0.29682) (0.29143) 40.22621)
Log (Area) -0.00097
(0.00084)
Investment 0.03968 ***0.14277 ***0.08876 ***0.12018
(0.02457) (0.03868) (0.03338) (0.02949)
Government 0.00076 -0.00658 -0.03278 *-0.04609
(0.01275) (0.01346) (0.03007) (0.02455)

 

 

Notes: ‘Grav_Open1 represents the difference between trade predicted by the gravity model and actual
trade assuming that trade is in differentiated products.
the Huber/White/Sandwich standard error estimates are in parentheses

*denotes significance at the 10% level
** denotes significance at the 5% level
"*denotes significance at the 1% level

Table A2.9 — Estimation Results with Initial Values and using instruments for y only

 

 

 

 

 

 

 

 

 

Regressor Cross- POLS FE-IV FD-IV
Sectional
Grav_Open2a -0.01774 *0.03339 ***0.18034 0.06155
(0.02328) (0.01962) (0.05540) (0.05074)
Log (Initial 0.00437 **-0.01532 ***-0.10195 ***-0.13549
Income) (0.00827) (0.00748) (0.03974) (0.0301 1)
Openness*log 0.00257 -0.00317 ** *-0.01928 -0.00307
(Income) (0.00269) (0.00231) (0.00696) (0.00638)
Population ***-1.10859 **-0.59167 *0.54416 "0.45365
Growth (0.26629) (0.27733) (0.29987) (0.22599)
Log (Area) *-0.00146 * * *-0.00246
(0.00084) (0.00055)
Investment 0.03208 * **0. 13477 * * *0.09009 ** *0.12129
(0.02391) (0.03969) (0.03333) (0.02946)
Government -0.00144 -0.00691 -0.03936 *-0.04658
(0.01371) (0.01358) (0.03151) (0.02461)

 

 

 

 

 

Notes: 'Grav_Open2 represents the difference between trade predicted by the gravity model and actual
trade assuming that trade is in homogeneous products.
the Huber/White/Sandwich standard error estimates are in parentheses

*denotes significance at the 10% level
"”" denotes significance at the 5% level
"*denotes significance at the 1% level

150

 

 

BIBLIOGRAPHY

151

Ahmad, J ., Hamhirun, S. “Cointegration and Causality Between Exports and
Economic Growth: Evidence From the ASEAN Countries,” Canadian Journal
of Economics 29 (1996): 3413-6

Anderson, J .E., Van Wincoop, E. “Borders, Trade and Welfare,” NBER Working
Paper 8515 (2001)

Arellano, Manuel and Stephen Bond, “Some Tests of Specification for Panel Data: Monte
Carlo Evidence and an Application to Employment Equations,” Review of
Economic Studies 58(2) (1991)

Arellano, Manuel and Olympia Bover, “Another Look at the Instrumental Variable
Estimation of Error-Components Models,” Journal of Econometrics 68(1) (1995):
29-5 1

Baillie, R.T., Pecchenino, RA. “The Search for Equilibrium Relationships in
International Finance: the Case of the Monetary Model,” Journal of

International Money and Finance, 10 (1991):582-593

Baillie, R.T., Selover, D.D., “Cointegration and Models of Exchange Rate
Determination,” International Journal of Forecasting, 3 (1987):43—51

Banerjee, A. “Panel Data Unit Roots and Cointegration: An Overview,” Oxford
Bulletin of Economics and Statistics, Special Issue (1999) 0305-9049

Barro, Robert J ., “Economic Growth in a Cross-Section of Countries,” Quarterly Journal
of Economics 106 (1991):407-444

Barro, Robert J ., “Inequality and Growth in a Panel of Countries,” Journal of Economic
Growth 5 (2000):5-32

Barro, Robert J ., and Xavier Sala-i-Martin, “Economic Growth,” McGraw Hill (1995)

Barro, Robert J ., and Xavier Sala-i-Martin, “Economic Growth,” MIT Press (2003)

Bergstrand, J .H. “The Gravity Equation in International Trade: Some Microeconomic
Foundations and Empirical Evidence,” The Review of Economics and

Statistics 67 (1985): 474-481

Bhagwati, J ., Panagariya, A., Srinivasan, T.N. “Lectures on International Trade,” MIT
Press 1983

Blundell, Richard and Stephen Bond, “Initial Conditions and Moment Restrictions in
Dynamic Panel Data Models,” Journal of Econometrics 87(1) (1998):]15-143

Bollerslev, T. “Generalized Autoregressive Conditional Heteroskedasticity,” Journal

152

 

of Econometrics 31 (1986): 307-327

Cass, David, “Optimum Growth in an Aggregative Model of Capital Accumulation,”
Review of Economic Studies 32 (1965):233-24O

Chinn, M. D., Meese, R.A., “Banking on currency forecasts: how predictable is
change in money?” Journal of International Economics, 38 ( 1995): 161-178

Choi, 1. “Unit Root Tests for Panel Data,” Journal of International Money and
Finance 20 (2001): 249-72

Cochran, S.J., Mansur, I. “The Interrelationships between US. and Foreign Equity
Market Yields,” Journal of International Business Studies 22 (1991): 723-736

Coe, D., Helpman, B. “International R&D Spillovers,” European Economic
Review 39 (1995): 859-87

DeJong, David N. and Marla Ripoll, “Tariffs and Growth: An Empirical Exploration of
Contingent Relationships,” The Review of Economics and Statistics 88(4)
(2006):625-640

Enders, W. “Applied Econometric Time Series,” John Wiley & Sons, Inc. (1995)

Engle, R.F. “Autoregressive Conditional Heteroscedasticity with Estimates of the
Variance of United Kingdom Inflation,” Econometrica 50 (1982): 987-1007

Engle, R. F ., Granger, C.W. “Cointegration and Error Correction: Representation,
Estimation and Testing,” Econometrica 55 (1987): 251-76

Evans, Paul, “Consistent Estimation of Growth Regressions,” manuscript, available at
(1997)

Evenett, S.J., Keller, W. “On Theories Explaining the Success of the Gravity
Equation,” NBER Working Paper 6529 (1998)

Feenstra, Robert C., Markusen, James K, Rose, Andrew K. “Using the Gravity
Equation to Differentiate Among Alternative Theories of Trade,” The
Canadian Journal of Economics 34 (2001): 430-447

Frankel, J. “On the Mark: A Theory of Floating Exchange Rates Based on Real
Interest Differentials,” The American Economic Review, 69 (1979):610-622

Frankel, J .A., Romer, D. “Does Trade Cause Growth?” American Economic
Review 89 (1999): 379-99

Garces-Diaz, D., “How Does the Monetary Model of Exchange Rate Determination

153

Look When It Really Works?” Econometric Society 2004 North American
Winter Meetings, 60 (2004)

Granger, C.W. “Investigating Causal Relations by Econometric Models and Cross
-Spectral Methods,” Econometrica 37 (1969): 424-38

Grossman, G.M., Helpman, E. “Trade, Innovation and Growth,” American Economic
Review 80 (1990): 86-91

Grossman, G.M., Helpman, E. “Endogenous Product Cycles,” The Economic Journal
101 (1991a): 1214-1229

Grossman, G.M., Helpman, E. “Innovation and Growth in the Global Economy,” MIT
Press (1991b)

Groen, J .J .J . “Long Horizon Predictability of Exchange Rates. Is it for Real?”
Empirical Economics, 24 (1999):451-469

Groen, J .J .J . “The Monetary Exchange Rate Model as a Long-Run Phenomenon,”
Journal of International Economics, 52 (2000): 299-319

Hallak, J. C., Levinsohn, J. “Fooling Ourselves: Evaluating the Globalization and
Growth Debate,” NBER Working Paper 10244 (2004)

Hamilton, J .D. Time Series Analysis, Princeton University Press (1994)

Harvey, AC. “The Econometric Analysis of Time Series,” John Wiley & Sons Inc
(1981)

Hayo, B. “Money-output Granger causality Revisited: an Empirical Analysis of EU
Countries,” Applied Economics 31 (1999): 1489-1501

Head, K., Ries, J. “Increasing Returns versus National Product Differentiation as an
Explanation for the Pattern of U.S.-Canada Trade,” The American Economic
Review 91 (2001): 858-876

Higgins, Matthew J ., Daniel Levy and Andrew T. Young, “Growth and Convergence
Across the United States: Evidence from County-Level Data,” The Review of
Economics and Statistics 88(4) (2006):671-681

Hoffman, R., Ramasamy, B., Yeung, M. “FDI and Pollution: a Granger
Causality Test Using Panel Data,” JEL, F18, F21, Q2

Hooper, P., Kohlhagen, S. “The Effect of Exchange Rate Uncertainty on Prices and
Volumes of International Trade,” Journal of International Economics (1978)

154

 

Hurlin, C. “Testing Granger Causality in Heterogenous Panel Data Models
with Fixed Coefficients,” JEL C23, C11

Hurlin, C., Venet, B. “Financial Development and Growth: A Re-
Examination Using a Panel Granger Causality Test,” Laboratoire
d’Economie d’Orlean — FRE CNRS 2783

Hurlin, C., Venet, B. “Granger Causality Tests in Panel Data Models with
Fixed Coefficients,” JEL, C23, C11

1m, K.S., Pesaran, M. H., Shin, Y. “Testing for Unit Roots in Heterogeneous
Panels,” Journal of Econometrics 115 (2003): 53-74

Kaminsky, G.L., Lewis, K.K. “Does Foreign Exchange Intervention Signal Future
Monetary Policy?” Journal of Monetary Econonrics, 37:2 (1996): 285-312

Kang, H., Reuveny, R. “International Trade, Political Conflict/Cooperation, and
Granger Causality,” American Journal of Political Science 40 (1996): 943-970

Kao, C. “Spurious Regression and Residual-Based Tests for Cointegration in Panel
Data,” Journal of Econometrics, 90 (1999): 1 — 44

Koopmans, Tjalling C., “On the Concept of Optimal Economic Growth,” The
Econometric Approach to Development Planning (1965)

Krugrnan, P.R. “A model of innovation, technology transfer and the world distribution
of income,” Journal of Political Economy 87 (1979): 253-266

Leichenko, R.M. “Exports, Employment, and Production: A Causal Assessment of
US. States and Regions,” Economic Geography 76 (2000): 303-325

Levin, A., Lin, C.F. “Unit Root Tests in Panel DatazAsymptotic and Finite Sample
Properties,” Unpublished manuscript, University of California, San Diego,
(1992)

Levin, A., Lin, C.F., Chu, C.S.J. “Unit Root Tests in Panel Data: Asymptotic and
Finite-Sample Properties,” Journal of Econometrics, 108 (2002): 1-24

Levine, Ross and David Renelt, “A Sensitivity Analysis of Cross-Country Growth
Regressions,” The American Economic Review September (1992):942-963

MacDonald, R., Taylor, M. “The Monetary Approach to the Exchange Rate:
Rational Expectations, Long-Run Equilibrium and Forecasting,” IMF Staff
Papers, 40 (1993): 89-107

Mankiw, Gregory N., David Romer and David Weil, “A Contribution to the Empirics of

155

Economic Growth,” Quarterly Journal of Economics May (1992):407-437

Mark, N.C. “Exchange Rates and F undarnentals: evidence on long-horizon
Predictability,” American Economic Review, 85 (1995): 201-218

McCallum, J. “National Borders Matter: Canada-US. Regional Trade Patterns,”
American Economic Review 85 (1995): 615-623

McNown, R.A., Wallace, M. “Cointegration tests for long-run equilibrium in the
monetary exchange rate model,” Econonrics Letters, 31 (1989): 263-267

Meese, R.A., Rogoff, K. “Empirical Exchange Rate Models of the seventies: Do
they fit out of sample?” Journal of International Economics, 14 (1983): 3-74

Pedroni, P. “Critical Values for Cointegration Tests in Heterogeneous Panels with
Multiple Regressors,” Oxford Bulletin of Econorrrics and Statistics, Special
Issue (1999) 0305-9049

Pedroni, P. “Panel Cointegration; Asymptotic and Finite Sample Properties of Pooled
Time Series Tests, With an Application to the PPP Hypothesis,” Indiana
University Working Papers in Economics, No. 95-013, (1995) June

Pedroni, P. “Panel Cointegration; Asymptotic and Finite Sample Properties of Pooled
Time Series Tests, With an Application to the PPP Hypothesis: New Results,”
Working paper, Indiana University, (1997a) April

Rapach, D.E., Wohar, M.E. “Testing the Monetary Model of Exchange Rate
Determination: New Evidence from a Centtu'y of Data,” Journal of
International Economics, 58 (2002): 359-385

Rodrik, D., Subramanian, A., Trebbi, F. “Institutions Rule: The Primacy of
Institutions over Geography and Integration in Economic Development,
NBER Working Paper 9305, (2002)

99

Rodriquez, Francisco and Dani Rodrik, “Trade Policy and Economic Growth: A
Skeptic’s Guide to the Cross-National Evidence,” NBER Macroeconomics
Annual (1999)

Rodriquez, F., Rodrik, D. “Trade Policy and Economic Growth,” A Skeptic’s
Guide to the Cross-National Evidence,” NBER Macroeconomics Annual,
(2001) '

Rose, A.K., Lockwood, B., Quah, D. “One Money, One Market: The Effect of
Common Currencies on Trade,” Economic Policy, 15 (2000): 7-45

Sachs, Jeffrey D. and Andrew M. Warner, “Economic Convergence and Economic

156

Policies,” NBER Working Paper No. 5039 (1995)

Sarantis, N. “The Monetary Exchange Rate Model in the Long-Run: an Empirical
Investigation,” Weltwirtschaftliches Archiv, 130 (1994): 698-711

Solow, Robert M., “A Contribution to the Theory of Economic Growth,” Quarterly
Journal of Economics 70 (l965):65-94

Sozovska, A. “Exchange Rate Regimes in Transition Econonries,” USAID Fiscal
Reform Project Report, (2004)

Tao, X., Zestos, G.K. “Trade and GDP Growth: Causal Relations in the United States
and Canada,” Southern Economic Journal 68 (2002): 859-874

Trefler, D. “International Factor Price Differences: Leontief was Right,” The
Journal of Political Economy 101 (1993) 961-987

 

Vernon, R., “International Investment and International Trade in the Product Cycle,”
Quarterly Journal of Econorrrics 80 (1966): 190-207

Wacziarg, R., Welch, K.H. “Trade Liberalization and Growth: New Evidence,” NBER
Working Paper 10152 (2003)

Wooldridge, Jeffrey M., “Econometric Analysis of Cross Section and Panel Data,”
MIT Press (2002)

Wooldridge, J. “Introductory Econometrics: A Modern Approach, 2e,” South-
Westem (2003).

157

IIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIII

uI(11111110111110)11191611211I