Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
THREE ESSAYS ON DEPENDENT PANELS: EMPIRICAL EVIDENCE
A DISSERTATION SUBMITTED TO THE GRADUATE DIVISION OF THE
UNIVERSITY OF HAWAI‘I AT MANOA IN PARTIAL FULFILLMENT OF THE
REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
IN
ECONOMICS
AUGUST 2014
By
Qianxue Zhao
Dissertation Committee:
Carl S. Bonham, Chairperson
Byron Gangnes
Peter Fuleky
Sumner La Croix
David S. McClain
Keywords: cross-sectional dependence, panel unit root test, panel estimator
Dedication
I dedicate my dissertation work to my family.
ii
Acknowledgements
I own many thanks to all those professors who instructed me to the best of their knowledge
and made this dissertation possible, especially my committee members: Professor Carl S.
Bonham, Professor Peter Fuleky, Professor Byron Gangnes, Professor Sumner La Croix,
and Professor David S. McClain.
My deepest gratitude is to two professors from whom I learn most. I first want to thank
Professor Carl S. Bonham, one of my advisors and the chairperson, who introduced me to
the fascinating area of dependent panel techniques. My indebtedness to professor Bonham
is also due to his supports of various softwares for all my projects and his effective guidance
on conducting empirical researches.
I am also indebted a lot to Professor Peter Fuleky, my other advisor, who introduced
me to the programing language R and helped me become acquainted with it. I also benefit
from Professor Fuleky in studying many other economic packages, econometrical methods
and writing skills.
I also want to thank professor Luigi Ventura, my co-author from University of Rome,
for providing me the opportunity of working jointly with him.
I would like to thank the Department of Economics in University of Hawaii, the Hung
Fellowship and the University of Hawaii Economic Research Organization (UHERO) for
providing financial supports throughout the years.
Finally, I own my greatest debt to my grand parents, my parents, my sister, my dog
back home and my forever-love husband. It is their strong love that backs me up in difficult
times. I cannot be what I am now without it.
iii
Abstract
The assumption of cross-sectionally independent units in the panel data may fail due to
common shocks and spillover effects. This dissertation mainly deals with the issue of cross-
sectional dependence when conducting empirical researches. This objective is accomplished
by utilizing advanced panel methods. This dissertation consists of three empirical stud-
ies exploring the improvements in econometric methods to investigate three different yet
equivalently interesting topics.
The first essay contributes to the literature of tourism studies. It is the first paper to
account for cross-sectional dependence when estimating the tourism demand elasticities.
Using a quarterly panel of 48 states on the mainland of the US form 1993Q1 to 2011Q2, I
found that the conventional estimation method is unable to control for unobserved common
factors in the variables appropriately. As a result, it leaves common factors that are non-
stationary in the regression errors and causes counter-intuitive estimations. To solve the
problem of cross-sectional dependence, I use advanced methods for dependent panel and
reestimate the tourism demand elasticities for Hawaii.
In the second essay, I study the degree of consumption smoothing through international
markets using a annual panel of 158 countries during the year of 1970 to 2010. To estimate
the degree of consumption smoothing, I compare different methods of separating the com-
mon and the idiosyncratic shocks from observed data. I show that the conventional method
fails to control for aggregate shocks completely. I reestimate the degree of consumption
smoothing with the statistically defensible CCE estimators.
In the last essay, I re-examine the degree of gasoline market integration in the US,
accounting for both cross-sectional dependence and structural breaks. I test for the law of
iv
one price (LOP) within state-level retail gasoline markets. To deal with the adverse effects
of cross-sectional dependence and structural breaks on the residuals of the LOP regression
model, I propose a hybrid panel unit root test. Using the hybrid method, I fail to find a
constant cointegrating relationship between state gasoline prices and the national average
price in the US.
v
Contents
Acknowledgements iii
Abstract iv
List of Tables viii
List of Figures x
1 Essay 1: Estimating Demand Elasticities in Non-Stationary Panels: TheCase of Hawaii’s Tourism Industry 11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Tourism Demand Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.3 Methodology Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31.4 Common Correlated Estimator . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4.1 CCE Estimators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71.5 Panel Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5.1 Cross-Section Dependence Test . . . . . . . . . . . . . . . . . . . . . 91.5.2 Panel Unit Root Tests . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.6 Data and Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . 111.6.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111.6.2 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.7 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141.7.1 Pre-test for Unit Root in Variables . . . . . . . . . . . . . . . . . . . 141.7.2 Estimations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171.8.1 Income . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171.8.2 Price . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.9 Robustness Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201.9.1 Other Substitute Price . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2 Essay 2: Common Correlated Effects and International Risk-sharing 262.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.2 Regression Equation for International Risk-sharing . . . . . . . . . . . . . . 27
2.2.1 Theoretical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272.2.2 Empirical Model in the Literature . . . . . . . . . . . . . . . . . . . 292.2.3 Empirical Model in this Paper . . . . . . . . . . . . . . . . . . . . . 30
2.3 The Common Correlated Effect Estimator . . . . . . . . . . . . . . . . . . . 322.3.1 Common Correlated Effect Estimator . . . . . . . . . . . . . . . . . 32
vi
2.3.2 Relationship to Consumption Risk-sharing . . . . . . . . . . . . . . . 352.4 Empirical Strategy and Data . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.4.1 Roadmap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362.4.2 Cross-sectional Dependence Test and Panel Unit Root Test . . . . . 372.4.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.5 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392.5.1 Variable Test and Residual Diagnostic Test . . . . . . . . . . . . . . 392.5.2 Estimation for the Overall β . . . . . . . . . . . . . . . . . . . . . . 482.5.3 The Change of the Overall β Over Time . . . . . . . . . . . . . . . . 492.5.4 Individual β . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3 Essay 3: How Integrated are US Gasoline Markets: An Empirical Testwith Cross-sectional Correlation and Structural Breaks 573.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573.2 Examination of the Degree of Gasoline Market Integration . . . . . . . . . . 593.3 Empirical Strategy for the Test for the LOP . . . . . . . . . . . . . . . . . . 62
3.3.1 The Regression Model . . . . . . . . . . . . . . . . . . . . . . . . . . 623.3.2 A Hybrid Unit Root Test . . . . . . . . . . . . . . . . . . . . . . . . 64
3.4 Data and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 663.4.1 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 673.4.2 Test of the LOP based on Relative Prices . . . . . . . . . . . . . . . 76
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Appendix A 82A.1 Additional Tables for Essay 1 . . . . . . . . . . . . . . . . . . . . . . . . . . 82A.2 Additional Tables for Essay 2 . . . . . . . . . . . . . . . . . . . . . . . . . . 84
Appendix B 86B.1 Additional Figures for Essay 1 . . . . . . . . . . . . . . . . . . . . . . . . . 86
Appendix C 89C.1 The Comparison Between the Pooled Estimator and the Mean Group Estimator 89
Appendix D 91D.1 Univariate Test for the Presence of Structural Breaks in the Time Series . . 91D.2 Univariate Test for a Unit Root with Structural Breaks . . . . . . . . . . . 92D.3 Panel Test for Unit Roots with Structural Breaks and Common Factors . . 94
Bibliography 97
vii
List of Tables
1.1 Raw Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.2 Test for additive outliers in individual variable . . . . . . . . . . . . . . . . 19
1.3 Tests for Individual Variables . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.4 Panel Estimates Comparison, CCE and OLS . . . . . . . . . . . . . . . . . 20
1.5 Residual Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.6 Panel Estimates Comparison, CCE and FMOLS . . . . . . . . . . . . . . . 22
1.7 Robustness Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.1 Tests for Individual Variables . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.2 Residual Diagnostic Tests, Weighted Averages . . . . . . . . . . . . . . . . . 41
2.3 Residual Diagnostic Tests, Simple Average . . . . . . . . . . . . . . . . . . . 44
2.4 Mean Group Coefficient Estimates for Sub-Samples . . . . . . . . . . . . . . 50
2.5 The Effect of Financial Liberalization . . . . . . . . . . . . . . . . . . . . . 52
2.6 CCEMG Coefficient Estimates for Sub-Samples, in Sub-Periods . . . . . . . 52
2.7 Homogeneous test for individual CCE estimates . . . . . . . . . . . . . . . . 53
2.8 Country-Specific Coefficient Estimates . . . . . . . . . . . . . . . . . . . . . 54
2.9 Comparison of country-specific coefficient estimates (continued) . . . . . . . 55
3.1 Possible Results for the LOP Test . . . . . . . . . . . . . . . . . . . . . . . 64
viii
3.2 Tests for Individual Variables . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.3 Reference Table for States . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.4 Test for break in individual log (price) . . . . . . . . . . . . . . . . . . . . . 71
3.5 Test for break in individual OLS residuals . . . . . . . . . . . . . . . . . . . 73
3.6 Diagnostic tests for OLS residuals . . . . . . . . . . . . . . . . . . . . . . . 74
3.7 Diagnostic tests for OLS residuals, excluding Hawaii and Alaska . . . . . . 75
3.8 Test for break in individual log relative price (price) . . . . . . . . . . . . . 77
3.9 Diagnostic tests for relative price level . . . . . . . . . . . . . . . . . . . . . 78
3.10 Diagnostic tests for relative price level, excluding Hawaii and Alaska . . . . 80
3.11 Estimated Dates of Break in Common Factors . . . . . . . . . . . . . . . . . 81
A.1 State Code and Regional CPI . . . . . . . . . . . . . . . . . . . . . . . . . . 82
A.2 State Code and Regional CPI, continued . . . . . . . . . . . . . . . . . . . . 83
A.3 Sub-Sample Country Group . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
A.4 Sub-Sample Country Group, continued . . . . . . . . . . . . . . . . . . . . . 85
ix
List of Figures
1.1 The roadmap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.2 Time plots of standardized logarithms of variables and the cross-sectionalaverages (in red) from 1993Q1 to 2012Q1. . . . . . . . . . . . . . . . . . . . 15
1.3 Time plots of standardized FMOLS and CCE residuals. . . . . . . . . . . . 18
2.1 The roadmap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.2 Distribution of the γci,c and γyi,y loading coefficient estimates in the first-stageequations of CCE, see equations (2.23) and (2.24). . . . . . . . . . . . . . . 42
2.3 Distribution of the γyi,c and γci,y loading coefficient estimates in the first-stageequations of CCE, see equations (2.23) and (2.24). . . . . . . . . . . . . . . 43
2.4 Distribution of correlation coefficients Corr(ξcit, cit− ct) and Corr(ξyit, yit− yt). 45
2.5 Scaled estimates of idiosyncratic components (left), ξcit and ξyit and the cross-sectionally demeaned variables (right), cit − ct and yit − yt. . . . . . . . . . 46
2.6 Estimates of idiosyncratic components, ξcit and ξyit, and the cross-sectionallydemeaned variables, cit − ct and yit − yt for representative countries. . . . . 47
2.7 Distribution of country specific coefficient estimates for long-run and short-run. 56
3.1 Flow chart for empirical strategy of panel unit root test . . . . . . . . . . . 65
3.2 Individual state-level gasoline prices and Break date (vertical line) . . . . . 69
B.1 Distribution of individual coefficient estimates. . . . . . . . . . . . . . . . . 86
B.2 Distribution of individual coefficient estimates. . . . . . . . . . . . . . . . . 87
B.3 Distribution of individual coefficient estimates. . . . . . . . . . . . . . . . . 88
x
Chapter 1
Essay 1: Estimating Demand Elasticities in Non-Stationary
Panels: The Case of Hawaii’s Tourism Industry
1.1 Introduction
Since the early 20th century, the tourism industry has expanded considerably together with
the development of worldwide economy. Along with this expansion, studies of tourism in
terms of explaining tourism demand started to flourish. Two major objectives of these
studies are to quantify the effects from determinants of tourism demand and to forecast
future tourism demand (Song and Li, 2008).
Due to data availability, early empirical studies of tourism demand often used time
series data from a single origin-destination pair. To avoid spurious regressions, studies have
paid special attention to the unit root and the cointegration properties of the data. As a
consequence, these studies applied advanced methods, such as the autoregressive distributed
lag model (ADLM)(Song et al., 2003), the error correction model (ECM)(Kulendran and
Wilson, 2000; Kulendran and Witt, 2003; Lim and McAleer, 2001), the vector autoregression
(VAR) model (Song and Witt, 2006) or the vector error correction model (VECM) (Allen
et al., 2009). However, estimations from the literature vary widely and their use in forecast
and policy-making are limited.
As data availability grew across regions, there was a trend to utilize panel data. Panel
datasets have many advantages over time series datasets: it provides richer information
with variations in both temporal and cross-sectional dimensions (Song and Li, 2008); and
it can overcome the problem of multicollinearity and lack of degrees of freedom. Yet,
the assumption of independent cross-sectional units in conventional panel techniques does
not always hold for macroeconomic studies due to the presence of common shocks and/or
spillover effects. Without adequately dealing with cross-sectional dependence, regression
results will be misleading (Westerlund and Urbain, 2011).
There is a strand of literature aiming at solving the cross-sectional dependence in panel
estimations (Kapetanios et al., 2011; Pesaran, 2007; Pesaran and Tosetti, 2011). However,
these cutting-edge methodologies have not yet been considered in the tourism economics
literature and I try to fill this gap in this paper. In this paper, I estimate tourism demand
elasticities for US visitors who travel to Hawaii, accounting for the possibility of cross-
1
sectional interdependence caused by non-stationary common factors.
The rest of this paper is organized as follows: Section 2 illustrates a theoretical tourism
demand model and summarizes estimated values from the literature. Section 3 goes over ex-
isting methods in estimating tourism demand. Section 4 and 5 discuss econometric method-
ologies used for unit root test and panel estimation. Section 6 explains the data and presents
the empirical strategy. Section 7 illustrates both results of unit root tests and regressions.
Section 8 discusses the economic meaning of estimated values. Then in Section 9, some
robustness tests are reported. Finally, this paper concludes in Section 10.
1.2 Tourism Demand Model
According to the demand theory, the budget line for a tourist is determined by his income
and the price of goods and services. Specifically, the demand for aggregate tourism flows
from origin i to destination j can be expressed as
Dij = f(Yi, Pi, Pj , Ps) , (1.1)
where Dij is the tourism demand in destination j by consumers from origin i; Yi is the level
of income at origin i; Pi is the price of goods and services at origin i; Pj is the price of
tourism goods and services at destination j; Ps is the price of tourism products at competing
destinations of place j (Bonham et al., 2009).
In most international tourism studies, the domestic destination is assumed to be a sub-
stitute for the destination abroad; thus, the price level in the origin place is considered
as a proxy for Ps (Witt and Witt, 1995). Similarly, to examine the tourism demand of a
domestic destination, the price level at the origin in this paper is assumed to be a proxy for
the price of the substitute Ps.1 As a result, equation (1.1) becomes
Dij = f(Yi, Pi, Pj) , (1.2)
Assuming a homogeneous demand function, tourism demand can be written as a function
of real income, and relative price level
Dij = f
(YiPi,PjPi
). (1.3)
In the literature, the most popular measurement for tourism demand Dij is the number
of visitors from the origin to the destination (Li et al., 2005; Song and Li, 2008); alternative
measurements include tourist expenditures and tourist nights spent (Witt and Witt, 1995).
Besides the dependent variable, the choice for measurements of Yi varies with the purpose of
traveling. Witt and Witt (1995) recommended including private consumption or disposable
1This is a strong assumption. But it may be possible if one believes that local travels within the originarea may substitute for travels to the destination place.
2
personal income to explain and predict holiday visits, and a more general income measure,
such as national income, for business travel. With regard to the choice for Pj , two price
variables are widely used in the literature: the cost of travel to the destination and the cost
of living at the destination.
Some researchers go beyond explanatory variables Yi and Pj mentioned above. For
example, Yap and Allen (2011) examined the potential effects of variables such as consumer
perceptions about the future of the economy, household debt, and the number of hours
worked in paid jobs. Meanwhile, other studies included dummy variables for special events
such as Olympic games, policy changes, or natural disasters (Falk, 2010; Kuo et al., 2009).
Estimated elasticities from the tourism demand model are often used to forecast tourism
demand in the future and to provide insight to policymakers. Therefore, the accuracy
of estimations are undoubtedly important. Even with a consensus in modeling tourism
demand, empirical studies found a wide range of estimated values of coefficients. Witt and
Witt (1995) reported from a summary of 30 years’ worth of international tourism demand
studies that income elasticity ranges from 0.4 to 6.6 with median value of 2.4. In addition,
Crouch (1995, 1996) found that nearly 5% of the estimates were negative while conventional
opinion indicates that income elasticity should be between one and two. Such large variation
in estimated elasticities therefore limits the value of these empirical studies in policymaking
and future prediction.2
Similarly, for transportation cost, estimations of price elasticity range from -0.04 to -4.3,
with median value -0.5 in Witt and Witt (1995), and values vary between 0.11 and -1.89
in Crouch (1995). For cost at the destination, price elasticity ranges from -0.05 to -1.5,
with median value -0.7 in Witt and Witt (1995), and about 29% of the estimates reviewed
in Crouch (1995, 1996) have positive estimates. Moreover, counterintuitive results are
found for other explanatory variables. For example, Yap and Allen (2011) found a positive
relationship between domestic tourism demand and working hours in Australia; Aslan et al.
(2009) obtained a positive relationship between earthquakes and tourism demand. Crouch
(1995, 1996) investigated a vast number of factors that might cause differences among
studies. By applying meta-analysis to all studies, the varying methods used in estimation
was found to be significant for inter-differences of estimated coefficients.
1.3 Methodology Development
The development of estimating methodology enables us to obtain more precise measures
for income and price elasticities in the tourism demand model. Due to the lack of adequate
data from multiple countries, early tourism demand studies relied on time series data from a
single pair of countries. Applications of the conventional time series approach are diversified:
2In these studies, income elasticity with a negative value and values less than one are respectivelyexplained by the arguments of “inferior” tourism destination and the necessity of traveling to adjacentdestinations, such as short-haul international trips from the US to Canada.
3
ranging from exponential smoothing to vector autoregressive and error correction models
(Li et al., 2005; Witt and Witt, 1995). Recently, some alternative quantitative tools, such as
artificial neural networks, fuzzy time series, and genetic algorithms, have also been applied
in the tourism literature. (For a comprehensive survey of recent developments in tourism
demand modeling, see Song and Li (2008).) However, a time series dataset may not be
sufficient to provide enough variations for estimating the parameter of interest. For example,
Bonham et al. (2009) found that “the income elasticity in the just-identified US demand
relationship is implausibly large and estimated quite imprecisely.”
Fortunately, there is a way to obtain a better estimate of the interested parameters
by taking advantage of panel data. The advantage of panel data over time series data
is straightforward: with variations in both cross-sectional and time dimensions, a shorter
size in the time dimension can be compensated by the cross-sectional dimension to ensure
enough variations, and vice versa. As more data became available and econometric tools
advanced, a trend to exploit the richness of panel data emerged in the tourism literature
(Seetaram and Petit, 2012; Song and Li, 2008). Song and Li (2008) noted an increasing
passion for panel methods in literature after reviewing recent methodological development of
tourism demand studies. Moreover, Song et al. (2012) suggested that “future studies should
pay more attention to the dynamic version of panel data analysis and to more advanced
estimation methods.”
One popular model used in many panel tourism studies (for example, Garın-Munoz and
Montero-Martın (2007), Aslan et al. (2009), Garın Munoz (2007), Naude and Saayman
(2005), Habibi et al. (2009), Kuo et al. (2009), and Brida and Risso (2009)) is a dynamic
model proposed by Morley (1998). It is widely used in tourism demand studies due to
its capability to account for the habit persistence of tourists. In this dynamic model, the
first lag of the dependent variable is included on the right hand side of tourism demand
equation. Due to the inclusion of lagged dependent variable, estimation from traditional
method (OLS, Random effects or Fixed effects) is inconsistent. To obtain consistent es-
timations, many studies assume that autocorrelation is within two periods and make use
of the proposed method from Arellano and Bond (1991), which uses historical dependent
variables as instruments.
With the awareness of unit roots in and co-integration relationship among variables,
Seetanah et al. (2010) examined the inbound tourism to South Africa using a panel of 38
origin countries in a gravity model. They found that all variables are non-stationary based
on the Im et al. (2003) panel unit root test and were able to reject the null hypothesis
of no cointegration at 5 % level when using the cointegration test proposed in Pedroni
(1999). Therefore, they estimated their tourism model by using the fully modified OLS for
heterogeneous panels developed by Pedroni (1999, 2001) to eliminate the likely endogeneity
of the regressors and serial correlation.
All conventional panel estimation techniques and the recently developed approach in
4
Arellano and Bond (1991) as well as the FMOLS are based on the assumption of cross-
sectional independence. Yet, empirical studies in macroeconomics (for example, Baltagi and
Moscone (2010) and Holly et al. (2010)), and results presented in this paper for the tourism
demand in Hawaii, show that cross-sectional dependence is very likely to be present. Cross-
sectional dependence occurs as a result of both simultaneous impacts on individual regions
from common shocks (for instance, business cycle) and spillover influences from regions
nearby, and it is very common for a panel of national or regional data. Mathematical proof
and simulations in theoretical studies (Kapetanios et al., 2011; Pesaran, 2006; Phillips and
Sul, 2003) found that neglecting cross-sectional dependence results in substantially biased
estimators and suffers from size distortions even with large N, T.
One way to account for cross-section correlation is through the spatial model, which
has been used in existing studies (for example, Baltagi and Moscone (2010)). To model
cross-sectional correlation, the error term in the regression model is assumed to follow a
certain type of spatial process with pre-specified weights.3 In many empirical work, the
choice for the weighting matrix is subjective, and it is usually based on the inverse of the
distance across units. From an economic perspective, the distance alone may not capture the
magnitude of the dependence between units completely and correctly, and the geographic
distance cannot fully represent the economic distance. Therefore, tourism demand studies
explored cross-sectional correlation in a more general way. Early studies have used seemingly
unrelated regressions (SUR) to deal with cross-sectional correlation (Allen and Yap, 2009;
Ledesma-Rodriguez et al., 2001; Yap and Allen, 2011). The GLS type transformation to
purge cross-sectional dependence from a SUR model is only satisfactory when the time
dimension is larger than the cross-sectional dimension.
For a dynamic panel model with autoregressive process, Phillips and Sul (2003) sug-
gested to use a common time effect with individual-specific loadings to model cross-sectional
dependence. This is justified by the fact that common shocks might cause co-movements
between multivariate time series. However, in the proposed methodology, the series of
the common time effect is restricted to have variance one and mean zero for identifica-
tion. Moreover, the iterative procedure of obtaining unbiased estimator is valid only when
cross-sectional dimension is large while the number of time periods is relatively small.
Similar to Phillips and Sul (2003), another approach to model the cross-sectional de-
pendence is to use a common factor structure. Common factors are frequently discussed
in international finance and macroeconomics where each individual region may be subject
to national or global shocks such as business cycles, technological innovations, oil crises or
national fiscal and monetary policies. For instance, Beck et al. (2009) investigated the co-
movement of inflation variables for both regions across Euro area and 11 metropolitan areas
in the US. By applying the principal components analysis, they found that there are at least
three common components shared among individual series, which explain a large portion
3For detail discussion of spatial model, see Chudik et al. (2011); Pesaran and Tosetti (2011).
5
(up to 80 %) of the variance of the data. However common shocks will cause cross-sectional
dependence of variables in the estimated model, violating the identically-independent dis-
tribution assumed in the conventional panel estimation techniques, and consequently will
cause problem in the estimation procedures.
Cross-sectional dependence in the panel cannot be dealt in a satisfactory way in clas-
sic estimators. To solve this problem, Pesaran (2006), Bai et al. (2009), and Kapetanios
et al. (2011) recently proposed some innovative methods. By using a factor structure,
recent development in econometric theory makes it possible to model the cross-sectional
dependence through a vector of unobserved common factors. Pesaran (2006) modeled the
cross-sectional dependence among units via a vector of unobservable stationary common
factors. To eliminate the cross-sectional dependence, he augmented the simple OLS re-
gressions with cross-sectional means of dependent and independent variables. The panel
estimation procedures provided in Bai et al. (2009), and Kapetanios et al. (2011) capture
cross-sectional dependence via even non-stationary common factors. By making use of a
factor model, they model the heterogeneity of cross-sectional interdependence for individ-
uals through differentiated loading parameters of common factors. Compared with early
solutions to the issue of cross-sectional dependence, the spatial model and the method in
Phillips and Sul (2003), the factor model is more applicable because it allows for a more gen-
eral type of serial correlation of variables (even non-stationary properties) and unit-specific
influence of common shocks. In the following section, I will focus on the panel estimation
technique that utilizes such factor structure from Pesaran (2006) and Kapetanios et al.
(2011).
1.4 Common Correlated Estimator
It is common for researchers to augment income and price variables in equation (1.3) with
deterministic variables such as time trends to capture evolving consumer tastes, secular
growth, or decline in an industry; a constant term to account for destination amenities
such as natural assets or other factors that are time invariant; dummies to account for
one-time events such as terrorism, natural disasters, major sporting events, and oil crises;
seasonality; or changes in data definitions or collection methods. These types of events, if
otherwise neglected, might lead to bias in the estimated parameters (Bonham et al., 2009).
The method described in this section deals with such deterministic effects the same way as
it deals with unobserved common factors such as business cycles, technological shocks, or
policy changes. As a result, I do not need to subjectively select deterministic proxies for
these events.
The long-run relationship compatible with the theoretical tourism demand model (1.3)
can be written in the following log-linear form,
yit = αi + β′ixit + uit , i = 1, 2, . . . , N , t = 1, 2, . . . , T , (1.4)
6
where yit = log(Dij,t), xit =
(log
(Yi,tPi,t
), log
(Pj,tPi,t
))′. Coefficient βi represents the
elasticity of demand with respect to the regressors xit.
Following Pesaran (2007) and Kapetanios et al. (2011), I model the dynamics and the
common unobserved factors in the error terms uit. In particular, I assume that uit has the
following structure
uit = γ ′ift + εit , i = 1, 2, . . . , N , t = 1, 2, . . . , T , (1.5)
in which ft is an m× 1 vector of unobserved common effects. εit are the individual-specific
(idiosyncratic) errors assumed to be distributed independently of xit and ft, but they can
be weakly dependent across i, and serially correlated over time. The parameter vector of
the slope coefficients, βi, is heterogeneous across units. But in order to assess the overall
effects of the demand determinants, I will focus on the estimation of its average value.
Assuming a random coefficient model, βi = β + wi, where wi ∼ IID(0,Vw), the overall
demand elasticities are β = E(βi).
The vector xit could be correlated with unobserved common factors, ft, and generated
as
xit = ai + Γ′ift + vit , i = 1, 2, . . . , N , t = 1, 2, . . . , T , (1.6)
where ai is a k × 1 vector of individual effects, Γi is a m× k factor loading matrix, vit are
the specific components of xit distributed independently of the common effects and across
i, but assumed to follow general covariance stationary processes. A valuable feature of the
model is that the error term, uit, is allowed to be correlated with the regressors, xit, through
the presence of the factors, ft, in both. The assumption of stationary εit implies that if ft
contains unit root processes then yit, xit, and ft must be cointegrated.
1.4.1 CCE Estimators
To estimate demand elasticity, β, I use the Common Correlated Effects (CCE) estimator of
Pesaran (2006), which asymptotically eliminates cross-sectional dependence from the errors.
Because the error term uit, contains non-stationary common factors that are correlated with
the regressors, conventional estimators of the model in (1.4) are biased. Pesaran (2006)
suggested using cross-sectional averages of yit and xit to deal with the effects of unobserved
factors. Kapetanios et al. (2011) proved that the CCE estimators are consistent regardless
of whether common factors ft, are stationary or non-stationary. They have further shown
that the CCE estimator of the mean of the slope coefficients β, is consistent for any number
of factors. Moreover, Pesaran and Tosetti (2011) showed later that results for the CCE
estimator hold even when the loading coefficients, γi and/or Γi, are zero (or in other words,
under the case of weak cross-sectional dependence). The good performance of the CCE
estimator sharply contrasts with the principal component approach of Bai et al. (2009),
7
which requires an estimate of the number of factors, and was shown to be more biased than
the CCE estimators in Westerlund and Urbain (2011).
For each individual unit, the CCE estimator for βi is
βi = (X ′iMXi)−1X ′iMyi , (1.7)
where Xi = (xi1,xi2, . . . ,xiT )′, yi = (yi1, yi2, . . . , yiT )′, and M = IT −H(H ′H)−1H ′ with
H = (ι, X, y), where ι is a T×1 vector of ones, X is a T×k matrix of cross-sectional means
of the k regressors, and y is a T × 1 vector of the cross-sectional mean of the dependent
variable.
The CCE estimator is equivalent to the Ordinary Least Squares (OLS) estimator applied
to an auxiliary regression that is augmented with cross-sectional means of all variables
in the regression, containing individual-specific loading coefficients of the cross-sectional
averages. Thus, the CCE estimator of the βi coefficients captures the effect of the demand
determinants after controlling for co-movement across units. By allowing for heterogeneous
loadings, γi and Γi, the CCE estimator allows for differentiated effects on individual units
of common factors.
To get the mean value of individual slope coefficients, Pesaran (2006) proposed two
estimators. The CCE mean group estimator (CCEMG) is a simple average of individual
CCE estimators, βi, defined by
βCCEMG =1
N
N∑i=1
βi , (1.8)
The estimator for the variance of βCCEMG is given by
V ar(βCCEMG) =1
N(N − 1)
N∑i=1
(βi − βCCEMG)(βi − βCCEMG)′ . (1.9)
When the slope coefficients, βi, are homogeneous across units, efficiency gains can be
achieved by pooling observations over cross section units. Pesaran (2006) developed a
pooled estimator as
βCCEP = (
N∑i=1
X ′iMXi)−1
N∑i=1
X ′iMyi , (1.10)
with variance
V ar(βCCEP ) =1
NΨ∗−1R∗Ψ∗−1 , (1.11)
where
Ψ∗ =1
N
N∑i=1
X ′iMXi
T, (1.12)
8
and
R∗ =1
N − 1
N∑i=1
(X ′iMXi
T)(βi − βCCEMG)(βi − βCCEMG)′(
X ′iMXi
T) . (1.13)
For the CCE estimator, unobserved common factors are assumed to be captured by
cross-sectional means. As long as the residuals are stationary, the CCE pooled (CCE-P)
and the CCE mean-group (CCE-MG) estimators are both consistent under the random
coefficient model assumption (Pesaran and Smith, 1995). Furthermore, with the help of
Monte Carlo simulations, Pesaran and Tosetti (2011) showed that the CCE method is still
able to provide consistent estimation in the presence of many types of cross-sectional depen-
dence: strong dependence (for example, due to common factors) and/or weak dependence
(for example, spatial correlation).
Compared to other approaches, the CCE estimator has many advantages. First of
all, the factor structure can stand for a variety of types of cross-sectional dependence.
Thus, CCE estimator is especially useful when the real type of cross-sectional dependence
is unknown. Second, the validity of CCE estimator does not require ex ante information
about the unobserved common factors, and it also allows the factors to contain unit roots
and to be correlated with the regressors. Finally, the CCE estimator offers good finite
sample properties (Kapetanios et al., 2011; Westerlund and Urbain, 2011), and is relatively
simple to implement.
1.5 Panel Data Analysis
1.5.1 Cross-Section Dependence Test
Without cross-sectional dependence, conventional estimators can provide valid estimates.
Therefore, the necessity for the CCE method is implied by the existence of cross-sectional
dependence. A pre-test for cross-sectional dependence in variables can provide insight for
the best method.
Pesaran (2004) proposed a cross-sectional dependence (CD) test which is applicable to
a variety of cases. The CD test is based on the average of pairwise correlations of individual
units
CD =
√2T
N(N − 1)(N−1∑i=1
N∑j=i+1
ρij) ∼a N(0, 1) , (1.14)
where ρij is the estimate of correlation between units yi and yj . Specifically,
ρij = ρji =
∑Tt=1 yityjt
(∑T
t=1 y2it)
1/2(∑T
t=1 y2jt)
1/2. (1.15)
9
It is shown in Pesaran (2004) that the CD test is valid whether or not the individual series
contains unit roots and the rejection of the null hypothesis of cross-sectional independence
will inform the presence of the cross-sectional dependence.
1.5.2 Panel Unit Root Tests
Breitung and Pesaran (2008) classified panel unit root tests into two generations. Methods
in the first generation assume that individual series are cross-sectionally independent. In
contrast, procedures in the second-generation relax this restrictive assumption and capture
the cross-sectional dependence through a factor structure. Using Monte Carlo simulation,
Gengenbach et al. (2006, 2010) showed that the dependence across units will cause size
distortion in panel unit root tests if it is overlooked. Therefore, if cross-sectional dependence
is found, a second-generation panel unit test is more desirable.
Several panel unit root tests accounting for the cross-sectional dependence have been
recently proposed in the literature. The PANIC test of Bai and Ng (2004) requires an
estimation for the number of common factors as an input and is based on the consistent
estimation of the unobserved components from the principal component analysis. Then, it
tests for unit roots in both components separately. To test for unit roots in the common
component, Bai and Ng (2004) suggested implementing the ADF test if the number of
common factor is estimated to be one and a rank-type unit root test for multiple common
factors.
Pesaran (2007) proposed a test for unit roots in the idiosyncratic component. In his
paper, the cross-sectional dependence is modeled through an unobservable common factor.
To deal with the shared common factor, he augmented each individual regression for the
standard ADF test with cross-sectional averages of lagged level and first-differences of the
series been tested. The covariate ADF test based on the augmented regression is then
implemented to each panel individual unit. Since Pesaran showed that individual statistics
from the covariate ADF test are independent from each other, he then derived the statistics
for the panel unit root test, CIPS, by averaging the covariate ADF test statistics computed
for each unit following Im et al. (2003). According to Gengenbach et al. (2010), the CIPS
test of Pesaran (2007) has higher power than the PANIC tests for testing a unit in the
idiosyncratic component.
However, it is important to note that the test in Pesaran (2007) misses unit roots
in the common component (Gengenbach et al., 2010) because it deals with cross-sectional
dependence by an orthogonal projection of the data on cross-sectional means. Consequently,
a unit root test for common factors is still required in order to draw a conclusion about the
unit root in a panel dataset.
The approach proposed by Sul (2009) uses the cross-sectional mean as the proxy for the
common factors and proposes a covariate unit root test with recursive mean adjustment
(CRMA) to test the null hypothesis that the common factors are integrated of degree one.
10
The proposed method for testing a unit root in common factors does not require estimation
of the number of common factors as in Bai and Ng (2004). More importantly, it is shown in
Sul (2009) and Gengenbach et al. (2010) that for unit root test for common factors, CRMA
test has similar size and power properties to Bai and Ng (2004).
In this paper, the test for unit roots in a panel dataset consists of two parts: the first,
which tests for unit roots in the idiosyncratic component from Pesaran (2007) and the
second, which tests for unit root in the common component from Sul (2009) and Bai and
Ng (2004). The panel is non-stationary if the null hypothesis of unit root is not rejected in
either component, and is stationary when both components are found to be I(0).
1.6 Data and Empirical Strategy
In this paper, I estimate the demand elasticities for tourism from the U.S. Mainland to
Hawaii. I use the number of visitors arriving to Hawaii (V IS) as my measure of tourism
demand and the total real personal income by state (Y ) as the income variable in the model.
My selection of the income variable is mainly owing to Witt and Witt (1995) as the authors
suggest to use personal income or disposable income for leisure travel, which accounts for
the majority (over 70%) of tourists’s purpose of visiting Hawaii. In addition, I include both
the transportation cost (airfare: PAIR) and the hotel room rate (PRM) as a proxy for
the cost of the trip. The nominal variables are deflated using the consumer price index at
the origin (CPI), so that prices enter the model in relative terms. Formally, one of the
estimated models (denoted as Model 2) in this paper is
log V ISit = αi+β1i log
(Yit
CPIit∗ 100
)+β2i log
(PAIRitCPIit
∗ 100
)+β3i log
(PRMt
CPIit∗ 100
)+uit .
(1.16)
In this regression, the hotel room rate at the destination is independent of trip origins and
therefore can be considered as an observed common factor. As already demonstrated, the
CCE method is based on an orthogonal projection onto proxy of all common factors (both
observed and unobserved). Therefore, the inclusion of the room rate should not have any
substantial effect on the CCE estimation of other coefficients. I verify this argument by
also estimating an alternative model below (denoted as Model 1) that does not include the
room rate.
log V ISit = αi + β1i log
(Yit
CPIit∗ 100
)+ β2i log
(PAIRitCPIit
∗ 100
)+ uit . (1.17)
1.6.1 Data
I obtained tourist arrivals by state from the first quarter of 1993 through the first quarter
of 2012 from annual reports of the Department of Business, Economic Development and
Tourism (DBEDT), Hawaii Visitors and Convention Bureau (HVCB), and Hawaii Visitor
11
Bureau (HVB). With the exception of year 1995 and year 1997, the data is available at the
monthly frequency. For the year 1995 and year 1997, I interpolated the annual values using
the pattern calculated from monthly data of all other years. The data of the total personal
income was collected from the Bureau of Economic Analysis (BEA). Airfares were obtained
from the DB1B Market database of the Bureau of Transportation Statistics (BTS), which
offers a 10% random sample of all domestic trips in a given quarter. To avoid the effect of
outliers, I calculated the median airfare per quarter for each state. The hotel room rate series
was obtained from Hospitality Advisors LLC. The consumer price index (CPI) was obtained
from the Bureau of Labor Statistics (BLS). Since the CPI series are only reported at the
metropolitan level, I assigned the values from a specific metropolitan statistical area to a
given state (see Table A.1 and A.2 for more information). In addition, because CPI series
are reported at monthly, bi-monthly and semiannual frequencies, I linearly interpolated the
low frequency series to approximate their values at the highest (monthly) frequency. For
my analysis, I converted all data to quarterly frequency and seasonally adjusted all series
using the X-12 ARIMA method. I excluded Delaware and the District of Columbia due to
the lack of airfare data. The raw data is summarized in Table 1.1.
Franses and Haldrup (1994) pointed out that additive outliers might produce spurious
stationarity, so the Dickey-Fuller unit root test will over-reject the null of unit-root. To
avoid the effect of outliers, I followed the procedure suggested by Perron and Rodrıguez
(2003) to remove additive outliers in all series entering the regression model. Results are
summarized in Table 1.2.
1.6.2 Empirical Strategy
The roadmap is illustrated as Figure 1.1. After removing outliers from each series, I pre-test
each variable for cross-sectional dependence by following Pesaran’s (2004) CD test. To test
for a unit root in the common component of variables, I implement the CRMA test from
Sul (2009). The rejection of the null hypothesis will indicate the presence of a stationary
common component. For the idiosyncratic component, I use the CIPS test from Pesaran
(2007), which will imply stationarity under alternative hypothesis. In addition, I compare
coefficient estimates from the OLS method and the FMOLS method to the CCE method,
with the former two serving as illustrations of the case when cross-sectional dependence
is neglected. Particularly, the FMOLS estimation is supposed to be valid since it aims at
avoiding spurious regression.
Moreover, for each regression, I use the panel unit root test on regression residuals as
the test for a cointegration relationship. In particular, I obtain the residuals of both the
OLS and FMOLS methods based on the following model,
uit = yit − αi − β′xit , (1.18)
12
Step 1 Pre-‐test of variables
Test for cross-‐sec1onal dependence (CD): H0: cross-‐sec/onal independence
In the common component (CRMA)
In the idiosyncra1c component (CIPS)
Test for unit root: H0: unit root
Step 2 Model es1ma1on
Step 3 Residual diagnos1c test (Test for cross-‐sec1onal
dependence and unit root)
reject
Conven1onal method (OLS, FMOLS): ignoring
cross-‐sec/onal dependence
Advanced method (CCE): dealing with cross-‐sec/onal dependence
Figure 1.1: The roadmap
13
where αi and β are estimated coefficients. Because the common factors in the variables are
ignored during these estimation process, they are likely to be left in the regression residuals,
uit.
In contrast, the CCE method is designed to get rid of unobserved common factors by
cross-sectional averages. Therefore, the CCE residuals, obtained by εit from equation (1.5),
is free of common factors. Nevertheless, it may still contain some weak cross-sectional
dependence which cannot be captured by common factors.
To account for cross-sectional dependence in residuals, I implement the following diag-
nostic tests for all regression residuals. The first diagnostic test is the CD test. Rejection
of the CD test will indicate the presence of common factors, which is further confirmed
by applying the approach in Bai and Ng (2002) to estimate the number of common fac-
tors. Since there is a tendency to overestimate the number of common factors (Gengenbach
et al., 2010), I only choose the minimum number estimated from all information criteria.
If there are any common factors detected in residuals from OLS and FMOLS, I will apply
the CRMA test in Sul (2009) or the test in Bai and Ng (2004) to examine common unit
root. For idiosyncratic components, I rely on the test of Pesaran (2007). On the other
hand, if there is any common factor detected in CCE residual, I argue that the factor is
probably due to some high degree of spatial correlation (or in other words, it is the weak
dependence discussed in Pesaran and Tosetti (2011)). Thus, I only apply the panel unit
root test in Pesaran (2007), which controls for cross-sectional dependence and tests for unit
root in idiosyncratic components.
1.7 Results
1.7.1 Pre-test for Unit Root in Variables
As results in Table 1.3 indicate, I reject the null hypothesis of cross-sectional independence
in all the variables. Furthermore, plots in Figure 1.2 graphically illustrate the existence of
the cross-sectional dependence in each variable. In each plot, there is a common movement
shared across units. Specifically, these common trends represent common shocks that affect
all individual series at the same time. For instance, the level of income in all states are
affected by the global economic recession in 2008-2010, causing a decline in all real income
series in the same period. Similarly, the crude oil price impacts the price level of airfare
in all states, which induces a co-movement of real airfare series. Interestingly, the cross-
sectional mean of individual series, highlighted in red, seems to be a good indicator of the
co-movement.
The presence of these common trends also points out that the appropriate panel unit
root tests should take into account the cross-sectional dependence among units. Therefore,
I implement Sul (2009)’s CRMA statistics to test for a unit root in the common component
of each variable. Since variables are I(1) when the common component and/or the idiosyn-
14
1995 2000 2005 2010
−4
−2
02
4
time
in lo
gs
log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)log of arrivals and cross−sectional means(red)
1995 2000 2005 2010
−2
−1
01
2
time
in lo
gs
log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)log of income and cross−sectional means(red)
1995 2000 2005 2010
−3
−2
−1
01
23
time
in lo
gs
log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)log of airfare and cross−sectional means(red)
1995 2000 2005 2010
−3
−2
−1
01
23
time
in lo
gs
log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)log of room rate and cross−sectional means(red)
log(V ISit)log
(Yit
CPIit∗ 100
)
log
(PAIRitCPIit
∗ 100
)log
(PRMt
CPIit∗ 100
)
Figure 1.2: Time plots of standardized logarithms of variables and the cross-sectional aver-ages (in red) from 1993Q1 to 2012Q1.
15
cratic component has a unit root, results in Table 1.3 indicate that I cannot reject the null
hypothesis of non-stationarity for any of the variables.
1.7.2 Estimations
I estimate equation (1.16) and (1.17) by both CCE and OLS estimators. For the CCE
method, I obtain the average effect of explanatory variables both by the pooled estimator
and the mean-group estimator. The estimation results are shown in Table 1.4.
As suggested, estimates of the income elasticity and the price elasticity (for airfare and
room rate, respectively) from the CCE method have expected signs. When comparing
CCE estimations between two model specifications (with and without lodging prices), the
estimates for the income and the airfare elasticity are similar. As explained, this is because
the hotel room rate in Hawaii is a common factor across states, which will be controlled
implicitly in the CCE estimation of Model 1. In contrast, the OLS estimators have an
unexpected sign on hotel room elasticity and a much smaller income elasticity compared to
the CCE estimations.
Residual diagnostic test of these regressions are in Table 1.5. It indicates that there
seems to be some cross-sectional dependence in both the OLS and the CCE residuals. The
information criteria for estimating the number of common factors from Bai and Ng (2002)
further suggest that with the exception of the CCE residuals from the pooled regression of
Model 1, there is no common factor found in the residuals of the CCE method. As common
factors in the regression are shown to be well controlled in the CCE method, I relate the
single factor in CCE residuals from the pooled regression of Model 1 to a high degree of
weak dependence. Therefore, I only report unit root test results from Pesaran (2007).
For the OLS regression, a single common factor is found in residuals of all specifications.
Thus, I use both the ADF test on the estimated common factor and the CRMA test on the
cross-sectional mean to test for a unit root in the common component. Because the null
hypothesis of a unit root cannot be rejected, the common component of the OLS residuals
is non-stationary. This result combined with unit root tests for idiosyncratic components
in Table 1.5 suggests that the OLS estimates are invalid because they are spurious. By
contrast, the CCE regressions are valid as residuals are stationary according to CIPS test
from Pesaran (2007).
The first panel of Table 1.6 compares the CCE estimates of Model 2 (with lodging
price) to the fully modified ordinary least squares (FMOLS) estimates commonly used
in the tourism literature (Seetaram and Petit, 2012). Same as the OLS estimates, the
FMOLS estimation of room price elasticity has a wrong sign. Because both the OLS and
the FMOLS methods ignore common factors, such as global shocks, in the variables, the
parameter estimates partly associate business cycle fluctuations in arrivals with business
cycle fluctuations in room prices. Therefore, the positive room rate elasticity may capture
the fact that the common factors in these two variables are positively correlated.
16
The middle pane of Table 1.6 presents unit root tests in the FMOLS residuals using
the conventional methodology. The tPP , and tADF are the Pedroni (1999, 2004) tests for
the null hypothesis of no-cointegration based on the Phillips and Perron t-statistics, and
the augmented Dickey Fuller t-statistic, respectively. Both tests assume cross-sectional
independence, and both reject the null of no cointegration. Thus, the conventional Pedroni
test leads to the acceptance of FMOLS results.
However, Pesaran’s (2004) CD test in the last pane rejects the null hypothesis of cross-
sectional independence, suggesting that the conventional tests are misleading due to their
disregard of common factors in the residuals. This is further proved by the procedure in
Bai and Ng (2002) and Sul (2009). In particular, the result of Bai and Ng (2002) indicates
that there is one common factor in the FMOLS residuals. In addition, Sul (2009)’s CRMA
statistic, which tests for unit roots in the common factors, fails to reject the null of a unit
root in the FMOLS residuals, implying that the FMOLS estimates are still spurious. In
Figure 1.3, I illustrate the comparison between the individual series of the CCE residuals
and the FMOLS residuals. In this figure, some co-movement can be clearly seen in the
FMOLS residuals.
In sum, the rejection of unit root in the residuals of CCE regressions, εi,t from equation
(1.5), suggests that CCE estimations are not spurious, implying that the observed variables
and the unobserved factors are cointegrated with the cointegrating vector given by the CCE
estimates (Kapetanios et al., 2011).
1.8 Discussion
1.8.1 Income
The estimated average income elasticity of tourism demand from the U.S. mainland to
Hawaii is slightly greater than unity, implying that travel to Hawaii is likely to be regarded
as a luxury good.4 Although Hawaii is a domestic tourism destination for US visitors, its
far distance from the mainland likely makes a trip to Hawaii very income elastic.
The CCE estimates of the income elasticity presented in this paper are similar to the
value of 0.996 in Nelson et al. (2011), which included some observed and deterministic
common factors in their model, such as oil prices and a non-linear time trend. However,
it is much lower than 3.5 found in Bonham et al. (2009) who estimated a VECM with
cointegrating relationships identified as supply and demand relations.5
4As illustrated in Figure ??, income elasticity estimates for individual units can be both positive andnegative, but most of them are positive and clustered around unity. Moreover, t test of the overall incomeelasticity indicates that the value is not significantly different from 1. Specifically, in Model 1 of Table 1.4,the t statistics for the null hypothesis of β1 = 1 from CCE-MG and CCE-P are 0.52 and 1.22, respectively;in Model 2, the t statistics for the null hypothesis of β1 = 1 from CCE-MG and CCE-P are 1.04 and 1.57,respectively.
5As noted in Section 1.4, the CCE estimator controls for global trends in the panel, and in general
17
Table 1.1: Raw DataVariable Source Frequency Seas. Adj.
Visitor Arrivals (VIS) HTA M NoPersonal Income (Y) BEA Q YesMedian Roundtrip Airfare (PAIR) BTS Q NoAverage Daily Room Rate (PRM) HA M NoConsumer Price Index (CPI) BLS M, BM, S NoNote: Visitors Arrivals (VIS) excludes estimated in-transit passengers, returning Hawai’i residents
and intended residents from airline passenger counts. I obtained monthly visitor arrivals for 1993-
1994 from HVCB, for 1996-1998 form HVB, for 1999-2010 from DBEDT, and for 2011-2012 from
HTA. For 1995 and 1997 only annual visitor arrivals are available, so I estimate the monthly series
by interpolation. The CPI series for the states are approximations based on data available for
metropolitan areas and geographic regions.
Acronyms: HVCB - Hawaii Visitors and Convention Bureau; HVB - Hawaii Visitor Bureau; HTA -
Hawaii Tourism Authority; DBEDT - Hawaii Department of Business, Economic Development and
Tourism; BEA - Bureau of Economic Analysis; BTS - Bureau of Transportation of Statistics; HA
- Hospitality Advisors, LLC; BLS - Bureau of Labor Statistics; M - Monthly; BM - Bi-Monthly; Q
- Quarterly; S - Semiannual.
1995 2000 2005 2010
−4
−2
02
4
time
in lo
gs
residual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmolsresidual from fmols
1995 2000 2005 2010
−4
−2
02
4
time
in lo
gs
residual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rateresidual from mean group estimator, with room rate
FMOLS residuals CCE residuals
Figure 1.3: Time plots of standardized FMOLS and CCE residuals.
18
Table 1.2: Test for additive outliers in individual variableVariable critical level No. of series tested Additive outliers ? No. of series with outliers
log V IS 0.01 48 yes 28
log YCPI 0.01 48 yes 4
log PAIRCPI 0.01 48 yes 20
log PRMCPI 0.01 22 no 0
Note: The regression model is illustrated in equation (1.16).
Table 1.3: Tests for Individual Variablesvariables y x1 x2 xa3CD 155.35∗ 285.51∗ 193.23∗ 130.12∗
CRMA -1.13 -0.45 -0.66 -0.03
Note: y = log(V IS), x1 = log(
YCPI ∗ 100
), x2 = log
(PAIRCPI ∗ 100
), x3 = log
(PRMCPI ∗ 100
). The
null hypothesis of the CD statistic is cross-sectional independence. The CRMA statistic tests for
unit roots in the cross-sectional mean; the null hypothesis is unit root; the lag length is chosen by
the BIC. Statistical significance at the 5% level or lower is denoted by ∗.a: I use 22 series to implement the test.
In addition, although Crouch (1996) found that the omission of a price variable might
cause a positive bias in income elasticity, I find that results from the CCE method are only
marginally affected by dropping the variable of room rate from the model. This is because
the CCE estimator controls for the omitted price variable via proxies.
1.8.2 Price
Results in this paper indicate that the tourism demand for Hawaii is inelastic with respect
to airfare. If airfare increases by 10%, arrivals to the state are expected to fall by a little
more than 2%. Again, this value is fairly close to -0.211, the airfare elasticity estimate of
Nelson et al. (2011). The estimated hotel room price elasticity suggests that tourists are
more responsive to changes in room rates than to fluctuations in airfare. Particularly, a
10% drop in the hotel room rate is expected to generate 12% higher visitor arrivals, over
five times more than a corresponding drop in airfare. Facing a $1000 airline ticket and a
daily price of $200 for a double occupancy room, a couple on a ten-day trip has to split
their budget evenly between airfare and accommodation.
The difference between airfare and room rate elasticities could also be explained by a two
stage decision making process undertaken by travelers: in the first stage they choose a desti-
nation from a range of competing locations, and in the second stage, they pick their flights.
The idea of two-stage decision-making process is related to the idea of multi-stage decision-
produces different results than conventional estimators of time series data lacking a cross-sectional dimension.
19
Table 1.4: Panel Estimates Comparison, CCE and OLS
Model1 Model 2Coefficient CCE-MG CCE-P OLS-MG CCE-MG CCE-P OLS-MG
β1 1.09∗ 1.18∗ 0.62 1.20∗ 1.27∗ 0.29β2 -0.32∗ -0.36∗ -0.35 -0.23∗ -0.26∗ -0.39β3 -1.23∗ -1.20∗ 0.56Note: Regression equations are illustrated in equation (1.16) and (1.17). Statistical significance at
the 5% level or lower is denoted by ∗. Since the OLS regression is invalid, I do not report the level
of marginal significance for OLS coefficient estimates.
making process discussed in Strotz (1957), who describes this process as first deciding how
to allocate a budget among several groups of goods and then making independent spending
decisions within the groups. Applying this idea to tourism, Syriopoulos and Thea Sinclair
(1993), and Song et al. (2012) all discuss the multi-stage decision-making process and Bon-
ham and Gangnes (1996), Nicolau and Mas (2005) and Eugenio-Martin and Campos-Soria
(2011) base their models on it. When deciding whether to make a trip to Hawaii, it is likely
that tourists first choose the destination among a group of competing locations, which may
be affected by promotional activities, such as free nights or the attracting package from
hotels. In the next stage, tourists select their favorite flights to the chosen destination by
minimizing their cost. As a result, the effect on tourism demand of changes in room rate
price is greater than the effect of changes in the airfare level.
1.9 Robustness Test
The criticism that the population growth might drive both aggregate income and aggregate
visitors can be avoided via the implementation of CCE estimator.6 This is because the
unobserved common factors shared by the number of visitors and total income is controlled
by the augmented cross-sectional means. Alternatively, population can be controlled by
transforming variables to per-capita terms. In tourism literature, there are many papers
specifying the tourism demand model in per-capita terms. However, this type of specifica-
tion constrains the elasticity of population to be one if a log-linear model is employed (Witt
and Witt, 1995).
1.9.1 Other Substitute Price
In the previous section, the model assumes that traveling to places near the origin is a
substitute for traveling to Hawaii. This assumption may be questionable since the moti-
vation for traveling to Hawaii is quite different from visiting places nearby. Considering
6As discussed in Witt and Witt (1995), including population as an additional explanatory variable mayalso induce multicollinearity problem.
20
Table 1.5: Residual TestsModel1 Model 2
CCE-MG CCE-P OLS-MG CCE-MG CCE-P OLS-MG
CD-3.32∗ -3.72∗ 87.81∗ -3.09∗ -3.75∗ 68.09∗
NO. of CF0 1 1 0 0 1
CIPSlag=1 -27.80∗ -24.49∗ -16.16∗ -28.76∗ -25.95∗ -14.52∗
lag=2 -21.26∗ -17.78∗ -10.31∗ -22.94∗ -19.73∗ -8.80∗
lag=3 -20.03∗ -16.88∗ -7.96∗ -22.69∗ -19.42∗ -6.69∗
lag=4 -15.58∗ -11.92∗ -4.73∗ -19.25∗ -14.59∗ -3.57∗
ADF— — -1.17 — — -1.17
CRMA— — -1.70 — — -1.55
Note: The null hypothesis of the CD statistic is cross-sectional independence. No. of common
factors is estimated following Bai and Ng(2002). The CIPS statistic is the cross-sectional average
of individually CADF(p) statistics. For the CIPS statistics the null hypothesis is that all series
are non-stationary; the alternative hypothesis is that some series are stationary; the maximum lags
of CIPS is set equal to T 1/3 = 4, which determines the lag length for all units in panel for their
individual Covariate ADF test. The ADF test is for the estimated single common factor. The
CRMA statistic tests for unit roots in the cross-sectional mean; the lag length is chosen by the
BIC. Statistical significance at the 5% level or lower is denoted by ∗.
21
Table 1.6: Panel Estimates Comparison, CCE and FMOLS
Model 2
Coefficient CCE −MG CCE − P FMOLS
β1 1.20∗ 1.27∗ 0.34∗
β2 -0.23∗ -0.26∗ -0.37∗
β3 -1.23∗ -1.20∗ 0.52∗
FMOLS Residual Diagnostics
First Generation tPP tADF-6.90∗ -4.65∗
SecondGeneration CD NO. of CF tADF CRMA68.93∗ 1 -1.13 -1.56
Note: Regression equation is illustrated in equation (1.16). FMOLS is Pedroni’s (2001)
fully modified OLS estimator for heterogeneous cointegrated panels. CCE-MG and CCE-
P are the mean group and pooled common correlated effects estimators of Pesaran (2006).
tPP , and tADF are the Pedroni (1999, 2004) cointegration tests based on the Phillips
and Perron t statistics, and the augmented Dickey Fuller t statistic, respectively. CD is
Pesaran (2004)’s cross-sectional independence test, and CRMA is Sul (2009)’s unit root
test for the common factors and the lag length was chosen by the BIC. NO. of CF is
the number of common factors estimated following Bai and Ng (2002), and tADF is the
standard ADF test implemented to the estimated single common factor following Bai and
Ng (2004). Lag length is determined by BIC. Statistical significance at the 5% level or
lower is denoted by ∗.
22
that Hawaii is famous for its seaside attractions, I choose another competing destination of
Hawaii - Miami, Florida. This makes more sense as both destinations attract US domestic
visitors with its beaches and sunny weather.
Specifically, I include another price variable, the airfare price to Miami, Florida, into
the regression equation (1.16) as a proxy for a substitute price. I select the airfare price
to Miami by three different ways: the median airfare to Miami International Airport, the
median airfare to Fort Lauderdale-Hollywood International Airport and the median airfare
to either airport. Estimation results are summarized in Table 1.7. Compared to estimation
for equation (1.16) in Table 1.5 and Table 1.6, estimates of income and price elasticity
only change slightly. The elasticity for substitute price measured in airfare to Miami is not
significant at 5% level for most regressions.7
1.10 Conclusion
Regional macroeconomic variables are likely to be affected by non-stationary common fac-
tors, which will lead to inconsistent and biased estimates if it is ignored. In this paper,
I explore the factor structure of regional variables in a tourism demand model, test the
series for common and idiosyncratic unit roots, and estimate tourism demand elasticities
by controlling for non-stationary common factors. To my knowledge, this is the first paper
in the tourism literature that accounts for non-stationary common factors in the data.
I apply the CCE estimators proposed by Pesaran (2006) and Kapetanios et al. (2011) to
a tourism demand model for Hawaii. The common factors controlled in the CCE method
could be population, business cycle and others. Using quarterly data spanning the period
from the first quarter of 1993 to the first quarter of 2012, I obtain income elasticity slightly
over unity, fairly high hotel room price elasticity, but relatively low airfare elasticity. A
sound explanation of these estimates is provided in the discussion section.
Although the CCE method has many merits, such as its easy implementation and its
availability under a variety of cross-sectional dependence, this paper still has room for
improvement. First of all, from a policy standpoint, it would be interesting to examine the
differentiated elasticities in segmented markets according to criteria such as geographical
segment, or income levels. By contrast, this paper treats the US mainland as a single
market, ignoring the differences in the behavior of tourist from different regions. Thus,
future research can be more concentrated on estimations in different markets.
Secondly, the log-linear constant elasticity demand model might be a misspecification of
demand behavior. According to Witt and Witt (1995), the popularity of the log-linear model
stems from its convenience in thinking in terms of elasticities and its superior empirical
7Of course, it may be more reasonable to state that the competing destination of Hawaii for US visitorsmight be Caribbean. Alternatively, I may follow Song et al. (2003) which uses tourist arrivals weighted CPIfrom a group of competing destinations as measure for substitution price. However, due to the lack of properdata, I leave this question for future work.
23
Table 1.7: Robustness TestsIncluding Substitution Price
airfare to mia-int airfare to fort-laud airfare to either airportCoefficient CCE-MG CCE-P CCE-MG CCE-P CCE-MG CCE-P
β1 1.16∗ 1.15∗ 1.14∗ 1.14∗ 1.18∗ 1.16∗
β2 -0.22∗ -0.26∗ -0.24∗ -0.24∗ -0.24∗ -0.24 ∗
β3 -1.27∗ -1.28∗ -1.73∗ -1.49 ∗ -1.50∗ -1.38∗
β4 0.03 0.08∗ 0.04c 0.06c 0.05 0.10∗
CD-3.18∗ -3.78∗ -2.98∗ -3.83∗ -2.96∗ -3.80∗
NO. of CF0 0 0 0 0 0
CIPSlag=0 -32.49∗ -31.41∗ -32.51∗ -31.60∗ -32.51∗ -31.69∗
lag=1 -29.05∗ -26.52∗ -29.72∗ -27.21∗ -29.40∗ -26.70∗
lag=2 -23.31∗ -20.36∗ -24.41∗ -20.54∗ -23.90∗ -20.33∗
lag=3 -23.37∗ -19.82∗ -24.76∗ -20.11∗ -23.76∗ -19.03∗
lag=4 -18.53∗ -14.36∗ -20.43∗ -15.63∗ -19.36∗ -13.93∗
Note: Regression equation is log V ISit = αi + β1i log(
Yit
CPIit∗ 100
)+ β2i log
(PAIRit
CPIit∗ 100
)+
β3i log(
PRMt
CPIit∗ 100
)+ β4i log
(PAIR−MIAit
CPIit
)+ uit. PAIR −MIAit is the airfare from mainland
states to Miami, Florida. The null hypothesis of the CD statistic is cross-sectional independence.
No. of common factors is estimated following Bai and Ng (2002). The CIPS statistic is the cross-
sectional average of individually CADF(p) statistics. Statistical significance at the 5% level or lower
is denoted by ∗. Statistical significance at the 10% level or lower is denoted by c.
24
results in terms of “correct” coefficient signs and model fit. Alternatively, when market
share is modeled, a semi-log functional form is generally specified. This is another area to
explore in future work.
Finally, the demand for the tourism in Hawaii may have changed during the past two
decades. Therefore, the tourism demand model with constant elasticities may be misspeci-
fied. To account for the change in the demand elasticity, a future extension of this study is
to examine the presence of structural breaks in the coefficients of tourism demand model.
25
Chapter 2
Essay 2: Common Correlated Effects and International
Risk-sharing
2.1 Introduction
Consumption is an integral part of everyday life and people pay attention to the level of
their consumption over time. Generally, consumers are risk averse; variation in their con-
sumption resulting from income shocks make them worse off. In order to avoid undesirable
welfare loss caused by fluctuations in consumption and obtain a relatively steady amount
of consumption from year to year, consumption smoothing activities must be considered.
With adequate consumption smoothing, consumers can insure themselves against unex-
pected shocks to their income and avoid large variation in their consumption. Within an
economy, consumers can rely on formal institutions such as charities, and private insurance
or on informal mechanisms such as loans from neighbors, friends and relatives (Cochrane,
1991) to insure against idiosyncratic shocks to their income and wealth. Similarly, in an
integrated world economy, residents of different countries can brace themselves against
country-specific shocks through international market: different countries can trade goods,
services and financial assets with each other to smooth country-specific shocks.
Under perfect international risk sharing, consumption is smoothed because idiosyncratic
risk in the income is diversified away. The theoretical implication of perfect risk sharing
states that individual consumption is independent of idiosyncratic income shocks and fol-
lows aggregate consumption only. However, in reality, perfect consumption smoothing may
fail due to incomplete financial and real markets, and limited participation in risk-sharing
activities.
Early papers such as Cochrane (1991), Mace (1991) and Obstfeld (1994) have derived
the regression model for empirical examinations of consumption smoothing. Based on the
same theoretical model, most of the later empirical studies on international risk sharing
test whether or not the idiosyncratic fluctuation in consumption is independent of the
idiosyncratic fluctuation in income. In such empirical test, the correct measurement of
idiosyncratic fluctuations in consumption and income is crucial. Yet, existing studies largely
neglect this important detail.
In this paper, I demonstrate that the conventional methods of controlling for the effects
26
of common shocks in consumption and income regression impose some ad-hoc homogeneity
assumptions. In addition, I propose a different method to isolate idiosyncratic fluctuations
in consumption and income. In particular, the proposed method deals with aggregate shocks
in the same way as in the Common-Correlated Effect (CCE) Estimator of Pesaran (2006)
and Kapetanios et al. (2011). Based on diagnostic tests, I show that the new method can
deal with aggregate shocks more appropriately than conventional methods.
With a panel of 158 countries over 40 years, I test for the implication of full risk sharing,
i.e., the idiosyncratic fluctuation in consumption is independent of idiosyncratic fluctuation
in income. Results from the CCE-type regression are compared with conventional regression
models. I find that: first, by using the CCE estimator, the regression residual is cross-
sectionally independent and stationary; second, loading coefficients of aggregates obtained
from the CCE estimator suggest that restrictions imposed in conventional regressions are
incorrect and may lead to spurious regressions; third, according to the homogeneity test of
coefficients across units, the degree of consumption risk sharing and the speed of adjustment
vary across countries.
This paper also contributes to the literature by examining the effect of globalization on
consumption risk sharing. Based on the CCE estimations, I estimate the degree of consump-
tion risk sharing in both pre-globalization and post-globalization sub-periods.1 Although
there is substantial removal of capital controls in the post-globalization period, I do not
find empirical support for an improvement of consumption risk sharing during the period
of financial globalization.
The following section is constructed as follows: In Section 2, I discuss the empirical
regression following the conventional method and the one following the CCE method. In
Section 3, I illustrate how the CCE method works and its relationship with the consumption
risk sharing regression. I describe the empirical strategy and data for this paper in Section
4, and report empirical results in Section 5.
2.2 Regression Equation for International Risk-sharing
2.2.1 Theoretical Model
In an economy with a single good and N countries, each country has a finitely lived repre-
sentative agent that makes decision about consumption under uncertainty, represented by
state-date event, st. At a given st, country i receives an endowment of good, eist and realizes
output Yst .
The representative agent in country i aims to maximize the discounted life-time utility
1The era of globalization refers to a period with increasing in cross-border trade and financial flows (seeKose et al. (2008), for instance).
27
as (Mace, 1991; Obstfeld, 1994) :
maxCi,0,Ci,st
(U i(Ci,0) +T∑t=1
ρti∑st
πstUi(Ci,st)) , i = 1 . . . N, (2.1)
Here, U i is the utility function function, and Ci,0, Ci,st are the consumption decisions. In
addition, T is the time horizon, ρi is the time preference coefficient for the agent in country
i and πst is the probability of date-event st. In such economy, countries can trade Arrow-
Debreu securities, which has the price of qst for the date-event st, payoff 1 if the date-event
st happens and 0 otherwise.
With budget constraints at time t = 0 and t = s, which are (Ci,0 +∑
stqstYst) = ei0 and
(Ci,st = eist + Yst), equation (2.1) is equivalent to
maxYst
(U i(ei0 −∑st
qstYst) +T∑t=1
ρti∑st
πstUi(eist + Yst)) , i = 1 . . . N. (2.2)
The solution to this maximization problem is obtained by taking first-order derivative with
respect to Yst , which leads to,
U i′(Ci,0)qst = ρtiπstU
i′(Ci,st). (2.3)
With the risk-aversion coefficient µi, a CRRA utility, U i(Ci,st) =C
1−µii,st
1−µi , is assumed. Further
assuming that U i′(Ci,0) is normalized to unity, equation (2.3) can be written as,
qst = ρtiπstC−µii,st
. (2.4)
Taking the logarithm transformation of equation(2.4), individual consumption satisfies,
log qst = t ∗ log ρi + log πst − µi logCi,st . (2.5)
When we equalize the left hand side of equation (2.5) for country i and j, it shows that the
consumption series in country i is correlated with the consumption in country j,
logCi,st =t(log ρi − log ρj)
µi+µjµi
logCj,st (2.6)
or
logCi,st = aijt + bij logCj,st (2.7)
Where aijt =t(log ρi−log ρj)
µiand bij =
µjµi
.
Because equation (2.5) also holds for an appropriately defined aggregate consumption,
28
the right hand side variables from equation (2.7) can be replaced by the aggregate con-
sumption and leads to,
logCit = ait + bi logCwt (2.8)
where Cit is a measure for country i’s consumption, Cwt is an aggregate measure of con-
sumption. Equation (2.8) implies that under a complete market, individual consumption
should move with aggregate consumption only. Moreover, it also implies that the logarithm
of individual consumption should not be influenced by any other idiosyncratic variables net
of the effects from aggregate consumption.
2.2.2 Empirical Model in the Literature
In the literature, the following regression is used as an empirical test for the null hypothesis
of perfect risk sharing,
logCit = αi + bc,i logCw,t + βi(log Yit − by,i log Ywt) + εit, (2.9)
where Yit is the country i’s income, Ywt is the aggregate income, and as a result, (log Yit −by,i log Yw,t) is considered to be the idiosyncratic income or output shock. Because country’s
consumption follows the aggregate consumption and is independent of idiosyncratic income
with perfect risk sharing, existing studies test two implied hypotheses of bc,i = 1 and βi = 0.
Others test the implication of prefect risk sharing by first differencing variables in equation
(2.9) in order to avoid the unit root in the regression and misleading inference of coefficients,
for instance Mace (1991).
The regression in levels emphasizes the co-movement between relative consumption and
income in the long-run while the regression in log differences applies for short-run. As argued
in Artis and Hoffmann (2012), there are many advantages when looking at a regression at
lower frequency. First and foremost, Lucas (1987) recognized that the welfare gain by
eliminating transitory variation in income is negligible, whereas there will be a susbstantial
benefit if idiosyncratic shocks are persistent. Secondly, regressions at higher frequency
(i.e., business frequency) is likely to be affected by the changes in the relative contribution
of permanent and transitory shocks to consumption. As consumption reacts primarily
to permanent shocks in income, the smaller volatility of the business cycle due to the
Great Moderation may offset the effect of financial globalization in terms of consumption
smoothing.
In equation (2.9) or its counterpart when using log-differenced variables, βi acts as a test
for the null hypothesis of perfect risk-sharing either in the long-run or in the short-run. In
particular, because the value of β is zero under the null hypothesis of perfect risk-sharing,
the rejection of the null of βi = 0 indicates the lack of perfect risk sharing. Moreover, βi
measures the transmission of idiosyncratic income to idiosyncratic consumption, as a result,
its value is also a measure for the degree of risk-sharing. A smaller value of βi suggests a
29
higher degree of risk sharing.
Following such implication of the coefficient βi, many existing studies directly test
whether or not the idiosyncratic income has significant effect on idiosyncratic consump-
tion. To accomplish this test, idiosyncratic consumption and income must be isolated from
observed data.
Existing literature (for example, Cochrane (1991), Asdrubali et al. (1996), Sorensen
and Yosha (1998), Crucini (1999)) often impose bc,i = 1, by,i = 1 and subtract aggregate
consumption (income) from individual consumption (income) to approximate measures of
idiosyncratic fluctuations. The imposed value for bc,i and by,i is derived assuming perfect
risk sharing and homogeneity in the discount factor ρ and the risk-aversion coefficients µ.
As a result, the regression models are:
logCit − logCw,t = αi + β(log Yit − log Yw,t) + εit, (2.10)
∆ logCit −∆ logCw,t = αi + β(∆ log Yit −∆ log Yw,t) + εit (2.11)
As shown, these assumptions lead to the absence of individual-specific effect of logCw,t and
log Yw,t, which implicitly assumes that aggregate fluctuation will have the same magnitude
of effect across countries.
In most papers based on Equation (2.10, 2.11), cross-sectional means of logCit and
log Yit are used to proxy for the aggregates, logCwt and log Ywt, respectively. As a result,
equation (2.10) and (2.11) are equivalent to regressions with cross-sectionally demeaned
variables (denoted as DEM in this paper). In addition, as no individual βi is allowed,
the assumption of a constant degree of consumption smoothing across country is imposed
in equation (2.10, 2.11). To allow for individual country’s heterogeneity, the fixed effect
estimator is used in most panel studies (denoted as DEM-FE).2
2.2.3 Empirical Model in this Paper
As noticed, there are some assumptions maintained in DEM-FE regressions. In the first
place, to obtain the idiosyncratic fluctuation in consumption, the DEM-FE regression as-
sumes identical and constant relative risk aversion coefficient µi across countries. Specifi-
cally, equation (2.7) shows that bc,i = µwµi
. Therefore, when µi is identical across units, i.e.,
µi = µj = µw, the loading coefficient bc,i in equation (2.9) is equal to one under the null
2Some studies (for example Asdrubali et al. (1996), Lewis (1997), Sorensen and Yosha (1998), andFratzscher and Imbs (2009)) replace this explicit cross-sectional demeaning with an implicit approach whichincludes a time dummy in the pooled regression to control for the effect of common trend in variables:
logCit = αi + dt + β log Yit + εit , (2.12)
where dt is a time dummy. Mathematically, the empirical estimation strategy for equation (2.12) with timedummies is the same as the regression with demeaned variables.
30
of perfect consumption smoothing. When perfect consumption smoothing is absent, the ef-
fect of aggregate fluctuation varies across countries, regardless of the value of risk aversion
coefficient. Additionally, when the attitude toward risk differs across countries, equation
(2.7) implies that the coefficient bc,i is individual-specific even under the situation of per-
fect consumption smoothing. In either case, imposing bc,i = 1 to obtain the idiosyncratic
consumption is likely to cause misspecification.
As for the idiosyncratic income, the coefficient for the effect of aggregate income, by,i
is also restricted to be one based on a similar homogeneity assumption as consumption.
Although it is reasonable to keep this assumption for a group of similar countries, it may
lead to the misspecification of the model when the group of countries under analysis is very
different from each other in terms of their productive and financial structures, regulations,
or participations in international trade. Especially in the case of a large and heterogeneous
panel used in this paper, it is more reasonable to relax the homogeneity assumption on the
loading coefficient and allow for country-specific influence of aggregate shocks.
Some studies, Pierucci and Ventura (2010) for instance, relax the assumptions on bc,i,
by,i and lead to regression equations as below (denoted here as Partial Pre-Filtering, or
PPF). But, these studies may be plagued by another problem.
log Y idioit = log Yit − by,i log Yw,t, (2.13)
logCidioit = logCit − bc,i logCw,t, (2.14)
logCidioit = αi + β log Y idioit + εit (2.15)
In particular, the way that both DEM and PPF regressions deal with aggregate shocks
implicitly assumes that the fluctuation in output is exogenous to consumption. As equa-
tion (2.13) illustrates, the idiosyncratic income is obtained by only removing the aggregate
income, log Yw,t. Similarly, in equation (2.14), the idiosyncratic consumption is extracted
from the data after controlling the aggregate consumption, logCw,t.
However, the assumption of uncorrelated consumption and output may fail due to rea-
sons other than market incompleteness (Becker and Hoffmann, 2006). One example will
be the non-separability between consumption and leisure in the utility function (Backus
et al., 1992). As a result, equation (2.10, 2.11) and (2.13 to 2.15) may be misspecified as it
neglects the possible correlation between fluctuations in consumption and income. Obstfeld
(1994) examined the effect of consumption from both aggregate consumption and aggregate
output for G-7 countries. Despite the high correlation between aggregate consumption and
output, the author still concluded that for France, Germany and Japan, consumption is
more correlated with aggregate output than with aggregate consumption.
31
Finally, there is another homogeneity assumption imposed in the existing test. Most
existing literature disregard heterogeneity in the coefficient β and estimate the overall β
from the fixed effect estimator. Pooling can bring efficiency gain when there is high degree of
homogeneity across individual units, otherwise, the ignored heterogeneity in the coefficient
of interest may lead to small sample bias in the estimation (See appendix for detailed
discussion about the pooled and mean-group estimator as in Coakley et al. (2001)). Given
a mix of over 100 countries in this paper, I also relax the assumption of a fixed β across
units and use the cross-sectional average of individual βi as a measure for the overall effect
(denoted as MG in this paper).
In this paper, I propose a different method to obtain measures of idiosyncratic fluctu-
ations in consumption and income. Recalling equation (2.9) above and rearranging terms,
the equation for risk-sharing test with fewest number of restrictions will be,
logCit = αi + γCi logCw,t + βilogYit + γYi logYw,t + εit , (2.16)
When cross-sectional means are used as proxy for aggregate series, equation (2.16) coincides
with a Common-Correlated-Effect (CCE) estimator for dependent panel regressions which
will be discussed in the following section.
2.3 The Common Correlated Effect Estimator
2.3.1 Common Correlated Effect Estimator
Pesaran’s (2006) common correlated effects (CCE) estimator, which deals with dependen-
cies across units in heterogeneous panel, is an ideal tool for an estimation with idiosyncratic
effects. The CCE estimator accomplishes this task because it accounts for common factors,
such as global business cycles or common shocks, allows for individual-specific effects of
these factors, and produces consistent coefficient estimates based on idiosyncratic fluctua-
tions in the data.
Specifically, the CCE estimator asymptotically eliminates the cross-sectional depen-
dence in a panel regression by augmenting it with cross-sectional means of variables in the
regression. Considering the following regression with panel dataset:
xit = αi + βizit + uit , i = 1, 2, . . . , N , t = 1, 2, . . . , T , (2.17)
where zit is a k× 1 vector of observed individual-specific regressors for the ith cross section
unit at time t. The dynamics and the unobserved common factors are modeled in the error
terms uit. In particular, uit is assumed to have the following structure
uit = γ ′ift + εit , (2.18)
32
where ft is an m× 1 vector of unobserved common effects with individual specific loading
vector γi. εit are the individual-specific (idiosyncratic) error, assumed to be distributed
independently of zit and ft. However, the εit are allowed to be weakly dependent across i,
and serially correlated over time.
The CCE estimator is based on the assumption that the regressor, zit, is generated as
zit = Ai + Γ′ift + vit , (2.19)
where Ai is the individual time-invariant effect, and Γi is a m × 1 loading vector for the
effect from common factors. The idiosyncratic component vit is distributed independently
of the common effects and across i. The error term in equation (2.17), uit, is allowed to
be correlated with the regressor, zit, through the presence of common factors in both, and
failure to account for this correlation will generally produce biased estimates of parameters
of interest.
Pesaran (2006) suggested using cross section averages of both regressant and regressors
to deal with the effect of unobserved factors. The CCE estimator is defined as,
βi = (z′iMzi)−1z′iMxi , (2.20)
where zi = (zi1, zi2, . . . , ziT )′, xi = (xi1, xi2, . . . , xiT )′, and M = IT − H(H ′H)−1H ′ with
H = (ι, z, x), and ι is a T × 1 vector of ones. z is a T × 1 matrix of cross-sectional means
of zit, and x is a T × 1 vector of cross-sectional means of the dependent variable.
While Pesaran (2006) derived the CCE estimator for stationary variables and factors,
Kapetanios et al. (2011) proved that the CCE estimator is consistent whether or not the
common factors, ft, are stationary. In the latter case, xit, zit, and ft are cointegrated if εit
(in equation 2.18) is stationary. The authors also showed that under a random coefficient
model, βi = β+ωi, where wi ∼ IID(0,Vw), both the individual and the mean of the slope
coefficient CCE estimates are consistent for any number of unobserved factors. To estimate
the mean value of the slope coefficient, two estimators are proposed. The CCE mean group
estimator (CCEMG) is a simple average of individual CCE estimators, βi,
βCCEMG =1
N
N∑i=1
βi . (2.21)
When slope coefficients, βi, are homogeneous, efficiency gains can be achieved by pooling
33
observations over cross section units, and the pooled estimator (CCEP) is defined as,3
βCCEP = (N∑i=1
y′iMyi)−1
N∑i=1
y′iMci . (2.22)
More specifically, the CCE estimation can be reconsidered as a two-stage procedure. In
the first stage, common effects are filtered out from the data by regressing each variable on
the cross-sectional averages of all variables in the model,
cit = αi,c + γci,cct + γyi,cyt + εcit , (2.23)
yit = αi,y + γci,y ct + γyi,yyt + εyit , (2.24)
where γci,c and γyi,c in equation (2.23) denote the individual-specific effect from aggregate
consumption and income, on individual consumption respectively; γci,y and γyi,y in equation
(2.24) capture the influence of aggregates on individual income, respectively. In the second
stage, the CCE estimate of an individual βi is obtained by regressing the residual εcit,
capturing idiosyncratic variation in consumption on the residual εyit, which captures the
idiosyncratic variation in income. Alternatively, βCCEP is obtained from a pooled regression
of εcit on εyit.
Although the CCE estimator admits both simple and weighted cross-sectional averages
in the M matrix, unequal weights may distort the inference if they overstate the importance
of outliers in the cross-sectional distribution of the data. For example, if a variable of interest
is in per capita terms, each country could be weighted by its population share, so that the
aggregate becomes a global per capita measure,
N∑i=1
(Cit ∗ wit) = Ct, wit =Nit
Nt, i = 1, 2, . . . , N , t = 1, 2, . . . , T , (2.25)
where C stands for consumption per capita, and N stands for population. This weight-
ing scheme overweighs countries with a large population. If some of these countries are
atypical, inferences will be distorted. Specifically, if the proxies for the common factors are
biased towards outliers, the CCE procedure will not be able to eliminate cross-sectional
dependence in the panel. Furthermore, most macroeconomic time series, such as consump-
tion and income, require a log-transformation before being fed to linear models. Such a
non-linear transformation will affect the location of the aggregate measure relative to the
distribution of the individual country level variables, and further distort inferences. In a
large panel dataset used in this paper, the size of population may not be a good weight
for consumption (income) per capita following all reasons mentioned above. Thus, simple
3In the CCEP, observations are pooled after a unit-by-unit orthogonal projection onto the cross-sectionalmeans. That is, the CCEP estimator allows for idiosyncratic loadings, γi, Γi, while estimating a commonslope coefficient for the variable of interest, β.
34
cross-sectional means entering the M matrix might be more appropriate to help handle
unobserved common factors.
2.3.2 Relationship to Consumption Risk-sharing
To examine the degree of international risk-sharing, empirical analysis requires control of
common shocks from individual variables and regressing idiosyncratic consumption on id-
iosyncratic income,
cidioit = αi + βiyidioit + uit , i = 1, 2, . . . , N , t = 1, 2, . . . , T , (2.26)
where cidioit , yidioit are idiosyncratic consumption and idiosyncratic income, respectively. With
the help of advanced techniques, aggregate shocks in individual consumption (income) can
be approximated by unobserved common factors. To be more specific, a particular country’s
observed income series, yit can be decomposed into two unobserved components. By defini-
tion, the common component can be captured by common factors, ft, and its contribution
to individual country’s observed income is measured by the factor loadings, λi,y,
yit = λ′i,yft + ξyit , (2.27)
where λi,y allows countries to be heterogeneous in terms of their sensitivity to global shocks.
The term λ′i,yft yields the amount of fully diversified income for country i, and the balance,
ξyit = yit − λ′i,yft, is the idiosyncratic component of observed income.
To control for common factors in both consumption and income and obtain the es-
timated coefficient for idiosyncratic variations, cross-sectional averages of both variables
can be incorporated in the regression following the CCE method from Pesaran (2006) and
Kapetanios et al. (2011), which leads to:
cit = αi + βiyit + γci ct + γyi yt + uit . (2.28)
Here cit and yit are observed individual consumption and income series in logarithm trans-
formation, and γci and γyi are the individual specific loading coefficients from aggregate
measures, respectively. Because Pesaran (2006) already showed that βi and the mean value,
β, can be consistently estimated by the augmented regression above, coefficient estimates
reflect the effect of income on consumption after controlling for common factors in the data.
Equation (2.28) takes the same functional form as equation (2.16) derived in the pre-
vious section, suggesting that the estimated coefficient βi from equation (2.16) measures
the transmission of idiosyncratic income to idiosyncratic consumption. This is because the
augmentation of aggregate consumption and income can proxy the common movement of
variables in the regression and help to isolate the idiosyncratic variations. It will be il-
35
lustrated later in the results section through some diagnostic tests that residuals from the
augmented regression are cross-sectionally independent. By contrast, DEM-type and PPF-
type regressions leave common factors in the regression residual which leads to a spurious
regression.
2.4 Empirical Strategy and Data
2.4.1 Roadmap
In this paper, I examine the degree of international risk-sharing in the long run using the
CCE method as equation (2.16) or (2.28). The CCE method allows me to handle all types
of aforementioned heterogeneity. To show that conventional demean (DEM) and partial
pre-filtering (PPF) methods have problems in dealing with aggregate shocks appropriately,
I perform residual diagnostics to each type of regression. As the roadmap (Figure 2.1)
illustrates, I first test for cross-sectional dependence in regression residuals using the CD
statistics proposed by Pesaran (2004). This is because cross-sectional dependence might
be present in regression residuals if the regression cannot appropriately control aggregate
shocks in the variables. Next, when the CD test rejects the null hypothesis of cross-sectional
independence, I use unit root tests that take into account cross-sectional dependence to
avoid distorted size and power discussed in Banerjee et al. (2004) and Gengenbach et al.
(2010). In particular, for the unit root test, I apply the CRMA test of Sul (2009) to test
for a unit root in common factors of the residual, and test for unit roots the idiosyncratic
component using the CIPS statistics of Pesaran (2007). As a robustness check of the
residual diagnostic test, I apply the same empirical procedure to a shorter sample which
excludes observations of the most recent recession.
For CCE regression, I also examine the degree of risk-sharing in the short-run using an
Error Correction Model (ECM) when the residual from the long-run regression is stationary,
and the equation for ECM under CCE estimator is:4
∆cit = βSRi ∆yit+Υci∆ct+Υy
i∆yt−κi(cit−1−αi−βLRi yit−1−γci ct−1−γyi yt−1)+eit (2.29)
To measure the mean value of the slope coefficient in the sample, I report βCCEMG
calculated by averaging individual βi. I compare estimations of CCEMG regression with
two alternatives: DEMFE and DEMMG. As discussed in Section 2 and 3, I suggest a
different method to obtain idiosyncratic consumption and income instead of using demeaned
4To estimate this model, I follow the two-stage estimation procedure from Engle and Granger (1987).In the first step, the long-run equilibrium relationship between common factors, idiosyncratic income andconsumption is estimated and the deviation is captured by the regression residual as (cit−1−αi−βLRi yit−1−γci ct−1 − γyi yt−1). In the second stage, the speed at which this equilibrium error is corrected, κ, can then beestimated along with the extent of risk sharing in the short run, βSR.
36
variables, a common approach in the literature. Moreover, I also relax the assumption of
a common coefficient β across countries. The comparison between DEMFE, DEMMG and
CCEMG will shed light on the relative merits of CCE method with such large panel data.
In addition to the mean effect for the whole sample, I also look at the subsample’s mean
effects. For each subsample, I further look at the change in the extent of risk sharing over
the past forty years to examine whether globalization has led to an increase in interna-
tional consumption smoothing. Finally, I examine individual estimates of βi and discuss
individual’s degree of risk-sharing.
2.4.2 Cross-sectional Dependence Test and Panel Unit Root Test
The international risk sharing hypothesis postulates that individual consumption and out-
put series across countries follow a similar pattern. The presence of common patterns across
countries can be tested by the cross-sectional dependence (CD) statistic of Pesaran (2004).
The test is based on the pairwise correlation of individual units, and has been shown to
have good finite sample properties in heterogeneous panels. Moreover, it is shown that the
CD test is valid regardless of whether individual series have unit roots. The statistic for
the CD test is:
CD =
√2T
N(N − 1)(
N−1∑i=1
N∑j=i+1
ρij) ∼a N(0, 1) , (2.30)
where ρij is the estimate of correlation between units i and j. Specifically,
ρij = ρji =
∑Tt=1 yityjt
(∑T
t=1 y2it)
1/2(∑T
t=1 y2jt)
1/2, (2.31)
where ys,t for s = i, j are individual series. The null hypothesis for the CD test is cross-
sectional independence, and the rejection suggests cross-sectional correlation.
Breitung and Pesaran (2008) classified panel unit root tests into two generations. The
first generation tests assume cross-sectional independence of individual series, while proce-
dures in the second-generation overcome this drawback by capturing cross-sectional depen-
dence through a factor representation. Using Monte Carlo simulations, Gengenbach et al.
(2006, 2010) showed that the dependence across units will cause size distortion in panel
unit root if it is overlooked. Thus, it is more desirable to use the second-generation panel
unit test when the null hypothesis of the CD test is rejected.
Several panel unit root tests that account for cross-sectional dependence have been
recently proposed in the literature. With the help of a factor structure, panel unit root test
can be implemented by testing for unit root in the common component and the idiosyncratic
component separately. The panel is found to be non-stationary if a unit root is found in
either component and stationary, if both components are stationary.
Pesaran (2007) proposes a test for unit root in the idiosyncratic component. To deal
37
Step 1 Pre-‐test of variables
Test for cross-‐sec1onal dependence (CD): H0: cross-‐sec/onal independence
In the common component (CRMA)
In the idiosyncra1c component (CIPS)
Test for unit root: H0: unit root
Step 2 Model es1ma1on
Step 3 Residual diagnos1c test
(Test for cross-‐sec1onal dependence and unit root)
reject
Advanced method (CCE):
relaxing homogeneity assump/ons
Conven1onal methods (DEM and PPF):
with homogeneity assump/ons
Figure 2.1: The roadmap
38
with cross-sectional correlation, the author first augmented individual standard ADF re-
gressions with cross-sectional averages of lagged levels and first-differences of the series been
tested. Then, he applies the covariate ADF test based on the augmented regression to each
individual series in the panel. Because individual covariate ADF test statistics are shown to
be independent from each other, he derives a panel test statistics, the CIPS, by averaging
individual statistics as in Im et al. (2003).
Because Pesaran’s (2007) CIPS test, which only focuses on the idiosyncratic component,
controls cross-sectional dependence through an orthogonal projection of the data on the co-
variates, it overlooks unit roots in common factors (Gengenbach et al., 2010). Therefore,
the unit root test for common factors is also required.
The approach proposed by Sul (2009) uses the cross-sectional mean as proxy for the
average of unobserved common factors and implements a covariate unit root test with
recursive mean adjustment (CRMA) to test the null hypothesis that the common factors
are integrated. The test for unit root in cross-sectional average overcomes the estimation of
the number of common factors required in other methods such as Bai and Ng (2004). Sul
(2009) and Gengenbach et al. (2010) showed that for the unit root test of common factors,
CRMA test has good size and power properties.
2.4.3 Data
In this paper, I use annual data from the Penn World Tables, version 7.1, released in
November 2012 (Heston et al. (2012)).5 This is a comprehensive dataset, covering more
than 170 countries over a fairly long time span. To maximize the number of countries with
continuously available data in the panel, I select the period from 1970 to 2010. All values are
expressed in US dollars, so as to make comparisons across countries and time feasible. From
the Penn World Tables, I collect the following variables: purchasing power parity (PPP)
converted GDP per capita at 2005 constant prices (Laspeyres), PPP-converted consumption
per capita at 2005 constant prices, and population.
2.5 Empirical Results
2.5.1 Variable Test and Residual Diagnostic Test
Considering the fact that individual consumption and income are both affected by common
shocks such as global crisis and world-wide technological change, it may be more reasonable
to use second generation panel unit root tests to examine the time series properties for each
variable. Pre-tests for logarithm transformed real consumption per capita and real GDP per
capita are summarized in Table 2.1. Both the CD test and the estimation for the number
5Analyzed series from Penn World Tables are comparable to those in other datasets. For example,growth rates of real GDP and real final consumption almost perfectly match those in the last release of theWorld Bank’s World Development Indicators.
39
of common factors indicate that there is cross-sectional interdependence for both series due
to the presence of common factors. Moreover, unit root tests on the common component
(CRMA) and the idiosyncratic component (CIPS) suggest that both series in log-level are
integrated of degree one.
To examine the degree of international risk sharing using non-stationary variables, the
first thing is to ensure that the residuals are stationary. Table 2.2 and 2.3 display residual
diagnostic tests for all three types of regressions: the CCE estimation, the estimation with
demeaned variables (DEM), and the PPF estimation. We run each regression with data
both in levels and in first-differences. Results in Table 2.2 are obtained when the global
shock is approximated by population weighted cross-sectional averages, whereas those in
Table 2.3 are based on cross-sectional averages with equal weights. Additionally, the lower
panels of both tables illustrate residual tests from regressions with a truncated sample
excluding the recent Great Recession. The extra column in Table 2.3 is the result for the
ECM model, described in equation (2.29).
The rejection of the CD test for residuals in all DEM and PPF regressions indicates
that these methods are not able to fully control cross-sectional dependence in the regres-
sion. This problem also remains for the case of CCE regression when the common trend
is approximated by population weighted cross-sectional averages. The latter result may be
explained by the fact that in such a large sample, some countries with a large population are
not typical in terms of common risks. Hence, population weights are inappropriate when
obtaining the approximation of common factors for real GDP per capita and consumption
per capita. From the residual diagnostic test, only the CCE estimation with simple averages
is statistically defensible.
As discussed, in contrast to the CCE method, the widely-used DEM method imposes
homogeneity assumption on individual’s link to aggregate shock and obtain the idiosyncratic
consumption (income) by simple cross-sectional demeaning. It has been shown in Table 2.2
and 2.3 that the simple demeaning leaves a unit root in the regression residuals. Here, I
include some visual comparisons between the idiosyncratic fluctuation obtained from the
CCE method and from the DEM method, to provide some intuition for the problem of
cross-sectional demeaning. Figure 2.2 illustrates the distribution of self-loading coefficients
γci,c and γyi,y in the first stage of CCE regressions (equation (2.23) and (2.24)). It reveals
that these loading coefficients vary sharply across countries and they are not all equal to
one, the value imposed by the DEM-type regression. Similarly, Figure 2.3 illustrates the
distribution of the cross-loading coefficients γyi,c and γci,y. They are not all equal to zero,
the value imposed by both DEM-type and PPF regressions. These loading coefficients
estimated from the data indicate that the values imposed in DEM-type fluctuation may
cause misspecification of the idiosyncratic fluctuations in consumption and income. Thus,
the DEM approach is incapable of filtering out the common factors from the panels.
In addition, Figure 2.4 visualizes the correlation between the idiosyncratic components
40
Table 2.1: Tests for Individual Variables
Levels DifferenceslogC log Y ∆ logC ∆ log Y
CD 253.45∗ 239.37∗ 22.93∗ 43.33∗
No. of CF 1 1 1 1CRMA 2.11 2.08 -3.92∗ -3.21∗
CIPS(4) -0.32 1.87 -9.72∗ -7.78∗
Note: Pesaran’s (2004) cross-sectional independence test (CD) follows a standard normal distribu-
tion. No. of CF denotes the estimated number of common factors in the panel following Bai and
Ng (2002). The 5% critical value for Pesaran’s (2007) panel unit root test (CIPS) is -2.06. The lag
length for the CIPS test is set to T 1/3 = 4. The 5% critical value for Sul’s (2009) unit root test for
the cross-sectional means (CRMA) is -1.88. The lag length for the CRMA test is determined by
the Bayesian Information Criterion with maximum 4 lags. Statistical significance at the 5% level or
lower is denoted by ∗.
Table 2.2: Residual Diagnostic Tests, Weighted Averages
Full Sample: 1970-2010
DEM PPF CCE DEM∆ PPF∆ CCE∆
CD 10.11∗ 6.76∗ 5.06∗ 17.24∗ 9.12∗ 3.80∗
CRMA -1.62 -1.57 -1.45 -4.69∗ -5.56∗ -4.97∗
CIPS(4) -3.31∗ -9.67∗ -7.39∗ -9.35∗ -8.83∗ -10.33∗
Truncated Sample: 1970-2007
DEM PPF CCE DEM∆ PPF∆ CCE∆
CD 11.00∗ 7.58∗ 5.55∗ 14.60∗ 6.33∗ 4.14∗
CRMA -1.78 -1.45 -1.15 -4.20∗ -4.55∗ -3.64∗
CIPS(4) -4.25∗ -8.83∗ -6.35∗ -7.31∗ -7.62∗ -8.79∗
Note: CCE and CCE∆ are the common correlated effects estimates for the data in log-levels and
in log-differences, respectively. DEM and DEM∆ are estimates from a cross-sectionally demeaned
regression with the data in log-levels and in log-differences, respectively. PPF and PPF∆ are
estimates from a partially pre-filtered regression with the data in log-levels and in log-differences,
respectively. ECM is the error correction model under CCE, equation (2.29). See note in Table 2.1.
41
CCE loading, individual cons on cross−sectional mean of cons
loading factor
Fre
quen
cy
−10 −5 0 5 10 15
05
1015
20
CCE loading,df, individual cons on cross−sectional mean of cons
loading factor
Fre
quen
cy
−5 0 5 10 15
05
1015
2025
3035
CCE loading, individual rgdp on cross−sectional mean of rgdp
loading factor
Fre
quen
cy
−10 −5 0 5 10 15
05
1015
20
CCE loading,df, individual rgdp on cross−sectional mean of rgdp
loading factor
Fre
quen
cy
−2 0 2 4
05
1015
γci,c γci,c
γyi,y γyi,y
Data in log-levels Data in log-differences
Figure 2.2: Distribution of the γci,c and γyi,y loading coefficient estimates in the first-stageequations of CCE, see equations (2.23) and (2.24).
42
CCE loading, individual rcons on cross−sectional mean of gdp
loading factor
Fre
quen
cy
−10 −5 0 5 10 15
05
1015
20
CCE loading,df, individual cons on cross−sectional mean of gdp
loading factor
Fre
quen
cy
−4 −2 0 2 4 6 8
05
1015
2025
30
CCE loading, individual rgdp on cross−sectional mean of cons
loading factor
Fre
quen
cy
−10 −5 0 5 10 15
05
1015
20
CCE loading,df, individual gdp on cross−sectional mean of cons
loading factor
Fre
quen
cy
−4 −2 0 2 4 6
05
1015
γyi,c γyi,c
γci,y γci,y
Data in log-levels Data in log-differences
Figure 2.3: Distribution of the γyi,c and γci,y loading coefficient estimates in the first-stageequations of CCE, see equations (2.23) and (2.24).
43
Table 2.3: Residual Diagnostic Tests, Simple Average
Full Sample (1970-2010)
DEM PPF CCE DEM∆ PPF∆ CCE∆ ECMCD 11.28∗ 15.27∗ 0.06 22.86∗ 10.17∗ -0.53 -0.83CRMA -1.74 -0.85 — -6.02∗ -5.45∗ — —CIPS -2.03 -5.13∗ -5.87∗ -8.66∗ -9.32∗ -10.09∗ -10.03∗
Truncated Sample (1970-2007)
DEM PPF CCE DEM∆ PPF∆ CCE∆ ECMCD 12.68∗ 16.41∗ 0.36 23.44∗ 8.74∗ -0.33 -0.48CRMA -1.86 -0.80 — -5.56∗ -5.19∗ — —CIPS -0.13 -3.96∗ -5.75∗ -5.34∗ -6.59∗ -8.36∗ -11.02∗
Note: See note in Table 2.1 and 2.2
based on the CCE first-stage (equation 2.23 and 2.24) and the cross-sectionally demeaned
variables. Again, with full risk sharing and when individual countries are homogeneous
in terms of the risk-aversion, time preference and endowments, a unit loading coefficient
is expected and the idiosyncratic components from the demeaned data will be close to
those from the CCE first-stage. Nevertheless, in the data, the idiosyncratic components
from equation (2.23) and (2.24) are different from the demeaned. According to Figure 2.4,
the correlation between the two types of idiosyncratic components is below 0.80 for most
countries.
Moreover, the misspecified idiosyncratic components from cross-sectional demeaning are
left with trending common factors and will lead to unexpected outcome in the estimation
of international risk sharing. Figure 2.5 compares the idiosyncratic measures from the first-
stage of the CCE method with the demeaned variables for all countries. This figure shows
that both demeaned consumption and income series behave quite differently from the first-
stage residuals of the CCE. In particular, for many countries, the idiosyncratic measures
based on cross-sectionally demeaned variables are trending, which may be either introduced
or not fully removed by the imposed and identical value of the effects of aggregate shocks.
When regressing the trending idiosyncratic consumption on income as in Figure 2.6 for two
representative countries, the trend from both dependent and independent variables is likely
to cause bias in the estimated β for two reasons. First, the trend effect in both variables may
dominate other types of variation and lead to a misleadingly estimated βi for the degree of
international risk sharing. Second, as reported in Table 2.3, there is no cointegration found
between these trending variables and the estimate for β in the long run is spurious.
44
Log Level Consumption
Correlation Coefficient
Fre
quen
cy
0.2 0.4 0.6 0.8 1.0
05
1525
Log Level Income
Correlation Coefficient
Fre
quen
cy
0.2 0.4 0.6 0.8 1.0
05
1020
CORR(ξcit, cit − ct)
CORR(ξyit, yit − yt)
Figure 2.4: Distribution of correlation coefficients Corr(ξcit, cit − ct) and Corr(ξyit, yit − yt).The idiosyncratic components, ξcit and ξyit, are estimated in (2.23) and (2.24), and the cross-sectionally demeaned variables, cit − ct and yit − yt, appear directly in (2.2). All analyzedseries are in log-levels.
45
1970 1980 1990 2000 2010
−3
−2
−1
01
23
Idiosyncratic Consumption
time
1970 1980 1990 2000 2010
−3
−2
−1
01
23
Demeaned Consumption
time
1970 1980 1990 2000 2010
−3
−2
−1
01
23
Idiosyncratic Income
time
1970 1980 1990 2000 2010
−3
−2
−1
01
23
Demeaned Income
time
ξcit in Log-Levels cit − ct in Log-Levels
ξyit in Log-Levels yit − yt in Log-Levels
Figure 2.5: Scaled estimates of idiosyncratic components (left), ξcit and ξyit and the cross-sectionally demeaned variables (right), cit − ct and yit − yt.Estimated in (2.23) and (2.24), and (2.10) respectively. All analyzed series are in log-levels.The estimates are highlighted for two representative countries: Singapore (dash-dotted line)and Cameroon (long-dashed line).
46
1970 1980 1990 2000 2010
−2
−1
01
2
Idiosyncratic Consumption,reprsentative country
time
1970 1980 1990 2000 2010
−2
−1
01
2
Demeaned Consumption,representative country
time
1970 1980 1990 2000 2010
−2
−1
01
2
Idiosyncratic Income,representative country
time
1970 1980 1990 2000 2010
−2
−1
01
2Demeaned Income,representative country
time
ξcit in Log-Levels cit − ct in Log-Levels
ξyit in Log-Levels yit − yt in Log-Levels
Figure 2.6: Estimates of idiosyncratic components, ξcit and ξyit, and the cross-sectionallydemeaned variables, cit − ct and yit − yt for representative countries.Estimated in (2.23), (2.24) and (2.2) respectively. All analyzed series are in log-levels. Theestimates are highlighted for two representative countries: China (solid line) and CentralAfrica Republic (dash-dotted line).
47
2.5.2 Estimation for the Overall β
In this section, I will focus on the discussion of the estimated β from the statistically valid
CCE method. In Table 2.4, I include the CCEMG estimation for the overall β both in the
long-run and in the short-run, as well as the average speed of adjustment from the ECM.
To compare and contrast with results following methods in the literature, I also estimate
two alternative specifications based on DEM-type idiosyncratic fluctuation. Particularly,
the DEMFE refers to a fixed effect estimator for the overall β when the idiosyncratic
consumption and income are obtained by cross-sectionally demeaning. In contrast to the
pooling feature in the fixed effect estimator, the DEMMG refers to a mean-group estimator
using DEM-type fluctuations.
The decision to estimate the overall β from the mean-group estimator instead of the
pooled estimator is based on Coakley et al. (2001), which examines the small sample prop-
erties of different panel estimators: the fixed effect estimator and the mean-group estimator.
By using Monte Carlo simulation, the authors show that when there is a large heterogene-
ity in the slope coefficient of a panel, the mean group estimator will be closer to the true
value of β. In Table 2.4, the sharp divergence between DEMFE and DEMMG for most
sub-groups of countries suggests a negative bias in DEMFE. As explained in the appendix,
this also implies a negative relationship between βi and the variance of demeaned income,
i.e., the country in a particular group with relatively large variation in its income will likely
have a small value of βi. This is consistent with bounded rationality discussed in Browning
and Collado (2001): countries with large variation in income tend to have a large degree of
consumption smoothing.
The valid and consistent CCE estimations provide many valuable results. First, in line
with earlier studies (Artis and Hoffmann, 2008; Becker and Hoffmann, 2006; Leibrecht and
Scharler, 2008), no evidence for perfect international risk-sharing is found, and on average,
consumption tends to be affected by idiosyncratic risks significantly in both short-run and
long-run. Second, the extent of risk-sharing tends to be higher in the short run than in
the long-run. In the long-run, the fraction of uninsured variation in GDP transmitted
to consumption is above 0.80, while it is about 0.70 in the short run. Third, partially
consistent with the findings in Kose et al. (2008), the degree of consumption smoothing is
inversely related to the level of development, both in the long-run and in the short-run.6
This finding is predictable economically since most developed economies have better access
to most risk-sharing channels: the credit and capital market, and international trade.
Finally, the estimates for OECD countries are quite similar to those obtained by Leibrecht
and Scharler (2008): our βLR = 0.68 and βSR = 0.80 fall slightly below their estimates of
about 0.7 and 0.9, respectively. However, the estimated speed of equilibrium-error correc-
6This argument is more obvious by comparing OECD v.s. Non-OECD and Developed v.s. Non-developed.
48
tion, κ = −0.31, deviates from their -0.1 estimate by a larger margin.7 Consequently, the
mean adjustment lag (computed as µ = (1− βSR)/(−κ) based on Hendry (1995) indicates
that in OECD countries an income shock exerts its full effect on consumption within about
a year according to our study, and in about three years according to the results of Leibrecht
and Scharler (2008).
2.5.3 The Change of the Overall β Over Time
Lane and Milesi-Ferretti (2007) has documented the intensity in the degree of financial
globalization for both industrial and developing countries since 1990s. For emerging mar-
kets, the improvement in financial globalization is characterized by the increasing equity
component in the external liabilities and official reserve assets. For major debtors, the debt
liabilities is their major source of the external finance.
By analyzing international risk sharing at low frequency, Artis and Hoffmann (2012)
examine the changes of β in periods of pre-1990 and post-19908 due to increasing financial
liberalization. Based on a fixed effect estimator of DEM-type fluctuation, the authors found
strong evidence of an improvement of international risk sharing in a group of OECD coun-
tries. As a preliminary analysis, I constructed the same dataset as in Artis and Hoffmann
(2012), and repeat their empirical strategy. Considering the problem of DEMFE estimation
discussed before, I conduct the CD test and second generation panel unit root tests for the
residuals. As Table 2.5 shows, the residuals are found to be cross-sectionally dependent and
non-stationary. Moreover, I reestimate their empirical model by relaxing the assumption of
a homogeneous β and by following the CCE method. In contrast to their findings, I could
not find any evidence of improved international risk sharing.
To acquire some insight into the change in the degree of international risk sharing with
the large panel, I follow Artis and Hoffmann (2012) and estimate the empirical model for
two sub-periods: 1970-1989, and 1990-2010. Results are reported in Table 2.6. Since the
1990s, the global financial market has become more integrated due to increasing capital flows
across countries. In principal, consumers can insure themselves against income shocks more
easily in a more integrated financial market. Hence, it is widely accepted that the degree
of international risk sharing should be higher in the period of 1990-2010. Nevertheless,
according to Table 2.6, results across all sub-groups do not suggest any evidence for a higher
degree of international risk sharing. The estimated β coefficient is not significantly lower
in the post-globalization period and this conclusion is robust to different sample groups of
countries. In fact, there are signs of a deterioration in international risk sharing in the EU
subsample.
The lack of the improved consumption risk sharing has been found in other studies.
7Notice, however, that the time period and the country analyzed by Leibrecht and Scharler (2008) aredifferent from ours.
8The dataset used in this paper is obtained form PWT 6.2, which ends in 2004. The two sub-periodsare 1960-1990 and 1990-2004.
49
Table 2.4: Mean Group Coefficient Estimates for Sub-Samples
Country Group βLRDEMFE βSR
DEMFE βLRDEMMG βSR
DEMMG βLRCCEMG βSR
CCEMG κ µWhole Sample1 0.78∗ 0.68∗ 0.85∗ 0.73∗ 0.83∗ 0.71∗ -0.39 0.74
High Income1,2,3,4 0.66∗ 0.74∗ 0.91∗ 0.72∗ 0.82∗ 0.65∗ -0.33 1.06UpperMid Income2 0.87∗ 0.78∗ 0.92∗ 0.74∗ 0.85∗ 0.73∗ -0.41 0.66LowerMid Income2,3 0.76∗ 0.54∗ 0.77∗ 0.67∗ 0.83∗ 0.70∗ -0.42 0.71Low Income 0.81∗ 0.71∗ 0.80∗ 0.77∗ 0.82∗ 0.77∗ -0.44 0.52
OECD3,4 0.81∗ 0.78∗ 0.87∗ 0.76∗ 0.80∗ 0.68∗ -0.31 1.03Non-OECD1 0.77∗ 0.68∗ 0.85∗ 0.72∗ 0.84∗ 0.72∗ -0.42 0.67
Developed1,2,3,4 0.85∗ 0.74∗ 0.87∗ 0.73∗ 0.78∗ 0.65∗ -0.32 1.09Developing1,3 0.78∗ 0.66∗ 0.88∗ 0.73∗ 0.87∗ 0.74∗ -0.41 0.63
EU4 0.87∗ 0.74∗ 0.88∗ 0.76∗ 0.83∗ 0.69∗ -0.34 0.92
Note: Country group definitions follow those used by the World Bank and OECD. βLRDEMFE
and βSRDEMFE are obtained by estimating equation (2.10) and (2.11), respectively. βLR
DEMMG and
βSRDEMMG are obtained by estimating equation (2.10) and (2.11) for each indiviudal, respectively,
and aggregating the βi using equation (2.21). βLRCCE is obtained by estimating equation (2.9) with
data in log-levels; βSRCCE is obtained by estimating the error-correction model in equation (2.29),
and aggregating βi using equation (2.21). κ denotes the estimated speed-of-adjustment coefficient
in the error-correction model. µ denotes the mean adjustment lag computed as µ = (1− βSR)/(−κ)based on Hendry (1995). The β coefficient measures the extent of income shocks transmitted toconsumption. The theory of international risk sharing implies H0: β = 0. Statistical significance atthe 5% level or lower is denoted by ∗.1 indicates that βLR
CCEMG and βLRDEMFE are different at the 10% level of marginal significance.
2 indicates that βLRCCEMG and βLR
DEMMG are different at the 10% level of marginal significance.
3 indicates that βSRCCEMG and βSR
DEMFE are different at the 10% level of marginal significance.
4 indicates that βSRCCEMG and βSR
DEMMG are different at the 10% level of marginal significance.
50
Pierucci and Ventura (2010) failed to conclude that the surge in international financial
liberalization improved long-run risk sharing. Kose et al. (2008) found that only industrial
countries have attained better risk sharing during the recent period of globalization, and that
developing countries have been mostly shut out of these benefits. In their paper, the reason
was imputed to the particular composition of capital flows, with external debt preventing
most emerging economies to efficiently share risks. At last, Bai and Zhang (2012) found that
consumption risk sharing shows little improvement for their whole sample, OECD countries
as well as emerging markets. Bai and Zhang (2012) suggests that it is the existence of
friction in the capital markets that limits the observed increase in capital flow under financial
liberalization. As a result, there is no significant improvement in consumption smoothing
and risk sharing.
2.5.4 Individual β
In this paper, I allow individual-specific coefficient of β in the regression model of consump-
tion smoothing considering the large and heterogenous panel used in this paper. In the
previous discussion about the estimation of the overall β, I attribute the difference between
the fixed effect estimator and the mean-group estimator partially to the heterogeneity in
terms of βi. Here, I provide additional evidence regarding the heterogeneity of βi. The
rejection of Pesaran and Yamagata’s (2008) test of parameter homogeneity in Table 2.7
implies that the heterogeneity of βi is statistically significant.
With the large degree of heterogeneity in βi across countries, it is also interesting to
examine individual βi. I report the estimations of βi in Table 2.8 and 2.9 and plot their
distribution in Figure 2.7. Notably, there is a remarkable heterogeneity in terms of countries’
risk-sharing coefficients, both in the long run and in the short run. This further suggests
that the pooled CCE estimator is less appropriate to measure the mean slope coefficient of
the regression for countries in such large sample.
Since the panel unit root test suggests stationarity in the CCE regression, I also test
for the presence of perfect risk-sharing hypothesis based on individual βi. To account for
heteroskedasticity and autocorrelation in individual regression residuals, I use the Newey
and West (1987, 1994) heteroskedasticity and autocorrelation consistent (HAC) covariance
matrix for the inference of βi.
Accordingly, more than 95% of the βLR and almost 90% of the βSR estimates are
significantly different from zero, indicating a widespread lack of consumption risk sharing.
Moreover, the degree of risk sharing tends to be lower in the long run for most countries.
In particular, 33 of the analyzed countries have 50% higher coefficient estimates in the long
run than in the short run, and over 45 of them appear to exhibit dis-smoothing behavior
(βi > 1) in the long run, whereas only 27 cases in the short run.
51
Table 2.5: The Effect of Financial Liberalization
Period: 1960-1990 Period: 1990-2004
βDEMFE 0.98 0.63CD test -3.74∗ -2.36∗
CIPS -1.65 5.09
Estimations from the Mean-group Estimators
βDEMMG 1.03 0.91βCCEMG 0.93 0.94
Note: Values obtained from βDEMFE are reported in Artis and Hoffmann (2012). All estimations are
significant at 5% level. Pesaran’s (2004) cross-sectional independence test (CD) follows a standard
normal distribution. The 5% critical value for Pesaran’s (2007) panel unit root test (CIPS) is -2.06.
The lag length for the CIPS test is set to T 1/3. Statistical significance at the 5% level or lower is
denoted by ∗.
Table 2.6: CCEMG Coefficient Estimates for Sub-Samples, in Sub-Periods
βLRCCEMG βSR
CCEMG
Country Group 1970-1989 1990-2010 1970-1989 1990-2010Whole Sample 0.82∗ 0.85∗ 0.70∗ 0.68∗
High Income 0.82∗ 0.83∗ 0.66∗ 0.62∗
UpperMid Income 0.88∗ 0.85∗ 0.75∗ 0.75∗
LowerMid Income 0.77∗ 0.85∗ 0.66∗ 0.66∗
Low Income 0.78∗ 0.87∗ 0.76∗ 0.70∗
OECD-AH 0.76∗ 0.78∗ 0.64∗ 0.56∗
OECD 0.80∗ 0.80∗ 0.68∗ 0.64∗
Non-OECD 0.82∗ 0.86∗ 0.71∗ 0.69∗
Developed 0.76∗ 0.80∗ 0.65∗ 0.63∗
Developing 0.85∗ 0.89∗ 0.72∗ 0.73∗
EU1 0.78∗ 0.86∗ 0.64∗ 0.69∗
Note: 1 indicates that the βLR estimates differ between two sub-periods at the 5% level of marginal
significance. See also the Table 2.4 notes.
52
Table 2.7: Homogeneous test for individual CCE estimates
CCELR CCEECM
∆ 69.79∗ 42.53∗
∆adj 72.43∗ 44.18∗
Note: Both ∆ and ∆adj have standard normal distribution. The null hypothesis is homogeneous
coefficient among individuals. ∗ suggests rejection of the null at the 5% level.
2.6 Conclusion
With perfect international risk sharing, the theoretical model implies that idiosyncratic con-
sumption should be independent of idiosyncratic income shocks. Accordingly, a prerequisite
of this test is the isolation of country-specific variation in the data.
In this paper, I show that the cross-sectional demeaning, a commonly used method
of the literature, is generally insufficient to eliminate the cross-sectional dependence in a
panel regression of the international risk sharing test. More importantly, I demonstrate that
the cross-sectional dependence left in the regression residuals leads to common unit roots.
Therefore, I suggest that a more statistically appropriate approach is the CCE methodology
of Pesaran (2006) and Kapetanios et al. (2011) that controls for global trends and allows
for heterogeneous effects across units. To estimate the degree of international risk sharing,
I use a relatively large panel dataset, which has never been considered in the literature.
Given such large number of different countries in the panel, I recommend the use of mean
group estimator for the overall risk sharing coefficient.
In the empirical analysis, I first conduct residual diagnostic tests for all types of regres-
sions and show that only residuals from the CCE method are cross-sectional independent
and stationary. Based on the statistically valid CCE mean group estimations, I measure the
average degree of international risk sharing for different subgroup countries both in the long
run and in the short run from an ECM. Similar to most existing literature, no evidence of
high degree of risk sharing is found. Moreover, I also investigate whether or not the degree
of international risk sharing is higher during the subperiod with more liberalization of inter-
national financial market. Consistent with a few conclusions in Kose et al. (2008) and Bai
and Zhang (2012), my empirical results cannot support the argument of an improvement in
the degree of risk sharing. Finally, I measure the risk sharing coefficient for each individual
country and find a large variation across countries.
53
Table 2.8: Country-Specific Coefficient Estimates
id country βLR βSR κ id country βLR βSR κ1 AFG 0.87∗ 0.95∗ -0.20 ∗ 41 DOM 0.90∗ 1.01∗ -0.21 ∗
2 AGO 2.00∗ 0.62 -0.58 ∗ 42 DZA 1.48∗ 0.24 -0.213 ALB 0.18∗ 0.29∗ -0.22 ∗ 43 ECU 0.71∗ 0.57∗ -0.58 ∗
4 ARG 0.97∗ 1.26∗ -0.19 ∗ 44 EGY 0.68∗ 0.37 -0.33 ∗
5 ATG 0.99∗ 1.58∗ -0.70 ∗ 45 ESP 0.89∗ 0.80∗ -0.32 ∗
6 AUS 0.42∗ 0.12 -0.10 46 ETH 1.10∗ 1.03∗ -0.43 ∗
7 AUT 1.11∗ 0.74∗ -0.35 ∗ 47 FIN 0.81∗ 0.46∗ -0.27 ∗
8 BDI 0.86∗ 0.79∗ -0.52 ∗ 48 FJI 0.64∗ 0.57∗ -0.37 ∗
9 BEL 0.98∗ 0.53∗ -0.35 ∗ 49 FRA 1.02∗ 0.73∗ -0.24 ∗
10 BEN 0.80∗ 0.66∗ -0.22 ∗ 50 FSM 1.03∗ 0.98∗ -0.81 ∗
11 BFA 1.96∗ 1.27∗ -0.52 ∗ 51 GAB 0.34∗ -0.09 -0.56 ∗
12 BGD 1.61∗ 1.10∗ -0.62 ∗ 52 GBR 1.13∗ 0.95∗ -0.55 ∗
13 BGR 1.01∗ 0.85∗ -0.39 ∗ 53 GER 1.16∗ 0.62∗ -0.28 ∗
14 BHR 0.64∗ 0.78∗ -0.42 ∗ 54 GHA 1.00∗ 0.88∗ -0.65 ∗
15 BHS 1.42∗ 1.29∗ -0.36 ∗ 55 GIN 1.56∗ 1.53∗ -0.33 ∗
16 BLZ 1.22∗ 1.11∗ -0.56 ∗ 56 GMB 0.85∗ 0.92∗ -0.35 ∗
17 BMU 1.47∗ 0.90∗ -0.38 ∗ 57 GNB 0.99∗ 0.74∗ -0.50 ∗
18 BOL 0.64∗ 0.86∗ -0.46 ∗ 58 GNQ 0.67∗ 0.72∗ -0.31 ∗
19 BRA 0.83∗ 0.90∗ -0.46 ∗ 59 GRC 0.20 0.49∗ -0.15 ∗
20 BRB 1.59∗ 1.29∗ -0.31 ∗ 60 GRD 0.46∗ 0.69∗ -0.53 ∗
21 BRN -0.90∗ 0.04 -0.57 ∗ 61 GTM 0.86∗ 0.75∗ -0.26 ∗
22 BTN 0.56∗ 0.57∗ -0.41 ∗ 62 GUY 1.15∗ 0.93∗ -0.26 ∗
23 BWA 0.40∗ 0.19∗ -0.34 ∗ 63 HKG 1.22∗ 0.85∗ -0.43 ∗
24 CAF 0.92∗ 0.83∗ -0.45 ∗ 64 HND 1.00∗ 0.19 -0.54 ∗
25 CAN 0.41∗ 0.48∗ -0.18 ∗ 65 HTI 0.97∗ 1.12∗ -0.46 ∗
26 CHE 0.23 0.23∗ -0.32 ∗ 66 HUN 1.02∗ 1.00∗ -0.20 ∗
27 CHL 0.96∗ 0.79∗ -0.22 67 IDN 1.09∗ 0.56∗ -0.27 ∗
28 CHN 1.00∗ 1.01∗ -0.42 ∗ 68 IND 0.95∗ 0.71∗ -0.58 ∗
29 CIV 0.76∗ 0.78∗ -0.48 ∗ 69 IRL 0.64∗ 0.62∗ -0.48 ∗
30 CMR 0.97∗ 0.79∗ -0.40 ∗ 70 IRN 0.59∗ 0.39∗ -0.27 ∗
31 COG 0.46∗ 0.30∗ -0.11 71 IRQ -0.08 0.27∗ -0.86 ∗
32 COL 0.92∗ 0.75∗ -0.40 ∗ 72 ISL 1.25∗ 1.01∗ -0.60 ∗
33 COM 0.40∗ 0.27 -0.38 ∗ 73 ISR 1.08∗ 0.83∗ -0.39 ∗
34 CPV 0.99∗ 0.67∗ -0.13 74 ITA 1.00∗ 0.76∗ -0.53 ∗
35 CRI 0.93∗ 1.13∗ -0.18 75 JAM 0.88∗ 0.80∗ -0.43 ∗
36 CUB 1.16∗ 1.18∗ -0.34 ∗ 76 JOR 1.34∗ 0.68∗ -0.30 ∗
37 CYP 0.87∗ 0.72∗ -0.48 ∗ 77 JPN 0.86∗ 0.63∗ -0.27 ∗
38 DJI 1.25∗ 0.96∗ -0.72 ∗ 78 KEN 1.08∗ 1.41∗ -0.43 ∗
39 DMA 0.62∗ 0.47∗ -0.70 ∗ 79 KHM 0.94∗ 0.97∗ -0.57 ∗
40 DNK 0.53∗ 0.67∗ -0.32 ∗ 80 KIR 0.67∗ 0.47∗ -0.30 ∗
Note: See Table 2.9 below.
54
Table 2.9: Comparison of country-specific coefficient estimates (continued)
id country βLR βSR κ id country βLR βSR κ81 KNA 0.80∗ -0.06 -0.51 ∗ 121 PRT 0.81∗ 0.50∗ -0.49 ∗
82 KOR 0.84∗ 0.78∗ -0.19 ∗ 122 PRY 0.77∗ 0.23 -0.30 ∗
83 LAO 0.85∗ 0.98∗ -0.39 ∗ 123 ROM 0.50∗ 0.61∗ -0.39 ∗
84 LBN 0.84∗ 0.84∗ -0.81 ∗ 124 RWA 0.59∗ 0.20∗ -0.11 ∗
85 LBR 1.15∗ 0.90∗ -0.28 ∗ 125 SDN 1.01∗ 1.43∗ -0.34 ∗
86 LCA 0.70∗ 0.89∗ -0.52 ∗ 126 SEN 0.88∗ 0.50∗ -0.21 ∗
87 LKA 1.10∗ 0.49∗ -0.62 ∗ 127 SGP 0.73∗ 0.46∗ -0.1388 LSO 1.05∗ 0.66∗ -0.35 ∗ 128 SLB 1.04 0.44∗ -0.17 ∗
89 LUX 0.69∗ 0.31∗ -0.26 ∗ 129 SLE 0.67∗ 0.79∗ -0.23 ∗
90 MAC 1.00∗ 0.25∗ -0.11 ∗ 130 SLV 1.32∗ 1.16∗ -0.55 ∗
91 MAR 0.34 0.44∗ -0.50 ∗ 131 SOM 0.95∗ 1.01∗ -0.48 ∗
92 MDG 1.17∗ 0.12 -0.10 132 STP 1.19∗ 1.31∗ -0.63 ∗
93 MDV 0.96∗ 0.66∗ -0.24 ∗ 133 SUR 1.30∗ 1.83 ∗ -0.68 ∗
94 MEX 0.80∗ 0.83∗ -0.19 ∗ 134 SWE 0.67∗ 0.58 ∗ -0.32 ∗
95 MHL 1.34∗ 0.56∗ -0.40 ∗ 135 SWZ 0.40∗ -0.08 -0.66 ∗
96 MLI -0.17 0.00 -0.44 ∗ 136 SYC 1.08∗ 0.89∗ -0.2397 MLT 0.75∗ 0.67∗ -0.13 137 SYR 0.78∗ 0.96∗ -0.20 ∗
98 MNG 0.67∗ 1.09∗ -0.52 ∗ 138 TCD 0.47∗ 0.51∗ -0.1999 MOZ 0.89∗ 0.70∗ -0.44 139 TGO 0.33 ∗ 0.54∗ -0.83 ∗
100 MRT 0.80∗ 0.83∗ -0.49 ∗ 140 THA 0.78∗ 0.61∗ -0.70 ∗
101 MUS 0.84∗ 0.60∗ -0.47 ∗ 141 TON 0.92∗ 0.65∗ -0.56 ∗
102 MWI 0.44∗ 0.54∗ -0.86 ∗ 142 TTO 0.93∗ 0.81∗ -0.77 ∗
103 MYS 0.54∗ 0.96∗ -0.26 ∗ 143 TUN 0.64∗ 0.28∗ -0.53 ∗
104 NAM 1.03∗ 0.88∗ -0.85 ∗ 144 TUR 0.72∗ 0.94∗ -0.38 ∗
105 NER 0.32∗ 0.67∗ -0.62 ∗ 145 TWN 1.12∗ 0.63∗ -0.37 ∗
106 NGA 1.13∗ 1.28∗ -0.45 ∗ 146 TZA -0.04 0.37∗ -0.33 ∗
107 NIC 0.72∗ 0.51∗ -0.36 ∗ 147 UGA 0.97∗ 0.94∗ -0.66 ∗
108 NLD 0.72∗ 0.71∗ -0.18 ∗ 148 URY 1.00∗ 0.95∗ -0.05109 NOR 0.38∗ 0.57∗ -0.26 ∗ 149 USA 0.91∗ 0.74∗ -0.21110 NPL 1.10∗ 1.13∗ -0.48 ∗ 150 VCT 0.74∗ 0.98∗ -0.36 ∗
111 NZL 0.95∗ 0.79∗ -0.36 ∗ 151 VEN 1.02∗ 0.77∗ -0.35112 OMN 1.26∗ 0.53 -0.57 ∗ 152 VNM 0.55∗ 0.77∗ -0.14113 PAK 0.59∗ 0.90∗ -0.49 ∗ 153 VUT 0.78∗ 0.76∗ -0.32 ∗
114 PAN 0.15 0.14 -0.65 ∗ 154 WSM 0.95∗ 0.92∗ -0.67 ∗
115 PER 0.95∗ 0.93∗ -0.36 ∗ 155 ZAF 0.67∗ 0.62∗ -0.38 ∗
116 PHL 0.29∗ 0.30∗ -0.21 ∗ 156 ZAR 0.63∗ 0.21 -0.55 ∗
117 PLW 0.45 -0.89∗ -0.12 157 ZMB 0.88∗ 1.34∗ -0.33 ∗
118 PNG 0.96 0.57 -0.08 158 ZWE 0.24 0.67∗ -0.52 ∗
119 POL 0.93∗ 1.08∗ -0.40 ∗
120 PRI 0.55∗ 0.34∗ 0.01
Note: βLR is obtained by estimating equation (2.28) with data in log-levels. βSR is obtained
by estimating the error-correction model in equation (2.29). κ denotes the estimated speed-of-
adjustment coefficient in the error-correction model. Inference is based on heteroskedasticity and
autocorrelation consistent robust standard errors. Statistical significance at the 5% level or lower is
denoted by ∗.
55
Histogram for beta, LR
Beta
Fre
quen
cy
−1.0 −0.5 0.0 0.5 1.0 1.5 2.0
010
2030
40
Histogram for beta, SR
Beta
Fre
quen
cy
−1.0 −0.5 0.0 0.5 1.0 1.5 2.0
010
2030
40
βLR βSR
Figure 2.7: Distribution of country specific coefficient estimates for long-run and short-run.βLR is obtained by estimating equation (2.28) with data in log-levels. βSR is obtained byestimating the error-correction model in equation (2.29).
56
Chapter 3
Essay 3: How Integrated are US Gasoline Markets: An
Empirical Test with Cross-sectional Correlation and
Structural Breaks
3.1 Introduction
Studying commodity prices across regional markets helps determine the extent of spatial
arbitrage and the resulting regional market integration. The degree of integration cross
regional markets produces insight to the effectiveness of policies which target all markets.
For example, when regional markets are well integrated, it implies that a national policy,
such as the Renewable Fuel Standard (RFS) program or the Keystone Pipeline project
will have a similar impact on local prices. Meanwhile, it is also an indication for market
liberalization policy and the market intervention from the government. When regional
markets are spatially integrated with each other, products are efficiently allocated among
markets, suppressing the need for additional government intervention.
Empirical researchers use the law of one price (LOP) to examine the degree of market
integration. The absolute LOP states that a good will sell for the same price across ge-
ographical regions when regional markets are perfectly integrated.1 When the LOP fails,
price changes in one market will not be transmitted on a one-for-one basis to prices in other
geographic markets either instantaneously, or over a number of periods. Thus, regional
prices will tend to diverge from each other in both the short and long run.
Since its deregulation in 1981, the U.S. retail gasoline market has been described as close
to perfectly competitive, suggesting that the LOP may hold. However, many other factors
may prevent the LOP from holding among regional markets. The price of retail gasoline in
a given region is set by the price of crude oil, refinery costs, the cost of transportation and
distribution as well as the cost of selling. Thus, the different structure of each intermediate
industry may affect the spatial relationship between retail gasoline prices over time. For
instance, refineries in different regions use differentiated sources of crude oil. As a result,
changes in the source of crude oil for US refineries (i.e. the decline in the imports from
African) will lead to mixed effects on local regular gasoline prices. Paul et al. (2001) and
1The relative LOP allows for the difference between regional prices in terms of transportation cost andother time invariant factors.
57
Holmes et al. (2013) argued that local gasoline retailers hold some degree of market power
due to factors ranging from the refinery’s purchasing price of crude oil to the demographic
characteristics (Iyer and Seetharaman, 2008; Temple, 1988). Because of these unknown
factors, the degree of spatial market integration for the retail gasoline in the US is unknown
and requires empirical examinations.
The empirical test of the LOP in the market for retail gasoline can be conducted at
different levels. For example, Paul et al. (2001) used data at the Petroleum Administrative
Defense District (PADD) level while Holmes et al. (2013), at the state-level. When state-
level data is used, the price of retail gasoline is influenced by many state-specific factors,
such as regulations on the gasoline content. When state markets are integrated, local
supply shocks may cause not only a price spike in a single state but in all other states as
well. Similarly, the change in federal regulations such as a nation-wide cap on Reid Vapor
Pressure (RVP) of gasoline,2 might influence the price level of gasoline in all states.
To examine the LOP for the gasoline market, some studies use a regression model as,
logPi,t = αi+βi logP ∗t + εi,t. In these studies, testing the LOP is based on hypothesis tests
on the regression coefficients, which examines the presence of an one-to-one relationship
between the regional price and the reference price. Since the price series may contain a
unit root, the test on the coefficient is valid only when the unit root test indicates that the
regression residuals donot have unit roots. To increase the power of the unit root test, panel
methods are used in more recent studies (Dreher and Krieger, 2005; Suvankulov et al., 2012).
However, these studies largely ignore two issues. First, given the fact that regional gasoline
prices are affected by common shocks such as a nation-wide policy change, world-wide
demand decline due to the global recession or a world-wide technological improvement,
the resulting residuals from the model of the LOP test are likely to be cross-sectionally
dependent. Second, the literature also overlooks the presence of structural breaks in the
regression residuals. One potential source of structural breaks in the regression residuals is
the misspecification of the regression model. The regression model used to test for the LOP
in the literature may be misspecified as it ignores changes in the degree of market integration
over time. Structural breaks in the regression residuals will affect the performance of unit
root test (Leybourne et al., 1998; Leybourne and Newbold, 2000; Perron, 1989). Thus,
it will also affect the consequent test on the coefficients. With structural breaks in the
regression residuals, tests of LOP may be misleading.
There are reasons to think that structural breaks in the degree of market integration
may exist. Early papers such as Borenstein et al. (1997), Lewis (2009) and Verlinda (2008)
among others have identified many regional factors that will impact gasoline price dynamics
over time. For instance, it is widely accepted that the rate of change of gasoline prices is
greater when prices are rising than when they are falling (Borenstein et al., 1997; Lewis,
2The policy was implemented in late 1989 and summer 1992, which led to changes in the amount of lightcomponents in gasoline.
58
2011). Thus, it is possible that the degree to which regional gasoline prices are cointegrated
differs in sub-periods. In addition, Lewis (2009) also found that there were long-lasting
geographical differences in retail prices after the supply shock of the Hurricane Rita. In
summary, structural breaks imply that the regression model with constant coefficients —
used to test for the LOP, is misspecified.
In this paper, I examine the degree of market integration across states for retail gasoline
in the US, taking into account two widely neglected issues: cross-sectional dependence and
structural breaks. I demonstrate that the test of the LOP is indeed affected by the issue
of cross-sectional dependence. Therefore, I propose a different panel unit root test to deal
with the resulting size distortion and reduced power discussed in Gengenbach et al. (2010).
In particular, I use a panel unit root test procedure that accounts for structural breaks in a
dependent panel. I first examine whether state-level retail gasoline prices are cointegrated
with the reference price (the national average price level). Then, I test for the LOP by
examining the slope coefficient when the null hypothesis of no cointegration in the first step
is rejected.
This paper is organized as follows: Section 2 discusses the literature that studies the
degree of gasoline market integration where they neglect both common cross-sectional de-
pendence and structural breaks; Section 3 discusses the empirical strategy for studying
the degree of gasoline market integration, with a focus on cross-sectional dependence and
structural breaks; Section 4 discusses the data and results.
3.2 Examination of the Degree of Gasoline Market Integra-
tion
As early as 1980s, there were studies that investigated the spatial relationship between
regional gasoline prices to examine the degree of market integration for gasoline in the US.
Stigler and Sherwin (1985) examined the co-movement of wholesale gasoline prices from
three cities - New Orleans, Chicago, and Detroit - during the period from October 1979
to October 1983. In this paper, they first calculated the correlation between pairs of price
series to quantify the similarity between city-level prices. Because common influences such
as the overall price level or the inflation of the crude oil price, may induce misleadingly
high correlation between individual prices, the authors therefore regressed the gasoline
price in each city on the crude oil price or on a national price level to get rid of these
common effects. Following this, the authors calculated the correlation between residuals
from these regressions for each city. Different from the analysis based on raw prices, these
regressions remove the common effects in the price of city gasoline, leaving behind the
desired price to study the provision of wholesaling services in local market. The authors
claimed that these regression residuals measure the price of wholesaling and are likely to
be equal across relevant geographic markets. As expected, their results show a fairly high
59
correlation between city’s residual series after removing the influence of common influences.
The simple comparison between local prices using correlation may be impaired by unit
roots in the price series. As an alternative to examine the relationship between local prices,
empirical analysis can also test whether these prices are diverging over-time. The co-
movement of local prices over time implies the presence of a long-run equilibrium relation-
ship between prices. From a time series perspective, such relationship between series implies
a cointegrating relationship. Therefore, cointegration techniques can be applied to study
the spatial relationship in gasoline prices. A cointegrating relationship allows testing of the
implication of perfect market integration, the presence of the LOP and the long-run price
parity.
Using two cointegration methods, Paul et al. (2001) studied the long-run price parity
among US retail gasoline markets using data at the Petroleum Administrative Defense
District (PADD) level. To detect the presence of the LOP as a consequence of a high degree
of price competition, they evaluated the price parity during the period between January
1983 and December 1998. They found a cointegrating relationship among PADD level price
series both from the Engle and Granger (1987) two-step method and the Johansen (1988)
rank approach, suggesting a co-movement between PADD-level gasoline prices in the long
run. However as the slope coefficient in the regression of pairs of prices was not equal to
unity, they failed to find the existence of long-run price parity. In other words, according
to their results, perfect market integration does not hold.
The long run price parity implies that relative prices or price differentials are mean-
reverting, or stationary. Therefore, unit root tests on relative prices are used as another
method to examine the LOP. Holmes et al. (2013) focused on the retail gasoline markets
in the US at the state-level. To test for price parity, they applied the Augmented Dickey-
Fuller test (ADF) to test for a unit root in the series of price differentials and calculated
the probability of rejecting of the ADF test in all pairs of state prices. Based on a rejection
rate for ADF tests of more than 95%, they concluded there was a strong support for the
law of one price between state-level markets on the mainland.
To increase the power of unit root tests, methods of panel unit root test were applied
in more recent literature to examine gasoline market integration in other areas. Dreher
and Krieger (2005) suspected the presence of a gasoline price convergence due to the ease
of cross-border purchases of both gasoline and oil products in Europe. They examined
the consequent degree of gasoline market integration using weekly prices of 15 European
countries during January 1994 to May 2004. Using panel unit root test in Maddala and Wu
(1999) and Levin et al. (2002) which include a constant term, Dreher and Krieger (2005)
rejected the null hypothesis of unit roots in a panel of relative price series, suggesting the
presence of a weak form of price parity, i.e., allowing for transportation cost differences.
Moreover, they also tested for the strong version of the LOP by implementing panel unit
root tests without a constant. In this latter case, the null hypothesis of unit roots cannot
60
be rejected.
The drawback of panel unit root tests used in Dreher and Krieger (2005) is discussed
and illustrated in Breitung and Pesaran (2008) and Gengenbach et al. (2010). Breitung and
Pesaran (2008) classified the literature of panel unit root test into two generations. The
first generation tests (for example, Maddala and Wu (1999) and Levin et al. (2002)) assume
that individual series are cross-sectionally independent, while procedures in the second
generation consider individual series to be cross-sectionally correlated with each other. As
a result, the panel data is modeled as a combination of the common component and the
idiosyncratic component. Using Monte Carlo simulations, Gengenbach et al. (2006, 2010)
showed that the dependence across units could cause size distortion and reduced power
in panel unit root tests when overlooked. Hence, if cross-sectional dependence exists in a
panel, the second-generation panel unit test is more desirable.
Accounting for cross-sectional dependence, Suvankulov et al. (2012) studied the LOP
in Canadian retail gasoline markets at the city level from January 2000 to October 2010.
In light of the threshold effect caused by transportation cost on the price differences in
the equilibrium, they tested for unit roots in relative prices using a non-linear panel unit
root test. Since the authors were able to reject the null of unit roots for most cities, they
suggested that the market is well integrated.
Another issue that may affect the test of the LOP and may interact with cross-sectional
dependence is the presence of structural breaks in the series being tested. Similar to the
problem caused by cross-sectional dependence, ignoring structural breaks in a time series
will also have size distortion effects on the unit root test.
To be more specific about the form of structural breaks considered in this paper, here I
assume the data-generating process of an interested series yt as,
yt = dt + ut, (3.1)
ut = αut−1 + vt, (3.2)
with −1 < α ≤ 1, and dt = µ + βt is the trend function of yt. In the trend function,
the intercept µ measures the level of the series, and the slope coefficient β represents the
average growth rate of the series.3 In a time series, the most obvious type of structural
breaks is the structural break in levels (µ) and / or in the growth rate (β). In this paper, I
consider the situation when structural breaks are both in levels and in the growth rate.
Perron (1989) found that the standard Dickey-Fuller (DF) type unit root test cannot
reject the null hypothesis of a unit root if the series tested is a stationary noise component
with a structural break in the slope of the trend function. This occurs because a structural
break in the trend function produces serial correlation properties that are similar to those
of a random walk. On the other hand, Leybourne et al. (1998) and Leybourne and Newbold
3Equivalently, the DGP illustrated in equation (1) and (2) can be rewritten as, yt = αyt−1 + (1−α)µ+αβ + (1 − α)βt+ vt for −1 < α < 1 and yt = yt−1 + β + vt for α = 1.
61
(2000) found that a structural break occurring early (in other words, with small value of
break fraction, τ , where breaking date is defined as (τ ∗ T ) ) in the series will cause an
over-rejection of null hypothesis in standard Dickey-Fuller unit root tests. Such “Converse
Perron Phenomena” happens because the limiting distribution of negative Dickey-Fuller
statistics varies with τ and is not symmetric around the point when τ = 1/2. As a result,
for small τ , the test statistics will take large negative values, leading to frequent rejections
of the null hypothesis. In general, if not taken into account, structural breaks in terms of
parameters in the trend function of a series will lead to misleading conclusions in the unit
root test.
In summary, the test of the LOP in the gasoline market is subject to problems caused
by both cross-sectional dependence and structural breaks. The literature has generally not
devoted much attention to these issues. Regional gasoline prices are likely to be cross-
sectionally correlated due to common shocks such as global economic recessions, or national
technology improvements in the gasoline industry. These unobserved common effects in
regional gasoline prices may result in cross-sectional dependence in the residuals of the
empirical model. Moreover, shocks to individual prices may change the relationship between
prices, affecting the degree of market integration. Failure to consider these changes in the
regression model may result in structural breaks in the regression residuals. To control for
the adverse effect of both cross-sectional dependence and structural breaks on the test of
the LOP, I propose a hybrid method in this paper.
3.3 Empirical Strategy for the Test for the LOP
3.3.1 The Regression Model
Following the literature, I test the LOP based on a regression model of regional prices and
the reference price, represented by
logPi,t = αi + βi logP ∗t + εi,t (3.3)
where logPi,t is the natural logarithmic transformation of price level in region i, and logP ∗t
is the natural logarithmic transformation of the corresponding price level in the reference
region.4 In equation (3.3), the coefficient αi measures the time-invariant difference between
the regional price and the reference price caused by factors such as transportation cost,
tax levels and the commodity’s quality. The coefficient βi measures how the regional price
co-moves with the reference price over time. When regional markets are well integrated, the
LOP holds and there will be a one-to-one relationship between the regional price and the
reference price, i.e., βi = 1. Theoretically, there are two forms of the LOP. The strong form
4Because price series are likely to have exponential growth rate, many existing studies and this paperuse a natural logarithmic transformation of prices in the regression.
62
of the LOP implies that αi = 0 and βi = 1, while the weak form allows αi to be non-zero.
When price series, logPi,t and logP ∗t , contain unit roots, equation (3.3) implies a long-
run equilibrium between logPi,t and logP ∗t when a cointegrating relationship exists. In
the presence of a cointegrating relationship, the LOP can be examined by testing the null
hypothesis of βi = 1.
To investigate the presence of an equilibrium relationship between logPi,t and logP ∗t , I
test for unit roots in a panel of regression residuals, εi,t from equation (3.3). The panel unit
root test on the residuals is subject to the presence of cross-sectional dependence resulting
from common shocks in individual prices. Hence, I consider a factor structure for the
regression residuals εi,t, which is εi,t = γiFt+vit. In such structure, the vector of unobserved
common factors Ft captures the presence of cross-sectional dependence. Meanwhile, the
idiosyncratic component, vit in the factor structure, is cross-sectionally independent. In
the panel of residuals, both Ft and vit may also contain structural breaks. This is because
structural breaks in the intercept and slope coefficients of equation (3.3) may occur; as a
result, the estimated regression model with time-invariant coefficients is misspecified and
structural breaks in the coefficients are left in the regression residuals.
In the case of the US gasoline market, structural breaks in the coefficients of equation
(3.3) are likely given that the degree of gasoline market integration in the US has changed
since the deregulation of gasoline price in the early 1980s. Muehlegger (2006) pointed out
that the US gasoline market became more differentiated because of the introduction of
more stringent and heterogenous federal and state regulations on gasoline content. There
are states that are required to use reformulated gasoline because they failed to meet air
quality standards. As a result, gasoline prices are higher in some states because they are
using gasoline products that have added operating and distribution costs. These regula-
tions lead to more differentiated gasoline products sold in regional markets and reduce the
substitutability of gasoline products between markets. Therefore, they are likely to cause a
change in the degree of market integration across states. Similarly, the closure of refineries,
production disruptions caused by natural disasters and construction of new distribution
systems may lead to permanent structural changes in the gasoline industry and affect the
degree of regional market integration. Yet, the empirical model, equation (3.3), does not
allow for any changes in its coefficients. Neglecting changes in the degree of market integra-
tion in the regression model will consequently result in structural changes in the regression
residuals of equation (3.3), i.e., in both components of the regression residuals εi,t.
The interactive effects from both unit roots and structural breaks on the regression
residuals will adversely affect the LOP test. In Table 3.1, I list four possible results for the
LOP test with both unit roots and structural breaks. Specifically, structural breaks in the
residuals indicate that the LOP regression model as equation (3.3) is misspecified. There-
fore, the LOP implied by this equation does not hold. On the other hand, the presence of
unit roots in the residuals suggests that price series do not have a cointegrating relation-
63
Table 3.1: Possible Results for the LOP Test
In the Residuals With Unit Root Without Unit Root
Structural Breaks LOP fails; Prices are not cointegrating LOP fails; Model is mis-specified
No Breaks LOP fails; Prices are not cointegrating Continue to test coefficients
ship. As a result, the test on the coefficient is invalid. To continue the test of LOP on the
regression coefficient, it requires that both unit roots and structural breaks are absent from
regression residuals.
In sum, the test for unit roots in εi,t is subject to the size distortion and reduced power
caused by both cross-sectional dependence and structural breaks. To deal with these issues,
I propose a hybrid method which combines some advanced unit root tests.
3.3.2 A Hybrid Unit Root Test
To test for unit roots in a panel with cross-sectional dependence and structural breaks,
I follow the second generation panel unit root test produce. I test for unit roots in the
common and the idiosyncratic components separately. The panel is I(0) when unit root
tests on both components reject the null hypothesis of unit roots.
More specifically, as Figure 3.1 illustrates, I first estimate equation (3.3) for each individ-
ual unit separately by OLS and saving the residuals as a panel. Then, I test for the presence
of cross-sectional dependence in the residuals through the CD test in Pesaran (2004) and
through estimating the number of common factors following Bai and Ng (2002). If the null
hypothesis of cross-sectional independence is rejected and the number of common factors
estimated is ≥ 1, I then use principal component analysis (PCA) to separate the common
and the idiosyncratic components in the panel.
I use the procedure of Bai and Carrion-i Silvestre (2009) to test for unit roots in the
idiosyncratic component. As shown in the Appendix, this method eliminates the common
component in the panel by the PCA and focuses solely on the idiosyncratic component.
For the idiosyncratic component, the unknown dates of structural breaks in each series
is estimated from the data and controlled by dummies. With individual-specific dates of
structural breaks in the idiosyncratic component, the method constructs individual modified
Sargan-Bhargava (MSB) statistics that are shown to be cross-sectionally independent. Later
on two panel test statistics are calculated using individual statistics or p values. Since this
method only tests for the unit root in the idiosyncratic component, it overlooks the unit
root in the common component.
The estimated common factors are independent from each other as a result of PCA.
64
T-‐test on coefficients:
Test for the LOP in US Gasoline Market Empirical Strategy
!
H0 :" i = 0#i =1
$ % &
Cross-‐sec/onal independence test of residuals
Reject
Obtain common and idiosyncra/c components by Principal Component Analysis
Not reject First genera/on panel unit root test with considera/on of breaks
Structural break test on each factor, Perron and Yabu (2009) ( No break )
No break
Unit root test on each factor, Augmented Dickey-‐Fuller (1976)
Break Common
!
H0 :
Es/mate
!
logPit = " i + #i logP t +$ it
Idiosyncra/c Component
Es/mate break points
Unit root test on the idiosyncra/c component
No
Factors
Reject
break
Break Not reject
The LOP is rejected
Reject
Not reject
Unit root test on the idiosyncra/c component
Unit root test on each factor, Carrion-‐I Silvestre(2009)
Panel unit root test, Bai and Carrion-‐I Silvestre (2009)
Save residuals
Figure 3.1: Flow chart for empirical strategy of panel unit root test
65
Next, to test for a unit root in the common component, I use univariate unit root tests and
test for a unit root in each of the estimated common factors separately. Prior to conducting
the unit root test, I test for the null hypothesis of no structural breaks in each factor
following Perron and Yabu (2009).5 Because the univariate test method offered in Carrion-i
Silvestre et al. (2009) will have serious size distortions when the series tested does not have
a structural break in the parameters of the trend function, I use the conventional ADF test
when the null hypothesis of no structural breaks cannot be rejected.
The presence of LOP based on equation (3.3) must satisfy two conditions sequentially:
first, there is a cointegrating relationship between prices, i.e., regression residuals do not
have unit roots; second, the β coefficient must be time-invariant and equal to one. When
the null hypothesis of unit root cannot be not rejected for the residuals, no cointegration
is found, implying that prices are diverging from each other in the long run. In contrast,
when unit root tests in both the common and the idiosyncratic components reject the
null hypothesis of unit roots, I conclude that a cointegrating relationship is supported by
equation (3.3) with estimated coefficients. Furthermore, if no structural break is found
in neither the common nor the idiosyncratic components, I continue to examine the LOP
by testing βi = 1. Otherwise, if a structural break is found in either component of the
regression residuals, it suggests that equation (3.3) is misspecified and the LOP does not
hold.
3.4 Data and Results
In this paper, I use data on monthly prices of regular gasoline sales to end users (measured
in dollars per gallon, excluding taxes) for all 50 states in the US.6 The data are available
from the Energy Information Administration (EIA), assembled by the survey form EIA-
782B. To calculate the gasoline price of a particular state, the Form EIA-782B is sent to
a scientifically selected sample of motor gasoline retailers. All sampled firms provide the
volume-weighted average pre-tax price of their sales during a particular month from all
their points of sale within a state. Therefore, the price of gasoline of a particular grade is
equal to the monthly revenue obtained from the sales of that gasoline grade divided by the
corresponding monthly volume. The reported state price is a weighted average of reports
of all the sampled firms.
Although data is available as early as January 1983, all regions have missing values
between July 1987 to December 1988. In addition, the survey was suspended by EIA in
5As explained in the appendix, the rejection of the null indicates that the relevant series contains atleast one structure break.
6Regular gasoline is defined as gasoline having an antiknock index (average of the research octane ratingand the motor octane number) greater than or equal to 85 and less than 88, and it is the most commontype of gasoline used. Moreover, the sales to end users is defined as: Sales made directly to the consumer ofthe product, including bulk consumers, such as agriculture, industry, and utilities, as well as residential andcommercial consumers. For more information, please refer to: http://www.eia.gov/dnav/pet/TblDefs/
pet_pri_allmg_tbldef2.asp
66
2011 due to the FY2011 funding level provided. Because of these reasons, my dataset ranges
from January 1989 to February 2011, which yields a total of 266 time observations for each
state.7 Following O’Connell and Wei (2002) and Suvankulov et al. (2012), the average price
level, log Pt, is chosen as the reference price.
Before estimating the empirical model of the LOP test, I first conduct some pre-tests
for the panel of individual gasoline prices. First, accounting for global shocks, I test for
the presence of cross-sectional dependence in the panel of state-level gasoline prices. The
rejection of the null hypothesis in the CD test suggests that individual log-transformed
gasoline prices are cross-sectionally dependent. As a result, I use panel unit root tests in
the second generation. In particular, the CRMA test of Sul (2009) is used to test for a unit
root in the common components of the panel and the CIPS test of Pesaran (2007) is for
unit root test in the idiosyncratic component. From Table 3.2, results indicate the panel of
logarithmic price is integrated of degree one.
Table 3.2: Tests for Individual Variables
Levels DifferenceslogP ∆ logP
CD 565.88∗ 485.75∗
CRMA -0.65 -6.97∗
CIPS(6) -18.43∗ -34.19∗
Note: Pesaran’s (2004) cross-sectional independence test (CD) follows a standard normal distribu-
tion. The 5 % critical value for Pesaran’s (2007) panel unit root test (CIPS) is -2.06. The lag
length for the CIPS test is set to T 1/3 = 6. The 5% critical value for Sul’s (2009) unit root test for
the cross-sectional means (CRMA) is -1.88. The lag length for the CRMA test is determined by
the Bayesian Information Criterion with maximum 6 lags. Statistical significance at the 5% level or
lower is denoted by ∗.
3.4.1 Main Results
Next, I investigate the LOP by taking into account both cross-sectional dependence and
structural breaks. As a preliminary examination for the presence of structural breaks, I
first test the null hypothesis of no structural breaks for each logarithmic price following the
procedure of Perron and Yabu (2009). Following the method described in the appendix
section D.1, I obtain an estimation for the location of the major structural break in each
price, which is illustrated in the plots of Figure 3.2. In Figure 3.2, I plot logarithmic state
gasoline prices over time. Because states in the West Coast region are shown to have a
7Despite the effort to obtain a balanced panel of data, few sporadic missing values remained: Arkansas(2008 m9, 2009 m8, 2009 m9, 2009 m10); Idaho (2010 m7); Montana (1996 m9, 1999 m4, 2001 m10, 2002m1, 2002 m4); North Dakota (1999 m4); New Jersey (2007 m1); Nevada (1997 m3, 1998 m6, 1998 m12,1999 m4); Oklahoma (1999 m7); Washington (2010 m2). These missing values were proxied using linearlyinterpolated values.
67
different estimated date of structural break than the rest of states, I plot the series in this
area separately.
The upper plot includes gasoline prices for all states except states in the West Coast
region8 and the vertical line is the estimated date of break9 for most states in this area.
The lower plot contains gasoline prices of states in the West Coast region. The dashed
series in black is the gasoline price for Alaska and the solid line in black is the gasoline
price for Hawaii. Vertical lines are estimated structural break dates, where the former one
is for the state of Alaska and the other is for most of other states in West Coast region.
From both plots, there are some features that should be noticed. First of all, there is co-
movement across states, illustrated by the thick parts in both plots. The common trends in
individual gasoline prices is likely to result in the presence of cross-sectional dependence in
the regression residuals of equation (3.3). Secondly, as Table 3.4 shows, the null hypothesis
of no structural break in state gasoline prices is rejected for most states, which may lead to
the presence of structural breaks in the relationship between gasoline prices, measured by
the coefficients of equation (3.3).10
Given the presence of cross-sectional dependence and structural breaks in state-level
gasoline prices, I begin testing for the LOP by first running regressions of state gasoline
prices on the reference price with a constant as equation (3.3). As mentioned before, state
gasoline prices can differ because of transportation cost, quality and etc, and these (time-
constant) differences are captured by the constant term αi in the regression.
Common shocks in individual prices may cause cross-sectional dependence in the re-
gression residuals of equation (3.3). Meanwhile, the presence of structural breaks in the
price variables may indicate structural changes in the intercept and slope coefficients, and
thus lead to structural breaks in the residuals of equation (3.3). Therefore, to deal with
both cross-sectional dependence and structural breaks in the regression residuals, I use the
hybrid unit root test method discussed in Section 3. Unlike the unit root test used in the
literature which overlooks both issues in the residuals, the hybrid method is able to correct
for adverse effects on the unit root test.
Table 3.5 and 3.6 summarize the residual diagnostic test for the panel of OLS residuals.
Particularly, Table 3.5 illustrates the preliminary test for the presence of structural break
in each residual series, while Table 3.6 summaries the result of testing the presence of
structural breaks in the common component and the result of unit root tests. As shown
in the first part of Table 3.6, the null hypothesis of cross-sectional independence in the
panel of residuals can be rejected according to the Pesaran’s (2004) CD test. Moreover,
the estimated number of common factors is five.11 With regard to structural breaks in
8See Table 3.3 for the information of states.9Details about the estimated dates of break for each state is reported in Table 3.4.
10Specifically, for most states in regions except the West Coast, the break date is at the time whenT = 127, corresponding to July, 1999. While the time of break is T = 183, which is March, 2004, for moststates in the West Coast, and it is T = 110, or February, 1998 for Alaska.
11These common factors may be due to some regional shocks that cannot be controlled by the national
68
−1.
0−
0.5
0.0
0.5
1.0
1.5
Individual Logarithm Price
States in East−Coast, Midwest, Gulf−Coast, Rocky−Mountain RegionsT
dolla
rs p
er g
allo
n in
loga
ritm
1989M1 1991M1 1993M1 1995M1 1997M1 1999M1 2001M1 2003M1 2005M1 2007M1 2009M1 2011M1
−1.
0−
0.5
0.0
0.5
1.0
1.5
Individual Logarithm Price
States in West−Coast RegionsT
dolla
rs p
er g
allo
n in
loga
ritm
1989M1 1991M1 1993M1 1995M1 1997M1 1999M1 2001M1 2003M1 2005M1 2007M1 2009M1 2011M1
Figure 3.2: Individual state-level gasoline prices and Break date (vertical line)
69
Table 3.3: Reference Table for StatesID State ID State
1 Connecticut 26 Nebraska2 Maine 27 North Dakota3 Massachusetts 28 Ohio4 New Hampshire 29 Oklahoma5 RhodeIsland 30 South Dakota6 Vermont 31 Tennessee7 Delaware 32 Wisconsin8 Maryland 33 Alabama9 New Jersey 34 Arkansas10 NewYork 35 Louisiana11 Pennsylvania 36 Mississippi12 Florida 37 New Mexico13 Georgia 38 Texas14 North Carolina 39 Colorado15 South Carolina 40 Idaho16 Virginia 41 Montana17 West Virginia 42 Utah18 Illinois 43 Wyoming19 Indiana 44 Alaska20 Iowa 45 Arizona21 Kansas 46 California22 Kentucky 47 Hawaii23 Michigan 48 Nevada24 Minnesota 49 Oregon25 Missouri 50 WashingtonNote: States 1 -17 belong to PADD-I, East Coast; States 18-32 belong to PADD-II, Mid-west; States 33 - 38 belong to PADD-III, Gulf Coast; States 39 - 43 belong to PADD-IV,Rocky Mountain; States 44 - 50 belong to PADD-V, West Coast.
70
Tab
le3.
4:T
est
for
bre
akin
ind
ivid
ual
log
(pri
ce)
IDB
reak
test
Bre
ak
dat
eB
reak
frac
tion
IDB
reak
test
Bre
akd
ate
Bre
akfr
acti
on1
8.47∗
Ju
l-99
127
2610
.36∗
Ju
l-99
127
25.7
7∗
Ju
l-99
127
277.
85∗
Dec
-97
108
35.
65∗
Ju
l-99
127
2810
.78∗
Ju
n-9
912
64
10.8
7∗
Ju
l-99
127
296.
14∗
Ju
l-99
127
57.6
8∗
Ju
l-99
127
308.
48∗
Ju
l-99
127
610
.63∗
Ju
l-99
127
3113
.17∗
Ju
l-99
127
712
.23∗
Ju
l-99
127
328.
92∗
Ju
l-99
127
812
.07∗
Ju
l-99
127
3313
.74∗
Ju
l-99
127
911.
48∗
Au
g-99
128
3418
.57∗
Ju
l-99
127
107.
96∗
Ju
l-99
127
3512
.97∗
Ju
l-99
127
117.
84∗
Ju
l-99
127
3613
.00∗
Au
g-99
128
1213.
25∗
Ju
l-99
127
378.
40∗
Mar
-04
183
1315.
35∗
Ju
l-99
127
389.
95∗
Ju
l-99
127
1419.
32∗
Ju
l-99
127
3913
.82∗
Mar
-04
183
1514.
46∗
Ju
l-99
127
4016
.05∗
Feb
-04
182
16
15.0
2∗
Ju
l-99
127
4112
.34∗
Mar
-14
183
1712.
08∗
Ju
l-99
127
4216
.99∗
Feb
-04
182
18
9.92∗
Ju
l-99
127
4314
.44∗
Mar
-14
183
1911.
42∗
Ju
l-99
127
448.
09∗
Feb
-98
110
20
7.01∗
Ju
l-99
127
456.
56∗
Feb
-03
170
21
7.32∗
Ju
l-99
127
469.
15∗
Mar
-99
123
2213.
92∗
Ju
l-99
127
4710
.48∗
Ap
r-04
184
239.
87∗
Ju
l-99
127
4810
.72∗
Feb
-04
182
24
8.07∗
Ju
l-99
127
4910
.87∗
Mar
-14
183
258.
92∗
Ju
l-99
127
5011
.56∗
Mar
-14
183
Note
:M
od
el-3
of
Per
ron
and
Yab
u(2
009)
isco
nsi
der
ed,
i.e.
,si
mult
aneo
us
bre
akb
oth
inth
ein
terc
ept
and
inth
esl
ope.∗
ind
icat
esth
enu
llhyp
oth
esis
of
no
stru
ctu
ral
bre
ak
isre
ject
edat
5%si
gnifi
cant
leve
l.
71
the residuals, results from Table 3.5 first indicate that the null of no structural break is
rejected for a majority of residual series. Therefore, during the panel unit root test, I use
the method of Bai and Carrion-i Silvestre (2009) to test for unit roots in the idiosyncratic
component and the univariate unit root test of Carrion-i Silvestre et al. (2009) for unit root
in the common component, respectively. Given Ttotal = 266 in the sample, I allow for an
arbitrary three structural breaks in the unit root test.
As explained in the appendix, the univariate unit root test of Carrion-i Silvestre et al.
(2009) requires a pre-test for the presence of structural breaks. Therefore, before unit root
test for the common component, I test for the null of no structural breaks following Perron
and Yabu (2009). The middle part of Table 3.6 suggests that for most estimated common
factors, I am able to reject the null of no break at the 5% significance level. As a result, I
use the univariate unit root test from Carrion-i Silvestre et al. (2009) to test for unit root in
each estimated common factors except the forth common factor, for which I use the ADF
test. Results from the last column of the middle part and the lower part report that the null
hypothesis of a unit root in the panel of regression residuals is not rejected due to the unit
root in one out of five common factors. In sum, following the hybrid unit root test, residuals
are not stationary because of the presence of both unit root and structural breaks in the
common factors. The unit root test on the residuals implies that there is no time-invariant
cointegrating relationship between state gasoline prices and the reference national average
price, suggesting that the regression model is misspecified and the test for the regression
coefficient is invalid.
One possible explanation for the absence of a time-invariant cointegration relationship
between the state-level price and the national average in the panel may be due to the
inclusion of prices for the state of Hawaii and Alaska. Both states are not geographically
contiguous with any other state on the mainland in the US. Therefore, the response of
gasoline price to common shocks in both states may require a longer period because some
inter-state transmission of gasoline products may not be available for them or because
the determinants of prices in both states may be different from states on the mainland.12
Excluding both states, the conclusion does not change. As Table 3.7 suggests, the null
hypothesis of unit root in the panel of residuals still cannot be rejected due to non-stationary
common factors. Thus, even for the contiguous 48 states on the mainland, individual
prices do not cointegrate with the national average price level with a constant cointegrating
vector.13
price level or may be caused by the structural breaks in the degree of market integration which are commonand left in the residuals.
12For instance, the price of gasoline in Hawaii may be different for many reasons: Hawaii is a very smallmarket that cannot easily import or export refined gasoline. In addition, the refining industry in Hawaii haslow cracking capacity, low desulfurization capacity and is undersized.
13This conclusion is in sharp contrast to the conclusion of Holmes et al. (2013) who could not reject thenull hypothesis of the presence of the LOP for state gasoline markets in the US with the same dataset.
72
Tab
le3.
5:T
est
for
bre
akin
ind
ivid
ual
OL
Sre
sid
ual
s
IDB
reak
test
Bre
ak
dat
eB
reak
frac
tion
IDB
reak
test
Bre
akd
ate
Bre
akfr
acti
on1
4.9
2∗
Au
g-99
128
261.
90S
ep-9
469
210.
56∗
Mar
-92
3927
10.0
7∗
Mar
-99
123
33.
45∗
Sep
-94
6928
11.6∗
Jan
-98
109
49.6
5∗
Mar-
92
3929
23.9
4∗
Dec
-95
845
5.8
5∗
Au
g-99
128
301.
44Ju
l-96
916
18.
10∗
Ju
l-99
127
315.
94∗
Dec
-01
156
71.
43S
ep-9
469
328.
30∗
Mar
-01
147
89.
46∗
Oct
-94
7033
12.2
6∗
Jan
-94
619
0.58
Sep
-06
213
346.
99∗
Sep
-99
129
102.7
9M
ar-
92
3935
19.8
8∗S
ep-0
520
111
2.5
3M
ar-
92
3936
7.42∗
Sep
-94
6912
10.8
7∗
Mar-
0215
937
1.55
Mar
-05
195
13
9.98∗
Sep
-98
117
385.
01∗
Feb
-02
158
14
8.53∗
Au
g-93
5639
8.26∗
Mar
-92
3915
6.2
2∗
Ju
l-99
127
406.
66∗
Nov
-99
131
16
5.54∗
Oct
-94
7041
2.66
Nov
-98
119
17
15.
57∗
Ju
l-97
103
422.
53D
ec-9
913
218
11.
40∗
Nov
-02
167
431.
68A
pr-
9240
19
2.25
Dec
-99
132
447.
37∗
Mar
-99
123
203.3
8∗
Mar-
94
6345
5.92∗
Mar
-92
3921
5.1
5∗
Jan
-00
133
463.
27∗
Oct
-98
118
223.3
5∗
Feb
-99
122
473.
44∗
Mar
-99
123
2313
.23∗
Oct
-99
130
4812
.70∗
Oct
-96
9424
5.2
4∗
Oct
-02
166
491.
08N
ov-9
913
125
5.8
8∗
Nov
-99
131
508.
67∗
Oct
-99
130
Note
:M
od
el-3
of
Per
ron
and
Yab
u(2
009)
isco
nsi
der
ed,
i.e.
,si
mult
aneo
us
bre
akb
oth
inth
ein
terc
ept
and
inth
esl
ope.∗
ind
icat
esth
enu
llhyp
oth
esis
of
no
stru
ctu
ral
bre
ak
isre
ject
edat
5%si
gnifi
cant
leve
l.
73
Table 3.6: Diagnostic tests for OLS residuals
Presence of Common Factors (CF)
Statistics Conclusion
CD 11.16∗ dependentNo. of CF 5
Test for Structural Break and Unit Root in CF
Factors Break Test Break Date (T ) Unit Root Test Unit Root?
CF1 5.49∗ May-00, Feb-06, Oct-08 (137, 206, 238) 8.91 YesCF2 8.98∗ Mar-99. Jun-01, Mar-04 (123, 150, 183) 5.57∗ NoCF3 3.49∗ Aug-94, Oct-98, Nov-01 (68, 118, 155) 4.07∗ NoCF4 2.52 — -4.64∗ NoCF5 5.41∗ Feb-98, Apr-01, Sep-03 (110, 148, 177) 3.42∗ No
Test for Unit Root in Idiosyncratic
Z statistics P statistics Pm statistics Unit root?
-5.25∗ 509.94∗ 28.98∗ No
Note: The null hypothesis of CD test is cross-sectional independent. The 5% critical valuefor break test (Exp −WFS) is 3.12 for model with simultaneous break in both interceptand slope. The number of structural breaks in unit root test is 3. The estimated fractionsof structural breaks are reported in the parenthesis in the third column of the middle part.Unit root test is based on the value of MPGLST (For Factor 4, I report the ADF statistics).For unit root test in idiosyncratic component, the number of structural breaks is 3 andstatistics are illustrated in Appendix equation D.25, D.26, and D.27 specifically.
74
Table 3.7: Diagnostic tests for OLS residuals, excluding Hawaii and Alaska
Presence of Common Factors (CF)
Statistics Conclusion
CD 19.06∗ dependentNo. of CF 4
Test for Structural Break and Unit Root in CF
Factors Break Test Break Date (T ) Unit Root Test Unit Root?
CF1 3.49∗ May-00, Feb-06, Oct-08 (137, 206, 238) 8.84 YesCF2 9.69∗ Dec-91, Aug-94, Mar-99 (36, 68, 123) 4.99∗ NoCF3 1.84 — -6.43∗ NoCF4 8.29∗ Mar-00, Apr-03, Sep-05(135, 172, 201) 3.37∗ No
Test for Unit Root in Idiosyncratic
Z statistics P statistics Pm statistics Unit root?
-5.14∗ 535.21∗ 31.69∗ No
Note: The null hypothesis of CD test is cross-sectional independent. The 5% critical valuefor break test (Exp −WFS) is 3.12 for model with simultaneous break in both interceptand slope. The number of structural breaks in unit root test is 3. The estimated fractionsof structural breaks are reported in the parenthesis in the third column of the middle part.Unit root test is based on the value of MPGLST (For Factor 3, I report the ADF statistics).For unit root test in idiosyncratic component, the number of structural breaks is 3 andstatistics are illustrated in Appendix equation D.25, D.26, and D.27 specifically.
75
3.4.2 Test of the LOP based on Relative Prices
The LOP implies that relative price series, logPi,t/ log Pt, is mean-reverting, or I(0). There-
fore, the LOP can also be examined by testing for unit roots in relative price series. Since
relative price series implicitly assume a fixed value of the coefficients (βi = 1) in equation
(3.3), it ignores the change in the degree of market integration. Moreover, unobserved com-
mon factors in individual price series will also lead to cross-sectionally dependent relative
price series. As a result, the test of the LOP based on relative price series is likely to be
affected by both cross-sectional dependence and structural breaks.
Table 3.8 presents the preliminary test for the structural break in each relative price
series. As suggested, the null hypothesis of no structural break is rejected for most states.
In addition, I construct a panel of relative price series and test for the presence of cross-
sectional dependence and structural breaks in both components of the panel. Results are
reported in Table 3.9. As previously, I allow for three structural breaks in the unit root
test.
Based on results in Table 3.9, I reject the null hypothesis of cross-sectional independence
in the CD test and estimate five common factors for the panel of relative price series. In
addition, I find rejections of the null of no structural break for three common factors.
Finally, based on the unit root test of both components, I find that the panel of relative
price series contains a common unit root. As in the previous section where I examine the
situation when excluding the state of Hawaii and Alaska, I also construct a panel of 48
states on the mainland and repeat the same empirical analysis. I report relevant results in
Table 3.10. The same conclusion for all 50 states that the panel of relative price series is
I(1) also holds when excluding Hawaii and Alaska.
In Table 3.11, I list the dates of structural break that are reported in both tests of the
LOP regression residuals and the relative prices. The dates of structural breaks estimated
from common factors coincide with some vital shocks to the petroleum industry in the US.
For example, March 1999 and May 2000 are indicated as two dates of structural break.
During this period, there were major mergers in the petroleum industry in the US. The
merger of petroleum company is likely to change the production and distribution of gasoline.
Thus, it is possible that it will also change the degree of state market integration. Moreover,
the common break date at February 2006 may be due to the effects hurricane Katrina
and Rita in the the September of 2005. Both disrupted the production of oil in the Gulf
Coast. Finally, the global recession in 2008-2009 seems also affected the structure of gasoline
industry in the US, through the possible channels such as the closure of refineries or the
modifications to the distribution system.
As reviewed in Section 2, many studies on the degree of gasoline market integration
are based on unit root tests of relative prices, ignoring the critical issues of cross-sectional
dependence and structural breaks. According to results in Table 3.8, 3.9 and 3.10, it is
likely that these studies are subject to size distortion and reduced power problems in the
76
Tab
le3.
8:T
est
for
bre
akin
ind
ivid
ual
log
rela
tive
pri
ce(p
rice
)
IDB
reak
test
Bre
ak
dat
eB
reak
frac
tion
IDB
reak
test
Bre
akd
ate
Bre
akfr
acti
on1
6.2
1∗
Au
g-99
128
262.
44D
ec-9
913
22
4.11∗
Mar
-92
3927
2.96
Mar
-99
123
34.
00∗
Au
g-99
128
289.
29∗
Jan
-98
109
45.
73∗
Mar
-92
3929
23.2
9∗
Oct
-93
585
6.80∗
Au
g-99
128
301.
42A
ug-
9692
612
.27∗
Mar
-92
3931
10.2
3∗
Au
g-99
128
72.
16S
ep-9
469
325.
42∗
Mar
-01
147
89.
31∗
Oct
-94
7033
12.8
2∗
Oct
-93
589
0.48
Oct
-99
130
3413
.61∗
Au
g-99
128
104.6
2∗
Mar-
92
3935
20.2
4∗
Dec
-03
180
114.0
0∗
Oct
-97
106
368.
47∗
Sep
-94
6912
19.4
7∗
Mar-
0215
937
1.44
Mar
-92
3913
18.
59∗
Ju
n-9
912
638
5.35∗
Feb
-02
158
14
13.
87∗
Au
g-99
128
3916
.64∗
Mar
-92
3915
15.0
6∗
Ju
l-99
127
4010
.31∗
Nov
-99
131
16
8.92∗
Oct
-94
7041
6.24∗
Nov
-98
119
17
22.
84∗
Ju
l-97
103
422.
94D
ec-9
913
218
6.94∗
Nov
-02
167
4312
.11∗
Nov
-99
131
19
1.51
Dec
-99
132
446.
96∗
Mar
-99
123
204.2
7∗
Mar-
94
6345
7.20∗
Ap
r-92
4021
7.5
6∗
Jan
-00
133
462.
61O
ct-9
811
822
8.3
2∗
Ju
n-9
242
474.
82∗
Mar
-99
123
2314
.25∗
Oct
-99
130
4815
.74∗
Au
g-96
9224
10.8
2∗
Oct
-02
166
491.
95O
ct-9
913
025
5.8
0∗
Nov
-99
131
5010
.14∗
Oct
-99
130
Note
:M
od
el-3
of
Per
ron
and
Yab
u(2
009)
isco
nsi
der
ed,
i.e.
,si
mult
aneo
us
bre
akb
oth
inth
ein
terc
ept
and
inth
esl
ope.∗
ind
icat
esth
enu
llhyp
oth
esis
of
no
bre
akis
reje
cted
at5%
sign
ifica
nt
leve
l.
77
Table 3.9: Diagnostic tests for relative price level
Presence of Common Factors (CF)
Statistics ConclusionCD 10.44∗ dependentNo. of CF 5
Test for Structural Break in CF
Factors Break Test Break Date (T ) Unit Root Test Unit Root?CF1 9.59∗ May-00, Feb-06, Oct-08 (137, 206, 238) 9.11 YesCF2 10.49∗ Mar-99, Jun-01, Mar-04 (123, 150, 183) 5.46∗ NoCF3 3.04 — -3.06∗ NoCF4 0.96 — -7.68∗ NoCF5 7.61∗ Feb-98, Sep-03, Apr-07 (110, 177, 220) 3.47∗ No
Test for Unit Root in Idiosyncratic
Z statistics P statistics Pm statistics Unit root?
-5.08∗ 355.42∗ 18.06∗ No
Note: The null hypothesis of CD test is cross-sectional independent. The 5% critical valuefor break test (Exp −WFS) is 3.12 for model with simultaneous break in both interceptand slope. The number of structural breaks in unit root test is 3. The estimated fractionsof structural breaks are reported in the parenthesis in the third column of the middle part.Unit root test is based on the value of MPGLST (For Factor 3 and 4, I report the ADFstatistics). For unit root test in idiosyncratic component, the number of structural breaksis 3 and statistics are illustrated in Appendix equation D.25, D.26, and D.27 specifically.
78
unit root test. Therefore, their conclusions about the LOP are questionable.
3.5 Conclusion
In this paper, I study the degree of retail gasoline market integration in the US taking into
account two important issues neglected in the literature: cross-sectional dependence and
structural breaks. To deal with adverse effects from both issues, I come up with a hybrid
panel unit root test by combining methods that are recently developed. Using this hybrid
method, I find that previous conclusions about perfect market integration or the LOP in
the US gasoline market may be due to the failure of considering these issues.
More specifically, I show that cross-sectional dependence and structural breaks are
present in the panel regression residuals of an empirical model and in the panel of rela-
tive price series. As a result, the test of the LOP is adversely affected if adhering to the
conventional method in the literature. I use a hybrid unit root test procedure to test for
unit roots in a dependent panel while at the same time accounting for structural breaks.
In contrast to previous literature that found a cointegrating relationship between gasoline
prices, I cannot reject the null of no cointegration.
My results also have implications for energy policy: results here suggest that state
gasoline markets are not spatially integrated. Therefore a nation-wide policy change such
as the Renewable Fuel Standard (RFS) program or the Keystone Pipeline project may not
affect all state gasoline prices with the same magnitude. Moreover, state gasoline prices used
in this paper do not contain tax at any level. Therefore, it is more likely that state-gasoline
prices with tax also diverge as the tax level varies a lot across states.
As mentioned earlier, the regression model used to test for the LOP may be misspecified
when there are changes in the degree of gasoline market integration. In this paper, I focus
on the test of a constant LOP and do not explicitly model these changes. Therefore, they
are left as structural changes in the regression residuals. Given the empirical findings in
this paper and the feature of the gasoline market in the US, a natural extension of this
paper is to model the structural break in the coefficient of regression model explicitly and
examine the LOP with the modified model.
Finally, the performance of the proposed hybrid panel unit root test may be better
evaluated with the help of some Monte Carlo simulations. With small sample size, the
estimated date of structural breaks may be biased, and reduced power and size distortion
may exist in the hybrid test method. Therefore, the results of structural break dates and
the conclusion of the absence of a cointegrating relationship should be used with cautions.
79
Table 3.10: Diagnostic tests for relative price level, excluding Hawaii and Alaska
Presence of Common Factors (CF)
Statistics ConclusionCD 19.44∗ dependentNo. of CF 5
Test for Structural Break in CF
Factors Break Test Break Date (T ) Unit Root Test Unit Root?CF1 8.24∗ May-00, Feb-06, Oct-08 (137, 206, 238) 9.08 YesCF2 13.23∗ Mar-99, Jun-01, Mar-04 (123, 150, 183) 5.60∗ NoCF3 1.54 — -6.04∗ NoCF4 12.06∗ Dec-98, Sep-03, Sep-08 (120, 177, 237 ) 3.39∗ NoCF5 5.48∗ Nov-91, Feb-94, Jun-01(35, 62, 150) 3.11∗ No
Test for Unit Root in Idiosyncratic
Z statistics P statistics Pm statistics Unit root?
-5.08∗ 355.42∗ 18.06∗ No
Note: The null hypothesis of CD test is cross-sectional independent. The 5% critical valuefor break test (Exp −WFS) is 3.12 for model with simultaneous break in both interceptand slope. The number of structural breaks in unit root test is 3. The estimated fractionsof structural breaks are reported in the parenthesis in the third column of the middle part.Unit root test is based on the value of MPGLST (For Factor 3, I report the ADF statistics).For unit root test in idiosyncratic component, the number of structural breaks is 3 andstatistics are illustrated in Appendix equation D.25, D.26, and D.27 specifically.
80
Table 3.11: Estimated Dates of Break in Common FactorsBreak date OLS residuals Relative prices OLS residuals Relative prices
(50) (50) (48) (48)
(Year-Month)1999-03 X X X XPotential Explanation Major mergers in the petroleum industry:
BP and Amoco in 1998; Exxon and Mobil in 1999
2000-05 X X X XPotential Explanation Major mergers in the petroleum industry:
BP/Amoco and Arco in 2000
2004-03 X X XPotential Explanation Nine million acres of Alaska’s National Petroleum Reserve opened
for long-term production
2006-02 X X X XPotential Explanation Hurricane Katrina &Rita in 2005 Aug-Sep
2008-10 X X X XPotential Explanation Worldwide recession
81
Chapter A
Appendix: Tables
A.1 Additional Tables for Essay 1
Table A.1: State Code and Regional CPI
States Regional Consumer Price Index
AK Anchorage, AK (MSA)AL SouthAR SouthAZ WestCA Los Angeles-Riverside-Orange County, CA (CMSA)CO Denver-Boulder-Greeley, CO (CMSA)CT New York-Northern New Jersey-Long Island, NY-NJ-CT-PA (CMSA)FL Miami-Fort Lauderdale, FL (CMSA)GA Atlanta, GA (MSA)IA MidwestID WestIL Chicago-Gary-Kenosha, IL-IN-WI (CMSA)IN Chicago-Gary-Kenosha, IL-IN-WI (CMSA)KS Kansas City, MO-KS (MSA)KY Cincinnati-Hamilton, OH-KY-IN (CMSA)LA SouthMA Boston-Brockton-Nashua, MA-NH-ME-CT (MSA)MD Philadelphia-Wilmington-Atlantic City, PA-NJ-DE-MD (CMSA)ME Boston-Brockton-Nashua, MA-NH-ME-CT (MSA)MI Detroit-Ann Arbor-Flint, MI (CMSA)MN Minneapolis-St. Paul, MN-WI (MSA)MO Kansas City, MO-KS (MSA)MS SouthMT West
82
Table A.2: State Code and Regional CPI, continued
States Regional Consumer Price Index
NC SouthND MidwestNE MidwestNH Boston-Brockton-Nashua, MA-NH-ME-CT (MSA)NJ New York-Northern New Jersey-Long Island, NY-NJ-CT-PA (CMSA)NM WestNV WestNY New York-Northern New Jersey-Long Island, NY-NJ-CT-PA (CMSA)OH Cleveland-Akron, OH (CMSA)OK SouthOR Portland-Salem, OR-WA (CMSA)PA New York-Northern New Jersey-Long Island, NY-NJ-CT-PA (CMSA)RI NortheastSC SouthSD MidwestTN SouthTX Dallas-Fort Worth, TXUT WestVA Washington-Baltimore, DC-MD-VA-WV(CMSA)VT NortheastWA Seattle-Tacoma-Bremerton, WA (CMSA)WI Chicago-Gary-Kenosha, IL-IN-WI (CMSA)WV Washington-Baltimore, DC-MD-VA-WV(CMSA)WY West
83
A.2 Additional Tables for Essay 2
Table A.3: Sub-Sample Country GroupGroup Name Country List
High Income Australia, Austria, Bahamas, Bahrain, Barbados, Belgium, Bermuda, Brunei, Canada,Cyprus, Denmark, Equatorial Guinea, Finland, France, Germany, Greece,Hong Kong, Hungary, Iceland, Ireland, Israel, Italy, Japan, Republic of Korea,Luxembourg, Macao, Malta, Netherlands, New Zealand, Norway, Oman, Poland,Portugal, Puerto Rico, Singapore, Spain, St. Kitts & Nevis, Sweden, Switzerland,Trinidad &Tobago, United Kingdom, United States
UpperMid Income Algeria, Angola, Antigua & Barbuda, Argentina, Botswana, Brazil, Bulgaria,Chile, China, Colombia, Costa Rica, Cuba, Dominica, Dominican Republic,Ecuador, Gabon, Grenada, Iran, Jamaica, Jordan, Lebanon, Malaysia, Maldives,Mauritius, Mexico, Namibia, Palau, Panama, Peru, Romania, Seychelles,South Africa, St. Lucia, St.Vincent & Grenadines, Suriname, Thailand, Tunisia,Turkey, Uruguay, Venezuela
LowerMid Income Albania, Belize, Bhutan, Bolivia, Cameroon, Cape Verde, Republic of Congo,Cote d‘Ivoire, Djibouti, Egypt, El Salvador, Fiji, Ghana, Guatemala, Guyana,Honduras, India, Indonesia, Iraq, Kiribati, Lesotho, Marshall Islands,Fed. Sts. Micronesia, Mongolia, Morocco, Nicaragua, Nigeria, Pakistan,Papua New Guinea, Paraguay, Philippines, Samoa, Sao Tome & Principe,Senegal, Solomon Islands, Sri Lanka, Sudan, Swaziland, Tonga, Vanuatu,Vietnam, Zambia
Low Income Afghanistan, Bangladesh, Benin, Burkina Faso, Burundi, Cambodia,Central African Republic, Chad, Comoros, Dem. Rep. Congo, Ethiopia, Gambia,The, Guinea, Guinea-Bissau, Haiti, Kenya, Liberia, Madagascar, Malawi, Mali,Mauritania, Mozambique, Nepal, Niger, Rwanda, Sierra Leone, Somalia, Tanzania,Togo, Uganda, Zimbabwe
Note: Sub-sample are grouped according to World Bank Country Groups and based on dataavailability of the whole sample.
84
Table A.4: Sub-Sample Country Group, continuedGroup Name List of Countries
OECD-AH Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Greece,Iceland, Ireland, Italy, Japan, Luxembourg, Netherlands, New Zealand,Norway, Portugal, Spain, Sweden, Switzerland, United Kingdom, United States
OECD Australia, Austria, Belgium, Canada, Chile, Denmark, Finland, France, Germany,Greece, Hungary, Iceland, Ireland, Israel, Italy, Japan, Republic of Korea, Luxembourg,Mexico, Netherlands, New Zealand, Norway, Poland, Portugal, Spain, Sweden,Switzerland, Turkey, United Kingdom, United States
Non-OECD All rest countries in the sample except OECD countries
Developed Australia, Austria, Belgium, Canada, Cyprus, Denmark, Finland, France, Germany,Greece, Hungary, Israel, Ireland, Italy, Japan, Luxembourg, Malta, Netherlands,New Zealand, Norway, Poland, Portugal, Romania, Spain, Sweden, Switzerland,United Kingdom, United States
Developing Algeria, Angola, Argentina, Bahrain, Bangladesh, Barbados, Benin, Bermuda,Bolivia, Botswana, Brazil, Burkina Faso, Burundi, Cameroon, Cape Verde,Central African Republic, Chad, Chile, China, Colombia, Comoros,Dem. Rep. Congo, Republic of Congo, Costa Rica, Cote d‘Ivoire, Cuba, Djibouti,Dominican Republic, Ecuador, Egypt, El Salvador, Equatorial Guinea, Ethiopia,Gabon, The Gambia, Ghana, Guatemala, Guinea, Guinea-Bissau, Guyana, Haiti,Honduras, Hong Kong, India, Indonesia, Iran, Iraq, Israel, Jamaica, Jordan, Kenya,Republic of Korea, Lebanon, Lesotho, Liberia, Madagascar, Malawi, Malaysia, Mali,Mauritania, Mauritius, Mexico, Morocco, Mozambique, Namibia, Nepal, Nicaragua,Niger, Nigeria, Oman, Pakistan, Panama, Papua New Guinea, Paraguay, Peru,Philippines, Rwanda, Sao Tome and Principe, Senegal, Sierra Leone, Singapore,Somalia, South Africa, Sri Lanka, Sudan, Taiwan, Tanzania, Thailand, Togo,Trinidad & Tobago, Tunisia, Turkey, Uganda, Uruguay, Venezuela, Vietnam,Zambia, Zimbabwe
EU Austria, Belgium, Bulgaria, Cyprus, Denmark, Finland, France, Germany,Greece, Hungary, Ireland, Italy, Luxembourg, Malta, Netherlands, Poland,Portugal, Romania, Spain, Sweden, United Kingdom
Note: Sub-sample are grouped according to World Bank Country Groups and based on dataavailability of the whole sample.
85
Chapter B
Appendix: Figures
B.1 Additional Figures for Essay 1
Histogram for income elasticity
Beta for income
Fre
quen
cy
−2 −1 0 1 2 3 4
05
1015
β1i
Figure B.1: Distribution of individual coefficient estimates.
The regression equation is log V ISit = αi + β1i log(
YitCPIit
∗ 100)
+ β2i log(PAIRitCPIit
∗ 100)
+
β3i log(PRMtCPIit
∗ 100)
+ uit.
86
Histogram for price elasticity of airfare
Beta for airfare
Fre
quen
cy
−1.0 −0.5 0.0 0.5
02
46
810
12
β2i
Figure B.2: Distribution of individual coefficient estimates.
The regression equation is log V ISit = αi + β1i log(
YitCPIit
∗ 100)
+ β2i log(PAIRitCPIit
∗ 100)
+
β3i log(PRMtCPIit
∗ 100)
+ uit.
87
Histogram for price elasticity of room rate
Beta for room_rate
Fre
quen
cy
−5 0 5
05
1015
β3i
Figure B.3: Distribution of individual coefficient estimates.
The regression equation is log V ISit = αi + β1i log(
YitCPIit
∗ 100)
+ β2i log(PAIRitCPIit
∗ 100)
+
β3i log(PRMtCPIit
∗ 100)
+ uit.
88
Chapter C
Appendix for Essay 2
C.1 The Comparison Between the Pooled Estimator and the
Mean Group Estimator
Coakley et al. (2001) explored the small sample properties of a mean group and two pooled
panel estimators of a regression coefficient under homogeneous and heterogenous coefficients
designs. For a static linear regression of a panel,
yit = αi + βixit + uit, (C.1)
three estimators are compared: the pooled OLS (POLS) estimator, the fixed effects (FE)
or ’within’ estimator, and the mean group (MG) estimator.
First, under the design of homogeneous coefficient βi, all three estimators are shown to
be unbiased by Monte Carlo simulations. In contrast, when βi is allowed to vary randomly
across groups, with E(βi) = β, it is shown in the simulations that the two pooled estimators
are likely to be biased when the regressor xit is correlated with βi.
In particular, under the assumption of heterogenous βi, the POLS and FE estimator
actually estimates
yit = αi + βxit + uit, (C.2)
with uit = βixit−βxit+uit. In such regression, the component (βi−β)xit is in the error term
and will introduce dependence between the estimated coefficient and the regressor. This is
because the fixed effect estimator βFE can be written as a weighted average of individual
estimates,
βFE =N∑i=1
(Si∑Ni=1 Si
)βi =N∑i=1
wiβi (C.3)
As a result, when the weight, defined as Si∑Ni=1 Si
, is positively correlated with βi, the βFE
is biased upward and vice visa.
89
With heterogenous βi, the mean group estimator,
βMG = N−1N∑i=1
βi (C.4)
coincides with βFE when Si = S or the weights wi are independent of βi. Otherwise, the
difference between the mean group estimator and the FE estimator exists when both N and
T go to infinity. The attraction of the mean group estimator is that since the individual
estimates tend to show extreme heterogeneity, averaging may produce better estimates.
This would be the case when the heterogeneity is the product of country-specific shocks
which happen to be correlated with the regressors but which cancel out when averaged
across countries.
90
Chapter D
Appendix for Essay 3
In this section, I provide details about three tests used in this paper. They are: the test
for the presence of structural break of Perron and Yabu (2009), the univariate test for a
unit root of Carrion-i Silvestre et al. (2009) and the panel test for unit roots of Bai and
Carrion-i Silvestre (2009).
D.1 Univariate Test for the Presence of Structural Breaks in
the Time Series
Perron and Yabu (2009) propose a Wald test for the null hypothesis of no structural breaks
in a univariate time series without knowing whether the relevant series is I(0) or I(1).
Specifically, the test is based on an assumed DGP of the series as,
yt = x′tΨ + ut, (D.1)
ut = αut−1 + et, with − 1 < α ≤ 1 (D.2)
where xt is a (r × 1) vector of deterministic components and Ψ is a (r × 1) vector of
unknown parameters. The null hypothesis is no structural break: RΨ = γ, with R is a
(q×r) full-rank matrix and γ is a (q×1) vector with q restrictions. Therefore, the rejection
of the null suggests that there is, at least, a structural break. In this paper, three types of
the break are considered (assuming the alternative is a single structure break):
Model-1 (Structural change in the intercept of a series or level shift):
xt = (1, DUt, t)′ and Ψ = (µ0, µ1, β0)′, where DUt = 1 for t > T1.
The hypothesis of interests is µ1 = 0.
Model-2 (Structural change in the slope or in the time trend):
xt = (1, t,DTt)′ and Ψ = (µ0, β0, , β1)′, where DTt = (t− T1) for t > T1.
The hypothesis of interests is β1 = 0.
Model-3 (Simultaneous structural changes both in the intercept and in the slope):
xt = (1, DUt, t,DTt)′ and Ψ = (µ0, µ1, β0, , β1)′.
The hypothesis of interests is µ1 = β1 = 0.
91
To test the null hypothesis of no structural break, the Wald statistics for testing the
restriction of parameters is obtained via feasible GLS estimation. The GLS estimation of
the parameters is obtained by applying OLS to the regression as,
(1− αL)yt = (1− αL)x′tΨ + (1− αL)ut, (D.3)
for t= 2, ....T and together with
y1 = x′1Ψ + u1. (D.4)
for t = 1. Under the case of unknown structural breaks, the estimation of the break date
is obtained by minimizing the sum of squared residuals from a regression of the relevant
series on a constant, a time trend, a level-shift dummy and a slope-shift dummy. The value
of α is obtained by a bias-corrected estimates, denoted as αM . When T12 |αM − 1| > 1,
αM = (
T∑t=2
utut−1/
T∑t=2
u2t−1) + C(τ)σα (D.5)
and otherwise αM = 1. In the above equation, C(τ) is calculated following equation (7) in
Perron and Yabu (2009) and σα is the standard deviation of the OLS estimate of α, which
is (∑T
t=2 utut−1/∑T
t=2 u2t−1). The Wald statistic for testing the null hypothesis is defined
as,
WFS(λ1) = [R(Ψ − Ψ)]′[s2R(X′X)−1R′]−1[R(Ψ − Ψ)] (D.6)
where X = (1− αML)xt for t = 2, ...T , and = x1 for t = 1 and s2 = T−1∑T
t=1 e2t .
The statistics defined in equation (9) assumes the prior knowledge of the date of struc-
tural break. To account for the unknown break date, an exponential functional form of the
Wald statistics is modified as,
Exp−WFS = log
[T−1
∑Λ
exp(1
2WFS(λ′1))
](D.7)
where Λ = ε ≤ λ′1 ≤ 1− ε for some ε > 0 and λ′1 denotes the generic break date. In this
paper, the value of the parameter ε is set to be 0.01.
D.2 Univariate Test for a Unit Root with Structural Breaks
Early unit root tests dealing with structural breaks did not model the breaks in the null
hypothesis. This means that, under the null hypothesis, a level shift in the series is modeled
as it comes from the tail of the distribution of the DGP, and a slope change is taken as
errors with a different mean in sub-samples (Kim and Perron, 2009). As argued by Carrion-i
92
Silvestre et al. (2009), Kim and Perron (2009) and many others, since a structural break will
happen regardless of the unit root in the series, unit root test which assume an unknown
date of structural break occurring only under the alternative hypothesis of stationarity are
undesirable for two reasons. First of all, these tests impose an asymmetric treatment when
allowing for a structural break. As a result, these tests are shown to be not invariant to the
structural break when the DGP of the series has a unit root and a structural break. Secondly,
early unit root tests do not exploit the information from the presence of a structural break
in the series, and are shown to have small power.
To avoid disadvantages of early unit root tests and to improve the power of tests, Kim
and Perron (2009) propose a new procedure for unit root test which allows for a structural
break under both hypothesis. And Carrion-i Silvestre et al. (2009) extend their method
to allow for an arbitrary number of structural breaks and to the use of the quasi-GLS de-
trending method to obtain better local asymptotic power for a variety of tests. In the paper
of Carrion-i Silvestre et al. (2009), the series yt is assumed to be generated by the following
DGP,
yt = dt + ut, (D.8)
ut = αut−1 + vt, (D.9)
and the deterministic component dt is given by
dt = z′tΨ. (D.10)
Under the situation of m structural breaks, zt = [z′t(T00 ), ..., z′t(T
0m)] and Ψ = (Ψ′0, ...,Ψ
′m),
with the first terms as deterministic components and corresponding coefficients: zt(T00 ) =
(1, t)′ and Ψ0 = (µ0, β0)′.
In this paper, three types of breaks are considered:
Model-1, break in the intercept only.
dt = µ0 + β0t+ µjDUt(T0j ), for 1 ≤ j ≤ m.
Model-2, break in the time trend only.
dt = µ0 + β0t+ βjDTt(T0j ), for 1 ≤ j ≤ m.
Model-3, simultaneous break in both the intercept and the time trend.
dt = µ0 + β0t+ µjDUt(T0j ) + βjDTt(T
0j ), for 1 ≤ j ≤ m.
Here, DUt(T0j ) = 1 and DTt(T
0j ) = (t − T 0
j ) for t > T 0j , and 0 otherwise and the T 0
j
denote actual break dates from the DGP. In this paper, the unknown fraction of structural
break, λ = Tj/T , is estimated from a global minimization of the sum of squared residuals
(SSR) of the GLS-detrended model discussed below, over all possible break fractions, i.e.,
λ = argminS(α, λ).
93
The feasible point optimal statistics, PGLST = {S(α, T 0) − αS(1, T 0)}/s2(T 0),1 to test
the null hypothesis of a unit root, α = 1 against the alternative hypothesis α = α < 1, is
constructed from quasi-GLS de-trending variables defined as yα1 = y1, zα1 = z1 and
yαt = (1− αL)yt, (D.11)
zαt (T 0) = (1− αL)zt(T0), (D.12)
α = 1 + c/T (D.13)
and the choice of c follows Elliott et al. (1996), which decides the alternative hypothesis as
α = 1 + c/T .
To overcome the size distortion of PGLST for the case with negative MA coefficients and
to account for the presence of multiple structural breaks, a new test statistics, MPGLST =
[c2T−2∑T
t=1 y2t−1 + (1 − c)T−1y2
T ]/s2(T 0), is considered in this paper. Carrion-i Silvestre
et al. (2009) shows that the limiting distribution of both PGLST and MPGLST under estimated
dates of structural break are the same as in the case of known dates of break. Moreover,
the simulation shows that the size of both tests is close to the normal size and the power
quickly approaches to the limit value suggested.
Yet, results from simulations also indicate that both statistics will exhibit size distortion
when the true DGP of the series tested has no structural break. This is because when
no structural break occurs, the estimate of the break fraction has a non-degenerate limit
distribution on the interval [0, 1] instead of converging to either 0 or 1 under the null
hypothesis and the asymptotic results derived under the case with structural break do not
hold. To deal with this issue, the authors suggests that a pre-test for the presence of
structural breaks following Perron and Yabu (2009) can help. They argue that If there is
no break in the trend function, the proper unit root test procedure is to simply apply a
standard Dickey and Fuller (1979) type test.
D.3 Panel Test for Unit Roots with Structural Breaks and
Common Factors
To test for unit roots in the idiosyncratic disturbance terms of a panel, taking in to ac-
count common trends and structural breaks in individual deterministic components, Bai and
Carrion-i Silvestre (2009) propose a modified Sargan-Bhargava test. In the panel dataset,
1Where S(α, T 0) is the minimum of function: S∗(Ψ, α, T 0) =∑Tt=1(yαt − Ψ′zαt (T 0)). And s2(T 0) is the
estimate of the spectral density at frequency zero of vt.
94
the DGP for each series Yi,t is assumed as:
Yi,t = Di,t + F ′tπi + ei,t, (D.14)
(1− L)Ft = C(L)ut, (D.15)
(1− ρiL)ei,t = Hi(L)εi,t (D.16)
with Di,t = µi +∑li
j=1 θi,jDUi,j,t, and DUi,j,t = 1 for t > T ij for model 1(level shifts),
and Di,t = µi + βit +∑li
j=1 θi,jDUi,j,t +∑mi
k=1 γi,kDTi,k,t, and DUj,t = 1 for t > T ia,k and
DTk,t = (t− Tb,k) for t > T ib,k for model 2 (level and slope changes).
The null hypothesis, ρi = 1, suggests a unit root in the idiosyncratic component of the
individual series with allowance for structural breaks. In the above set-up, common factors,
Ft, capture the co-movement of individual series as well as cross-sectional correlation. To
estimate unobserved factors and idiosyncratic components consistently, procedures from Bai
and Ng (2004) are applied to differenced data. Differencing equation (17), the model for
level shifts becomes,
∆Yi,t = ∆F ′tπi + ∆e∗i,t, (D.17)
∆e∗i,t = ∆ei,t +
li∑j=1
θi,jI(T ij ) (D.18)
With I(T ij ) being impulses, obtained by differencing the mean breaks DUi,j,t. To be more
compact, equation for differenced model can be rewritten as:
yi = fπi + zi (D.19)
Where yi = ∆Yi = (∆Yi,2,∆Yi,3, ...,∆Yi,T )′, and f = ∆F = (∆F2,∆F3, ...,∆FT )′, and
zi = ∆e∗i = (∆e∗i,2,∆e∗i,3, ...,∆e
∗i,T )′.
On the other hand, the model for both level slope changes becomes,
∆Yi,t = ∆F ′tπi + βi +
mi∑k=1
γi,kDUi,k,t + ∆e∗i,t, (D.20)
∆e∗i,t = ∆ei,t +
li∑j=1
θi,jI(T ij ) (D.21)
And in a matrix format, it becomes,
yi = fπi + aiδi + zi (D.22)
Where yi, f and zi are defined earlier, and δi = (βi, γi,1, ..., γ′i,mi
) and ai = (ai,2, ..., ai,T )′
with ai,t = (1, DUi,t, ..., DUi,mi,t)′.
95
Because the principal components method provides consistent estimates for f, πi, zi as
in Bai and Ng (2004), the panel unit root test is based on independent series of idiosyncratic
components computed as,
ei,t =t∑
s=2
zi,s (D.23)
The Modified Sargan-Bhargava (MSB) test statistics for each individual i is constructed as,
MSBi =T−2
∑Tt=1 e
2i,t−1
σ2i
(D.24)
Where σ2i is a consistent estimator of the long-run variance of ei,t − ρiei,t−1.
The authors show that with level shifts, the limiting distribution for individual statistics
is invariant to structural breaks and argue that there is no need to estimate dates of level
shifting.2 With slope changes, the method first estimates the dates of structural breaks
for each individual unit equation-by-equation. Then the vector of common factors and
coefficients in the equation (25) are obtained following an iteration procedure.
Constructed following equation (27), individual MSB statistics are shown to be cross-
sectionally independent. To increase the power of unit root test, the authors provide two
approaches to construct statistics for panel test. The first statistics of panel test calculates
the average of individual statistics
Z =√NMSB − ξ
ζ→ N(0, 1) (D.25)
Where MSB = N−1∑N
i=1MSBi, ξ = N−1∑N
i=1 ξi, ζ = N−1∑N
i=1 ζi, with ξi and ζi denote
the mean and the variance of individual MSB statistics respectively, given by ξi = 12 and
ζ2i = 1
3 . The second approach pools individual p-values from MSBi.
P = −2N∑i=1
lnpi → χ22N (D.26)
or for large panels,
Pm =−2∑N
i=1 lnpi − 2N√4N
→ N(0, 1) (D.27)
Based on Monte Carlo simulation, Bai and Carrion-i Silvestre (2009) shows that all statistics
have non-trivial power when the autoregressive parameter is close to the null hypothesis,
and P and Pm statistics have empirical size close to the nominal level.
2However, the estimation of structural breaks might help to improve finite sample property.
96
Bibliography
Allen, D. and Yap, G. (2009). Modelling australian domestic tourism demand: A panel data
approach. School of Accounting, Finance and Economics & FEMARC Working Paper
Series No. 0910.
Allen, D., Yap, G., and Shareef, R. (2009). Modelling interstate tourism demand in aus-
tralia: A cointegration approach. Mathematics and Computers in Simulation, 79(9):2733–
2740.
Arellano, M. and Bond, S. (1991). Some tests of specification for panel data: Monte carlo
evidence and an application to employment equations. The Review of Economic Studies,
58(2):277–297.
Artis, M. and Hoffmann, M. (2012). The home bias, capital income flows and improved long-
term consumption risk sharing between industrialized countries. International Finance,
14(3):481–505.
Artis, M. J. and Hoffmann, M. (2008). Financial globalization, international business cycles
and consumption risk sharing. The Scandinavian Journal of Economics, 110(3):447–471.
Asdrubali, P., Sorensen, B. E., and Yosha, O. (1996). Channels of interstate risk sharing:
United States 1963-1990. The Quarterly Journal of Economics, 111(4):1081–1110.
Aslan, A., Kaplan, M., and Kula, F. (2009). International tourism demand for Turkey: A
dynamic panel data approach. Research Journal of International Studies, (Issue 9):65–73.
Backus, D. K., Kehoe, P. J., and Kydland, F. E. (1992). International real business cycles.
Journal of Political Economy, 100(4):745–775.
Bai, J. and Carrion-i Silvestre, J. L. (2009). Structural changes, common stochastic trends,
and unit roots in panel data. The Review of Economic Studies, 76(2):471–501.
Bai, J., Kao, C., and Ng, S. (2009). Panel cointegration with global stochastic trends.
Journal of Econometrics, 149(1):82–99.
Bai, J. and Ng, S. (2002). Determining the number of factors in approximate factor models.
Econometrica, 70(1):191–221.
97
Bai, J. and Ng, S. (2004). A panic attack on unit roots and cointegration. Econometrica,
72(4):1127–1177.
Bai, Y. and Zhang, J. (2012). Financial integration and international risk sharing. Journal
of International Economics, 86(1):17–32.
Baltagi, B. and Moscone, F. (2010). Health care expenditure and income in the oecd
reconsidered: Evidence from panel data. Economic Modelling, 27(4):804–811.
Banerjee, A., Marcellino, M., and Osbat, C. (2004). Some cautions on the use of panel meth-
ods for integrated series of macroeconomic data. The Econometrics Journal, 7(2):322–340.
Beck, G., Hubrich, K., and Marcellino, M. (2009). Regional inflation dynamics within and
across euro area countries and a comparison with the united states. Economic Policy,
24(57):141–184.
Becker, S. O. and Hoffmann, M. (2006). Intra- and international risk-sharing in the short
run and the long run. European Economic Review, 50(3):777–806.
Bonham, C., Gangnes, B., and Zhou, T. (2009). Modeling tourism: A fully identified vecm
approach. International Journal of Forecasting, 25(3):531–549. Special Section: Time
Series Monitoring.
Bonham, C. S. and Gangnes, B. (1996). Intervention analysis with cointegrated time series:
the case of the hawaii hotel room tax. Applied Economics, 28(10):1281–1293.
Borenstein, S., Cameron, A. C., and Gilbert, R. (1997). Do gasoline prices respond asym-
metrically to crude oil price changes? The Quarterly Journal of Economics, 112(1):305–
339.
Breitung, J. and Pesaran, M. (2008). Unit roots and cointegration in panels. The Econo-
metrics of Panel Data, 46:279–322.
Brida, J. and Risso, W. (2009). A dynamic panel data study of the german demand for
tourism in south tyrol. Tourism and Hospitality Research, 9(4):305–313.
Browning, M. and Collado, M. D. (2001). The response of expenditures to anticipated
income changes: Panel data estimates. American Economic Review, 91(3):681–692.
Carrion-i Silvestre, J. L., Kim, D., and Perron, P. (2009). Gls-based unit root tests with mul-
tiple structural breaks under both the null and the alternative hypotheses. Econometric
Theory, 25(6):1754–1792.
Chudik, A., Pesaran, M. H., and Tosetti, E. (2011). Weak and strong cross-section depen-
dence and estimation of large panels. The Econometrics Journal, 14(1):C45–C90.
98
Coakley, J., Fuertes, A.-M., and Smith, R. (2001). Small sample properties of panel time-
series estimators with I(1) errors. Computing in Economics and Finance 2001 191, Society
for Computational Economics.
Cochrane, J. H. (1991). [pitfalls and opportunities: What macroeconomists should know
about unit roots]: Comment. NBER Macroeconomics Annual, 6:201–210.
Crouch, G. (1995). A meta-analysis of tourism demand. Annals of Tourism Research,
22(1):103–118.
Crouch, G. (1996). Demand elasticities in international marketing: a meta-analytical ap-
plication to tourism. Journal of Business Research, 36(2):117–136.
Crucini, M. J. (1999). On international and national dimensions of risk sharing. The Review
of Economics and Statistics, 81(1):73–84.
Dreher, A. and Krieger, T. (2005). Do gasoline prices converge in a unified europe with
non-harmonized tax rates? KOF Working Paper No. 114.
Elliott, G., Rothenberg, T. J., and Stock, J. H. (1996). Efficient tests for an autoregressive
unit root. Econometrica, 64(4):813–836.
Engle, R. F. and Granger, C. W. J. (1987). Co-integration and error correction: Represen-
tation, estimation, and testing. Econometrica, 55(2):251–276.
Eugenio-Martin, J. L. and Campos-Soria, J. A. (2011). Income and the substitution pattern
between domestic and international tourism demand. Applied Economics, 43(20):2519–
2531.
Falk, M. (2010). A dynamic panel data analysis of snow depth and winter tourism. Tourism
Management, 31(6):912–924.
Franses, P. and Haldrup, N. (1994). The effects of additive outliers on tests for unit roots
and cointegration. Journal of Business & Economic Statistics, 12(4):471–478.
Fratzscher, M. and Imbs, J. (2009). Risk sharing, finance, and institutions in international
portfolios. Journal of Financial Economics, 94(3):428–447.
Garın Munoz, T. (2007). German demand for tourism in spain. Tourism Management,
28(1):12–22.
Garın-Munoz, T. and Montero-Martın, L. (2007). Tourism in the balearic islands: A
dynamic model for international demand using panel data. Tourism Management,
28(5):1224–1235.
Gengenbach, C., Palm, F., and Urbain, J. (2006). Cointegration testing in panels with
common factors. Oxford Bulletin of Economics and Statistics, 68(s1):683–719.
99
Gengenbach, C., Palm, F., and Urbain, J. (2010). Panel unit root tests in the presence of
cross-sectional dependencies: Comparison and implications for modelling. Econometric
Reviews, 29(2):111–145.
Habibi, F., Rahim, K., Ramchandran, S., and Chin, L. (2009). Dynamic model for interna-
tional tourism demand for malaysia: Panel data evidence. International Research Journal
of Finance and Economics, 33:207–217.
Hendry, D. F. (1995). Dynamic Econometrics. Oxford University Press.
Heston, A., Summers, R., and Aten, B. (2012). Penn world table version 7.1. Center
for International Comparisons of Production, Income and Prices at the University of
Pennsylvania.
Holly, S., Pesaran, M., and Yamagata, T. (2010). A spatio-temporal model of house prices
in the usa. Journal of Econometrics, 158(1):160–173.
Holmes, M. J., Otero, J., and Panagiotidis, T. (2013). On the dynamics of gasoline market
integration in the united states: Evidence from a pair-wise approach. Energy Economics,
36:503–510.
Im, K., Pesaran, M., and Shin, Y. (2003). Testing for unit roots in heterogeneous panels.
Journal of econometrics, 115(1):53–74.
Iyer, G. and Seetharaman, P. (2008). Too close to be similar: Product and price competition
in retail gasoline markets. Quantitative Marketing and Economics, 6(3):205–234.
Johansen, S. (1988). Statistical analysis of cointegration vectors. Journal of economic
dynamics and control, 12(2):231–254.
Kapetanios, G., Pesaran, M. H., and Yamagata, T. (2011). Panels with non-stationary
multifactor error structures. Journal of Econometrics, 160(2):326–348.
Kim, D. and Perron, P. (2009). Unit root tests allowing for a break in the trend func-
tion at an unknown time under both the null and alternative hypotheses. Journal of
Econometrics, 148(1):1–13.
Kose, M. A., Prasad, E. S., and Terrones, M. E. (2008). Does financial globalization promote
risk sharing? Journal of Development Economics, 89(2):258–270.
Kulendran, N. and Wilson, K. (2000). Modelling business travel. Tourism Economics,
6(1):47–59.
Kulendran, N. and Witt, S. F. (2003). Leading indicator tourism forecasts. Tourism Man-
agement, 24(5):503–510.
100
Kuo, H.-I., Chang, C.-L., Huang, B.-W., Chen, C.-C., and McAleer, M. (2009). Estimating
the impact of avian flu on international tourism demand using panel data. Tourism
Economics, 15(3):501–511.
Lane, P. R. and Milesi-Ferretti, G. M. (2007). The external wealth of nations mark ii:
Revised and extended estimates of foreign assets and liabilities, 1970–2004. Journal of
international Economics, 73(2):223–250.
Ledesma-Rodriguez, F. J., Navarro-Ibanez, M., and Perez-Rodriguez, J. V. (2001). Panel
data and tourism: a case study of tenerife. Tourism Economics, 7(1):75–88.
Leibrecht, M. and Scharler, J. (2008). Reconsidering consumption risk sharing among
oecd countries: some evidence based on panel cointegration. Open Economies Review,
19(4):493–505.
Levin, A., Lin, C., and Chu, J. (2002). Unit root tests in panel data: asymptotic and
finite-sample properties. Journal of econometrics, 108(1):1–24.
Lewis, K. K. (1997). Are countries with official international restrictions ’liquidity con-
strained’? European Economic Review, 41(6):1079–1109.
Lewis, M. S. (2009). Temporary wholesale gasoline price spikes have long-lasting retail
effects: The aftermath of hurricane rita. Journal of Law and Economics, 52(3):581–605.
Lewis, M. S. (2011). Asymmetric price adjustment and consumer search: An examination of
the retail gasoline market. Journal of Economics & Management Strategy, 20(2):409–449.
Leybourne, S. J., C Mills, T., and Newbold, P. (1998). Spurious rejections by dickey–fuller
tests in the presence of a break under the null. Journal of Econometrics, 87(1):191–203.
Leybourne, S. J. and Newbold, P. (2000). Behaviour of the standard and symmetric dickey–
fuller-type tests when there is a break under the null hypothesis. The Econometrics
Journal, 3(1):1–15.
Li, G., Song, H., and Witt, S. (2005). Recent developments in econometric modeling and
forecasting. Journal of Travel Research, 44(1):82–99.
Lim, C. and McAleer, M. (2001). Cointegration analysis of quarterly tourism demand by
hong kong and singapore for australia. Applied Economics, 33(12):1599–1619.
Lucas, R. E. (1987). Models of business cycles, volume 26. Basil Blackwell Oxford.
Mace, B. J. (1991). Full insurance in the presence of aggregate uncertainty. Journal of
Political Economy, 99(5):928–56.
Maddala, G. S. and Wu, S. (1999). A comparative study of unit root tests with panel data
and a new simple test. Oxford Bulletin of Economics and statistics, 61(S1):631–652.
101
Morley, C. (1998). A dynamic international demand model. Annals of Tourism Research,
25(1):70–84.
Muehlegger, E. (2006). Gasoline price spikes and regional gasoline content regulation: A
structural approach. mimeo, JFK School, Harvard University.
Naude, W. and Saayman, A. (2005). Determinants of tourist arrivals in africa: a panel data
regression analysis. Tourism Economics, 11(3):365–391.
Nelson, L., Dickey, D., and Smith, J. (2011). Estimating time series and cross section
tourism demand models: Mainland united states to hawaii data. Tourism Management,
32(1):28–38.
Newey, W. K. and West, K. D. (1987). A simple, positive semi-definite, heteroskedasticity
and autocorrelation consistent covariance matrix. Econometrica, 55(3):703–708.
Newey, W. K. and West, K. D. (1994). Automatic lag selection in covariance matrix esti-
mation. The Review of Economic Studies, 61(4):631–653.
Nicolau, J. L. and Mas, F. J. (2005). Stochastic modeling: a three-stage tourist choice
process. Annals of Tourism Research, 32(1):49–69.
Obstfeld, M. (1994). Are industrial-country consumption risks globally diversified? NBER
Working Papers 4308, National Bureau of Economic Research, Inc.
O’Connell, P. G. and Wei, S.-J. (2002). The bigger they are, the harder they fall: Retail
price differences across us cities. Journal of International Economics, 56(1):21–53.
Paul, R. J., Miljkovic, D., and Ipe, V. (2001). Market integration in us gasoline markets.
Applied Economics, 33(10):1335–1340.
Pedroni, P. (1999). Critical values for cointegration tests in heterogeneous panels with
multiple regressors. Oxford Bulletin of Economics and Statistics, 61(S1):653–670.
Pedroni, P. (2001). Purchasing power parity tests in cointegrated panels. Review of Eco-
nomics and Statistics, 83(4):727–731.
Pedroni, P. (2004). Panel cointegration: asymptotic and finite sample properties of pooled
time series tests with an application to the ppp hypothesis. Econometric Theory,
20(3):597–625.
Perron, P. (1989). The great crash, the oil price shock, and the unit root hypothesis.
Econometrica: Journal of the Econometric Society, 57(6):1361–1401.
Perron, P. and Rodrıguez, G. (2003). Searching for additive outliers in nonstationary time
series. Journal of Time Series Analysis, 24(2):193–220.
102
Perron, P. and Yabu, T. (2009). Testing for shifts in trend with an integrated or stationary
noise component. Journal of Business & Economic Statistics, 27(3):369–396.
Pesaran, M. and Smith, R. (1995). Estimating long-run relationships from dynamic hetero-
geneous panels. Journal of Econometrics, 68(1):79–113.
Pesaran, M. and Yamagata, T. (2008). Testing slope homogeneity in large panels. Journal
of Econometrics, 142(1):50–93.
Pesaran, M. H. (2004). General diagnostic tests for cross section dependence in panels.
CESifo Working Paper Series No. 1229; IZA Discussion Paper No. 1240.
Pesaran, M. H. (2006). Estimation and inference in large heterogeneous panels with a
multifactor error structure. Econometrica, 74(4):967–1012.
Pesaran, M. H. (2007). A simple panel unit root test in the presence of cross-section
dependence. Journal of Applied Econometrics, 22(2):265–312.
Pesaran, M. H. and Tosetti, E. (2011). Large panels with common factors and spatial
correlation. Journal of Econometrics, 161(2):182–202.
Phillips, P. and Sul, D. (2003). Dynamic panel estimation and homogeneity testing under
cross section dependence. The Econometrics Journal, 6(1):217–259.
Pierucci, E. and Ventura, L. (2010). Risk sharing: a long run issue? Open Economies
Review, 21(5):705–730.
Seetanah, B., Durbarry, R., and Ragodoo, J. (2010). Using the panel cointegration approach
to analyse the determinants of tourism demand in south africa. Tourism Economics,
16(3):715–729.
Seetaram, N. and Petit, S. (2012). Panel data analysis. In Dwyer, L., Gill, A., and Seetaram,
N., editors, Handbook of Research Methods in Tourism: Quantitative and Qualitative
Approaches, pages 127–144. Edward Elgar Publishing.
Song, H., Dwyer, L., Li, G., and Cao, Z. (2012). Tourism economics research: A review and
assessment. Annals of Tourism Research, 39(3):1653–1682.
Song, H. and Li, G. (2008). Tourism demand modelling and forecasting–A review of recent
research. Tourism Management, 29(2):203–220.
Song, H. and Witt, S. F. (2006). Forecasting international tourist flows to macau. Tourism
Management, 27(2):214–224.
Song, H., Wong, K. K., and Chon, K. K. (2003). Modelling and forecasting the demand for
hong kong tourism. International Journal of Hospitality Management, 22(4):435–451.
103
Sorensen, B. E. and Yosha, O. (1998). International risk sharing and european monetary
unification. Journal of International Economics, 45(2):211–238.
Stigler, G. J. and Sherwin, R. A. (1985). The extent of the market. Journal of Law and
Economics, 28(3):555–585.
Strotz, R. H. (1957). The empirical implications of a utility tree. Econometrica: Journal
of the Econometric Society, 25(2):269–280.
Sul, D. (2009). Panel unit root tests under cross section dependence with recursive mean
adjustment. Economics Letters, 105(1):123–126.
Suvankulov, F., Lau, M. C. K., and Ogucu, F. (2012). Price regulation and relative price
convergence: Evidence from the retail gasoline market in Canada. Energy Policy, 40:325–
334.
Syriopoulos, T. C. and Thea Sinclair, M. (1993). An econometric study of tourism de-
mand: the aids model of us and european tourism in mediterranean countries. Applied
Economics, 25(12):1541–1552.
Temple, B. (1988). Gasoline Marketing in the 1980s: Structure, Practices, and Public
Policy. American Petroleum Institute.
Verlinda, J. A. (2008). Do rockets rise faster and feathers fall slower in an atmosphere of
local market power? evidence from the retail gasoline market. The Journal of Industrial
Economics, 56(3):581–612.
Westerlund, J. and Urbain, J. (2011). Cross sectional averages or principal components?
Maastricht: METEOR, Maastricht Research School of Economics of Technology and
Organization.
Witt, S. and Witt, C. (1995). Forecasting tourism demand: A review of empirical research.
International Journal of Forecasting, 11(3):447–475.
Yap, G. and Allen, D. (2011). Investigating other leading indicators influencing australian
domestic tourism demand. Mathematics and Computers in Simulation, 81(7):1365–1374.
104