cover

Papers Factors Affecting an Economys Tolerance and Delay of Response to the Impact of a Positive Oil Price Shock Bwo-Nung Huang High Frequency Export and Price Responses in the Ontario Electricity Market Angelo Melino and Nash Peerbocus Efforts and Efficiency in Oil Exploration: A Vector Error-Correction Approach Klaus Mohn Threshold Cointegration Analysis of Crude Oil Benchmarks Shawkat M. Hammoudeh, Bradley T. Ewing and Mark A. Thompson An Empirical Analysis of Urban Form, Transport, and Global Warming Fabio Grazi, Jeroen C.J.M. van den Bergh and Jos N. van Ommeren Economies of Scale and Scope in Multi-Utilities Mehdi Farsi, Aurelio Fetz and Massimo Filippini Identifying the Rebound: Evidence from a German Household Panel Manuel Frondel, Jrg Peters, and Colin Vance Book Reviews Electricity and Energy Policy in Britain, France and the United States since 1945, by Martin Chick Richard L. Gordon The Strategic Petroleum Reserve: U. S. Energy Security and Oil Politics, 1975-2005, by Bruce A. Beaubouef Peter Van Doren A Handbook of Primary Commodities in the Global Economy, by Marian Radetzki Roberto F. Aguilera

Referee Acknowledgments Volume 29, Number 4 International Association for Energy Economics

Index to Volume 29 2008

President President-elect Immediate Past President Past President Vice President and Treasurer Vice President for Development & International Affairs Vice President for Publications Vice President and Secretary Vice President for Conferences Appointed Council Member Appointed Council Member Appointed Council Member Appointed Council Member Appointed Council Member Appointed Council Member

Carlo Andrea Bollino Georg Erdmann Andr Plourde Jean-Philippe Cueille Jan Myslivec Sophie Meritet Christian von Hirschhausen Yunchang J. Bor Lars Bergman Larry Chow Vincenzo De Bustis Reinhard Haas Einar Hope Wumi Iledare Mine Ycel

The Energy Journal is a publication of the Energy Economics Education Foundation, Inc., 28790 Chagrin Blvd., Suite 350, Cleveland, OH, 44122-4630, USA, a non-profit organization, in association with the International Association for Energy Economics. Membership dues for the IAEE include subscriptions to The Energy Journal and the IAEE Newsletter. Subscriber & Membership matters should be sent to IAEE, 28790 Chagrin Blvd., Suite 350, Cleveland, OH, 44122, USA; phone 216-464-5365, fax 464-2737. Non-member subscriptions to The Energy Journal are $425 for institutions, libraries, and individuals. Postage and handling are paid by the publisher. Outside the United States and Canada, non-member subscriptions are $450. Articles appearing in The Energy Journal are listed in both Envir. Per. Bibliography and The Journal of Economic Literature and are indexed/abstracted in ABI/Inform, EconLit, JEL on CD, ISI Journal Citation Reports, Web of Science, SciSearch, Research Alert, Ei COMPENDEX, Trade and Industry Index, PAIS International, DOE Energy, Energy/Enviroline, and Wilson Business Abstracts, and the following EBSCO databases: Academic Search Premier and Business Source Premier. Authorization to photocopy articles appearing in The Energy Journal for internal or personal use, or the internal or personal use of specific clients, is granted by IAEE provided that you register with the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923, phone 978-750-8400. The Energy Journal is available in reference systems (e.g., microfilm, CD-ROM, full text etc.) produced by the ProQuest Co. (ProQuest Information & Learning), the Gale Group (Info Trac databases) and the H.W. Wilson Companys Business Periodicals Index. For further information please contact ProQuest Company at 734-761-4700, www.proquestcompany.com, the Gale Group at 248-699-4253, www.galegroup.com, or H.W. Wilson Co., at 718-588-8400, www.hwwilson.com. AUTHORS: Please see back pages for information on the proper style of submissions. ADVERTISING INFORMATION: For a complete media kit, contact Marketing Department, IAEE Headquarters, 28790 Chagrin Blvd., Ste. 350, Cleveland, OH, 44122-4630. Phone: 216-464-5365, fax: 216-464-2737. E-mail: iaee@iaee.org. Homepage: http://www.iaee.org. Copyright 2008 by the International Association for Energy Economics. All rights reserved. The editors and publisher assume no responsibility for the views expressed by the authors of articles printed in The Energy Journal. International Standard Serial Number: 0195-6574. Printed in the U.S.A

Editor-in-Chief, Adonis Yatchew Editors, Lester C. Hunt and James L. Smith European Editor, Yves Smeers Associate Editor, Geoffrey Pearce Book Review Editor, Richard L. Gordon Assistant Book Review Editor, Carol Dahl Board of Editors
Lars Bergman Stockholm School of Economics Carol A. Dahl Colorado School of Mines, Golden, CO Joy C. Dunkerley Consulting Economist Washington, D.C. A. Denny Ellerman MIT - Center for Energy and Environmental Policy Research Massimo Filippini University of Lugano ETH Zurich CEPE Richard L. Gordon Pennsylvania State University James M. Griffin Texas A & M University College Station, Texas William W. Hogan Harvard University John F. Kennedy School of Government Einar Hope Norwegian School of Economics and Business Administration Mark K. Jaccard Simon Fraser University British Columbia, Canada Edward Kahn Analysis Group, Inc. San Francisco, CA J. Daniel Khazzoom College of Business San Jose State University Kenichi Matsui Institute of Energy Economics, Tokyo Mohan Munasinghe Energy Advisor Colombo, Sri Lanka Rajendra K. Pachauri Tata Energy Research Institute, New Delhi Francisco R. Parra Consultant F.R. Parra Inc. Peter J.G. Pearson Imperial College of Science, Technology and Medicine, London Andr Plourde University of Alberta Alberta, Canada Hossein Razavi The World Bank Washington, D.C. Ali M. Reza College of Business San Jose State University Geoffrey Rothwell Stanford University USA Dieter Schmitt Universitt Essen Germany Margaret F. Slade The University of Warwick Coventry, UK Thomas Sterner Gteborg University Sweden Catherine Waddams University of East Anglia United Kingdom W. David Walls University of Calgary Alberta, Canada Leonard Waverman University of Calgary Alberta, Canada John P. Weyant Stanford University, CA Franz Wirl University of Vienna, Department of Business Studies, Austria Catherine Wolfram Haas School of Business University of California Berkeley, CA Chi-Keung Woo Energy and Environmental Economics, Inc., San Francisco, CA

ii

THE ENERGY JOURNAL Volume 29, Number 4, 2008 Table of Contents

Papers Factors Affecting an Economys Tolerance and Delay of Response to the Impact of a Positive Oil Price Shock Bwo-Nung Huang High Frequency Export and Price Responses in the Ontario Electricity Market Angelo Melino and Nash Peerbocus Efforts and Efficiency in Oil Exploration: A Vector Error-Correction Approach Klaus Mohn Threshold Cointegration Analysis of Crude Oil Benchmarks Shawkat M. Hammoudeh, Bradley T. Ewing and Mark A. Thompson An Empirical Analysis of Urban Form, Transport, and Global Warming Fabio Grazi, Jeroen C.J.M. van den Bergh and Jos N. van Ommeren 1 35 53 79 97

Economies of Scale and Scope in Multi-Utilities 123 Mehdi Farsi, Aurelio Fetz and Massimo Filippini Identifying the Rebound: Evidence from a German Household Panel 145 Manuel Frondel, Jrg Peters, and Colin Vance

Book Reviews Electricity and Energy Policy in Britain, France and the United States 165 since 1945, by Martin Chick Richard L. Gordon The Strategic Petroleum Reserve: U. S. Energy Security and 167 Oil Politics, 1975-2005, by Bruce A. Beaubouef Peter Van Doren A Handbook of Primary Commodities in the Global Economy, 168 by Marian Radetzki Roberto F. Aguilera

Referee Acknowledgments Index to Volume 29 iii

171 185

iv

Factors Affecting an Economys Tolerance and Delay of Response to the Impact of a Positive Oil Price Shock
Bwo-Nung Huang* This paper applies a multivariate threshold model to estimate a countrys threshold level of economic tolerance (c) and delay of response (d) to the impact of a positive price change and its shock. Regression analysis is employed to investigate the factors affecting c and d. We find: (1) as a country becomes more advanced in economic development and acquires a lower ratio of energy use in its industry and transportation sectors, the threshold of tolerance is greater as evidenced by the positive impact of an oil price change and its shock; (2) if a country has a lower ratio of energy use in the industry sector, a lower energy import ratio and is more advanced in economic development, it will have a longer delay; and (3) as an economy becomes more advanced, the length of the response time from the impact of the shock of an oil price change will be longer. 1. Introduction Since the global energy crises of the 1970s and 1980s and their effects on the world economy, the impact of an oil price change and its shock on economic activities have been a focus of research over the past three decades. Following the 1992 Gulf War, oil prices remained relatively stable and discussions on the subject were not as frequent. More recently, because of the increased demand for oil from developing countries, China and India in particular, and the uncertainty in the Middle East fueled by the tensions arising because of Irans alleged nuclear ambitions, the oil price began to rise again in early 2004. By March 13, 2008, the oil price had spiked to a historical high of $110.21 per barrel (WTI spot). With
The Energy Journal, Vol. 29, No. 4. Copyright 2008 by the IAEE. All rights reserved. * Department of Economics, West Virginia University and Department of Economics and Center for IADF, National Chung Cheng University, Chia-yi 621, Taiwan. Email: ecdbnh@ccu.edu.tw.

Financial support from the National Science Council (NSC94-2415-H-194-002) is gratefully acknowledged. The valuable comments from the editor and four referees are greatly appreciated. Any errors are my own. This paper was completed during my visit to West Virginia University during the 2006 school year. In particular, I would like to thank the Department of Economics at WVU for the many resources it provided and Professor Ming-Jeng Hwang for his helpful suggestions.

2 / The Energy Journal oil prices increasing, economists may refocus their attention on the issue of an oil price change and its impact on economic activities. While there are related studies on the use of the nonlinear model to investigate the impact of an oil price change on the economy, they do not take into consideration the difference in tolerance and the speed of adjustment due to an oil price change resulting from the differences in economic development, energy dependence, and the efficiency of energy use in each country.1 Is there a degree of economic tolerance which shields the economy from the impact of an oil price change or its shock? 2 If there is, what are the factors affecting the degree of economic tolerance? The major contribution of our paper is that it uses Tsays MVTAR (Multivariate Vector Threshold Autoregressive) model to estimate the threshold value (c) regarding the impact of an oil price change or its shock, and the length of response (d) of the threshold variable in each of the 21 countries.3 We then employ the multiple regression model to test and identify the relationship between energy variables and c or d in each country. The monthly data used are the real stock price (rstkpt), industrial production (yt), the real oil price (roilpt), and the interest rate (rt). We find: (1) as a country becomes more advanced in economic development (real per capita GDP is used to denote the degree of economic development) and acquires a lower ratio of energy use in its industry and transportation sectors, the threshold of tolerance is greater as evidence by the positive impact of an oil price change and its shock; (2) if a country has a lower ratio of energy use in the industry sector, a lower energy import ratio and is more advanced in economic development, it will have a longer delay in terms of its economic response from the positive impact of an oil price change; and (3) as an economy becomes more advanced, the length of response time from the impact of the shock of an oil price change will be longer. Such discoveries will help policy planners in different countries design their own energy policy based on the degree of economic tolerance and the speed of response from the impact of an oil price shock. This paper is organized as follows. Section I states the motivation for this research. Section II reviews the literature. Section III introduces the empirical models and related procedures. Section IV describes the data, and tests for the unit root and cointegration. In Section V, the MVTAR model is used to estimate the
1. The important literature on using the nonlinear model to investigate the impact of an oil price change on the macroeconomy includes Mork (1989), Lee et al. (1995), Hamilton (1996, 2003) and Jimnez-Rodriguez and Snchez (2005). By separating a price increase from a price decrease or a positive price change (positive shock) from a negative price change (negative shock), they then explore the impact of both a positive shock and a negative shock on aggregate economic activities. However, no consideration is given as to the differences in response to the oil price shocks based on the differences in economic development, energy dependability, and the efficiency of energy use in each country. 2. Following Lee et al. (1995), standardized residuals estimated from the GARCH model of oil price changes are used to represent oil price shocks where the oil price change is defined as percentage changes of the oil price. 3. We define the value of c as a measure of the threshold of tolerance, with a larger value of c indicating that it requires a larger oil price increase to have an effect, while the value of d measures the length of delay of the economic response, with a larger value of d indicating a longer delay.

Factors Affecting an Economys Tolerance / 3 threshold values for each country. Section VI analyzes the estimated results based on threshold levels (c) and the delay (d) in each country. Section VII concludes the analysis. 2. Literature Review Prior research on the relationship between an oil price change and economic activities can be traced back to the pioneering work of Hamilton (1983),4 who was the first to indicate that oil price increases partly accounted for every U.S. recession after World War II (except for one in 1960). Since then, using alternative estimation procedures on new data such as the U.S., the U.K. and Japan, researchers have extended and largely reinforced Hamiltons results (Gisser and Goodwin, 1986; Burbidge and Harrison, 1983). Gradually, attention has shifted to the relationship between an oil price change and GDP. Gilbert and Mork (1986) and Mork (1989) first reasoned that the impact of an oil price increase or decrease on production may be nonlinear. An oil price change may increase the cost of resource allocation regardless of the direction of the price change. When the oil price is decreasing, the decrease in the cost of production and the increase in the cost of resource allocation often offset each other. With the increase in the oil price, these two forces reinforce each other and the corresponding impact on GDP can be significant. Hence, the oil price change should have an asymmetrical effect on the economy. By separating oil price changes into price increases and price decreases, Mork (1989) found that there exists an asymmetrical relationship between an oil price change and real output. That is, there is a negative relationship between an oil price increase and GDP. Lee et al. (1995) and Hamilton (1996) investigated the impact of an oil price change on aggregate economic activities after first going through certain nonlinear transformations of oil price data. They discovered that, when the oil price is raised, there is a negative relationship between the oil price and real activity in the U.S. economy. On the other hand, when the oil price falls, there is an insignificant a (2003) investigated relationship between the two. Cuado and Prez de Graci the asymmetric relationship between an oil price change and real activity in 14 European countries by following three indexes often used in the literature. They discovered that, if the problem of asymmetry is not considered, seven out of the fourteen countries will indicate that an oil price change does not lead (negatively) industrial production. When a positive oil price change is considered, only three countries do not indicate that an oil price change leads (negatively) industrial production. By using the same indexes with the multivariate VAR model to study the relationship between an oil price change and aggregate economic activities in eight OECD countries, Jimnez-Rodriguez and Snchez (2005) found that the impact of an oil price increase on economic growth is far greater than the impact of an oil price decrease. In addition, they discovered that an oil price rise in five oil
4. Brown and Ycel (2002) provide an excellent related survey on the relationship between an oil price change and aggregate economic activity.

4 / The Energy Journal importing countries (excluding Japan) has a negative impact on economic growth. In two oil exporting countries out of the eight OECD countries, an oil price increase was found to have a different impact. In the U.K., an oil price rise was found to have a negative impact on GDP, but Norway was found to have a positive impact between the two. A few economists have claimed in the literature that an economic recession does not originate from an oil price increase, but from the governments tight monetary policy. Bernanke, Gertler, and Watson (1997, BGW) employed the VAR model, using different policy options, to show that the decrease in GDP from the oil price increases of 1973, 1979-1980, and 1990 was caused by monetary policy, and not by the oil price increase itself. The results of BGW (1997), of course, gave rise to further discussions. Subsequently, Hamilton and Herrera (2004, H&H) employed the same data as BGW and the same estimating procedures, but used two different assumptions and came to a different conclusion. H&H first used a lag of 12 periods, rather than seven periods as used by BGW. Second, BGW assumed that the Federal Reserves increasing of the federal funds rate by 900 (basis points) was deemed impractical. As a result, H&H came to support the negative correlation between the oil price increase and GDP. Balke, Brown, and Ycel (2002) used a model similar to the BGW one by adding certain variables and allowed for a nonlinear impact relationship. They found that oil price changes may affect GDP even without any change in monetary policy. In addition, an asymmetrical relationship was found between an oil price change and GDP or a short-term interest rate. Hooker (1999) added an interest rate variable in a multivariate VAR model to investigate the role of monetary policy in the correlation between an oil price change and GDP growth. He indicated that the impact of an oil price change on GDP is significantly reduced. He therefore concluded that an oil price increase indirectly affects GDP through monetary policy. Tatom (1988, 1993) found that the asymmetric response of economic activity to oil shocks disappears when the monetary policy or changes in the misery index (which combines unemployment and inflation rates) are taken into account. On the other hand, Brown and Ycel (2002) also showed that the Federal Reserves response to oil price shocks is not the cause of the asymmetry. They found that the asymmetry does not go away when both the federal funds rate is held constant or the federal funds rate and expectations of the federal funds rate are held constant. Hence, monetary policy does not appear to be the sole cause of asymmetry on the real side. Besides discussing the relationship between an oil price change and GDP, some economists have emphasized the relationship between an oil price change and other macro-economic variables. Davis and Haltiwanger (2001) used plant-level census quarterly data to investigate the relationship between an oil price change and job creation or job loss. Keane and Prasad (1996) used the individual data of the U.S. National Longitudinal Survey of Young Men to investigate the relationship between an oil price change and the real wage rate. They found that an oil price increase causes the overall wage rate to decline but the wage rate

Factors Affecting an Economys Tolerance / 5 of skilled workers to increase. Carruth, Hooker, and Oswald (1998) employed the error correction model (ECM) to test the relationship between an oil price change and unemployment. They found that the oil price has more of an impact on unemployment than the interest rate. The research on the impact of an oil price change on the capital market started relatively late compared to its impact on the labor market and product market. Jones and Kaul (1996) were the first to test the impact of an oil price change on the stock market using 1947-1991 data and discovered that the oil price did have an impact on the overall stock return. However, Huang et al. (1996) employed 1979-1990 daily future oil price data only to find no evidence of the impact of an oil price change on the overall stock return. Subsequently, Sadorsky (1999) used 1947-1996 data on U.S. monthly interest rates, the oil price, industrial production, and stock returns to construct a VAR model to investigate the impact of an oil price change on the overall stock return and industrial production. He found that the oil price change or its shock does have an impact on the real stock returns. In particular after 1986, the oil price, rather than the interest rate, better explains the predicted error of the stock return. Besides the variable for an oil price change, some researchers have used the variable of its volatility (uncertainty) to investigate its impact on certain macro-economic variables. Both Sadorsky (1999) and Kaul and Seyhum (1990) used the volatility of oil price changes to test its relationship with the stock returns. The literature on the subject is generally based on U.S. data. However, some researchers used data from different countries. For example, Mork et al. (1994) used data for seven OECD countries; Papapetrou (2001) used data for Greece; Cuado a (2003) used data from 14 European countries; de Miguel, and Prez de Graci Manzano, and Martin-Moreno (2003) used Spanish data; Jimnez-Rodriguez and Snchez (2005) used data for eight OECD countries; and Kilian (2008) employed G7 data. Some studies not only investigated the relationship between the oil price and aggregate economic activity, but also looked at the extent of an economic response from an oil price change (e.g., Mork et al., 1994; Cuado and Prez de a, 2003; Jimnez-Rodriguez and Snchez, 2005 and Kilian, 2008). Graci There is an abundant literature on the nonlinear impact of an oil price shock on economic activity. However, due to the differences in the degree of economic development, energy dependence, and the efficiency of energy use, the economic tolerance and the speed of economic response in each country as a result of the impact of a positive oil price change and its shock are expected to be different. One important issue lacking in the literature is that we are unable to calculate appropriate threshold values for the economic tolerance in relation to an oil price shock in each country. Huang et al. (2005) employed the multivariate threshold autoregressive model (MVTAR) of Tsay (1998) to find the threshold value of an oil price change and its shock in each country. They found: (1) the most appropriate threshold value seems to vary according to how an economy depends on imported oil and its attitude towards adopting energy-saving technology; (2) an oil price change or its shock has a limited impact on the economy if

6 / The Energy Journal the change is below the threshold levels; (3) if the change is above the threshold levels, it appears that the change in the oil price explains the macroeconomic variables better than the shock caused by the oil price; and (4) if the change is above the threshold levels, a change in the oil price or its shock explains the variation in GDP growth better than the real interest rate. Huang et al. (2005) have also taken into consideration the economic tolerance of the impact of an oil price change and its shock based on different natural endowments in each country. However, only three countries (the U.S., Canada, and Japan) are included in the study and no statistical tests are used to verify its validity. More countries and more refined statistical techniques are needed to test the relationship between the economic tolerance emanating from the impact of an oil price change (or its shock) and the degree of dependence on crude oil. 3. Empirical Models and Related Procedures Due to the differences in the degree of economic development, energy dependence, and the efficiency of energy use, the economic tolerance and the speed of economic response in each country from the impact of a positive oil price change and its shock are expected to be different. It is necessary to use the MVTAR model to first estimate the threshold of tolerance (c) and the length of delay (d) in each country. Then, we search for energy related or growth related variables in an attempt to explain how an oil price change and its shock have different impacts on the threshold of tolerance and length of delay in each country. The origin of the MVTAR model of Tsay (1998) dates back to the univariate threshold autoregressive model (TAR) of Tong (1978). The objective of a TAR model is to divide data into different regimes through a threshold variable qt-d, where d represents the delay periods of the threshold variable. At the beginning, the threshold variable uses the delay (d) period of the dependent variable; hence, the univariate TAR model is also referred to as SETAR (self-exciting TAR). The two-regime model or SETAR(1) can be expressed as yt = (0,1 + 1,1yt1)(1 I [qt1 > c]) + (0,2 + 1,2yt1) I [qt1 > c] + et, (1)

where et epresents an iid white-noise process and qt1 = yt1. I [.] is an index function, which equals 1 or I [.] = 1 if the relation in the brackets holds. I [.] equals zero otherwise. Equation (1) can be viewed as a multivariate threshold VAR(1) if yt is a vector with qt1 a scalar based on one of the elements of yt. If there is a cointegration relationship among the variables, the multivariate threshold error correction (MVTEC) model can be expressed as yt = (c1 + d1zt1 + i,1yti)(1 I [qtd > c]) + i =1 p (c2 + d2zt1 + i,2yti) I (qtd > c]) + et,
i =1 p

(2)

Factors Affecting an Economys Tolerance / 7 where E(e) = 0, E(ee ') = , yt is a (4x1) vector, and zt1 is an error correction term. If there is no cointegration among the variables (d1 = d2 = 0), equation (2) becomes the MVTAR model. In order to estimate the threshold of tolerance (c) and length of delay (d) from an oil price change or shock in each country, the estimation of the nonlinear relationship in (2) must include the restriction that the column of i,1 corresponding to the effect of oil prices is identically zero for i=1, , p.5 If only the oil price variable and production variable are used to investigate the impact of an oil price change on economic activity, misleading results may appear as an oil price change may affect other economic variables such as stock prices and interest rates. Since the capital market plays an increasingly important role in economic activity, it is necessary to include the stock price and interest rate variables to study the impact of an oil price change on aggregate economic activities as was already indicated in the literature (Jones and Kaul, 1996; Sadorsky, 1999 or Huang et al., 2005). Thus, the study of the response of aggregate economic activities from an oil price change should include in the vector yt items such as the relationships among an oil price change (lroilpt), an interest rate change (rt), a real stock price change (lrstkpt), and a production change (lyt). If the study seeks to investigate the impact of an oil price shock on aggregate economic activities, the standardized residual (vt) is used for the shock. Prior to the estimation of equation (2), we need to test if the nonlinear relation indeed exists. That is, the null hypothesis: H0 : i,1 = i,2, i = 1,2..., , c1 = c2, and d1 = d2 should be tested. The C(d) statistic based on the arranged regression (Tsay, 1998) can be used to test the H0: linear relation vs. HA: nonlinear relation. Note that C(d) is a random variable that follows an asymptotically chi-square distribution with k(pk+1) degrees of freedom in which p is the lag length of the MVTEC or MVTAR model and k is the number of variables in yt. In order to estimate the optimal threshold level c* for a given p and d, we arrange qt-d in ascending order to search for c* based on the following criterion (Weise, 1999): (c ) , (c*, d ) = arg min det i (3)

is the determinant value of the variance and covariance matrix of where det equation (2).6 Equation (2) is a MVTEC model with the presence of a cointegration relationship. Sometimes, if the data are not in long-run equilibrium (i.e., d1 = d2 = 0), equation (2) becomes a MVTAR model. In order to apply our data to either the MVTEC or the MVTAR model, we need to test the four variables for a unit root. These four variables are the log real oil price (lroilpt), interest rate (rt), log
5. The suggestion made by one of the referees for the restriction is greatly appreciated. 6. See Appendix A for the detail on how to calculate the C(d) statistic and how to find the optimal threshold level c*.

8 / The Energy Journal real stock price (lrstkpt), and log industrial production index (lyt). Of these four variables, if at least two variables are in the form of I(1), tests of the cointegration relationship are needed to determine the use of either the MVTEC or MVTAR model. To investigate the relationship between an oil price shock and other economic variables, the GARCH(1,1) AR(p) model (Lee, Ni and Ratti, 1995) is used to estimate the standardized residual (vt) as the proxy for an oil price shock and the related threshold levels of the oil price shock. The GARCH model of the real oil price change (Dlroilpt) is defined as Dlroilpt = b0 + Dlroilpt + et
i =1 p

et It1 ~ N (0,ht)

(4)

ht = a0 + a1et12 + a2 ht1, where It1 represents the information set at time t-1. The oil price shock vt is estimated through the standardized residual error (et1 / ht1 ) of equation (4). 4.  Data Sources and Results of Tests for Unit Roots and Cointegration The monthly data for 21 countries are used in this research. The oil price data (oilp) are obtained from the average crude price in the commodity price section of the IFS CD-ROM provided by International Monetary Fund. With the exception of Taiwan, the data for the other three variables are also obtained from the IFS CD-ROM. The money market rate (line 60b) is used as a proxy for the interest rate. The share price (line 62) is used to reflect the stock price and industrial production (line 66) is used to approximate output. The data for Taiwan are taken from AREMOS. Logarithmic transformations of industrial production, the real stock price, and real oil prices are all taken before conducting the analysis. The real oil price and real stock price are deflated by the base year 2000 consumer price index (cpi). The length of the available data in each country is different. Among these 21 countries, 6 countries (Austria, France, Japan, the Netherlands, South Africa and the U.S.) have the longest data (1970.1~2005.4 for 424 data points). Portugal has the shortest data (1988.1~2005.4 for 208 data points). The data for all 21 countries are available up to 2005.4 except for India (only up to 1998.5). 7 Table 1 displays the data periods for all 21 countries. Since the VAR or VEC (vector error correction) model is used to estimate the nonlinear relation, and the delay periods cannot be too short for the statistical analysis, at least 200 data points are needed for a delay of 12 periods as suggested by Hamilton and Herrera (2004). The availability of data for the industrial produc7. In order to collect all the necessary monthly data with the required length of time for the stock return, industrial production, the real oil price, and the interest rate, only 21 countries are found to meet these requirements in the International Financial Statistics (IFS) database. Many countries lack stock index data or industrial production index data, or the length of the stock index data is too short.

Factors Affecting an Economys Tolerance / 9 Table 1. Sample Countries and Data Periods
Country Austria Canada Denmark Finland France India Israel Italy Japan Korea Malaysia Mexico Netherlands Norway Philippines Portugal South Africa Spain United Kingdom USA Taiwan Abbreviation AT CA DK FI FR IN IL IT JP KR MY MX NL NO PH PT ZA ES GB US TW Data period 1970:01~2005:04 1975:01~2005:04 1974:01~2005:04 1977:12~2005:04 1970:01~2005:04 1970:01~1998:05 1984:06~2005:03 1971:01~2005:04 1970:01~2005:04 1978:01~2005:04 1980:01~2005:04 1984:01~2005:04 1970:01~2005:04 1971:08~2005:04 1981:01~2005:04 1988:01~2005:04 1970:01~2005:04 1974:01~2005:04 1972:01~2005:04 1970:01~2005:04 1981:01~2005:04 Number of Observations (T) 424 364 376 329 424 341 250 412 424 328 304 256 424 405 292 208 424 376 400 424 292

Note: ISO-3166 list of two-letter abbreviations for country code is used.

tion index, the interest rate, and stock price are also needed in choosing countries. After taking all these into consideration, only 21 countries are selected. Before the MVTEC or MVTAR model is formally employed in the statistical analysis, all the variables need to be tested for unit roots. Since the proxy for the oil price shock is estimated from the standardized residuals of equation (4), it needs to be I(0) and, as such, no unit root test is necessary. We use the DF-GLS of Elliott et al. (1996) and the MZd of Ng and Perron (2001) to test for the existence of a unit root for the time series data in each country. The two unit root test results for all 21 countries are displayed in Appendix Table 1. According to Appendix Table 1, among the 21 countries, 13 countries exhibit I(1) for all four variables, while eight countries exhibit I(1) for three variables. Overall, for eight countries with I(0) for at least one level variable (variable in its level form), the interest rate variable is not I(1) in such countries. Since we are investigating the impact of an oil price change (Model I) and the impact

10 / The Energy Journal of an oil price shock (Model II) on the interest rate, stock price, and industrial production, all the I(1) variables (lroilpt, rt, lrstkpt, lyt in Model I and rt, lrstkpt, lyt in Model II) need to be examined regarding the existence of a cointegration relation. We apply the trace method and maximum eigenvalue proposed by Johansen (1988) to test for the existence of a cointegration relation for these I(1) variables. Appendix Table 2 displays Johansens cointegration test results with the model types used. The third column of Appendix Table 2 lists the estimated trace statistics; the fourth column consists of the maximum eigenvalues and related statistics; the fifth column lists the variables of I(1) for which cointegration tests are needed; and the sixth column reports the model selected to estimate threshold levels according to the cointegration test results. Appendix Table 2-1 shows estimated results from Model I (lroilpt, rt, lrstkpt, lyt) and Appendix Table 2-2 displays estimated results from Model II (rt, lrstkpt, lyt). When the trace statistics and maximum eigenvalues reject H0 at less than the 5% significance level, there exist cointegration relations among variables. As shown in Appendix Table 2-1 under Model I, Israel, Mexico, Norway, the U.K. and Taiwan provide evidence of a cointegration relation and the VEC model is used for these five countries. The remaining 16 countries do not exhibit a cointegration relation and the VAR model is sufficient for the statistical analysis. Under Model II, Israel, Korea, Mexico, Spain, the U.K., and Taiwan show the existence of a cointegration relation and as such the VEC model is used, while the VAR model is used in the other 15 countries for the purposes of the analysis. 5.  Estimating the Threshold Levels and the Delay of Threshold Variables For countries with the presence of a cointegration relation as indicated in Appendix Table 2, we need to use the MVTEC model to analyze the data: yt = ( a1j dtj + d1zt1 + i,1yti)(1 I [qtd > c]) + j =1 i =1 p 12 ( a2j dtj + d2zt1 + i,2yti) I [qtd > c] + et,
j =1 i =1 12 p

(5)

where yt is (lroilpt, rt, lrstkpt, lyt) or (vt, rt, lrstkpt, lyt). dtj = 1 if observation t is characterized by month j and is zero otherwise. zt-1 denotes the cointegration vector among the I(1) variables, qt-d denotes the threshold variable with a delay of d periods, c is the threshold level, and t is assumed to follow an iid with N(0,1) distribution. For countries without a cointegration relation, the following MVTAR model is used: yt = ( a1j dtj + i,1yti)(1 I [qtd > c]) + j =1 i =1 p 12 ( a2j dtj + i,2yti) I [qtd > c] + et,
j =1 i =1 12 p

(6)

Factors Affecting an Economys Tolerance / 11 where yt is (lroilpt, rt, lrstkpt, lyt) or (vt, rt, lrstkpt, lyt). Before equation (5) or (6) is used for the estimation, the C(d) test proposed by Tsay (1998) for the testing of nonlinearity needs to be performed first. Note that C(d) is a random variable that follows an asymptotically chi-square distribution with k(pk+1) degrees of freedom in which p is the lag length of the VAR model and k is the number of variables in yt. We set p at 12 periods (one year) so that seasonal and dynamic relations can be detected by the data.8 With p = 12, the upper bound of the delay d for the threshold variable is also set at 12 periods. Since our focus is on the impact of an oil price change (percentage) and its shock (size) on economic activities, the threshold variable (qt-d) is chosen as lroilpt-d for Model I and vt-d for Model II, respectively. Appendix Table 3-1 (Model I) and 3-2 (Model II) display the test results of the C(d) statistics. Some statistical problems related to the testing of the nonlinear threshold model warrant further discussion. The multivariate threshold model proposed by Tsay (1998) did not take into consideration the cointegration relation because it gives rise to the complicated statistical testing problem of linearity vs. nonlinearity and cointegration vs. non-cointegration. At the present time, the two-step test approach adopted by Balke and Fomby (1997) is widely used. That is, under the linear assumption, we test for the existence of a cointegration relation first. If there is a cointegration relation, the pre-specified cointegration vector needs to be included in the model to test whether it is a linear or nonlinear relation (Lo and Zivot, 2001; Seo, 2006). Our test follows the two-step approach.9 Based on Appendix Tables 3-1 and 3-2, every country, under some delay of d periods, indicates the existence of the nonlinear relation as the impact of an oil price change or its shock reaches certain threshold levels. Using the estimation method proposed by Tsay (1998), the estimated results of a nonlinear relation based on the delay selected by the C(d) test statistics in each country under Models I and II are displayed in Tables 2-1 and 2-2. Since we assume that the threshold variable is lower than a certain level, an oil price change or shock will not bring about changes in the interest rate, stock price or production variables. Thus, Table 2 only displays Granger causality test results when the threshold variable is higher than a certain level for both the linear model and the nonlinear model.

8. We have used the AIC or SIC to select the optimal lag for each country. However, the estimated results from these criteria are often too short to handle the dynamic relation between an oil price change and aggregate economic activities well. By following the suggestion of Hamilton (1994) with monthly data, it is a good idea to include at least 12 lags in the regression (p. 583). In addition, Hamilton and Herrera (2004) also indicated that the impact of an oil price change on production may have a delay of three or four quarters. Therefore, we select a lag of p = 12 to estimate the MVTAR or MVTEC model. 9. Theoretically, the test involves four categories: (1) linear without a cointegration relation; (2) nonlinear without a cointegration relation; (3) linear with a cointegration relation; and (4) nonlinear with a cointegration relation. So far, the literature does not seem to have a good test in the case of the nonlinear with or without a cointegration relation as it involves a different distributional problem (Seo, 2006).

12 / The Energy Journal Table 2-1. E  stimated Results of Nonlinear Multivariate Threshold Model and Directions of Causality Tests (Model I: Dlroilp, Dr, Dlrstkp, Dly)
Linear M AT CA DK FI FR IN IL IT JP KR MY MX NL NO PH PT ZA ES GB US TW A A A A A A E A A A A E A E A A A A E E d 9 3 8 2 9 1 6 2 2 5 4 3 1 3 1 1 1 2 9 2 c 5.0001 2.1460 4.4314 0.5710 4.2763 0.2587 0.0747 4.5631 2.5405 1.0570 0.3367 0.4835 0.9370 2.3615 0.3248 0.1920 3.4484 0.7209 0.3435 1.5473 n2 79 117 80 137 87 120 108 77 124 120 136 98 155 122 121 95 97 144 153 120 110 Dlroilp / Dr 9.37 8.66 19.94* 9.56 12.34 18.67 28.02 30.64
* *** ***

R2 Dlroilp / Dly 23.72** 9.62 7.99 12.56 10.52 18.49

* *

Dlroilp / Dlrstkp 15.23 18.33 13.03 20.75** 14.89 9.23 20.23 22.93
**

Dlroilp / Dr 9.42 8.62 46.43*** 17.46 4.21 18.84 37.15 23.55

* *** **

Dlroilp / Dlrstkp 9.18 22.75** 22.96** 26.30*** 15.54 20.15 18.36

*

Dlroilp / Dly 28.58*** 11.65 6.80 11.61 32.98*** 24.05** 24.84** 10.61 11.21 17.87 33.40*** 12.90 29.04*** 4.71 4.98 18.33* 7.40 21.53** 41.35*** 21.73** 28.47***

25.78 23.14

** **

17.27
*

15.10 9.87 17.17 12.19 9.59 16.64 11.98 11.64 9.09 6.35 24.63** 23.53
**

10.97 16.40 3.63 33.48*** 24.58

**

7.61 10.83 31.15*** 12.41 15.76 9.07 15.66 7.81 8.66 14.48 45.53*** 27.73
*** **

23.66** 13.75 18.50* 12.51 24.75

**

28.01* 25.79*** 9.38 34.15*** 11.82 30.87*** 3.29 12.40 21.55

**

24.16** 3.03 14.63 10.03 24.68** 30.05*** 31.81

***

11.46 21.79
**

24.89** 10.20 15.87 30.07*** 22.56

**

13.98 38.45*** 21.51

**

A 11 2.3829

15.05

7.94

24.13

7.54

10.29

Note: d is the delay of threshold variable; c is the optimal upper bound threshold level; n2 represent the sample sizes of regime 2 (R2); / denotes does not Granger cause; and *, **and *** represent 10%, 5%, and 1% significance levels, respectively. M=A denotes VAR model while M= E denotes VECM model.

From the estimated results of Model I in Table 2-1, the optimal threshold levels c are as follows: the highest level is Austria at 5.0001%, the next highest is Italy at 4.5631%, and the lowest level is Israel at 0.0747%. The next lowest is Portugal at 0.1920% and India is at 0.2587%. The threshold levels c for all other countries are in between. For the delay d of the threshold variables, the shortest period is found for India, the Netherlands, the Philippines, Portugal, and South Africa for 1 period and the longest is the U.S. for 11 periods. For some countries, the impact of a positive oil price change on production and the stock price is rapid (one month), while for other countries the impact is slow (almost one year). The other part of Table 2-1 investigates the Granger causality of the impact of a positive oil price change on the interest rate, production, and stock price. As expected, there is no causal relationship in many countries under the linear model. When

Factors Affecting an Economys Tolerance / 13 Table 2-2. E  stimated Results of Nonlinear Multivariate Threshold Model and Directions of Causality Tests (Model II: v, Dr, Dlroilp, Dly)
Linear M AT CA DK FI FR IN IL IT JP KR MX NL NO PH PT ZA ES GB US TW A A A A A A E d 9 1 3 7 3 3 3 c 0.0313 0.5494 0.3968 0.4063 0.4625 0.0574 0.0198 0.0455 0.3820 0.0472 0.1327 0.0189 0.8608 0.0217 0.4311 0.0067 0.0637 0.0191 0.0066 0.2317 0.0057 n1 182 91 119 106 102 136 111 186 112 146 125 111 62 177 90 93 187 171 187 147 136 v / Dr 8.71 12.56 22.20** 4.72 23.17** 16.54 20.89* 31.64
***

R2 v / Dly 20.41* 12.03 10.10 17.70 6.18 14.34 18.58 21.58

**

v / Dlrstkp 14.35 14.13 12.78 15.32 11.46 6.66 20.45* 21.72

**

v / Dr 7.87 12.11 12.00 14.48 14.07 15.20 7.56 35.72

***

v / Dlrstkp 18.45* 26.02

***

v / Dly 20.71** 16.05 15.75 30.09*** 9.45 23.92** 12.23 17.15 5.79 28.47*** 29.07*** 13.12 97.78*** 9.63 8.45 23.36** 5.54 28.62*** 31.77*** 24.92** 26.80***

21.13** 20.51* 27.61*** 12.94 26.77*** 26.84*** 7.33 19.18 11.87 26.08
**

A 10 A 12 E E A A A A E 3 1 1 7 3 6 4 2

11.05 24.81** 16.02 14.26 16.76 11.88 15.00 18.23 12.26 11.54 17.56 18.69 20.94
* *

11.69 12.71 5.57 28.58

***

4.19 13.71 41.90*** 12.25 12.92 2.45 15.49 6.73 9.50 26.91 21.45
**

26.92*** 12.09 6.41 16.99 3.45 18.64* 9.80 9.65 14.90 10.12 19.09 18.54
*

MY A

A 10

23.62** 19.78* 4.96 13.77 8.95 32.83 27.92

***

12.08 24.06** 23.73** 11.38 27.73*** 28.86*** 22.23** 25.47

***

E 12 A 11 E 5

27.24**
***

40.28***
** **

6.78

25.14

17.85

10.56

Note: d is the delay of threshold variable; c is the optimal upper bound threshold level; n2 represent the sample sizes of regime 2 (R2); / denotes does not Granger cause; and *, **and *** represent 10%, 5%, and 1% significance levels, respectively. M=A denotes VAR model while M=E denotes VECM model.

the data are above the threshold level (regime 2), certain causal relationships appear. In the linear case, and for nine out of 21 countries, the causality test does not find any linkage between an oil price change and the interest rate, or stock price, or production. Under regime 2, all countries display at least one causal relation: the impact of an oil price change leads the interest rate, stock price, or production. Nine countries show that the oil price change leads the interest rate change (four countries in the linear case); nine countries indicate that the oil price change leads the stock price change (eight countries in the linear case); and ten countries suggest that the oil price change leads industrial production (seven countries in the linear case). It is clear that the economic impacts of an oil price change are nearly unavoidable when the oil price change exceeds a certain threshold. For some countries, the economic impact of an oil price change is immediate, but it

14 / The Energy Journal may take six months or more for some other countries. For some countries, the oil price change needs to exceed the threshold level of 4% or 5% to exert any economic impact. Yet, for other countries, it may take less than 1% or 2% of an oil price change to have an economic impact. What are the factors affecting the differences in the threshold level and the speed of response in different countries? These discussions are at the center of our contribution and are also applicable to the economic impact emanating from the shock of an oil price change. Table 2-2 displays the economic impact of the shock of an oil price change. The fastest response is found in Canada, Malaysia and Mexico with one period, while the U.K. and Japan are the slowest with 12 periods. Some economies can sustain the greater shock of an oil price change and some cannot. The greatest degree of sustainability is in the Netherlands at 0.8608. The next is in Canada at 0.5494 followed by France at 0.4625. The lowest sustainability of the economy to the shock is found in the U.K. at 0.0066. The next lowest is in Portugal at 0.0067. The Granger causality test demonstrates that ten out of 21 countries under the linear case exhibit no causal relationship. However, if the data are classified as higher than the threshold value (regime 2), the impact of the shock of an oil price change on the interest rate change, stock price change, or industrial production change is much greater. For 12 countries, the shock of an oil price change is found to lead the stock price change under regime 2, but only impacts six countries in the linear case. The shock of an oil price change leads industrial production change in 11 countries under regime 2, but only in six countries in the linear case. The relationship between the shock of an oil price change and interest rate change indicates that an oil price change leads the interest rate in four countries under the linear model. Yet, under the nonlinear model (regime 2), the shock of an oil price change leads the interest rate change in only two countries. Similarly, we need to know whether the economic impact of the shock of an oil price change is immediate or takes more than six months, and whether the tolerance of the impact is different among countries. Our results show that, overall, the statistical results from the impact of an oil price change are stronger than those from the impact of the shock of an oil price change. These findings are consistent with the findings of a three-country study by Huang et al. (2005). Furthermore, the previous literature has been lacking in terms of taking into consideration the differences in economic endowments and energy conditions among countries; therefore, earlier studies find little relationship between an oil price change or the shock of an oil price change and its impact on economic variables. With the threshold regression, we are able to identify the economic impact of a positive oil price change and its shock on the speed and tolerance of the impact in each country.

Factors Affecting an Economys Tolerance / 15 6.  Elements Affecting the Threshold Level and the Delay Periods of Threshold Variables The delay (d) of the threshold variable (qt-d) reflects the speed of response based on the economic impact of a positive oil price change and its shock. The value of the threshold level (c) reflects the tolerance of the impact. Due in part to the difference in economic development, energy dependence, and the efficiency of energy use, the speed of economic response (d) and the degree of tolerance (c) from an oil price change and its shock are expected to be different in each country. Tables 2-1 and 2-2 display our estimates of the speed of response (delay periods d) and the degree of tolerance (or threshold level c) as a consequence of the impact of a positive oil price change and its shock. The next step is to investigate the factors affecting the speed and the tolerance of the impact. The possible potential factors affecting the speed and tolerance of the impact of an oil price change and its shock are the degree of economic development, oil imports, and the proportion of the use of energy in the industry and/or transportation sector to the total consumption of energy. Those factors could all be used to partially explain the impact of a positive oil price change and its shock on the speed of response (d) and the tolerance level (c) in each country.10 We collect the data from the Energy Balance published by the International Energy Agency (IEA). Those data include: (1) the percentage of energy imports to total final energy consumption (imp/tfc); (2) the percentage of energy use in industry to total energy consumption (ind/tfc); (3) the percentage of energy use in transportation to total energy consumption (tra/tfc); and (4) the log of real per capita GDP (lkgdp) to represent the degree of economic development. Because the values of c and d in Table 2 are obtained from cross-sectional data consisting of 21 countries, the variables used to interpret c and d need to be cross-sectional as well. Most of our data start in the 1970s with some starting in the 1980s; hence, we need to calculate the yearly average of the four above-mentioned variables from 1971-2003 and 1981-2003 to calculate their cross-sectional values. The reasons for using the periods 1971-2003 and 1981-2003 are: (1) the availability of most of the data for the 21 countries studied in those two periods; and (2) to study whether there is a change in imports of energy and final energy use. Table 3 shows the means and standard deviations of the related energy data in those two periods. In terms of the percentage of oil imports to total energy consumption (imp/tfc), the 1981- 2003 averages in most countries exhibit a downward trend (9 countries indicate an upward trend). Among the 21 countries, Denmark (since 1996), Canada, the U.K., Mexico, Malaysia, and Norway are all oil exporting
10. Other variables such as different monetary policy responses, may well impact both the magnitude and the length of the effect emanated from an oil price change ( Bernanke et al., 1997; Hamilton and Herrera, 2004). In addition, Kilian (2007) categorizes source of the shock into demand and supply driven. While these variables are of great value in the analysis, they are not easily accessible and quantifiable.

16 / The Energy Journal Table 3. Basic Statistics of Related Energy Data During 1971-2003 and 1981-2003
imp/tfc -71 AT 0.886 (0.045) CA 0.314 (0.079) DK 1.248 (0.179) FI 0.955 (0.065) FR 0.988 (0.094) IN 0.278 (0.057) IL 1.519 (0.518) IT 1.282 (0.085) JP 1.289 (0.053) KR 1.202 (0.260) MY 0.736 (0.304) MX 0.098 (0.077) NL 1.898 (0.257) NO 0.450 (0.203) imp/tfc -81 0.892 (0.048) 0.287 (0.066) 1.166 (0.143) 0.946 (0.060) 0.934 (0.043) 0.259 (0.054) 1.734 (0.125) 1.240 (0.036) 1.264 (0.033) 1.322 (0.219) 0.604 (0.117) 0.111 (0.083) 2.016 (0.193) 0.325 (0.049) tra/tfc -71 0.241 (0.015) 0.274 (0.011) 0.276 (0.038) 0.168 (0.019) 0.258 (0.042) 0.213 (0.083) 0.377 (0.038) 0.278 (0.045) 0.241 (0.030) 0.200 (0.042) 0.361 (0.024) 0.353 (0.030) 0.198 (0.033) 0.216 (0.018) tra/tfc -81 0.244 (0.017) 0.276 (0.009) 0.297 (0.021) 0.179 (0.009) 0.282 (0.023) 0.181 (0.080) 0.359 (0.030) 0.304 (0.021) 0.258 (0.016) 0.219 (0.034) 0.369 (0.022) 0.362 (0.031) 0.214 (0.025) 0.223 (0.016) ind/tfc -71 0.294 (0.028) 0.354 (0.010) 0.195 (0.013) 0.405 (0.049) 0.305 (0.033) 0.393 (0.084) 0.257 (0.049) 0.364 (0.047) 0.441 (0.067) 0.407 (0.031) 0.398 (0.020) 0.385 (0.030) 0.361 (0.029) 0.416 (0.037) ind/tfc -81 0.278 (0.011) 0.358 (0.008) 0.189 (0.009) 0.429 (0.037) 0.286 (0.016) 0.376 (0.095) 0.242 (0.051) 0.336 (0.018) 0.403 (0.032) 0.404 (0.036) 0.398 (0.019) 0.385 (0.036) 0.350 (0.025) 0.400 (0.031) kgdp -71 kgdp -81

18329.94 20218.06 (3818.89) (2860.66) 18364.89 19802.54 (3009.36) (2335.49) 23829.15 25563.23 (3670.35) (2979.00) 17605.87 19413.59 (3546.73) (2548.34) 17301.14 18752.13 (2882.70) (2089.78) 304.02 (94.13) 343.96 (85.69)

14110.50 15363.66 (2550.21) (1972.14) 14707.32 16215.32 (2876.60) (1927.70) 29025.01 32530.81 (6749.16) (4739.12) 6127.57 7608.81

(3259.24) (2788.25) 2480.83 (934.78) 5017.76 (529.05) 2916.33 (771.17) 5269.06 (333.13)

17830.93 19217.39 (3182.83) (2789.29) 26537.28 30129.89 (7258.46) (5465.38)

Note: imp/tfc = the ratio of energy import to total energy consumption; tra/tfc = the ratio of energy use in transportation sector to total energy consumption; ind/tfc = the ratio of energy use in industry sector to total energy consumption; 71 = 71-03 yearly average; and 81 = 81-03 yearly average. kgdp= per capita real GDP (2000 USD). Numbers inside ( ) are standard deviation.

Factors Affecting an Economys Tolerance / 17 Table 3. Basic Statistics of Related Energy Data During 1971-2003 and 1981-2003 (continued)
imp/tfc -71 PH 0.855 (0.069) PT 1.213 (0.082) ZA 0.313 (0.070) ES 1.198 (0.036) GB 0.603 (0.156) US 0.334 (0.084) TW 1.287 (0.201) imp/tfc -81 0.849 (0.081) 1.236 (0.077) 0.300 (0.080) 1.205 (0.034) 0.530 (0.062) 0.355 (0.085) 1.369 (0.169) tra/tfc -71 0.220 (0.076) 0.309 (0.016) 0.229 (0.019) 0.348 (0.033) 0.278 (0.045) 0.361 (0.030) 0.221 (0.036) tra/tfc -81 0.232 (0.087) 0.309 (0.019) 0.224 (0.015) 0.364 (0.023) 0.303 (0.027) 0.377 (0.021) 0.237 (0.031) ind/tfc -71 0.291 (0.026) 0.421 (0.029) 0.481 (0.047) 0.390 (0.054) 0.303 (0.056) 0.286 (0.029) 0.550 (0.035) ind/tfc -81 0.283 (0.027) 0.417 (0.033) 0.480 (0.050) 0.361 (0.036) 0.270 (0.022) 0.272 (0.023) 0.538 (0.031) kgdp -71 914.99 (74.94) 7378.58 (1912.21) 3110.92 (207.76) 10399.46 (2288.29) 18591.30 (3797.38) 26471.13 (5136.59) 7138.59 (3771.57) kgdp -81 934.19 (63.83) 8254.91 (1610.51) 3048.78 (204.41) 11375.86 (2048.30) 20330.48 (3198.49) 28942.87 (4054.37) 8891.22 (3148.52)

countries. The remaining 15 countries are oil importing countries. When a country is dependent on oil imports, this country will be more sensitive in response to an oil price change and the degree of tolerance is expected to be smaller. If these two are related, we expect a negative relationship between the imp/tfc ratio and c or d. With regard to energy consumption in the industry and transportation sectors in each country, we do not see a noticeable change before and after the energy crisis, but the relative importance of the industry and transportation sectors to energy demand in these 21 countries is different. In some countries, such as Finland, Japan, South Korea, Norway, Portugal, South Africa, and Taiwan, energy consumption in industry is the largest sector in terms of energy use and the percentage of energy use in industry to total energy consumption (ind/tfc) is over 40%. However, in Israel, Malaysia, Mexico, Spain and the U.S., the energy use in the transportation sector (tra/tfc) accounts for over one-third of total energy consumption. The literature on the relationship between energy consumption and economic growth such as Akarca and Long (1979) indicates that energy consumption can bring about economic growth. If that is the case, GDP is directly related to industry output and indirectly related to transportation, and the impact of an oil price change and its shock for countries with a high percentage of energy con-

18 / The Energy Journal sumption (ind/tfc or ind/tfc+tra/tfc) should be greater than countries with a low percentage of energy consumption. If these two are related, we expect a negative relationship between energy consumption (ind/tfc or ind/tfc+tra/tfc) and c or d. Finally, in a more developed country, the economic impact of an oil price change (or its shock) can be more easily adjusted through economic policy and technological advances. Hence, we expect a positive relationship between the economic development variable and c or d. The log of real per capita GDP (deflated by the 2000 consumer price index for average values of 1971-2003 and 1981-2003) is used to represent different degrees of economic development. In order to explore how these variables explain c and d, the regression model is used in the statistical analysis. Since c is a random variable, a multiple regression is adequate. However, d is a multiple-response type variable and the size of d indicates the response periods; an ordered multiple-choice model needs to be used (e.g., ordered probit or logit). The multiple regression of c can be shown as: cDlroilp (or cv) = a0 + a'Xi + ei, (7)

where Xi represents explanatory variables (imp/tfc, ind/tfc+tra/tfc, lkgdp). 0 is a constant term and i follows an iid distribution. cDlroilp and cv represent tolerance for the threshold level from an oil price change and from the shock of an oil price change, respectively. Since we use cross-sectional data for 21 countries, we may run into the problem of heteroskedasticity. The heteroskedasticity-consistent covariance model is employed to correct the data for heteroskedasticity. Since d is an ordered variable, the ordered multiple-choice model is established as:
* dD (or d* ) = b'Xi + ei, lroilp v

(8)

where * denotes an unobserved latent variable; Xi denotes the explanatory variables (imp/tfc, ind/tfc, lkgdp); and i is assumed to follow an iid process. dDlroilp and dv represent the speed of response from an oil price change and from the shock of an oil price change, respectively. If the cumulative probability density function is assumed to follow a normal distribution, we have an ordered probit model. If i is a logistic type distribution, we have an ordered logit model. Table 4 displays the estimated results of cDlroilp and cv from the multiple regression model, and the estimated results of dDlroilp and dv from the ordered variable model with a normal distribution.11 From the column for cDlroilp in Table 4, we observe a significant correlation between cDlroilp and two variables (ind/tfc+tra/tfc and lkgdp). Both the ind/ tfc+tra/tfc and lkgdp variables are statistically significant at less than 5% and 1%,
11. Since we cannot reject the null hypothesis that residuals are normally distributed, Table 4 only reports the estimated results from the probit model. The logit model is also used in our estimation. Both sets of results are very much alike. To save space, only the estimated results from the probit model are displayed in Table 4. Interested readers may request the results from the logit model.

Table 4. The Estimated Results from the Threshold Level of the Multiple Regression Model and the Delay of Threshold Variables from the Ordered Dependent Variable Model

(probit) 71-03 0.7763*** (2.86) -0.7043** (-2.38) -8.5145** (-2.40) -1.1551** (-2.03) -1.0802*** (-2.49) 0.14 2.21 [0.11] 2.03 [0.36] 0.21 0.22 2.18 [0.34] 1.21 [0.55] 0.0988 (0.51) -8.7384*** (-2.49) 0.0837 (0.60) 0.10 2.80 [0.06] 1.23 [0.54] 0.67 [0.72] 0.06 -0.6566** (-2.07) 1.8236 (0.76) 0.3225 (0.60) 0.7769*** (2.87) 0.0606** (2.26) 0.0581** (2.12) 81-03 71-03 81-03 71-03 0.4400** (1.96)

dep

clroilp

d lroilp

cv

Indep.

(probit) 81-03 0.4208** (2.02)

dv

71-03

81-03

lkgdp

0.6268*** (3.31)

0.6527*** (3.23)

ind/tfc+ tra/tfc

-5.3468** (-2.23)

-5.7096** (-2.20)

ind/tfc

0.6003 (0.25) 0.2596 (0.56)

imp/tfc

-0.5675 (-0.91)

-0.6595 (-1.16)

2 R

0.20

0.21

White F test

0.84 [0.56]

0.89 [0.53]

J-B test

1.00 [0.61]

0.92 [0.63]

0.76 [0.68] 0.05

Pseudo R2

Factors Affecting an Economys Tolerance / 19

Note: numbers inside ( ) are t statistics; numbers inside [ ] are p-value; *, **, *** represent 10%, 5%, and 1% significance levels, respectively; dep. and indep. represent the dependent and independent variables, respectively; probit represent the ordered dependent variable model based on normal error distributions i; J-B = Jarque-Bera normality test; and White F test = White heteroskedasticity test. lkgdp= log of per capita GDP.

20 / The Energy Journal respectively. The estimated signs are as expected. In other words, the countries with a higher level of economic development tend to have a higher level of tolerance in responding to a positive oil price change or shock. If the industry and transportation sectors consume more energy in a country, the ability to tolerate a positive oil price change or shock will be lower. Irrespective of whether the countries are oil importing or oil exporting countries, we do not find the relation to be significant for the imp/tfc variable to interpret cDlroilp. In the cv column, we observe that there is a negative significant relationship (lower than 5%) between ind/tfc+tra/tfc and cv, and a positive significance relationship (lower than 5%) between cv and lkgdp. Overall, both variables explain about 20% and 14% of the variation in cDlroilp and cv, respectively. The test results for heteroskedasticity using the White test indicate that our model is free of heteroskedasticity. Furthermore, the Jarque-Bera (J-B) statistic indicates that the residual follows a normal distribution. The explanatory power of the model using 1971-2003 data is about the same as that of the model using 1981-2003 data. For the interpretation of the delay period (d), we use lkgdp, ind/tfc and imp/tfc variables to explain dDlroilp and the pseudo R2 of 0.21 (based on the probit model) is much higher than the 0.05 obtained from dv . When the economic development is more advanced in a country with lower energy consumption in the industrial sector and lower energy imports, the response period (d) from a positive oil price change should be longer. By using the same variables to explain dv, only lkgdp can significantly explain the variation in dv. The other two variables (ind/ tfc and imp/tfc) cannot significantly explain the variation in dv. Overall, lkgdp explains approximately 5% of the variation in dv. The J-B test results indicate that both the residuals of dDlroilp or dv cannot reject the null hypothesis of normality. It is clear that the use of the probit model to estimate the delay period is appropriate. From the above regression estimation, economic development (per capita real GDP) is the most important factor affecting the threshold of tolerance and response time emanating from the impact of a positive price change and its shock (proxied by the standardized residuals). In other words, the more a countrys economic growth advances, the better the country can withstand the impact of an oil price change and its shock. The percentage of energy use in the industrial sector and oil imports as a percentage of total consumption can be used to explain the delay in the response to an oil price change. The percentage of energy consumption in both the industry and transportation sectors can be used to explain the threshold of tolerance from an oil price change and its shock. 7. Concluding Remarks Subsequent to the 9/11 terrorist attack in 2001, the Dow Jones Industrial Average finally broke the 12,000 mark on October 19, 2006. The economic impact of a recent oil price hike was not as severe as that of the previous energy crises. The recent price hike did not seem to harm the U.S. economy as much as was expected. For some countries, however, the impact was more severe. It is

Factors Affecting an Economys Tolerance / 21 clear that the impact of a positive oil price change on the threshold of tolerance and the speed of response is different for each country. This paper first employs a nonlinear model to estimate threshold level (c) and the delay of response (d) from the impact of an oil price change and its shock. We also employ energy use and economic growth-related variables to explain different possible values of c and d in each country. In general, the length of the response to the economic impact of an oil price change is about 3 ~ 4 quarters (Hamilton and Herrera, 2004). Hence, a delay of 12 periods in monthly data is used in four-variable threshold VAR or threshold VEC analysis. Since the required sample period cannot be too short, only 21 countries with at least 200 data points in each country are available to collect the necessary data. Using the delay (d) of an oil price change (lroilpt-d) and the shock of an oil price change (vt-d) as threshold variables, the nonlinear statistical test in each country supports the existence of the nonlinear relation. We, therefore, can estimate without much difficulty the threshold level (c) and the delay of threshold variable (d) from the nonlinear model. The purpose of this paper is to identify evidence affecting different values of c and d for each country. We find possible explanatory variables from energy balance data published by the IEA. A nonlinear model is used to estimate the threshold level (c) and the delay of threshold variable (d) in each country. A multiple regression model is then employed to test and identify the relationship between energy variables and c or d in each country. We find: (1) as a country becomes more advanced in economic development (real per capita GDP is used to denote the degree of economic development) and acquires a lower ratio of energy use in its industry and transportation sectors, the threshold of tolerance is greater as evidence by the positive impact of an oil price change and its shock; (2) if a country has a lower ratio of energy use in the industry sector, a lower energy import ratio and is more advanced in economic development, it will have a longer delay in terms of its economic response from the positive impact of an oil price change; and (3) as an economy becomes more advanced, the length of the response time from the impact of the shock of an oil price change will be longer. The results from the regression analysis suggest that the economic development variable is the most important variable affecting the threshold of tolerance (cDlroilp and cv) and the speed of response (dDlroilp and dv) emanating from the impact of a positive price change and its shock. Although we have found possible factors to explain the threshold level and the speed of adjustment from the impact of positive oil price shock, this research is not without limitations. The potential shortcomings of this research are as follows: (1) With the regression results based on 21 countries, the estimation results might not be as reliable. (2) With such a sample size, it might be a bit of a stretch to rely on asymptotic theory to justify the use of t-statistics in conducting inference. (3) There is a host of possible omitted variables that may affecting the tolerance level and the delay of the effect. Some possibilities pertain to monetary policy and the type of shocks, but there are other possibilities such as

22 / The Energy Journal degree of openness of the economy, fiscal and exchange rate policy, etc. (see for instance Bohi, 1991) These omitted variables may be included in future analyses to test the robustness of the result. References
Akarca, A. T. and T. V. Long (1979). Energy and Employment: A Time Series Analysis of the Causal Relationship. Resources and Energy 2: 151-162. Balke, N. S., S. P. A. Brown, and M. K. Ycel (2002). Oil Price Shocks and the U.S. Economy: Where Does the Asymmetry Originate? Energy Journal 23(3): 27-52. Balke, N. and T. Fomby (1997). Threshold Cointegration. International Economic Review 38: 627-645. Bernanke, B. S., M. Gertler, and M. Watson (1997). Systematic Monetary Policy and the Effects of Oil Price Shocks. Brookings Papers on Economic Activity 1: 91-157. Bohi, D. R. (1991). On the Macroeconomic Effects of Energy Price Shocks. Resources and Energy 13: 145-162. Brown, S. and M. K. Ycel (2002). Energy Prices and Aggregate Economic Activity: An Interpretive Survey. Quarterly Review of Economics and Finance 42: 193-208. Burbidge, J. and A. Harrison (1984). Testing for the Effects of Oil Price Rises Using Vector Autoregression. International Economic Review 25: 459-484. Carruth, A. A., M. A. Hooker, and A. J. Oswald (1998). Unemployment Equilibria and Input Prices: Theory and Evidence from the United States. Review of Economics and Statistics 80(4): 621-28. a, (2003). Do Oil Price Shocks Matter? Evidence from Some Cuado, J. and Fernando Prez de Graci European Countries. Energy Economics 25: 137-154. Davis, S. J. and J. Haltiwanger (2001). Sectoral Job Creation and Destruction Response to Oil Price Changes. Journal of Monetary Economics 48: 465-512. de Miguel, C., B. Manzano, and J. M. Martin-Moreno (2003). Oil Price Shocks and Aggregate Fluctuations. Energy Journal 24(2): 47-61. Elliott, G., T. J. Rothenberg and J. H. Stock (1996). Efficient Test for an Autoregressive Unit Root. Econometrica 64: 813-836. Gilbert, R. J. and K. A. Mork (1986). Efficient Pricing During Oil Supply Disruptions. The Energy Journal 7(2): 51-68. Gisser, M. and T. H. Goodwin (1986). Crude Oil and the Macroeconomy: Tests of Some Popular Notions. Journal of Money, Credit and Banking 18: 95-103. Hamilton, J. D. (1983). Oil and the Macroeconomy Since World War II. Journal of Political Economy 91: 228-248. Hamilton, J. D. (1994). Time Series Analysis. New Jersey: Princeton University Press. Hamilton, J. D. (1996). This is What Happened to the Oil Price-Macroeconomy Relationship? Journal of Monetary Economics 38(2): 215-220. Hamilton, J. D. (2003) What is an Oil Shock? Journal of Econometrics 113: 36398. Hamilton, J. D. and A. M. Herrera (2004). Oil Shocks and Aggregate Macroeconomic Behavior: The Role of Monetary Policy: A Comment. Journal of Money, Credit and Banking 36(2): 265-286. Hooker, M. A. (1999). Oil and the Macroeconomy Revisited. Mimeo, Federal Reserve Board, Washington, D.C., August. Huang, B. N., M. J. Hwang, and H. P. Peng (2005). The Asymmetry of the Impact of Oil Price Shocks on Economic Activities: An Application of the Multivariate Threshold Model. Energy Economics 27: 455-76. Huang, R. D., R. W. Masulis, and H. R. Stoll (1996). Energy Shocks and Financial Markets. Journal of Futures Markets 16(1): 1-27. Jimnez-Rodriguez, R. and M. Snchez (2005). Oil Price Shocks and Real GDP Growth: Empirical Evidence for Some OECD Countries. Applied Economics 37: 201-228.

Factors Affecting an Economys Tolerance / 23

Johansen, S. (1988). Statistical and Hypothesis Testing of Cointegration Vectors. Journal of Economic Dynamics and Control 12: 231-254. Jones, C. M. and G.. Kaul (1996). Oil and the Stock Market. Journal of Finance 51: 463-491. Kaul, G. and H. N. Seyhun (1990). Relative Price Variability, Real Shocks, and the Stock Market. Journal of Finance 45: 479-496. Keane, M. P. and E. S. Prasad (1996). The Employment and Wage Effects of Oil Price Changes: A Sectional Analysis. Review of Economics and Statistics 78: 389-399. Kilian, Lutz (2008). A Comparison of the Eects of Exogenous Oil Supply Shocks on Output and Ination in the G7 Countries. Journal of the European Economic Association March: forthcoming. Kilian, Lutz (2007). "Not All Oil Price Shocks Are Alike: Disentangling Demand and Supply Shocks in the Crude Oil Market. Accessed at http://www-personal.umich.edu/~lkilian/ aer111507r1.pdf Lee, K., S. Ni, and R. A. Ratti (1995). Oil Shocks and the Macroeconomy: The Role of Price Variability. Energy Journal 16(4): 39-56. Lo, M. and E. Zivot (2001). Threshold Cointegration and Nonlinear Adjustment to the Law of One Price. Macroeconomic Dynamics 5: 533-576. Mork, K. A. (1989). Oil and the Macroeconomy when Prices Go Up and Down: An Extension of Hamiltons Results. Journal of Political Economy 97: 740-744. Mork, K., O. Olsen, and H. Mysen (1994). Macroeconomic Responses to Oil Price Increases and Decreases in Seven OECD Countries. Energy Journal 15: 19-35. Ng, S. and P. Perron (2001). Lag Length Selection and the Construction of Unit Root Tests with Good Size and Power. Econometrica 69: 1519-1554. Papapetrou, E. (2001). Oil Price Shocks, Stock Markets, Economic Activity and Employment in Greece. Energy Economics 23: 511-532. Sadorsky, P. (1999). Oil Price Shocks and Stock Market Activity. Energy Economics 21: 449-469. Seo, M. (2006). Bootstrap Testing for the Null of No Cointegration in a Threshold Vector Error Correction Model. Journal of Econometrics 134: 129-50. Tatom, J. A. (1988). Are the Macroeconomic Effects of Oil Price Changes Symmetric? CarnegieRochester Conference Series on Public Policy 28: 325368. Tatom, J. A. (1993). Are There Useful Lessons from the 199091 Oil Price Shock? Energy Journal 14 (4): 129 150. Tong, H. (1978). On a Threshold Model. in C. D. Chen (ed.), Pattern Recognition and Signal Processing, Amsterdam: Sijthoff and Noordhoff, 101-41. Tsay, R. S. (1998). Testing and Modeling Multivariate Threshold Models. Journal of the American Statistical Association 93: 1188-1202. Weise, C. L. (1999). The Asymmetric Effects of Monetary Policy: A Nonlinear Vector Autoregression Approach. Journal of Money, Credit, and Banking 31(1): 85-108.

Appendix A The Definition of C(d) Statistic: The C(d) statistic is calculated as follows. First, the linear multivariate model can be denoted in matrix form as shown below: y't = X 't + et, t = 1,....,n (A-1)

where h=max(p, d), Xt=(1, y't1, y't2,..., y'tp) ' is a (pk+1)-dimensional regressor, and denotes the parameter matrix. The data are arranged in the order of small to large according to the values of the threshold variable qt-d. The arranged regression becomes:

24 / The Energy Journal y't(i)+d = X 't(i)+d + et(i)+d, i = 1,....,n h (A-2)

where t(i) is the time index of q(i). Tsay (1998) used the recursive least squares approach to estimate (A-2) and, at the same time, predictive residuals were obtained. If the threshold relation does not exist, the predictive residuals approach white noise. Consequently, the predictive residuals are uncorrelated with the regressor Xt(i)+d. On the other hand, if yt follows a threshold model, then the predictive residuals are no longer white noise, because the least squares estimator is biased. In this case, the predictive residuals are correlated with the regressor Xt(i)+d. Let m be the least squares estimate of of equation (A-2) with i=1, , m; i.e., the estimate of the arranged regression using data points associated with the m smallest values of qt-d. Let ' X e t (m +1)+ d = yt (m +1)+ d m t (m +1)+ d and t (m +1)+ d = e t (m +1)+ d /[1 + X 't (m +1)+ dVmXt (m +1)+ d]1/2,
m

(A-3)

(A-4)

where Vm = [i=1 Xt(i)+d X't(i)+d]1 is the predictive residual and the standardized predictive residual of regression (A-2). Next, consider the regression t (l)+ d = X 't (l)+ d + w't (l)+ d, l = m 0 + 1,...,n h, (A-5)

where m0 denotes the starting point of the recursive least squares estimation. The problem of interest is then to test the hypothesis H0: = 0 versus the alternative Ha: 0 in regression (A-5). The C(d) statistic is therefore defined as: C (d ) = [n h m 0 (kp + 1)] {ln[det(S0)] ln[det(S1)]} (A-6)

where the delay d signifies that the test depends on the threshold variable qt-d, det(A) denotes the determinant of the matrix A, and 1 S0 = n h m0 and 1 S1 = n h m0
nh nh

l=m 0 +1

t (l)+ d 't (l)+ d

l=m 0 +1

t (l)+ d w 't (l)+ d, w

where w t is the least squares residual of regression (A-5). For more details, see Tsay (1998).

Factors Affecting an Economys Tolerance / 25 Finding the Optimal Threshold Level: The data are arranged in ascending order based on the threshold variable qt-d. Depending on the abundance of the data, 15% to 25% of the smallest and largest values are discarded. Let the largest value be c or the 85th percentile (if 15% of th the larger data are discarded) or the 75 percentile (if 25% are discarded). Let the smallest value c be the 15th percentile (if 15% of the smaller data are discarded) th or the 25 percentile (if 25% of the values are discarded). Then, a grid search proAppendix Table 1. Results of Unit Root Tests
AT(T=424) Level DFGLS r lroilp lrstkp ly -1.89 -1.42 -1.20 -1.59 MZ -8.69** -4.03 -3.77 -5.04 Difference DFGLS -2.54** -16.70*** -15.96*** -2.06** DK(T=376) Level DFGLS r lroilp lrstkp ly -5.14*** -2.33 -1.54 -2.95** MZ Difference DFGLS MZ -70.07*** -134.86*** -287.44*** -8.38** DFGLS -1.89 -2.05 -2.29 -1.72 Level MZ -7.48 -7.24 -8.44 -13.54 MZ -146.65*** -202.29*** -198.24*** -89.51*** DFGLS -2.15** -2.31 -2.54 -2.51 Level MZ -11.93** -9.06 -16.08* -13.61 CA(T=364) Difference DFGLS MZ

-22.61*** -72.90*** -14.36*** -127.57*** -16.48*** -160.40*** -8.87*** -295.98***

FI(T=329) Difference DFGLS MZ

-75.09*** -5.21*** -9.55 -9.08 -14.71*** -17.78***

-12.24*** -135.40*** -12.35*** -115.97*** -12.65*** -131.18*** -19.15*** -72.78***

-177.74*** -6.83*** FR(T=424) Level

IN(T=341) Level DFGLS -2.72*** -0.81 -1.36 -1.59 MZ Difference DFGLS MZ

Difference DFGLS -4.44*** -16.77*** -5.63*** -2.45** MZ -111.79*** -206.68*** -185.78*** -65.80***

DFGLS r lroilp lrstkp ly -1.31 -1.50 -1.21 -1.36

MZ -4.09 -4.12 -3.05 -14.95*

-42.29*** -15.43*** - 63.85*** -1.51 -3.43 -12.33 -14.47*** - 161.70*** -11.92*** - 111.05*** -3.19*** - 42.63*** continued

26 / The Energy Journal Appendix Table 1. Results of Unit Root Tests (continued)
IL(T=250) Level DFGLS r lroilp lrstkp ly -0.89 -0.22 MZ -1.31 0.30 Difference DFGLS -2.35** -2.36** -7.42*** -2.00** JP(T=424) Level DFGLS r lroilp lrstkp ly -0.98 -1.60 -1.44 -1.87 MZ -2.28 -4.43 -4.36 -3.57 Difference DFGLS -3.42*** -17.00*** -13.49*** -2.08** MY(T=304) Level DFGLS r lroilp lrstkp ly MZ Difference DFGLS -4.08*** -12.71*** -2.81*** -3.02*** MZ -93.91*** -114.60*** -43.71*** -94.79*** DFGLS -3.12** -0.23 -1.71 -2.17 Level MZ -13.61 0.39 -4.81 -10.16 MZ -191.98*** -210.10*** -165.15*** -297.71*** DFGLS -1.00 -2.33 -1.70 -1.45 Level MZ -2.91 -9.43 -5.42 -7.90 MZ -8.83** -41.10*** -77.62*** -34.71*** DFGLS -1.46 -1.46 -1.26 -1.85 Level MZ -4.11 -4.18 -3.52 -8.85 IT(T=412) Difference DFGLS -4.89*** -2.27** -6.54*** -4.59*** MZ -86.64*** -43.67*** -179.45*** -83.73***

-4.35*** -23.90*** -1.61 -11.78

KR(T=328) Difference DFGLS -3.49*** -13.21*** -12.57*** -21.81*** MZ -81.16*** -108.87*** -140.86*** -202.23***

MX(T=256) Difference DFGLS -11.59*** -10.30*** -3.48*** -1.98** MZ -32.02*** -101.44*** -61.84*** -44.05*** continued

-3.02*** -26.45*** -1.79 -3.90

-2.86*** -10.73 -2.13 -15.13*

cedure targets the middle 70% or 50% of the rank-ordered data set to estimate the according to difdeterminant of the variance and the covariance matrix det ferent threshold levels ci. Finally, equation (3) is used to find the optimal threshold value c*. Since we are interested only in the positive (price rise) impact of an oil

Factors Affecting an Economys Tolerance / 27 Appendix Table 1. Results of Unit Root Tests (continued)
NL(T=424) Level DFGLS r lroilp lrstkp ly -2.97*** -1.49 -1.19 -1.42 MZ Difference DFGLS MZ DFGLS Level MZ -8.32** -4.28 -3.10 -3.21 NO(T=405) Difference DFGLS -18.26*** -16.22*** -7.78*** -15.56*** MZ -245.00*** -192.73*** -84.72*** -666.96***

-18.09*** -22.78*** -4.04 -1.81 -14.59* -16.63*** -4.70*** -2.82***

-208.71*** -2.02** -196.21*** -1.46 -94.96*** -51.31*** -1.17 -0.75

PH(T=292) Level DFGLS r lroilp lrstkp ly -4.41*** -1.08 -1.68 -1.30 MZ Difference DFGLS MZ DFGLS Level

PT(T=208) Difference MZ DFGLS MZ -25.50*** -85.35*** -7.47** -47.61***

-35.50*** -14.17*** -3.72 -5.58 -3.69 -10.51*** -16.68*** -16.88*** ZA(T=424)

-822.94*** -1.13 -116.05*** -2.18 -144.56*** -1.15 -144.99*** -1.94

-21.37*** -3.11*** -11.59 -2.56 -7.34 -9.22*** -1.64** -21.30***

ES(T=376) Level DFGLS MZ Difference DFGLS MZ -15.35*** -227.17*** -168.20*** -16.51*** continued

Level DFGLS r lroilp lrstkp ly -1.59 -1.20 -2.68*** -0.97 MZ -6.71 -2.70 -14.46* -3.79

Difference DFGLS -20.13*** -16.79*** -2.70*** -13.19*** MZ

-210.59*** -4.47*** -203.32*** -1.94 -40.92*** -0.52

-33.55*** -11.80*** -6.79 -0.50 -3.88 -13.29*** -13.89*** -4.01***

-120.80*** -2.02

price change on economic activities, the lower bound ( c ) of the grid search is set at the 50th percentile of all the points and, therefore, the search area ( c, c ) should be between the 85th percentile of the points (or the 75th percentile of the points) for the largest value, and the 50th percentile of the points for the smallest value.

28 / The Energy Journal Appendix Table 1. Results of Unit Root Tests (continued)
GB(T=400) Level DFGLS r lroilp lrstkp ly -1.45 -1.63 -1.42 -2.20* MZ -6.92 -4.82 -3.15 -10.69 Difference DFGLS -19.56*** -5.97*** -13.53*** -3.15*** MZ Level DFGLS MZ -12.78 -3.30 -2.35 -11.88 US(T=424) Difference DFGLS -13.67*** -16.78*** -3.38*** -7.79*** MZ -109.91*** -205.77*** -86.59*** -319.59***

-144.36*** -2.80* -125.44*** -1.40 -151.33*** -1.14 -10.97** -2.60*

TW(T=292) Level DFGLS r lroilp lrstkp ly -1.40 -1.57 -1.24 -1.62 MZ -3.75 -3.74 -4.26 -3.92 Difference DFGLS -13.96*** -5.70*** -15.50*** -4.20*** MZ -72.15*** -61.83*** -143.82*** -33.38*** Level DFGLS MZ Difference DFGLS MZ

Note: r is the interest rate. lroilp, lrstkp and ly represent respectively the real oil price, real stock price, and industrial production index after logarithmic transformations; *, **, and *** represent respectively 10%, 5%, and 1% significance levels; T denotes sample size.

Factors Affecting an Economys Tolerance / 29 Appendix Table 2-1.  Results of the Johansen Cointegration Test (lroilp, r, lrstkp, ly model)
H0 r=0 AT r1 r2 r3 r=0 CA r1 r2 r=0 DK r1 r2 r=0 FI r1 r2 r3 r=0 FR r1 r2 r3 r=0 IN r1 r2 r=0 IL r1 r2 r3 r=0 IT r1 r2 r3 r=0 JP r1 r2 r3 r=0 KR r1 r2 r3 r=0 MY r1 r2 Eigen value 0.05 0.03 0.02 0.00 0.04 0.01 0.00 0.04 0.01 0.00 0.07 0.06 0.01 0.00 0.05 0.05 0.01 0.00 0.03 0.03 0.00 0.91 0.05 0.01 0.01 0.06 0.03 0.02 0.01 0.05 0.03 0.01 0.00 0.07 0.04 0.01 0.00 0.05 0.03 0.00 Trace 47.31 24.70 10.31 0.02 21.47 6.38 1.17 18.27 5.25 1.39 46.53 22.52 4.11 0.00 44.62 24.83 6.44 0.63 21.03 10.04 0.02 581.99 17.18 5.73 2.17 46.69 21.68 11.02 2.33 41.17 19.22 5.94 0.85 39.91 16.71 3.86 1.15 24.21 9.69 1.16 p-value 0.06 0.17 0.26 0.88 0.33 0.65 0.28 0.55 0.78 0.24 0.07 0.27 0.89 0.99 0.10 0.17 0.64 0.43 0.36 0.28 0.89 0.00 0.63 0.73 0.14 0.06 0.32 0.21 0.13 0.18 0.48 0.70 0.36 0.23 0.66 0.91 0.28 0.19 0.31 0.28 Eigen value 0.05 0.03 0.02 0.00 0.04 0.01 0.00 0.04 0.01 0.00 0.07 0.06 0.01 0.00 0.05 0.05 0.01 0.00 0.03 0.03 0.00 0.91 0.05 0.01 0.01 0.06 0.03 0.02 0.01 0.05 0.03 0.01 0.00 0.07 0.04 0.01 0.00 0.05 0.03 0.00 MaxEigen 22.60 14.39 10.29 0.02 15.09 5.21 1.17 13.02 3.86 1.39 24.01 18.40 4.11 0.00 19.79 18.39 5.81 0.63 11.00 10.02 0.02 564.81 11.44 3.57 2.17 25.01 10.66 8.70 2.33 21.95 13.27 5.09 0.85 23.21 12.84 2.71 1.15 14.52 8.54 1.16 p-value 0.19 0.33 0.19 0.88 0.28 0.72 0.28 0.45 0.87 0.24 0.13 0.12 0.85 0.99 0.36 0.12 0.64 0.43 0.65 0.21 0.89 0.00 0.60 0.90 0.14 0.10 0.68 0.31 0.13 0.22 0.43 0.73 0.36 0.16 0.47 0.96 0.28 0.32 0.33 0.28 I(1) Variable r lroilp lrstkp ly lroilp lrstkp ly lroilp lrstkp ly r lroilp lrstkp ly r lroilp lrstkp ly lroilp lrstkp ly r lroilp lrstkp ly r lroilp lrstkp ly r lroilp lrstkp ly r lroilp lrstkp ly lroilp lrstkp ly continued VAR VAR VAR VAR VECM VAR VAR VAR VAR VAR VAR Model

30 / The Energy Journal Appendix Table 2-1.  Results of the Johansen Cointegration Test (lroilp, r, lrstkp, ly model) (continued)
H0 r=0 MX r1 r2 r3 r=0 NL r1 r2 r=0 NO r1 r2 r=0 PH r1 r2 r=0 PT r1 r2 r3 r=0 ZA r1 r2 r3 r=0 ES r1 r2 r=0 GB r1 r2 r3 r=0 US r1 r2 r3 r=0 TW r1 r2 r3 Eigen value 0.12 0.06 0.04 0.01 0.04 0.02 0.00 0.07 0.04 0.02 0.07 0.05 0.02 0.13 0.06 0.01 0.01 0.04 0.04 0.01 0.00 0.06 0.01 0.01 0.09 0.04 0.02 0.00 0.05 0.04 0.02 0.00 0.11 0.10 0.02 0.00 Trace 57.82 26.17 10.54 1.47 26.35 10.33 0.68 51.94 23.49 7.98 38.52 19.44 5.90 41.80 15.50 3.79 1.38 37.17 21.41 6.01 0.71 27.54 6.89 2.32 56.93 20.98 7.18 0.85 49.34 27.39 8.98 0.05 68.45 34.96 4.84 0.01 p-value 0.00 0.12 0.24 0.23 0.12 0.26 0.41 0.01 0.10 0.25 0.13 0.26 0.47 0.16 0.75 0.92 0.24 0.34 0.33 0.69 0.40 0.09 0.59 0.13 0.01 0.36 0.56 0.36 0.04 0.09 0.37 0.83 0.00 0.01 0.83 0.93 Eigen value 0.12 0.06 0.04 0.01 0.04 0.02 0.00 0.07 0.04 0.02 0.07 0.05 0.02 0.13 0.06 0.01 0.01 0.04 0.04 0.01 0.00 0.06 0.01 0.01 0.09 0.04 0.02 0.00 0.05 0.04 0.02 0.00 0.11 0.10 0.02 0.00 MaxEigen 31.65 15.63 9.07 1.47 16.02 9.65 0.68 28.45 15.51 7.98 19.09 13.54 5.90 26.30 11.71 2.42 1.38 15.76 15.40 5.30 0.71 20.65 4.56 2.32 35.95 13.80 6.33 0.85 21.95 18.41 8.93 0.05 33.49 30.12 4.83 0.01 p-value 0.01 0.25 0.28 0.23 0.22 0.24 0.41 0.02 0.17 0.25 0.30 0.29 0.47 0.07 0.58 0.98 0.24 0.69 0.26 0.70 0.40 0.06 0.80 0.13 0.00 0.38 0.57 0.36 0.22 0.12 0.29 0.83 0.01 0.00 0.76 0.93 I(1) Variable r lroilp lrstkp ly lroilp lrstkp ly lroilp lrstkp ly lroilp lrstkp ly r lroilp lrstkp ly r lroilp lrstkp ly lroilp lrstkp ly r lroilp lrstkp ly r lroilp lrstkp ly r lroilp lrstkp ly VECM VAR VECM VAR VAR VAR VAR VECM VAR VECM Model

Note: Trace = Trace statistics; Max-Eigen= Maximum eigenvalue Statistics; VAR=Vector Autoregressive Model; VECM=Vector Error Correction Model. r=number of cointegating relation(s).

Factors Affecting an Economys Tolerance / 31 Appendix Table 2-2.  Results of the Johansen Cointegration Test (v , r, lrstkp, ly model)
H0 r=0 AT r1 r2 r=0 CA DK FI r1 r=0 r1 r=0 r1 r2 r=0 FR r1 r2 r=0 IN IL r1 r=0 r1 r2 r=0 IT r1 r2 r=0 JP r1 r2 r=0 KR r1 r2 r=0 MY MX r1 r=0 r1 r2 r=0 NL NO r1 r=0 r1 Eigen value 0.05 0.02 0.01 0.05 0.02 0.03 0.02 0.07 0.04 0.01 0.04 0.03 0.00 0.03 0.00 0.90 0.03 0.00 0.03 0.02 0.00 0.03 0.01 0.00 0.07 0.01 0.00 0.03 0.00 0.11 0.02 0.00 0.03 0.00 0.03 0.02 Trace 32.93 11.65 2.90 24.61 7.87 19.27 7.62 38.88 17.37 3.92 27.73 12.61 0.25 11.31 0.99 556.49 7.96 0.58 21.74 10.14 0.18 17.22 5.74 1.46 29.23 5.17 0.76 8.75 0.76 34.55 5.10 0.65 13.79 0.75 19.89 6.73 p-value 0.34 0.84 0.89 0.07 0.26 0.27 0.28 0.12 0.39 0.75 0.09 0.13 0.62 0.19 0.32 0.00 0.47 0.45 0.31 0.27 0.68 0.62 0.73 0.23 0.06 0.79 0.38 0.39 0.38 0.01 0.80 0.42 0.09 0.39 0.23 0.37 Eigen value 0.05 0.02 0.01 0.05 0.02 0.03 0.02 0.07 0.04 0.01 0.04 0.03 0.00 0.03 0.00 0.90 0.03 0.00 0.03 0.02 0.00 0.03 0.01 0.00 0.07 0.01 0.00 0.03 0.00 0.11 0.02 0.00 0.03 0.00 0.03 0.02 MaxEigen 21.28 8.75 2.90 16.74 7.87 11.64 7.62 21.50 13.46 3.92 15.12 12.36 0.25 10.32 0.99 548.53 7.38 0.58 11.60 9.97 0.18 11.48 4.28 1.46 24.06 4.41 0.76 7.99 0.76 29.46 4.45 0.65 13.04 0.75 13.15 6.73 p-value 0.18 0.75 0.89 0.12 0.26 0.45 0.28 0.17 0.29 0.75 0.28 0.10 0.62 0.19 0.32 0.00 0.45 0.45 0.59 0.21 0.68 0.60 0.83 0.23 0.02 0.81 0.38 0.38 0.38 0.00 0.81 0.42 0.08 0.39 0.32 0.37 I(1) Variable r lrstkp ly lrstkp ly lrstkp ly r lrstkp ly r lrstkp ly lrstkp ly r lrstkp ly r lrstkp ly r lrstkp ly r lrstkp ly lrstkp ly r lrstkp ly lrstkp ly lrstkp ly continued VAR VAR VECM VAR VECM VAR VAR VECM VAR VAR VAR VAR VAR VAR Model

32 / The Energy Journal Appendix Table 2-2.  Results of the Johansen Cointegration Test (v , r, lrstkp, ly model) (continued)
H0 r=0 PH PT r1 r=0 r1 r2 r=0 ZA r1 r2 r=0 ES GB r1 r=0 r1 r2 r=0 US r1 r2 r=0 TW r1 r2 Eigen value 0.04 0.01 0.10 0.02 0.01 0.04 0.03 0.00 0.05 0.00 0.07 0.02 0.00 0.05 0.02 0.00 0.11 0.05 0.00 Trace 14.72 3.92 26.43 4.89 1.00 28.16 12.75 1.90 19.53 0.39 38.31 8.46 0.85 28.09 7.03 0.07 45.74 14.47 0.49 p-value 0.60 0.75 0.12 0.82 0.32 0.08 0.12 0.17 0.01 0.53 0.00 0.42 0.36 0.08 0.57 0.79 0.00 0.07 0.48 Eigen value 0.04 0.01 0.10 0.02 0.01 0.04 0.03 0.00 0.05 0.00 0.07 0.02 0.00 0.05 0.02 0.00 0.11 0.05 0.00 MaxEigen 10.80 3.92 21.54 3.89 1.00 15.41 10.84 1.90 19.14 0.39 29.85 7.61 0.85 21.06 6.95 0.07 31.27 13.98 0.49 pvalue 0.53 0.75 0.04 0.87 0.32 0.26 0.16 0.17 0.01 0.53 0.00 0.42 0.36 0.05 0.49 0.79 0.00 0.06 0.48 I(1) Variable lrstkp ly r lrstkp ly r lrstkp ly lrstkp ly r lrstkp ly r lrstkp ly r lrstkp ly VECM VAR VECM VECM VAR VAR Model VAR

Note: Trace=Trace statistics; Max-Eigen= Maximum eigenvalue Statistics; VAR=Vector Autoregressive Model; VECM=Vector Error Correction Model. r=number of cointegating relation(s).

Factors Affecting an Economys Tolerance / 33 Appendix Table 3-1. Results of the Tsay Nonlinear C(d) Test (Dlroilp, Dr, Dlrstkp, Dly model)
d=1 AT 269.19 (0.09) CA 320.89 (0.00) DK 431.45 (0.00) FI 258.04 (0.20) FR 297.63 (0.01) IN 383.54 (0.00) IL 258.63 (0.25) IT 260.79 (0.17) JP 324.23 (0.00) KR 307.72 (0.00) MY 301.18 (0.00) MX 325.36 (0.00) NL 283.30 (0.03) NO 292.30 (0.02) PH 333.42 (0.00) PT 314.44 (0.00) ZA 302.90 (0.00) ES 257.33 (0.21) GB 269.14 (0.13) US 293.75 (0.01) TW 282.37 (0.05) d=2 281.24 (0.03) 292.10 (0.01) 280.01 (0.04) 310.05 (0.00) 405.93 (0.00) 350.19 (0.00) 323.55 (0.00) 337.50 (0.00) 333.89 (0.00) 355.52 (0.00) 405.25 (0.00) 345.79 (0.00) 241.84 (0.45) 349.14 (0.00) 451.17 (0.00) 297.94 (0.01) 274.57 (0.06) 316.11 (0.00) 354.96 (0.00) 324.70 (0.00) 331.58 (0.00) d=3 253.69 (0.26) 283.83 (0.03) 277.76 (0.05) 283.19 (0.03) 300.11 (0.01) 293.87 (0.01) 284.76 (0.04) 241.84 (0.45) 287.43 (0.02) 275.08 (0.06) 285.23 (0.02) 325.12 (0.00) 266.23 (0.12) 291.53 (0.02) 283.34 (0.03) 284.36 (0.03) 242.27 (0.45) 284.65 (0.03) 263.88 (0.18) 375.66 (0.00) 282.87 (0.04) d=4 206.65 (0.94) 339.81 (0.00) 243.40 (0.43) 248.94 (0.33) 266.78 (0.11) 305.30 (0.00) 254.49 (0.31) 215.22 (0.87) 236.84 (0.55) 260.09 (0.18) 352.91 (0.00) 361.74 (0.00) 237.33 (0.54) 279.07 (0.06) 258.82 (0.19) 278.57 (0.04) 230.22 (0.66) 272.64 (0.07) 296.23 (0.01) 322.59 (0.00) 278.37 (0.06) d=5 253.98 (0.26) 285.03 (0.02) 263.83 (0.14) 235.10 (0.58) 284.63 (0.03) 284.50 (0.03) 361.78 (0.00) 256.95 (0.22) 274.74 (0.06) 277.27 (0.05) 350.43 (0.00) 413.80 (0.00) 250.81 (0.30) 260.35 (0.23) 255.52 (0.23) 296.79 (0.01) 338.26 (0.00) 275.78 (0.06) 282.22 (0.05) 281.59 (0.03) 293.33 (0.02) d=6 269.50 (0.09) 296.26 (0.01) 397.10 (0.00) 274.53 (0.06) 292.70 (0.01) 337.83 (0.00) 355.76 (0.00) 310.17 (0.00) 261.23 (0.17) 366.03 (0.00) 374.32 (0.00) 343.51 (0.00) 261.11 (0.17) 344.10 (0.00) 310.12 (0.00) 330.77 (0.00) 253.77 (0.26) 310.16 (0.00) 268.06 (0.14) 233.69 (0.60) 308.40 (0.00) d=7 248.54 (0.34) 313.26 (0.00) 262.59 (0.15) 261.73 (0.16) 228.20 (0.70) 357.42 (0.00) 288.59 (0.03) 237.34 (0.54) 311.94 (0.00) 353.74 (0.00) 253.18 (0.27) 353.87 (0.00) 217.27 (0.85) 286.88 (0.03) 409.29 (0.00) 265.62 (0.12) 304.78 (0.00) 247.01 (0.36) 253.97 (0.32) 268.42 (0.10) 296.57 (0.01) d=8 257.82 (0.20) 318.69 (0.00) 312.80 (0.00) 251.92 (0.29) 217.99 (0.84) 280.58 (0.04) 268.47 (0.14) 236.56 (0.55) 274.82 (0.06) 327.97 (0.00) 263.13 (0.15) 271.48 (0.11) 226.90 (0.72) 219.59 (0.87) 315.74 (0.00) 259.55 (0.18) 224.58 (0.75) 267.13 (0.11) 291.87 (0.02) 260.67 (0.17) 300.97 (0.01) d=9 302.08 (0.00) 263.85 (0.14) 467.03 (0.00) 281.34 (0.03) 297.62 (0.01) 341.07 (0.00) 326.23 (0.00) 233.06 (0.61) 221.74 (0.80) 300.06 (0.01) 296.45 (0.01) 315.23 (0.00) 285.29 (0.02) 301.61 (0.01) 398.33 (0.00) 328.32 (0.00) 269.13 (0.10) 215.40 (0.87) 281.27 (0.05) 307.00 (0.00) 322.96 (0.00) d=10 253.02 (0.27) 290.06 (0.01) 294.55 (0.01) 281.48 (0.03) 282.65 (0.03) 351.37 (0.00) 371.55 (0.00) 224.61 (0.75) 264.38 (0.13) 273.40 (0.07) 293.60 (0.01) 292.80 (0.02) 225.93 (0.73) 268.70 (0.13) 341.04 (0.00) 321.32 (0.00) 250.99 (0.30) 235.39 (0.57) 283.64 (0.04) 282.91 (0.03) 259.92 (0.23) d=11 238.64 (0.51) 225.83 (0.74) 237.72 (0.53) 263.71 (0.14) 235.51 (0.57) 306.16 (0.00) 323.15 (0.00) 260.98 (0.17) 205.43 (0.95) 227.47 (0.71) 271.59 (0.08) 252.28 (0.34) 254.86 (0.24) 292.44 (0.02) 339.47 (0.00) 267.65 (0.11) 261.84 (0.16) 215.73 (0.87) 255.33 (0.30) 281.80 (0.03) 233.13 (0.68) d=12 256.87 (0.22) 267.50 (0.11) 234.46 (0.59) 256.99 (0.22) 255.51 (0.23) 253.48 (0.26) 325.98 (0.00) 230.88 (0.65) 278.40 (0.04) 304.89 (0.00) 287.72 (0.02) 289.35 (0.02) 254.31 (0.25) 232.69 (0.69) 234.75 (0.58) 263.93 (0.14) 265.80 (0.12) 265.57 (0.12) 216.97 (0.89) 284.41 (0.03) 245.07 (0.47)

34 / The Energy Journal Appendix Table 3-2. Results of the Tsay Nonlinear C(d) Test (Dlroilp, Dr, Dlrstkp, Dly model)
d=1 248.88 (0.33) CA 323.78 (0.00) DK 418.93 (0.00) FI 284.87 (0.02) FR 283.18 (0.03) IN 299.57 (0.01) IL 223.93 (0.82) IT 218.61 (0.84) JP 262.80 (0.15) KR 315.91 (0.00) MY 291.54 (0.01) MX 320.24 (0.00) NL 216.81 (0.86) NO 289.87 (0.02) PH 374.07 (0.00) PT 284.02 (0.03) ZA 252.19 (0.28) ES 275.13 (0.08) GB 310.06 (0.00) US 233.71 (0.60) TW 312.75 (0.00) AT d=2 306.30 (0.00) 286.92 (0.02) 241.24 (0.47) 298.64 (0.01) 281.56 (0.03) 293.46 (0.01) 298.09 (0.01) 277.12 (0.05) 316.70 (0.00) 288.53 (0.03) 341.98 (0.00) 329.19 (0.00) 254.84 (0.24) 284.40 (0.03) 410.33 (0.00) 278.28 (0.05) 274.06 (0.06) 287.12 (0.03) 338.53 (0.00) 283.22 (0.03) 287.04 (0.03) d=3 242.79 (0.44) 273.90 (0.07) 298.29 (0.01) 327.66 (0.00) 335.69 (0.00) 328.52 (0.00) 297.32 (0.01) 234.44 (0.59) 262.60 (0.15) 283.64 (0.04) 324.59 (0.00) 360.80 (0.00) 240.61 (0.48) 262.00 (0.16) 284.47 (0.03) 257.55 (0.21) 281.36 (0.03) 295.83 (0.01) 197.31 (0.99) 288.78 (0.02) 315.66 (0.00) d=4 262.72 (0.15) 264.80 (0.13) 235.46 (0.57) 278.39 (0.04) 281.67 (0.03) 312.76 (0.00) 296.06 (0.01) 211.90 (0.90) 241.61 (0.46) 265.20 (0.17) 347.30 (0.00) 369.97 (0.00) 261.15 (0.17) 229.42 (0.68) 241.57 (0.46) 286.72 (0.02) 316.33 (0.00) 254.28 (0.31) 269.51 (0.13) 300.43 (0.00) 297.36 (0.01) d=5 279.77 (0.04) 279.44 (0.04) 235.79 (0.56) 243.13 (0.43) 274.92 (0.06) 313.81 (0.00) 288.55 (0.03) 241.36 (0.46) 325.53 (0.00) 234.27 (0.66) 338.37 (0.00) 413.49 (0.00) 223.08 (0.78) 215.26 (0.87) 237.83 (0.53) 253.61 (0.26) 236.90 (0.54) 265.68 (0.16) 262.69 (0.20) 244.36 (0.41) 308.11 (0.00) d=6 245.85 (0.38) 299.11 (0.01) 255.33 (0.24) 267.08 (0.11) 302.04 (0.00) 372.20 (0.00) 302.49 (0.01) 274.14 (0.06) 253.14 (0.27) 296.28 (0.01) 285.88 (0.02) 295.30 (0.01) 245.97 (0.38) 341.19 (0.00) 274.42 (0.06) 336.76 (0.00) 237.69 (0.53) 251.15 (0.36) 248.07 (0.42) 230.69 (0.66) 306.92 (0.00) d=7 278.60 (0.04) 294.64 (0.01) 260.36 (0.18) 316.64 (0.00) 275.64 (0.06) 325.36 (0.00) 302.07 (0.01) 228.76 (0.69) 282.72 (0.03) 276.74 (0.07) 272.87 (0.07) 341.85 (0.00) 203.39 (0.96) 296.15 (0.01) 377.63 (0.00) 285.60 (0.02) 263.92 (0.14) 236.26 (0.63) 254.31 (0.31) 247.70 (0.35) 310.78 (0.00) d=8 256.98 (0.22) 327.22 (0.00) 256.25 (0.22) 207.89 (0.93) 197.41 (0.98) 323.47 (0.00) 254.46 (0.31) 234.32 (0.59) 256.60 (0.22) 295.88 (0.01) 331.74 (0.00) 286.30 (0.03) 207.79 (0.93) 270.32 (0.09) 366.60 (0.00) 269.61 (0.09) 197.39 (0.98) 313.45 (0.00) 283.01 (0.04) 227.22 (0.71) 288.60 (0.03) d=9 307.11 (0.00) 301.73 (0.00) 250.12 (0.31) 265.85 (0.12) 286.00 (0.02) 341.80 (0.00) 300.21 (0.01) 250.53 (0.31) 258.63 (0.20) 258.47 (0.25) 328.91 (0.00) 311.16 (0.00) 279.36 (0.04) 281.67 (0.03) 403.58 (0.00) 307.66 (0.00) 253.41 (0.26) 245.20 (0.47) 262.06 (0.20) 242.35 (0.45) 332.28 (0.00) d=10 285.37 (0.02) 296.67 (0.01) 235.01 (0.58) 229.90 (0.67) 316.42 (0.00) 344.09 (0.00) 359.37 (0.00) 288.19 (0.02) 237.79 (0.53) 228.65 (0.75) 262.98 (0.15) 291.42 (0.02) 287.46 (0.02) 240.82 (0.47) 257.73 (0.21) 288.10 (0.02) 242.74 (0.44) 221.76 (0.84) 265.76 (0.16) 329.60 (0.00) 269.14 (0.13) d=11 268.19 (0.10) 242.78 (0.44) 238.20 (0.52) 227.55 (0.71) 224.09 (0.76) 332.04 (0.00) 343.48 (0.00) 273.83 (0.07) 203.98 (0.96) 246.47 (0.44) 300.54 (0.00) 269.11 (0.13) 248.24 (0.34) 263.76 (0.14) 222.49 (0.78) 301.94 (0.00) 268.40 (0.10) 238.08 (0.59) 267.61 (0.14) 294.99 (0.01) 251.95 (0.35) d=12 271.73 (0.08) 331.44 (0.00) 219.39 (0.83) 265.63 (0.12) 307.24 (0.00) 277.16 (0.05) 309.94 (0.00) 244.65 (0.40) 282.42 (0.03) 294.50 (0.01) 331.11 (0.00) 270.08 (0.12) 279.51 (0.04) 277.92 (0.05) 254.74 (0.25) 282.65 (0.03) 297.59 (0.01) 280.87 (0.05) 302.59 (0.01) 303.81 (0.00) 277.19 (0.07)

Note: d = the delay periods of threshold variable; numbers inside ( ) are p-value.

High Frequency Export and Price Responses in the Ontario Electricity Market
Angelo Melino* and Nash Peerbocus** Export responses to unanticipated price shocks can be a key contributing factor to the rapid mean reversion of electricity prices, a phenomenon often seen in electricity markets. In this paper, we use event analysis to demonstrate how hourly export transactions respond to negative supply shocks in the Ontario electricity market. Although event analysis has been used for many years in other applications, particularly finance, to our knowledge this is the first time that this technique has been applied to price response analysis in the electricity market. The analysis clearly demonstrates the sensitivity of export volume to price changes, and more generally, the responses of prices and quantities to an unexpected supply shock. Introduction In the discussion of competitive reforms in electricity markets, demand response is often mentioned as one of the key requirements that can facilitate the effective functioning of the market. In this context, demand response entails reduction of electricity consumption in relation to anticipated or realised rapid price increases. Contemporary demand response programs focus on demand response from internal domestic consumers.1 But in todays interconnected electricity markets with extensive interregional trades, a complete analysis of demand response must also include the behavioural response of export demand
The Energy Journal, Vol. 29, No. 4. Copyright 2008 by the IAEE. All rights reserved. * ** Department of Economics, University of Toronto and RCEA. Corresponding author: Senior Economist, Market Evolution, Analysis and Research, Independent Electricity System Operator; Tel 905-855-6174. Email; nash.peerbocus@ieso.ca.

The authors wish to thank three anonymous referees whose suggestions improved the paper. 1. The US Department of Energys report to Congress on the benefits of demand response described various dimensions of demand response. Export demand as a component of demand response is not discussed in the report. See Benefits of Demand Response in Electricity Markets and Recommendations for achieving them; a report to the US Congress. US Department of Energy, February, 2006.

35

36 / The Energy Journal to price changes. Export responses to unanticipated price shocks can be a key contributing factor to the rapid mean reversion of electricity prices, a phenomenon often seen in electricity markets. In times of stress, system operators have a number of tools at their disposal to manage system reliability. In the presence of transparent price signals, the response of price- sensitive export demand can be another critical and reliability-enhancing market mechanism that can contribute to stabilisation of the interconnected electrical grid. In this paper, we use event analysis to demonstrate how hourly export transactions respond to negative supply shocks in the Ontario electricity market. Although event analysis has been used for many years in other applications, particularly finance, to our knowledge this is the first time this technique has been applied to price response analysis in the electricity market. The analysis clearly demonstrates the negative relationship between export volume and high prices. Our work also contributes to the small but important literature on the integration of electricity markets. So far, these papers have focused exclusively on price differentials and whether such differentials are small enough to be consistent with arbitrage bounds. Our paper augments this literature by showing direct evidence of the quantity responses that are implicitly assumed in previous work. The paper is organised as follows. Section I provides an overview of the Ontario electricity market. In section II, we describe our empirical methodology. Section III and IV discuss the results of the event analysis. Finally, section V concludes and discusses some policy implications. 1. The Ontario Electricity Market In December 2004, the Government of Ontario passed the Electricity Restructuring Act, 2004 which reorganized the province's electricity sector effectively creating a hybrid electricity market: a competitive wholesale market coexisting with regulated prices in parts of the electricity sector. The Ontario market operates under a single market clearing price auction. Suppliers submit offers to sell energy. Domestic demand consists of metered market participants, wholesale customers, and low volume and designated consumers. The first two pay the hourly spot market prices while the others pay fixed rates set by the Ontario Energy Board. In the Ontario market, domestic demand plus export demand constitute total market demand. The intersection of market demand with the least cost supply offer determines the market clearing price.2 The market clearing price (MCP) is calculated every five minutes of the day. The average of these five minute MCPs across the hour is called the hourly Ontario electricity price (HOEP) (henceforth the Ontario price). The spot market clears each hour with a price and a dispatch schedule-a list of what each supplier will provide to the system over the next hour and a list of
2. In fact the IESO operates an energy and an operating reserve market. The price is determined through the joint optimisation of the energy and reserve markets.

High Frequency Export and Price Responses / 37 which demanders will receive electricity.3 However bids and offers are submitted in advance. To facilitate planning and to provide market participants with useful information, the system operator produces a number of pre-dispatch schedules with associated price projections. The critical schedule is the 3 hour ahead predispatch schedule because this schedule provides the latest information available to market participants to make their trades before the market closes. In other words all offers and bids must be submitted two hours ahead of the dispatch hour.4 Market participants who export electricity from Ontario buy the electricity at the real-time hourly spot market price, i.e. the HOEP. Market participants that import power into Ontario are given a price guarantee for their energy. This is done largely for reliability reasons. Importers typically use the 3-hour ahead pre-dispatch price projections to place their trades. Thus, whereas exporters are sensitive to the spot market prices, importers are hedged from any downside realtime price risk. In other words if the real-time spot price turns out to be lower than the price at which they offered to sell, they do not incur the loss on the transaction. They still get their offer price. The 2007 Ontario energy supply mix consisted of 51 per cent nuclear, 21 per cent hydro, 18 per cent coal, 9 per cent oil, gas and other fuels, and wind at 1 per cent. Ontario is connected to five other control areas or jurisdictions: New York, Michigan,5 Quebec, Minnesota, and Manitoba. Table 1 shows the total annual trade volumes from Ontario since 2003. Annual net export volume increased from negative 4.1 TWh (net importer) in 2003 (2.7 per cent of the 2003 Ontario demand) to 5.1 TWh in 2007 (3.4 per cent of the 2007 Ontario demand). Table 1. Ontario Electricity Trade Volumes in Terawatt-hours (TWh) 6
2003 2004 2005 2006 2007 Import(TWh) 10.4 9.8 11 6.2 7.2 Export(TWh) 6.3 9.5 10.2 11.4 12.3 Net Export -4.1 -0.3 -0.8 5.2 5.1

The New York market is the largest export market for Ontario. Approximately 80 to 85 per cent of export trades from Ontario flow to the New York market. This study focuses on export trades to the New York market.
3. A second algorithm chooses suppliers in a way that respects physical constraints on the system, such as transmission capacity and ramping rates. A set of side payments is made to compensate for the differences in the allocations from the market clearing ones. Although important for many purposes, the outcome of this second algorithm is not of interest for this paper. 4. Under exceptional circumstances identified in the Market Rules, the IESO can allow offers within the two hour window. 5. The Michigan control area is part of the MISO organization (Midwest Independent Transmission System Operator). 6. Data obtained from the IESO.

38 / The Energy Journal 1.1 Arbitrage Activity between Ontario and New York The rational profit-maximising trader always buys power from the low price market and sells it to the high price market. This arbitrage activity continues until all profit opportunities are eroded. An unexpected negative supply shock in the Ontario market places upward pressure on the Ontario price. As the Ontario price increases relative to the New York price,7 traders schedule less export from Ontario to New York. Initially and immediately after the negative supply shock, the price gap between Ontario and New York deviates from its pre-shock equilibrium level. However as the New York market receives less export from Ontario, there is initial upward pressure on the price in New York. This is because in the short-run fast-ramping suppliers are required to meet demand. The price in New York rises to the point where the price gap between Ontario and New York narrows down to its pre-shock trend level. The ultimate impact on the New York price depends on the supply elasticity in New York. In the short-run the ability of traders to promptly respond to the supply shocks depends on a number of factors including inter-jurisdictional trade protocols, transmission congestion, expectations about short-term future price movements and general uncertainty about the duration of the supply shocks. In the long-run fundamental market characteristics determine the equilibrium price gap between the two markets. 2. Methodology of Event Analysis Event analysis is widely used in finance and other applications. However to our knowledge this is the first application of event analysis to electricity market analysis. We identify episodes with large observable shocks to supply and then trace out the responses of prices and quantities to these shocks. In this section we provide a detailed explanation of how event analysis is used in this study. Figure 1. Stylized Equilibrium in Electricity Market

7. The New York price used in our analysis is the real-time New York price in zone west, the zone adjacent to the Ontario market near Niagara Falls.

High Frequency Export and Price Responses / 39 Figure 1 provides the well known intuition about how to estimate demand responses given equilibrium prices and quantities. Suppose we have an idealized electricity market with a market demand schedule MD and supply schedule S. Equilibrium is given by their intersection at the point A. The demand schedule itself is the horizontal sum of the export schedule (X) and the domestic demand schedule (DD). Suppose we could exogenously shift the supply curve from S to S* holding the components of demand constant. The shift from A to the new equilibrium B would trace out the slope of the demand curve. At the same time, the movement from a to b would trace out the slope of the export demand curve. Unanticipated nuclear outages provide an excellent example of an exogenous shift in the supply curve. The outages occur for technical reasons and are unrelated to market prices and quantities. In the idealized market, we could learn all we want from a single shift in the supply curve (or, at most, from single shifts of a given size if there are non-linearities in demand). However life is more complicated with data from real markets. There are always other observable and unobservable shocks hitting the market at the same time as the nuclear outages, so both the demand and supply curves are shifting before and after the nuclear outages. And with high frequency data, we have to face up to the fact that responses take time. Rather than look at a single event, we look at the response path of prices and quantities to many unanticipated nuclear outages and then average these responses. Because they shift the supply curve but are unrelated to movements in the demand curve, unanticipated nuclear outages could be used as instruments to identify the slope of the demand curve. But they are also suited to an event analysis which provides some advantages. In particular, because they are random and unpredictable, we do not have to explicitly model the demand curve. This is particularly useful with high frequency data where expectations and costs of adjustments are important but difficult to model well. In addition, although calendar effects( time of the day, day of the week, month of the year, season), weather, etc. are obviously important shifters of the demand curve, the randomness in the timing of the nuclear outages means that we do not have to explicitly model and control for these demand shifters well or even at all. The randomness in the timing of nuclear outages eliminates the usual left out variable bias. To illustrate these various points, and to provide some intuition on the advantages of an event analysis compared to a structural modeling approach, it is useful to briefly consider the latter. Suppose we posit a demand equation in period h of the form g + Xd l + ud qh = bph + Y d h h h (1)

where q denotes the quantity of electricity demanded, p denotes price, Y d is a vector of predetermined variables (such as lagged prices and quantities, prices in neighbouring jurisdictions, etc.), and Xdis a vector of strictly exogenous variables (weather, day of the week, and other seasonals). The allocation of variables into Y d and Xd depends on the orthogonality assumptions pertaining to each. By definition, ud is uncorrelated with current and lagged values of Y d , but future values of h h

40 / The Energy Journal Yd may respond to the current demand shock. Variables in Xd are uncorrelated at h all leads and lags with the demand shock. Without loss of generality, we can assume all the variables are demeaned and detrended. Let Nh0 denote a dummy variable that indicates the period in which an unplanned nuclear outage occurs. If we know the form of the demand equation, including all the components of Y d and Xd , and have the data, then it is relatively straightforward to estimate the parameters of the demand equation, using Nh0 as an instrument for ph. Premultiply both sides of the demand equation by Nh0 and average over the sample to obtain T1hN 0 q = T1hN 0 p + T1hN 0 Y d g + T1hN 0 Xd l + T1hN 0 ud (2) h h h h h h h h h h For the nuclear outage dummy to be a valid instrument, we require the maintained assumption that plim T1hN 0 ud = 0 h h This is not a testable hypothesis, but it is plausible if we believe that shocks to demand do not contribute to the probability of an unplanned nuclear outage. However, even without a complete specification of the demand curve, we can estimate the impact demand elasticity parameter, , by exploiting the property that unplanned nuclear outages occur randomly. If unplanned nuclear outages occur for physical reasons that are unrelated to Y d and Xd , then all but h h the first term on the right-hand side of eq [2] will also converge to zero. While plausible a priori, in principle, this is testable to the extent that we have data on the components of Y d and Xd . If unplanned nuclear outages occur randomly, then h h on average the movements of prices and quantities that occur during the periods with an unplanned nuclear outage trace out the demand curve. So a consistent estimator of will be given by b = T1hN 0 q / T1hN 0 p h h h h In fact, the average movements of prices and quantities that occur before the components of Yd respond to the nuclear outage are also movements along the demand curve and can be used to estimate . If we use the difference in prices and quantities before and after the nuclear outage to proxy for the difference in the demeaned and detrended movements, we get the estimate shown in Table 2 below for the short-run elasticity parameter . Although the immediate response of prices and quantities are movements along the demand curve, in subsequent periods we will observe only equilibrium responses that incorporate the reactions of suppliers and system operators, as well as electricity demanders, and the dynamic feedback through the predetermined variables. With a fully specified dynamic model of demand and supply, and a speci-

High Frequency Export and Price Responses / 41 Table 2. Short-run Elasticity Estimates
Export HOEP Mean Before 1296 MW $54 0.84 Value After 1062 MW $66

Elasticity

fication of how the market and the system operator co-ordinate these two schedules, we could trace out the dynamic response of prices and quantities to the unplanned nuclear outage shock. However, just as it allows us to uncover the impact responses of the nuclear shutdown on prices and quantities, event analysis allows us to estimate these dynamic responses without relying on a fully specified model. Event analysis focuses on six basic steps organised around three phases. Table 1 below provides details of the three phases and the six steps. The first phase is the definition phase and it consists of the first three steps. The first step is to define the event to the market. In this analysis the event is the occurrence of a forced nuclear outage. The second step is to define the event time. The event time is the time when the event hits the market. The third step is to define the study period which is the period before (i.e. pre-event period) and after the event (i.e. the post-event period). The pre-event period is used to establish the normal level of the variables absent the disturbance to the market. In this analysis the study period is the ten hours before and the ten hours after the disturbance. The second phase is the measurement phase and it involves the fourth and the fifth step. The fourth step is to define a baseline value for the variable in the pre-event period. In this analysis we use the mean value of the variable in the pre-event period as the baseline value. The fifth step is to define the deviation metric in the post-event period. The metric we use is the difference between the level of the variable in the post-event period and the baseline value of the variable in the pre-event period. The last phase includes the sixth step which is the analysis of the trend in the deviation metric relative to the baseline value used. We outline the implementation of this methodology below. Table 3. Application of Event Analysis in Export Response Analysis
Definition Phase I Phase Phase II Phase III Measurement Phase Analysis Phase Event Definition The occurrence of an unanticipated nuclear outage Event Time The interval in which the outage occurred Study Period The ten hours before and the hours after the event Baseline Metric The mean of the variable in the ten hours before the outage

Deviation Metric The distance between the variable and the baseline value Metric Comparison Comparison of the baseline and the post-deviation metric

42 / The Energy Journal 2.1 Timeline of an Unanticipated Nuclear Outage Before describing our empirical results, it is useful to illustrate the timeline of an unanticipated nuclear outage, and the behaviour of prices and quantities in response to it, by looking in detail at one particular incident, chosen more or less at random. Unfortunately, the date of this event cannot be given for confidentiality reasons. The nuclear generating station was forced out of service because of an unanticipated technical problem. This forced outage resulted in a substantial total generation loss of over 500MW. The chain of events is described below to give a sense of the exogeneity and randomness of the event. At 6:23 in the morning, staff at the nuclear plant informed the system operator that they were investigating a technical failure in a component of the system. At 6:44 the nuclear plant operator informed the system operator that the unit would reduce its output to zero. At 6:46 the nuclear unit was taken out of service. Therefore in a very short period of time (23 minutes) the unexpected technical outage forced the unit out of service. The Ontario price increased from $39 in delivery hour 6 to $213 in delivery hour 8. The amount of export transaction dropped from 1928 MW in delivery hour 6 to 959 MW in delivery hour 11. Import volume increased from 1160 MW in delivery hour 6 to 1630 MW in delivery 10. This single event clearly demonstrates the path of price and quantities that one expects to see following the unplanned nuclear outage. 2.2 Application of Event Analysis We analyse the path of prices and quantities before and after an unanticipated nuclear outage occurs in the Ontario market. We have a sample of 40 independent events over the period January 2005 to June 2007 where a nuclear unit was forced out of service. The nuclear outages were randomly distributed8 and were not driven by market prices. The generation loss in each case was unanticipated and it was greater than 400MW. A generation loss of this magnitude would have a sizeable impact on the HOEP. In short each outage was significant and its occurrence was exogenous. This scenario provides an ideal setting to conduct an event analysis. At time zero, the event hits the market. (Time zero is the hour in which the event occurs in the market). We used the ten hours before the forced outage occurred (t-1 to t-10) and the ten hours after the forced outage occurred (t +1 to t+10) to specify the study period. Next we calculate the baseline metric which in this case is the average value of the variable in the ten pre-event hours. The preevent deviation metric value is the difference between the pre-event value of the variable and the baseline metric. The post-deviation metric value is the difference between the value of the post-event variable and the baseline metric. The analytic8. Pearson Chi-square test using cells formed from days of the week or from hours of the day did not reject the null that the unexpected nuclear outages are uniformly distributed over the week. Results available on request from the authors.

High Frequency Export and Price Responses / 43 al steps are summarised in Table 3. Intuitively this methodology describes how the path of the variable deviates from its normal values over time. Our main focus is the response of export volume and the Ontario price but we also look at a variety of relevant prices and quantities that help us understand how the system adapts to the supply shock. The variables that we track are the Ontario price, the New York price in zone west, the Ontario-New York price difference, the volume of export from Ontario to New York, the volume of import from Michigan to Ontario, and the price inelastic Industrial Load.9 3. Results of the Event Analysis The results of the analysis are summarised in figures 2 to 7 below. Figures 2 to 6 shows the response paths for the Ontario price, Export volume, New York price, the Ontario-New York price gap and import volume from Michigan. We plot the estimated response paths (solid line) as well as the estimated response paths plus one standard deviation (dotted line). At time t+1, following the nuclear outage, the Ontario price rises sharply. As a consequence the price gap shown in Figure 5 deviates largely from its preevent level. This large gap is mainly driven by the large increase in the Ontario price at time t+1. The drop in export shown in Figure 3 that results in that hour (and for the next two hours until time t+3) may partly be the result of actions taken by the system operator.10 Import volume increased substantially from time t+2 onward as shown in figure 6. As the information is absorbed by the market, traders adjust their export bids to reflect their expectations about future prices (and profits). In the Ontario market traders at time t+1 can only change their export bids for time t+4 onwards. However at time t+1, traders have the ability to mitigate price risks by adjusting their transactions in the New York market for time t+2 onwards.11 This implies that some of the export reductions observed in t+2 and t+3 may be related to the implementation of this risk-mitigation strategy.

9. The price inelastic industrial load represents an aggregate of electricity consumption by several large industrial customers which pay the hourly Ontario price. Based on their historical profiles these industrial companies are largely price inelastic users of electricity. 10. The Ontario system operator, under the Market Rules, has the ability to recall or curtail export transactions for either system adequacy or reliability reasons. 11. In Ontario, inter-jurisdictional trades are fixed one hour ahead of real-time. The trader finalises her export bid two hours ahead of real-time i.e. this is the time when the Ontario market bid/offer window closes. The New York market on the other hand allows bidding up to 75 minutes before realtime. In addition the trader can adjust information about the trade (the NERC tag) in order to nullify the transaction half an hour before real-time. Thus a trader has options to cancel her import offer into New York (export from Ontario) after the Ontario market closes. If she does so, her action is viewed as an export transaction failure by the Ontario system operator. In the final communications between the two system operators, the transaction is then removed from both the Ontario market and the New York market. In this sense the trader, after seeing an Ontario price spike can mitigate her price risk by cancelling her trade in the New York market.

44 / The Energy Journal Figure 2. Ontario Price Path

Figure 3. Export Volume to New York

Figure 4. New York Price Path

High Frequency Export and Price Responses / 45 Figure 5. OntarioNew York Price Gap

Figure 6. Import Volume from Michigan

Figure 7. Price Inelastic Industrial Load

46 / The Energy Journal The largest reduction in export volume occurs at time t+4 and it primarily reflects the response of traders to the price shock within the current design of the Ontario market. In other words the volume of export at time t+4 reflects the action taken by traders in the Ontario market at time t+1. The response of the New York price exhibits an interesting path. The range of variation of the New York price seems similar to what it was before the nuclear outage. Within this range of variation, the largest impact seems to occur at time t+4. Recall that this is precisely the time when the largest export volume reduction from Ontario occurs. This supports the view that the two markets are linked. Were this not to be the case the New York price would be largely unchanged following the change in export volume. Also of interest is the behaviour of the Ontario price inelastic industrial load before and after the event as shown in figure 7. This large group of consumers effectively act as a control group for this study. There does not appear to be any reduction in electricity consumption in the immediate hours following the unanticipated nuclear outage. This is consistent with the economic characteristics of these consumers i.e. they are largely insensitive to the electricity price. Their consumption pattern is however in sharp contrast to the behaviour of price sensitive exporters as illustrated in figure 3. 4. C  omparison of Event Analysis and Regression Method How do our results using event analysis compare to more familiar regression methods? As a partial but useful exercise, in this section we investigate the behaviour of the HOEP using both methods. We choose to examine the HOEP because we have available an alternative forecast for this series that we can use as a proxy for what its path would like in the absence of a supply shock; we do not have access to comparable forecasts for the other prices and quantities. Unfortunately, the available data series for the HOEP price forecast covers only a subperiod of our sample. But even this partial exercise using only one series and a subsample is useful. We show that both event analysis and regression methods lead to very similar qualitative and quantitative estimates of the response of the HOEP to an unexpected nuclear outage. In other work for the IESO, we have constructed a day-ahead forecast for the HOEP. Prices for hours 1 to 24 are forecasted using information available as of 4pm on the previous day. The forecasted price series is constructed from 24 separate regressions that project the realized HOEP for that hour for the next day on a variety of information variables. The information variables used are (i) the IESOs pre-dispatch prices as of 4pm day-ahead (ii) the day-ahead market demand in Ontario , (iii) the day-ahead market supply cushion which is a measure of the excess energy available to meet demand and reserve requirements, (iv) the day-ahead price in zone west from the New York market, (iii) the day-ahead Michigan Hub price (iv) seasonal variables (e.g. day of the week, month of the year). The dayahead forecast model was estimated from June 2006 to July 2007.

High Frequency Export and Price Responses / 47 Unfortunately, because of the shorter sample period, and the fact that the model was estimated just for weekdays, we have only eight unanticipated nuclear outage events in our subsample. This makes it difficult to estimate accurately the response of the HOEP to the supply shock event, but we have enough events to provide a useful comparison between event analysis and regression method estimates of the response. Table 4 reports the regression results. As we would expect if the price forecast provides a good base value for the path of the HOEP, the coefficient on the day-ahead HOEP price forecast is insignificantly different from 1 and the estimated intercept is very close to zero. The coefficients on the variables Di estimate the deviation of the HOEP from the base path in hour i after the unanticipated nuclear outage. The estimated response from this regression model is qualitatively similar to the estimates obtained using event analysis on the whole sample. The price jumps by a significant amount in hour D0 and D1. Curiously, it is estimated to then fall back to its base value for two hours, but jump up again in hours D4-D7 before finally dissipating. Table 4. Estimates of Regression Coefficients
C FORECAST D0 D1 D2 D3 D4 D5 D6 D7 D8 D9 D10 Coefficient -1.15 1.02 14.95 21.90 7.12 0.73 2.73 13.13 14.02 11.15 8.88 2.34 3.25 Std. Error 0.50 0.01 5.36 5.36 5.36 5.36 5.36 5.36 5.36 5.36 5.36 5.36 5.36 t-Statistic -2.31 110.14 2.79 4.09 1.33 0.14 0.51 2.45 2.62 2.08 1.66 0.44 0.61

We ran the same regression including an additional ten dummy variables for the ten hours before the nuclear outage, as a test of the specification. Consistent with the hypothesis that the outages were unexpected, the coefficients on these additional variables are insignificantly different from zero. We also tried various specifications in which we made the coefficients on the variables D0-D10 vary with the size of the loss of generating capacity from the unanticipated nuclear outage. We found no statistical improvement in fit from this more general specification, probably reflecting the limitations of having only eight events in our subsample and a very narrow range for the size of the outages. To compare with the event analysis methodology, we repeated our procedure from Section III using the HOEP and the same subsample used for the regression estimates. The results are displayed in figure 8 along with the coefficients from D0-D10 from Table 4.

48 / The Energy Journal Figure 8. Comparison of Estimated Responses from Event Analysis with Estimated Coefficients from Regression Analysis

Using the same subsample, we see that both approaches lead to very similar estimates for the response of the HOEP to an unanticipated nuclear outage. This should provide some additional confidence in the event analysis results from Section III, which we were able to apply to more variables and for which we could use a longer sample period. 4.1 Electricity Market Integration One of the goals behind the reorganisation of the electricity industry is improvement in the short-run operational efficiency of electricity systems. Markets play a fundamental role in this process. In a deregulated market environment the price signal plays a critical role in allocating resources. Price differences between regions serve as strong signals for trading activities. Arbitragebounded trades enable price convergence between geographically dispersed markets and bring home the benefits of competition. Market integration can provide a measure of market performance. Our evidence that export moves with the inter-market price differential supports the hypothesis of market integration, corroborating the empirical findings in Woo et al. (1997), DeVany and Walls (1999), Bower (2002), Worthington et al. (2005) and Jamasb and Pollit (2005). The event analysis that we have done provides some new evidence on the integration of the New York and Ontario electricity markets. In fact the existence of arbitrage and integration of the two markets are illustrated in figures 3, 5 and 6. If there were no arbitrage pressures we expect the negative price shock to result in a persistent large price difference between Ontario and New York. However this is not the case over the sample period examined. The negative supply shock is reflected in the large increase in the Ontario price at time t+1. As discussed above, traders used available risk-mitigation strategies to cut export transactions and lim-

High Frequency Export and Price Responses / 49 it their risks and export volumes fall from time t+1 to time t+4. During that time the New York market receives less Ontario supply and the New York price trends upward. This initial adjustment narrows the price difference between Ontario and New York. The largest price effect occurs at time t+4 coinciding with the biggest export volume reduction in the Ontario market. Note also, in figure 6, how import volume from Michigan increased following the price shock. The very fact that export volume declines whilst import volume increases following the positive price shock provides robust evidence that the Ontario, New York and Michigan markets are interrelated. Furthermore the fact that New York price trends upward when the New York market does not receive Ontario exports provides evidence that the two markets are integrated. Were this not to be the case, the price in New York in the initial hours after the market disturbance would be independent of the Ontario exports. Therefore it must be the case that the two markets are linked through the export channel (absent congestion and other impediments to trade on the intertie). The increase in the New York price from time t+1 to t+4 partly reflects adjustment in supply in the New York market. In the immediate hours after the export reduction (the short-term), relatively fast-ramping units are required to replace the supply lost from Ontario. From time t+4 onwards, as the market information about the forced outage is absorbed, suppliers in New York re-adjust their offers and units are brought online to replace the reduction in imported power from Ontario. The supply elasticity in the New York market ultimately determines the level of the New York price after the Ontario market price shock. The fact that the price gap returns to some trend level (as opposed to a rising trend) further suggests the existence of arbitrage-bounded trades between the Ontario and New York electricity markets. Electricity markets are complex and present formidable challenges to the analyst. The event analysis is a simple, intuitive yet persuasive demonstration of export demand response to price increases. The results of the analysis indicate that export volume respond to an increase in the Ontario price relative to the New York price. Furthermore the analysis demonstrates that arbitrage activities narrow the price difference between the Ontario and New York markets. Finally the analysis provides evidence that the Ontario, New York and Michigan markets are interrelated. 5.  Conclusions, Policy Implications and Future Research Demand response has become increasingly important in todays electricity markets. Much focus has been placed on the domestic dimension of demand response. Export demand response is another potential, yet unexploited avenue of demand response which exists where markets are interconnected. In this study we used event analysis to demonstrate the export demand response that occurred following a price increase in the Ontario electricity market. We also showed how the Ontario, New York and Michigan markets are interconnected such that a positive

50 / The Energy Journal price shock in Ontario caused a reduction in export volume to New York and an increase in import volume from Michigan. The findings of this study have implications for policy development in the electricity sector. First in evaluating the options for demand response initiatives, the incremental cost/benefit of improved interconnector capability which would improve export (and import) response capability should be evaluated against the incremental cost/benefit of payment-based demand response programs. The elastic response of trade transactions could potentially yield substantial efficiency gains. On the other hand, unless the value of electricity consumption is known, the efficiency gains provided by payment-based programs are difficult to assess. In many cases such payments may even be highly inefficient. For example payments made to a participant who reduced its electricity consumption in the presence of high prices that would have caused the participant to cut consumption anyway (i.e. in the absence of the payments), would result in an inefficient use of resources. Second it is known that demand response can help to mitigate market power concerns in electricity markets. The existence of an effective and competitive regional electricity network with sufficient interconnector capacity serves as a powerful deterrent to abnormal price deviations in the home market. Arbitrage forces will not allow persistent abnormal high prices to prevail. Therefore the benefit of adequate infrastructure investment in interconnector transmission capacity goes far beyond the efficiency gains of export demand response. It actually expands the definition of the market to allow others to compete with the domestic generation and thus help to mitigate regulatory concerns of market power. In April 2001, the US energy regulator imposed a must-offer obligation on generators in the California electricity market.12 This was done primarily to address concerns about high prices in the market. Our analysis for the Ontario market suggests that high prices are quickly arbitraged away through an effective regional market response. This certainly demonstrates the benefits of having regional market integration as a policy objective. Looking ahead there is a dire need to understand how interconnected electricity markets evolve over time and how these markets cope with stressful events. Future research could examine the diffusion of price shocks across interties and across markets. Understanding the dynamics behind these complex price movements can help to clarify the extent to which electricity market integration can yield efficiency gains that can ultimately lead to cost reductions for the consumer.

12. According to FERC, The purpose of the Commissions must-offer obligation is to ensure that all units that are able to run but are not already scheduled to run (with the exception of hydroelectric power) are in fact made available to the ISO in the real-time market. Docket No. EL00-95-012, et al. ORDER ESTABLISHING PROSPECTIVE MITIGATION AND MONITORING PLAN FOR THE CALIFORNIA WHOLESALE ELECTRIC MARKETS AND ESTABLISHING AN INVESTIGATION OF PUBLIC UTILITY RATES IN WHOLESALE WESTERN ENERGY MARKETS.

High Frequency Export and Price Responses / 51 References

Arciniegas Ismael, Barrett Chris and Marathe Achla (2003). Assessing the efficiency of US electricity markets. Utilities Policy 11(2003) 75-86. Borenstein, Severin, Bushnell James, Knittel Christopher R, Wolfram, Catherine (2001). Death of a market: Trading inefficiencies in Californias electricity markets. University of California Energy Institute, PWP-086. Bower, John (2002). Seeking the European Electricity Market: Evidence from am empirical analysis of wholesale market prices. Oxford Institute for Energy Studies, EL-01. DeVany, Arthur S and Walls, David W (1999). Price dynamics in a network of decentralized power markets. Journal of Regulatory Economics, 15:123-140(1999). DeVany, Arthur S and Walls, David W (1999). Cointegration analysis of spot electricity prices: insights on transmission efficiency in the western US. Energy Economics, 21(1999) 435-448. Jamasb, Tooraj and Pollitt, Michael (2005). Electricity market reform in the European Union: review of progress toward liberalization and Integration. The Energy Journal, European Energy Liberalization Special Issue, 2005, 11-41. Woo Chi-Keung, Lloyd-Zanneti Debra, and Horowitz Ira (1997). Electricity market integration in the Pacific Northwest. The Energy Journal, Vol. 18, No 3. Worthington, Andrew Kay-Spratley Adam, Higgs Helen (2003). Transmission of prices and price volatility in Australian electricity spot markets: a multivariate GARCH analysis. Energy Economics 27 (2005) 337-350.

52 / The Energy Journal

Efforts and Efficiency in Oil Exploration: A Vector Error-Correction Approach

Klaus Mohn*

High oil prices and gradual resource depletion have raised global concerns for security of energy supply. Successful exploration activity is a critical factor for future oil production. Based on standard neoclassical producer behavior and modern time series econometrics, this study reveals new insights into the process of oil and gas exploration. I find that reserve additions are enhanced by an increase in the oil price, due to responses both in effort and efficiency of exploration. Moreover, oil companies accept higher exploration risk in response to an oil price increase, implying lower success rates and higher expected discovery size. 1. INTRODUCTION With brisk economic growth and high oil prices, issues of energy supply have ascended on the geopolitical agenda. Global oil demand has been fuelled by strong growth, both in the OECD area and in emerging economies like Brazil, Russia, India and China. Consequently, the sharp increase in the oil price over the last years has raised concerns among consumer interests about the security of supply (e.g., IEA 2006). However, no oil and gas can be produced before the reserves are uncovered and proved through exploration activities. It is therefore interesting that the most important strategic challenge for the oil industry also relates to reserve replacement. Oil and gas reserves in market-oriented economies like USA, Canada, United Kingdom and Norway are faced with depletion, and exploration activities are gradually redirected towards resource-rich regions in other parts of the world (e.g., Russia, Latin America, and the OPEC countries). Unless production volumes are replaced through successful exploration efforts, the basis for future production will be undermined.

The Energy Journal, Vol. 29, No. 4. Copyright 2008 by the IAEE. All rights reserved. * University of Stavanger (Department of Industrial Economics), 4036 Stavanger, Norway. Phone: + 47 95 44 18 36. Email: klaus.mohn@uis.no.

53

54 / The Energy Journal The reserve concept is one of the factors that distinguish non-renewable resource industries from other industries. Due to this defining characteristic, oil companies engage in extremely risky exploration activities to support and grow their base of oil and gas reserves, and to sustain production activity over the longer term. Among the oil companies, the set of exploration opportunities is subject to continuous evaluation and management based on a range of criteria relating to geology, technology, economic factors, and government policies. The result of this balancing act is a dynamic exploration strategy. Moreover, the implied portfolio of exploration drilling activities yields a certain average finding rate, a particular distribution of discovery size, and ultimately, a specific rate of gross reserve additions. Consequently, the data we observe for efforts and efficiency in oil exploration are formed by simultaneous decisions in each company. This simultaneity should be appreciated also in economic models of the exploration process. Drilling efforts have been subject to a wide range of econometric studies since the mid 1960s, especially for the US (Dahl and Duggan, 1998). Less attention has been paid to the success and efficiency of oil and gas exploration, not to mention the interaction between efforts and efficiency. Moreover, only a few empirical exploration studies cover more than one oil price cycle. To bridge these gaps in the empirical literature on oil and gas exploration, this study examines three components of reserve growth simultaneously in an integrated novel modelling approach. A simple model for reserve-generation is derived from standard neoclassical assumptions of producer behavior. Fishers (1964) decomposition of reserve growth is then applied to specify annual reserve additions as a result of drilling activity, success rates and average discovery size. Co-integration techniques (Engle and Granger, 1987; Johansen, 1995) are applied to estimate a vector error-correction model with three simultaneous equations for drilling activity, success rates and average discovery size, linking the three components of reservegeneration to relevant economic, geological and technology variables. The under-explored data set of the econometric application covers the full history of oil and gas exploration on the Norwegian Continental Shelf (1969-2004), including the last oil price increase. The estimated model gives a breakdown of effects from explanatory variables between key components of reserve-generation, as well as a systematic separation between short-term (temporary) effects and long-term (persistent) effects in the exploration process. The proposed model framework allows a detailed analysis of factors behind exploration behavior and reserve generation, including effects of oil price changes, historical exploration success, licensing policies, seismic surveys, depletion, and technological progress. In terms of results, the preferred econometric models suggest that reserve-generation from exploration is influenced by variables of geology, technology, economics, and regulation. In the short-term, drilling success has a temporary feedback effect on subsequent drilling activity and reserve additions. Licensing policies has a potential to stimulate efforts and efficiency of exploration, with

Efforts and Efficiency in Oil Exploration / 55 somewhat higher impact in the short term than in the longer term, according to the results. Moreover, accelerating seismic surveying activities have had a positive influence on exploration success, which may be seen as a reflection of the importance of technological progress in the oil exploration process. Finally, the results also suggest that an increase in the oil price has a positive effect on both efforts and efficiency in oil exploration. The estimated model provides a strong indication that the companies appetite for exploration risk is increasing in the oil price. Specifically, an oil price increase has a negative effect on average success rates, but a positive effect on average discovery size. Still, the net effect is an increase in yield per effort. Consequently, the results for oil price effects suggest a pro-cyclical pattern not only for exploration drilling, but also for displayed exploration risk and realised reserve additions. The paper is organized as follows. Section 2 gives an overview of previous research on exploration productivity. A simple model of exploration is outlined in Section 3, before the econometric model is specified in Section 4. The data set is presented in Section 5 and results are outlined and discussed in Section 6. Concluding remarks are offered in Section 7. 2. PREVIOUS RESEARCH The combination of geological and economic variables in empirical models of exploration dates back some 40 years. Fisher (1964) opened this field of research with his seminal econometric studies of US oil and gas exploration, estimating equations for the drilling rate, success rate and the discovery rate for different US Petroleum Administration Defence Districts (PADD) over the period 1946-1955. Explanatory variables included oil prices, seismic crews and proxy variables for drilling costs. Based on extended versions of the same data set, refined updates of the Fisher framework were presented in a range of papers for the US oil and gas industry over the following 15 years (e.g., Erickson and Spann, 1971; Pindyck 1974, 1978; Kolb, 1979). Gradual improvements were due to improved data availability, more sophisticated modelling and additional explanatory variables.1 Into the 1990s, some studies also emerge for the exploration and production of oil and natural gas on the United Kingdom Continental Shelf (UKCS; e.g., Pesaran, 1990; Favero and Pesaran, 1994). These models depart from an integrated, dynamic optimisation problem, and produce plausible econometric equations for exploration, development and production. However, their explanations of exploration efficiency are quite simple, and they also fail to establish robust estimates in support of inter-temporal maximisation. Later studies, especially for the US oil and gas industry, have therefore relaxed the assumption of inter-temporal
1. A requirement for theory-based econometric research on the oil and gas industry is that the data set is generated in an industrial environment mainly guided by commercial principles. Further, the quality of the data will usually increase, the longer the history of the actual activity. This explains why most applications of economic theory to observed exploration behavior is based on data from USA.

56 / The Energy Journal behavior, and several have returned to the period-by-period optimisation approach (e.g., Iledare and Pulsipher, 1999; Farzin, 2001). Implicitly, this behavioral assumption is also adopted by Ringlund et al. (2007). Still, their econometric specification of oilrig activity for a panel of non-OPEC countries is indeed dynamic. The present study follows a similar line of thought.2 Empirical exploration models for the US oil and gas industry are surveyed by Dahl and Duggan (1998), who conclude that acceptable models have been obtained for drilling efforts, with long-term oil price elasticities above one. On the other hand, Dahl and Duggan recommend that more work should be done on drilling efficiency. The role of the oil price and technological change is especially interesting in this respect, as a role for these variables would open for policy implications beyond the stimulation of drilling efforts. Iledare (1995) include exploration efficiency in a two-equation system of reserve generation. However, economic variables are restricted to the drilling equation in his model, with no effects from the oil price on yield per effort. Kemp and Kasim (2006) estimate a system of equations for the exploration process based on data for five regions on the United Kingdom Continental Shelf (UKCS), and also do not include oil price variables in the equations for the discovery process. As oil companies adapt their exploration strategies according to changes in economic and financial conditions, recordings of exploration efficiency will also reflect oil price changes. To some extent, this point is acknowledged by Iledare and Pulsipher (1999), who estimate drilling effort and yield per effort in onshore gas exploration in Louisiana (USA). In their estimated yield-per-effort equation, the oil price takes a negative coefficient, and the authors suggest the reason is a negative link between the oil price and average discovery size. However, without a split of exploration efficiency between the discovery rate and average discovery size, such an assertion remains conjectural. Learning-by-doing, business intelligence and technological advances play an important role in the exploration process (e.g., Quyen, 1991; Hendricks and Porter, 1996). However, finding variables to identify these effects precisely remains an unresolved issue. A key problem is that variables that capture technological progress tend to move monotonically over time, and therefore correlate with indicators of maturation and depletion. Moreover, the explanatory content of trending variables may also be mixed. As an example, cumulative drilling may be seen both as an indicator of accumulated knowledge (Uhler, 1979; Iledare, 1995), and a measure of depletion (e. g., Pesaran, 1990). Thus, a statistically significant time trend, as well as any other trending variable may combine the information of different (opposing) effects. These challenges are hardly overcome by the recent studies of success rates in US oil and gas exploration by Forbes and Zampelli (2002, 2002), who suggest the use of annual time dummies to assess the impact of technology change. A more promising attempt is offered by Managi et al. (2005),
2. To test the validity of this assumption, a variety of interest rate variables were included in preliminary estimations. However, plausible and robust estimates could not be established for any of their coefficients.

Efforts and Efficiency in Oil Exploration / 57 who construct an exploration-specific index of technology diffusion for their econometric studies of yield per effort in GoM exploration. A natural extension of this approach is to study the impact of specific technologies through special variables. Still, Forbes and Zampelli (2000, 2002) and Managi et al. (2005) limit their scope to exploration efficiency. An econometric model of the exploration process should acknowledge the simultaneous interaction between efforts and efficiency, and reveal responses from economic, geological and technological variables in the short term as well as in the long term. The modelling approach of the present study represents an indicative step in this direction. 3. A SIMPLE MODEL OF EXPLORATION BEHAVIOR The reserve concept is a key characteristic of the production technology in the oil and gas industry. Oil companies invest in risky exploration activities to grow their reserve base, a critical requirement to sustain production over the longer term. Over the last years, markets have developed for oil and gas reserves, license shares and other petroleum assets. The implication is that exploration decisions do not necessarily require a commitment to future development and production. Rather, any discovery is evaluated on a more or less independent basis, and a range of strategies is available for value optimisation. Oil exploration may therefore be studied as a distinct profit-generating activity among oil companies.3 A simple model for oil exploration may therefore be based on the technology structure of reserve-generation. With gross reserve additions (R) as the output of the exploration process, the point of departure for our theoretical model is a well-behaved neo-classical production function R = F (L, X, H, t), where L represents variable inputs, X represents fixed inputs and state variables, H is an indicator for the depletion of exploration opportunities, and t is an index of technological progress, usually represented by a simple time trend. Both the efficiency of this process and the implied accumulation of oil and gas reserves contribute significantly to the value of the firm. Consequently, oil companies maximize profits from reserve generation subject to output prices, input prices, and constraints relating to geology, technology, and policy regulations. Under these assumptions, a restricted profit function (( )) of reserve generation may be derived from a representative oil companys profit maximisation problem: = (P, W, X, H, t) = max R,L {P R W L} s.t. F (L, X, H, t) R, (1)

3. An increasingly liquid asset market is available to the oil and gas industry, enabling companies to optimize their portfolios of different reserves and projects. Oil and gas discoveries with a poor strategic fit may thus be traded for cash or for shares in alternative assets (swap transactions).

58 / The Energy Journal where P is the marginal value of new oil and gas reserves4 and W represents traditional input prices. Previous literature suggests that the role of traditional inputs is dominated by other factors in the process of oil and gas exploration (e.g., Dahl and Duggan, 1998). More sophisticated models of capital inputs normally involve explicit dynamic behavior in terms of intertemporal profit maximisation. However, the assumption of period-by-period optimisation is justified on several grounds. First, this kind of myopic behavior is well supported by previous studies of exploration behavior, whereas the hypothesis of dynamic optimisation is not. In an econometric study of US oil and gas supply, Farzin (1986) obtains estimates for the discount rate among 33.4 per cent, implying a time horizon of approx. 4 years for their extraction decisions. Another example is a study of exploration and production in the United Kingdom by Pesaran (1990), where present-value maximisation is also not supported by the data. Second, the reserve definition of this study requires that recovery is highly likely at current economic and operating conditions. This reserve concept implies a higher degree of flexibility than a geophysical definition of reserves, and the difference is influenced by continuous development in variables relating to technology, economics, and government policy. Third, the hypothesis of discretionary profit-maximisation does not entirely exclude the idea of dynamic optimisation. As pointed out by Farzin (2001), period-by-period maximisation may rather be seen as a special case, whereby the flow of future profits is expected to grow at a rate below the discount rate. A few theoretical studies of oil exploration have been occupied with human capital and learning-by-doing. Quyen (1991) stresses that the discovery of new reserves is only one of the motives of oil exploration. In addition, exploration drilling will reveal information and insights that may prove useful for future exploration activities.5 These learning-by-doing mechanisms are linked to key personnel in the exploration process. At the same time, the information revealed through drilling efforts has a potential value for both partners and competitors in the actual exploration area. Hendricks and Porter (1996) apply a game-theoretical modelling framework to analyse how information externalities and potential freeriding affect the optimal timing of drilling decisions. As expected gains from drilling have to be balanced against expected gains from waiting for drilling informa4. The price of crude oil serves as a proxy in this respect. Following Iledare (1995), econometric tests have been performed for a range of more sophisticated unit cash-flow variables, including a variety of adaptive expectations hypotheses for the oil price. These variables are typically put together in relations like: Ve = Pe(1-c)(1-), where Ve is a proxy for the expected marginal value of reserves, Pe is the expected oil price, c is a unit cost variable and is a corresponding unit tax. However, none of the measures that incorporate various types of adaptive price expectations, unit costs and tax payments were able to outperform the simple oil price variable in the estimated models. The outlined plain formulation is therefore maintained. 5. A popular analogy is found in the classic board game Battleship. In the early phases of the game, with many ships on the board, expected rewards from bombing are high, with major learning effects involved whenever a new ship is hit. However, expected marginal gains, as well as the learning effects, drop towards the end of the game, when the majority of ships have been sunk.

Efforts and Efficiency in Oil Exploration / 59 tion in neighbouring areas, the result is a war of attrition, giving rise to U-shaped patterns of drilling activity in each tract. This kind of strategic interaction also provides an important role for human resources in the exploration process. Still, a clear role for labour and labour costs in empirical models of exploration behavior is yet to be established (Dahl and Duggan, 1998). Following the mainstream of recent research, the simple theoretical model of this paper therefore assumes that costs associated with capital and labour stem from a standby capacity, and that they are insensitive to fluctuations in exploration activity from one year to another.6 The implication is that the portfolio of exploration activities is subject to discretionary evaluation and adjustment from year to year, based on market developments and other changes in the opportunity set given a fixed capacity of capital and labour. Consequently, variable inputs (L) and their prices (W) are disregarded in the further development of our model. To approach the empirical specification, we now assume the following multiplicative form for the profit function in Equation (1): ~ ~ n (P, X, H, t) = KP a Xii egH+dt,
i =1

(2)

~ ~ and represent elasticities of profit with reWhere K is a constant term, and a i spect to the oil price (P) and state and policy variables (Xi), respectively. With a negative coefficient for , the specified mechanism implies an exponential decline in the exploration opportunity set. As found by Iledare (1995), this physical rate of decline could be offset by technological progress, captured in Equation (2) by the time trend t. Hotellings lemma may no be applied to the restricted profit function in Equation (2) to derive the optimal supply schedule for gross reserves. Specifically, differentiation in Equation (2) with respect to the marginal value of new reserves (Pt) now yields: (P,X ) = R* (P, X, H, t) = KP a Xii egH+dt i P (3)

~ ~K ~ 1. An asterisk is added to R* to distinguish desired where K = a and a = a reserve requirement from the observed level. Equation (3) may be interpreted in terms of optimal supply, and establishes a theoretically consistent relationship between reserve generation on the one hand and the oil price and various state variables on the other, derived directly from the companys profit maximisation problem.7 Positive changes to the marginal value of reserves (P) are expected to stimulate reserve additions and exploration efforts. The marginal impact of
6. Wage and interest rate variables were included in preliminary stages of the estimation process. However, based on statistical evaluation, none of these variables could justify a position in the preferred empirical models. 7. The equivalent dual approach is to depart from the cost-minimisation problem of exploration activities, and derive a similar relation from the corresponding first-order conditions, whereby the marginal value of gross reserve generation is equated to its marginal cost (e. g., Farzin, 2001).

60 / The Energy Journal changes to the state variables (X) will depend on the specific nature of these variables. The depletion mechanism (H) is expected to have a dampening effect on reserve-generation, partially offset by technological progress (t). In a number of previous empirical studies of exploration behavior, oil price effects are confined to drilling activities, whereas success rates and discovery size are determined by physical variables (e. g., Pesaran, 1990; Iledare, 1995; Kemp and Kasim, 2003, 2006). However, there is reason to believe that oil price fluctuation has a direct effect also on the productivity of the exploration process (Iledare and Pulsipher, 1999; Forbes and Zampelli, 2000). Decision-makers in the oil companies face a set of exploration opportunities which is subject to continuous shocks, due to market developments, changes in regulations, new information, and technological innovations. Consequently, the portfolio of exploration activities requires continuous evaluation and management. The outcome of this balancing act is a dynamic exploration strategy. Moreover, any established combination of drilling activities will yield a certain expected finding rate, a particular distribution of discovery size, and ultimately, a specific rate of reserve additions. For these reasons, the data we observe for efforts and efficiency in oil exploration are generated by simultaneous decisions in each company. To grasp the complexity of the exploration process, we need a model that takes explicit account of both the breakdown of reserve-generation and the simultaneity between its components. We therefore apply the useful decomposition introduced by Fisher (1964), whereby gross reserve additions (R) may be seen as the product of the rate of drilling activity (D), the success rate (S), and average discovery size (M): R (P, X, H, t) = D (P, X, H, t) S (P, X, H, t) M (P, X, H, t). (4)

Potentially, the oil price (P) and the vector of state variables (X) may influence the growth of oil and gas reserves via each of these three components, and an integrated model should therefore provide a portrayal of these variables. For each of the components of reserve generation, Equation (4) implies a relationship similar to Equation (3): D* (P, X, H, t) = KdP a Xii eg
i d d d H + ddt

(5) (6) . (7)

S* (P, X, H, t) = KsP a Xii eg

i m i

s H + dst

M* (P, X, H, t) = KmP a Xii eg

m H + dmt

where superscripts d, s, and m is introduced to separate coefficients associated with drilling efforts (d) from the corresponding coefficients of the drilling success (s) and discovery size (m) equations, respectively. With small-caps for natural logs, gross reserve-generation can now be represented by the following sum:

Efforts and Efficiency in Oil Exploration / 61 r ( p, x, H, t) = d ( p, x, H, t) + s ( p, x, H, t) + m ( p, x, H, t), where: d* ( p, x, H, t) = kd + adp + idxi + g H + d t,

d d i =1 n

(8)

(9) (10) (11)

s* ( p, x, H, t) = ks + asp + isxi + g H + d t,
s s i =1 m m m* ( p, x, H, t) = km + amp + imxi + g H + d t. i =1 n

Consequently, the elasticity of reserve additions (Rt) with respect to the different explanatory variables may now be studied as the sum of three partial elasticities. For the oil price elasticity (p), this implies: DR (P, X, t) P Dr d s m ep = = = + + = ad + as + am DP R Dp p p p (12)

Corresponding relations apply for the vector of state variables (ex = id + is + i im), and for the semi-elasticities associated with depletion (eH = gd + gs + gm) and technological progress (et = dd + ds + dm). This integrated approach to the exploration process should be fruitful both for industry leaders and policy-makers. As an example, consider the influence on the exploration process from an oil price shock. An increase in the oil price is likely to stimulate drilling efforts. However, if the general risk inclination in the industry is affected by the oil price, an oil price increase may also induce oil and gas companies to take on higher exploration risk. This would imply a selection of more risky exploration wells, a potential reduction in discovery rates, and an increase in average discovery size. In turn, these processes could influence on subsequent drilling efforts in the short term and/or in the long term. The outcome in terms of reserve generation is unclear. However, the involved effects may be addressed explicitly within the outlined model specification. We now proceed to an econometric specification that expands the modelling approach even further, to take account of potential interaction between the three components of reserve-generation, as well as sluggishness and dynamics in the underlying data-generating process. 4. ECONOMETRIC SPECIFICATION Drilling efforts (Dt), drilling success (St) and average discovery size (Mt) are the endogenous variables of our model. The proposal of this study is an econometric specification that meets requirements of tractable estimation, takes proper account of the dynamics of the data-generating process, and appreciates the simultaneity of efforts and efficiency in oil and gas exploration.

62 / The Energy Journal From the oil companies point of view, the set of available exploration opportunities is subject to constant change in external conditions. Examples include oil price shocks, changes in regulations, new information, and technological advances. Managers respond to these changes through a continuous management of their exploration portfolio. The presence of price expectations and/or adjustment lags in the impact of the explanatory variables may also cause deviation between the actual level of reserve generation and the desired level. More specifically, the process of oil exploration may be disturbed by contractual obligations, leads and lags in the development of new technology, uncertainty about future prices, as well as regulatory constraints and other economic, psychological, and technological factors. Thus, the data-generating process may well be characterized by sluggishness and dynamics, even if the data is not the result of a dynamic optimisation procedure. With annual time series, the variables of my model are also likely to be non-stationary. Moreover, their linear combination may also be non-stationary. In that case, direct regressions on variable levels yields inefficient coefficient estimates, threatening the validity of statistical inference. Simple estimation in the levels of the variables will also fail to describe the dynamics of the data-generating process. However, if the variables are integrated of degree 1 (I(1)), their difference will be stationary: yt~I(1) yt~I(0). A corollary from the literature on co-integration is that a linear combination of co-integrated non-stationary variables will produce a stationary error-term. A pioneering reference to this literature is Engle and Granger (1987), who also demonstrate that any set of co-integrated variables has a valid error-correction specification.8 The point of departure for the econometric specification is given by the following system of dynamic equations: yt = jytj + kxtk + ut
j =1 k =0 n n

(13)

where yt=[dt,st,mt]' is the vector of endogenous variables and xt=[pt,et,zt,Ht,t] is the vector of explanatory variables, including the oil price (pt), the depletion indicator (Ht) and the technology index (t) for simplicity of exposition. Other explanatory variables include a variable for licensed exploration acreage (et) at the beginning of year t, and a measure of seismic surveying activity (zt). All these variables will be explained in further detail below. Lag coefficients on endogenous variables and exogenous variables are given by the j and k matrices, respectively. Finally, ut is a 3x1 vector of white-noise residuals. If the variables of Equation (13) are co-integrated, both the dynamics and the underlying structure of its data-generating process may be estimated in a vector error-correction model (VECM) with n lags: Dyt = aiDytj + bkDxtk + lyt1 + dxt1 + ut ,
j =1 k =0 n1 n1

(14)

8. Hendry and Juselius (2000, 2001) provide a modern introduction to the econometrics of cointegration and error-correction models.

Efforts and Efficiency in Oil Exploration / 63 where i are 33 matrices of short-term coefficients from lagged changes in endogenous variables, and the i matrices contain short-term coefficients from other explanatory variables. The 33 matrix of error-correction coefficients is given by the 33 matrix, whereas persistent effects from the explanatory variables are represented by the 35 matrix. Equation (14) contains a large number of variables and coefficients, with high requirements for the data set. Our application is based on annual time series data from 1969-2004 (n=36), and the modelling ambitions will have to be adjusted accordingly. With one lag and five exogenous variables the following model version is the starting point for our econometric application:9 Dyt = aDyt1 + bDxt1 + lyt1 + dxt1 + ut (15)

With variables in natural logs, the coefficients of Equation (15) can be interpreted in terms of short-term and long-term elasticities. More specifically, the vector contains short-term elasticities of drilling efforts, discovery rates and average discovery size with respect to the explanatory variables of the model. In the long run, all change variables in Equation (15) approach zero. This property can be utilized to bring out the long-term parameters () directly from the estimated error-correction model (Brdsen 1989). Eliminating all changes and solving for the dependent variables, the matrix of long-run coefficients can be derived as: = l1d . Following Equation (4), total elasticities of reserve generation are computed as the sum of partial elasticities, both for short-run (temporary) effects and for long-term (persistent) effects. Let js , jl represent total elasticities of reserve generation (Rt) wrt. to explanatory variable j in the short run and long run, respectively. This yields: ejs = bjd + bjs + bjm, ejl = jd + js + jm, j = p, e, z, H, t, (16) j = p, e, z, H, t,

where [jd,js,jm] and [jd,js,jm] represent short-term and long-term effects from variable j on drilling effort (d), success rate (s), and average discovery size (m), respectively. This calculation of total elasticities of reserve generation gives a systematic account for the interaction between efforts, efficiency, and other explanatory variables in the exploration process. 5. DATA SET AND VARIABLES Time series are retrieved from the data bases of The Norwegian Petroleum Directorate, who has collected and processed information and statistics on
9. Contemporaneous changes in the explanatory variables were also evaluated in the process of estimation. However, based on statistical inference, a position could not be justified for any of these variables in the preferred model. For simplicity of exposition, they are therefore left out in Equation (15).

64 / The Energy Journal Norwegian oil and gas activities since the mid 1960s. The top left-hand panel of Figure 1 illustrates total efforts in terms of exploration wells drilled per year (Dt), along with accumulated number of exploration wells (Ht), a common proxy for the depletion mechanism involved in the exploration process (e. g., Pesaran, 1990; Iledare and Pulsipher, 1999). Annual well-count is a simple and plain activity measure, has an easy interpretation, and agrees well with the selected theoretical specification. A typical feature for the NCS is that no clear ex-ante distinction is made between exploration wells for oil and exploration wells for natural gas. In the present study, the total annual sum of exploration wells therefore serves as the key indicator for exploration efforts. Exploration activity peaked during the 1980s. Over the last 20 years, exploration efforts have stagnated, and so has the number of discoveries.10 Key components of exploration efficiency, the average success rate (St) and average discovery size (Mt) are illustrated in the top right-hand panel of Figure 1. The largest discoveries were made between 1969 and 1980, and recent discoveries are miniscule compared to the typical field size of the 1970s. On the other hand, the discovery rate has not come down. This suggests that shortage of opportunity is offset by improved exploration technology and the development of specific geological competence (Forbes and Zampelli 2002, 2002; Managi et al. 2005). As oil and gas companies minimize their costs of reserve generation, the most prospective areas are normally drilled first (e.g., Iledare, 1995). The result is large average discoveries and high reserve growth in the early phases of an oil and gas province. Over time, the pool of undiscovered resources is gradually depleted by exploration activities. Success rates may be upheld through learning-by-doing and technological progress, but average discovery size will normally fall. The first exogenous variable is the oil price (Pt). The preferred choice for this study is Brent blend, the standard reference for North Sea crude oil. Preliminary estimations were run on both NOK and USD denominations for the oil price, and the properties of nominal versus the real price have been thoroughly tested. Based on these tests, a real USD denomination is favoured.11 Technological progress is captured by a linear time trend. However, the model also includes a variable for the intensity of information gathering and processing, measured by seismic surveying activity.12 The Zt variable represents
10. Observe that the annual number of discoveries (Ft) is given by the product of annual drilling activity (Dt) and the annual success rate (St): Ft= Dt St. In a similar fashion, exploration efficiency (Gt), or yield per effort, may be defined as the product of the annual success rate (St) and average discovery size (Mt): Gt=StMt. 11. Statistical inference was the key criteria for this selection. However, I also looked at the explanatory power of the various model versions, and how different oil price variables interfered with the quality of the other coefficient estimates of the model. Based on these considerations, a real USD denomination was preferred in the final version of the model. 12. Seismic profiles of the underground are acquired by transmitting sound waves from a source above or in the substratum. The sound waves travel through the rock layers which reflect them up to sensors on the sea bed or at the surface, or down in a borehole. This enables an image of formations in the substratum to be formed. The seismic mapping of the Norwegian continental shelf started as early as 1962 (Norwegian Petroleum Directorate, 2007).

Efforts and Efficiency in Oil Exploration / 65 Figure 1. Key Variables of the Data Set

Sources: Oil price: http://www.EcoWin.com. All other numbers: Norwegian Petroleum Directorate.

annual seismic activity measured as square kilometres of survey coverage per year. As we see from Figure 1, Panel 3, the Zt variable has a significant time trend, and therefore correlates with common measures of depletion, which also tend to move monotonically over time. A widely applied indicator of depletion is accumulated drilling activity (e. g., Iledare and Pulsipher 1999; Managi et al. 2005; Kemp and Kasim 2006), illustrated by accumulated number of exploration wells in the top left-hand panel of Figure 1. The bottom right-hand panel of Figure 1 also illustrates how exploration acreage (Et) has been regulated by the Norwegian Government. Some 42,000 km2 were awarded in the 1st licensing round in 1965, ahead of the opening of the Norwegian Continental Shelf. Licenses that were handed back to the Government towards the mid 1970s reduced the cumulative open exploration acreage, before a series of licensing rounds added new frontier acreage in the Norwegian Sea and in the Barents Sea from 1980. Licensing policies have also been adjusted over the

66 / The Energy Journal last few years to spur exploration activity, and large areas were awarded in mature areas and frontier areas both in 2003 and 2004. Descriptive statistics for all model variables are offered in Table 1. Table 1. Descriptive Statistics for Data Sample
Variable D t S t M t P t E t Z t H t Obs. 36 36 36 36 36 36 36 Mean 27.000 0.3575 53.632 31.201 38348 213787 486.55 St. dev. 12.89 0.1364 81.385 19.179 14874 249760 367.87 Min. 0 0.09 1.74 8.8 0 48 0 Max. 50 0.71 335 81.2 65335 803936 1080

Source: Norwegian Petroleum Directorate (D, S, M, E, Z, H; http://www.npd.no), ReutersEcowin (P; http://www.ecowin.com).

6. ESTIMATION AND RESULTS The econometric analysis aims at establishing robust statistical relations for the three endogenous variables of the exploration model, explaining their development over time in terms of dynamic interaction and influence from explanatory variables.13 As described by Equation (14), changes in drilling efforts (dt), discovery rates (st) and average discovery size (mt) are regressed against both the changes and the lagged levels of the dependent (dt-1, st-1, mt-1) and independent variables (pt-1, et-1, zt1, Ht-1, t). The error-correction approach allows a separation between short-term effects, and long-term impact from the explanatory variables. Further, the error-correction model provides explicit estimates for the interlinked adjustment pattern of the data-generating process. Due to excessive volatility in drilling rates, discovery rates, and average discovery size (cf. Figure 2), the estimated models are based on three-year moving averages for the endogenous variables. All estimation procedures are performed with the full-information maximum likelihood procedures for dynamic VAR models, as implemented in PcGive 10 (Dornik and Hendry 2001). A requirement for our econometric model specification is that the variables are non-stationary, and integrated of the same order. We therefore stop for
13. The size of my data set suggests that a simpler model with one or two equations might involve gains both in terms of economic explanation and not least in terms of statistical quality. Consequently, my estimations were supplemented with a corresponding single-equation ECM model for grossreserve growth (Rt), as well as a two-equation VECM model for efforts (Dt) on the one hand and efficiency (Gt = St.Mt) on the other. However, none of these alternative specifications of gross reserve growth offers an explanation that could be preferred to the presented three-equation model. Rather, these simpler specifications show weaker econometric performance, and seem to blur the behavioral detail and complexity of the exploration process.

Efforts and Efficiency in Oil Exploration / 67 a moment to investigate the stationarity properties of the model variables. Augmented Dickey-Fuller tests, Dickey-Fuller generalized least squares tests, and Phillips-Perron tests are carried out for all model variables, and their results are presented in Appendix 1. Non-stationarity is rejected in only 2 out of 21 cases, and we therefore conclude that the variables of our model are indeed non-stationary. To rule out higher order integration, the same battery of tests is run on the changes of our model variables. Non-stationarity in the changes of the variables is rejected for 19 out of 21 cases. Consequently, we base our subsequent analysis on the presumption that the variables contain a unit root, and specify the exploration process as a model of three simultaneous error-correction equations. The estimation procedure draws on the general-to-specific approach (Hendry 1995). The starting point of this procedure is a full version of the econometric model, including the full set of variables and lags. A sequential reduction is then pursued, whereby parameter estimates are eliminated one by one, based on statistical significance and their contribution to the general quality of the model. However, the present data set calls for modification, as a full-fledged version of Equation (14) would exhaust the degrees of freedom. Starting with the error-correction matrix (), the various coefficient matrices have therefore been exposed to estimation, testing, and reduction one by one in a sequential procedure.14 The result was a reduced version of Equation (14), with a narrow set of significant parameters and favourable econometric properties. To control for potential interaction between groups of variables, each of the previously removed variables were re-entered one by one in the final parsimonious model, to verify that they really had no significant role to play in the preferred model. The result of this process is a preferred model, represented by the following three estimated equations: = 0.2 4** Dd + 0.2 8** Ds + 0.26** De Dd t t1 t1 t2 (0.02) (0.02) (0.05) 0.2 1*** dt1 0.1 6** st1 + 0.0 8* pt1 + 0.03* et2
(0.00) (0.04) (0.06) (0.10)

(17)

= 1.1 9*** + 0.33** De + 0.11*** Dz Ds t t2 t1 (0.00) (0.05) (0.01) 0.37*** st1 0.05 mt1 0.0 9* pt1 + 0.0 8*** zt1
(0.01) (0.16) (0.06) (0.00)

(18)

= 0.7 3*** m + 0.40**p + 0.21**e 0.9 1*** H Dm t t1 t1 t2 t

(0.00) (0.05) (0.01) (0.00)
*)

(19)

Significant at 90, **) 95 and

***)

99 per cent confidence level, respectively. p-values in brackets.

14. For the nested models in each of these stages, elimination of variables was based on consistently calculated confidence levels. Each of the stages in the reduction procedure was also monitored and estimation results were recorded for progress appraisal. Improvement in model quality was evaluated through the log-likelihood, as well as the Schwartz, Hanna-Quinn, and Akaike information criteria.

68 / The Energy Journal The estimated system of equations passes the usual specification tests for autocorrelated residuals (LM test statistic: 0.84; p-value = 0.65), normality of residuals (LM test statistic: 9.65; pvalue = 0.14), and heteroskedasticity (F test statistic: 0.38; pvalue=1.00).15 Estimated parameters take plausible values and signs. Note that the error-correction coefficients are negative and highly significant for all the three equations. This is also an indication for co-integration (Kremers, Ericsson and Dolado 1992), and therefore supports our VECM specification. Short-term dynamics play a modest role in the exploration process, with an exception for drilling activity (dt). Long-run oil price effects are significant in all three equations, whereas no short-term effects can be detected for the oil price. New exploration acreage (et-1) largely has a temporary effect on drilling efforts (dt) and success rates (st), whereas the impact on average discovery size (mt) is persistent. Moreover, drilling efforts seem insensitive to changes in average discovery size. On the other hand, an increase in average discovery size has a weak positive effect on subsequent success rates.16 Changes in the oil price (pt-1) have no short-term effect on either of the three components of reserve generation, according to the results. On the other hand, licensing rounds (et-1) exert a significant short-term impetus to drilling efforts (0.26), to the discovery rate (0.33), and thereby also to reserve generation. Following Equation (17), the short term elasticity of reserve generation with respect to new acreage (es) can be computed to 0.68 (p-value < 0.00). This suggests an effective role for licensing policies in the short-term.17 Finally, an increase in seismic surveying activity (zt1) has a modest, instant effect on the finding rate (st), but no effect on drilling activity (dt) or average discovery size (mt). Observe also that the estimated drilling equation implies a significant temporary impulse from historical drilling success. As the number of discoveries (Ft) is the product of drilling efforts (Dt) and the success rate (St), the impact on drilling from historical exploration success can be caluculated as the sum of the two relevant short-term elasticities. For the short-term elasticity of drilling (Dt) with respect to lagged exploration success (Ft-1), this yields: 0.23 + 0.28 = 0.51 (p < 0.00). A temporary feedback from exploration success on subsequent drilling efforts and reserve additions is in accordance with previous results by Mohn and Osmundsen (2008). Based on the theoretical model specification, a time trend was entered in preliminary estimations, to allow for technological progress. However, this vari15. See Dornik and Hendry (2001) for theoretical background and specific procedures for standard specification tests and model diagnostics in PCGive 10. More detailed specification tests are presented in Appendix 2. 16. With a p-value of 0.16, the coefficient on mt-1 in Equation [18] is barely significant in statistical terms. However, the variable is still retained in the preferred model to accommodate our hypothesis of three co-integrating vectors, as indicated by preliminary tests of cointegration rank performed on the full set of time series variables. 17. Historically, access to exploration acreage on the Norwegian Continental Shelf has been subject to strict regulations and gradual opening of new areas. See Ministry of Petroleum and Energy (2007) and Norwegian Petroleum Directorate (2007) for details.

Efforts and Efficiency in Oil Exploration / 69 able could not be retained in the preferred system of estimated equations. Technological progress plays a potentially important role in the exploration process. However, the exact identification of the empirical effects remains an unresolved issue (e. g., Forbes and Zampelli, 2000, 2002; Managi et al. 2005). The reason is that variables that capture technological progress tend to move monotonically over time, and therefore correlate with indicators of maturation and depletion. In the presented model, there is reason to believe that technological progress is captured by the variable for seismic surveying activity (Zt). Costs of seismic surveying activity have collapsed over the 40-year history of the Norwegian Continental Shelf. Previous simple and costly methods have gradually been replaced by modern technologies for the collection and processing of enormous amounts of geological information. As we see from Equation (18), these activities have a positive and highly significant effect on the success rate in NCS oil and gas exploration. On the other hand, the results do suggest a significant role for the depletion mechanism. As we see from Equations (17)-(19), the depletion indicator (Ht) is retained only for the field-size equation. As drilling activities cumulate, average discovery size tends down. This follows from an exploration behavior whereby prospects with high reserve potential are drilled first. Moreover, this part of the results also indicates that the most important role for the depletion mechanism in NCS oil and gas exploration relates to exploration success, and not so much to drilling efforts. At this point, the results differ from the greater part of previous research on drilling activity (e. g., Dahl and Duggan, 1998; Mohn and Osmundsen, 2008).18 One possible explanation for this result relates to differences in modelling approach. Previous studies have typically focused on single-equation models of drilling activity, whereas the present study captures effort and efficiency in a simultaneous model. Another possible explanation may be due to institutional differences between the NCS and other more market-oriented petroleum provinces. Specifically, the gradual opening and strict regulation of exploration activities on the NCS may act as a disturbing factor for the structural depletion mechanism observed in other less regulated petroleum provinces. Finally, depletion mechanisms may also be picked up by the increase over time in the success rate (St), which again exerts a negative influence on drilling efforts, according to the results. Persistent effects on the three components of reserve generation from the exogenous variables are defined by the co-integrating vectors of the estimated system, and correspond directly to the long-term elasticities presented in Table 3. Partial and total effects are also illustrated in Figure 2. Following Equation (16), significance tests for the long-term parameters are computed by testing the validity of equivalent general restrictions on the estimated system. The first line of Table 3 contains a summary of estimated oil price elasticities (Appendix 3 accounts for the derivation). The oil price exerts a persistent influence on all the three components of reserve generation in the exploration process (Dt, St, and Mt).
18. However, the previous literature also offers examples where cumulated drilling activities exert a positive influence on current drilling. For example, Iledare and Pulsipher (1999) argue that this might be due to a dominance of learning-by-doing effects over the standard depletion mechanism.

70 / The Energy Journal As illustrated by Equation (16), the total elasticity of reserve generation with respect to the oil price is given by the sum of the three partial elasticities. The results imply that an oil price increase of 1 per cent will produce an increase in annual reserve additions by 0.89 per cent in the long term.19 Table 3. Long-Term Elasticities of Reserve-Generation
Reserve generation (R) Oil price (P) Licensed acreage (E) Seismic surveys (Z) Depletion (H)a) 0.89*** (0.00) 0.44*** (0.00) 0.05 (0.62) -1.29*** (0.00) Exploration drilling (D) 0.51*** (0.00) 0.11 (0.16) -0.16 (0.16) 0.13 (0.35) Drilling success (S) -0.17* (0.10) 0.04 (0.28) 0.22*** (0.00) -0.17 (0.26) -1.25*** (0.00) Average discovery size (M) 0.55** (0.02) 0.29*** (0.00)

*) Significant at 90, **) 95 and ***) 99 per cent confidence level, respectively. p-values in brackets. a) Semi-elasticity: percentage change in dependent variable from absolute change in depletion indicator.

Table 3 and Figure 2 also illustrate that the oil price effect varies distinctly across the three equations. Obviously, drilling efforts are invigorated by oil price increases, although the effect is quite modest.20 On the other hand, discovery rates are suppressed when the oil price increases. Finally, a positive and highly significant link is established between the oil price and average discovery size. A likely interpretation is that oil companies adjust their portfolio of exploration activities according to changes in economic and financial conditions (Reiss, 1990). In times of high oil prices, high cash-flows and high risk appetite, companies tilt their exploration activities towards risky areas (frontier exploration), with rela19. Recent trends in oil and gas exploration activity indicate that companies respond swiftly to any oil price drop, whereas the reaction to an oil price increase is more sluggish. This suggests the presence of asymmetries in oil and gas exploration behavior, feeding into oil and gas supply (e. g., Kaufmann and Cleveland, 2001). Previous econometric applications include studies of price transmissions in commodity markets (e. g., Meyer and von Cramon-Taubadel, 2004; Grasso and Manera, 2006). Due to the limitation of the present data set, these ideas are left for future research. 20. Surveying econometric studies of US exploration activity, Dahl and Duggan (1997) find the average corresponding elasticity to exceed one. However, our results resemble previous studies of the Norwegian Continental Shelf (Mohn and Osmundsen, 2008) and also compare well to international assessments of exploration drilling activity (Ringlund et al., 2007). Both these studies suggest that the oil price elasticity of drilling could vary inversely with the degree of regulation. With high tax protection (78 per cent marginal tax), oil companies tend to shift exploration spending into Norway when the oil price is low, and out of Norway when the oil price is high. This tendency contributes to the explanation of moderate oil price elasticities of exploration drilling on the NCS.

Efforts and Efficiency in Oil Exploration / 71 tively low discovery rates, and with high expected discovery size. When oil prices are low, cash flows are constrained, and the risk appetite is more modest, exploration strategies are typically more cautious. Exploration efforts are reduced, and focused in areas with higher discovery rates and smaller expected field sizes (mature areas).21 Figure 2. Decomposed Elasticities of Reserve-generation Estimated partial and total elasticities by explanatory variable (per cent)*)

a) Semi-elasticity: percentage change in dependent variable from absolute change in depletion indicator.

The second line of Table 3 summarizes the role of access to exploration acreage. In the long run, an increase in the amount of available exploration acreage by 1 per cent will produce an increase in annual reserve additions by 0.44 per cent, according to the results. This effect has two significant sources. First, a modest increase in drilling activity is sustained even in the long-term when new acreage is offered, although the long-term impact is far smaller than the shortterm effect, and also not statistically significant. Second, new licensing rounds
21. As opposed to frontier exploration areas, mature areas are typically characterized by proven exploration models, producing fields, well-developed infrastructure, transport facilities and market access. Moreover, exploration activities in these areas are usually directed at smaller satellite fields which can be tied back to already producing facilities of larger reservoirs (in decline), without the large investments involved by stand-alone field developments in new oil and gas regions (Norwegian Petroleum Directorate, 2007).

72 / The Energy Journal have a positive effect on average discovery size. With drilling efforts focusing on the most prospective available blocks at any time, it is natural that new licensing rounds will result in higher average discovery size. 7. CONCLUSION An accurate understanding of the economics of oil and gas exploration is crucial to provide a complete representation of oil and gas supply. Based on standard assumptions of neoclassical producer behavior and modern time series econometrics, this study reveals new insights into the process of oil and gas exploration. The proposed econometric model provides a framework for specific analysis of exploration behavior, allowing detailed studies of oil price changes, exploration success, licensing policies, technological progress, and resource depletion. The application on NCS data suggests that material contributions to total resource growth work via exploration efficiency, and that studies of exploration efforts fall short of explaining the full process of reserve generation. The results establish a relation between exploration efficiency, oil price, licensing policies and seismic surveying activity. These insights from a highly regulated oil and gas province are useful both for companies and policy-makers, as oil and gas investment now find their way into new regions, with a higher degree of regulation and government intervention than Western oil and gas companies have been used to. My results clearly suggest that variables relating to economics, regulation, and technology play an important role for reserve generation through exploration, in addition to the physical process of geological depletion. Specifically, short-term fluctuations in drilling and reserve-generation are influenced by historical exploration success. Moreover, licensing policy also has a potential to affect both drilling activity and success rates in the short-term, but without continuous addition of new acreage, the long-term effects are muffled. The estimated effects for seismic surveying activity on the development of the success rate also suggest a significant role for technological progress in the model. According to the preferred econometric model, reserve-generation is enhanced by an increase in the oil price, with a long-term elasticity of 0.89. Both effort and efficiency in oil exploration is stimulated by an oil price increase, according to the results. When the oil price increases, oil companies take on more exploration risk. Consequently, discovery rates will fall whereas the average discovery size will increase. The net effect is an increase in yield per effort. The implication is a pro-cyclical pattern for both exploration risk and reserve-generation. There are also interesting policy implications to be made. First, our modelling approach clearly illustrates that policy measure need not and should not be limited to the regulation of drilling activity. With reserve additions as the ultimate target, the potential gains from measures to stimulate the success rate or average field size may well exceed the importance of drilling activity. Second, our results suggest that annual reserve additions are procyclical, due to the strong positive link between the oil price and average discovery size. If the government interest is

Efforts and Efficiency in Oil Exploration / 73 to stabilize reserve growth over time, this result provides a case for countercyclical exploration policies. A caveat of the present study relates to data sufficiency. The presented application is done on a rather small data set from the Norwegian Continental Shelf. The estimated models provide interesting new insights on the process of oil and gas exploration, but the empirical foundation of the results calls for additional assurance. An appraisal of the usefulness of the proposed modelling framework should therefore await the application of the model to other non-OPEC oil and gas provinces. An interesting route for further sophistication is to test the symmetry of the various sources of oil and gas resource growth. As an example, we suspect that an oil price drop has a different effect on exploration activity than an equivalent increase, at least in the short term. Corresponding asymmetries may also be relevant for accumulated discoveries and exploration acreage. To develop their base of oil and gas reserves, oil and gas companies balance their drilling efforts between exploration and production drilling. A topic for further research would also be to study how the balance between production and exploration drilling is affected by economic, geological and policy variables at best in a combined framework of investment behavior. ACKNOWLEDGMENTS This study has gained significantly from comments and suggestions from Frank Asche, Gang Liu, Petter Osmundsen, Knut Einar Rosendahl, delegates at the 29th IAEE International Conference in Potsdam 7-10 June 2006, seminar participants at the University of Stavanger, Statistics Norway, The Norwegian Petroleum Directorate, The Norwegian School of Economics and Business Administration, as well as three anonymous referees. The usual disclaimer applies. REFERENCES
Brdsen, G. (1989). Estimation of Long-Run Coefficients in Error-Correction Models. Oxford Bulletin of Economics and Statistics 51: 345-350. Dahl, C. and T. E. Duggan (1998). Survey of Price Elasticities From Economic Exploration Models of US Oil and Gas Supply. Journal of Energy Finance and Development 3 (2): 129-169. Dickey, D. A. and W. A. Fuller (1981). Likelihood Ratio Statistics for Autoregressive Time Series With a Unit Root. Econometrica 49 (4): 1057-1071. Dornik, J. A. and H. Hansen (1994). A Practical Test for Univariate and Multivariate Normality. Discussion Paper. Nuffield College. Dornik, J. A. and D. F. Hendry (2001). Modelling Dynamic Systems Using PCGive 10. London: TCL. Elliott, G., Rothenberg, T., and J. Stock. (1996). Efficient Tests for an Autoregressive Unit Root. Econometrica 64: 813-836. Engle, R. F. (1982). Autregressive Conditional Heteroscedasticity, with Estimates of the Variance of United Kingdom Inflation. Econometrica 50: 987-1007. Engle, R. F. and C. W. J. Granger (1987). Cointegration and Error-Correction: Representation, Estimation and Testing. Econometrica 55: 251-276.

74 / The Energy Journal

Erickson, E.W. and R. M. Spann (1971). Supply In a Regulated Industry: The Case of Natural Gas. Bell Journal of Economic and Management Science 2: 94-121. Farzin, Y. H. (1986). Competition in the Market for an Exhaustible Resource. JAI Press: Greenwich. Farzin, Y. H. (2001). The Impact of Oil Prices on Additions to US Proven Reserves. Resource and Energy Economics 23: 271-291. Favero, C. A. and M. H. Pesaran (1994). Oil Investment in The North Sea. Economic Modelling 11 (3): 308-329. Fisher, F. M. (1964). Supply and Costs in The US Oil and Gas Industry: Two Econometric Studies. Baltimore: Johns Hopkins Press. Forbes, K. F. and E. M. Zampelli (2000). Technology and the Exploratory Success Rate in the U.S. Offshore. The Energy Journal 21 (1): 109-120. Forbes, K. F. and E. M. Zampelli. 2002. Technology and the Exploratory Success Rate in The U.S. Onshore. Quarterly Journal of Economics and Finance 42 (2): 319-334. Grasso, M. and M. Manera (2006). Asymmetric Error Correction Models for the Oil- Gasoline Price Relationship. Economic Policy 35 (1): 156-177. Hendry, D. F. (1995). Dynamic Econometrics. Oxford: Oxford University Press. Hendry, D. F. and K. Juselius (2000). Explaining Cointegration Analysis: Part I. The Energy Journal 21 (1): 1-42. Hendry, D. F. and K. Juselius. (2001). Explaining Cointegration Analysis: Part II. The Energy Journal 22 (1): 75-120. Hendricks, K. and R. H. Porter (1996). The Timing and Incidence of Exploratory Drilling on Offshore Wildcat Tracts. American Economic Review 86 (3): 388-407. Hubbert, M. K. (1962). Energy Resources. A Report to the Committee on Natural Resources. Washington, D. C. National Academy of Sciences, National Research council. Iledare, O. O. (1995). Simulating the Effect and Policy Incentives on Natural Gas Drilling and Gross Reserve Additions. Resource and Energy Economics 17: 261-279. Iledare, O. O. and A. Pulsipher (1999). Sources of Change in Petroleum Drilling Productivity in Onshore Lousiana in the US, 1977-1994. Energy Economics 21: 261- 271. Johansen, S. (1995). Likelihood-Based Inference in Cointegrated Vector Auto-Regressive Models. Oxford: Oxford University Press. International Energy Agency (2006). World Energy Outlook. Paris: IEA. Kaufmann, R.K. and C. J. Cleveland (2001). Oil Production in the Lower 48 States: Economic , Geological, and Institutional Determinants. The Energy Journal 22: 27-49. Kemp, A. and S. Kasim (2003). An Econometric Model of Oil and Gas Exploration Development and Production in the UK Continental Shelf: A Systems Approach. The Energy Journal 24 (2): 113-141. Kemp, A. and S. Kasim (2006). A Regional Model of Oil and Gas Exploration in the UKCS. Scottish Journal of Political Economy 53 (2): 198-221. Kolb, J. A. (1979). An Econometric Study of Discoveries of Natural Gas and Oil Reserves in the United States 1948 to 1970. New York: Arno Press. Kremers, J. J. M., Ericsson, N. R., and J. Dolado (1992). The Power of Cointegration Tests. Oxford Bulletin of Economics and Statistics 54: pp 325-348. Managi, S., J. J. Opaluch, D. Jin and T. A. Grigalunas (2005).Technological Change and Petroleum Exploration in the Gulf of Mexico. Energy Policy 33 (5): 619-632. Meyer, J. and S. von Cramon-Taubadel (2004). Asymmetric Price Transmission: A Survey. Journal of Agricultural Economics 55 (3), 581-611. Ministry of Petroleum and Energy (2007). Facts 2007. The Norwegian Petroleum Sector. (http://www. petrofacts.no). Mohn, K. and P. Osmundsen (2008). Exploration Economics in a Regulated Petroleum Province: The Case of The Norwegian Continental Shelf. Energy Economics 30 (2): 303-320. Norwegian Petroleum Directorate (2007). The Petroleum Resources on The NCS 2005. (http://www. npd.no/English/Frontpage.htm). Pesaran, M. H. (1990). An Econometric Model of Exploration and Extraction of Oil in the UK Continental Shelf. Economic Journal 100: 367-390.

Efforts and Efficiency in Oil Exploration / 75

Phillips, P. C. B. and P. Perron (1988). Testing for a Unit Root in Time Series Regression. Biometrika 75: 335-346. Pindyck, R. S. (1974). The Regulatory Implications of Three Alternative Economic Supply Models of Natural Gas. The Bell Journal of Economics and Management Science 5: 633-645. Pindyck, R. S. (1978). Higher Energy Prices and the Supply of Natural Gas. Energy Systems and Policy 2 (2): 177-209. Quyen, N. V. (1991). Exhaustible Resources: A Theory of Exploration. Review of Economic Studies 58: 777-789. Reiss, P. C. (1990). Economic and Financial Determinants of Oil and Gas Exploration. In Hubbard, R. G., ed., Asymmetric Information, Corporate Finance and Investment. Chicago: University of Chicago Press. pp. 181-206. Ringlund, G. B., K. E. Rosendahl and T. Skjerpen (2007). Does Oilrig Activity React to Oil Price Changes? An Empirical Investigation. Energy Economics, 30 (2): 371-396. Uhler, R. S. (1979). The Rate of Petroleum Exploration and Extraction. in Pindyck, R. S. ed., Advances in the Economics of Energy and Resources. Vol. 2. Connecticut: JAI Press. pp. 93-118. White, H. (1980). A Heteroskedastic-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity. Econometrica 48: 817-838.

Appendix 1. Stationarity Tests for Model Variables Augmented Dickey-Fuller tests (ADF), Dickey-Fuller GLS tests (DF GLS), and Phillips-Perron tests (PP). Variable levels a)
Dependent variables dt ADF DF GLS PP -0.16 -1.01 0.39 st 0.82 -3.32* -1.29 mt -1.06 -3.67* -1.08 pt 0.36 -1.59 0.36 Explanatory variables et 0.45 -1.42 0.40 zt 0.83 -2.17 0.74 ht 3.38 -0.56 1.20

Variable changes a)
Dependent variables dt ADF DF GLS PP -3.1
***

Explanatory variables mt pt -4.3

***

st -4.1 -2.9 -8.0

*** ***

et -3.7
***

zt -6.3
***

ht -4.7*** -1.41 -11.7***

-5.1

***

-3.1** -3.4
***

-5.2*** -6.2
***

-4.3*** -6.0
***

-4.1*** -5.8
***

-8.3*** -9.3
***

Note: Dickey and Fuller (1981) introduced a popular procedure to test that a variable follows a unit-root process. The null hypothesis is that the variable is contains a unit root, with a stationary data-generating process as the alternative. With the augmented Dickey-Fuller test (ADF), a regression is run of the differenced

76 / The Energy Journal variable on its lagged level, as well as its lagged differences (sometimes also with a time trend): Dxt = 0 + 1xt1 + jxtj + vt
j =2 n

The unit root test involves a null hypothesis that 1 = 0, against 1 < 0. Non-rejection of the null is an indication that the variable expressed in levels (xt) is non-stationary. Elliot, Rothenberg and Stock (1996) propose a modified version of the Dickey-Fuller test, based on detrending via generalized least squares (DF GLS), for improved power and precision. The Dickey-Fuller test accounts for serial correlation by use of additional lags of the first-difference variable. Phillips and Perron (1988) introduced a modified variant of this test (PP), whereby a nonparametric correction of the standard errors is applied to capture serial correlation of the above regression.
*) a)

Significant at 90, **) 95 and ***) 99 per cent confidence level, respectively.  ll variables in natural logs; drilling effort (dt), success rate (st), average field size (mt), real oil price A (pt), licensed exploration acreage (et), seismic surveying activity (zt), depletion indicator (ht).

Appendix 2. Specification Tests and Model Diagnostics

Drilling activity dt LM (1,2) ARCH (1,2) Normality (LM) F(*,*) F(*,*) 2(*) 2.47* (0.10) 0.62 (0.44) 6.13** (0.05) 0.38 (0.96) Drilling success st 2.73* (0.09) 1.30 (0.26) 0.38 (0.83) 1.11 (0.48) Average discovery size mt 2.00 (0.16) 0.18 (0.68) 5.53* (0.06) 0.20 (1.00) Full model VECM 0.84 (0.65) .. 9.67 (0.14) 0.38 (1.00)

Heteroskedasticity F(*,*)

Note: LM (1,2) represents the Lagrange Multiplier test for autocorrelated residuals (Dornik and Hendry 2001), whereby estimated residuals are regressed on original variables and lagged residuals. The null is no autocorrelation. There are weak indications of autocorrelation in the drilling and success rate equations, but not for the systems version (VECM) of the test. The second line is the autoregressive conditional heteroskedasticity (ARCH) test (Engle 1982). The present form tests the joint significance of lagged squared residuals in the regression of squared residuals and lagged squared residuals. Again, the null is that the model is well-behaved. The ARCH test can not reject the null that the model is ok for any of the equations of the system. The third line is the Lagrange multiplier test

Efforts and Efficiency in Oil Exploration / 77 for normality of the residuals. This follows the procedure by Doornik and Hansen (1994), and tests whether the skewness and kurtosis of the residuals correspond to those of a normal distribution. Under the null, the residuals comply with the normal distribution. Again, we have indications of non-normality in the drilling and field-size equations. However, normality of the residuals is not rejected for the systems version of the test. The fourth line presents the test for heteroskedasticity proposed by White (1980). An auxiliary regression is involved, whereby the squared residuals are regressed on the original variables and their squares. By rejection, this test implies that the variance of the error process depends on the regressors and their squares. Non-rejection implies that the model is OK. No indications of heteroskedasticity are revealed by this test.
*)

Significant at 90, **) 95 and

***)

99 per cent confidence level, respectively. p-values in brackets.

APPENDIX 3.  LONG-TERM ELASTICITIES OF RESERVE GENERATION Letting all changes in Equations (17)-(19) approach zero, and solving for the dependent variables yields the following system of long-term equations: = 0.7 6* s + 0.38** p + 0.14** e d t t t t
(0.09) (0.01) (0.01)

(A4.1) (A4.2) (A4.3)

= 0.1 4 m 0.2 4** p + 0.2 2*** z s t t t t

(0.26) (0.08) (0.00)

= 0.5 5** p + 0.2 9*** e 1.2 5*** H m t t t t

(0.02) (0.00) (0.00)
*)

Significant at 90, **) 95 and

***)

99 per cent confidence level, respectively. p-values in brackets.

The long-term impact on reserve generation from a change in the oil price is the sum of the impact on each of the three long-term equations (cf Equation (12)). The straightforward solution would be to pick out the oil price coefficients directly from the long-term equations and calculate their sum. However, this approach will neglect the simultaneity of the three equations, and more sophistication is therefore required. As an example, the oil price effect on drilling activities (dt) will depend not only on the estimated direct oil price effect (0.38), but also on repercussions via the success rate (st) and average discovery size (mt). Accordingly, the overall oil price effect on drilling is calculated as: Dd d d s d s m = + + Dp p s p s m p (A4.4)

Applied to the estimated system of long-term equations, this yields:

78 / The Energy Journal Dd = 0.38 0.76 (0.24) 0.76 0.14 0.55 Dp (A4.5)

The result is an estimated long-term oil price elasticity of drilling at 0.51. A restriction that this combined coefficient equals zero is clearly rejected (p < 0.00). Consequently, the estimated long-term oil price elasticity of drilling is sig ), this nificant at a 99 per cent confidence level. For the success rate equation ( s t procedure gives an estimated long-term oil price elasticity at -0.17 (p=0.10), and ). In summary, we now have: 0.55 (p<0.02) for the field size equation ( m t Dr Dd Ds Dm = + + = 0.51*** 0.17* + 0.5 5** = 0.8 9*** 0.00) (0.10) (0.02) (0.00) Dp Dp Dp Dp (A4.6)

This is the procedure applied to derive the long-term elasticities of Table 3.

Threshold Cointegration Analysis of Crude Oil Benchmarks

Shawkat M. Hammoudeh*, Bradley T. Ewing** and Mark A. Thompson***

The paper examines the dynamic relationship between pairs of four oil benchmark prices (i.e., West Texas Intermediate, Brent, Dubai, and Maya), which have different physical properties and locations. The results indicate that there is a long-run equilibrium relationship between different benchmarks, regardless of their properties and locations. More importantly, there is asymmetry in the adjustment process that is specifically modeled and implications are discussed. 1. Introduction There are virtually hundreds of different grades of crude oil found under the surface of the earth.1 The different oil grades can be classified into three groups based on two basic physical properties: specific gravity and sulphur content.2 In terms of pricing, each of the three major oil groups has a benchmark or a marker, which is used as the reference in an equation that determines the grades own price after factoring in quality and transportation costs for the different locations. The benchmarks are economically important because they are traded on the commodity centers, as well as the spreads (or relationships) between two benchmarks are also traded and vital in the price discovery process of crude oil and its derivatives.
The Energy Journal, Vol. 29, No. 4. Copyright 2008 by the IAEE. All rights reserved. * Professor of Economics, LeBow College of Business, Drexel University, Philadelphia, PA 19104-2875. Phone: (215) 895-6673. E-mail: hammousm@drexel.edu. ** Rawls Professor in Operations Management, Rawls College of Business, Texas Tech University, Lubbock, TX 79409-2101. Phone: (806) 742-3939. E-mail: bradley.ewing@ttu.edu. *** Corresponding author. Cree-Walker Chair of Business Administration, Hull College of Business, Augusta State University, Augusta, GA 30904. Phone: (706) 667-4589, Fax: (706) 667-4064. E-mail: mthompson@aug.edu. 1. The 2006 International Crude Oil Handbook provides a description of more than 160 grades of crude oil that are traded on the world markets. 2. The standard measure of specific gravity is American Petroleum Institute (API) gravity; the higher this API number, the lighter the crude oil. The API gravity categorizes crude into three main types: Light, Medium and Heavy. The second property grades oil into sweet crudes that have relatively lower naturally occurring sulphur content or sour crudes that are higher in sulphur.

79

80 / The Energy Journal The benchmarks for the light groups are the West Texas Intermediate (WTI) in North America and Brent blend in Europe and Africa. WTI is typically a lighter and sweeter crude than Brent and it should naturally have a higher price but this is not always the case.3 West Texas Intermediate (WTI) and North Sea Brent (Brent) are the best-known oil price markers for crude oil produced on the world markets. The WTI-Brent (light/light) spread usually reflects a quality differential but is also affected by the existing fundamental and transitory factors in both markets. Natural disasters and refinery shutdowns and outages in the United States left more WTI idle on the market and helped tilt the spread in favor of Brent in 2007. Oil peaking in some North Sea oil fields also helped increase the price of Brent. A strike in the Fos-Lavera oil terminal near Marseille (France) also affected the spread in favor of Brent. Nigerian and Iranian oil grades are also priced according to a Brent-based formula. As the political situation gets tense in these countries, Brent shows a greater increase in price. The pricing difference between these two benchmarks may decline as US refiners finish maintenance and maximize gasoline output in time. Moreover, some countries may be faced with higher imports bill depending on which side of the spread moves against the oil they refine and consume. The Dubai benchmark representing the medium, sour crudes is priced in balance to WTI and Brent.4 This benchmark crude (which is now supplemented by Oman crude) is currently traded at the Dubai Mercantile Exchange (DME) and Londons International Commodity Exchange (ICE).5 As the world becomes more critically reliant on heavier and higher-sulfur streams, the emphasis is placed more on sour crudes, since the heavy and medium grades discussed above are assuming more importance in the price discovery process. Over half of the worlds oil supply these days is heavy sour, and more significantly, incremental production is primarily of sour crude from the Middle East and Latin America. Maya is a heavy, sour crude and sells at a significant discount to WTI and Brent.6 Roughly half of Mexicos oil exports are heavy Maya crude, which is exported to the United States, Europe and Asia. Mexican Maya is also benchmarked to WTI, forming the WTI- Maya (light/heavy) spread. In contrast to the previous benchmarks, it is worth noting that Maya is illiquid because it is not actively traded on any oil futures market.
3. Brent crude typically has an API gravity averaging 38.3 degrees and a sulphur content of 0.4 percent; and it is considered a light, sweet crude oil. WTI has an average API gravity of 39.6 degrees and a sulphur content of 0.2 percent, and thus is lighter and sweeter than Brent. 4. The medium crude group is benchmarked by Dubai-Oman crude, which has an API gravity of 31.0 degrees and sulphur content of 2.0 percent. Almost all the crude oil exported from the Middle East to Asian markets, which is currently over 10 million barrels a day, has been directly or indirectly tied to the spot value of Dubai crude oil (Montepeque, 2005). In 2001, Platts introduced Oman as an alternative delivery grade in Dubai contracts because the production of Dubai crude dwindled to 100,000 barrels a day from about 400,000 barrels in the 1980s. 5. Initially, DME ownership was equally shared by NYMEX and Tatweer (a part of Dubai Holding). These two partners currently hold 32.5 percent each, while Oman Investment Fund holds 30 percent and the remaining 5 percent stake has been allocated to the Exchanges floor members. 6. The heavy crude group is benchmarked by Mexican Maya that has an API gravity of 22 degrees and a sulphur content of 3.3 percent.

Threshold Cointegration Analysis of Crude Oil Benchmarks / 81 Thus, understanding the behaviors of the benchmarks is important not just because there are trades on their own contracts and their spreads in providing profitable opportunities but also because of their role in pricing other crude oil grades. Since changes in the spreads could bring arbitrage opportunities, it is important to understand the differential between these four oil markers in order to be able to maintain a balanced pricing relationship among the different grades of crude in their categories. However, because of active arbitrage activity in the oil market, these benchmark prices form a long-run relationship leading to an equilibrium (spread) position. Several important questions arise. First, how do these markers move into a long-run relationship following a disturbance, i.e., what is the nature of the adjustment? Second, do they move at the same speed of adjustment or does the adjustment process differ depending on whether the spread is widening or narrowing? Finally, can the long-run threshold for the spread be identified? There are several reasons to expect that the nature of the adjustment process in these oil markets to be asymmetric. In general, the presence of risk-averse (financial) market participants with heterogeneous expectations in North American, European, Middle Eastern, Asian and Latin American markets can lead to price asymmetry in these different grades. Some compulsive or noisy trading activities may also result in asymmetric adjustment process where a price is pushed up (or down) causing the spread to widen (or narrow) until informed traders react to the temporary deviation and push prices back to the equilibrium position (Cootner, 1962). Dumas (1992, 1994), Sercu et al. (1995), and McMillan (2005) have suggested that transaction costs may explain the asymmetric adjustment back to the equilibrium position. Financial market frictions and futures contracts availability, as well as institutional and regulatory constraints, are also major factors in oil market price mechanisms and may affect the convergence to equilibrium. Previous literature on oil spread asymmetry examines asymmetric transmission in the upstream and downstream markets by defining alternate regimes as periods of either rising or falling upstream prices. More recently Chen et al. (2005) and Ewing et al. (2006) investigated spread asymmetry in those markets using the threshold cointegration analysis and found strong evidence of asymmetric adjustment, but these studies did not include an examination of the spreads between different stream benchmarks. Milones and Henker (2001) examined the factors that affect the spread between WTI and Brent; however, they did not examine the asymmetric adjustment in global oil benchmarks. Moreover, examining a threshold value of zero, as some of the previous studies did, may lead to biased estimates of the cointegration model. The purpose of this paper is to model the dynamic adjustment process of the four major benchmark spreads using threshold cointegration techniques. It will be interesting and useful to know how the different spot prices adjust relative to different benchmarks (i.e., how the different spreads behave over time). In particular, the momentum-threshold autoregression (M-TAR) model of Enders and Siklos (2001) is employed in this analysis. We also estimate the asymmetric (M-TAR) error-correction model in order to examine the speeds of adjustment

82 / The Energy Journal of the different benchmarks in the four pairs when the spread is widening versus narrowing. Improvements in modeling these benchmark spreads should lead to improved forecasting models. The importance of forecasting oil benchmarks is beneficial for managing risk in investment portfolios, dynamic pricing models, and hedging and market timing strategies. As such, models that do not account for significant asymmetry in the adjustment process will lead to biased or skewed distribution of forecast errors. Therefore, our results indicate how forecasts could improve by incorporating asymmetry in the spreads. 2. Data and Methodology The data are comprised of daily time series for the closing spot price of the four oil grade benchmarks: WTI, Brent, and Dubai over the period January 2, 1990 to May 1, 2006 and Maya starting January 1, 1991. The spot prices for the crude oil benchmarks are prices quoted for immediate delivery of WTI at Cushing, Oklahoma; Brent at Londons International Petroleum Exchange (IPE), Dubai -Oman at Dubai Mercantile Exchange (DME), International Commodity Exchange (ICE) and Singapore Exchange among other Asian commodity centers; and Maya at NYMEX. Prices of crude oil are expressed in U.S. dollars per barrel. Data for WTI and Brent prices are accessed from the EIA website, while data for Dubai and Maya are obtained from Deutsche Bank. Not surprisingly, the averages of the four benchmark prices measured in US dollars over the sample period follow their physical properties (see Table 1). The lightest and sweetest WTI yields the highest mean, followed by Brent, Dubai and Maya in this order. Since we use prices in dollars instead of natural logs, we should define volatility in terns of the coefficient of variation so that the standard deviations will be scaled by their respective means. The coefficient of variation reveals that Brent, Maya and Dubai have the (relatively) higher volatility followed by WTI. Brent is the most volatile because there is a shift in its oil field towards production of less light and less sweet crude since its production peaked in 1995. Brent is also relatively affected by the geopolitical events in the oil producing countries (e.g., Iran and Nigeria) whose oil is benchmarked to it. The relatively high volatility of Maya is perhaps due to its location in a very active hurricane area, the complex refineries needed to process this oil type and the higher-than-expected decline in Mexican oil production. Dubai is influenced by the geopolitical events in the Middle East and by its dwindling production. It will be interesting to see what will happen to its volatility after the establishment of Dubai Mercantile Exchange (DME). WTI is still considered volatile because it is particularly sensitive to the size of its buildup storage in Cushing, Oklahoma. It is also influenced by the speculative behavior of financial institutions that use the stored WTI in Cushing. This price is also affected by the transitory factors that affect refineries in the Gulf of Mexico.

Threshold Cointegration Analysis of Crude Oil Benchmarks / 83 Figure 1. Crude Oil Benchmark Prices

Note: The graphs are in natural logs of the dollar prices over the sample period.

Table 1. Descriptive Statistics of Benchmarks and their Spreads in US Dollars

Benchmarks Mean Maximum Minimum WTI Brent Dubai Maya 26.332 24.736 22.512 19.646 73.750 73.960 67.110 58.030 10.82 9.100 9.340 5.990 Standard Deviation 12.079 11.901 10.737 9.392 Coefficient of Variation 0.459 0.481 0.477 0.478 0.674 0.563 0.808 0.525

Spreads WTI-Brent 1.596 7.020 -5.430 1.076 WTI-Dubai 3.820 18.280 -0.940 2.149 Brent-Dubai 2.224 14.530 -2.170 1.796 WTI-Maya 6.807 21.440 0.910 3.574

Note: The data are comprised of daily closing spot price of the four oil grade benchmarks: WTI, Brent, and Dubai over the period January 2, 1990 to May 1, 2006 and Maya starting January 1, 1991.

84 / The Energy Journal A casual review of Figure 1 illustrates that the four oil benchmark price series move closely together over time despite their very different physical properties and locations, i.e., they may be cointegrated with each other. Moreover, the highest contemporaneous correlation among the four benchmark prices is between the two light, sweet benchmarks WTI and Brent (0.996), and the lowest (0.970) is between the medium, sour (Dubai) and heavy, sour (Maya) as shown in Table 2-Panel A. It is interesting to see that Maya has consistently the lowest correlations with the other three benchmarks. The differences in correlations reflect the differences in the physical property, location, and volatility. Figure 2 displays the movements of the spreads between the four benchmarks. As expected, the highest variation in spread movements is between the WTI-Brent spread and that of the WTI-Maya spread due to differences in the physical properties of the benchmarks. The figure illustrates that at certain times the spreads narrow or widen considerably while at other times they turn negative, providing plenty of arbitrage opportunities for oil traders. Recently, the spreads have narrowed significantly across the board regardless of the physical properties and location because of the tightness and backwardation in the oil market.7 In fact, the spread between WTI and Brent temporarily reversed the course and turned negative, which bestowed the title broken benchmark or the weakest link on WTI.8 In comparing the contemporaneous correlations between those spreads, the highest correlation is between (WTI, Dubai) and (Brent, Dubai) spreads, but the lowest positive correlation is between (WTI-Maya) and (Brent-Dubai) spreads reflecting greater disparity in terms of physical properties and locations (see Table 2-Panel B). The correlation between (WTI, Brent) and (Brent, Dubai) spreads is negative, possibly qualifying those spreads as candidates in a diversified portfolio. In addition, unit-root testing confirms that each of the (in natural logarithm) spot prices are I(1). Full results of the augmented Dickey-Fuller (1981), Phillips-Perron (1988), and the Kwiatkowski et al. (1992) unit root tests are reported in Table 3. Given that each series is stationary in first-differences, cointegration analysis is usually employed. Many previous studies have found that oil prices are cointegrated in the conventional symmetric sense (e.g., Hammoudeh et al., 2003). However, conventional cointegration tests are misspecified when the adjustment process is asymmetric. The Enders-Siklos (2001) technique extends the popular two-step Engle-Granger (1987) symmetric cointegration methodology to test for long-run relationships between two time series while allowing for adjustment asymmetry. The resulting M-TAR testing procedure has shown good power and size properties relative to the assumption of symmetric adjustment, and is thus the basis for our analysis (see Enders and Granger, 1998). The first step in the symmetric Engle-Granger (1987) and asymmetric Enders-Siklos (2001) framework is to estimate the following model representing the long-run relationship between two crude oil spot benchmark prices using ordinary least squares:
7. Backwardation is a market situation in which the spot price exceeds the futures price. 8. For more information on the broken benchmark, please see Fattouh (2007).

Threshold Cointegration Analysis of Crude Oil Benchmarks / 85 Figure 2. Crude Oil Benchmark Spreads (US $)

Note: Each of the oil benchmark spreads is in US dollars over the sample period of January 2, 1990 to May 1, 2006 (except for the WTI-Maya spread, which starts January 1, 1991).

Table 2. Contemporaneous Correlations between the Oil Benchmark Prices and Spreads
Panel A: Correlations between prices WTI Brent Dubai Maya WTI-Brent WTI-Dubai Brent-Dubai WTI-Maya WTI Brent Dubai Maya 1 0.996 1 0.988 0.990 1 0.971 0.973 0.970 WTI-Brent WTI-Dubai Brent-Dubai

1 WTI-Maya

Panel B: Correlations between spreads 1 0.4408 1 -0.243 0.763 1 0.319 0.295 0.089

Note: The correlations are estimated for WTI, Brent, and Dubai, over the daily period January 2, 1990 to May 1, 2006 and for Maya over the period starting January 1, 1991 to May 1, 2006.

86 / The Energy Journal Table 3. Unit Root Tests

WTI WTI Brent Brent Dubai Dubai Maya Maya ADF -2.156 -41.761** -2.109 -62.792** -2.309 -42.237** -2.325 -70.064** PP -2.105 -66.673** -2.075 -62.760** -2.255 -71.569** -2.445 -70.027** KPSS 1.206** 0.193 1.177** 0.1811 1.085** 0.162 0.954** 0.128

Notes: * (**) denotes significance at the 5 (1) percent level. Both the ADF and PP test the null hypothesis of unit root, whereas the KPSS tests the null hypothesis of stationarity.

Pti = c + b1Ptj + et

(1)

where Pti and Ptj are the logarithmic values of the i and j spot benchmark prices at time t.9 Four sets of oil spot benchmark spreads are examined: (i, j) = (WTI, t, derived Brent), (WTI, Dubai), (Brent, Dubai), and (WTI, Maya).10 The residuals, e from equation (1) are then used to estimate the following Engle-Granger second step (from equation 2a) and the Enders-Siklos (2001) M-TAR model (from equation 2b): t = r0 e t1 + gt D e t1 + et De
i =1 n

(2a)
n i =1

t = r1 Mt e t1 + r2 (1 Mt) e t1 + gt D e t1 + et De

(2b)

where et ~ I.I.D(0, s2), 0, 1 and 2 are the speeds of adjustment parameters for the symmetric (i.e., 0) and asymmetric (i.e., 1 and 2) cointegration models, re are meant to yield uncorrelated residuals. spectively, and the lagged values of D e The heaviside indicator function for equation (2b) is denoted as follows: t1 1 if D e Mt = t1 < 0 if D e

(3)

t is greater than the threshold, the spread is widening and indicating when D e t is less than the threshold, the spread is narrowing. Assuming a threshold when D e value of 0 may lead to a biased estimate if there is asymmetry in the adjustment

9. Transforming the data series into natural logarithms has the advantage of distinguishing between given changes in spot prices when prices are high versus when they are low. 10. Figure 2 illustrates the arithmetic benchmark spread for each of the four cases.

Threshold Cointegration Analysis of Crude Oil Benchmarks / 87 process.11 Therefore, the threshold,, is endogenously determined using Chans (1993) method to find the consistent estimate of the threshold. This method art} in ascending order, excludes the smallest and largest 15 ranges the values {D e percent, and the parameter that yields the smallest sum of squared residuals over the remaining 70 percent is the consistent estimate of the threshold. The critical values to test the joint null hypothesis of no cointegration (i.e., 1 = 2 = 0) are reported in Enders and Siklos (2001). Since the adjustment t1 (r2 e t1) when the D e t1 is process in the M-TAR equation is modeled by r1 e above (below) the endogenous threshold and if the null hypothesis of no cointegration is rejected, then the null hypothesis of symmetric adjustment (i.e., 1 = 2) can be tested using the usual F-statistic. If the joint null hypothesis of symmetric adjustment is rejected and | r1 | > | r2 |, then increases in the spread tend to revert back toward the threshold faster than decreases. In other words, the speed of adjustment when the spread is widening is faster than when the spread is narrowing. This implies that traders and the different oil markets respond differently to the direction the spread is moving, and not just if the spread is out of alignment (or in disequilibrium). 3. Empirical Results and Discussion The two step Engle-Granger (1987) cointegration approach inherently assumes symmetric adjustment back to the long-run equilibrium position. If there is asymmetry in the adjustment process where adjustments move at different speeds depending on the position of the spread relative to the threshold, then symmetric cointegration and conventional error-correction models may be misspecified. This model misspecification would lead to biased pricing models and generate skewed forecasting errors. Therefore, the M-TAR model is estimated and the results are provided in Table 4. The endogenously-derived threshold is estimated for each oil benchmark spread. While these four thresholds may appear not to be much different than 0, the AIC indicates that these models with the consistent estimate of (daily) threshold are better than assuming a threshold value of 0. An out-of-sample forecast comparison was performed using the M-TAR and Engle-Granger model for each of the benchmark spreads. While we found no statistical difference in the outof-sample comparisons, the in-sample evidence suggests the M-TAR model be used for future modeling of oil benchmarks. Each threshold signals the change in the spread needed to adjust asymmetrically back to the long-run position. The joint null hypothesis of no cointegration with M-TAR adjustment (i.e., 1 = 2 = 0) is rejected for each oil benchmark spread given that the -statistic exceeds the respective critical value (Enders and Siklos, 2001). Given this finding, the following null hypothesis of symmetric adjustment (i.e., 1 = 2) was tested and
11. For example, one possible explanation is that transaction costs may differ depending on the spread widening or narrowing due to the different benchmarks and their properties within the spread or traders may respond differently to positive and negative shocks in certain spreads.

88 / The Energy Journal Table 4. M-TAR Cointegration Tests

1 2 1= 2 lags WTI-Brent 0.017 -0.157 -0.075 64.420** 19.41** 5 WTI-Dubai 0.011 -0.055 -0.025 20.842** 6.984** 9 Brent-Dubai -0.018 -0.023 -0.079 23.216** 14.823** 9 WTI-Maya -0.011 -0.006 -0.063 35.348** 26.483** 4

Notes: is the estimated threshold. The represents the sample values for the joint hypothesis of 1 = 0 and 2 = 0 (see Enders and Siklos, 2001 for critical values). 1= 2 tests the null hypothesis of symmetric adjustment. * (**) denotes significance at the 5 (1) percent level.

Figure 3. Asymmetric Adjustment Paths of WTI-Brent Spread

Notes: The top half illustrates the adjustment path after a positive shock (i.e., spread widens relative to long-run position). The adjustment path results in the spread narrowing back to its longrun equilibrium spread. The bottom half (negative shock) illustrates the speed of adjustment as the spread widens back to its long-run position.

also rejected in favor of the alternative for all four oil benchmark spreads. Note that the estimates for 1 and 2 indicate faster convergence toward long-run equilibrium spread when the spread is increasing or widening than when the spread is decreasing or narrowing for the WTI-Brent and WTI-Dubai price spreads. How-

Threshold Cointegration Analysis of Crude Oil Benchmarks / 89 ever, the opposite is the case for the Brent-Dubai and WTI-Maya spreads. These results imply that arbitrageurs are more active in exploiting larger profitable opportunities depending on the spread and the direction the spread is moving from its long-run position. In addition, these results indicate that the relationship between benchmarks is dependent on the physical properties of the benchmarks themselves. Moreover, in any of the four spreads the more recognized leader does most of the adjustment. The benchmark that does most of the adjustment tends to be the less volatile one. It also seems that the transaction costs vary among the different spreads depending on whether they are above or below the threshold. The costs, which are determined by the fundamental and transitory forces of the paired benchmarks, affect profitability and influence the direction of the adjustment towards equilibrium over time. In order to illustrate this finding, we computed and compared the adjustment paths from a positive and negative shock (of equal magnitude) to the WTI-Brent spread based on the results from equation (2b) (see Figure 3). After a positive (negative) shock, the spread will adjust back to the long-run equilibrium relationship. However, the adjustment back to the equilibrium position is slower when the spread narrows from a positive shock than when the WTI-Brent spread widens from a negative shock. The asymmetric adjustment paths (of the spread) further highlight the potential forecasting errors when using a model that assumes symmetric adjustments. The finding of cointegration with the M-TAR adjustment lends itself to estimating the asymmetric error-correction models.12 The M-TAR vector errorcorrection model differs from the conventional error-correction model by allowing asymmetric adjustments toward the long-run equilibrium. Such a specification recognizes the fact that traders respond differently depending on whether the different oil benchmark spreads widen or narrow. 4. M-TAR Error-Correction Models In order to better understand the relationships between the four different oil benchmarks, error-correction models are estimated. The results are given as follows:13

12. The null hypothesis of no cointegration was also rejected using the Engle-Granger (equation 2a) model and the M-TAR (equation 2b with =0) model for each oil benchmark spread. According to AIC, the M-TAR model with consistent estimate of the threshold fits the data better than the Engle-Granger and M-TAR (=0) models. Empirical results of these other versions are available upon request. 13. In each equation, Aij(L) represent the first-order polynomials in the lag operator L as suggested by the Akaike information criterion (AIC). The Fij represents the p-value for the joint null hypothesis that all coefficients of Aij in the polynomials are equal to zero on the lagged values of the changes in the futures and the spot prices, respectively. The t-statistics are reported in parentheses. Serial LM test fails to reject the null hypothesis of no serial correlation; lnL is the log likelihood; the F-statistic is reported with * (**) representing significance at the 5 (1) percent level or less.

90 / The Energy Journal Case 1. WTI-Brent Spread t1 0.0 381(1 Mt) e t1 DPtWTI = 0.00 06 0.13 6 9 Mt e
(1.51) ( 4.67) (2.63)

WTI + A11 (L ) DPt + A12 (L ) DPt1Brent + e1t 1 (F11< 0.01) (F1 2<0 .61)

Serial LM test =4.39 ( p value = 0.12); ln L = 9,706.52; F statistic = 11.12**

(0.74) ( 1.60) (3.76)

t1 + 0.0 487(1 Mt) e t1 DPtBrent = 0.00 03 + 0.0 416 Mt e

WTI Brent + A21 (L ) DPt + A22 (L )DPt + e2t 1 1 (F21< 0.01) (F22< 0. 01)

Serial LM test =2.58 ( p value = 0.27); ln L = 10,165.11; F statistic = 94.71**

As previously mentioned, WTI and Brent are similar in quality, highly liquid, and actively traded on oil futures markets. The contract tradability reduces transaction costs, facilitates arbitrage activity, and spurs error-correction dynamics. The results show that WTI takes on an active role and responds in the long run when the spread is narrowing and widening but at different speeds of adjustments. While traders of WTI are active in both directions of the threshold, WTI adjusts slower to narrowings of the spread than widenings. WTI also responds in the short-run to its own past changes (but not to past Brent changes), probably reflecting the historical leadership position of the broken benchmark. The arbitrage opportunities on WTIBrent spread may be limited due to the low share (0.6%) of Brent in total US oil imports (Fattouh, 2007). In contrast, Brent responds to the spread narrowing only. In the short run, Brent responds to past changes in WTI and Brent spot prices. The North Sea benchmark nudges the spread to adjust to its long-run equilibrium by increasing; hence, the spread narrows toward the threshold. Case 2. WTI-Dubai Spread t1 0.0 218(1 Mt) e t1 DPtWTI = 0.0004 0.04 54 Mt e
(1.06) ( 3.06 ) (2.61)

WTI Dubai + A11 (L )DPt + A12(L ) DPt + e1t 1 1 (F11< 0.05) (F12< 0. 47)

Serial LM test =1.40 ( p value = 0.50); ln L = 9,702.14; F statistic = 7.89**

Threshold Cointegration Analysis of Crude Oil Benchmarks / 91 t1 + 0.0 121(1 Mt) e t1 DPtDubai = 0.00 03 + 0.0174 Mt e
(0.77) ( 1. 07) (1.44)

WTI Dubai + A21 (L ) DPt + A22 (L )DPt + e2t 1 1 (F21< 0.01) (F22< 0. 01)

Serial LM test =2.82 ( p value = 0.24); ln L = 9,531.21; F statistic = 38.41** This is a case of a spread between light and medium heavy benchmarks where the Light is more actively traded than the Medium Heavy. In this case, WTI responds to both widening and narrowing of the spread and to its own past changes, again reflecting its prominent leadership role and high liquidity. The WTI speed of adjustment to long-run equilibrium is faster when the spread is widening as in Case 1 for the light sweet benchmarks. The less liquid and less actively traded Dubai only responds to past changes in WTI and Dubai (but not significantly in the long run to positive and negative disequilibria), reflecting higher transaction and refining costs. Therefore, major differences between the medium Case 2 and the light Case 1 are that WTI adjusts much more sluggishly from both directions of the threshold in the medium case than in Case 1, and the medium Dubai benchmark does not adjust to long-run equilibrium as is the case with Brent. The Dubai benchmark adjusts to short run changes in WTI and its own past changes, because it is not as actively traded and does not have the leadership prominence as WTI. In fact, its Oman futures contact suffers from occasional erratic price behavior, dominance of few traders, and overemphasis on physical delivery (Fatouh, 2008).14 Case 3. Brent-Dubai Spread t1 0.0 795(1 Mt) e t1 DPtBrent = 0.00 01 0.0 1 61 Mt e
(0.30) ( 2.13 ) (4.09)

Brent Dubai + A11 (L )DPt + A12(L ) DPt + e1t 1 1 (F11< 0.01) (F12< 0. 01)

Serial LM test =0.52 ( p value = 0.77); ln L = 10,022.21; F statistic = 32.49**

t1 + 0.0 062(1 Mt) e t1 DPtDubai = 0.00 04 + 0.0 17 4 Mt e

(0.93) (1.03) (0.27)

Brent Dubai + A21 (L ) DPt + A22 (L )DPt + e2t 1 1 (F21< 0.01) (F22< 0. 01)

Serial LM test =3.64 ( p value = 0.16); ln L = 9,473.33; F statistic = 14.93**

14. The ratio of physical delivery to totally volume for Dubai/Oman contract is 22 percent, which is the highest in the world for oil futures contracts.

92 / The Energy Journal This case is relevant to the 10 million barrels a day of Middle Eastern oil that ships primarily towards Asia. The spread in this case is more popular in the commodity centers than in Case 2. In Case 3, the very actively traded Brent responds to both positive and negative disequilibria as is with WTI in Case 2 and also to past changes in both Brent and Dubai. However, in contrast to WTI which adjusts faster from widenings in the spread in the previous case, Brent in Case 3 adjusts much faster to narrowings in the spread. In addition, the less liquid and actively traded Dubai responds just to past changes in Brent and Dubai (and not to the long-run disequilibria), which is similar to the WTI and Dubai benchmarks in Case 2, reflecting less world-wide leadership and higher transaction costs. For the DME futures price to adjust more efficiently in the short and long run, Oman should abandon the price retroactivity by relinquishing its parallel official pricing system and using DMEs futures prices. To increase liquidity, DME should concentrate on introducing more financially settled contracts to attract more financial players who will trade spreads. Case 4. WTI-Maya Spread t1 0.0 252(1 Mt) e t1 DPtWTI = 0.00 03 0.0 0 3 0 Mt e
(0.81) ( 0.68) (2.31)

WTI Maya + A11 (L )DPt + A12(L ) DPt + e1t 1 1 (F11< 0.01) (F12< 0. 02)

Serial LM test =3.04 ( p value = 0.22); ln L = 9,311.73; F statistic = 7.66**

t1 + 0.0396(1 Mt) e t1 DPtMaya = 0.00 02 + 0.0 37 Mt e

(0.53) (0.61) (1.74)

WTI Maya + A21 (L ) DPt + A22 (L ) DPt + e2t 1 1 (F21< 0.01) (F22< 0. 02)

Serial LM test =1.19 ( p value = 0.55); ln L = 8.600.97; F statistic = 22.44**

This case that involves the light WTI and the heavy Maya, is different from Case 2 (the WTI-Dubai case) because Maya is not traded on any futures markets, while Dubai is. As mentioned before, contract tradability and liquidity facilitates arbitrage activity and enhances error-correction dynamics. In contrast to Case 2, WTI adjusts from narrowings of the spread, and at a slower pace. WTI also responds to past changes in WTI and Maya. While in Case 2 Dubai benchmark does not adjust to long-run disequilibrium, the heavy Maya responds to spread narrowings (at 10%) and to past changes in both WTI and Maya. In this Case 4, Maya marginally adjusts upward to narrow the spread in order to move

Threshold Cointegration Analysis of Crude Oil Benchmarks / 93 toward its long-run position, still reflecting lack of strong leadership, liquidity, higher transaction, and refining bottleneck costs. 5. Concluding Remarks In the oil market, prices of hundreds of crude oil grades extracted from different locations on this planet are based on prices of few benchmarks. Each pricing equation takes into account the benchmark price pertinent to the group in which the oil grade in mind belongs, as well as the differentials due to API gravity, sulphur content, and geographic location. This paper examines the spreads between four major benchmarks: WTI, Brent, Dubai (Oman) and Maya. These benchmarks play a major role in the price/discovery process of the oil market as they are traded daily at the worlds major commodity centers. The oil participants also trade these benchmarks based on their own prices and also against each other in the form of spreads when they search for arbitrage opportunities. Our results indicate that each of the benchmark pairs forming a spread in the four bivariate groups has a long run, stable relationship. This is not surprising because they belong to one great pool and the oil markets are globalized (Gulen, 1997; 1999). When these spreads get out of alignment (i.e., deviate from the longrun position), there may be arbitrage opportunities enhanced by understanding the asymmetric adjustment process for each spread. It is possible that the magnitude of transaction costs vary depending on whether the benchmarks are liquid and actively traded in futures markets and on the position of the spread from the threshold. This helps produce the asymmetric adjustment to equilibrium. While WTI and Brent are highly liquid, very actively traded, and adjust to equilibrium in the long run, Dubai is not as liquid. In addition, Dubai has relatively high transaction (arbitrage) costs making it less responsive to spread changes over time. The establishment of DME in Dubai and the supplement Dubai crude with Oman crude may result in the Dubai benchmark becoming more liquid, more actively traded and being one to watch for in the near future. DME should work on creating more financially settled futures contracts to boast liquidity. Maya is not actively traded on futures markets and seems to adjust marginally to the long-run equilibrium but still surprisingly does better than Dubai up to date. Therefore, these results reflect less liquidity and lower leadership stature on part of Dubai and Maya and more on WTI and Brent. It is possible Dubai and Maya are likely to undergo adjustment changes in the future as their institutional and fundamental factors change. For Dubai, the change may come soon after the successful trading at DME. Furthermore, Mayas spread with WTI may change and have different adjustments because Mexican oil production is falling more than expected and the world will build more complex refineries.15
15. Millard and Luhnow (2008) contend that the Mexicos oil industry is in decay. Its oil output has declined steadily from a peak of 3.4 million barrels a day in 2004 and is expected to fall to 2.7 million by the end of 2008. It this pattern continues, Mexico will likely stop exporting oil in seven years. The problem is not related to reserves but to the oil industry management.

94 / The Energy Journal The production of the broken benchmark, WTI, has fallen to about 400,000 barrels a day and it has infrastructure logistics, which has cast some doubts on its leadership and benchmark positions. We must, however, note that the broken benchmark phenomenon for WTI is transitory and can still occur sporadically particularly if the pipeline logistics are solved. There is also a seasonality factor to this phenomenon occurring in the shoulder periods of spring and autumn when refinery capacity in the Midwest is taken out for maintenance and not much can be done with WTI inventories, giving Brent a higher premium. However, the results show clear adjustment to equilibrium and strong leadership for WTI and Brent in the long run. Since the four benchmarks have different liquidity, adjustments and spreads, traders strategies to take advantage of profit opportunities should be different and spread-dependent. As such, an avenue for future research would incorporate these asymmetric error-correction models to evaluate the forecasting ability compared to the symmetric approach. Acknowledgement The authors wish to thank the editor, James Smith, associate editor, Geoffrey Pearce, and two anonymous reviewers for helpful comments on an earlier draft, as well as, Adam Sieminski and Amanda Lee of Deutsche Bank for providing the data. References
Balke, N.S., S.P.A. Brown, and M.K. Ycel (1998). Gasoline and Crude Oil Prices: An Asymmetric Relationship? Federal Reserve Bank of Dallas Economic Review, First Quarter: 2-11. Borenstein, S., A.C. Cameron, and R. Gilbert (1997). Do Gasoline Prices Respond Asymmetrically to Crude Oil Price Changes? Quarterly Journal of Economics 112: 305-339. Chan, K. (1993). Consistency and Limiting Distribution of the Least Squares Estimator of a Threshold Autoregressive Model. Annals of Statistics 21: 520-533. Chaudhuri, K. (2001). Long-run Prices of Primary Commodities and Oil Prices. Applied Economics 33: 531-538 Chen, L-H., F.M. Finney, and S.L. Kon (2005). A Threshold Cointegration Analysis of Asymmetric Price Transmission from Crude Oil to Gasoline Prices. Economics Letters 89: 233-239 Dickey, D. and W. Fuller (1979). Likelihood Ratio Statistics for Autoregressive Time Series with a Unit Root. Econometrica 49: 1057-1072. Dumas, B. (1992). Dynamic Equilibrium and the Real Exchange Rate in a Spatially Separated World. Review of Financial Studies 5: 153-180. Dumas, B. (1994). Partial Equilibrium Versus General Equilibrium Models of the International Capital Market. In F. Van Der Ploeg, ed., The Handbook of International Macroeconomics. Oxford: Blackwell. Enders, W. and C. Granger (1998). Unit-Root Tests and Asymmetric Adjustment with an Example Using the Term Structure of Interest Rates. Journal of Business and Economic Statistics 16: 304-311. Enders, W. and P. Siklos (2001). Cointegration and Threshold Adjustment. Journal of Business and Economic Statistics 19: 166-176. Engle, R. and C. Granger (1987). Cointegration and Error Correction Representation, Estimation and Testing. Econometrica 55: 251-276.

Threshold Cointegration Analysis of Crude Oil Benchmarks / 95

Ewing, B., S. Hammoudeh, and M. Thompson (2006). Examining Asymmetric Behavior in U.S. Petroleum Futures and Spot Prices. The Energy Journal 27: 9-23. Fattouh, B. (2007). WTI Benchmark Temporarily Breaks down: Is it Realy a Big Deal? Middle East Economic Survey 49(20), May 14. Fattouh, B. (2008). Prospects of the DME Oman Crude Oil Futures Contract. Middle East Economic Survey 51(13), March 31. Godby, R., A. M. Lintner, T. Stengos, and B. Wandschneider (2000). Testing for Asymmetric Pricing in the Canadian Retail Gasoline Market. Energy Economics 22: 349-368. Gulen, S. G. (1999). Regionalization in the World oil Market: Further Results. The Energy Journal 20: 125-39. Gulen, S.G. (1997). Regionalization in the World oil Market. The Energy Journal 18: 109-126. Hammoudeh, S., H. Li, and B. Jeon (2003). Causality and Volatility Spillovers among Petroleum Prices of WTI, Gasoline and Heating oil in Different Locations. North American Journal of Economics and Finance 14: 89-114. Kwiatkowski, D., P. Phillips, P. Schmidt, and Y. Shin (1992). Testing the Null Hypothesis of Stationarity Against the Alternative of a Unit Root: How Sure are we that Economic Time Series have a Unit Root? Journal of Econometrics 54:159-178. McMillan, D. (2005). Non-Linear Dynamics in International Stock Market Returns. Review of Financial Economics 14: 81-91. Millard, P. and D, Luhnow (2008). Crunch Time for Mexican Oil. Energy Bulletin, February 16. www.energybulletin.net/40416.html Milones, M.T. and T. Henker (2001). Price Spread and Convenience Yield Behavior in the International Oil Market. Applied Financial Economics 11: 23-36. Montepeque, J. (2005). Sour Crude Pricing: A Pressing Global Issue. Middle East Economic Survey 48(14), April 4. Phillips, P. and P. Perron (1988). Testing for a Unit Root in Time Series Regression. Biometrika 75: 335-346. Sercu, P., R. Uppal, and C. Van Hulle (1995). The Exchange Rate in the Presence of Transaction Costs: Implications for Tests of Purchasing Power Parity. Journal of Finance 50: 1309-1319. Xiaowen, S. and M. Tamvakis (2001). Spillover Effects in Energy Futures Markets. Energy Economics 23: 43-56.

96 / The Energy Journal

An Empirical Analysis of Urban Form, Transport, and Global Warming

Fabio Grazi* Jeroen C.J.M. van den Bergh** and Jos N. van Ommeren*** Does urban form affect travel choices and thus CO2 emissions by individuals? If this is the case, then urban form and policies that influence it deserve serious attention in the context of long-term climate policy. To address this issue, we examine the impact of urban density on commuting behavior, and the consequences for CO2 emissions. The empirical investigation is based on an instrumental variable approach (IV), so as to take account of endogeneity of residence location. We decompose travel demand into components related to modal split and commuting distance by each mode. 1. INTRODUCTION Despite its short history, research on climate policy has generated an immense literature. This paper addresses a theme that has received little attention so far, namely the effect of urban form on greenhouse gas (GHG) emissions through transport. Urban form covers such aspects as density, geometric shape, use of land (residential, industrial), and infrastructure (road, rail, waterway),
The Energy Journal, Vol. 29, No. 4. Copyright 2008 by the IAEE. All rights reserved. * International Research Centre on the Environment and Development (CIRED, umr CNRS/ EHESS), 45 bis Avenue de la Belle Gabrielle, 94736 Nogent sur Marne CEDEX, France. Paris Institute of Technology, Paris, France. Author for correspondence. Fax: +33 (0)1 4394 7370. E-mail: grazi@centre-cired.fr (F. Grazi). ** ICREA, and Institute of Environmental Science and Technology & Department of Economics and Economic History, Autonomous University of Barcelona, 08193 Bellaterra (Cerdanyola), Spain. Department of Spatial Economics, Faculty of Economics and Business Administration, Free University De Boelelaan 1105, 1081 HV Amsterdam, The Netherlands. Institute of Environmental Studies, Free University, Amsterdam, The Netherlands. *** Department of Spatial Economics, Faculty of Economics and Business Administration, Free University De Boelelaan 1105, 1081 HV Amsterdam, The Netherlands. Jos van Ommeren wishes to thank the Frish Centre in Oslo, Norway, for hospitality.

97

98 / The Energy Journal with implications for indicators such as density, fragmentation and accessibility.1 Urban form may be affected by policies in a number of ways. In Europe, spatial (or physical) planning is the most common approach, which covers the set of policy instruments available to regulate and manage land use in the broadest sense of the word, including infrastructure, settlements, parking spaces, etc (Grazi and van den Bergh, 2008). Spatial planning can take the form of zoning of business areas, regulating density (and thus height) of buildings, and investment in infrastructure. Studying the role of urban form in global warming is relevant for two reasons. First, there is likely a close relationship between urban structure and global warming through GHG emissions and energy consumption from economic activities located in cities. Second, urban form affects transport activity, which in turn contributes to global warming (Greene and Schafer, 2003). The latter relationship is the focal point of this article. Worldwide CO2 emissions from transport have been increasing rapidly over the last two decades. According to last available data, in 2004 the transport sector accounted for one-fifth of the world energy-related CO2 emissions (IEA, 2006). This illustrates the relevance of analyzing transport in the context of global warming. What is more, this share is expected to grow at a rate of 1.7% per year up to 2030 (IEA, 2006). The annual growth rates of CO2 emissions by the transport sector in the developing world and in economies in transition are projected to be even higher, namely 3.4% and 2.2%, respectively (IEA, 2006). These increases are mainly due to increases in the volume of road transport, with passenger movements accounting for 6070% of increases in total emissions (IEA, 2004). Reducing GHG emissions by transport leads to considerable ancillary benefits, as other environmentally harmful gases (notably NOx and SO2 and particulate matter) resulting from fossil fuel combustion will be reduced as well (e.g., Elkins, 1996). This in turn improves local health and well-being. In addition, spatial planning aimed at mitigating GHG emissions might reduce traffic congestion. All in all, this suggests that direct local and global benefits of climate policy through spatial and transport planning go well together. Recognizing this potential, leading international institutions (OECD, 2007) and national governmental agencies (EPA, 2001) consider land-use policies especially in urban areas as an effective instrument to combat the contribution of transport to global warming. The impact of urban form on urban transport and associated GHG emissions follows a number of routes. Urban form affects transport distances and the number of trips made between home, work and service areas like city centers and shopping areas.2 In particular, residential density affects travel behavior by getting people in close proximity to destinations and consequently increasing the
1. Harvey and Clark (1965), Anas et al. (1989), and Burchfield et al. (2006) offer excellent syntheses and economic interpretations of the genesis and development of urban forms over the last two centuries. 2. A choice between urban concentration and urban sprawl is relevant here, which involves both environmental and welfare considerations. In the US this has given rise to a debate about Smart Growth strategies (Bertaud, 2003; Handy et al., 2005).

An Empirical Analysis of Urban Form, Transport, and Global Warming / 99 number of possible destinations that can be reached within the same range of distance. Here we focus on the impact of residential density on commuting behavior, but our findings seem to be extendable to other travel choices.3 A number of theoretical studies suggest a negative relationship between commuting distance and residential density (e.g., Fujita, 1989). The main reason is that workers choose a residential location (conditional on workplace location) by making a trade-off between generalized commuting cost (or distance or duration) and housing prices. In high density areas, workers benefit from a shorter commuting distance but face higher housing prices, whereas in low density areas they enjoy lower prices but travel longer distances. In addition, changes in modal split, notably a shift from car to more energy-efficient transport modes like walking, biking and public transport will contribute to a reduction in transport-related GHG emissions. Here we intend to study to what extent urban form affects individual travel behavior and consequently the transport-induced level of CO2 emissions. In line with other studies, urban form is measured through density (Pushkarev and Zupan, 1977; Niemeier and Rutherford, 1994; Schimek, 1996; Kenworthy and Laube, 1999; Schafer and Victor, 2000). Although this does not capture all aspects of spatial organization, residential density can in fact serve as a reliable proxy for urban form when measured at a sufficiently disaggregated level (see for example, Pushkarev and Zupan, 1977; Niemeier and Rutherford, 1994; Schimek, 1996; Giuliano and Narayan, 2003). Higher density generally goes along with more efficient public transport services, better spatial coordination between services, activities and individuals, and better accessibility to infrastructure. Our method of analysis consists of three steps. First, we undertake an econometric analysis of the impact of density on commuting distance using OLS and instrumental variable (IV) approaches. Second, we decompose the specific effects into the contribution of choice of transport mode and travel distance. Third, we assess the impact of urban density on CO2 emissions. The relevance of devoting policy attention to urban density so as to gear processes of urban re-organization towards more efficient energy use (from building and transport) is underlined by an increasing number of ongoing national projects on this issue (in France, The Netherlands, Germany, etc.). In many countries in Europe and in particular in The Netherlands, urban density is strongly determined by national governmental policy, through spatial planning. As an extreme case, consider Almere, one of the youngest cities in The Netherlands, which was built on land gained from the sea about 50 years ago. It was almost entirely planned by the national government so that the location, size, type of residences, and therefore urban density are more the result of planning than of autonomous development.
3. Urban density has been shown to cause negative effects on other types of trips (Greenwald and Boarnet, 2001; Bagley and Mokhtarian, 2002; Khattak and Rodriguez, 2005; Schwanen and Mokhtarian, 2005; Cao et al., 2007). Hence, we might underestimate the effect on all trips by considering only commuting. The latter in fact accounts for a quarter of total individuals travel distance (Hamilton and Roell, 1982; van Dender, 2006).

100 / The Energy Journal Obtained insights from the present empirical analysis can support policies affecting urban form such as spatial planning aimed at reducing energy consumption or CO2 emissions. On the other hand, if the results suggest that in making travel decisions individuals are hardly affected by urban structure, then these policies cannot be expected to play an important role in combating GHG emissions. Note finally that our approach does not address the impact of urban form on social welfare or individual happiness. This is no shortcoming, as we focus on carbon dioxide emissions that create external costs which in turn harm social welfare. Evidently, such external costs are by definition undesirable for a society and must be taken care of regardless of other effects of urban form. The remainder of this paper is structured as follows. Section 2 outlines the theoretical background of the relationship between urban density and travel behavior. Section 3 presents the results of the econometric analysis and performs a decomposition analysis. Section 4 calculates the impact of differences in urban density on CO2 emissions. Section 5 concludes. 2. Urban density and travel behavior The study of the relationship between urban form and transportation demand has a long history, starting with the study of Pushcarev and Zupan (1977), and it covers both aggregate and disaggregate studies of individual travel behavior. In disaggregate studies (e.g., Boarnet and Sarmiento, 1998, Crane and Crepeau, 1998, Giuliano and Narayan, 2003; Handy et al., 2005; Bento et al., 2005) the dependent variable (e.g., commuting) is measured at the individual level, whereas in aggregate studies (e.g., Newman and Kenworthy, 1989a and 1989b; Cervero and Gorham, 1995; Kenworthy and Laube, 1999; Newman and Kenworthy, 1999) the dependent variable is measured at the level of the city or region (e.g., the average commuting distance within a city). This literature covers at least 70 studies, which deliver mixed results (see Ewing and Cervero, 2001). The majority of studies find a statistically significant but usually small effect of density on transport (e.g., Handy, 1996; Levinson and Kumar, 1997; Boarnet and Sarmiento, 1998; Ewing and Cervero, 2001). Other studies find a negative correlation. For example, Newman and Kenworthy (1999) used a sample of 32 cities in four continents and found a negative statistical correlation between residential density and transport demand. Gordon et al. (1989) find that there is a positive relation between metropolitan residential densities and commuting times. The variety of findings may be due to data limitations, the use of different density measures, and differences in statistical methods. For example, some studies do not account for socio-demographic variables, whereas others only include attitudes and preferences of respondents. Few studies distinguish clearly between statistical correlation and causality in terms of the effect of urban form on travel (see Handy et al., 2005).4 We then
4. By causality, we do not mean the behavioral process but the direction of the influence (Singleton and Straits, 1999). Hence we aim to show to what extent an exogenous change in urban form affects travel.

An Empirical Analysis of Urban Form, Transport, and Global Warming / 101 intend to verify that the relationship between the two variables considered is such that individual travel choices are the result of urban form rather than the other way around. It is important to note that urban form only gradually changes over time, whereas simultaneously many other changes occur in the economy at large. As a result, in terms of a particular time order, the effect of urban form on transport is difficult to identify. Hence, we rely on cross-section data, which is common in the literature (see, e.g., Handy and Clifton, 2001; and Bagley and Mokhtarian, 2002).5 In the case of cross-section data, a number of criteria have to be met then, in order to show causality between an independent variable (the cause) and a dependent one (the effect): (1) association: a statistical association between the cause and effect is found; (2) a causal mechanism: the mechanism by which the cause influences the effect is known; and (3) non-spuriousness: no third factor contributes to an accidental relationship between the variables (Singleton and Straits, 1999; Wooldridge, 2002). Correlation is a minimum condition and satisfies only the first criterion. That the second criterion is satisfied is clear from the discussion in Section 1 in which the various routes were outlined through which urban form affects urban transport and associated GHG emissions. Both the first and the second criteria have often been addressed in the literature on urban form, while the third one has received less attention. It is tested for in the current study, so that in effect we address all three criteria and therefore be able to assess whether the impact of urban form on GHG emissions is of a causal nature. The problem associated with the third criterion requires special attention and can be illustrated as follows. Individuals who do not mind or even enjoy traveling driving long distances are more likely to live relatively far from job centers. Therefore, such individuals would sort themselves into residential areas where housing prices are relatively low. This phenomenon is known in the literature as self-selection, and it is influenced by such factors as preferred lifestyle, accessibility to educational and recreational facilities and security concerns. Whether, and the extent to which, this process occurs is an issue that has been long debated in transport economics, urban economics and regional science (Handy et al., 2005). The literature that considers residential self-selection is rather small, and the results are somehow mixed. For example, Boarnet and Sarmiento (1998) and Greenwald and Boarnet (2001) both use an instrumental variable technique to test the impact of the urban structure at the neighborhood scale on travel decisions regarding non-working purposes. But while the former study considers automobile trips and shows that urban structure has little impact, the latter study focuses only on walking as a travel mode and finds that urban structure affects travel choices considerably. Krizek (2003) applies linear regression to a longitudinal design to show that modifications in neighborhood and regional residential structure are responsible for most changes in travel behavior. Yet, the study does not account for individual characteristics, which may lead to biased interpretation of the findings.
5. Changes over time in urban form and travel may also provide evidence of any causal effect, but identifying the effect of urban form on travel may even be more difficult, as urban form and travel patterns change rather slowly over time.

102 / The Energy Journal Finally, Schwanen and Mokhtarian (2005) show that the influence of urban structure is stronger for suburban residents than for urban residents, once self-selection of residence location is controlled for. Not accounting for the effect of this residence location selection mechanism may lead to spurious interpretations of the effect of urban form on individual travel behavior. Self-selection is ultimately concerned with the issue of causality or, in a model-technical sense, endogeneity. To be able to exclude endogeneity, and as a result a spurious relationship between variables related to urban form and travel preferences of individuals, we employ standard instrumental variable (IV) techniques (e.g., Angrist and Evans, 1998; Angrist and Krueger, 2001; Wooldridge, 2002). In this way we can address causality criterion 3 (cf. Frankel and Rose, 2005). Standard regression techniques would in case of endogeneity most likely overestimate the effect of urban form on distance. An instrumental variable approach addresses the causal nature of the relationship between residential density and travel by taking into account that the urban density is self-selected (and therefore potentially endogenous with respect to individuals commuting behavior). In addition, our study assesses the underlying mechanism by which the impact of urban density on commuting distance occurs, by decomposing the effect on travel demand into effects on modal split and travel distance per transport mode. This further supports criterion 2. In effect, we address all three criteria and therefore are able to establish the causal direction of the effect of urban form on GHG emissions. 3. Econometric analysis 3.1 Data and Variables The data used for this study is on individual workers and comes from the Dutch housing survey 1998. It covers housing characteristics, individual job characteristics, and individual commuting behavior. The dataset contains 25,991 valid observations. Urban density is used as the main independent variable of interest. It is measured at the municipality level as the percentage of the population living in an area with a specific local density. For this purpose, the number of addresses per square kilometer (addresses/km2) is categorized into five discrete classes at the municipality level. The data covers 458 municipalities. A municipality in the Netherlands contains on average about 12,000 residences.6 Three dependent variables of interest are use of travel mode, (the logarithm of) commuting distance, and commuting distance by travel mode. As a result, three separate models are estimated, i.e. one for each of these variables.

6. For discussion of urban density in relation to patterns of residential land use and measures of urban density see McDonald (1989), Fujita (1989), Fujita et al. (1999), Schafer and Victor (2000), and Bento et al. (2005).

An Empirical Analysis of Urban Form, Transport, and Global Warming / 103 Car is the most important travel mode from the perspective of CO2 emissions.7 Table 1 shows that the probability of using a car to commute rapidly decreases with an increase of urban density. This suggests a strong impact of urban density on the use of car, but as emphasized above it is an insufficient basis to infer the presence of a causal effect. Table 1 further shows that urban density is negatively related to commuting distance, distance traveled by car, and use of car. Finally, note that given the choice of the car, density is negatively related to distance, but that given the choice of other modes, a reverse relation is found. The average value of the commuting distance in our sample is in line with what is generally found in the literature, which is about 21 kilometers (CBS, 2004; van Ommeren and Dargay, 2006).8 Table 1. Descriptive Data on the Relationship Between Urban Density and Use of Car, Commuting Distance, Commuting Distance by Car and by Other Modes
Urban density Mean Label Very low Low Moderate High Very high Unit Commuting Commuting (addresses/ Use of Commuting distance distance by km2) car distance by car other modes < 500 500 1,000 1,000 1,500 1,500 2,500 > 2,500 0.67 0.62 0.57 0.53 0.44 0.57 24.5 21.8 21.6 20.3 18.3 21.5 32.0 29.8 28.9 27.8 25.2 29.3 9.4 8.9 11.9 12.0 13.1 11.1 5,192 6,159 5,762 5,762 3,116 25,991 Number of observations

Average

3.2 The Estimation Procedure In previous studies, it was assumed that unobserved variables that affect both density and travel are absent. This assumption is unlikely to hold as has been emphasized in the introduction. It is difficult in the empirical analysis to fully control for the variation in the spatial environment (e.g., supply of public transport, congestion, motorway accessibility, etc.), which may cause urban environment to become endogenous. In particular, it is extremely plausible that the individuals' selection of a certain residence location depends on the individuals travel behavior. Consequently, households sort themselves in neighborhoods based on travel preferences that are reflected by the urban form variables (self-selection).
7. The category car includes not only travel by car but also travel by other motorized commuting modes, namely motorbike, scooter and moped. Due to the low proportion of the other motorized subcategories in our sample (less than 3%), it does not make sense to treat these subcategories separately. 8. The share of the population using a car to commute in our sample is in line with other Dutch data, and may thus be considered to be representative for the Netherlands as a whole.

104 / The Energy Journal This would imply that the estimated coefficients on these variables are inconsistent due to their correlation with the error term. In other words, the relationship between the urban form variables and individual travel preferences might be spurious. Hence, in case of endogeneity, criterion (3) is not met and the causal nature of the relationship cannot be assessed. To avoid such endogeneity, we estimate a series of models using an instrumental variable approach (IV) (Boarnet and Sarmiento, 1998; Boarnet and Greenwald, 2000; Angrist a Krueger, 2001; Frankel and Rose, 2005). The accuracy of IV analysis depends on the validity of instruments that are employed. As instruments for density, we employ the following variables: same gender of children; number of children in the household; number of adults in the household; partners type of job; partners sector of employment; partners specialization level of employment. From a theoretical perspective, these variables should not directly influence the commuting behavior of workers (commuting distance, use of the car) but only indirectly via the choice of residential density. The choice of these specific instruments can be motivated as follows: Same gender of children The same gender of children reduces the need for rooms, because bedrooms can be more easily shared among children of the same sex. This in turn allows households to search for a residence in urban centers, where houses are generally smaller and with fewer rooms (Angrist and Evans, 1998). Consequently, household with children of the same gender are more likely to live in higher density neighborhoods. Number of children in the household and number of adults in the household The number of people in a household generally implies more need for space. This translates in such households being more likely to be found in residences in areas with a low price per square meter of housing, i.e. in suburbs. Characteristics of the partners employment It is supposed that the qualification level and sector of employment of the partner do not directly influence individual travel behavior but do affect the geography of labor opportunities of the partner and therefore are likely to influence the choice of residence location. In order to test the strength of the instrumental variables, we regress urban density on these three types of instruments and exogenous regressors. Results are reported in Table A1 in the Appendix. The instruments show a statistically significant effect. This implies that the first condition for validity of the instruments (instrument relevance) is satisfied. Moreover, to check whether our instruments satisfy the second condition for validity (instrument exogeneity), we perform a standard Hausman test for over-identifying restrictions (Wooldridge, 2002, pp 118122). The chi-square is less than X247, ensuring this condition to be satisfied.

An Empirical Analysis of Urban Form, Transport, and Global Warming / 105 3.3 Regression Analysis of Distance Both OLS and IV estimates of the effect on (the logarithm of) commuting distance are presented in Table 2, along with goodness-of-fit and test statistics.9 The estimated parameters are generally close to each other for the two approaches, the main exception being urban density.10 The effect of density on commuting distance is estimated to be 0.063 employing OLS, and 0.139 employing IV. A test based on the Hausman t statistic (Wooldridge, 2002, pp 122124)11 indicates that the difference between the OLS and IV estimates is just not statistically significant at the 5% confidence level but significant at the 10% level, suggesting that the bias in the OLS estimate cannot be ignored. Recall that commuting distance is expressed in logarithmic terms and that the commuting distance is on average 21.6 kilometers. The marginal effects of the explanatory variables on commuting distance can then be easily calculated and are presented in Table 2. These marginal effects are 1.355 and 2.989, respectively. These estimates imply that workers living in the most dense locations tend to travel about 5.3 km and 11.9 km less, respectively, than workers in the least dense location. 3.4 Estimates of Use of Travel Modes Next, we consider the four main categories of commuting travel mode in more detail. The relevant dependent variables here are car, public transport (except train), train, and slow modes (bike and foot).12 Table A3 in the Appendix reports the descriptive statistics of the travel modes in terms of frequency. We perform a multinomial Probit analysis employing a standard and IV approach. Given the small number of alternative categories, this allows for assessing the across effect between alternatives. As a result, we can examine substitution between modes, i.e. modal split (Winston, 1985). Also here we compare standard OLS and IV approaches.

9. We also show results of OLS because the IV approach is conceptually not more appropriate a priori. OLS estimates are usually preferred when they are statistically equal to the IV estimates, because the former have smaller standard errors. If the IV estimates are statistically different from the OLS estimates, the latter must be inconsistent and cannot be used. 10. Note that in the results shown for the IV estimates, urban density is assumed to be a continuous variable. We have also estimated models taking into account the discrete character of all observations of urban density. It appears that the results are very similar (see Table A2 in the Appendix). 11. Rather than comparing the OLS and IV estimates of a particular linear combination of parameters, as the standard Hausman test does, it makes more sense to test for the change in the estimate of the coefficient of interest, which in our case is the parameter of urban density. The Hausman t-statistic reduces then to: (b1,IV b1,OLS )/{[se (b1,IV )]2 [se (b1,OLS )]2 }1/2 , which gives a value of 1.92. 12. The variable public transport combines the use of tram and metro, while train is separately addressed, because tram and metro mainly relate to intra-urban transport, whereas train covers mainly inter-urban transport (longer travel distances). The distinction evidently is relevant for assessing environmental impacts.

106 / The Energy Journal Table 2. Parameter Estimates and Statistical Tests (Dependent Variable: (Logarithm of) Commuting Distance)
Variables Household characteristics Urban density Two earners in household Individual characteristics Man Dutch nationality Age: 2635 years Age: 3645 years Age: over 46 years Secondary education University education Higher education Gross income of the respondent (in log) Number of weekly hours worked: 1323 Number of weekly hours worked: 2432 Number of weekly hours worked: 3340 Number of weekly hours worked: > 40 Production-based sector of employment Profession-based sector of employment Qualification level of employment Number of observations Adjusted R
2

OLS estimation t-value marg

IV estimation t-value marg

coefficient

coefficient

0.063 0.034 9.718 1.854 1.355 0.731 0.139 0.022 3.396 2.989 1.169 0.473

0.135 0.115 0.172 0.095 0.045 0.156 0.164 0.236 0.058 0.439 0.662 0.773 0.464 6.459 2.181 6.381 3.497 1.693 0.156 3.795 6.057 6.935 11.828 12.370 17.975 23.226 11.702 2.902 2.473 3.698 1.995 0.967 0.107 3.354 3.526 5.074 1.247 9.439 14.233 16.619 9.975 0.130 0.161 0.172 0.089 0.052 0.011 0.154 0.158 0.246 0.058 0.443 0.674 0.781 0.461 6.193 2.765 6.356 3.286 1.934 0.361 3.745 5.794 7.159 11.922 12.427 2.795 3.461 3.698 1.913 1.118 0.237 3.311 3.397 5.289 1.247 9.524

Lower secondary education 0.005

17.978 14.491 23.224 11.586 included included included 25,991 0.076 16.792 9.911

included included included

25,991 0.079

Notes: (1) Reference categories for model estimates: age: 1825 years; primary education; number of weekly hours worked : <13. (2) marg = marginal effect.

Table 3. Multinomial Probit Model Estimates of Travel Modes: Standard Estimation

Use of car Use of public transport t-value 17.60 7.16 6.05 0.020 0.110 2.80 0.013 0.068 3.96 0.021 0.029 7.09 0.023 0.187 6.03 0.018 0.318 3.50 0.009 0.084 0.01 0.001 0.262 3.46 0.010 0.261 2.55 0.009 0.633 3.64 0.000 0.078 3.20 0.007 0.333 5.48 0.004 0.561 5.64 0.002 0.554 0.82 0.020 0.371 included included included 25,991 22,002.077 0.057 0.053 0.246 0.129 14.98 2.84 0.016 0.012 0.144 0.159 8.43 3.43 2.17 0.47 0.44 2.79 4.67 0.98 2.65 4.23 10.32 4.61 3.37 5.73 6.06 3.25 included included included marg coefficient t-value marg coefficient t-value Use of train marg 0.011 0.003 0.002 0.008 0.013 0.013 0.016 0.002 0.019 0.014 0.043 0.002 0.003 0.010 0.004 0.012 coefficient 0.160 0.184

Variables

Household Characteristics Urban density Two earners in household

Individual Characteristics Man 0.167 5.65 0.054 0.291 Dutch nationality 0.173 2.28 0.042 0.311 Age: 2635 years 0.484 12.54 0.144 0.242 Age: 3645 years 0.244 6.36 0.085 0.443 Age: over 46 years 0.147 3.90 0.058 0.358 Lower secondary education 0.047 1.22 0.006 0.248 Secondary education 0.082 1.51 0.033 0.001 University education 0.030 0.97 0.006 0.186 Higher education 0.021 0.63 0.025 0.147 Gross income of the respondent (in log) 0.068 10.12 0.017 0.050 Number of weekly hours worked: 1323 0.451 9.09 0.112 0.261 Number of weekly hours worked: 2432 0.597 11.63 0.138 0.451 Number of weekly hours worked: 3340 0.828 17.82 0.214 0.425 Number of weekly hours worked: > 40 1.070 19.22 0.266 0.091 Production-based sector of employment included Profession-based sector of employment included Qualification level of employment included

Number of observations

An Empirical Analysis of Urban Form, Transport, and Global Warming / 107

Loglikelihood

Notes: (1) Reference categories for Multinomial Probit estimates: age: 1825 years; primary education; number of weekly hours worked: <13. (2) marg = marginal effect.

Table 4. Multinomial Probit Model Estimates of Travel Modes: IV Estimation

Use of car Use of public transport t-value 0.594 0.244 8.68 5.10 0.034 0.008 0.526 0.073 8.47 1.59 3.16 1.04 0.30 2.23 3.91 0.42 2.31 4.45 7.96 4.26 2.84 4.60 5.29 3.35 included included Included 25,991 22,351.767 marg coefficient t-value marg coefficient t-value marg 0.034 0.010 0.005 0.001 0.012 0.012 0.014 0.004 0.017 0.015 0.033 0.001 0.001 0.004 0.001 0.011 Use of train coefficient 0.208 5.63 0.093 0.173 6.60 0.047

Variables

108 / The Energy Journal

Household Characteristics Urban density Two earners in household

Individual Characteristics Man 0.158 5.36 0.050 0.257 5.42 0.021 0.161 Dutch nationality 0.208 2.63 0.063 0.372 2.48 0.018 0.119 Age: 2635 years 0.474 12.35 0.142 0.211 3.53 0.023 0.020 Age: 3645 years 0.234 6.12 0.082 0.389 6.40 0.025 0.149 Age: over 46 years 0.136 3.61 0.054 0.310 5.33 0.019 0.266 Lower secondary education 0.063 1.57 0.014 0.179 2.52 0.008 0.037 Secondary education 0.082 1.51 0.030 0.047 0.55 0.001 0.228 University education 0.025 0.79 0.004 0.184 3.49 0.012 0.273 Higher education 0.006 0.18 0.012 0.247 4.22 0.014 0.502 Gross income of the respondent (in log) 0.069 10.18 0.017 0.044 3.35 0.000 0.070 Number of weekly hours worked: 1323 0.450 9.09 0.114 0.227 2.84 0.002 0.280 Number of weekly hours worked: 2432 0.602 11.64 0.144 0.378 4.63 0.000 0.453 Number of weekly hours worked: 3340 0.827 17.77 0.215 0.398 5.36 0.004 0.483 Number of weekly hours worked: > 40 1.059 19.09 0.264 0.113 1.04 0.02 0.378 Production-based sector of employment included included Profession-based sector of employment included included Qualification level of employment included included

Number of observations

Loglikelihood

Notes: (1) Reference categories for Multinomial Probit estimates: age: 1825 years; primary education; number of weekly hours worked: <13. (2) marg = marginal effect.

An Empirical Analysis of Urban Form, Transport, and Global Warming / 109 Tables 3 and 4 show the results. The reference category is slow mode. Marginal effects are also presented. The standard estimates suggest that a marginal increase (i.e. changing one class up out of five) in density corresponds with a decrease of 5.7 % in the use of car, whereas for public transport and train a response of 1.6 % and 1.1 % are found, respectively. The IV estimates have the same sign but a different magnitude: the marginal effect on car use is 9.3%, whereas the marginal effect on the use of public transport and train is equal to 3.4%. Hausman tests indicate that the standard estimate of car is not biased, but the marginal effects of density on train and other public transport modes are substantially higher. These results make sense and are in line with a utility maximizing framework where commuters choose between different modes (e.g., Jara-Diaz, 2007). In high-density areas, the utility of driving versus cycling or the use of public transport is substantially reduced because of higher congestion costs and costs related to parking. This is likely reinforced because with higher urban density, the time-cost of public transport is substantially reduced due to a more detailed public transport network and more frequent services. The predicted modal choice outcomes are shown in Figure 1. As a result, in more densely populated areas workers tend to shift from car to other travel modes, notably public transport (metro and tram). Figure 1. The Impact of Urban Density on Modal Shift

Note: Class 1 denotes lowest degree of density, 5 denotes highest.

110 / The Energy Journal 3.5 Estimates of Commuting Distance by Travel Mode Previous subsections have analyzed the impact of urban density on commuting distance and use of travel mode. In order to complete the picture, we analyze the effect of density on a third variable, namely commuting distance by specific mode (car, public transport, train, and slow mode). For example, the commuting distance by car is zero when another mode is used. The same holds for commuting distance by other modes. We therefore use a range of Tobit models. The full set of results for the Tobit analysis of commuting distance by car is reported in Table 5. The marginal effect on commuting distance by car is equal to 3.374 using the IV approach, and 2.706 ignoring endogeneity.13 Therefore, by ignoring endogeneity the effect is slightly underestimated. When comparing the most dense location to the least dense one, OLS and IV estimates indicate that workers tend to travel about 10.8 km and 13.5 km more by car, respectively, which is substantial. Results for the impact of urban density on traveled distance by the other modes are derived similarly and presented in Table 6. To avoid repetition, we only report findings for the main independent variable urban density on (a) commuting distance by public transport, (b) commuting distance by train, and (c) commuting distance by slow mode. Hausman tests indicate that the difference between the standard and IV estimates is not statistically significant, except for slow mode, where according to IV estimates changes in density have hardly an impact on commuting distance by bike or foot. The marginal effects of urban density on the three dependent variables for standard Tobit are 0.28 km, 0.34 km and 0.23 km, respectively. This implies increases in commuting distance by public transport means, commuting distance by train and commuting distance by slow mode of 1.1 km, 1.4 km, and 0.9 km, respectively, when passing from least dense to most dense locations. According to the IV approach, the increased commuting distance associated with one unit increase in urban density (out of five classes) are 0.37 km, 0.9 km, and 0.06 km for public transport, train, and slow mode, respectively. The overall marginal effects increase up to 1.6 km, 3.6 km, and 0.24 km, respectively, for a change in density from the lowest to the highest density level. 3.6 Decomposition of Travel Demand In order to investigate some of the underlying mechanisms by which the impact of urban density on average commuting distance occurs, we provide a decomposition of the effect of density on travel demand. In doing so, we capture the effect of urban density on commuting distance via mode choice. We write the expected distance as the sum of the conditional expected distances, where the condition is the use of mode to commute.
13. The marginal effect on commuting distance by car given the use of car is equal to the coefficient reported.

An Empirical Analysis of Urban Form, Transport, and Global Warming / 111 Table 5. Tobit Model Estimates (Dependent Variable : Commuting Distance by Car)
Variables Standard estimation Household Characteristics Urban density Two earners in household Individual Characteristics Man Dutch nationality Age: 2635 years Age: 3645 years Age: over 46 years Secondary education University education Higher education Gross income of the respondent (in log) Number of weekly hours worked: 1323 Number of weekly hours worked: 2432 Number of weekly hours worked: 3340 Number of weekly hours worked: > 40 Production-based sector of employment Profession-based sector of employment Qualification level of employment Number of observations Log likelihood coefficient t-value marg IV estimation t-value marg coefficient

5.132 2.540 25.62 4.54 2.706 1.338 6.400 2.295 4.97 3.85 3.374 1.208

3.860 4.590 12.304 8.353 5.979 1.952 3.455 4.783 1.427 11.913 17.080 22.572 23.149 5.93 2.84 14.48 9.75 7.01 0.18 1.51 4.05 4.49 9.18 10.04 13.99 20.18 17.98 included 2.026 2.529 6.837 4.569 3.240 0.092 1.048 1.846 2.591 0.752 6.824 10.152 11.849 14.301 3.727 5.452 12.199 8.160 5.659 0.494 1.954 3.403 5.157 1.449 12.014 17.378 22.728 23.060 5.63 3.02 14.20 9.37 6.50 0.49 1.50 3.90 4.70 9.17 10.00 13.90 19.99 17.69 included included included 25,991 78,593.409 1.956 3.034 6.768 4.556 3.059 0.259 1.048 1.817 2.797 0.764 6.876 10.329 11.928 14.206

Lower secondary education 0.175

included included

25,991 78,276.23

Notes: (1) Reference categories for Tobit model estimates: number of adults in the household = 1; no child in the household; age: 1825 years; primary education; number of weekly hours worked: <13. (2) marg = marginal effect.

112 / The Energy Journal E ( y) = E ( y|m = 1) p (m = 1) + E ( y|m = 0) p (m = 0). (1)

Here E(y) is the expected value of commuting distance and p indicates the probability of using a certain mode (m) to commute. The marginal effect of density on commuting distance can then be decomposed as follows: E ( y) E ( y|m = 1) E ( y|m = 0) = p (m = 1) + p (m = 0) + x x x p (m = 1) p (m = 0) + E ( y|m = 1) + E ( y|m = 0) . x x

(2)

The term on the left-hand side of equation (2) is the marginal effect of density on distance. The first two components on the right-hand side indicate the marginal effect on distance per mode keeping mode choice constant, whereas the last two terms capture the marginal effect on modal choice while keeping the distance per mode constant. Equation (2) can be easily generalized to more than two modes. We use the four modes described above. The decomposition reflects some of the complex mechanisms by which urban density acts on commuting distance. Using the estimates presented above, the decomposition indicates that the effect of urban density on commuting distance occurs by approximately equal magnitude through a change in commuting distance given the mode and a change in the probability of using a certain mode.14 4.  Calculating CO2 emissions for different urban forms To analyze the potential effect of spatial organization on carbon dioxide emission abatement we employ conversion factors. We adopt these from TREMOVE baseline version 2.4, a policy evaluation model of the effects of spatial, transport and environmental measures on the emissions of the transport sector (EC, 2006).15 We use data on CO2 emissions by transport activity for the Netherlands, for the year 2005, which includes both exhaust (or tailpipe) and lifecycle emissions. The latter represents emissions that occur during the production of fuels and electricity. Since the operational emissions tend to decrease in the future, the relative share of lifecycle emissions will increase and may become substantial.
14. The values for the predicted scenario of modal shift are calculated by multiplying the mean of all the variables in the sample (except urban density) by the coefficients from the Probit model and adding the results. For urban density, which is the scenario variable, we calculate the product of the marginal effect from Probit and each class of density. 15. This model includes both passenger and freight transport in the EU 15, covering a period of 35 years, from 1995 (historical data) to 2030 (forecasted data).

Table 6. Tobit Model Estimates of Urban Density Parameter for Different Travel Modes
Distance by train (km) marg 0.340 0.917 0.716 0.184 coefficient t-value 10.86 0.45 Distance by slow mode (km) marg 0.226 0.058

Distance by public transport (km) coefficient 11.652 13.458 25,991 14.76 3.27 0.273 0.367 18.568 46.699 10.04 4.41 t-value marg coefficient t-value

Urban density

Standard estimation IV estimation

Number of observations

Note: marg = marginal effect.

Table 7. CO2 Emissions by Transport Modes

Exhausted emissions (t CO2) 23,119,855.4 22,586,389.9 22,281.6 49,023.8 462,160.1 0 82,491.9 0 Lifecycle emissions (t CO2) CO2 emission factor (g CO2 /pkm) 4,471,841.7 4,388,527.5 4,687.7 10,313.9 68,312.6 12,061.4 824,733.8 0 201.5 178.6 69.6 113.6 73.8 9.1 65.6 0

Travel mode

Transport activity (million pkm)

Car private automobile moped motorcycle bus/coach

159,129.9 151,026.6 393.1 526.8 7,183.3

Public transport (metro/tram)

1,326.6

Passenger train

13,827.5

Slow mode (bike and foot)

16,788.6

An Empirical Analysis of Urban Form, Transport, and Global Warming / 113

Note: Data from TREMOVE 2.4 (EC, 2006)

114 / The Energy Journal Table 7 reports indicative values for total (exhaust and lifecycle) CO2 emissions (in tons of CO2 [tCO2]) and transport activity (in passenger kilometers [pkm]), for all four categories of travel modes under consideration. Transport volume is measured in passenger kilometers instead of vehicle kilometers to incorporate the contribution of occupancy rate. To show the magnitude of the impact of a specific travel mode on CO2 emissions, all the variables related to CO2 emissions, namely transport activity, (exhaust and lifecycle) emissions and CO2 emission factors, are reported at the subcategory level (see Table A3 in the Appendix). Exhaust CO2 emissions of slow transport and of electric public transport means (metro and tram) are zero, whereas train has positive values for exhaust CO2 emissions. The last column of the table shows the CO2 emission conversion factors. Their values are derived for each transport mode by dividing total CO2 emissions by the total travel distance by that mode (in pkm). The factors are expressed in grams of CO2 per traveled km (gCO2/pkm). The subdivision of the travel demand by travel modes is derived from the descriptive statistics in Table A3 in the Appendix.16 Table 8 presents predicted CO2 emissions for each travel mode under different classes of urban density based on the IV estimates as reported in Tables 5 and 6. Figure 2 shows the trend graphically. A difference in urban density between its lowest and highest level is associated with a considerable reduction in CO2 emissions by motorized (car) mode. Electric public transport and train show relatively small increases in CO2 emissions due to the increase of both use and Table 8. CO2 Emissions (in gCO2 /pkm) by Urban Density
Travel mode Car Public transport Passenger train Slow mode Total effect Very low 4217.38 0.00 188.56 0 4405.94 Low 3672.10 0.00 210.87 0 3882.97 Urban density Moderate 3126.81 0.00 233.18 0 3359.99 High 2581.53 2.41 255.49 0 2839.42 Very high 2036.24 4.89 277.80 0 2318.93

distance by those modes. Note that the values for public transport are very low. For this reason, the corresponding line in Figure 2 cannot be well observed. The net difference in total CO2 emissions from lowest to highest urban density level is equal to 2087 gCO2/pkm (see bottom row of Table 8), which amounts to 47% reduction. The CO2 reduction associated to a difference from the current situation (i.e. the sample) with urban density = 3 and urban density = 5 is equal to 1041.6 gCO2/pkm (31% reduction).
16. Note that the TREMOVE model computes emissions as dependent on travel speed, and thus accounts for congestion effect in denser areas.

An Empirical Analysis of Urban Form, Transport, and Global Warming / 115 Figure 2. CO2 Emissions as a Function of Urban Density

Note: Class 1 denotes lowest degree of density, 5 denotes highest.

5. Conclusions In order to answer the question whether urban form affects travel behavior by individuals and consequently environmental quality, this paper has performed an analysis of the influence of urban density on GHGs emissions through commuting behavior. From a policy standpoint, our study has indicated the potential contribution of policies that affect urban form (notably spatial planning) to reduce CO2 emissions by the transport sector. Such policies may complement more direct regulation of CO2 emissions, such as through incentives for energy efficiency. To address this issue, we examined the impact of urban density on commuting behavior, and subsequently determined the consequences for CO2 emissions. The econometric approach adopted here has involved a range of techniques, including OLS, Probit and Tobit. An instrumental variable (IV) technique served the purpose of addressing the endogeneity of urban density, which potentially invalidates previous estimates. Our findings indicate that with respect to commuting distance, endogeneity problems with urban density are small. Based on our results for distance, we therefore believe that the OLS estimates are quite accurate. In the case of

116 / The Energy Journal the choice of travel mode, however, OLS generates inconsistent results and IV estimation is preferred. In particular, the effect on the use of public transport and the train are far more pronounced. In addition, a decomposition of travel demand into modal split and commuting distance per transport mode has been performed so as to shed light on the underlying mechanisms by which the impact of urban density on commuting distance occurs. It has been shown that the effect of urban density on commuting distance arises by approximately equal magnitude through the effect on commuting distance, given the transport mode and the effect on the probability of using a mode. Combining the IV results of the travel demand analysis with mode specific emissions factors, the impact of a difference in urban density on CO2 emissions was calculated for different values of urban density. Our results show that in the densest urban locations CO2 emissions by motorized (car) mode are considerably reduced. Electric public transport and train, as opposed, show increases in CO2 emissions due to an increase in both use and distances by these modes. In terms of policy implications, the lesson emerging from this study is that a higher urban density is likely to lead to a change in travel behavior. The magnitude and direction of this change are observed by modal shifts in individual travel choices, from motorized vehicle use to other transport modes, notably public transport, train and slow modes (bike and foot). Our estimates imply that in locations where density is 500 addresses per square kilometer higher, CO2 emissions from transport are on average 15% lower. These results also imply that the commuting-related CO2 emissions associated with a high-density city such as Amsterdam (more than 2500 addresses per square kilometer) are about half the emissions from low-density villages (less than 500 addresses per square kilometer). The CO2 reduction associated with a difference between the current average urban density situation (about 1250 addresses per square kilometer) and an extremely high density situation (more than 2500 addresses per square kilometer) is 1041.6 grams of CO2 per passenger kilometer, which means a 31% reduction. The main implication of this finding is that policies that try to enforce or stimulate a higher density of activities may have a rather small but favorable effect in terms of reduction of CO2 emissions. To give an example, if due to policy 10% of the workforce would live in high-density instead of low-density areas, the reduction in CO2 would be about 5%. Note that if one is interested in the effect of more substantial changes in density due to purposeful policies, indirect or general equilibrium type of effects may have to be taken into consideration. All taken together, urban form, and therefore policies that affect urban form, such as spatial and transport planning, deserve more attention in climate policy debates, as they can contribute to a reduction in greenhouse gases. For example, transport planning may try to stimulate modal shift by increasing density through the development of new public transport, such as the planned additional subway line in the centre of Amsterdam, and thus allow the design of a more effective transport infrastructure network as well as the creation of fast lanes for buses and separate lanes for bicyclists. Therefore, a key challenge is to find

An Empirical Analysis of Urban Form, Transport, and Global Warming / 117 ways to increase density in existing urban areas in the face of financial, political, historical, social and environmental constraints. Density-increasing policies are feasible in countries like The Netherlands, as spatial planning here traditionally is very effective in influencing urban density. The results presented in this study may be of particular interest when considering urban and transport planning in the context of new, emerging urban areas and rapidly growing cities, notably in developing countries. For instance, combating CO2 emissions through spatial and transport planning may be a wise second-best strategy as other types of climate policies, such as externality taxes on fuels, are perceived as politically unacceptable and are severely hampered by vested interests and public good features of global warming. Nevertheless, the two types of policy are generally complementary and in the long run may need to be implemented simultaneously. ACKNOWLEDGEMENTS The authors are grateful to Malcolm O. Asadoorian, Ada Ferrer-i-Carbonell, Thomas de Graaff, Eva Gutirrez Puigarnau, Piet Rietveld, three anonymous referees, and the Editor of the Journal for helpful comments and suggestions. REFERENCES
Anas, A., R. Arnott and K.A. Small (1989). Urban Spatial Structure. Journal of Economic Literature 36(3): 14261464. Angrist, J. D. and W.N. Evans (1998). Children and their Parents Labor Supply: Evidence from Exogenous Variation in Family Size. American Economic Review 88(3): 450477. Angrist, J. D. and A.B. Krueger (2001). Instrumental Variables and the Search for Identification: From Supply and Demand to Natural Experiments. The Journal of Economic Perspectives 15(4): 6985. Bagley, M. N. and P.L.Mokhtarian (2002). The Impact of Residential Neighborhood Type on Travel Behavior: A Structural Equations Modeling Approach. The Annals of Regional Science 36: 279297. Bento, A.M., M.L. Cropper, A.M. Mobarak, and K. Vinha (2005). The Effects of Urban Spatial Structure on Travel Demand in the United States. The Review of Economics and Statistics 87(3): 466478. Bertaud, A. (2003). Clearing the Air in Atlanta: Transit and Smart Growth or Conventional Economics? Journal of Urban Economics 54(3): 379400. Boarnet, M.G. and M.J. Greenwald (2000). Land Use, Urban Design, and Non-Work Travel: Reproducing for Portland, Oregon Empirical Tests from Other Urban Areas. Transportation Research Record 1722: 2737. Boarnet, M. and S. Sarmiento (1998). Can Land Use Policy Really Affect Travel Behavior? A Study of the Link between Non-Work Travel and Land Use Characteristics. Urban Studies 35(7): 11551169. Burchfield, M., H.G. Overman, D. Puga, and M.A. Turner (2006). Causes of Sprawl: A Portrait from Space. The Quarterly Journal of Economics 121(2): 587633. Cao, X., P. L. Mokhtarian, and S.L. Handy (2007). Do Changes in Neighborhood Characteristics Lead to Changes in Travel Behavior? A Structural Equations Modeling Approach. Transportation 34(5): 535556. CBS (2004). Statistics Netherlands Web Magazine. Central Bureau of Statistics, CBS. The Netherlands. http://www.cbs.nl/

118 / The Energy Journal

Crane, R. and R. Crepeau (1998). Does Neighborhood Design Influence Travel? A Behavioral Analysis of Travel Diary and GIS data. Transportation Research D 3(4): 225238. Dender, K. van (2003). Transport Taxes with Multiple Trip Purposes. Scandinavian Journal of Economics 105(2): 295310. EC (2006). TREMOVE Transport Model. European Commission, EC-DG Environment, Brussels. Elkins, P. (1996). How Large a Carbon Tax is Justified by Secondary Benefits of the CO2 Abatement? Resource and Energy Economics 18(2): 161187. EPA (2001). EPA Guidance: Improving Air Quality through Land Use Activities. Environmental Protection Agency, EPA, Washington, DC. Ewing, R. and R. Cervero (2001). Travel and the Built Environment: A Synthesis Discussion. Transportation Research Record 1780: 87114. Frankel, J.A. and A.K: Rose (2005). Is Trade Good or Bad for the Environment? Sorting Out the Causality. Review of Economics and Statistics 87(1): 8591. Fujita, M. (1989). Urban Economic Theory. Cambridge: Cambridge University Press. Fujita, M., P.R. Krugman, and A.J. Venables (1999). The Spatial Economy: Cities, Regions, and International Trade. MIT Press, Cambridge MA, US. Giuliano, G. and D. Narayan (2003). Another Look at Travel Patterns and Urban Form: The US and Great Britain. Urban Studies 40(11): 22952312. Gordon, P., A. Kumar, and H.W. Richardson (1989). Congestion, Changing Metropolitan Structure, and City Size in The United States. International Regional Science Review 12(1). 4556. Grazi, F. and J.C.J.M. van den Bergh (2008). Spatial Organization, Transport, and Climate Change: Comparing Instruments of Spatial Planning and Policy. Ecological Economics, doi:10.1016/j.ecolecon.2008.01.014 Greene, D.L. and A. Schafer (2003). Reducing greenhouse gas emissions from U.S. transportation. Report, Pew Center on Global Climate Change, Arlington, VA. Greenwald, M.J. and M.G.Boarnet (2001). Built Environment as Determinant of Walking Behavior: Analyzing Nonwork Pedestrian Travel in Portland, Oregon. Transportation Research Record 1780: 3342. Hamilton, B. and A. Roell (1982). Wasteful Commuting. The Journal of Political Economy 90(5): 10351053 Handy, S., X. Cao, and P. Mokhtarian (2005). Correlation or Causality between the Built Environment and Travel Behavior? Evidence from Northern California. Transportation Research D 10(6): 427444. Handy, S.L. and K.J.Clifton (2001). Local Shopping as a Strategy for Reducing Automobile Travel. Transportation 28: 317346. Handy, S. (1996). Methodologies for Exploring the Link between Urban Form and Travel Behavior. Transportation Research D 1(2): 151165. Harvey, R.O. and W.A.V. Clark (1965). The Nature and Economics of Urban Sprawl. Land Economics 41(1): 19. IEA (2006). World Energy Outlook. International Energy Agency, IEA Paris. IEA (2004). Energy Technologies for a Sustainable Future: Transport. IEA Technology Briefs. Report, International Energy Agency, IEA, Paris. Jara-Diaz, S. (2007). Transport Economic Theory. Elsevier. Kenworthy, J. and F. Laube (1999). A Global Review of Energy Use in Urban Transport Systems and its Implications for Urban Transport and Land-Use Policy. Transportation Quarterly 53(4): 2348. Krizek, K. (2003). Residential Relocation and Changes in Urban Travel: Does Neighborhood-Scale Urban Form Matter? Journal of the American Planning Association 69:, 265281. Levinson, D.M. and A. Kumar (1997). Density and the Journey to Work. Growth and Change 28(2): 147172. McDonald, J. (1989). Econometric Studies of Urban Population Density: A Survey. Journal of Urban Economics 26(3): 361385.

An Empirical Analysis of Urban Form, Transport, and Global Warming / 119

Newman, P. and J. Kenworthy (1999). Sustainability and Cities: Overcoming Automobile Dependence. Washington, DC: Island Press. Newman, P. and J. Kenworthy (1989a) Gasoline Consumption and Cities. Journal of the American Planning Association Winter: 2437. Newman, P. and J. Kenworthy (1989b). Cities and Automobile Dependence: An International Sourcebook. Brooksfield VT: Gower Technical. Niemeier, D. and G. Rutherford (1994). Nonmotorized Transportation, NPTS: Travel Mode Special Reports. Report, US Department of Transportation, Federal Highway Administration, Office of Highway Information Management, Washington, DC. OECD (2007). 20072008 Programme of Work of the OECD Territorial Development Policy Committee. Report, Organization for Economic Co-operation and Development, OECD, Paris. Ommeren, J. van, and J. Dargay (2006). The Optimal Choice of Commuting Speed: Consequences for Commuting Time, Distance and Costs. Journal of Transport Economics and Policy 40(2): 279296. Pushkarev, B. and J.M. Zupan (1977). Public Transportation and Land Use Policy. Bloomington, IN: Bloomington University Press. Schafer, A. and D.G. Victor (2000). The Future Mobility of the World Population. Transportation Research A 34(3): 171205. Schimek, P. (1996) Household Motor Vehicle Ownership and Use: How Much Does Residential Density Matter? Transportation Research Record 1552: 120125. Schwanen, T. and P.L. Mokhtarian (2005). What if You Live in the Wrong Neighborhood? The Impact of Residential Neighborhood Type Dissonance on Distance Travelled. Transportation Research D 10 (2):127151. Singleton, R.A. and B.C. Straits (1999). Approach to Social Research. New York and Oxford: Oxford University Press. Winston, C. (1985). Conceptual Developments in the Economics of Transportation: An Interpretive Survey. Journal of Economic Literature 23(1): 5794. Wooldridge, J.M. (2002). Econometrics Analysis of Cross Section and Panel Data. Cambridge, MA: MIT press.

120 / The Energy Journal APPENDIX Table A1. Parameter Estimates And Statistical Tests Of Instruments (Dependent Variable: Urban Density)
Variables coefficient OLS estimation t-value std. error

Instrumental Variables Same gender of children 0.0522 2.01 0.026 Number of adults in the household = 2 0.442 14.85 0.029 Number of adults in the household 3 0.468 12.73 0.037 Number of children in the household = 1 0.078 3.07 0.025 Number of children in the household = 2 0.185 6.88 0.027 Number of children in the household = 3 0.293 8.91 0.033 Number of children in the household 4 0.359 6.80 0.053 Production-based sector of employment of partner included Profession-based sector of employment of partner included Qualification level of employment of partner included Exogenous variables Dutch nationality 0.616 Two earners in household 0.019 Man 0.009 Dutch nationality 0.078 Age: 2635 years 0.091 Age: 3645 years 0.079 Age: over 46 years 0.149 Lower secondary education 0.215 Secondary education 0.033 University education 0.115 Higher education 0.114 Gross income of the respondent (in log) 0.008 Number of weekly hours worked: 1323 0.046 Number of weekly hours worked: 2432 0.102 Number of weekly hours worked: 3340 0.0245 Number of weekly hours worked: > 40 0.095 Production-based sector of employment Profession-based sector of employment Qualification level of employment Number of observations 12.29 0.57 0.43 3.07 3.27 2.81 5.78 7.33 0.84 4.34 3.43 1.72 1.35 2.88 0.76 2.48 included included included 25,991 0.050 0.035 0.022 0.025 0.028 0.027 0.026 0.029 0.039 0.026 0.033 0.005 0.034 0.035 0.032 0.038

Notes: (1) Reference categories for estimates: number of adults in the household = 1; no child in the household; age: 1825 years ; primary education ; number of weekly hours worked : <13. (2) marg = marginal effect.

An Empirical Analysis of Urban Form, Transport, and Global Warming / 121 Table A2. P  arameter Estimates (Dependent Variable: (Logarithm of) Commuting Distance)
Variables Household characteristics Urban density: low Urban density: moderate Urban density: high Urban density: very high Two earners in household Individual characteristics Man Dutch nationality Age: 2635 years Age: 3645 years Age: over 46 years Lower secondary education Secondary education University education Higher education Gross income of the respondent (in log) Number of weekly hours worked: 1323 Number of weekly hours worked: 2432 Number of weekly hours worked: 3340 Number of weekly hours worked: > 40 OLS estimation t-value coefficient 0.149 0.189 0.217 0.276 0.034 0.135 0.113 0.171 0.093 0.045 0.007 0.158 0.166 0.237 0.058 0.439 0.661 0.771 0.464 Included Included Included 25,991 0.079 6.491 2.137 6.352 3.473 1.697 0.219 3.846 6.129 6.961 11.922 12.385 17.933 0.143 11.688 6.029 7.448 8.499 9.041 1.885

Production-based sector of employment Profession-based sector of employment Qualification level of employment Number of observations Adjusted R 2

Note: Reference categories for model estimates: urban density: very low; age: 1825 years ; primary education ; number of weekly hours worked : <13.

122 / The Energy Journal Table A3. Composition of the Sample in Terms of Travel Modes
(Commuting) travel mode Car Moped Motorbike, scooter Personal automobile Coach, bus Public transport (tram, metro) Passenger train Slow mode Bike Foot Total Number of observations Mean Mean 1.7 1.1 52.5 2.1 14,931 57.4 451 293 13,648 539 1056 830 4.1 3.2 30.5 4.9

9,174 35.3 7,894 1,280 25,991 100.0

Economies of Scale and Scope in Multi-Utilities

Mehdi Farsi, Aurelio Fetz and Massimo Filippini* This paper explores the economies of scale and scope in the electricity, gas and water utilities. These issues have a crucial importance in the actual policy debates about unbundling the integrated utilities into separate entities, a policy which has often been supported by the ongoing reforms in the deregulation of network industries. This paper argues that the potential improvements in efficiency through unbundling should be assessed against the loss of scope economies. Several econometric specifications including a random-coefficient model are used to estimate a cost function for a sample of utilities distributing electricity, gas and/or water to the Swiss population. The estimates of scale and scope economies are compared across different models and the effect of heterogeneity among companies are explored. While indicating considerable scope and scale economies overall, the results suggest a significant variation in scope economies across companies due to unobserved heterogeneity. 1. Introduction In Switzerlands energy sector, there is a certain tendency that local utility companies operate in both electricity and gas distribution as well as in the provision of water. Generally, this horizontal integration strategy allows the local multi-utility companies to save on costs by exploiting the economies of scope and to provide customers with an integrated set of services. As pointed out by Baumol, Panzar et al. (1982), economies of scope can result from sharing or joint utilization of inputs such as labor and capital. The distribution companies use similar equipment such as wires, overhead line and similar skills such as those required
The Energy Journal, Vol. 29, No. 4. Copyright 2008 by the IAEE. All rights reserved. * Department of Management, Technology and Economics, ETH Zurich, Zurichbergstr. 18, CH-8032 Zurich, Switzerland. Tel. +41-44-632 06 50, Fax. +41-44-632 10 50 and Department of Economics, University of Lugano. Email: mfarsi@ethz.ch, afetz@ethz.ch, mfilippini@ethz.ch.

This study has benefited from the financial support of the Swiss National Science Foundation through research grant 100012-108288 and also that of the State Secretariat for Economic Affairs (SECO), which is gratefully acknowledged. The authors also wish to thank Adonis Yatchew and two anonymous reviewers for their very helpful suggestions.

123

124 / The Energy Journal for network operation and maintenance. Synergies also exist in advertising and billing activities. Another source of cost savings is due to economies of massed reserves (Waldman and Jensen (2001)). Multi-utility companies can make use of the same reserve capacity for maintenance and emergency repair activities. During the last two decades the introduction of high levels of competition in the electricity and gas sectors of several EU-member countries has raised the general question of the necessity of unbundling services of utility companies. The regulatory reforms have been toward a separation of activities in the form of functional, legal or ownership unbundling, which are often believed to lower entry barriers and boost competition. However, the importance of the potential synergies through horizontal integration has been recognized in the recent European regulatory recommendations (cf. DG Energy & Transport (2004)). An effective policy for unbundling multi-utilities, requires a reliable assessment of the scope economies and their variation with the companys size and other characteristics. Despite its policy importance, there are only a few studies that have studied the issue of scope economies in multi-utilities. In general, these studies suggest that the scope economies are considerable at least for relatively small companies. However, the evidence as to the extent and statistical significance of the scope economies is rather mixed. A major difficulty in estimating scope and scale economies is the fact that utilities operate with different networks with various environmental and technical characteristics, which might induce various levels of synergies across different services. Many of these characteristics are not observed or difficult to measure. Such omitted variables could bias the estimation results. Moreover, the differences among companies could be beyond their variation in output and size. In fact, the strong heterogeneity among utilities operating in such different environments, suggests that a cost function with constant coefficients might be inadequate for a reliable analysis of scope economies. Given that such network characteristics can be considered more or less constant over time, panel data can be used to account at least partially, for such heterogeneity and perhaps assess the potential biases. However, to our knowledge none of the previous studies in this field has used the advantages of panel data models to account for heterogeneity among companies. Benefiting from a data set from 87 companies over a nine-year period, this paper applies two panel data models, a GLS model with random intercept and a random coefficient model, to estimate the scope and scale economies for individual firms. The variation across individual companies has been studied regarding both observed and unobserved heterogeneity. The results suggest significant scope and scale economies at most output levels and regardless of the variation in observed characteristics. The analysis also highlights the effect of unobserved heterogeneity across companies, suggesting that sophisticated econometric specifications such as random coefficients may be superior for analyzing the potential variation in scope and scale economies beyond the observed characteristics such as output patterns and customer density.

Economies of Scale and Scope in Multi-Utilities / 125 The rest of the paper is organized as follows. Section 2 presents the background along with a brief review of previous literature. The model specification and methods are presented in Section 3. Section 4 describes the data and Section 5 presents the regression results. The definition of scale and scope economies and their estimates are discussed in Section 6. The paper ends with a summary of main results and policy conclusions. 2. Background The ongoing regulatory reforms in the energy sector in many countries have adopted measures toward unbundling public utilities into separate operations. The traditional models based on vertical integration in single sectors are often rejected. Especially, in the electricity sector the vertically integrated companies are generally required to unbundle the production, transmission and distribution functions. For instance, the directive 2003/54/EC of the European Parliament and of the EU Council of 26 June 2003 requires a legal and functional unbundling of the utilities operating in a single sector. As opposed to vertical unbundling that is generally being promoted by the ongoing reforms, the horizontal unbundling of multi-utilities has remained an open question with less clear-cut recommendations. The unbinding guidelines released by the EU Directorate-General of Energy and Transport (DG Energy & Transport (2004)) state that the extent of management separation between activities related to different sectors can only be decided on a case by case basis. Further it is highlighted that a clear answer to this unbundling question requires a balanced assessment of, on the one hand, the need for independence and, on the other hand, the interest of multi-utility operators to look for possible synergies. While allowing certain flexibility in unbundling multi-utilities, this note requires the policy makers to assess the extent of the economies of scope before taking policy decisions. According the EU policy directive all the utilities with fewer than 100,000 customers can be exempt from any functional unbundling requirement. The distinction of small and large companies is based on the relative insignificance of scope economies in large companies that exploit scale economies. Such discriminative policies allow small companies to benefit from other synergies than scale economies. Since Switzerland is among the European countries with a large number of small companies in its energy sector, it provides a policy-relevant context for exploring the economies of scope. Moreover, although Switzerland does not belong to the European Union, the Swiss unbundling requirements upcoming in the near future, will probably reflect those discussed in the European directives. This study can provide the Swiss policy-makers with some insight concerning the effectiveness of similar regulatory measures in Switzerland. Unbundling the services into separate functions allows a greater efficiency through stronger and more transparent competition that can be separately introduced in electricity, gas and water sectors. However, the implementation of the unbundling requirements will reduce the possibility of exploiting the economies

126 / The Energy Journal of scope. The analysis of scope economies and its assessment across different companies can have important policy implications for the actual policy debates on the regulatory reforms in the Swiss gas and electricity sectors. Therefore, it is relevant for the Swiss federal authorities to identify if and to what extent multiutility companies are able to use the scope and scale economies to reduce their costs in comparison to a group of single-utility companies. This question is in line with the issue of multiproduct natural monopoly raised by Baumol, Panzar et al. (1982), which has been applied to local public services. In the presence of economies of scope a multiproduct firm is more economical than separate specialized firms. As first identified by Mayo (1984a), such economies are especially significant in relatively small companies. Therefore, the choice to exempt small and medium-size companies from the unbundling requirements could be sustained by economic arguments. In the literature there are only a few studies on the economies of scope in multi-utilities: Mayo (1984a), Chappell and Wilder (1986) and Sing (1987) in electricity and gas distribution, and Fraquelli, Piacenza et al. (2004) and Piacenza and Vannoni (2004) in electricity, gas and water sectors. Mayo (1984a) and Chappell and Wilder (1986) estimate a quadratic cost function for two cross sectional data sets from the US electricity and gas distribution sectors. Mayo (1984a) reports scope economies only for small companies, whereas Chappell and Wilder (1986) conclude significant scope economies over most of output ranges. Sing (1987), also using a cross-sectional data set including electricity and gas distributors, estimates a generalized translog cost function with a Box-Cox transformation for outputs. In addition to the factor prices of labor, capital and fuel, he includes the customer density as an output characteristic. While reporting diseconomies of scope for the sample mean Sing (1987) finds scope synergies for certain output combinations, without any clear pattern with respect to the outputs magnitude. The relatively recent papers by Fraquelli, Piacenza et al. (2004) and Piacenza and Vannoni (2004) use data from 90 Italian electricity, gas and water distributors over 3 years. However the data is pooled across the years and no panel data models are applied. They compare different functional forms such as the translog cost function with a small value transformation, the generalized translog, the separable quadratic and the composite cost function introduced by Pulley and Braunstein (1992). They conclude that economies of scope exist but their statistical significance can only be asserted over small outputs. A summary of the above studies and their main results is presented in Table 1. As we can see, panel data has hardly been utilized to date. The short panels used in the recent studies by Fraquelli, Piacenza et al. (2004) and Piacenza and Vannoni (2004) probably have not allowed the authors to account for unobserved heterogeneity and correlation in the error terms. Another interesting study is Yatchew (2000) who applied a semi-parametric model to a 3-year panel data set of Canadian electricity distributors. Focusing on scale economies that author uses an additional dummy variable to account for the economies gained by joint distribution of water and electricity.

Economies of Scale and Scope in Multi-Utilities / 127 Table 1. Summary of Previous Empirical Studies of Multi-Utilities
Mayo (1984a) Chappell and Wilder (1986) Crosssection 1981, US) Sing (1987) Fraquelli, Piacenza et al. (2004) Piacenza and Vannoni (2004)

Data

Cross-section (1979, US)

Cross-section 1981, US)

Pooled Pooled (1994-96, Italy) (1994-96, Italy) Translog, generalized translog, separable quadratic, composite and general form (Pulley and Braunstein (1992)) NLSUR

Quadratic and Functional flexible fixed form costs quadratic

Quadratic

Generalized translog

Translog, generalized translog, separable quadratic and composite NLSUR

Estimation OLS method Electricity and gas distribution

OLS Electricity and gas distribution Exist over most of the output ranges, +12% for small, -10% for largest companies Global and productspecific economies of scale exist

SUR

Outputs

Electricity and Electricity, gas distribution gas and water distribution Labor, capital, fuel Customer density Output combinations of both scope economies and diseconomies, no economies of scope for the mean output (-7.2%) Product-specific economies of scale for electricity, diseconomies for gas Labor, other inputs -

Electricity, gas and water distribution

Factor prices Other factors

Labor, fuel -

Labor, other inputs Exist with all the models except with the translog cost function. For the median output between 16 and 64%

Exist only for small companies (+0.77%), Economies for large of scope companies diseconomies (up to -11.7%) Product-specific economies of scale for gas Economies over all outputs, of scale for electricity only for small companies

Exist, but significant only for companies producing less than the median output

Exist, but significant only for companies producing less than the median output

All the models show economies of scale except the translog model

Given that the energy distribution companies operate in strongly heterogeneous environments, accounting for firm-specific unobserved factors might change the estimates of scope and scale economies. The moderately long panel data set used in this study allows the use of panel data models that can account for such heterogeneity and assess their effects on the estimations.

128 / The Energy Journal Before turning to the model it is worth noting that the Swiss energy sector is a fragmented market characterized by a strong heterogeneity across the 3,023 communities. With a total of 940 electricity utilities, 124 gas companies and 2,995 water distributors Switzerlands energy sector is characterized by its staggeringly large number of distributors with a prevalence of small and medium size companies throughout the 3,023 Swiss communities (cf. Dymek and Glaubitz (2003), VSG (2007) and Fllmi and Meister (2005)). Multi-utilities play an important role in all three sectors: The share of multiproduct utilities in the electricity and gas sectors is respectively about 35 and 75 percent of the total national consumption. With a roughly estimated share of 80 percent of the total national consumption, multi-utilities are also dominant in the water sector.1 In general multi-utilities tend to be active in all three sectors. The share of double-output utilities is relatively low (limited to a few percentage points), especially in the gas sector. 3. Model Specification and Estimation Method The model specification is based on a cost function with three outputs (electricity, gas and water). The model also includes a measure of the characteristic of the service area and three sector-specific linear time trends capturing technological changes. Moreover, four input prices are also included in the model. As in Sing (1987) customer density is introduced as a service area characteristic. This variable should capture, at least partially, the impact on costs of the heterogeneity of the service area of the companies. In fact, differences in networks and environments influence the production process and, therefore, the costs. Of course, we are aware that the heterogeneity of the service area cannot be summarized into one single variable. Unfortunately, the information is not available for all network and environmental characteristics. Thus, many of these characteristics are omitted from the cost function specifications. As we see later these omitted factors are represented by firm-specific stochastic components in the adopted panel data econometric models. If it is assumed that the firm minimizes cost and that the technology is convex, a total cost function can be written as: C = C (q(1), q(2), q(3), w(1), w(2), w(3), w(4), t(1), t(2), t(3)) (1)

where C represents total costs; q(1), q(2) and q(3) are respectively the distributed electricity, gas and water during the year, w(1), w(2), w(3) and w(4) are respectively the input factor prices for labor and capital services and the purchased electricity and gas; r is the customer density measured by the number of customers divided by the size of the service area measured in square kilometers; and the sector-specific
1. The numbers for electricity and gas are based on the data from 127 electricity distributors and 80 gas companies that respectively provide about 90% of electricity and gas consumption in Switzerland. The share in water distribution is estimated based on the available data from 95 companies that provide about 41 percent of the national water consumption.

Economies of Scale and Scope in Multi-Utilities / 129 linear trends are represented by t(1), t(2) and t(3) respectively for electricity, gas and water sectors. Following Baumol, Panzar et al. (1982) and Mayo (1984a) we use a quadratic cost function. This form has been considered as one of the most relevant options for estimating scope economies (Tovar, Jara-Diaz et al. (2007)). Unlike logarithmic forms, this functional form accommodates zero values for outputs thus, allows a straightforward identification of scope economies. Although logarithmic functions could be used with an arbitrary small value transformation for zero values, it has been shown that this approach could result in large errors in the estimation of scope economies (Pulley and Humphrey (1993)). As in our case, many output values for electricity, gas and water distribution are zero, such estimation errors may lead to misleading conclusions about scope economies. The choice of the quadratic functional form has been also in close relationship with the econometric specification possibilities for the available panel data that will be described later. In fact, unlike other functional forms, the quadratic functional form can be easily estimated with panel data models. For instance, the application of panel data models (especially the random effects models) in non-linear models such as Box-Cox or the composite model (Fraquelli, Piacenza et al. (2004), Mayo (1984b)) is not straightforward. Given the potential importance of the unobserved heterogeneity in the data we focused on the quadratic functional form that is readily adaptable to panel data models. Especially as the utilities operate in environments characterized by strong heterogeneity and given the fact that the integrated companies as well as specialized utilities are included in the data, the omitted variables could have an important effect that can be better accounted for in panel data models. By a similar argument we excluded the equation system approach with factor share equations as this approach cannot easily accommodate random effects specification. One disadvantage of the quadratic form is that the linear homogeneity of the cost function in input prices cannot be imposed by parametric restrictions without compromising the flexibility of the functional form (Caves, Christensen et al. (1980)). A fairly common approach around this issue is the normalization of all monetary variables by one of the common factor prices referred to as numeraire price (see Farsi, Fetz et al. (2007), Featherstone and Moss (1994) and Jara-Diaz, Martinez-Budria et al. (2003)). However, depending on which input factor is chosen as the numeraire, the normalized model has non-unique solutions that might result in certain discrepancy across the estimates.2 Considering this drawback, we favored the non-normalized version of the model that has a greater flexibility as well as a better robustness. Especially in the context of this paper, in which the main focus is on the output coefficients that determine the economies of scope and scale, imposing the linear homogeneity restriction does not appear

2. Because of its additive form the obtained quadratic models are not equivalent. This is in contrast with multiplicative models such as translog in which normalization is perfectly invariant to the choice of the numeraire and equivalent to a single parametric restriction.

130 / The Energy Journal to provide any added value into the analysis.3 The adopted quadratic cost function using a random effects specification can be written as follows: Cit = 0 + mqit(m) + 1/2 mnqit(m) qit(n) + bpwit(p)Di(p)
m m n p M MM P

(2)

+ rrit + gmt(tm) Di(m) + ui + eit

m

where superscripts m and p denote respectively, the number of products (1, 2, 3) and the number of input factors (1, 2, 3, 4), and subscripts i and t denote respectively the company and year. The stochastic terms ui and it represent respectively the firm-specific individual effects and the error term. The factor prices w and the density variable r are introduced in a linear way (following Mayo (1984a)). The dummy variables Di(p) take one if the corresponding input factor has been used in the production. These dummies, relevant only for electricity and gas prices, allow to exclude the corresponding term if the company does not distribute electricity or gas (see Isaacs (2006) for this approach). The linear trends tt(m) are specific to the sector as each one of the sector might be subject to a different technological progress. Similarly, dummy variables Di(m) represent the cases in which the company distributes the corresponding product (electricity, gas and water). Finally 0 is the intercept. The alternative specification would be a flexible fixed cost model as in Mayo (1984a) and Panzar (1989), which includes several intercepts depending on the sector or the utilitys output combination. We explored this possibility, but given that the estimated intercepts are not significantly different from each other, we favored the simpler model with a single intercept. The quadratic form is a flexible functional form that can be considered as a second-order Taylor approximation of any arbitrary function around a local approximation point. In this paper following the commonly used approach in the literature (e.g. Jara-Diaz, Martinez-Budria et al. (2003)), the sample mean has been used as the approximation point. This normalization has been obtained by demeaning all the included explanatory variables (subtracting from their mean values). Therefore the intercept 0 captures the total costs of production at the sample mean. The above cost function has been specified as a random effect GLS model with: ui ~ iid(0,su2).4 This model has a clear advantage over an alternative crosssectional model that pools the data across companies, thus simplifies the firmspecific effects in a constant intercept. Using individual effects ui, the GLS model allows for certain variation among companies regarding the models intercept,
3. This has been confirmed by a supplementary analysis (available upon request) in which we have considered normalizing the costs and input prices by the labor price. The results suggest no significant change as far as the scope and scale economies are concerned. 4. We have also estimated an alternative random effects model with AR1 serial correlation. The results (available upon request) do not show any significant difference between the corresponding coefficients.

Economies of Scale and Scope in Multi-Utilities / 131 that as pointed out by Jara-Diaz, Martinez-Budria et al. (2003), has an important effect on the estimates of economies of scope. The main assumption is that the random effects ui are uncorrelated with the explanatory variables, a restriction that could be relaxed in a fixed-effects specification.5 However, the reliability of fixed-effects estimators depends on the extent of within-company variations that is, the variation of costs and outputs of given companies over time. As Cameron and Trivedi (2005) pointed out, the fixed-effects approach has an important weakness in that the coefficients of explanatory variables are very imprecise if the variables variation over time is dominated by that across companies (between variation).6 The data used in this study show a relatively low within variation (variation over time) in some of the variables, especially, the ratios between the three outputs remain more or less constant within a given company. The extremely low variation in some of the variables coupled with the presence of the second-order terms in the quadratic functional form also exacerbate the risk of multicollinearity, thus unreliable results.7 Moreover, the fixed-effects estimators are strongly conditioned upon the companies included in the sample, thus not convenient for boundary predictions at output bundles with zero values that are required for the estimation of scope economies.8 In fact the definition of the economies of scope relies on a comparison of the companys costs of producing all outputs with those of the same company with zero production in certain outputs. However, changes from positive output to zero output usually do not occur within a specific company. Therefore, the economies of scope can only be identified through the variations between a given company and other companies that are similar in all aspects but have little or zero production in those outputs. In the fixed-effect model such between variations are entirely captured by the companys individual effect, thus excluded from the cost function. Considering the above discussion, we excluded the fixed-effect model and focused on the random effects framework. We recognize however, the limitation of the adopted models concerning the assumption that omitted factors are uncorrelated with the explanatory variables.

5. Such correlation might create heterogeneity bias in the estimates (more on this later). The term heterogeneity bias probably coined by Chamberlain (1982), has also been used for the bias due to ignoring variation of regression coefficients across individuals (e.g. Asteriou and Hall (2007)). 6. Johnston and DiNardo (1997) also show that the attenuation bias due to measurement errors is exacerbated in the fixed-effects models depending on the fraction of the within variation due to mismeasurement especially when the explanatory variables are correlated across time. In our case it is plausible that the reporting errors have a contribution in the observed within variations. 7. Following a referees suggestions we estimated several fixed-effects models. The results (available upon request) indicate that the estimates of the main output coefficients are quite sensitive to the included variables and occasionally counter-intuitive, suggesting that the within variation is not sufficient in order for the fixed-effects model to provide sensible results. 8. As pointed out by Hsiao (2003), while the fixed-effects model is more appropriate for conditional predictions for individuals, the random effects is a better specification for unconditional (populationaveraged) analysis provided that the random effects are uncorrelated with the explanatory variables. See also Cameron and Trivedi (2005) and Verbeek (2004) for a discussion of this issue.

132 / The Energy Journal In the random effects model the unobserved firm-specific heterogeneity is accounted for by individual effects. These factors might be correlated with the explanatory variables, in which case the estimations might be affected by heterogeneity bias. One improvement over the GLS model in this respect could be obtained by including random coefficients for those explanatory variables. The variation of these coefficients should capture part of the correlation of the random intercept with the corresponding variables. Moreover, the unobserved firm-specific heterogeneity could also apply to marginal costs represented by the coefficients of the cost function. Therefore, we also estimate the cost function using a random coefficient (RC) model.9 In this model the three output coefficients, the intercept and the output characteristics are assumed to be random variables with a normal distribution across companies. The quadratic cost function with the adopted random coefficient specification can be written as follows: Cit = 0i + imqit(m) + 1/2 mnqit(m) qit(n) + bpwit(p)Di(p)
m m n p M MM P

(3)

+ irrit + gmt(tm) Di(m) + eit

m

where superscripts m and p denwhere im ~ N(am,s2 ), for m=0,1,2,3, and am ir ~ N(ar,s2 ). Similar to the GLS model, all the explanatory variables are norar malized to their sample means. The above random coefficient model has been estimated using a simulated maximum likelihood method. The firm-specific parameters are estimated for individual companies as their conditional expectation. The random coefficient model described above provides a relatively rich specification that allows for interaction of unobserved factors such as network characteristics with outputs and customer density. However it has a shortcoming in that it imposes the normality assumption on the random intercept. Therefore the choice of the best model between the two depends on the trade-off between refining the econometric specification against the distribution restrictions. As we see later, as far as the estimates of the economies scale and scope are concerned, the results are not sensitive to this choice. Another important issue is that the specification of random coefficients can be extended to other variables. The benefits of such extensions should be assessed against the entailed numerical difficulties as well as the interpretation problems.10

9. For a presentation of this model see Cameron and Trivedi (2005). See also Birn, Lindquist et al. (2002) for an application of this model in the estimation the returns to scale among heterogeneous technologies. 10. Following the suggestion of a referee we estimated several alternatives in which the input prices especially capital price have also random coefficients. The results (available upon request) indicate that adding random coefficients to the model can cause convergence problems and numerical instability, otherwise, counterintuitive results that are difficult to interpret. These problems could be explained by the relatively large number of explanatory variables in the model and the fairly limited number of companies included in the data.

Economies of Scale and Scope in Multi-Utilities / 133 4. Data The unbalanced panel data set used for this analysis contains financial and technical information from 87 companies observed during the nine-year period between 1997 and 2005. The companies in the sample cover about 42% of total electricity, 67% of total gas and 22% of total water distribution in Switzerland. Among these companies, 33 are fully integrated and offer electricity, gas and water. 11 companies offer electricity and water, 3 companies distribute gas and water and 2 companies electricity and gas. The remaining companies are specialized companies from which 23 are active only in electricity distribution, 12 only in gas distribution and 3 only in water distribution. The presence of just a few number of specialized water distributors could be considered as a drawback for the estimation of economies of scope. However, this limitation should be considered together with the fact that in a fair number of companies in the sample, the distribution of gas and electricity constitutes a small fraction of the total output.11 The data were collected from the companies annual reports containing financial and technical information.12 As pointed out by Kaserman and Mayo (1991), the degree of vertical integration can have an important impact on costs, thus affecting the estimates of economies of scope. The problem does not arise in gas and water sectors, in which companies have a uniform level of integration with the generation section (fully integrated in the case of water and completely separate in gas companies). In order to abstract from the effect of vertical integration in electricity distribution, companies with more than 10% self-generation of total electricity distribution were excluded. The variables for the cost function specification were constructed as follows. Total costs (C) are calculated as the total expenditures of the energy and water distribution firms in a given year. The outputs q(m) are measured by the total quantity delivered to the customers. The measurement units are GWh for electricity and gas and million cubic meters for water.13 Input prices are defined as factor expenditures per factor unit. Labor price (w(1)) is defined as the ratio of annual labor costs to the total number of employees as full time equivalent. As data on full time equivalent was not available for 40 companies and taking the number of employees including part time workers would underestimate the labor price, a correction was done by taking the mean with the labor price of the companies within the same canton. Following Friedlaender and Chiang (1983), the capital price (w(2)) is calculated as residual cost
11. The number of these companies depends on the units used for measuring the various outputs. For instance, if we choose the units such that the sample median values will have the same order of magnitude (GWh for gas and electricity 104 cubic meters for water output) there are 14 companies whose water output is more than two third of their total output. 12. Information on the size of the firms distribution area is from the Arealstatistik 2002 from the Federal Statistical Office and from the Preisberwacher. 13. The distributed gas is generally reported in energy units rather than volume units. Given that the gas distributors in Switzerland mainly use the same source of imported natural gas with a uniform quality, we do not expect that the change of measurement unit has any effect on the results.

134 / The Energy Journal (where residual cost is total cost minus labor and electricity and gas purchases) divided by the network length.14 For the multi-utilities, the prices were weighted by the share of the residual costs in each sector to the total residual costs in all sectors (see also Fraquelli, Piacenza et al. (2004) for this approach). The electricity and gas price is defined as the expenditures of purchasing the input factors divided by the amount purchased (in MWh). Table 2 provides the samples descriptive statistics. All the costs and prices are adjusted for inflation using consumer price index and are measured in year 2000 Swiss Francs (CHF). As can be seen in the table, the sample shows a considerable variation in all three outputs. Table 2. Descriptive Statistics (622 observations)
Variable C q
(1)

Unit CHF Million GWh GWh Million m3 CHF/ employee CHF/ km CHF/ MWh CHF/ MWh Customers/ km
2

Min. 1.52 0 0 0 75,575 8,165 43.5 16.3 2.3

Median Mean 35.7 115.4 78.8 1.1 26,421 103.2 29.1 230.1 79.0 405.9 363.8 3.4 34,018 101.0 30.3 348.0

Max. 611.8 6,177.0 6,665.0 62.6 153,673 234,796 163.5 63.2 3,619.3

Total cost Electricity distribution Gas distribution Water distribution Labor price Capital price Electricity price Gas price Density

q (2) q (3) w (1) w (2) w (3) w

(4)

103,610 104,863

5. Results The estimation results obtained from the GLS model are given in Table 3. These results show that the output and input price coefficients are highly significant and have the expected positive sign. As expected, the effect of customer density (coefficient r), is negative, showing that an increase in the customer density decreases costs. The coefficients of the linear trends suggest different technological progress across the three sectors. The results, while suggesting a cost decrease in the electricity networks, indicate a growth in operating costs in both gas and water sectors. These differences might also be related to the differences in the regulation systems for these sectors. It is interesting to note that although almost all public utilities are undergoing regulatory reforms, the electricity distributors have been subject to a relatively more advanced de-regulation process.15 However, the relative growth of costs in
14. More precise measures of capital stock and expenditures can be obtained from a perpetual inventory approach. Unfortunately such inventory data was not available. 15. The first official attempt for the de-regulation of the Swiss electricity market dates back to 2002.

Economies of Scale and Scope in Multi-Utilities / 135 gas and water networks might be related to the relatively higher age of these networks, thus a more accentuated need for new investments. Another interesting observation is the considerable variation of the random intercept as reflected in the estimate of su. The significant variation of the fixed costs across companies might be considered as a support for models with flexible fixed costs suggested by Mayo (1984a) and Panzar (1989). However, our additional estimations with a similar model but a varying intercept across different sectors suggest no statistically significant differences across sectors. This result combined with a relatively important variation in the random effects indicates that the variation in the fixed costs across companies might be mainly due to unobserved variations across companies. However, as we will see later from the random coefficient models results the GLS model could overstate the variation of intercept because it assumes constant slopes for all companies. Table 3. Regression Results (GLS model)
Quadratic cost function (GLS)
1 2 3

b b b b g g g

(Electricity output) (Gas output) (Water output)

152,698 42,659 2,266,445 -22.33

** ** ** **

(3,318) (4,210) (504,478) (1.33) (1.54) (22,532) (3.91) (366) (230.47) (75.77) (33.78) (52,794) (111,478) (2,973.91) (331,928) (390,922) (461,136) (1,926,850)

11 22 33 12 13 23

0.18 -43,314 21.27 -1,687 -970.71 132.75 139.85 127,777 562,209 -7,207.54 -2,639,987 945,850 1,544,447 90,140,600 * ** ** ** * ** ** ** ** ** ** ** **

1 2 3

(Labor price) (Capital price) (Electricity price) (Gas price)

14 r

(Customer density)

1 2 3

(Electricity sector) (Gas sector) (Water sector)

s s

u e

10,586,724 9,411,338

** and * indicate 5% and 10% significance level respectively. Standard errors are given in parentheses.

136 / The Energy Journal Table 4. Regression Results (Random Coefficient Model)
Quadratic cost function (RC) Mean Standard deviation (1,533) (1,639) (184,820) (0.95) (0.539) (8,684) (2.80) (154) (150.02) (30.94) (15.97) (23,154) (64,763) (983.31) (198,250) (424,060) (432,471) (682,477) 108,524 (355,020) 14,200.7 ** (1,157.4) 15,652 ** 14,921 ** 9,289 (524) (882) (48,903)

1 (Electricity output) 2 (Gas output) 3 (Water output) 11 22 33 12 13 23 b1 (Labor price) b2 (Capital price) b3 (Electricity price) b4 (Gas price) ar(Customer density) g1 (Electricity sector) g2 (Gas sector) g3 (Water sector) 0

162,889 ** 50,132 ** 1,562,760 ** -32.21 ** -0.63 -12,262 25.78 ** -1,704 ** -399.68 ** 126.56 ** 128.91 ** 91,957 ** 522,290 ** -3,829.95 ** -2,488,370 ** 916,995 ** 1,323,520 ** 93,564,700 **

** and * indicate 5% and 10% significance level respectively. Standard errors are given in parentheses.

In the random coefficient model, it is assumed that the intercept and the first-order coefficients of output and customer density vary across companies. The random coefficient model was estimated with the simulated likelihood method using quasi-random Halton draws.16 Table 4 lists the regression results obtained from the random coefficient model. The first important observation is that the estimated coefficients are slightly (but mostly not significantly) different from those obtained from the GLS model. However, the estimated standard deviations of the random coefficients are all statistically significant for electricity and gas output as well as the customer density. This suggests that there is a significant variation in the output and density coefficients across companies. As for the intercept the standard deviation shows a in Table 3), considerably lower value than that obtained from the GLS model (s u suggesting that ignoring the heterogeneity in slopes can result in an overestima16. The number of draws has been fixed to 1000. The model was also estimated with several numbers of draws between 100 and 1,000. The results indicate that after 500 draws, the estimations become stable.

Economies of Scale and Scope in Multi-Utilities / 137 tion of the variations of the fixed costs across companies. The random coefficient estimators can be used to estimate the conditional expectation of firm-specific coefficients. These estimates show that for the intercept and the output coefficients, all the coefficients are positive, while for the customer density coefficient, 13 companies (out of 87) have positive coefficients. This can be explained by the fact that as customer density increases, certain companies incur extra costs through congestion effects or some unobserved network characteristics.17 The estimates of the variances of the random effects in both models (Table 3 and Table 4) show that there is a considerable unobserved firm-specific heterogeneity. We will see later if and how ignoring this heterogeneity could affect the estimates of scale and scope economies. The estimation results presented in Table 3 and Table 4 can be used to compute the estimated of the economies of scale and scope. These results along with a formal description of the concepts will be presented in the following section. 6. Scale and Scope Economies Following Baumol, Panzar et al. (1982) the global economies of scale in a multi-output setting are defined as: C(q) SL = M q(m) * C / q(m)
m (1) (2)

(4)

where q=(q , q , q(3)) for m=1 (electricity), 2 (gas) and 3 (water). Global economies of scale describe the cost behavior due to proportional changes in the entire production. The returns to scale are increasing, constant or decreasing if the corresponding ratio (SL) is greater, equal or less than one. Economies of scope are present when costs can be reduced by joint production of multiple outputs. Following Baumol, Panzar et al. (1982) the degree of global economies of scope across three products is defined as the ratio of excess costs of separate production to the costs of joint production of all outputs: C(q (1) , 0, 0) + C(0, q (2) , 0) + C (0, 0, q (3)) C(q) SC = C(q) (5)

A positive (negative) value for the above expression implies the existence of global economies (diseconomies) of scope.
17. We explored the possibility that the congestion effect might be related to some observed variables by including a square term for customer density and accounting for the network location in rural/urban areas. The results do not show statistical significant effect which could lead to any conclusive evidence in this regard.

138 / The Energy Journal Scope and scale economies are usually estimated using the deterministic part of the cost function at some representative outputs. In previous studies these representative outputs are generally obtained by setting the outputs at different points of their sample distribution such as median and other quartiles. As seen in Equation (5), a correct estimation of economies of scope relies on adequately predicting of costs at certain points that are at the sample boundary or completely out of the sample. The precision of such predictions depends on the econometric specification. As discussed earlier, a GLS model provides a relatively accurate out-of-sample prediction. The random-coefficient model has an additional advantage with respect to heterogeneity bias in the coefficients. The predictions required for estimating scope and scale economies in Equations (4) and (5), can also be conducted at the individual company level, using the individual estimates of company-specific random effects and coefficients. The individual company-level estimates can better represent the actual output patterns. The company-level cost predictions might however entail relatively large estimation errors. In this paper, we have used both approaches. Using Equations (4) and (5) and the regression results, the values of scope and scale economies have been estimated for five hypothetical companies with representative output combinations. These companies are characterized by the 1st, 2nd, 3rd and 4th quintiles and the sample median of the non-zero output values and customer density. A summary of these results is provided in Table 5. These results suggest the presence of scope and scale economies at most output levels. The estimates also show a well-behaved variation: as outputs increase (decrease) both scale and scope economies fall (rise). Table 5. Point Estimates of Global Economies of Scope and Scale
Representative firm 1st Quintile 2nd Quintile Median 3rd Quintile 4th Quintile Economies of Scope GLS 0.37 0.22 0.17 0.11 0.03 RC 0.27 0.16 0.12 0.07 -0.003 Economies of Scale GLS 1.24 1.14 1.10 1.07 1.06 RC 1.17 1.09 1.07 1.04 1.03

The representative points are based on positive values of the three outputs as well as the customer density. Input prices and time trends are kept constant at their sample mean values. The random effects (and coefficients) are assumed to be at their mean values. Representative sample points such as output quintiles correspond to hypothetical productions that vary in overall scale and density as they represent a more or less similar ratio between all outputs. In this case the firms with non-typical mixtures of outputs and customer density would not be represented. In order to study the variation of scale and scope economies in the sample, based on the actual

Economies of Scale and Scope in Multi-Utilities / 139 levels of production rather than hypothetical values, we computed the economies of scope and scale for each individual company. Note that the definitions of global economies of scope and scale as defined in Equations (4) and (5) is directly applicable only to all-positive-output combinations. In order to extend the estimates to other companies we have chosen a hypothetical all-positive output for each one of these companies. While keeping the positive observed values, we replaced the zero values by a positive value constructed based on the companys overall scale relative to all the companies in the sample. For any given company the overall scale factor is defined as that companys maximum output standardized by the mean value and standard deviation of that output observed in the sample. For any given company the hypothetical output of a given zero output is constructed by multiplying the companys overall scale factor by the sample mean value of that output. An alternative method would be to limit the estimates to the companies with all-positive outputs. However, the fact that the fully integrated companies might be a selection of companies in that they exploit the economies of scope and might have a lower fixed costs, could distort the estimates of scope economies.18 Table 6 and Table 7 respectively provide a summary descriptive of the distribution of the estimates of the global economies of scope and scale across the companies included in the sample. The results obtained from both GLS and RC models are listed. The first and third columns provide the estimates obtained by ignoring the random effects, namely the means of the random coefficients are considered. In the second and fourth columns, the firm-specific random effects are included in the calculation of scale and scope economies. The input prices and the time trends have been set equal to their mean values over the entire sample. Both GLS and RC estimates suggest the existence of scope and scale economies across a major part of the sample. Looking across the numbers from both models indicate that more than 60 percent of the companies can exhibit economies of scope and at least 80 percent can benefit from economies of scale. Table 6. Distribution of Global Economies of Scope Estimated for Individual Companies
1st Quintile 2nd Quintile Median 3rd Quintile 4th Quintile
a)

GLS a 0.05 0.09 0.14 0.17 0.25

GLS b -0.11 0.05 0.15 0.19 0.33

RC a -0.02 0.05 0.10 0.11 0.18

RC b -0.18 0.02 0.04 0.10 0.29

Individual random effects are not taken into account. b) Individual firm-specific random effects are included in the computations. The values are estimated for all individual observations. Input prices and time trends are kept constant at their sample mean values. 18. We have also estimated these values for the 33 fully integrated companies. The results do not show much difference.

140 / The Energy Journal

Table 7. Distribution of Global Economies of Scale Estimated for Individual Companies

1st Quintile 2nd Quintile Median 3rd Quintile 4th Quintile
a)

GLS a 1.08 1.11 1.12 1.13 1.22

GLS b 0.97 1.09 1.15 1.19 1.28

RC a 1.04 1.06 1.07 1.08 1.13

RC b 1.00 1.05 1.07 1.09 1.24

Individual random effects are not taken into account. b) Individual firm-specific random effects are included in the computations. The values are estimated for all individual observations. Input prices and time trends are kept constant at their sample mean values.

Assuming that the larger companies have a lower potential of scale and scope economies (as suggested by Table 5), these results indicate that all small and moderate-sized utilities can benefit from significant savings through scale and scope economies. However, as seen in Table 6 and Table 7 the extent of these economies can vary depending on the adopted model and the approach used for accounting the estimated effects of unobserved factors. The first and third columns in both tables indicate that if the random effects are not considered in the computations, GLS and RC models provide a quite similar distribution of scale and scope economies across companies. However, a comparison of the first and third columns with the second and fourth ones respectively, suggests that including the individual random effects results in a wider range of variation in scale and scope economies. These results indicate that the economies of scope and scale could be influenced by unobserved factors beyond output and density. We could not find any conclusive pattern suggesting a one-sided bias because of ignoring such unobserved heterogeneity. The results suggest however that compared to GLS model, the RC model provides a lower overall estimate of both economies, as seen in slightly lower median values. This could be explained by the fact that the RC model gives a relatively lower weight to differences regarding fixed costs because part of these costs might be captured by random coefficients. However, it should be noted that some of the observed variation in the above tables might be related to the relatively large estimation errors of the fixed costs across all models. Considering that the reliability of the individual estimates remains a contentious issue, we contend that the extreme values especially those of scope economies should be considered with caution. Overall these results suggest that a great majority of the companies can benefit from significant economies of scope and scale. Considering the median values these savings vary depending on the model, from 4 to 15 percent for scope economies and 7 to 15 percent for scale economies. Especially the small multi-utilities benefit from considerable scope economies that could reach 20 to 30 percent of total costs.

Economies of Scale and Scope in Multi-Utilities / 141 7. Conclusions Using a panel data set from the distribution utilities operating in water, gas and electricity sectors this paper has studied the economies of scope and scale in multi-output utilities. A random effect panel data (GLS) model and a random-coefficient (RC) model have been used to explore the effect of unobserved heterogeneity across different networks. While the GLS model considers the unobserved heterogeneity as various cost shifts across companies, the RC model includes variations in marginal effects of outputs and customer density. Compared to cross-sectional model, the GLS specification provides a better control for omitted variables. The RC model provides an additional improvement regarding the potential heterogeneity bias in the coefficients estimates. This paper also shows that the computation of the economies of scope and scale can be extended to include the estimates of firm-specific individual effects, namely the conditional expectation of the random intercept and random coefficients. While admitting that such company-level estimates may entail relatively large estimation errors at the individual level, we assert that the overall results could represent a better picture of scope and scale economies based on actual levels of outputs and network characteristics rather than simplified hypothetical values. From the results three general observations can be pointed out. First, the results confirm the existence of significant scope and scale economies in a majority of multi-utilities, which can be considered as suggestive evidence of natural monopoly in multi-utilities. This conclusion is confirmed across the two models and regardless of whether the individual firm-specific stochastic terms are included in the estimations. Secondly, considerable variation of the estimated values among individual companies suggests that the economies of scope and scale can depend on unobserved network characteristics as well as output patterns and customer density. Finally, the variations across the models indicate that the overall point estimates are not very sensitive to the specification of unobserved firm-specific factors. The results of this paper show that even after accounting for unobserved heterogeneity, the scope economies exist in a majority of the multi-utilities, suggesting that additional costs could result from unbundling the multi-utility companies. In the actual situation many companies avoid these additional costs through scope economies. Especially for small companies the savings associated with scope economies are considerable. In this study it is assumed that there is no functional separation between distribution and supply to end-use consumers. This assumption closely reflects Switzerlands actual situation and most probably, its future development. In fact, under the EU policy directive the utilities with fewer than 100,000 customers can be exempt from any functional unbundling requirement. As most of the distribution companies in Switzerland are relatively small with only a few companies having more than 100,000 customers, with a likely adoption of policies similar to those of EU, the multi-utilities are likely to remain integrated in the future. There-

142 / The Energy Journal fore, the results of this study are especially relevant for the context of Switzerland as well as in many similar cases in other countries. References
Asteriou, D. and S. G. Hall (2007). Applied econometrics: A modern approach. New York: Palgrave Macmillan. Baumol, W. J., J. C. Panzar and R. D. Willig (1982). Contestable Markets and the Theory of Industry Structure. New York: Hartcourt Brace Jovanovich. Birn, E., K.-G. Lindquist and T. Skjerpen (2002). Heterogeneity in Returns to Scale: A Random Coefficient Analysis with Unbalanced Panel Data. Journal of Productivity Analysis 18(1): 39-57. Cameron, A. C. and P. K. Trivedi (2005). Microeconometrics Methods and Applications. Cambridge: Cambridge University Press. Caves, D. W., L. R. Christensen and M. W. Tretheway (1980). Flexible Cost Functions for Multiproduct Firms. Review of Economics and Statistics 62(3): 477-81. Chamberlain, G. (1982). Multivariate Regression Models for Panel Data. Journal of Econometrics 18(1): 5-46. Chappell, H. W. and R. P. Wilder (1986). Multiproduct Monopoly, Regulation, and Firm Costs: Comment. Southern Economic Journal 52(4): 1168-74. DG Energy & Transport (2004). Note of DG Energy & Transport on the Directives 2003/54/EC and 2003/55/EC on the internal Market in Electricity and Natural Gas, The Unbundling Regime, Brussels. Dymek, B. and D. Glaubitz (2003). Rtteln am eidgenssischen Tor. Energie & Management 19: 4-5. Farsi, M., A. Fetz and M. Filippini (2007). Economies of scale and scope in local public transportation. Journal of Transport Economics and Policy 41(3): 345-361. Featherstone, A. M. and C. B. Moss (1994). Measuring Economies of Scale and Scope in Agricultural Banking. American Journal of Agricultural Economics 76(3): 655-61. Fllmi, R. and U. Meister (2005). Product-Market Competition in the Water Industry: Voluntary Nondiscriminatory Pricing. Journal of Industry, Competition and Trade 5(2): 11535. Fraquelli, G., M. Piacenza and D. Vannoni (2004). Scope and Scale Economies in Multi-Utilities: Evidence from Gas, Water and Electricity Combinations. Applied Economics 36(18): 2045-57. Friedlaender, A. F. and W. Chiang (1983). Productivity Growth in the Regulated Trucking Industry. Research in Transportation and Economics 1: 149-84. Hsiao, C. (2003). Analysis of panel data. Cambridge: Cambridge University Press. Isaacs, J. P. (2006). Determining the Regulatory Impact on Vertical Economies in the Electric Utility Industry, Fourth Annual International Industrial Organization Conference, Boston, Massachusetts. Jara-Diaz, S. R., E. Martinez-Budria and F. J. Ramos-Real (2003). Adapting Productivity Theory to the Quadratic Cost Function: An Application to the Spanish Electric Sector. Journal of Productivity Analysis 20(2): 213-29. Johnston, J. and J. DiNardo (1997). Econometric methods. New York: McGraw-Hill. Kaserman, D. L. and J. W. Mayo, (1991). The measurement of Vertical economies and the Efficiency Structure of the Electricity Utility Industry Journal of Industrial Economics 34 (5): 483-502. Mayo, J. W. (1984a). Multiproduct Monopoly, Regulation, and Firm Costs. Southern Economic Journal 51(1): 208-18. Mayo, J. W. (1984b). The Technological Determinants of the US Energy Industry Structure. Review of Statistics and Economics 66 (1): 51-58. Panzar, J. C. (1989). Technological Determinants of Firm and Industry Structure, Handbook of Industrial Organization, R. Schmalensee and R. D. Willig, Amsterdam, Elsevier, Volume 1: 3-59. Piacenza, M. and D. Vannoni (2004). Choosing among Alternative Cost Function Specifications: An Application to Italian Multi-utilities. Economics Letters 82(3): 415-22.

Economies of Scale and Scope in Multi-Utilities / 143

Pulley, L. B. and Y. M. Braunstein (1992). A Composite Cost Function for Multiproduct Firms with an Application to Economies of Scope in Banking. Review of Economics and Statistics 74(2): 221-30. Pulley, L. B. and D. B. Humphrey (1993). The Role of Fixed Costs and Cost Complementarities in Determining Scope Economies and the Cost of Narrow Banking Proposals. Journal of Business 66(3): 437-62. Sing, M. (1987). Are combination Gas and Electric Utilities multiproduct natural Monopolies? Review of Economics and Statistics 69(3): 392-8. Tovar, B., S. R. Jara-Diaz and L. Trujillo (2007). Econometric Estimation of Scale and Scope Economies Within the Port Sector: A Review. Maritime Policy & Management 34(3): 203-23. Verbeek, M. (2004). A guide to modern econometrics. Chichester: Wiley. VSG (2007). Erdgas - Daten und Fakten, Tabellen und Grafiken 2006/07, Verband der Schweizerischen Gasindustrie (VSG), Zrich. Waldman, D. E. and E. J. Jensen (2001). Industrial organization theory and practice. Boston: Addison-Wesley. Yatchew, A. (2000). Scale Economies in Electricity Distribution: A Semiparametric Analysis. Journal of Applied Econometrics 15(2): 187-210.

144 / The Energy Journal

Identifying the Rebound: Evidence from a German Household Panel

Manuel Frondel, Jrg Peters, and Colin Vance* Using a panel of household travel diary data collected in Germany between 1997 and 2005, this study assesses the effectiveness of fuel efficiency improvements by estimating the rebound effect, which measures the extent to which higher efficiency causes additional travel. Following a theoretical discussion outlining three alternative definitions of the rebound effect, the econometric analysis generates corresponding estimates using panel methods to control for the effects of unobservables that could otherwise produce spurious results. Our results, which range between 57% and 67%, indicate a rebound that is substantially larger than obtained in other studies, calling into question the efficacy of policies targeted at reducing energy consumption via technological efficiency. 1. Introduction The improvement of energy efficiency is often asserted to be one of the most promising options to reduce both the usage of energy and associated negative externalities, such as carbon dioxide emissions (CO2). Ever since the creation of the Corporate Average Fuel Economy (CAFE) standards in 1975, this assertion has been a mainstay of energy policy in the United States. In recent years, it has also found increasing currency in Europe, as attested to by the voluntary agreement negotiated in 1999 between the European Commission (EC) and the European Automobile Manufacturers Association, stipulating the reduction of average emissions to a target level of 140g CO2/km by 2008. Additionally, the EC is considering legislation that would set a target of 120g CO2/km by 2012.
The Energy Journal, Vol. 29, No. 4. Copyright 2008 by the IAEE. All rights reserved. * Corresponding author: Dr. Manuel Frondel, Rheinisch-Westflisches Institut fr Wirtschaftsforschung (RWI Essen), Hohenzollernstr. 1-3, D-45128 Essen. E-mail: frondel@rwi-essen.de.

We are very grateful for valuable comments and suggestions by five anonymous referees, and the editor, Lester HUNT. We also would like to thank participants of the Ninth European IAEE Conference 2007 in Florence, Italy, and the participants of a seminar at Clark University for constructive discussions.

145

146 / The Energy Journal Although such technological standards undoubtedly confer benefits via reduced per-unit prices of energy services, the extent to which they reduce energy consumption, and hence pollution, remains controversial. It is plausible, for instance, that the owner of a more fuel-efficient car will ceteris paribus drive more in response to lower per-kilometer traveling costs relative to other modes. This increase in service demand is called the rebound effect, alternatively referred to as take back of efficiency improvements. Khazzoom (1980) was among the first to study the rebound effect at the microeconomic level of households, focusing on the effects of increases in the energy efficiency of a single energy service, such as space heating and individual conveyance. The rebound, however, is a general economic phenomenon, diminishing the potential gains of innovations that may reduce the usage of resources such as water, as well as of time-saving technologies (e. g. Binswanger 2001). The significance of the rebound has been hotly debated among energy economists, as documented in several surveys of the relevant literature (e.g. Brookes 2000, Greening et al. 2000, and Sorrell 2007). Though the basic mechanism is widely accepted, the core of the controversy lies in the identification of the magnitude of the direct rebound effect, which describes the increased demand for an energy service whose price shrinks due to improved efficiency.1 This substitution mechanism in favor of the energy service works exactly as would the price reduction of any commodity other than energy, and suggests that price elasticities are at issue when it comes to the estimation of direct rebound effects. Some analysts, most notably Lovins (1988), but also Greene (1992), and Schipper and Grubb (2000), maintain that these effects are so insignificant that they can safely be ignored. Other authors argue that they might be so large as to completely defeat the purpose of energy efficiency improvements (Brookes 1990, Saunders 1992, Wirl 1997). Support for both views is found in the available empirical evidence. In the case of personal automotive transport, for example, an international survey by Goodwin, Dargay, and Hanly (2004) indicates rebound effects varying between 4% and 89% from studies using pooled cross-section/time-series data. Results from subsequent studies are equally wide-ranging. Using cross-sectional micro data from the 1997 U.S. Consumer Expenditure Survey, West (2004) finds a rebound effect that is 87% on average, while Small and Van Dender (2007), who use a pooled cross-section of U.S. states for 1966-2001, uncover rebound effects varying between 2.2% and 15.3%. Aside from differences in the level of data aggregation, one major reason for the diverging results of the empirical studies is that there is no unanimous definition of the direct rebound effect. Instead, several definitions have been
1. The indirect rebound effect and general equilibrium effects have also been distinguished in the literature (see, e.g., Greening and Greene 1997, Greene et al. 1999). The former arises from an income effect: lower per-unit cost of an energy service implies ceteris paribus that real income grows. The latter arises from innovations, such as James Watts famous steam engine, that increase societys aggregate income potential. Given that both indirect and general equilibrium effects are difficult to quantify, the overwhelming majority of empirical studies confines itself to analyzing the direct rebound effect.

Identifying the Rebound: Evidence from a German Household Panel / 147 employed as determined by the availability of price and efficiency data, making comparisons across studies difficult. The variety of definitions used in the economic literature is summarized and analyzed in an illuminating way by Sorrell and Dimitroupoulos (2008), who argue that it is particularly due to the omission of potentially relevant factors, such as capital cost, that the size of the rebound effect might be frequently overestimated in empirical studies. Greene, Kahn, and Gibson (1999) and Small and Van Dender (2007) express similar reservations, noting in particular the shortcomings of cross-sectional or pooled approaches that fail to control for the time-invariant effects of neighborhood design, infrastructure, and other geographical features, which are likely to be strongly correlated with fuel economy and travel. Departing from the theoretical grounds provided by Beckers (1965) classical household production function approach and drawing on a panel of household travel data, this paper focuses on estimating the direct rebound effect from variation in the fuel economy of household vehicles. Several features distinguish our analysis. After cataloguing three commonly employed definitions of the direct rebound effect in the theoretical section of the paper, econometric estimates corresponding to each of the three definitions are provided in the empirical section. These estimates are generated from panel models of micro-level data, thereby bypassing aggregation problems while at the same time controlling for time-invariant omitted variables. Our results, which range between 57% and 67%, indicate a rebound that is substantially larger than the typical effects obtained from the U.S. transport sector. Based on household survey data, Greene, Kahn, and Gibson (1999:1), for instance, find a long-run take back of about 20% of potential energy savings, confirming the results of other U.S. studies using national and or state-level data. While this issue has received relatively less scrutiny in the European context, our results are also substantially larger than those of Walker and Wirl (1993), who estimate a long-run rebound effect of 36% for Germany using aggregate timeseries data. The following section presents three definitions of the direct rebound effect, building the basis for the empirical estimation. Section 3 describes the econometric specifications and estimators. Section 4 describes the panel data base used in the estimation, followed by the presentation and interpretation of the results in Section 5. The last section summarizes and concludes. 2. A Variety of Direct Rebound Effect Definitions Along the lines of Beckers seminal work on household production, we assume that an individual household derives utility from energy services, such as mobility or comfortable room temperature. A specific service is taken to be the output of a production function f, where f describes how households produce the service in the amount of s by using energy, e, time, t, capital, k, and other market goods o:

148 / The Energy Journal s = f (e,t,k,o). (1)

Using this framework, we begin by drawing on the definition of energy efficiency typically employed in the economic literature (e.g. Wirl 1997): = s/e > 0, (2)

where the efficiency parameter characterizes the technology with which a service is provided. For the specific example of individual conveyance, parameter designates fuel efficiency, which can be measured in terms of vehicle kilometers per liter of fuel input. Efficiency definition (2) assumes proportionality between service level and energy input regardless of the level a simplifying assumption that may not be true in general, but provides for a convenient first-order approximation of the relationship of s with respect to e. Efficiency definition (2) reflects the fact that the higher the efficiency of a given technology, the less energy e = s/ is required for the provision of a certain amount s of energy service. Hence, the concept of energy efficiency is perfectly in line with Beckers idea of household production, according to which households are, ultimately, not interested in the amount of energy required for a certain amount of service, but in the energy service itself. Based on efficiency definition (2), it follows that the price ps per unit of the energy service, given by the ratio of service cost to service amount, is smaller the higher the efficiency is: e pe e pe ps = = pe = . s s (3)

We now provide a concise summary of three widely known definitions of the direct rebound effect that are based on efficiency, service price, or energy price elasticities. Using each of these definitions and data on fuel efficiency, fuel prices, distance driven, and fuel consumption for household vehicles, we will estimate the direct rebound effect. The proofs of the propositions that complement the rebound definitions are given in Appendix A. Definition 1: The immediate and most general measure of the direct rebound effect see e.g. Berkhout et al. (2000) is given by (s) = lns/ln, the elasticity of service demand with respect to efficiency, reflecting the relative change in service demand due to a percentage increase in efficiency. Proposition 1: Having (s) in hand, we obtain the relative reduction in energy use due to a percentage change of efficiency: (e) = (s) 1 . (4)

Only if (s) equals zero, that is, only if there is no direct rebound effect, (e) amounts to 1, indicating that 100% of the potential energy savings due to an efficiency improvement can actually be realized.

Identifying the Rebound: Evidence from a German Household Panel / 149 Definition 2: Instead of (s), empirical estimates of the rebound effect are frequently based on ps (s), the negative price elasticity of service demand see e.g. Binswanger (2001) and Greene et al. (1999). Major reasons for this preference are that data on energy efficiency is often unavailable or data provides only limited variation in efficiencies. The basis for this definition is given by the following proposition. Proposition 2: If energy prices pe are exogenous and service demand solely depends on ps, then (s) = ps (s). (5)

That the rebound may be captured by ps (s) reflects the fact that the direct rebound effect is, in essence, a price effect, which works through shrinking service prices ps. Definition 3: Empirical estimates of the rebound effect are sometimes necessarily based on pe (e), the negative own-price elasticity of energy consumption, rather than on ps (s), because data on energy consumption and prices is more commonly available than on energy services and service prices. It was this definition of the rebound that was originally introduced by Khazzoom (1980:38) and is also employed by Wirl (1997:30). The basis for Definition 3 is given by the following proposition. Proposition 3: If the energy efficiency is constant, then pe (e) = ps (s). (6)

It bears emphasizing that Definitions 2 and 3 are based on the assumption that service demand solely varies with the energy input e, or alternatively is only a function of service price ps, as is the conventional assumption in the literature. In other words, the possibility that efficiency improvements may not only increase service demand, but also determine other factors, such as the time usage required by an energy service, the use of other commodities, or capital cost, is not considered. In practice, however, more energy efficient appliances frequently have higher fixed costs while simultaneously reducing operating costs through lower fuel and time requirements, a point to which we return in Section 5. 3. Methodology Our empirical methodology proceeds with two principle aims: (1) to compare alternative model specifications that yield estimates corresponding to each of the three definitions of the rebound effect explicated in the theoretical discussion; (2) to generate these estimates using various panel data estimators that control for the omission of potentially relevant factors varying across observations and over time.

150 / The Energy Journal Referring to Definition 1, the first specification regresses the log of monthly kilometers traveled, ln(s), on the log of kilometers traveled per liter, ln(), the coefficient of which yields the rebound effect: (s). As control variables, we additionally include the logged price of fuel per liter, ln(pe), and a set of household- and car-level variables designated by the vector x. Model 1: ln(sit) = a0 + a ln(it) + ape ln( peit) + aTx xit + i + it . (7)

Subscripts i and t are used to denote the observation and time period, respectively. i denotes an unknown individual-specific error term, and it is a random component that varies over individuals and time. The second model generates estimates of the rebound corresponding to Definition 2, which involves regressing ln(s) on the logged price of fuel per kilometer, ln(ps), and the vector of control variables x. In this model, the rebound effect is obtained according to Proposition 2 by the negative coefficient of ln(ps): (s) = ps (s) = aps. Model 2: ln(sit) = a0 + aps ln(psit) + aTx xit + i + it . (8)

Recognizing that ps = pe /, and that ln(ps) = ln(pe) ln(), it can be seen that the specification of Model 2 is functionally equivalent to that of Model 1. In fact, if we impose the restriction H0 : a = ape (9)

on Model 1, we exactly get Model 2. Hence, testing the null hypothesis H0 using Model 1 allows for a simple examination of whether both models are equivalent. Moreover, the anti-symmetry reflected by H0 is intuitive: for constant fuel prices pe, raising the energy efficiency should have the same effect on the service price ps, and hence on the distance traveled, as falling fuel prices pe given a constant energy efficiency . Corresponding to our third definition of the rebound effect, the final specification regresses the logged monthly liters of fuel consumed, ln(e), on ln(pe) and the vector of control variables x. Model 3: ln(eit) = a0 + ape ln(peit) + aTx xit + i + it . (10)

According to Propositions 2 and 3, the rebound effect results from the negative of the price coefficient: (s) = pe (e) = ape. It also bears noting that

Identifying the Rebound: Evidence from a German Household Panel / 151 it is possible to examine whether Model 3 differs from Model 2 by testing the hypothesis H0 : ape = 1 (11)

on the basis of the estimates of Model 3 and, additionally, by testing the hypothesis H0 : aps = 1 (12)

on the basis of the estimates of Model 2. Only if both hypotheses were to hold would Model 2 be identical to Model 3, as can be seen by inserting restriction (11) into equation (10) and by applying restriction (12), efficiency definition (2), = s/e, and the price relationship ps = pe/ to equation (8). Panel data affords three principle approaches for econometric modeling that can each be applied to any of the three models of the rebound effect: the fixed-, between-groups, and random-effects estimators. The key advantage of using the fixed-effects estimator is that it typically includes dummy variables to capture the influence of time-invariant, unobservable factors i, such as topography and urban form, that are potentially correlated with the explanatory variables, thereby producing consistent estimates. In contrast, random effects treats the i as part of the disturbances, thereby assuming that their correlation with the regressors is zero. If this assumption is met, the random-effects estimator is a viable alternative, as it confers the advantage of greater efficiency over the fixed-effects estimator. Violation of the assumption, however, implies biased estimates. While most analyses neglect between-groups effects, instead focusing on the choice between fixed and random effects, we see merit in applying all three estimators to the three model specifications. For starters, our relatively short panel of three years means that some of the regressors may have insufficient variability to be precisely estimated using the fixed-effects estimator, a problem that does not afflict the between-groups effects estimator given its reliance on cross-sectional information. Beyond this, the between-groups effects estimation, which is equivalent to an OLS regression of averages across time, conveys valuable economic content that is not otherwise revealed. Specifically, while fixed effects tell us the impact of an explanatory variable as it inter-temporally changes within subjects, between-groups effects tell us the impacts of cross-sectional changes in an explanatory variable between subjects. Last but equally important, we distinguish between fixed and random effects using a test that, in essence, is based on the comparison of the fixed- and between-groups effects. This test is an alternative to the Hausman test commonly employed to test the null hypothesis that the fixed effects are equal to the random effects, which, if not rejected, would suggest adoption of the random-effects estimator due to its higher efficiency. Yet, testing the hypothesis that the fixed- and the random effects are equal is numerically identical to testing that the between-groups and fixed ef-

152 / The Energy Journal fects are equal see e.g. Baltagi (2005:67) and thus that the inter-temporal within-subject effects are the same as the cross-sectional effects across subjects. As there is rarely a theoretical basis for this assumption, it must not be surprising if the null hypothesis of the Hausman test is not found to withstand empirical scrutiny. Exploiting the equivalence of between-groups and fixed effects under the null, we thus implement modified versions of our models that easily allow us to examine both the equality of the fixed- and between-groups effects for individual variables as well as that of the whole range of coefficients (see Proposition 4 of Appendix B). Chi square tests can then be used to determine for which variables the assumption of equivalence holds and which variables require separate specification of the fixed- and between-groups effects. Using, say, Model 3, testing the null of the standard Hausman test on the basis of the following specification, T ln(eit) = a0 + abe ln (p ei) + awe (ln( pei) ln ( pei) ) + a bx xi + aT wx (xit xi) + i + it . (13)

translates to examining H0 : abe = awe, abx = awx. (14)

Estimated using OLS, specification (13) retrieves the entire set of fixedand between-groups effects estimates, where awe and awx designate the fixed-effects coefficients and abe and abx the between-groups effects coefficients. 4. The German Mobility Panel Data Set The data used in this research is drawn from the German Mobility Panel (MOP 2007), an ongoing travel survey that was initiated in 1994. The panel is organized in overlapping waves, each comprising a group of households surveyed for a period of one week in autumn for three consecutive years. All households that participate in the survey are requested to fill out a questionnaire eliciting general household information, person-related characteristics, and relevant aspects of everyday travel behavior. In addition to this general survey, the MOP includes another survey focusing specifically on vehicle travel among a sub-sample of randomly selected car-owning households. This survey takes place over a roughly six-week period in the spring, during which time respondents record the price paid for fuel, the liters of fuel consumed, and the kilometers driven with each visit to a gas station and for every car in the household. The data used in this paper cover nine years of the survey, spanning 1997 through 2005, a period during which real fuel prices rose 3.2% per annum on average. To avoid complications of multiple car ownership due to substitution effects among cars, we focus on single-car households, which comprise roughly 53% of the households in Germany (MiD 2007). The resulting sample includes 547 house-

Identifying the Rebound: Evidence from a German Household Panel / 153 holds, 254 of which appear two consecutive years in the data and 293 of which appear three consecutive years. To correct for the non-independence of repeated observations from the same households over the years of the survey, the regression disturbance terms are clustered at the level of the household, and the presented measures of statistical significance are robust to this survey design feature. We used the travel survey information, which is recorded at the level of the automobile, to derive the dependent and explanatory variables required for estimating each of the three variants of the rebound effect. The two dependent variables, which are converted into monthly figures to adjust for minor variations in the survey duration, are the total monthly distance driven in kilometers (Definitions 1 and 2) and the total monthly liters of fuel consumed (Definition 3). The three explanatory variables for identifying the direct rebound effect are the kilometers traveled per liter (Definition 1), the price paid for fuel per kilometer traveled (Definition 2), and the price paid for fuel per liter (Definition 3).2 Table 1. Variable Definitions and Descriptive Statistics
Variable Name s e ps pe car age diesel car premium car household size # high school diploma # employed car vacation children job change Variable Definition Monthly kilometers driven Monthly fuel consumption in liters Kilometers driven per liter Real fuel price in Euros per kilometer Real fuel price in Euros per liter Age of the car Dummy: 1 if fuel type is diesel Dummy: 1 if car is a sports- or luxury model Number of household members Number of household members with a high school diploma Number of employed household members Dummy: 1 if household undertook car vacation during the survey period Dummy: 1 if children younger than 12 live in household Dummy: 1 if an employed household member changed jobs within the preceding year Mean 1156.82 94.82 12.51 0.08 0.96 6.13 0.10 0.21 1.98 0.48 0.71 0.24 0.13 0.10 Std. Dev. 714.55 59.33 2.76 0.02 0.13 4.04 0.30 0.40 1.04 0.65 0.74 0.43 0.34 0.30

The remaining suite of variables selected for inclusion in the model measure the socio-demographic and automobile attributes that are hypothesized to influence the extent of motorized travel. Table 1 contains the definitions and de2. The price series was deflated using a consumer price index for Germany obtained from Destatis (2007).

154 / The Energy Journal scriptive statistics of all the variables used in the modeling. To control for the effects of quality (Wirl 1997:14), the age of the automobile and a dummy indicating luxury models is included.3 Although income is not directly measured, an attempt is made to proxy for its influence via measures of the number of employed residents and the number with a high school diploma living in the household. Finally, controls are included for household size, the presence of children, whether the household undertook a vacation with the car during the survey period, and whether any employed member of the household changed jobs in the preceding year. 5. Empirical Results Our empirical analysis involved the estimation of two sets of models, one in which the individual-specific component was specified at the level of the household and one in which it was specified at the level of the automobile. Noting that this distinction had little bearing on the qualitative conclusions of the analysis, the following discussion focuses on the estimates generated at the household level. This focus facilitates the comparison of the three estimators, as it ensures that each uses the same sample of observations. Were the individual component set at the level of the automobile, then observations in which the household changes automobiles from one year to the next would drop out in the case of the fixedeffects estimator.4 Table 2 presents estimates of the rebound effect corresponding to Model 1, using fixed, between-groups, and random-effects estimators. Several features of the results bear highlighting. First, we confirm that the impact of efficiency improvements on traveled distance is of the same order as the effect of fuel prices: As reported in the penultimate row of the table, upon testing the null hypothesis H0: a = ape, we cannot reject the anti-symmetry given by H0 for any of the estimation techniques. Hence, there is no reason, neither on a theoretical nor an empirical basis, to assume that Model 1 and 2 are principally different, implying the conclusion that it is equally well-founded to estimate the direct rebound effect on the basis of either Definition (Model) 1 or Definition (Model) 2. Second, the estimated rebound effects are considerably higher than most estimates reported elsewhere in the literature, and suggest that some 58% of the potential energy savings due to an efficiency improvement is lost to increased driving. Finally, these effects are of a strikingly similar magnitude across the three estimators, differing by less than a percentage point. This similarity does not hold for many of the remaining coefficients. A particularly stark difference is seen for the effect of the number of employed household members, which has a counterintuitive and negative coefficient in the fixed-effects model, but is positive in the between-group and random-effects models. All else equal, we would expect that a greater number of employed persons
3. We also tested for quality-dependent rebound effects by interacting the efficiency measure with the luxury dummy, but found this to be insignificant in all of the specifications. 4. Roughly 18% of households changed cars at least once over the three years of the survey.

Identifying the Rebound: Evidence from a German Household Panel / 155 Table 2. Estimation Results for Model 1 and the Rebound Effect based on Definition 1
Fixed-Effects Estimator Between-Group Estimator Random-Effects Chi Square Estimator Test Std. 2 (1) Errors Statistics (0.105) (0.127) (0.005) (0.090) (0.052) (0.025) (0.030) (0.029) (0.031) (0.067) (0.055) (0.219) 0.01 0.02 0.88 2.86 0.47 0.05 2.82 ** 33.8 3.16 0.75 0.03

ln(s) Coeff.s ln() ln( pe) car age diesel car premium car household size # employed vacation with car children job change constants H0 : a = ape

Std, Errors Coeff.s

Std. Errors Coeff.s (0.139) ** 0.584 (0.211) ** -0.588 (0.006) ** -0.018 (0.102) (0.062) (0.031) (0.036) (0.037) (0.070) (0.091) (0.089) (0.284) 0.014 ** 0.220 * 0.040 * 0.070 ** 0.092 ** 0.300 0.036 ** 0.145 ** 7.854
2

** 0.585 (0.122) ** 0.575 **-0.622 (0.164) ** -0.595 -0.011 (0.006) ** -0.020 -0.232 0.140 (0.182) 0.120 0.003 (0.147) ** 0.252

-0.013 (0.050)

# high school diploma -0.012 (0.055) ** 0.095 ** -0.135 (0.046) ** 0.208 ** 0.275 (0.037) -0.064 0.113 (0.118) (0.071) ** 0.411 0.062 0.127 ** 7.768

F(1, 545) = 0.03

F(1, 545) = 0.01 2 (11) = 55.19 **

(1) = 0.02

Standard Hausman Test

Note: * denotes significance at the 5 %-level and ** at the 1 %-level, respectively. Number of observations used for the estimations: 1,375. Number of households: 546.

in the household would increase the dependency on the automobile. The negative fixed-effects estimate is difficult to interpret, but may be the result of temporary disruptions associated with changes in the household labor force that reduce automobile travel. The remaining control variables have either intuitive effects or are statistically insignificant. Referencing the random-effects coefficients, older cars are seen to be driven less, while premium cars are driven more. Another important determinant is whether a vacation with the car was undertaken over the survey period, which results in a roughly 35% (= exp(0.30) - 1) increase in distance traveled. Aside from the fuel price and the number of employed household members, this is the only control variable also found to be significant in the fixed-effects model. We also explored models in which time dummies were included to control for autonomous changes in the macroeconomic environment. As these were found to be jointly insignificant across all of the models estimated, they were excluded from the final specifications. Not unexpectedly, the Hausman test rejects the null hypothesis that the fixed- and random-effects coefficients are jointly equal for all significance levels. Whether this result therefore implies that equality fails to hold for each of the vari-

156 / The Energy Journal Table 3. Estimation Results for Model 2 and the Rebound Effect based on Definition 2
Fixed-Effects Estimator Between-Group Estimator Random-Effects Chi Square Estimator Test Std. 2 (1) Errors Statistics (0.084) (0.005) (0.089) (0.051) (0.025) (0.030) (0.029) (0.032) (0.066) (0.055) (0.174) 0.02 0.87 2.87 0.48 0.05 2.81 ** 33.8 3.20 0.74 0.03

ln(s) Coeff.s ln( pe) car age diesel car premium car household size # employed vacation with car children job change constants H0 : aps = 1

Std, Errors Coeff.s

Std. Errors Coeff.s (0.122) ** -0.585 (0.006) ** -0.019 (0.098) (0.061) (0.031) (0.036) (0.037) (0.070) (0.091) (0.088) (0.249) 0.017 ** 0.220 * 0.033 * 0.072 ** 0.082 ** 0.306 0.046 ** 0.144 ** 7.867

** -0.596 (0.099) ** -0.581 -0.011 (0.006) ** -0.020 -0.226 (0.180) 0.140 0.013 (0.050) 0.121 0.003 (0.147) ** 0.252

# high school diploma -0.012 (0.055) ** 0.095 ** -0.135 (0.046) ** 0.208 ** 0.274 (0.037) -0.063 0.113 (0.117) (0.071) ** 0.411 0.062 0.126 ** 7.779

F(1, 545) = 16.75 ** F(1, 545) = 11.88 ** 2 (11) = 55.31 **

2 (1) = 24.37 **

Standard Hausman Test

Note: * denotes significance at the 5 %-level and ** at the 1 %-level, respectively. Number of observations used for the estimations: 1,375. Number of households: 546.

ables individually is, however, not immediately clear. To pursue this issue further, we estimated the model in equation (13) and proceeded to test the equality restrictions (14) using individual chi-square tests, the results for which are presented in the final column of Table 2. These findings confirm what was already evident from casual inspection: the difference between the fixed- and between-groups effects estimates of the rebound effect are statistically insignificant. In fact, this conclusion applies to several of the other explanatory variables, with the one clear exception being the number of employed people in the household. Table 3 presents estimates of the rebound effect corresponding to Definition 2, which employs the price of fuel per kilometer as the key explanatory variable. As expected from the results of the anti-symmetry test reported in Table 1, the overall pattern is similar to that of Table 2. Again, the estimated rebound effects are high, roughly on the order of 59%. The remaining coefficient estimates are also similar to the first specification. The Hausman test rejects equality of the fixed- and random-effects models for all significance levels. The only variable for which differences are clearly evident is again the number of employed household members. Table 4 presents estimates of Model 3, which is distinguished by the use of total fuel consumption as the dependent variable and the price of fuel per liter

Identifying the Rebound: Evidence from a German Household Panel / 157 Table 4. Estimation Results for Model 3 and the Rebound Effect based on Definition 3
Fixed-Effects Estimator Between-Group Estimator Random-Effects Chi Square Estimator Test Std. 2 (1) Errors Statistics (0.128) (0.005) (0.091) (0.050) (0.024) (0.029) (0.029) (0.032) (0.066) (0.055) (0.063) 0.15 0.84 1.89 1.30 0.16 1.59 ** 33.4 * 4.19 0.25 0.01

ln(e) Coeff.s ln( pe) car age diesel car premium car household size # high school diploma # employed vacation with car children job change constants H0 : ape = 1 ** -0.569 -0.297 0.149

Std, Errors Coeff.s (0.165) ** -0.671 (0.195) 0.001 0.027

Std. Errors Coeff.s (0.211) ** -0.594 (0.006) ** -0.018 (0.094) (0.057) (0.031) (0.036) (0.036) (0.070) (0.092) (0.089) (0.073) -0.076 ** 0.258 * 0.045 * 0.065 ** 0.093 ** 0.289 0.055 ** 0.137 ** 4.098

-0.010 (0.007) ** -0.019 (0.149) ** 0.330

0.002 (0.048)

0.002 (0.055) ** 0.081 ** 0.254 (0.036) ** 0.412 - 0.020 (0.118) 0.053 0.103 0.116 (0.070)

** -0.117 (0.044) ** 0.219

** 4.003

F(1, 545) = 6.82 ** F(1, 545) = 2.44 ** 2 (11) = 57.57 **

2 (1) = 10.09 **

Standard Hausman Test

Note: * denotes significance at the 5 %-level and ** at the 1 %-level, respectively. Number of observations used for the estimations: 1,375. Number of households: 546.

as the key regressor. That this model is not identical to Models 1 or 2 is confirmed by the rejection of the null hypotheses that the price coefficients in Models 2 and 3 are both -1. The estimates reported in Table 4 are remarkably similar to those of Tables 2 and 3, albeit with a larger range across the fixed- and between-groups effects estimators. In this instance, the estimated rebound effect is seen to vary between 57% and 67%; but even here we cannot reject the hypothesis that the coefficients are equal based on a chi square test. Likewise, with the exception of the number of employed household members and the dummy indicating a car vacation, the other coefficients also appear to be equal despite the rejection of the Hausman test for all significance levels. We thus conclude that although our estimates of the rebound effect are high, they appear to be robust to both the estimator and the specification. Whether the model controls for time-invariant factors that vary across cases as with the fixed-effects estimator or case-invariant factors that vary over time as with the between-groups effects estimator has no substantial impact on the key results. Perhaps even more notable is the similarity of the estimates corresponding to Definition 3 with those of Definitions 1 and 2. While the latter definitions incorporate efficiency either directly via the kilometers per liter traveled or indirectly

158 / The Energy Journal via the service price per kilometer, Definition 3 relies exclusively on the price mechanism, suggesting that this information can serve as a useful substitute in the absence of data on technology.5 A few caveats should be recognized, perhaps the strongest of which is the assumption that automobile efficiency is exogenous. If, for example, individuals who drive more also select more fuel-efficient vehicles, then we might expect an upward bias imparted on the rebound effect estimated by the coefficients of ln() and ln(ps). However, there are two reasons why we do not deem endogeneity to be a serious concern here. First, any time-invariant unobservable factors that would otherwise induce correlation between the rebound effect and the error term (e.g. proximity to public transit, environmental attitudes) will be captured by the fixed-effects model. Although we cannot exclude the possibility of relevant time-variant unobservables, we believe the range of included explanatory variables the presence of young children, job changes, and the number of employed provides reasonably good coverage of temporal changes whose absence could induce biases. Second, endogeneity problems relating to vehicle choice would not be expected to afflict the estimates from Definition 3, as these are based on the price of fuel. The fact that these estimates are of roughly the same magnitude as those from Definitions 1 and 2 provides some confirmation that any upward bias from endogeneity is negligible. An additional caveat arises from the analysis' neglect of changes in capital costs due to efficiency improvements in automotive technology. Specifically, more costly equipment reduces disposable income, thereby triggering an income effect that mitigates the rebound. Several authors, such as Henley et al. (1988), therefore argue that neglecting capital cost would lead to an overestimation of the rebound when relying on definitions of the direct rebound effect, such as Definitions 1, 2, and 3. On the other hand, the correlation of energy and time efficiency may lead to an underestimation of the rebound: If more efficient cars also reduce transport time, the increased time for leisure, for example, may imply a higher energy consumption due to energy-intensive leisure activities. As the mobility data analyzed here includes no information on annualized capital cost ki, nor time requirements, nor the consumption of other goods oi, we are not able to pursue these issues. To date, only a few empirical studies, such as Brnnlund, Ghalwash, Nordstrm (2007), have availed data that allows for explicit consideration of income effects on the consumption of goods other than fuel, thereby enabling estimation of the net rebound effect, including the indirect effect. This is clearly an area warranting further study.

5. To the extent that data on liters of fuel consumed and unit fuel prices is easier to collect, our results indicate that it is equally well-founded to estimate the rebound effect on this empirical basis, as is stated by Proposition 3, rather than with data on fuel efficiency and traveled distances.

Identifying the Rebound: Evidence from a German Household Panel / 159 6. Summary and Conclusion Industrialized countries are increasingly struggling both to ensure their security of energy supply and to reduce emissions of greenhouse gases. It is commonly asserted that efficiency-increasing technological innovations, particularly in the transport sector, are an important pillar in this process. This assertion underpins the CAFE standards in the United States as well as the legislation proposed by the European Union in December 2007, stipulating the reduction of average emissions in the new car fleet. Although increased efficiency confers economic benefits in its own right, its effectiveness in reducing fuel consumption and pollution critically depends on how consumers alter behavior in response to cheaper energy services due to improved efficiency. To the extent that service demand increases via rebound effects, gains in reducing environmental impacts and energy dependency will be offset. The results presented in this paper, based on the analysis of a German household panel, suggest that the size of this offset is potentially quite large, varying between 57% and 67%. While these estimates are considerably higher than those found elsewhere in the literature, with most empirical evidence originating from the U.S., our results are robust to both alternative panel estimators and to alternative measures of the rebound effect. Moreover, our results are consistent with recent anecdotal evidence from Germany. Between 2004 and 2005, fuel prices increased by 5% while average road mileage decreased by 3% (MWV 2007), suggesting a sizeable price elasticity of -0.6. One possible explanation for the discrepancy between the estimates in this study and those from the U.S. may be superior access to public transport in Germany. Related to this, longer trip distances in the U.S., particularly for commuting (Stutzer, Frey 2007), likely decrease the responsiveness of mode choice to changes in automobile efficiency. As this is one of the few studies conducted on this issue in a European context, it would be of interest to see whether the findings presented here are corroborated by studies using other data sets from within Germany and other European countries. If this is found to be the case, it would suggest that policy interventions targeted at technological efficiency be they voluntary agreements or command and control measures may have only muted effects in reducing fuel consumption. At the very least, our results indicate that the current emphasis on efficiency as the principal means for policy-makers to address environmental challenges may be misplaced. Given the strong responses to prices found here, price-based instruments such as fuel taxes would appear to be a more effective policy measure.

160 / The Energy Journal References

Baltagi, B.H. (2005). Econometric Analysis of Panel Data. Third edition, John Wiley & Sons, Ltd. Becker, G.S. (1965). A Theory of the Allocation of Time. Economic Journal 75, 493-517. Berkhout, P.H.G., Muskens, J.C., Velthuisjen, J.W. (2000). Defining the Rebound Effect. Energy Policy 28, 425-432. Binswanger, M. (2001). Technological Progress and Sustainable Development: What About the Rebound Effect? Ecological Economics 36, 119-132. Brnnlund, R., Ghalwash, T., Nordstrm, J. (2007). Increased Energy Efficiency and the Rebound Effect: Effects on Consumption and Emissions. Energy Economics 29, 1-17. Brookes, L. (2000). Energy Efficiency Fallacies Revisited. Energy Policy 28, 355-367. Brookes, L.G. (1990). The Greenhouse Effect: The Fallacies in Energy Efficiency Solution. Energy Policy 18 (2), 199-201. Destatis (Federal Statistical Office Germany) (2007). http://www.destatis.de/ Goodwin, P., Dargay, J., Hanly, M. (2004). Elasticities of Road Traffic and Fuel Consumption with Respect to Price and Income: A Review. Transport Reviews 24 (3), 275-292. Greene, D.L., Kahn, J.R., Gibson, R.C. (1999). Fuel Economy Rebound Effect for U.S. Household Vehicles. The Energy Journal 20 (3), 1-31. Greene, D.L. (1992). Vehicle Use and Fuel Economy: How big is the Rebound Effect? The Energy Journal 13 (1), 117-143. Greening, L.A., Greene, D.L. (1997). Energy Use, Technical Efficiency, and the Rebound Effect: A Review of the Literature. Report to the Office of Policy Analysis and International Affairs, US Department of Energy, Washington DC. Greening, L.A., Greene, D.L., Difiglio, C. (2000). Energy Efficiency and Consumption the Rebound Effect A Survey. Energy Policy 28 (6-7), 389-401. Henley, J., Ruderman, H., Levine, M.D. (1988). Energy Saving Resulting from the Adoption of More Efficient Appliances: A Follow-Up. The Energy Journal 9 (2), 163-170. Khazzoom, D.J. (1980). Economic Implications of Mandated Efficiency Standards for Household Appliances. The Energy Journal 1, 21-40. Lovins, A.B. (1988). Energy Saving Resulting from the Adoption of More Efficient Appliances: Another View. The Energy Journal 9 (2), 155-162. MiD (Mobility in Germany) (2007). German Federal Ministry of Transport, Building and Urban Affairs. http://www.kontiv2002.de/publikationen.htm MOP (German Mobility Panel) (2007). http://www.ifv.uni-karlsruhe.de/MOP.html MWV (Association of the German Petroleum Industry) (2007). http://www.mwv.de Saunders, H.D. (1992). The Khazzoom-Brookes Postulate and Neoclassical Growth. The Energy Journal 13 (4), 131-148. Small, K.A., Van Dender, K. (2007). Fuel Efficiency and Motor Vehicle Travel: The Declining Rebound Effect, The Energy Journal 28 (1), 25-52. Schipper, L., Grubb, M. (2000). On the Rebound? Feedback between energy intensities and energy uses in IEA countries. Energy Policy 28, 367-388. Sorrell, S. (2007). The rebound effect: An assessment of the evidence for economy-wide energy savings from improved energy efficiency. UK Energy Research Centre (UKERC). Review of evidence for the rebound effect. www.ukerc.ac.uk. Sorrell, S., Dimitroupoulos, J. (2008). The Rebound Effect: Microeconomic Definitions, Limitations, and Extensions. Ecological Economics 65 (3), 636-649. Stutzer, A., Frey, B.S. (2007). Commuting and life satisfaction in Germany. Informationen zur Raumentwicklung Heft 2/3, 179-189. Walker, I.O., Wirl, F. (1993). Asymmetric Energy Demand due to Endogenous efficiencies: An empirical investigation of the transport sector. The Energy Journal 14, 183-205. West, S.E. (2004). Distributional Effects of Alternative Vehicle Pollution Control Policies. Journal of Public Economics 88 (3-4), 735-757. Wirl, F. (1997). The Economics of Conservation Programs. Kluwer Academic Publishers, London.

Identifying the Rebound: Evidence from a German Household Panel / 161 Appendix A: Proofs Proof of Proposition 1: Employing efficiency definition (2) and taking logarithms, we obtain ln e = ln s ln . Logarithmic differentiation with respect to yields the claim: lne lns ln lns (e) = = = 1 = (s) 1, ln ln ln ln thereby exploiting the fact that, for a desired amount of service, the input e of energy is, ultimately, a function of the parameter , making it meaningful to take the derivative of s = f(e(), t, k, o) with respect to . Proof of Proposition 2: Using price relationship ps = pe/, the chain rule, and the explicit assumption that the service amount s solely depends on the price ps, we obtain lns lns lnps ln( pe/) (s) = = = ps(s) = ln lnps ln ln lnpe ln lnpe (s) = ps(s) = ps(s) 1 . ln ln ln If energy prices pe are exogenous, it follows that lnpe/ln = 0, so that (s) equals ps (s). Proof of Proposition 3: Using the chain rule and the definition (2) of energy efficiency, = s/e, as well as price relationship ps = pe/, we obtain lne lne lnps ln(s/) ln( pe/) pe(e) = = = lnpe lnps lnpe lnps lnpe ln (s) = ps(s) lnps

ln 1 . lnpe

Hence, only if ln/lnps = 0 and ln/lnpe = 0, which holds true if the energy efficiency is constant, both elasticities are equal: pe (e) = ps (s).

162 / The Energy Journal Appendix B. Hausman Test for Individual Variables The following Proposition 4 is a generalization of Baltagis (2005:76) exercise formulated in a bivariate context and allows for examining the null hypothesis of the classical Hausman test that is typically employed for the distinction of fixed versus random effects. The null of this test is equivalent to the hypothesis that the set of fixed-effects coefficients w equals the set of between-group effects coefficients b (Baltagi 2005:67): H0 : w = b. In addition to testing H0, estimating representation (13) via OLS also allows for examining the equality of the fixed- and between-group effects for individual variables. The basis for these tests is given by the following proposition. Proposition 4: Departing from a standard panel data model, yit = 0 + Txit + i + it, i = 1,..., N, t = 1,...,T, and estimating the regression model yit = 0 + Tw (xit xi ) + Tb xi + i + it . (16) (15)

via OLS yields at once estimates of the between-group and fixed effects, where the OLS estimates of w provides for the fixed-effects estimates and the betweengroup effects estimates is given by the OLS estimates of b. Proof of Proposition 4: First, the between-group effects estimator of parameter vector emerging from model (15) can be obtained by averaging (15) over time and estimating the result via OLS: yi = 0 + T xi + i + i. (17)

Second, the fixed-effects estimates of are to be retrieved by subtracting (17) from (15) and estimating the result via OLS: yit yi = T (xit xi ) i + it. (18)

Third, instead of estimating either (17) or (18), we alternatively suggest in Proposition 4 estimating (16) via OLS in order to at once get both the fixed- and between-group effects estimates of . This can be seen as follows: Upon averaging (16) over time, the term related to w wipes out so that the result is, aside from the notation of the parameter vector, identical to (17): yi = 0 + Tw 0 + Tb xi + i + i. (19)

Identifying the Rebound: Evidence from a German Household Panel / 163 Hence, whether averaging either (15) or (16) over time and estimating the results via OLS must yield the same estimates, namely the between-group effects estimates of . Next, inserting (19) into (16) yields (18) with w instead of as parameter vector. In other words, either demeaning (15) or (16) and estimating the result via OLS provides for the same estimates of , namely the fixed-effects estimates.

164 / The Energy Journal

Book Reviews

Electricity and Energy Policy in Britain, France and the United States since 1945, by Martin Chick. (Cheltenham and Northampton: Edward Elgar,
2007) 205 pages. Hardcover, ISBN 978-1-84542-111-3. A glaring disparity prevails between what the books title promises and the books size. The author obviously chose to be selective but still deals with topics that each is broader than any one short book can handle. Indeed, all his five topics have been the subject of multiple separate books. He covers the displacement of coal by oil, the European Coal and Steel Community, electricity pricing, electricity investment, and electricity restructuring. A further limitation is that except in the last case, he stresses coverage of the years for which the British and French thirty-year archive-disclosure restrictions have ended. Curiously, the thirty-year limit is more strictly applied to the readily updatable realm of data than to the actual developments. The effort to cover so much in so little space unsurprisingly produces breathlessness that makes the book of limited interest. Choppy organization aggravates the problems. In each chapter, he jumps too rapidly among topics. Too many inadequate treatments of details arise. The book starts with an introduction stressing an overview of the electric power industries in the three countries. He gets off to a bad start by not making clear that the U.S. organization was well established well before 1945 in contrast to Britain and France, which were about to nationalize. His overall take on the most complicated of the cases, the United States, misses too much. In particular, his discussion of municipals fails to note that most are distribution companies buying from others, particularly various entities of higher levels of government such as the Tennessee Valley Authority. The existence of cooperatives is ignored. Chapter 2 Moving from Coal to Oil: What Price Security?starts with review of the rapid substitution of oil for coal in the three countries, turns to the rise of Middle Eastern oil, and then summarizes the policy response in the three countries. On the last, he sees France as more willing to accept imported energy, perhaps because it had less domestic energy demanding protection. The bulk of the chapter deals with security. The discussion begins and ends with recognition that supply-disruption dangers were used as an excuse for protectionism. In between, he leafs through numerous supply crises and policy initiatives in the three countries. Chapter 3 deals with the European Coal and Steel Community (ECSC). It starts off with a clear discussion of the French motives for promoting the Community and the British reasons for resistance. The characteristic jumpiness then returns. He moves among the later initial rejected British effort to join ECSC and

165

166 / The Energy Journal the two newer Communities and the inability of the Communities to agree on an energy policy in the face of the vigorous competitive pressures on ECSC coal industries. He never gets to eventual British membership or the death of ECSC. Chapter 4 deals with marginal-cost pricing of electricity. It starts with review of the development of the germane theories, moves to the contrast between French embrace and British rejection of the ideas, and then reviews U.S. developments. Chapter 5 turns to investment decisions in the electric-power sectors of the three countries. The fullest attention is given to British efforts. Discussions of the other two countries are strewn about the chapter. Chapter 6 contrasts among the three countries in their approach since 1978 to electric-industry organization. As is amply discussed elsewhere, Britain privatized, some U.S. states restructured, and France was unchanged. The U.S. case gets by far the most attention, but the results are problematic. He relates the familiar story of how a provision of the 1978 U.S. Public Utility Regulatory Policies Act (PURPA) encourage electric companies to purchase electricity from specified types of independently-owned facilities led to concerns about electricutility industry organization. He moves on to a sketch of how some states responded by imposing reorganization plans on the states utilities and notes federal initiatives. He ends with review of how the California fiasco discouraged further changes. While he gets the broad picture right, he stumbles on the details. An overriding problem is failure to deal with the efficacy of decentralized solutions. He seems to consider the decentralized development of the U.S. grid as inherently inferior to the fully centralized U.K. and French systems and barely aware of the U.S. reliability council approach to coordination. His treatment of PURPA implementation simultaneously stresses efforts to encouraging environmentally friendly technologies and the actual dominance of cogeneration. Worse, he talks of use of small-scale conventional plants instead of the gas turbines either as separate units or combined with boilers (combined cycle) that actually dominated. (Since combined cycles are discussed later in the book, the problem is inadequate editing rather than ignorance.) He repeatedly assigns a further key role to the 1992 Energy Policy Act. By the time, he gets to noting the key provisions of that Act he botches the description of one; he talks of encouraging independent generators without mentioning that the act was focused on permitting existing utilities to act as independent generators in other regions. His treatment of U.K. privatization appears as two interruptions of his review of the U.S. The first comes between an initial treatment of PURPA and review of the longer-term problems. Additional material appears as discussion of new institutions that U.S. restructuring necessitated. The French situation appears a discrete conclusion to the chapter. His final chapter is not a clear summing up but a mlange of further thoughts. He does start with reiteration of the major conclusions of the prior chapter but quickly moves on to other points, many extensions of prior chapters. Thus, he discusses the relevant French traditions in public-utility economics, the effects

Book Reviews / 167 of British privatization, and further aspects of supply security. With the last, he expands to encompass depletion, disruption, adequacy of electricity investment, and global warming. The result then is a collection of largely unsatisfactory recapitulations of part of a vast literature. Curiously, he does provide an extensive and generally well selected bibliography. Richard L. Gordon The Pennsylvania State University ***

The Strategic Petroleum Reserve: U. S. Energy Security and Oil Politics, 1975-2005, by Bruce A. Beaubouef. (College Station, Texas: Texas A & M
Press, 2007) Hardcover, ISBN 978-1-58544-600-1, $49.95. Of the major policy responses to the first oil shock of 1973-74, the only one still in existence is the Strategic Petroleum Reserve. Unlike other consuming countries, which imposed storage requirements on industry, the U.S. instead chose to maintain supply reserves exclusively in publicly owned and funded storage facilities (underground hollowed-out salt domes). This timely book supplies a definitive chronological history of the reserve from early discussions in white papers in the Nixon administration to the present. The early years were filled with extreme political pressures to fill the reserve and a series of technical and management setbacks that prevented the political goals from being realized. The Iran-Iraq war occurred too soon for the reserve to be used. In fact, filling the reserve was suspended in 1979 to ease pressure on world markets and prices. While Beaubouef supplies all the details of interest to energy professionals, he also demonstrates how the reserve becomes part of the political struggles of each era. Thus in the 1980s when deficits were a source of political dispute and spending for the reserve was the largest component of the Department of Energy budget, the spending for the reserve was put off budget and borrowed without explicit on-budget appropriations. When spending constraints were in vogue under rules created by Senators Gramm, Rudman, and Hollings, the Reserve was placed back on budget. The first real test of the Reserve occurred after the Iraqi invasion of Kuwait and the subsequent brief war in 1991. President Bush ordered a drawdown of the SPR on the same day he ordered the invasion of Iraq. Over the next two days, the price of oil dropped immediately from $32 to around $19. Because the war was short and successful, isolating the effect of the reserve announcement isnt possible scientifically. But many economists are sceptical of any effect. Beaubouef quotes Phil Verleger as saying that that the SPR release announcement

168 / The Energy Journal had nothing to do with the price decrease, but for the public and politicians, the SPR had done its job. The Energy Policy Conservation Act of 1990 amendments gave the President more explicit authority to use the SPR, and both the Congress and the President took advantage. They proposed oil sales to raise revenues and lower the deficit. In 1999 Congress wanted to refill what was sold in 1996. The explicit goal was oil security, but their implicit goal was to raise oil prices (remember when low prices were a problem). They wanted to assist oil producers without explicitly raising taxes so they changed the royalty payments from offshore wells from cash to in-kind oil contributions to the SPR. Finally, in the heat of the close 2000 Presidential election, Vice President Gore explicitly called for the use of the SPR to lower oil prices even though earlier in the same year, he had opposed such use. President Clinton ordered the exchange of 30 million barrels in an attempt to lower heating oil prices. The SPR had been transformed from an emergency device for use during war to a politicized price smoother. The author is a historian and journalist and not an economist. Thus the reader will not find an extensive welfare-economic analysis of the SPR. Instead he repeats the conventional view that the economic costs associated with oil supply disruptions are large and that public stockpiling is necessary policy response. For an analysis of the generic problem of optimal inventories and the experience with government stockpiles of other (strategic) commodities, the reader will have to look elsewhere. Peter Van Doren Cato Institute ***

A Handbook of Primary Commodities in the Global Economy, by Marian Radetzki (Cambridge University Press: 2008), 233 pages, ISBN 978-0-521-88020-6 hardback.
A handbook often signifies a book of guidance. In reading the book of widely esteemed Professor of Economics Marian Radetzki, from the Lulea University of Technology in Sweden, readers are guided through the turbulent global markets of primary commodities. The current global commodity boom has become a source of interest, and concern, more than any boom in history. It is therefore hard to think of a better time for this sort of work to emerge. Other books have been written to address some of the issues. I would say that some are lacking in substance and seem to have been written by authors hoping to somehow benefit suddenly from the turmoil. By contrast, Radetzkis book is based on four decades of research and experience. In fact, the antecedent to this new book is entitled A Guide to Primary

Book Reviews / 169 Commodities in the World Economy (Radetzki, 1990). The topics from the older book have been completely rewritten, with several new topics covered. A wide audience with interest in mineral and energy issues is the audience for this book which includes students, executives, government officials, and members of the general public who are concerned about high energy prices and the depletion of exhaustible resources. He is successful in writing to this readership, thanks in part to the clear and concise nature of the exposition. Despite some of the complexities of commodity markets, he effectively communicates his findings by kindly avoiding the use of overly technical language. Chapter 1, the historical framework, deals with the important role of primary commodities in economic development, the decline of transport costs that has dramatically increased world trade, and the history of resource nationalism (which has started to resurge this decade). The chapter is concise, which is pleasant considering there are already many books adequately covering the history of primary commodities. Chapter 2, which covers the geography of commodity production and trade, efficiently highlights major geographical issues. Chapter 3, comparative advantage and trade policy distortions, discusses public policies used by nations in commodity production and trade. The reasons for the establishment of such policies are also explored. There is a useful section in the chapter that provides numerical measures of trade restriction in global trade. Readers will find that the final section on the effects of tariff escalation in commodity processing is interesting and well-explained. Chapter 4 explains price formation and price trends in commodities with a practical and basic lesson in microeconomics. An area related to prices that falls outside of traditional economic fundamentals and has received vast attention this decade, is the role of commodity exchanges, investments, and speculation. Chapter 5 does a good job of explaining these complex subjects and their effects on commodity markets. Chapter 6, the economics of exhaustible resource depletion, provides persuasive arguments countering the fears of many concerned experts who fear depletion is near. As the name of the chapter implies, depletion is discussed from an economic rather than physical point of view. Therefore, in spite of the appropriateness of dealing with depletion economically, the argument may do little to relieve the pessimists worries. Chapter 7, fears of and measures to assure supply security, provides an informative review of the consequences of supply disruptions and the response measures designed to mitigate the negative effects. The chapter leaves us thinking that the response measures, although costly and not often used, resemble an insurance policy that is probably worth having. Chapter 8, producer cartels in international commodity markets, discusses the principles of cartels and why they have usually failed. OPEC is the exception, with evidence provided that the cartel has been mostly successful for over thirty years. The assertion of OPECs effectiveness is different from most experts views, but I believe Radetzkis arguments will leave the reader convinced.

170 / The Energy Journal Chapter 9, public ownership in primary commodity production, discusses the presently relevant topic of national oil companies and state mining enterprises. The chapter correctly suggests that in general these corporations are not well run. Still, there are a number of exceptions in the energy and metal industries, so it is nice that Radetzki does not completely discount public ownership. Chapter 10, the monoeconomies (issues raised by heavy dependence on commodity production and exports), discusses the important mineral/energy economic topics of commodity instability, generation of public revenue, the Dutch disease, the resource curse, and exchange rate policies. Diversification is the usual solution offered by many to the problems raised by heavy dependence on a few commodities. However, Radetzki provides several arguments in concluding that a country like Venezuela, for instance, which is heavily dependent on the export of essentially one commodity, may not benefit by reducing this dependence and increasing the contribution of the manufacturing industry. In writing this book, Radetzki provides several valuable contributions related to primary commodities in international markets. To reiterate, the book comes at an opportune point in time, provides a wealth of information, and is useful as a guide to really understanding the themes addressed. REFERENCES
Radetzki, M. (1990), A Guide to Primary Commodities in the World Economy, Blackwell, Oxford.

Roberto F. Aguilera Mining Center, School of Engineering Pontificia Universidad Catolica de Chile

REFEREE ACKNOWLEDGMENTS
All papers received and published by The Energy Journal are sent to external referees for anonymous peer review. Without the help of hundreds of expert referees, the quality and standards of the Journal could not be maintained. Below is a list of referees who have volunteered their time and talents to appraise submissions in the past year. The editors sincerely thank all referees for their efforts. Referees included as of June 1, 2008. Philip Adams Centre of Policy Studies, Monash University Seabron Adamson CRA International, Cambridge MA Bahram Adrangi University of Portland Paolo Agnolucci Policy Studies Institute Per Agrell Catholic University of Louvain Erik Ahlgren Dept of Energy and Environment, Chalmers University Jim Airola Naval Postgraduate School Viren Ajodhia University of Delft Vicent Alcantara Autonomous University of Barcelona Joseph Aldy Resources for the Future A.F. Alhajji Ohio Northern University Stefan Ambec University of Toulouse Glen Ames University of Georgia Kathryn Anderson Vanderbilt University Dennis Anderson Imperial College, ICCEPT, London UK Philip Andrews-Speed University of Dundee, CEPMLP Beng Wah Ang National University of Singapore Jay Apt Carnegie Mellon University Seth Armitage University of Edinburgh Angel Arcos Endesa, Spain Daniel Aromi University of Maryland Pablo Arocena University of Navarra Hiroshi Asano University of Tokyo Frank Asche University of Stavanger John Ashton University of East Anglia, Centre for Competition & Regulation Emil Attanasi U.S. Geological Survey Hans Auer Vienna University of Technology Maximilian Auffhammer University of California Berkeley Shimon Awerbuch University of Sussex, UK Denis Babusiaux French Petroleum Institute Lance Bachmeier Kansas State University Necmi Bagdadioglu University of East Anglia Mohsen Bahmani-Oskooee University of Wisconsin-Milwaukee Elizabeth Bailey NERA Economic Consultants

171

172 / The Energy Journal

Richard Baillie Erin Baker Fridrik Baldursson Ross Baldick Ed Balistreri Charles Ballard Sara Banazsak Spencer Banzhaf Terry Barker Edmond Baranes Andrea Baranzini Douglas Barnes Mark Barrett Robert Bartels Adolfo Benavides Fred Benth Jan Bentzen Douglas Berg Sandford Berg Jean-Thomas Bernard Paul Bernstein Geoff Bertram Paolo Bertoldi John Besant-jones Subhes Bhattacharyya Andrea Bigano Petter Bjerksund Mette Bjorndal David Bjornstad Thomas B Bjorner David Bjornstad Odile Blanchard Kornelis Blok Gordon Bloomquist Seth Blumsack Carl Blumstein Sven Bode Frederico Boffa Christoph Bohringer David Bonilla Annette Boom Severin Borenstein Valentina Bosetti Audun Botterud Francois Bouffard Pal Boug Alain Bousquet Joseph Bowring Michigan State University University of Massachusetts Amherst University of Reykjavik The University of Texas at Austin Colorado School of Mines Michigan State University American Petroleum Institute Resources For The Future University of Cambridge University of Montpelier HEG Geneva The World Bank University College London University of Sydney, Economics & Business Texas A&M University University of Oslo Aarhus School of Business Sam Houston State University Public Utility Research Center, Florida University Laval, Department of Economics Charles Rivers Associates Victoria University of Wellington European Commission The World Bank University of Dundee FEEM, Milan, Italty NHH, Norway NHH, Norway Oak Ridge National Laboratory AKF Inst. of Local Govt Studies Denmark Oak Ridge National Laboratory University Pierre Mendes-France Ecofys, The Netherlands Washington State University Carnegie Mellon University UC Energy Institute, Berkeley Hamburg Germany Northerwestern University Centre for European Economic Research ZEW University of Cambridge Copenhagen Business School Haas School of Business and UC Energy Institute Fondazione Eni Enrico Mattei Argonne National Laboratory University of Manchester Statistics Norway Universite de la Rochelle France PJM Interconnection

Acknowledgments / 173
John Boyce Gale Boyd Steven Braithwait Runnar Brannlund Thierry Brechet Tim Brennan Dagobert Brito Gregory Brown Stephen Brown Gert Brunekreeft Dallas Burtraw Torstein Arne Bye Robert Cairns Peter Cappers Pantelis Capros Jared Carbone Judith Cardell Bjorn Carlen Paul Carpenter Carlo Carraro Richard Carson Colin Carter David Carter Aitor Carrieta Steve Casler Jennifer Castle Miguel Cerrutti Halim Ceylan Ujayant Chakravorty Hing Lin Chan Duane Chapman Yiuhsu Chen Janie Chermak Menzie Chinn In-Koo Cho Sam Choi Aitor Ciarreta Charlie Cicchetti Joe Clarke Leon Clarke Tim Coelli Jon Conrad Timothy Considine Frank Convery Robert Cope Aad Correlje Ken Costello Simon Cowan University of Calgary Argonne National Laboratory Christensen Associates Omea University Sweden Catholic University of Louvain University of Maryland Baltimore County Rice University University of North Carolina at Chapel Hill Federal Reserve Bank of Dallas, TX Tilburg University, The Netherlands Resources for the Future Statistics Norway McGill University UtiliPoint International, NY National Technical University of Athens Williams University Smith College Dept of Economics, Stockholm University Brattle Group University of Venice University of California San Diego University of California Davis Oklahoma State University Universidad del Pais Vasco Allegheny College Nuffield College, Oxford State of California Pamukkale University, Turkey University of Central Florida Hong Kong Baptist University Cornell University John Hopkins University University of New Mexico University of Wisconsin University of Illinois University of Nevada University of the Basque Country University of Southern California University of Strathclyde Pacific Northwest Lab University of Queensland, Australia Cornell University The Pennsylvania State University University College Dublin Southeastern Louisiana University Delft University of Technology National Regulatory Research Institute Oxford University

174 / The Energy Journal

Claude Crampes Peter Cramton Jesus Crespo Anna Creta Patrick Criqui Ian Cronshaw John Cuddington Juncal Cunado Daniel Czamanski Pradeep Dadhich Carol Dahl Joyce Dargay Susmita Dasgupta Douglas Davis Graham A. Davis Neal Davis Mary Deily Shi-Jie Deng Stephen DeCanio Jeroen de Joode Laurens de Vries Antonia Diaz John Dimitropoulos David Dismukes Yazid Dissou Otto Doering Linda Doman James Dooley K Doroodian Joseph Doucet Stratford Douglas Dominique Dupont Mark Duvall Minh Duong Isaac Dyner Robert Earle Anton Eberhard Andrew Eckert Richard Eckaus Huw Edwards Ruud Egging Vernon Eidman Paul Ekins Denny Ellerman Caroline Elliott Molley Espey Bob Ethier Joseph Eto CICT France University of Maryland University of Innsbruk Universit L. Bocconi LEPII-EPE Grenoble, France International Energy Agency Colorado School of Mines University of Navarra Czamanski Ben-Shahar & Co Ltd TERI, New Delhi, India Colorado School of Mines Transport Studies, University of Oxford The World Bank Virginia Commonwealth University Colorado School of Mines EIA US Dept of Energy Lehigh University Georgia Institute of Technology University of California ECN The Netherlands TU Delft Universidad Carlos III de Madrid SPRU University of Sussex Louisiana State University University of Ottawa Purdue University U.S. Department of Energy Batelle Pacific NW Labs Ohio University School of Business, University of Alberta West Virginia University University of Twente EPRI, Palo Alto, CA Centre-CIRED France National University of Colombia Charles River Associates University of Cape Town Dept of Economics, University of Alberta Massachusetts Institute of Technology Loughborough University ECN The Netherlands University of Minnisota Policy Studies Institute, UK Massachusetts Institute of Technology Lancaster University, Department of Economics Clemson University Market Monitoring, ISO New England Lawrence Berkeley National Laboratory

Acknowledgments / 175
Bradley Ewing Sebastian Eyre Kira Fabrizio Natalia Fabra Simon Fan Barbara Farhar Mehdi Farsi Frank Felder Peter Ferderer Patrick Feve Denzil Fiebig Frank Figge Massimo Filippini Dominique Finon Jeffrey Fischer Karen Fisher-Vanden SteinErik Fleten Kevin Forbes Chris Forde Finn Forsund Roger Fouquet Giovanni Fraquelli Mark French Manuel Frondel Steven Gabriel Beatriz Gaitan-Soto Marzio Galeotti Paul Gallagher Richard Garbaccio Dermot Gately Richard Gerlach Reyer Gerlagh John German Mike Giberson Vincent Gitz Rajeev Goel Jose Goldemberg Charles Goldman Rolf Golombek Noriyuki Goto Daniel Graham Wafik Grais Ove Grande Ryan Grant Richard Green David Greene Lorna Greening Paul Gribik Texas Tech University John Hall Associates Emory University Carlos III University Madrid Lingnan University National Renewable Energy Laboratory ETHZ Switzerland Rutgers, State University of New Jersey Macalester College University of Toulouse University of NSW, School of Economics St Andrews University University of Lugano Energy Program of CNRS, CIRED France Federal Trade Commission, Washington Dartmouth College, Hanover, NH NTNU, Norway Catholic University of America, Washington University of Leeds University of Oslo University of the South Pacific, Fiji University of Eastern Piedmont Federal Reserve Board RWI, Essen, Germany University of Maryland University of Hamburg University of Milan Iowa State University United States EPA Dept. of Economics, New York University University of Sydney Vrije University Amsterdam American Honda Motor Co Giberson Co. Centre CIRED, France Illinois State University, Dept of Economics University of Sao Paulo, Brazil Lawrence Berkeley National Laboratory Frischsenteret, Oslo Norway University of Tokyo, Japan Imperial College, London UK The World Bank SINTEF Norway Analysis Group Inc University of Birmingham, UK Oak Ridge National Laboratory Energy and the Environment, Los Alamos, NM Midwest ISO

176 / The Energy Journal

James Griffin Henri de Groot Willem van Groenendaal George Gross Omrane Guedhami Reinhard Haas Stanton Hadley Lester Hadsell Catherine Hagem Niels Haldrup James Hamilton Henrik Hammar Shawkat Hammoudeh Sue Handy Nick Hanley Rognvaldur Hannesson Philip Hanser Frank Harris Nick Hartley Peter Hartley Toru Hattori Scott Harvey Michael Heely Peter Helby Jerome Hericourt Ana Maria Herrera Sebastian Hers Borge Hess James Hewlett Pierre Hillion Lennart Hjalmarsson Vladimir Hlasny Chun-Yu Ho William Hogan Stephen Holland Daniel Hollas Par Holmberg Franziska Holz Einar Hope Stephen Horan Marvin Horowitz Richard Howarth Roy Hrab Bwo Nung Huang Franz Hubert Randy Hudson Jonathan Hughes Benjamin Hunt Bush School, Texas A&M University Free University, The Netherlands Tilburg University University of Illinois at Urbana Champaign University of South Carolina Technical University of Vienna Oak Ridge National Laboratory University at Albany NY Statistics Norway University of Aarhus, Denmark University of California San Diego Gothenburg University Drexel University University of California Davis University of Sterling Norwegian School of Economics and Business Admin Brattle Group Southern California Edison Oxera Consulting Ltd Rice University CRIEPI Japan Law & Economics Consulting Group, Cambridge Colorado School of Mines Lund University, Environmental & Energy Systems Universite Paris 1 Michigan State University ECN The Netherlands Dresden University of Technology Energy Information Administration, US DOE INSEAD, France Goteborg University Michigan State University Boston University Harvard University University of North Carolina Greensboro University of Texas San Antonio Uppsala University German Institute of Economic Research NHH, Norway St. Bonaventure University Demand Research Environmental Studies, Dartmouth College Ontario Energy Board National Chung Cheng University Humboldt University Berlin Oak Ridge National Laboratory University of California Davis International Monetary Fund

Acknowledgments / 177
Lester Hunt John Hunter Hillard Huntington Omowumi Iledare Marija Ilic Mariana Iootty Jun Ishii Marc Ivaldi Maithili Iyer Christopher Jablonowski Mark Jaccard Amy M. Jaffe Tooraj Jamasb James Jensen Stine Jensen Yanbo Jin Donald W. Jones Clifton T. Jones Paul L. Joskow Frederick Joutz Guy Judge Martin Junginger Katarina Juselius Daniel Kaffine Edward Kahn Mikiko Kaimuma Mark Kaiser Robert Kaufmann Yoichi Kaya Lynne Keisling Claudia Kemfert Jussi Keppo Madhu Khanna Neha Khanna Daniel Khazzoom A Kildergaard Lutz Killian Brendan Kirby Andrew Kleit Gernot Klepper Christopher Knittel Bertrand Koebel Tina Koljonen Jonathan Koomey Tarjei Kristiansen Kerry Krutilla Subal Kumbhakar Rolf Kunneke University of Surrey Brunel University Energy Modeling Forum, Stanford University Louisiana State University Carnegie Mellon University Federal University of Rio de Janeiro Brazil University of California at Irvine CICT, France Lawrence Berkeley Lab University of Texas at Austin Simon Fraser University Rice University, Texas University of Cambridge Jensen Associates Inc RISOE Denmark California State University Northridge RCF Economics & Financial Consulting Stephen F. Austin State University Massachusetts Institute of Technology George Washington University University of Portsmouth Utrecht University University of Copenhagen Colorado School of Mines Analysis Group/Economics National Institute for Environmental Studies, Japan Louisianna State University, Center for Energy Studies Boston University RITE, Japan Northwestern University Oldenburg University, Germany University of Michigan University of Illinois, Urbana-Champaign Cornell University San Jose State University, California Ris National Laboratory University of Michigan Oak Ridge National Laboratory Pennsylvania State University IFW University of Kiel University of California at Davis Universite Louis Pasteur VTT Finland Lawrence Berkeley National Laboratory KEMA Research Consultants University of Indiana State University of New York, Economics TU Delft

178 / The Energy Journal

Ken Kurani Gerard Kuper Atsushi Kurosawa Snorre Kverndokk John Kwoka Andy Kydes Socrates Kypreos Xavier Labandeira Stephane Lambrecht Andreas Lange Ian Lange Markku Lanne Frederic Lantz Alessandro Lanza Eric Larsen Don Larson Harri Laurikka Lester Lave Thomas Leautier Chloe Le Coq Chien Chiang Lee Dean Leistikow Jeff Leitzinger Francois Lescaroux Steve Letendre Soren Leth-Petersen Matt Lewis Gary Libecap Jiang Lin Pedro Linares Matti Liski Stephen Littlechild Gang Liu Wilson Liu Asa Lofgren David Loughran Knox Lovell Mark Lowry Tom Lundgren Robert Lundmark Matilde Machado Dave Maddison Jeff Makholm Mike Maloney Erin Mansur Jose Mantovani Osmel Manzano Gregg Marland University of California Davis Groningen University, The Netherlands Institute of Applied Energy Japan Ragnar Frisch Centre for Economic Research Northeastern University US Dept of Energy Geneva Switzerland University of Vigo, Spain University of Lille, France University of Maryland US Environmental Protection Agency University of Jyvaskyla Finland Institut Francais du Petrole FEEM, Italy University of Lugano The World Bank GreenStream Network Ltc Carnegie Mellon University University of Toulouse University of California Berkeley National Chung Hsing University Fordham University Econ One French Petroleum Institute Green Mountain College AKF Denmark University of California Energy Institute University of Arizona Lawrence Berkeley National Laboratory Pontifical University of Comillas Spain Helsinki School of Economics, Finland Judge Institute of Management, University of Cambridge Statistics Norway James Madison University Goteborg University RAND, California The University of Georgia, College of Business Pacific Economics Group Swedish University of Agricultural Sciences Lulea University of Technology University Carlos III de Madrid University of Birmingham NERA Clemson University School of Management, Yale University Sao Paulo State University Brazil Corporacion Andina de Fomento Oak Ridge National Laboratory

Acknowledgments / 179
Chris Marnay Joan Martinez Isamu Matsukawa Scott Matthews John Mayo Tanga McDaniel Ross McKitrick Stuart McMenamin Robert McNown Kenneth Medlock Bill Megginson Robert Mendelsohn Flavio Menezes Sophie Meritet Gilbert Metcalf Robert J. Michaels Augusto R. Micola Carlos de Miguel David Mills John Miranowski Franklin Mixon Juan Pablo Montero Imad A. Moosa Alan Moran Juan Moreno Richard Morgenstern Matt Morey Knut Anton Mork Saeed Moshiri Diana Moss Frederic Murphy Scott Murtishaw Masao Nakamura Carole Nakhle Erkka Nasakkala Peter Navarro Greg Nemet Bernie Neenan Karsten Neuhoff Anne Neumann David Newbery Shawn Xiaoguang Ni Alessandro Nicita Ole E. Barndorff-Nielsen Oivind Nilsen Michael Noel Pierre Noel Robert Noland Lawrence Berkelely National Laboratory Autonomous University of Barcelona Musashi University, Japan Carnegie Mellon University Georgetown University Appalachian State University Dept of Economics, University of Guelph Itron Co University of Colorado, Economics Dept Rice University, James A Baker III Institute Universite Paris Dauphine Yale University University of Queensland Universite Dauphine, France Tufts University California State University, Fullerton IMD International, Switzerland University of Vigo Spain University of Virginia, Economics Dept Iowa State University University of Southern Mississippi Pontifical Catholic University of Chile La Trobe University, Australia IPA Australia University of Calgary Resources for the Future Christensen Associates Handelsbanken Norway University of Manitoba American Antitrust Institute Dept of Economics, Temple University Lawrence Berkeley National Laboratory University of British Columbia University of Surrey, UK Helsinki University, Finland University of California Irvine University of Wisconsin Utilipoint Cambridge University, Economics Dresden University of Technology Cambridge University University of Missouri The World Bank University of Aarhus NHH Norway University of California, San Diego University of Cambridge Imperial College London

180 / The Energy Journal

Nikos Nomikos William Nordhaus Richard Norgaard Gbadebo Oladosu Luis Olmos Ole Jess Olsen Frans Oosterhuis Petter Osmundsen Andrew Oswald Hugh Outhred Anthony Owen Emilio Padilla Karen Palmer Sergey Paltsev Agis Papadopoulos Evangelia Papapetrou Carl Pasurka Robert Patrick Peter Pedroni Yannick Perez Yezid Orlando PerezAleman Roger Perman Adriaan Perrels Daniel Petrolia Frederik Petterson Wolfgang Pfaffenberger Andy Philpott Jeremy Piger Robert Pindyck Andr Plourde Steve Polasky Michael Pollitt Carlos Pombo Aude Pommeret Julia Popova Edward Porter Paul Preckel Geoff Pritchard Stef Proost Steven Puller John Quiggin Kevin Quinn Philip Quirion Nuria G Rabanal Stanislav Radchenko Ricardo Raineri Arvid Raknerrud Reza Ramazani Cass Business School, London Yale University University of California Berkeley Environmental Sciences Division, ORNL University of Comillas Roskilde University VU University Amsterdam University of Stavanger University of Warwick University of New South Wales, Australia Curtin University of Technology UAB Spain Resources for the Future, Washington Massachusetts Institute of Technology Aristole University Thessaloniki Bank of Greece, University of Athens US Environmental Protection Agency Rutgers University Williams University U-PSUD France Pontificia University Javeriana University of Strathclyde VATT, Finland Mississippi State University Lulea University International University and Bremer Energy Institute University of Auckland University of Oregon Massachusetts Institute of Technology University of Alberta, Canada University of Minnisota Judge Inst of Management, Cambridge University University of the Andes HEC University of Lausanne West Virginia University American Petroleum Institute Purdue University University of Auckland Catholic University of Leuven Texas A&M University, Dept of Economics University of Queensland St. Norbert College Centre-CIRED France University of Leon Spain University of North Carolina Charlotte Pontificia Universidad Catolica de Chile Statistics Norway Saint Micheals College, Vermont

Acknowledgments / 181
Kevin Rask Ronald Ratti Micheal Rauscher Dennis Ray Hossein Razavi Grant Read Amy Richmond Fieke Rijkers Francisco Rodriguez Joseph Roop Stephanie Ropenus Fabien Roques Adam Rose Kenneth Rose Juan Rosellon Roberto Roson Knut Einar Rosendahl Michael Rothkopf Geoffrey Rothwell Bryan Routledge Jonathan Rubin Sophia Ruester Larry Ruff Carlos Rufin Milton Russell Tom Rutherford Emile Rutledge David Ryan Perry Sadorsky Francois Salanie Ronald Sands Jayant Sathaye Vladislav Savin Carlos Scarpa Lee Schipper Benjamin Schlesinger Dieter Schmitt Craig Schulman Edwardo Schwartz Peter Schwarz Frank Scrimgeour Yadviga Semikolenova Apostolos Serletis Deepak Sharma Hany Shawky Robert Shelton Afzal Siddiqui Jos Sijm Colgate Unviersity, Economics University of Missouri-Columbia Rostock University Germany Power Systems Engineering Research Center, Madison WI The World Bank University of Canterbury United States Military Academy CASEMA The Netherlands Wesleyan University Pacific Northwest National Laboratory RISOE Denmark International Energy Agency University of Southern California National Regulatory Research Institute, Ohio CIDE Mexico Universita Ca Foscari di Venezia Statistics Norway Rutgers University Stanford University Carnegie Mellon University University of Maine Dresden University of Technology Energy and Economic Consulting, Windsor CA Babson College University of Tennessee Knoxville University of Colorado UAE University University of Alberta Schulich School of Business, York University University of Toulouse Joint Global Change Research Institute PNL Lawrence Berkeley National Laboratory ISEER, Russia Univesit degli Studi di Brescia EMBARQ the WRI Center for Sustainable Transport Benjamin Schlesinger and Associates Inc Essen University Law and Economics Consulting Group Anderson School, UCLA University of North Carolina Charlotte University of Waikato Management School Colorado School of Mines University of Calgary University of Technology, Sydney School of Business, University at Albany, NY Oak Ridge National Laboratory University College London ECN, The Netherlands

182 / The Energy Journal

Julian Silk Boris Silverstovs Fereidoon Sioshansi George Siotis Tomas Sjogren James Skea Terje Skjerpen Margaret Slade Kenneth Small James E. Smith James L. Smith Thomas Smith Ron P Smith Vaclav Smil Patrik Soderholm Ron Soligo Ugur Soytas Carine Staropoli John Staub Spiro Stefanou Thanasis Stengos David Stern Jonathan Stern Thomas Sterner Paul Stevens Steven Stoft Paul Stoneman Karl Storchmann Mark Strazicich Eric Strobl Steinar Strom Hiroaki Suenaga Andrew Sweeting Richard Tabors Kaisa Tavanainen Massimo Tavoni Jerry Taylor Timothy Taylor Kjetil Telle Dek Terrell Leigh Tesfatsion Philip Thompson Elspeth Thomson Mark Thurber Tom Tietenberg Asher Tishler Tom Tietenberg Sigve Tjotta George Washington University DIW Germany Henwood Energy Services Inc Universidad Carlos III de Madrid UMU Sweden UK Energy Research Centre Statistics Norway Warwick University University of California Irvine Fuqua School of Business, Duke University Southern Methodist University, Texas Zayed University, Dubai Birkbeck College, Dept of Economics University of Manitoba Lulea University of Technology, Sweden Rice University, Texas METU Turkey University of Paris I EIA US Dept of Energy Pennsylvania State University University of Guelph, Department of Economics Rensselaer Polytechnic Institute Oxford Institute for Energy Studies Gothenburg Univeristy University of Dundee, Scotland Nantucket, MA Warwick Business School Whitman College, WA Appalachian State University Catholic University of Louvain University of Oslo Curtin University of Technology Northwestern University, Dept of Economics Charles River Associates Lappeenranta University of Technology FEEM, Italy CATO Institute University of Florida, Gainsville Statistics Norway Louisiana State University Iowa State University, Dept of Economics Central Michigan University National University of Singapore Stanford University Colby College Tel Aviv University, Faculty of Management Colby College University of Bergen

Acknowledgments / 183
Asgeir Tomasgard NTNU, Trondheim Norway Christoph Tonjes Clingendael Intl Energy Programme Kenneth Train University of California Berkeley Michael Trebilcock University of Toronto Lenos Trigeorgis Sloan School of Management, MIT Stefan Trueck Queensland University of Technology Mike Tsionas Athens University of Economics and Business Wallace Tyner Purdue University Kazushi Uemura Dept Economics, Sophia University, Japan Ismet Ugursal Dalhousi University Rocio Uria University of California Colin Vance RWI Essen Germany Kurt Van Dender University of California Irvine Rick van der Ploeg IUE, Italy Peter Van Doren Cato Institute Willem Van Groenendaal Tilburg University Frits Van Oostvoorn ECN The Netherlands Angel dela VegaNavarro Universidad Nacional Autonomo Mexico Shree Vikas SAIC Company Ingo Vogelsang Boston University, Dept of Economics Klaus Vogstad NTNU Norway Herman Vollebergh Erasmus University Rotterdam Nils-Henrik Von der Fehr University of Oslo, Dept of Economics Christian Von Hirschhausen DIW, Germany Laurens de Vries Delft University of Technology Catherine Waddams Centre for Competition & Regulation, Univ of East Anglia Zia Wadud Imperial College, London Liam Wagner University of Queensland Martin Wagner Institute for Advances Studies Vienna Christine Wallich The World Bank W. David Walls Centre for Regulatory Affairs, University of Calgary Margaret Walls Resources for the Future Jianhui Wang Argonne National Lab Zhongmin Wang Northeastern University Linda Warell Lulea University of Technology Mike Waterson University of Warwick William Watson Energy Information Administration, US DOE Jason Wei University of Toronto Heinz Welsch University of Oldenburg, Economics Dept Rafal Weron Hugo Steinhaus Center Poland Douglas West University of Alberta Hege Westskog CICERO Norway Tom Weyman-Jones Loughborough University Michelle White University of California San Diego Bert Willems Catholic University of Louvain, Belgium Eric Williams Arizona State University Jeffrey Williams University of California, Davis Robert Williams University of Texas

184 / The Energy Journal

Ian Sue Wing Franz Wirl Ryan Wiser Wiepke Wissema Norbert Wohlgemuth Catherine Wolfram Chi-Keung Woo Ernst Worrell Andrew Worthington Steve Wright Kang Wu June Xie Kuan Xu Lin Xu Jian Yao Adonis Yatchew Gary Yohe Mine Yucel Georges Zaccour Jay Zarnikau Hisham Zerriffi Yabei Zhang Yanhua Zhang Zhen Zhu Jifang Zhuang Radu Zmeureanu Bob van der Zwaan Boston University, Massachusetts University of Vienna, Austria Lawrence Berkeley National Lab WUR The Netherlands University of Klagenfurt Haas School of Business, UC Berkeley Energy and Environmental Economics, Inc. Utrecht University University of Wollongong, Australia Birkbeck College Research Program, East-West Center, Honolulu California ISO Co Dalhousie University University of Texas UC Berkeley, IEOR University of Toronto, Department of Economics Wesleyan University, Economics Federal Reserve Bank of Dallas, Texas GERAD, Ecole des Hautes Etudes Commerciale Frontier Associations, Austin Texas Stanford University University of Maryland University of Toulouse C.H. Guernsey Co, University of Central Oklahoma University of Maryland Concordia University Energy Research Centre of the Netherlands

VOLUME 29 INDEX, 2008

AUTHORS
AGUILERA, ROBERTO F.: Book review, A Handbook of Primary Commodities in the Global Economy, (by Marian Radetzki), 4, 168-170 ASCHE, FRANK; NILSEN, ODD BJARTE; and TVETERS, RAGNAR: Natural Gas Demand in the European Household Sector, 3, 27-46 AUFFHAMMER, MAXIMILIAN; BLUMSTEIN, CARL; and FOWLIE, MEREDITH: Demand-Side Management and Energy Efficiency Revisited, 3, 91-104 BALDICK, ROSS see JOUNG, MANHO BALISTRERI, EDWARD J.: Book review, Hacia El Futuro: Energy, Economics and the Environment in 21st Century Mexico, (by Maria Eugenia Ibarrarn and Roy Boyd), 1, 173-174 BELLAS, ALLEN; and LANGE, IAN: Impacts of Market-based Environmental and Generation Policy on Scrubber Electricity Usage, 2, 151-164 BLUMSTEIN, CARL see AUFFHAMMER, MAXIMILIAN BOYD, GALE A.: Estimating Plant Level Energy Efficiency with a Stochastic Frontier, 2, 23-44 BROWN, STEPHEN P.A.; and YCEL, MINE K.: What Drives Natural Gas Prices? 2, 45-60 DEWEES, DONALD N.: Pollution and the Price of Power, 2, 81-100 DREHER, AXEL; and KRIEGER, TIM: Do Prices for Petroleum Products Converge in a Unified Europe with NonHarmonized Tax Rates?, 1, 61-88 EWING, BRADLEY T. see HAMMOUDEH, SHAWKAT FARSI, MEHDI; FETZ, AURELIO; and FILIPPINI, MASSIMO: Economies of Scale and Scope in the Swiss Multi-Utilities Sector, 4, 123-144 FELDER, FRANK A.: Book review, Competitive Electricity Markets and Sustainability, (edited by Franois Lvque), 3, 177-179 FELDER, FRANK: Book review, Electric Choices: Deregulation and the Future of Electric Power (edited by Andrew N. Kleit), 1, 175-176 FLETEN, STEIN-ERIK see MARIBU, KARL MAGNUS FOSTER, JOHN; HINICH, MELVIN J.; and WILD, PHILLIP: Randomly Modulated Periodic Signals in Australias National Electricity Market, 3, 105-130 FOWLIE, MEREDITH see AUFFHAMMER, MAX FRONDEL, MANUEL; PETERS, JORG; and VANCE, COLIN: Identifying the

Boldface numbers refer to issue number: 1 (January); 2 (April); 3 (July); 4 (October).

185

186 / The Energy Journal

Rebound: Evidence from a German Household Panel, 4, 145-170 GIBSON, JOHN see OLIVIA, SUSAN GILLINGHAM, KENNETH see VAN BENTHEM, ARTHUR GORDON, RICHARD L.: Book review, The Economics of Public Utilities, (by Ray Rees) and The Political Economy of Regulation, (edited by Thomas P. Lyon), 1, 177-178 GORDON, RICHARD L.: Book review, Electricity and Energy Policy in Britain, France and the United States since 1945, (by Martin Chick), 4, 165-166 GRAZI, FABIO; VAN DEN BERGH, JEROEN C.J; and VAN OMMEREN, JOS: An Emprical Analysis of Urban Form, Transport, and Global Warming, 4, 97-122 GREEN, RICHARD: Carbon Tax or Carbon Permits: The Impact on Generators Risks, 3, 67-90 GRONWALD, MARC: Large Oil Shocks and the US Economy: Infrequent Incidents with Large Effects, 1, 151-172 HAMILTON, JAMES D.: Book review, China and the Global Energy Crisis: Development and Prospects for Chinas Oil and Natural Gas, (by Tatsu Kambara and Christopher Howe), 2, 185-186 HAMMOUDEH, SHAWKAT M.; EWING, BRADLEY T.; and THOMPSON, MARK A.: Threshold Cointegration Analysis of Crude Oil Benchmarks, 4, 79-96 HARTLEY, PETER R.; MEDLOCK III, KENNETH B.; and ROSTHAL, JENNIFER E.: The Relationship of Natural Gas to Oil Prices, 3, 47-66 HINICH, MELVIN J. see FOSTER, JOHN HOLLAND, STEPHEN P.: Modeling Peak Oil, 2, 61-80 HOPE, EINAR: Book review, The Political Economy of Power Sector Reform: The Experiences of Five Major Developing Countries, (by David G. Victor and Thomas C. Heller), 2, 186-187 HUANG, BWO-NUNG: Factors Affecting an Economys Tolerance and Delay of Response to the Impact of a Positive Oil Price Shock, 4, 1-34 HUGHES, JONATHAN E.; KNITTEL, CHRISTOPHER R.; and SPERLING, DANIEL: Evidence of a Shift in the Short-Run Price Elasticity of Gasoline Demand, 1, 113-134 JOUNG, MANHO; BALDICK, ROSS; and SON, YOU SEOK: The Competitive Effects of Ownership of Financial Transmission Rights in a Deregulated Electricity Industry, 2, 165-184 KAFFINE, DANIEL: Book review, The Politics of the Environment: Ideas, Activism, Policy, (by Neil Carter), 3, 180-182 KNITTEL, CHRISTOPHER R. see HUGHES, JONATHAN LANGE, IAN see BELLAS, ALLEN LAVE, LESTER see SPEES, KATHLEEN LESCAROUX, FRANOIS; and RECH, OLIVIER: The Impact of Automobile Diffusion on the Income Elasticity of Motor Fuel Demand, 1, 41-60 LINN, JOSHUA: Technological Modifications in the Nitrogen Oxides Tradable Permit Program, 3, 153-176 MARIBU, KARL MAGNUS; and FLETEN, STEIN-ERIK: Combined Heat and Power in Commercial Buildings: Investment and Risk Analysis, 2, 123-150

Volume 29 Index / 187

MAYERES, INGE; and VAN REGEMORTER, DENISE: Modelling the Health Related Benefits of Environmental Policies and Their Feedback Effects: A CGE Analysis for the EU Countries with GEM-E3, 1, 135150 MEDLOCK III, KENNETH B. see HARTLEY, PETER R. MELINO, ANGELO; and PEERBOCUS, NAWAZ: High Frequency Export and Price Responses in the Ontario Electricity Market, 4, 35-52 METCALF, GILBERT E.: An Empirical Analysis of Energy Intensity and Its Determinants at the State Level, 3, 1-26 MOHN, KLAUS: Efforts and Efficiency in Oil Exploration: A Vector ErrorCorrection Approach, 4, 53-78 NAWAZ, PEERBOCUS see MELINO, ANGELO NILSEN, ODD BJARTE see ASCHE, FRANK OLIVIA, SUSAN; and GIBSON, JOHN: Household Energy Demand and the Equity and Efficiency Aspects of Subsidy Reform in Indonesia, 1, 21-40 PERCEBOIS, JACQUES: Electricity Liberalization in the European Union: Balancing Benefits And Risks, 1, 1-20 PETERS, JORG see FRONDEL, MANUEL RECH, OLIVIER see LESCAROUX, FRANOIS ROSNES, ORVIKA: The Impact of Climate Policies on the Operation of a Thermal Power Plant, 2, 1-22 ROSTHAL, JENNIFER E. see HARTLEY, PETER R. SCHOLTENS, BERT; and WANG, LEI: Oil Risk in Oil Stocks, 1, 89-112 SON, YOU SEOK, see JOUNG, MANHO SPEES, KATHLEEN; and LAVE, LESTER: Impacts of Responsive Load in PJM: Load Shifting and Real Time Pricing, 2, 101-120 SPERLING, DANIEL see HUGHES, JONATHAN E. SWEENEY, JAMES see VAN BENTHEM, ARTHUR THOMPSON, MARK A. see HAMMOUDEH, SHAWKAT TVETERS, RAGNAR see ASCHE, FRANK VAN BENTHEM, ARTHUR; GILLINGHAM, KENNETH; and SWEENEY, JAMES: Learning-by-Doing and the Optimal Solar Policy in California, 3, 131-152 VAN DEN BERGH, JEROEN C.J. see GRAZI, FABIO VAN DOREN, PETER: Book review, The Strategic Petroleum Reserve: U.S. Energy Security and Oil Politics, 1975-2005, (by Bruce A. Beaubouef), 4, 167-168 VAN OMMEREN, JOS see GRAZI, FABIO VAN REGEMORTER, DENISE see MAYERES, INGE VANCE, COLIN see FRONDEL, MANUEL WANG, LEI see SCHOLTENS, BERT WILD, PHILLIP see FOSTER, JOHN YCEL, MINE K see BROWN, STEPHEN P.A.

188 / The Energy Journal TITLES

Carbon Tax or Carbon Permits: The Impact on Generators Risks, Richard Green, 3, 67-90 China and the Global Energy Crisis: Development and Prospects for Chinas Oil and Natural Gas, (by Tatsu Kambara and Christopher Howe), book review by James D. Hamilton, 2, 185186 Combined Heat and Power in Commercial Buildings: Investment and Risk Analysis, by Karl Magnus Maribu and Stein-Erik Fleten, 2, 123-150 Competitive Electricity Markets and Sustainability, (edited by Franois Lvque), book review by Frank A. Felder, 3, 177-179 The Competitive Effects of Ownership of Financial Transmission Rights in a Deregulated Electricity Industry, by Manho Joung, Ross Baldick and You Seok Son, 2, 165-184 Demand-Side Management and Energy Efficiency Revisited, Maximilian Auffhammer, Carl Blumstein and Meredith Fowlie, 3, 91-104 Do Prices for Petroleum Products Converge in a Unified Europe with NonHarmonized Tax Rates? by Axel Dreher and Tim Krieger, 1, 61-88 The Economics of Public Utilities, (by Ray Rees) and The Political Economy of Regulation, (edited by Thomas P. Lyon), book reviews by Richard L. Gordon, 1, 177-178 Economies of Scale and Scope in the Swiss Multi-Utilities Sector, by Mehdi Farsi, Aurelio Fetz, and Massimo Filippini, 4, 123-144 Efforts and Efficiency in Oil Exploration: A Vector Error- Correction Approach, by Klaus Mohn, 4, 53-78 Electric Choices: Deregulation and the Future of Electric Power (edited by Andrew N. Kleit), book review by Frank Felder, 1, 175-176 Electricity and Energy Policy in Britain, France and the United States since 1945, (by Martin Chick), book review by Richard L. Gordon, 4, 165-166 Electricity Liberalization in the European Union: Balancing Benefits And Risks, by Jacques Percebois, 1, 1-20 An Empirical Analysis of Energy Intensity and Its Determinants at the State Level, by Gilbert E. Metcalf, 3, 1-26 An Emprical Analysis of Urban Form, Transport, and Global Warming, by Fabio Grazi, Jeroen C.J van den Bergh, and Jos van Ommeren, 4, 97122 Estimating Plant Level Energy Efficiency with a Stochastic Frontier, by Gale A. Boyd, 2, 23-44 Evidence of a Shift in the Short- Run Price Elasticity of Gasoline Demand, by Jonathan E. Hughes, Christopher R. Knittel and Daniel Sperling, 1, 113-134 Factors Affecting an Economys Tolerance and Delay of Response to the Impact of a Positive Oil Price Shock, by Bwo-Nung Huang, 4, 1-34 Hacia El Futuro: Energy, Economics and the Environment in 21st Century Mexico, (by Maria Eugenia Ibarrarn and Roy Boyd), book review by Edward J. Balistreri, 1, 173-174 A Handbook of Primary Commodities in the Global Economy, (by Marian Radetzki), book review by Roberto F. Aguilera, 4, 168-170

Volume 29 Index / 189

High Frequency Export and Price Responses in the Ontario Electricity Market, by Angelo Melino and Nawaz Peerbocus, 4, 35-52 Household Energy Demand and the Equity and Efficiency Aspects of Subsidy Reform in Indonesia, by Susan Olivia and John Gibson, 1, 21-40 Identifying the Rebound: Evidence from German Household Panel, by Manuel Frondel, Jorg Peters, and Colin Vance, 4, 145-170 The Impact of Climate Policies on the Operation of a Thermal Power Plant, by Orvika Rosnes, 2, 1-22 The Impact of Automobile Diffusion on the Income Elasticity of Motor Fuel Demand, by Franois Lescaroux and Olivier Rech, 1, 41-60 Impacts of Market-based Environmental and Generation Policy on Scrubber Electricity Usage, by Allen Bellas and Ian Lange, 2, 151-164 Impacts of Responsive Load in PJM: Load Shifting and Real Time Pricing, by Kathleen Spees and Lester Lave, 2, 101-120 Large Oil Shocks and the US Economy: Infrequent Incidents with Large Effects, by Marc Gronwald, 1, 151-172 Learning-by-Doing and the Optimal Solar Policy in California, by Arthur van Benthem, Kenneth Gillingham and James Sweeney, 3, 131-152 Modeling Peak Oil, by Stephen P. Holland, 2, 61-80 Modelling the Health Related Benefits of Environmental Policies and Their Feedback Effects: A CGE Analysis for the EU Countries with GEM-E3, by Inge Mayeres and Denise Van Regemorter, 1, 135-150 Natural Gas Demand in the European Household Sector, by Frank Asche, Odd Bjarte Nilsen and Ragnar Tveters, 3, 27-46 Oil Risk in Oil Stocks, by Bert Scholtens and Lei Wang, 1, 89-112 The Political Economy of Power Sector Reform: The Experiences of Five Major Developing Countries, (by David G. Victor and Thomas C. Heller), book review by Einar Hope, 2, 186187 The Politics of the Environment: Ideas, Activism, Policy, (by Neil Carter), book review by Daniel Kaffine, 3, 180-182 Pollution and the Price of Power, by Donald N. Dewees, 2, 81-100 Randomly Modulated Periodic Signals in Australias National Electricity Market, by John Foster, Melvin J. Hinich and Phillip Wild, 3, 105-130 The Relationship of Natural Gas to Oil Prices, by Peter R. Hartley, Kenneth B Medlock III and Jennifer E. Rosthal, 3, 47-66 The Strategic Petroleum Reserve: U.S. Energy Security and Oil Politics, 1975-2005, (by Bruce A. Beaubouef), book review by Peter Van Doren, 4, 167-168 Technological Modifications in the Nitrogen Oxides Tradable Permit Program, by Joshua Linn, 3, 153-176 Threshold Cointegration Analysis of Crude Oil Benchmarks, by Shawkat M. Hammoudeh, Bradley T. Ewing, and Mark A. Thompson, 4, 79-96 What Drives Natural Gas Prices? by Stephen P. A. Brown and Mine K. Ycel, 2, 45-60

190 / The Energy Journal

The International Association for Energy Economics

is pleased to announce the winners of the

The Energy Journal Campbell Watkins Best Paper Award

for 2007
Massachussetts Institute of Techonology

A. DENNY ELLERMAN
and

JUAN-PABLO MONTERO
Universidad Catolica de Chile

for their paper

The Efciency and Robustness of Allowance Banking in the U.S. Acid Rain Program
Established in 1989, the Campbell Watkins Best Paper Award is presented each year to the author(s) of the paper judged to be the best from among those published in the IAEEs premier energy publication, The Energy Journal. The selection is made by a committee chaired by the Associations Vice President for Publications and the award is presented at the Associations International Conference. The Award carries a stipend of $2500. The Energy Journal, a leader in its eld, is a refereed publication and is published quarterly by the International Association for Energy Economics. More information on The Energy Journal and the IAEE may be had by contacting the Association at
28790 Chagrin Blvd., Suite 350 Cleveland, OH 44122 Phone 216-464-5365 Fax 216-464-2737 iaee@iaee.org www.iaee.org

/ 191
Master your power. And boost it.

Professional MBA Energy Management

A unique and tailor-made MBA program that sets world-class standards in Energy Management studies. Target Group: Industry professionals (middle- and top-management) Professionals of international organizations Technical professionals moving to a general management position Professionals of international nancial institutions Consultants and project managers Duration: 15 months, part-time (17 sessions blocked to 5 modules) Venue: 4 modules in Vienna, home of many international organizations such as IAEA, OPEC etc. and 1 module in the Gulf Region Faculty: world-renowned professors and experts Start: November 2008 Contact: AnnaMaria.Nusko@wu-wien.ac.at +43 1 313 36 5225 www.executiveacademy.at

192 / The Energy Journal

IAEE/USAEE CONFERENCE PROCEEDINGS ORDER FORM

Developing & Delivering Affordable Energy in the 21st Century Securing Energy in Insecure Times
27th USAEE/IAEE North American Conference, Houston, TX, September 1619, 2007 This CD-Rom includes articles on the following topics:
Crude Oil and Petroleum Product Price Dynamics Economics of the LNG Industry Energy Efficiency and the Economy Large-Scale, Low Carbon Energy Technologies Unconventional Fossil Fuel Resources: Challenges and Opportunities Political Economy of Energy Energy Policy and Price Effects on Economic Growth Impact of International Environmental Agreements on Reducing Carbon Emissions Global Perspectives on Electric Power Transmission Infrastructure Distributed Energy Resources & Renewables Price Impact on Upstream Petroleum Industry Investments

29th IAEE International Conference, Potsdam, Germany, June 7-10, 2006 This CD-Rom includes articles on the following topics:
Energy in an Insecure World Securing Oil and Gas Supplies Kyoto and Beyond Sustainable Transportation Renewables Role in Securing Energy Long-term Technology and Policy Choices Carbon Modeling Combined Heat and Power - CHP Hydrogen Wholesale Electricity Models Wind Integration Power and Gas Distribution Electricity Market Design Simulation of Efficient Energy Use Long-term Contracts, Vertical Integration, and Competition in Electricity and Gas Markets

From Restructuring to Sustainability: Energy Policies for the 21st Century

30th IAEE International Conference, Wellington, New Zealand, February 1821, 2007 This CD-Rom includes articles on the following topics:
Taking Stock: Two decades of Deregulation and Restructuring in Energy Markets Looking Forward: Energy, Poverty, and Sustainable Development New Market Drivers: Emerging Global Markets for Carbon and LNG Old Habits Die Hard: The Future of Oil and Coal The World Energy and Environmental Outlook at the Commencement of Kyoto 1 Energy Market Design: Lessons and Issues from Around the World Network Expansion, Infrastructure Adequacy and Crisis Management in Electricity and Gas Markets Energy, Poverty and Sustainable Development The Global Market for the LNG: Technologies, Arbitrage Opportunities, and Consequences for National and Regional Energy Markets The Future of Decentralised Energy Systems New Zealand and Australian Energy and Environmental Perspectives

Fueling the Future: Prices, Productivity, Policies and Prophecies

25th USAEE/IAEE North American Conference, Denver, CO, Sept. 18-21, 2005
Natural Gas Industry Economics of Electric and Gas Utilities International Energy Economics Energy Industry Finance Market for Motor Vehicle Fuels How Credible are Proven Oil Reserves? High Oil Prices: A Non-OPEC Capacity Game Restoring the Nuclear Option in the U.S. Energy Data and Modeling Demand Estimation Economics of New Energy Technologies Energy and the Environment Oil Industry: E&P, Transportation, Refining, etc. Are High Oil Prices Here to Stay? The Hydrogen Futures Simulation Model Natural Gas Market Volatility Deregulation and Restructuring in Power Markets

This CD-Rom includes articles on the following topics:

Globalization of Energy: Markets, Technology and Sustainability

Energy in a World of Changing Costs and Technologies

28th IAEE International Conference, Taipei, Taiwan, June 3-6, 2005 This CD-Rom includes articles on the following topics:
Global Warming and Energy Natural Gas (including LNG) Green and Renewable Energy Technology Liberalization and Market Power Restructuring and Deregulation Energy Pricing, Taxation, and Subsidy Renewable Energy and New Energy Oil and Coal: Today and Tomorrow International Electricity Research & Concerns Energy Statistics and Efficiency Indicators Global Warming and Energy Sustainable Energy Development Energy Modeling, Simulation, and Forecasting Energy Planning and Policy Options Conservation Know-how and R&D Distributive Energy Systems and New Fuels

26th USAEE/IAEE North American Conference, Ann Arbor, MI, September 2427, 2006 This CD-Rom includes articles on the following topics:
Transportation Vehicle Technologies Future Trends in Transportation Oil Market Security and Reliability Crunch Time for North American Natural Gas Science and Technology Policy Natural Gas Supply and Pricing Issues Investment Decisions Under Risk & Uncertainty Electricity Investment, Reliability, and Environmental Effects Regulatory or Market Economics: Which Really Maximizes Electric Utility Consumer Benefits? Energy, Economic Development & Energy Poverty The Global Oil Market: Issues and Trends Carbon Emissions and Energy Modeling Carbon Sequestration Modeling & Analysis

continued on next page

/ 193

IAEE/USAEE CONFERENCE PROCEEDINGS ORDER FORM

Energy, Environment and Economics in a New Era
24th USAEE/IAEE North American Conference, Washington, DC, July 8-10, 2004 This CD-Rom includes articles on the following topics:
Natural Gas Industry Energy Data and Modeling Demand Estimation Economics of Electric and Gas Utilities Economics of New Energy Technologies & Conservation International Energy Economics Energy and the Environment Energy Industry Finance Oil Industry: E&P, Transportation, Refining, etc. Regulation of Energy and Mineral Leasing Market Power in Deregulated Electricity Markets The Role of Solar Energy Sustainable Energy Climate Policy Uncertainty Electricity Reliability in the States OPEC Production and Reserves Gas Supply Security in Europe

Integrating the Energy Markets in North America: Issues & Problems, Terms & Conditions

23rd IAEE North American Conference, Mexico City, Mexico, Oct. 19-21, 2003 This CD-Rom includes articles on the following topics:
Oil Prices and Markets North American Energy Security & Reliability Climate Change Energy Efficiency as a Resource Hydrogen Economy Energy Trade and Transportation Transmission Issues in Electricity Industry Emissions and Energy Mexican Power Prospects for Green Power Retail Electricity Issues Gas & Power Convergence or Divergence NAFTA and Energy Restructuring Electricity Markets

IAEE/USAEE CONFERENCE PROCEEDINGS - ORDER FORM ********************************************************************************

To order, please send check (payable to IAEE in U.S. dollars, drawn on a U.S. bank) or credit card order to:
Proceedings Order Department, IAEE, 28790 Chagrin Blvd., Ste. 350, Cleveland, OH 44122 USA Phone: 216-464-5365 Fax: 216-464-2737 E-mail: iaee@iaee.org Method of payment: Check Credit Card

Name on Card (print):___________________________________________________________________________ Visa MasterCard

Number: __________________________________________________________ Exp. Date _________________

MEMBERS NON-MEMBERS

____ Developing & Delivering Affordable Energy in the 21st Century ____ From Restructuring to Sustainability: Energy Policies for the 21st Century ____ Energy in a World of Changing Costs and Technologies ____ Securing Energy in Insecure Times ____ Fueling the Future: Prices, Productivity, Policies and Prophecies ____ Globalization of Energy: Markets, Technology and Sustainability ____ Energy, Environment and Economics in a New Era ____ Integrating the Energy Markets in North American: Issues & Problems, Terms & Conditions Please send publication(s) to:

$130.00 $130.00 $130.00 $130.00 $130.00 $100.00 $100.00 $100.00

$180.00 $180.00 $180.00 $180.00 $180.00 $150.00 $150.00 $150.00

Name:_______________________________________________________________________________________ Company:____________________________________________________________________________________ Mailing Address: ______________________________________________________________________________ Country:_ ____________________________________________________________________________________ Phone: ___________________ Fax: _________________ E-mail:_ ______________________________________

194 / The Energy Journal

BOOK REVIEWERS NEEDED

Read a good book related to energy economics?
Why not tell everyone what you thought. The Energy Journal is seeking book reviews between 500-1000 words. Any energy economists interested in reviewing new books (in any language) for The Energy Journal are invited to send a onepage C.V. along with a letter listing their research interests to either one of the Journals book review editors: Dr. Richard L. Gordon, Professor Emeritus The Pennsylvania State University 214 Horizon Drive, State College, PA 16801 E-mail: rlg3@psu.edu Professor Carol A. Dahl, Director CSM/IFP Petroleum Economics & Management Colorado School of Mines, Division of Economics & Business Golden, Colorado 80401, USA Email: cadahl@Mines.edu Publishers, Academic Presses, and authors are invited to submit new energy economics books for review to the addresses above.

REQUEST FOR REFEREES for The Energy Journal

Any energy economist interested in refereeing research papers submitted to The Energy Journal are invited to send a brief C.V. and letter describing their research interests to: The Energy Journal International Association for Energy Economics 28790 Chagrin Blvd., Suite 350 Cleveland, OH 44122, USA Phone: 216-464-5365 Fax: 216-464-2737 Email: energy_journal@iaee.org Web site: http://www.iaee.org Topics regularly addressed include the following: technology change, energy efficiency, renewables, natural gas, electricity markets, oil demand/supply, climate change, nuclear power, energy taxation, regulation, OPEC, crude oil markets, economic modeling, energy and environmental issues, coal, energy futures, DSM.

/ 195

Information for Authors

The Energy Journal publishes theoretical and empirical papers devoted to the advancement and dissemination of knowledge concerning energy and related topics. Content and editorial board composition are international in scope. Although most articles published will emphasize the economic aspects of their subjects, coverage is interdisciplinary. The Energy Journal quarterly issues include: Major, refereed articles (25-35 pages) that deal with theoretical or applied problems of continuing interest in energy economics and related disciplines. Shorter, refereed papers and notes (10-15 pages) that are more topical and focus on the theory and practice of economics as related to current energy policy problems. An Energy Perspectives section which provides current readable (and refereed) papers on important issues. A review section examining books and topical government reports, national and international, which assess issues in energy policy and prospects. Submissions: Authors please submit online at http://www.iaee.org/en/publications/submit. aspx two electronic copies of your paper in PDF format; one with the authors names included and one with authors names removed. Articles will not be considered if multiple files are submitted e.g., text files along with separate files with graphics, tables, equations, etc. Each of the two electronic files must incorporate all text, tables, equations and graphics. Assurance should be made that the paper has not been submitted elsewhere for publication. Revised articles should be emailed directly to the Associate Editor at geoffrey.pearce@ utoronto.ca Authors submitting a manuscript do so on the understanding that if it is accepted for publication, copyright in the article, including the right to reproduce the article in all forms and media, shall be assigned exclusively to The Energy Journal. The Energy Journal will not refuse any reasonable request by the author for permission to reproduce any of his or her contributions to the journal. Manuscripts should be prepared according to the following style rules which are based on The Chicago Manual of Style. Form. Manuscripts should be double-spaced on one side of 8.5 x 11 inch (or similar) white paper. The first page should contain the article title, author(s) name(s) and complete affiliation(s), title(s), and complete mailing address of the person to whom all correspondence should be addressed. An abstract of no more than 150 words must be included. The manuscript with the names removed will be sent to anonymous referees for double-blind peer review. Math. Mathematical expressions should be set in italic type with all equations numbered consecutively on the right hand side of the page. Tables. Tables should be numbered consecutively. All tables should have concise titles. Footnotes. Footnotes should be numbered consecutively with superscript Arabic numerals. Footnotes are placed at the bottom of each page. Figures. All figures should be numbered consecutively throughout the text. Figures must be submitted as camera-ready copy or high resolution PDF format in a form suitable for reproduction. References. Please use the author-date citation system and include an alphabetical reference list of all works cited. Examples of reference list style are as follows: Chandler, Alfred (1977). The Visible Hand. Cambridge: Belknap Press. Federal Trade Commission, Bureau of Economics (1979). The Economic Structure and Behavior of the Natural Gas Production Industry. Staff Report. Henderson, J.S. (1986). Price Determination Limits in Relation to the Death Spiral. The Energy Journal 7(3): 150-200. Proofs. Page proofs will be sent to the corresponding author for approval before publication. Offprints. One set of 10 offprints without covers will be provided free of charge to the corresponding author. Additional article reprints may be purchased from IAEE. Book Reviews: Books for review should be sent to Dr. Richard L. Gordon, Professor Emeritus, The Pennsylvania State University, 214 Horizon Drive, State College, PA 16801, USA. Email: rlg3@psu.edu or Professor Carol Dahl, Colorado School of Mines, Division of Economics & Business, Golden, CO 80401, USA. Email: cadahl@mines.edu.

196 / The Energy Journal

SUBSCRIPTION ORDER FORM

THE ENERGY JOURNAL

Published by International Association for Energy Economics

OVERVIEW: The Energy Journal is published by the International Association for Energy Economics. It is
the Associations premier journal and is published four times a year. The journal encompasses a broad spectrum of energy economics topics such as: Alternative Transportation Fuels Hydrocarbon Issues Conservation of Energy International Energy Issues Economic Energy Modeling Issues in Exploration & Production Energy Demands and Usage Markets for Crude Oil Electricity and Coal Natural Gas Topics Energy & Economic Development Energy Management Energy Policy Issues Environmental Issues & Concerns Greenhouse Effect & Change Nuclear Power Topics Petroleum Topics Renewable Energy Issues Theory & Practice of Energy Related Issues Developing World Energy Issues

This professional journal carries a variety of other topics as well. Also included are topics of timely energy interest, research papers and reports, book reviews, government reports on energy policy and prospects and advertisements of interest to readers. This is one publication every learned library should have in its collection.

SUBSCRIPTION PRICE: U.S. and Canada subscriptions: $425.00; all other Countries: $450.00. METHOD OF PAYMENT:
Check payable to IAEE, enclosed. Note that checks must be drawn on a U.S. bank in U.S. dollars. Payment by wire transfer. Account Name: International Association for Energy Economics; Account Number: 108-551-4996; Bank Name: Key Bank National Association; Routing Number: 041001039; Bank City: Cleveland, Ohio, USA. Fax this subscription order form to 216-464-2737 upon transmission of your payment to this bank account. MasterCard Visa Card No. _______________________________________ Exp. Date: ____________

Name on Card: ________________________________________ Todays Date: ______/______/_________ Total Amount to be charged: $____________ Please invoice to the following attention. Signature: _______________________________________

SUBSCRIPTION ADDRESS:
NAME:________________________________________________________________________________________________________________ POSITION:_____________________________________________________________________________________________________________ COMPANY ____________________________________________________________________________________________________________ ADDRESS:_____________________________________________________________________________________________________________ ADDRESS:_____________________________________________________________________________________________________________ POSTAL CODE/COUNTRY: ______________________________________________________________________________________________ PHONE NUMBER:____________________________________________ FAX NUMBER: _____________________________________________

Send this form to:

International Association for Energy Economics The Energy Journal Subscription Dept. 28790 Chagrin Blvd., Ste. 350 Cleveland, OH 44122, USA

Phone: 216-464-5365 Fax: 216-464-2737 E-Mail: iaee@iaee.org

Institutional Membership
IAEE is very grateful for the support and involvement of our Institutional Members listed below. Institutional membership carries with it many special benefits. For a full listing of benefits and services please visit http://www.iaee.org/en/membership/institutional.aspx
Aramco Services Co., USA AREVA, France BP Plc, UK Center for Energy Marine Transportation and Public Policy and School for International and Public Affairs, Columbia University Council on Energy & Environment, Korea Curtin Business School, Perth, Australia GSE S.p.A, Italy Institute for Economics, Transport & Energy, Czech Republic Institut Francais Du Petrole, France Institute of Energy Economics, Tokyo National Energy Board, Alberta, Canada Pacific Isalnds Forum Secretariat, Fiji Islands Platts, USA Rice University, Baker Institute, USA Sandia National Laboratories, USA Shell International, UK University of Alberta, Canada University of Auckland Business School, New Zealand University of Georgia, USA University of Stavanger, Norway

ISSN 0195-6574 Energy Economics Education Foundation, Inc. 28790 Chagrin Blvd, Suite 350 Cleveland, OH 44122 USA

Non-Profit Org. U.S. POSTAGE PAID Lillington, NC Permit No. 32