Estimating fuel demand elasticities to evaluate CO2 emissions: Panel data evidence for the Lisbon Metropolitan Area

Transportation Research Part A 67 (2014) 30–46

Contents lists available at ScienceDirect

Transportation Research Part A journal homepage: www.elsevier.com/locate/tra

Estimating fuel demand elasticities to evaluate CO2 emissions: Panel data evidence for the Lisbon Metropolitan Area Patricia C. Melo a,⇑, Ahmad Razi Ramli b,c a

Social, Economic and Geographical Sciences, The James Hutton Institute, Craigiebuckler, Aberdeen AB15 8QH, United Kingdom Centre for Transport Studies, Dept of Civil and Environmental Engineering, Imperial College London, London SW7 2AZ, United Kingdom c Centre for Transport, Logistics and Operations Management, Faculty of Business Management, Universiti Teknologi MARA, 42300 Puncak Alam, Malaysia b

a r t i c l e

i n f o

Article history: Received 4 August 2012 Received in revised form 25 November 2013 Accepted 2 June 2014

Keywords: Fuel demand Road transport CO2 greenhouse gas emissions Lisbon Metropolitan Area

a b s t r a c t This paper estimates fuel demand models for the Lisbon Metropolitan Area (AML) and uses the demand elasticities obtained to predict future levels of road transport CO2 greenhouse gas emissions. Data for the municipalities constituting the AML and the period 1993–2010 are analysed using static and dynamic panel data models to measure the relative importance of fuel price, income, vehicle stock, the price of public transport, and the availability of urban and suburban rail networks on fuel demand. To the best of our knowledge, this is the first study in the Portuguese context to produce fuel demand elasticities for a specific metropolitan area, as opposed to the estimation of country-level aggregate elasticities. Our findings indicate that the elasticity of fuel demand with respect to fuel price ranges between 0.48 and 0.72 in the short run and between 1.19 and 1.82 in the long run. Income elasticities are found to range between 0.51 and 0.54 in the short run and between 1.26 and 1.37 in the long run. The elasticity of fuel demand with respect to vehicle stock (keeping population constant) is 0.57 in the short run and 1.43 in the long run. There is only weak evidence of a reduction in fuel demand as a result of a decrease in the price of public transport, and no effect of greater availability of rail networks. Based on the elasticities estimated, we predict road transport CO2 emissions for the AML according to different macroeconomic scenarios. The results indicate that the emissions target is only achieved in the scenario of poor economic performance. In the presence of medium and strong economic growth, fuel prices would need to increase by about 7% and 11% per year respectively in order to meet the emissions target. Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction Excessive road transport results in urban and environmental problems, including unsustainable levels of energy consumption and negative externalities such as congestion and greenhouse gases (GHG). The National Inventory Report on Greenhouse Gases produced by the Portuguese Environment Agency (Agência Portuguesa do Ambiente, 2011) reports that transport sources, of which 97% refer to road transport, represented 25.3% of total greenhouse gases emissions in 2009, while in 1990 the share was equal to 17%.1

⇑ Corresponding author. Tel.: +44 (0)12 2439 5316. 1

E-mail address: [email protected] (P.C. Melo). Total emissions excluding LULUCF (land use, land use change and forestry) on a carbon equivalent basis.

http://dx.doi.org/10.1016/j.tra.2014.06.001 0965-8564/Ó 2014 Elsevier Ltd. All rights reserved.

P.C. Melo, A.R. Ramli / Transportation Research Part A 67 (2014) 30–46

31

225 Road transport GHG emissions Road transport CO2 GHG emissions 200 Total GHG emissions CO2 GHG emissions

Index GHG emissions (1990=100)

175

150

125

100

75

50 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 Fig. 1. Evolution of total and road transport GHG emissions in Portugal (1990 = 100). Source: Authors based on figures reported in the Portuguese National Inventory Report on Greenhouse Gases produced by the Portuguese Environment Agency (Agência Portuguesa do Ambiente, 2011).

The evolution of road transport and total GHG emissions is depicted in Fig. 1. It can be observed that road transport emissions have increased at a much faster rate than total emissions.2 Total and CO2 related GHG emissions in Portugal increased by 26% and 28% between 1990 and 2009 respectively. Road transport total GHG emissions increased by 84% between 1990 and 2009, while CO2 related GHG emissions (about 99% of total road transport emissions) increased by 95%. The strongest increase in road transport GHG emissions was registered during the 1990s, with maximum value in 2002, and was mainly a result of the growth in car ownership and road traffic levels, supported by large investments in road infrastructure during the 1990s (Agência Portuguesa do Ambiente, 2011). Since 2002, road transport emissions have remained stable and even reduced slightly in 2008 and 2009. As part of the Kyoto Protocol, Portugal committed to limit its total GHG emissions for the period 2008–2012 to no more than 27% above the level of emissions registered in 1990 (Agência Portuguesa do Ambiente, 2011). Although this target appears to be achievable for total GHG emissions, road transport emissions are well above the agreed target. Moreover, because road transport represents 25% of all GHG emissions it is crucial to ensure that road transport GHG emissions are kept at sustainable levels. The new Transport White Paper sets a European target of reducing GHG emissions of the transport sector by at least 20% and 60% until 2020 and 2050 respectively, compared to 1990 levels (EC, 2011). To help promote the development of sustainable urban mobility measures and the reduction of levels of car dependence, the European Commission (EC) issued an Action Plan on Urban Mobility in 2009 which contains several transport, and nontransport, related measures to be implemented by the various local, regional, and national authorities. Some of the transport related measures refer to changes in modal share in favour of public transport modes, fuel efficiency improvements in private and public transport fuel powered vehicles, old car scrappage schemes, tax systems favouring more fuel efficient vehicles and more responsible driving patterns. In order to design policies aiming to create sustainable cities, it is important to understand which factors can be used to change behaviour towards more sustainable choices. This can be achieved through the estimation of fuel demand elasticities in urban contexts. The sort of questions we hope to answer with this analysis can be formulated as follows. Can fuel taxation play a significant role in curbing fuel consumption and CO2 emissions? By how much more would fuel prices have to increase in order to act against the positive effect of rising incomes on fuel demand? Is there scope for reducing fuel emissions by promoting a shift from (predominantly fuel based) private modes of transport to electricity based urban and suburban transit systems? In this paper we estimate fuel demand elasticities for the Lisbon Metropolitan Area (AML) and predict road transport CO2 emissions for the AML according to different macroeconomic scenarios. The objectives of the analysis are to provide a better 2

Total emissions refer to the following sectors: energy (including transport), industrial processes, solvent use, agriculture, and waste.

32

P.C. Melo, A.R. Ramli / Transportation Research Part A 67 (2014) 30–46

understanding of fuel consumption in the largest metropolitan region of Portugal and to contribute to the design of policy instruments aimed at improving the sustainability of urban and regional mobility patterns. This paper makes two contributions. It generates new empirical evidence on fuel demand elasticities for the Lisbon Metropolitan Area using a dataset which has been compiled for the first time and specifically for this analysis. The second contribution consists of the design of a comprehensive model specification which accounts simultaneously for the role of the price and the availability of public transport on fuel demand. Although some of the previous studies have considered the effect of public transport prices on fuel demand (e.g. Crôtte et al., 2010), we are not aware of previously published studies controlling for both the price and the availability of public transport simultaneously. We estimate short-, medium- and long-run fuel demand elasticities with respect to fuel price, income, vehicle stock, availability of urban and suburban rail networks, and the price of public transport. Considerable attention and discussion is dedicated to the appropriateness of the various panel data estimators in the context of the estimation of static and dynamic demand models and the structure of our dataset. Based on the fuel demand elasticities, it is possible to construct policy scenarios to evaluate future levels of fuel consumption and associated CO2 GHG emissions. This requires making assumptions about the growth of average real incomes, vehicle stock, and fuel prices. The structure of the paper is as follows. Section 2 provides a brief overview of existing empirical evidence on fuel demand elasticities. Section 3 describes the Lisbon Metropolitan Area and the data available for this study. Section 4 presents the empirical model specification and econometric techniques used to obtain fuel demand elasticities from panel data. Section 5 reports and discusses the main results, while Section 6 provides an illustrative exercise of how the estimated fuel demand elasticities can be used to predict future levels of road transport CO2 emissions. Finally, Section 7 provides a summary of the main conclusions. 2. Overview of empirical evidence There are a number of comprehensive review articles that evaluate the extensive empirical evidence on the price and income elasticities of fuel demand. The following paragraphs, and Table 1, provide a summary of the range of values for the short-run and long-run elasticities reported in recent surveys (e.g. Espey, 1998; Graham and Glaister, 2002a,b; Goodwin et al., 2004; Basso and Oum, 2007; Brons et al., 2008) and meta-analyses (e.g. Espey, 1998; Hanly et al., 2002; Brons et al., 2008) of the literature. We also consider existing evidence on price and income elasticities of fuel consumption for Portugal. Although there has been extensive research in fuel demand over the last decades, there is only very limited evidence for Portugal. To our knowledge, only one of the studies reviewed in previous survey articles and meta-analyses reports elasticity estimates for Portugal. The only other source of evidence comes from a government document published in 2001 (De Oliveira, 2001). Fuel demand is usually measured in terms of total fuel consumption (Virley, 1993; Samimi, 1995; Al-faris, 1997; Li et al., 2010), fuel consumption per capita (Polemis, 2006; Hughes et al., 2008; Liddle, 2009; Park and Zhao, 2010), or fuel consumption per vehicle (Baltagi et al., 2003; Crôtte et al., 2010; Pock, 2010). The most common model specification accounts for the effect of fuel price and average income level, while some studies have also considered the role of vehicle stock, the price of possible fuel substitutes (e.g. gasoline versus diesel), and other travel related characteristics such as the availability and cost of public transport. Although the exclusion of some of these variables can lead to omitted variable bias, their inclusion can also prevent researchers from capturing the more long term response of fuel consumption to changes in fuel price and/or income. For example, models that include vehicle stock in the basic fuel demand model specification are likely to measure more short-run responses to changes in fuel price and hence fail to capture the more long-run effects deriving from changes in vehicle ownership and use. This is specially the case in models that also account for the role of income on fuel demand, as part of the income effect takes place through changes in vehicle stock. Similarly, the income elasticity will also tend to capture only short-run effects related to increased fuel consumption per vehicle.

Table 1 Summary of gasoline demand elasticities. Review article Espey (1998) Hanly et al. (2002) Graham and Glaister (2002b) Graham and Glaister (2002a, 2004) Goodwin et al. (2004) Basso and Oum (2007) Brons et al. (2008)

Price elasticity Short run Long run Short run Long run Short run Long run Short run Long run Short run Long run Short run Long run

0.23 (median); 0.43 (median); 0.25 (mean) 0.64 (mean) 0.2 to 0.3 0.6 to 0.8 0.21 (median); 0.55 (median); 0.25 0.6 0.2 to 0.3 0.6 to 0.8 0.53 (mean)

Income elasticity 0.26 (mean) 0.58 (mean)

0.25 (mean) 0.77 (mean)

0.39 (median); 0.81 (median); 0.39 (mean) 1.08 (mean) 0.35 to 0.55 1.1 to 1.3 0.42 (median); 0.91 (median); 0.4 1.0 0.3 to 0.5 0.9 to 1.3 –

0.47 (mean) 0.88 (mean)

0.47 (mean) 0.93 (mean)

P.C. Melo, A.R. Ramli / Transportation Research Part A 67 (2014) 30–46

33

Overall, the range of elasticity values reported in the recent review articles is consistent across studies. This appears to be especially true for price elasticities. Fuel demand elasticities are generally higher in the long run than in the short run because the variety of responses to changes in fuel price and income is wider in the long term (Espey, 1998; Graham and Glaister, 2002b, Hanly et al., 2002; Graham and Glaister, 2004; Brons et al., 2008). In the short run people adjust to rising fuel prices not necessarily by driving less, but by adopting a more fuel efficient driving behaviour and/or improving car maintenance. In the long run, people may choose to move to more efficient cars, and/or adopt alternative travel patterns based on public transport or different home/work locations. Similarly, income elasticities are also higher in the long run because of changes in car ownership and vehicle stock composition. Another consensual result is that price elasticities tend to be lower than income elasticities, sometimes by a large factor. As a result, to reduce fuel consumption levels fuel prices need to increase at a faster rate than income. The review of empirical estimates of price and income elasticities of fuel demand conducted by Graham and Glaister (2002b) reports that short-run price elasticities range between 0.2 and 0.3, while long-run price elasticities range between 0.6 and 0.8. Income elasticities are reported to range between 0.35 and 0.55 in the short run and between 1.1 and 1.3 in the long run. Based on an update of this survey, Graham and Glaister (2002a) summarise 600 price elasticities of fuel demand and 483 income elasticities of fuel demand obtained from 113 studies published between 1966 and 2000. The mean short-run (long-run) price elasticity equals 0.25 (0.77), while the mean short-run (long-run) income elasticity equals 0.47 (0.93). Goodwin et al. (2004) provide a summary of 175 estimates for the period between 1929 and 1998. The survey reports that studies using a dynamic model specification based on panel data tend to produce lower elasticity estimates. The average value of the price elasticity estimates included in the review is 0.25 for the short run and 0.6 for the long run. The average value of the short- and long-run income elasticity is 0.4 and 1.0 respectively. Basso and Oum (2007) provide a survey of the literature with a special attention to the role of the different empirical approaches and methods. They report values for fuel price elasticities between 0.2 and 0.3 in the short run, and between 0.6 and 0.8 in the long run. As for income elasticities, their values tend to fall between 0.3 and 0.5 in the short run and between 0.9 and 1.3 in the long run. Espey (1998), Hanly et al. (2002), and Brons et al. (2008) conduct meta-analyses of the empirical evidence to explain the causes of variation among estimates of price and income elasticities of fuel consumption. Espey (1998) examines a sample of 640 price elasticities of gasoline consumption and 590 income elasticities of gasoline consumption obtained from 101 studies published between 1966 and 1997. The mean (median) short-run and long-run price elasticities are equal to 0.26 (0.23) and 0.58 (0.43) respectively. The mean (median) short-run and long-run income elasticities are equal to 0.47 (0.39) and 0.88 (0.81) respectively. Hanly et al. (2002) review a sample of 491 fuel demand elasticities obtained from 69 studies covering the period from 1929 to 1991. Of the 491 fuel demand related elasticities, 121 are price elasticities of total fuel consumption and 115 are income elasticities of total fuel demand. The mean short-run (long-run) price and income elasticities are 0.25 (0.64) and 0.39 (1.08) respectively. Brons et al. (2008) consider a sample of 312 estimates of fuel consumption obtained from 43 studies, of which 158 are price elasticities of total gasoline demand. The mean price elasticity of total gasoline consumption is 0.53. Elasticity estimates may also be dissimilar due to the contextual characteristics inherent in a particular study. The usual differentiating factors include the spatial (e.g. city, state or national level data) and temporal (time period) aspects involved. For example, the magnitude of the price and income elasticities are generally found to differ according to the country studied. Lower price elasticities have been reported by recent review articles for the US, Canada, and Australia, compared to higher estimates for European countries (e.g. Espey, 1998; Hanly et al., 2002; Brons et al., 2008). Recent studies also found that fuel elasticities are also subject to change due to behavioural and structural factors. For example, Hughes et al. (2008) found a shift in the fuel demand elasticities in the US between two different time periods that they attribute to land-use and social factors. Similarly, Dargay et al. (2007) and Dahl (2012) suggest that there may be evidence to indicate that travel saturation (as income increases) may explain the decline of income elasticities in developed countries as compared to developing countries. Including measures of vehicle ownership and/or vehicle characteristics (e.g. fuel efficiency) in the model specification tends to produce lower short-run price elasticity estimates. Similarly, the inclusion of measures of vehicle ownership also results in lower income elasticity estimates, both for the short and long run. While the exclusion of vehicle stock can produce upward biased estimates of the income elasticity, including it makes the model unable to identify the indirect effect of income (through changes in vehicle stock) on fuel demand. A similar argument can be made with respect to the exclusion of fuel efficiency in the model specification. Some of the review articles have also decomposed the price elasticity of gasoline consumption into the price elasticities of fuel efficiency, mileage per car, and car ownership (Hanly et al., 2002; Graham and Glaister, 2002a; Brons et al., 2008). Brons et al. (2008) found that changes in gasoline consumption due to changes in price are mainly driven by changes in fuel efficiency and car ownership, while Hanly et al. (2002) estimated a stronger influence of changes in mileage per car and car ownership, and Graham and Glaister (2002a) reported stronger responses in fuel efficiency. We only found two studies reporting fuel demand elasticities for Portugal. Sterner et al. (1992) estimate dynamic gasoline demand models for 20 OECD countries for the period 1960–1985 and obtain short-run (long-run) price and income elasticities for Portugal equal to 0.13 (0.67) and 0.37 (1.93) respectively. In a more recent study, De Oliveira (2001) uses time series cointegration techniques to estimate price and income elasticities of gasoline and diesel consumption for

34

P.C. Melo, A.R. Ramli / Transportation Research Part A 67 (2014) 30–46

Portugal during the period from 1977 to 2000. The short-run (long-run) elasticity of price for gasoline demand is around 0.4 (0.75 to 0.9), while the short-run (long-run) income elasticity is 0.32 (0.65–0.96). The author also experimented including (gasoline powered) vehicle sales and vehicle stock per capita in the model specification together with gasoline price and income but could not identify a statistically significant effect. The short- and long-run elasticities of diesel consumption with respect to price are lower and equal 0.07 and 0.17 respectively. The income elasticities tend to be higher and are equal to 0.63 and 1.45 for the short and long run respectively. As for the role of (diesel powered) vehicle stock per capita the short- and long-run elasticities are estimated to be 0.10 and 0.24 respectively. The price elasticities of gasoline consumption obtained by Sterner et al. (1992) are relatively more inelastic than those estimated by De Oliveira. It is difficult to explain the reasons underlying the differences in the value of the price elasticities because the studies differ in many dimensions, including the type of data, the period of the analysis, the model specification, and the statistical estimators. One possible explanation can relate to differences in the availability of public transport in the periods analysed by Sterner et al. and De Oliveira, which may influence the responsiveness of demand to increases in fuel prices. The short-run income elasticities estimated in both studies are similar and in line with the ranges reported in the recent review articles, although the long-run income elasticity produced by Sterner et al. (1992) appears to be overvalued.

3. Data and variables In this section we provide a detailed discussion of the data and variables used in the estimation of fuel demand for the Lisbon Metropolitan Area. The AML, shown in Fig. 2, is divided by the river Tagus into two areas, AML-North and AML-South, which are connected by three bridges. According to the Portuguese Office for National Statistics (INE), the land area of the AML accounts for about 3% of continental Portugal territory and nearly 3 million inhabitants (2007), representing more than

Fig. 2. View of the Lisbon Metropolitan Area (AML) and its relative size in Portugal (top right corner). Source: Google maps.

35

P.C. Melo, A.R. Ramli / Transportation Research Part A 67 (2014) 30–46 Table 2 Summary statistics. Variable

Units

Obs.

Mean

Median

SD

Min

Max

Fuel Fuel/Vstock Fuel/Pop Pfuel/CPI Vstock Vstock/Pop RInc/Pop PTprice/CPI Rail/Pop

Kl (000 litres) Kl per vehicle Kl per capita €/ltr Number Number € € Per 10,000 people

306 306 306 306 306 306 306 306 306

109,000 1.57 0.757 0.63 82,401 0.49 622 34.50 0.31

74,509 1.33 0.636 0.60 55,792 0.46 594 30.08 0.29

137,000 0.99 0.47 0.08 96,618 0.13 132 13.55 0.30

4900 0.52 0.180 0.52 3037 0.19 411 14.27 0.00

1,180,000 7.23 3.296 0.80 491,000 0.91 1063 75.38 1.31

25% of the total population. The AML is the largest economic region of Portugal and accounts for 33% of total employment and more than 36% of the GDP. The main economic centre of the region is Lisbon, which is also the capital of the country. The dataset includes data on fuel demand and related variables for a balanced panel of 17 municipalities for the period from 1993 to 2010.3 Fuel demand (in litres) refers to both gasoline and diesel and was obtained from the AML’s Observatory for Economic and Social Development (ODES/AML). Fuel price data (in Euros) were obtained from the Directorate-General for Energy and Geology (DGEG) of the Ministry of Economy, Innovation and Development. Prices were obtained separately for diesel and gasoline and a weighted fuel price was computed by taking into account the levels of diesel and gasoline consumption. Before 2004 fuel prices were set by the government and did not vary across municipalities. Although fuel prices have been freely set since 2004, we observe that there is little variation in average prices across municipalities (see Table 3). Data for population were obtained from INE, and (monthly) average income data were obtained from the Ministry of Labour and Social Solidarity (MTSS). Inflation data based on the Consumer Price Index (CPI) were also obtained from INE to calculate real income and real price data. Real incomes and prices are expressed in terms of 1993 values. Fig. 3 shows average values of several fuel related variables for the AML during the period 1993–2010. The variables displayed in the graph refer to total fuel consumption, fuel consumption per vehicle, vehicle stock per capita (car ownership), real fuel price, real average monthly income, and real price of the monthly travelcard to/from Lisbon. The values are indexed to 100 in 1993. The variable with the strongest increase during the period under analysis is car ownership. The highest value was registered in 2002, and was 49% higher than in 1993. There was a very strong reduction in car ownership levels in 2003 as a result both of the international and internal economic crises of 2002–2003. According to the Bank of Portugal, the annual growth rate of GDP per capita for the Portuguese economy was of 0.03% and 1.6% in 2002 and 2003 respectively. Between 2004 and 2008 car ownership levels increased gradually, although remaining below 2002 levels, after which they fell again as a result of the economic crisis of 2008. An increase of 12% in car ownership was registered in 2010, in spite of the weak economic growth. This increase happened mainly for two reasons, both related to announcements made by the government of changes in fiscal regimes affecting the cost of buying a car in 2011. The first was the announcement of the end of the old car scrappage scheme implemented in 2000 which offered fiscal discounts in the acquisition of new cars.4 The second was the announcement of the increase of the VAT rate from 21% to 23%. Both factors motivated consumers who were considering buying a new car to do it before the start of 2011. Similarly to car ownership, total fuel consumption has also increased considerably during this period. The maximum values were registered in 2001 and 2004 (45% and 47% respectively compared to 1993 levels). Since 2004, fuel consumption levels have been falling very significantly, with exception of 2009 during which fuel prices were reduced in real terms. Fuel consumption per vehicle has either been decreasing or remained similar to 1993 levels. The maximum level of fuel consumption per vehicle was registered in 2004, but has been falling considerably thereafter. In 2010 it accounted for 55% and 63% of the level registered in 2004 and 1993 respectively. Two effects help understand this trend. First, as the number of cars per household increases, there is a reduction in the utilisation of each vehicle (i.e. doubling the number of cars owned by a household less than doubles the amount of kilometres driven). Second, increased fuel efficiency makes fuel consumption more productive (i.e. same kilometres can be covered with less fuel). We now turn to fuel price. The average price of fuel remained below 1993 levels, in real terms, until 2004 (fuel prices were regulated by the government until late 2003). Between 2004 and 2008, real fuel prices increased by a total of 27% (also 27% higher than in 1993). The price of fuel fell by about 15% in 2009, while increasing back to 2008 levels in 2010.5 It can also be observed that between 2004 and 2010 there is a strong negative relationship between fuel consumption per vehicle and fuel prices. Although the (real) price of gasoline has been greater than the (real) price of diesel, there has been a convergence in price levels since 2004. In 1993 the cost of one litre of diesel was about 69% of the price of one litre of gasoline, while in 2008 3 The municipality of Odivelas was created in 1998 and was previously part of Loures. To ensure data consistency we merged data for Odivelas with the municipality of Loures from 1999. 4 Programa de Incentivos ao Abate de Veículos em Fim de Vida (VFV). 5 Although not shown in the graph, fuel prices have continued to increase steadily since 2010.

36

P.C. Melo, A.R. Ramli / Transportation Research Part A 67 (2014) 30–46

Table 3 Variation in data (N = 17, T = 1993–2010). SD

Fuel/Vstock

Fuel/Pop

Pfuel/CPI

Vstock

Vstock/Pop

RInc/Pop

PTprice/CPI

Rail/Pop

Overall Between Within

0.99 0.80 0.61

0.47 0.38 0.29

0.08 0.02 0.08

96,617.54 97,748.48 17,684.59

0.13 0.11 0.07

131.59 120.06 60.86

13.55 13.27 4.18

0.30 0.28 0.13

SD – standard deviation.

175 Total Fuel Consumption Fuel Consumption per Vehicle Vehicle Stock per capita Real Average Price of Fuel 150

Real Average Monthly Income

Index (1993=100)

Real Average Travelcard to/from Lisbon

125

100

75

50 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 Fig. 3. Total fuel consumption, fuel consumption per vehicle, real price of fuel, vehicle stock per capita, real average income, and real price of monthly travelcard to/from Lisbon in AML (1993=100).

(2010) the ratio increased to 91% (84%). Unfortunately, data issues prevented us from studying the effect of the shift from gasoline to diesel vehicles on fuel consumption and CO2 emissions.6 This shift is reflected in the changes in diesel and gasoline consumption during the period. In 2010 total diesel consumption was 47% higher than in 1993, while total gasoline consumption was only 60% of 1993 gasoline consumption levels. Turning to income levels, it can be observed from Fig. 3 that average real incomes have risen progressively during the period under analysis. In 2009 (no data available for 2010 yet) the average monthly real income per capita was 27% higher than in 1993. Rising income levels are generally strongly correlated with increasing car ownership levels and vehicle kilometres driven, both increasing fuel consumption. The data displayed in Fig. 3 appears to indicate that the positive effect of rising real income levels is stronger for car ownership than fuel consumption per vehicle, although fuel efficiency and rising fuel prices can also help explain the negative trend in fuel consumption per vehicle. The use of public transport, particularly electricity based transport modes such as urban and suburban rail systems, can affect fuel consumption negatively. We consider the evolution of the average cost incurred by passengers using the public transport system to/from Lisbon. Fig. 3 shows the average price (in real terms) of the monthly travelcard to/from Lisbon in the AML. The price of monthly travelcards remained below 1993 levels until 2000 and remained at that level until 2002, increasing gradually thereafter. The real value of the monthly travelcard in 2010 was on average 25% higher than in 1993 (and 2000).7

6 In order to investigate the importance of differences in gasoline and diesel prices and associated changes in the composition of fuel consumption we need separate data for gasoline and diesel vehicle stock for the areas under study. Unfortunately this information was not available at the time of this study. 7 In 2011 and 2012 monthly travelcards increased in some cases by 25–50% as part of the austerity measures required by the International Monetary Fund (IMF), European Central Bank (ECB), and the European Commission (EC).

37

P.C. Melo, A.R. Ramli / Transportation Research Part A 67 (2014) 30–46

Besides its cost, the availability and quality of public transport networks can also affect fuel consumption levels. We measure the accessibility to public transport networks for both the urban and suburban rail systems in the AML. There are two suburban commuter rail systems, operated by CP – Comboios de Portugal and Fertagus. CP – Comboios de Portugal operates mainly in the AML-North and has very limited (and infrequent) services in the AML-South. While the municipalities in the AML-North have been connected to Lisbon for several decades, the municipalities in the AML-South have only recently become connected to Lisbon across the river Tagus with Fertagus (operations started in 1999 for a limited number of municipalities and were fully extended in 2004). There is also a metro system operated by Metropolitano de Lisboa (ML), with services predominantly within the municipality of Lisbon (in 2004 the system was extended to two neighbouring municipalities). In 2007 a new light rail system (MTS – Metro Transportes do Sul) started operations in the AML-South. The services take place predominantly within the municipality of Almada while also reaching some parts of Seixal. The metro (ML) and light rail (MTS) networks are depicted in Fig. 2 above. Table 2 summarises the variables used in the empirical analysis. The table reports the mean, median, standard deviation (SD), minimum and maximum values for the variables described above. Table 3 summarises the degree of overall, betweenmunicipality and within-municipality variation in the data. There is stronger between-municipality variation for the dependent variable, vehicle stock, real average income, price of public transport travelcard, and the availability of urban and suburban rail networks. In contrast, there is stronger within-municipality variation in the data for fuel price. Table 4 reports the coefficients of pairwise correlation between the variables used in the fuel demand models. We observe the following relationships: average real income levels are positively correlated with total fuel consumption but there is almost no correlation between average real income and fuel consumption per capita; real average fuel prices are negatively associated with fuel consumption per vehicle and fuel consumption per capita; car ownership (vehicle stock per capita) is positively correlated with total fuel consumption and fuel consumption per capita, whilst there is a weak and negative correlation with fuel consumption per vehicle; average real price of public transport is positively associated with fuel consumption per vehicle and fuel consumption per capita but negatively correlated with total fuel consumption; the availability of public transport (per 10,000 people) is negatively associated with fuel consumption per vehicle and fuel consumption per capita but not with total fuel consumption; finally, the pairwise correlation between average income and vehicle stock (0.55) suggests that multicollinearity is not a problem in our dataset. 4. Empirical methodology This section presents the model specification and discusses the model estimation of the panel data fuel demand model for the Lisbon Metropolitan Area. We estimate both static and dynamic fuel demand models, which produce both medium-run and short- and long-run elasticity estimates respectively. 4.1. Model specification The choice of model specification used was greatly informed by the discussions on model specification presented in some of the recent survey and review papers of the empirical literature (e.g. Espey, 1998; Goodwin et al., 2004; Basso and Oum, 2007). We follow Baltagi and Griffin (1997) and Baltagi et al. (2003) model specification of demand for fuel consumption, while also considering the effect of the price and availability of public transport on fuel demand. We define fuel consumption per vehicle (Fuel/Vstock) to be a log-linear function of the real price of fuel (Pfuel/CPI), real average income (RInc/Pop), and vehicle stock per capita (Vstock/Pop). In addition to this baseline model, we also consider the effect of the price of the public transport travelcard to/from Lisbon (PTprice/CPI) and the availability of rail based (i.e. urban and suburban) public transport networks (Rail/Pop). Eq. (1) below illustrates the more comprehensive model specification of the fuel consumption model, including the two public transport variables.



  b  b  b  b  b Fuel Pfuel 1 RInc 2 Vstock 3 PTprice 4 Rail 5 ¼a Vstock CPI Pop Pop CPI Pop

ð1Þ

Table 4 Correlation matrix. Variable (1) (2) (3) (4) (5) (6) (7) (8) (9)

Fuel Fuel/Vstock Fuel/Pop Pfuel/CPI Vstock Vstock/Pop RInc/Pop PTprice/CPI Rail/Pop

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

1 0.07 0.15 0.01 0.93 0.55 0.51 0.41 0.48

1 0.92 0.14 0.21 0.06 0.01 0.23 0.19

1 0.07 0.00 0.30 0.16 0.18 0.06

1 0.10 0.17 0.31 0.14 0.21

1 0.58 0.55 0.46 0.50

1 0.50 0.11 0.30

1 0.18 0.52

1 0.17

1

38

P.C. Melo, A.R. Ramli / Transportation Research Part A 67 (2014) 30–46

To allow for adjustment in fuel consumption per vehicle to desired fuel consumption per vehicle over time we also adopt a flow adjustment model assumed to follow a first-order process (e.g. Baltagi and Griffin, 1997; Baltagi et al., 2003).



"



Fuel Vstock t  Fuel  Vstock t1

¼

 #h

Fuel  Vstock t  Fuel  Vstock t1

;0 < h < 1

ð2Þ

Substituting Eq. (2) in Eq. (1) and log-linearizing it using pooled data for the AML municipalities i and time t (year), we obtain the dynamic demand equation for fuel consumption per vehicle:

ln



         Fuel Fuel Pfuel RInc Vstock ¼ h ln a þ ð1  hÞ ln þ hb1 ln þ hb2 ln þ hb3 ln Vstock i;t Vstock i;t1 CPI i;t Pop i;t Pop i;t     PTprice Rail þ ei;t þ hb4 ln þ hb5 ln CPI Pop i;t

ð3Þ

where ei,t is the disturbance term which is commonly specified as a two-way error component model:

ei;t ¼ li þ kt þ ui;t ; i ¼ 1; . . . ; I; t ¼ 1; . . . ; T

ð4Þ

where li is a municipality-specific effect, kt is a time-specific effect, and ui,t is the random independent and identically distributed (i.i.d.) error term, which can be relaxed to accommodate heteroskedasticity, within municipality correlation, and serial correlation. Both short-run and long-run elasticities can be obtained from Eq. (3). The short-run elasticities of fuel consumption per vehicle with respect to real fuel price, real average income, vehicle stock per capita, real public transport travelcard, and availability of urban and suburban rail networks are hb1, hb2, hb3, hb4, and hb5 respectively. The long-run elasticities of fuel consumption per car with respect to real fuel price, real average income, vehicle stock per capita, real public transport travelcard, and availability of rail networks are b1, b2, b3, b4, and b5 respectively, with speed of adjustment to the long-run equilibrium equal to (1  h). Note that vehicle stock enters the model both as part of the dependent variable and as an explanatory variable. Thus, the appropriate short-run and long-run elasticities of total fuel consumption with respect to vehicle stock are (1 + hb3) and (1 + b3) respectively. We also consider a model specification based on fuel consumption per capita (Fuel/Pop) instead of fuel consumption per vehicle (Fuel/Vstock). In summary, we estimate static and dynamic fuel demand models using both fuel consumption per vehicle (Fuel/Vstock) and fuel consumption per capita (Fuel/Pop) as the dependent variable. The following paragraphs provide a brief discussion of the main estimation issues and the statistical estimators used. 4.2. Model estimation We estimate the static fuel demand models using the pooled OLS (OLS), the random-effects (RE) and the fixed-effects (FE) estimators. In the presence of correlation between the municipality-specific effect (and hence the error term) and the covariates, only the FE produces consistent estimates. However, in the absence of correlation, both the OLS and the RE estimators provide consistent parameter estimates, but only the RE estimator provides efficient parameter estimates. To assess the appropriateness of the RE or the FE model a Hausman test (Hausman, 1978) of the null hypothesis of no correlation between the individual effects and the covariates can be performed. Not accounting for the presence of time-dependent serial correlation in the residuals can lead to inefficient parameter estimates of the regression coefficients (Baltagi, 2008). To test for the presence of serial error autocorrelation, we use the modified Bhargava, Franzini, and Narendranathan (BFN) Durbin–Watson test with null hypothesis of serial independence (Bhargava et al., 1982). To accommodate first-order serial correlation in the error term we re-estimate the RE and FE models with the appropriately modified error structure; these models are named RE-AR(1) and FE-AR(1) respectively. In the context of the estimation of dynamic demand models neither of the standard panel data estimators (i.e. the random-effects and the fixed-effects estimators) can ensure consistent parameter estimates. The main issue to be addressed is the correlation between the lagged dependent variable and the error term, which cannot be removed through the use of a fixed-effects (or within-groups) estimator. The bias of the fixed-effects estimator does not disappear with increases in the number of cross-sectional units, and hence the estimator is inconsistent for samples with large N and small T. As the time dimension T increases the estimator becomes consistent (Baltagi, 2008). The Generalised Method of Moments (GMM) can offer a means of obtaining consistent parameter estimates. The basic idea is to construct a set of valid instruments based on the time series nature of the dataset, which are correlated with the covariates but uncorrelated with the error term. Arellano and Bond (1991) and Blundell and Bond (1998) proposed two different dynamic GMM estimators, the differenceand system-GMM respectively. The difference-GMM uses first-differences to remove unobserved time-invariant individualspecific effects, and then instruments the lagged dependent variable in the first-differenced equation using levels of the series lagged two periods or more, under the assumption that the time-varying disturbances in the original levels equations are not serially correlated. The system-GMM combines the standard set of equations in first-differences with suitably lagged levels as instruments, with an additional set of equations in levels with suitably lagged first-differences as instruments. In

39

P.C. Melo, A.R. Ramli / Transportation Research Part A 67 (2014) 30–46

the presence of data with little variation over time, the system-GMM estimator has been shown to be preferred to the difference-GMM, on the grounds that it can provide increased efficiency and less finite sample bias (e.g. Arellano and Bover, 1995; Blundell and Bond, 1998; Blundell and Bond, 2000). In order to produce consistent parameter estimates, the dynamic GMM estimators need to conform to two main criteria. There should be no first-order serial autocorrelation in the errors of the level equation, and the set of instruments used should be uncorrelated with the residual term. To assess the presence of serial correlation we use the Arellano and Bond serial autocorrelation tests with null hypothesis of no second-order serial correlation in the first differenced residuals, implying that the errors from the levels equations are serially uncorrelated (Arellano and Bond, 1991). To evaluate instrument exogeneity, we consider the Hansen test of overidentifying restrictions, which tests the null hypothesis that instruments are orthogonal to the error term (Hansen, 1982). In the presence of samples with moderate or small cross-sectional dimension (N), the properties of the dynamic GMM estimators may no longer hold (e.g. Arellano and Bond, 1991; Blundell and Bond, 1998; Bun and Kiviet, 2006). In such cases, a bias-corrected within-groups estimator (LSDVc) may be a better alternative if the length of the time dimension (T) is long. Judson and Owen (1999) conducted Monte Carlo simulations to compare different estimators of dynamic models for macro type panels of various cross-sectional (N = 20, 100) and time (T = 5, 10, 20, 30) dimensions. Their results suggest (for balanced panels) that this estimator performs at least as well as GMM based alternative estimators when T = 30. For T 6 10 and T = 20 the results favoured the use of the LSDVc estimator (see Judson and Owen, 1999). The length of the cross-sectional and time dimensions of our sample is very close to 20 so we decided to also estimate the fuel demand model using the LSDVc technique. We used the LSDVc estimator based on a bias approximation term up to order T1 and initial values as obtained from the two consistent dynamic GMM estimators: Arellano and Bond (1991) first-difference GMM estimator, and the Blundell and Bond (1998) system-GMM estimator.

5. Results and discussion This section presents and discusses the main results obtained from the estimation of the static and dynamic fuel demand models for the Lisbon Metropolitan Area, as described in the previous section. The static fuel demand models were estimated using the following estimators described in the previous section: OLS, RE, FE, RE-AR(1), and FE-AR(1). These standard panel data estimators produce inconsistent model parameter estimates for the dynamic fuel demand model because they do not correct for the correlation between the lagged dependent variable and the error term in Eq. (3). To address the issue of endogeneity of the lagged dependent variable we use the two dynamic GMM estimators described in the previous section – Arellano and Bond (1991) first-difference GMM estimator (Diff-GMM) and the Blundell and Bond (1998) system-GMM estimator (SYS-GMM) –, as well the bias-corrected fixed-effects estimator using initial values obtained from the difference-GMM estimator (LSDVc-AB) and the system-GMM estimator (LSDVc-BB). Table 5 shows the (medium-run) elasticity estimates for the preferred static fuel demand model based on the RE-AR(1) and the FE-AR(1) estimators for the baseline model specification including fuel price, vehicle stock per capita and average income, as well as the extended model specification which also includes the two public transport variables (i.e. price of the travelcard to/from Lisbon and the availability of urban and suburban rail networks). Table 6 shows the short- and long-run elasticities obtained from the preferred dynamic fuel demand model for both the baseline and the extended model specifications. The elasticity estimates shown in each table refer to the models for fuel demand per vehicle (Fuel/Vstock) and fuel demand per capita (Fuel/Pop) as the dependent variable, and are identified accordingly. The full set of results for the static and dynamic fuel demand models is reported in Tables A.1–A.4 in Appendix A. Overall, the medium-run elasticities obtained from the static fuel demand models lie between the short- and long-run elasticities estimated in the dynamic fuel demand models, and the long-run elasticities are about 2.5 times higher than the respective short-run elasticities. We find that the magnitude of the price elasticity estimates for the AML is generally higher than the range of values obtained in previous studies and summarised in Table 1 of the literature review section,

Table 5 Elasticity estimates of fuel demand obtained from the preferred static models.

ln(Pfuel/CPI) ln(RInc/Pop) ln(Vstock/Pop) ln(PTprice/CPI) ln(Rail/Pop) ***

Significance at 1%.

Fuel demand per vehicle (Fuel/Vstock)

Fuel demand per capita (Fuel/Pop)

Baseline model

Baseline model

Extended model

Extended model

RE-AR(1)

FE-AR(1)

RE-AR(1)

FE-AR(1)

RE-AR(1)

FE-AR(1)

RE-AR(1)

FE-AR(1)

0.606*** 0.806*** 0.600*** – –

0.586*** 0.962*** 0.613*** – –

0.755*** 0.786*** 0.563*** 0.375*** 0.214

0.735*** 0.861*** 0.604*** 0.355 0.017

0.606*** 0.806*** 0.400*** – –

0.586*** 0.962*** 0.387*** – –

0.755*** 0.786*** 0.437*** 0.375*** 0.214

0.735*** 0.861*** 0.396*** 0.355 0.017

40

P.C. Melo, A.R. Ramli / Transportation Research Part A 67 (2014) 30–46

Table 6 Elasticity estimates of fuel demand obtained from the preferred dynamic models. Fuel demand per vehicle (Fuel/Vstock) Baseline model Short run ln(Pfuel/CPI) ln(RInc/Pop) ln(Vstock/Pop) ln(PTprice/CPI) ln(Rail/Pop) * ** ***

***

0.619 0.5044*** 0.472*** – –

Fuel demand per capita (Fuel/Pop)

Extended model Long run ***

1.649 1.344*** 1.259*** – –

Short run ***

0.722 0.541*** 0.435*** 0.382 0.101

Baseline model Long run ***

1.822 1.367*** 1.099*** 0.965 0.256

Short run 0.354 0.444* 0.162 – –

Extended model Long run 0.919 1.152* 0.420 – –

Short run **

0.484 0.510** 0.225 0.470* 0.051

Long run 1.194** 1.259** 0.554 1.158* 0.126

Significance at 10%. Significance at 5%. Significance at 1%.

while the estimates of the income elasticity tend to be at the upper bound of the values reported in Table 1 (see Graham and Glaister, 2002b; Basso and Oum, 2007). Starting with the static fuel demand models, the Hauman test indicates that the FE estimator should be preferred to the RE estimator, while the BFN Durbin–Watson test statistic suggests there is serial correlation in the residuals (see Tables A.1 and A.2 in Appendix A). We therefore select the FE-AR(1) as our preferred estimator. The magnitude of the elasticity estimates obtained from the extended models tends to be relatively higher than that of the baseline models: about 25% higher (in absolute value) for fuel price and 10% and 7% higher for income. The results obtained from the preferred static fuel demand model indicate that the fuel price elasticity estimate is equal to 0.586 and 0.735 for the baseline and extended model specifications respectively. This suggests that a 10% increase in fuel prices is associated with a reduction of 7.4% (extended model) in fuel consumption per vehicle and per capita. As for the elasticity of income, our findings indicate a value equal to 0.786 and 0.861 for the baseline and extended models respectively, suggesting that a 10% increase in average income levels is associated with an increase of 8.61% in fuel consumption (extended model). A 10% increase in car ownership levels is associated with an increase in fuel demand per capita of about 4% (both in the baseline and extended models), but a reduction in fuel consumption per vehicle of 6% (both in the baseline and extended models). This suggests there is a less than proportional increase in fuel consumption levels compared to increases in vehicle stock. Although the RE-AR(1) model suggests that increasing the price of public transport (+10%) is associated with an increase of fuel demand (+3.75%), the coefficient is not statistically significant in our preferred FE-AR(1) model. On the other hand, the models suggest that there is no effect from increased availability of urban and suburban rail networks. We believe that this result may be partly due to the fact that the measure used (i.e. density of urban and suburban rail stations per 10,000 people) does not capture actual service levels but only available infrastructure. Although the availability of public transport is an important facilitator of public transport use, it is actual levels of service that people consider when deciding about whether to use public transport or private transport. Unfortunately, we could not measure urban and suburban rail service levels for the different municipalities over the period of analysis in this study. We now consider the dynamic fuel demand models, for which only the difference- and system-GMM estimators and the bias-corrected LSDV estimators based on initial values obtained from the difference- and system-GMM estimators can provide consistent parameter estimates. Tables A.3 and A.4 in Appendix A show that the Hansen test fails to reject the null hypothesis of instrument exogeneity for all the GMM models, confirming that the instruments used are valid. Overall the GMM models pass the Arellano and Bond tests, AB AR(1) and AB AR(2), for serial correlation in the error term. The exceptions are the difference-GMM and LSDVc – AB models. As discussed in the previous section, the difference- and system-GMM estimators have been shown to lead to biased and inefficient parameter estimates in the presence of samples with a small number of cross-sectional units (e.g. Bun and Kiviet, 2006). We therefore select the coefficients obtained from the bias-corrected LSDV estimators using initial values from the system-GMM estimator (i.e. LSDVc-BB) as our preferred estimates of the fuel demand elasticities. Table 6 summarises the short- and long-run elasticity values obtained from the baseline and extended model specifications for the models using fuel demand per vehicle (Fuel/Vstock) and fuel demand per capita (Fuel/Pop) as the dependent variable. The results for the extended fuel demand model based on the LSDVc – BB estimator indicate that a 10% increase in fuel prices is associated with a reduction in fuel consumption per vehicle (fuel consumption per capita) of 7.22% (4.84%) in the short run and of 18.22% (11.94%) in the long run. This indicates that the demand for fuel is inelastic with respect to fuel prices in the short run, but significantly more elastic in the long run. As for the response to changes in income levels, our findings indicate an income elasticity of 0.541 and 1.367 for the fuel demand per vehicle for the short and long run respectively (extended demand model). The results for the models based on fuel demand per capita indicate an income elasticity of 0.510 and 1.259 for the short and long run respectively (extended demand model). Similarly to fuel prices, the demand for fuel is inelastic with respect to income in the short run but reasonably elastic in the long run. A 10% increase in income is associated with an increase of fuel consumption per vehicle (fuel consumption per capita) of 5.41% (5.10%) in the short run and 13.67% (12.59%) in the long run.

41

P.C. Melo, A.R. Ramli / Transportation Research Part A 67 (2014) 30–46

The elasticity estimate of car ownership (i.e. vehicle stock per capita) obtained from the extended model of fuel demand per vehicle indicates that increasing car ownership by 10% reduces fuel consumption per vehicle by 4.35% and 10.99% in the short and long run respectively. This is in agreement with the findings obtained in the static models, which also indicated a less than proportional increase in fuel consumption levels compared to increases in vehicle stock. We can also derive the effect of increasing vehicle stock on total fuel consumption from these figures: a 10% increase in total vehicle stock (keeping population constant) is associated with an increase of 5.65% and 14.27% in total fuel demand in the short and long run respectively. As for the impact on fuel demand per capita, the results indicate that it increases by 2.25% and 5.54% in the short and long run respectively given a 10% increase in car ownership levels (although the coefficients are not statistically significant). Turning finally to the role of the cost and availability of public transport on fuel demand, the findings obtained from the preferred model appear to be mixed. Although the direction of the effect is as expected, that is, an increase in the price of public transport increases fuel demand while an increase in the availability of public transport reduces fuel demand, the estimates tend to be statistically insignificant (as with the static demand models). The exception is for the effect of the price of public transport on fuel demand per capita, which is positive and equal to 0.470 in the short run and 1.158 in the long run. These fairly high values would suggest a strong substitution effect between public transport and private transport in the Lisbon Metropolitan Area; however the effect tends to be insignificant in the other fuel demand models. The inclusion of the price of public transport and the availability of rail networks in the fuel demand models does not generally change the statistical significance of the other covariates (except perhaps for the models of fuel demand per capita), but can affect the magnitude of the elasticity estimates. The elasticity estimates for the fuel price elasticity are somewhat higher (in absolute values), about 17% (10%) higher in the short run (long run) for the model of fuel demand per vehicle and 37% (30%) higher in the short run (long run) for the model of fuel demand per capita. The income elasticities appear to be affected only marginally, between 2% and 15% depending on the fuel demand model and the time horizon. It is not easy to compare our findings to previous evidence for Portugal. As discussed in Section 2, we know only of two studies (Sterner et al., 1992; De Oliveira, 2001) producing evidence on price and income elasticities for Portugal. These studies, however, use very different time periods and empirical methods, making direct comparisons difficult. Sterner et al. (1992) used pooled OLS to estimate dynamic gasoline demand models for 20 OECD countries during the period 1965– 1985. De Oliveira (2001) used time series cointegration techniques to estimate gasoline and diesel demand models for Portugal between 1977 and 2000. The elasticity estimates of fuel price obtained in our models are higher than the gasoline price elasticity obtained by Sterner et al. (1992) and De Oliveira (2001). Sterner et al. (1992) obtained short- and long-run elasticities equal to 0.13 and 0.67, respectively. De Oliveira (2001) estimated considerably higher elasticity estimates, around 0.40 in the short run and between 0.75 and 0.90 in the long run. Our values are closer to those obtained by De Oliveira (2001) and range between 0.48 and 0.72 in the short run and between 1.20 and 1.82 in the long run. Our income elasticity estimates range between 0.51 and 0.54 in the short run and between 1.26 and 1.37 in the long run, and are considerably higher than the elasticities produced by De Oliveira (2001): 0.32 in the short run and between 0.65 and 0.96 in the long run. 6. Using fuel demand elasticities to evaluate road transport CO2 emissions Given the fuel demand elasticities estimated in the previous section, it is possible to construct scenarios for future fuel consumption and associated CO2 GHG emission levels. This requires making assumptions about the growth of average real incomes, vehicle stock, and average fuel prices in the short and long run. To illustrate how the fuel demand elasticity values can be used to predict road transport CO2 GHG emissions we use CO2 emission factors for gasoline and diesel provided by the DGEG in national legislation (Agência Portuguesa do Ambiente, 2011) and apply them to fuel consumption levels as predicted according to different scenarios for growth of real incomes, vehicle stock, and fuel prices. We consider the scenarios shown in Table 7 below. Scenario A represents a weak growth of real income and vehicle stock levels, and a strong increase in fuel prices. This macroeconomic environment is similar to Portugal’s current economic situation, characterised by poor economic performance and high fuel prices. Scenario C predicts an optimistic scenario characterised by strong growth of real income and vehicle stock, while keeping fuel prices high. Scenario B describes a macroeconomic setting between scenarios A and C, characterised by medium growth of real income, vehicle stock, and fuel prices. By combining the parameters reported in Table 7 with the preferred fuel demand elasticities estimated in the previous section, we obtain fuel consumption levels which are then converted to CO2 emissions using official fuel CO2 emission factors for Portugal until the year 2020. In addition, we also consider the increase in fuel prices required to ensure that road transTable 7 Assumptions on the growth of real incomes, vehicle stock, and price of fuel. Annual growth rate

Scenario A

Scenario B

Scenario C

Real income Vehicle stock Real price of fuel

0.5 1.5 7.5

1.5 2.5 5

2.5 7.5 7.5

42

P.C. Melo, A.R. Ramli / Transportation Research Part A 67 (2014) 30–46

200

180

160

Index (1990=100)

140

120

100

80 2020 Target 60

40

20 Scenario A

Scenario B

Scenario C

0 1990

1992

1994

1996

1998

2000

2002

2004

2006

2008

2010

2012

2014

2016

2018

2020

Fig. 4. Evolution of road transport CO2 GHG emissions in the AML based on the estimated fuel demand elasticities and alternative scenarios for growth of real incomes, vehicle stock, and fuel prices (1990 = 100).

port CO2 emissions meet the European 2020 target in the context of medium and strong economic growth (as depicted in scenarios B and C respectively). Portugal is undergoing a strong economic recession, with annual real GDP growth rates of 0.1% in 2008, 2.6% in 2009, +1.4% in 2010, 1.6% in 2011, and 2.8% in 2012. There is great uncertainty about future performance and the latest projections conducted by the Portuguese government, together with the European Central Bank (ECB), the European Commission (EC), and the International Monetary Fund (IMF), anticipate real GDP growth rates of 1.8% and +0.8% for 2013 and 2014 respectively. In addition, average income levels have been negatively affected by strong wage cuts in the public sector, while fuel prices have also increased considerably.8,9 This suggests that scenario A, based on a poor growth of income levels and vehicle stock, and a strong increase in fuel prices may be a reasonable scenario in the short to medium term. Given the current circumstances of the Portuguese economy, characterised by reduced average incomes and a strong increase in fuel prices, it is expected that road transport CO2 emissions will follow the recent downward trend depicted in Fig. 1. Fig. 4 shows the evolution of road transport CO2 emissions for the AML between 1990 and 2020. The values shown for CO2 emissions between 2010 and 2020 are based on the three scenarios – A, B, and C – described above. The values are indexed to 100 in 1990.10 CO2 greenhouse gas emissions from gasoline and diesel consumption in the AML increased sharply during the 1990s, following the rising trend in real incomes, car ownership, and cheap real fuel prices (see Fig. 3). The peak of CO2 emissions occurred in 2004, +85% than in 1990, after a slight reduction during the economic crisis of 2002– 2003. Since 2004, CO2 emissions have been falling gradually, with the exception of 2009 which experienced a short lived recovery. Fig. 4 shows CO2 emissions as predicted according to scenarios A, B and C. Only scenario A predicts that CO2 emissions will meet the European target (20% of 1990 levels) established by the new Transport White Paper (EC, 2011). According to scenario A road transport CO2 emissions are predicted to fall to 80% of CO2 emissions levels in 1990 between 2017 and 2018; and to 67% of 1990’s CO2 emissions levels in 2020. Scenario B predicts that by 2020 road transport CO2 emissions will be very similar to CO2 emissions in 1990 (only 4% higher). To meet the European target of not exceeding 1990 CO2 emissions levels by 80%, fuel prices would have to rise, on average, by 7.4% per year. Scenario C predicts that road transport CO2 emissions in 2020 will be 15% higher than in 1990. This outcome is a result of the 8 One of the main sources of reduction in public sector wages was the abolition of both the Holiday subsidy and the Christmas subsidy, which account for 14% of the annual salary. 9 Gasoline and diesel prices increased by 12–13% (18–20%) and 19% (26%) in 2011 (2012) respectively, compared to 2010 (see http://www.pordata.pt/). 10 The values refer to gasoline and diesel consumption only.

P.C. Melo, A.R. Ramli / Transportation Research Part A 67 (2014) 30–46

43

positive economic environment, characterised by rising real incomes and vehicle stock. In order to meet the CO2 emissions target, fuel prices would have to rise, on average, by 10.8% per year. These results can only provide an indicative interval for future road transport CO2 emissions, framed by the three scenarios specified. There are, of course, other factors that can affect the level of CO2 emissions which we could not incorporate in our scenarios. We have assumed that changes in emissions due to vehicle and fuel type composition, driving style and weather remain unaltered during the period 2010–2020. Although future improvements in fuel efficiency are expected to deliver energy savings, these may be offset by rebound effects resulting from more frequent and longer journeys. In addition, although the increase in electric cars may also contribute to the reduction of road transport CO2 emissions, so far the uptake of this technology has been rather disappointing and low. Less than 1% of passenger vehicles in 2011 ran on fuels other than gasoline or diesel. In addition, hybrid and electric vehicles accounted for only about 0.5% of the total sales of new passenger vehicles in 2011 (INE, 2012). Nevertheless, and in spite of the uncertainty about future improvements in fuel efficiency and the adoption of green fuel technology, such improvements are likely to lead to more optimistic reductions in road transport CO2 emissions for each of the three scenarios considered here.

7. Conclusions The empirical analysis conducted in this paper contributes to the understanding of fuel consumption in the largest metropolitan region of Portugal and can be used in the design of policies aimed at improving urban and regional sustainable mobility patterns producing lower road transport CO2 greenhouse gas emissions. The paper makes two contributions to the literature on fuel demand. It generates new empirical evidence on short-, medium-, and long-run fuel demand elasticities for the Lisbon Metropolitan Area using a dataset which has been compiled for the first time and specifically for this analysis. The second contribution consists of the use of a comprehensive model specification which accounts simultaneously for the effect of the price of public transport and the availability of urban and suburban rail networks on the demand for fuel. Our findings for the Lisbon Metropolitan Area for the period 1993–2010 indicate that both fuel price and income play an important role in determining fuel consumption levels. The elasticity of fuel demand per capita with respect to income is about 5% higher than the elasticity of fuel price (in absolute value), suggesting that fuel prices have to increase proportionally more than income to reduce fuel consumption per capita. Compared to existing evidence for fuel demand elasticities, our findings suggest a more elastic fuel demand in the Lisbon Metropolitan Area, especially with respect to fuel prices. The magnitude of the price elasticity tends to be higher than the range of values obtained in previous studies, while the income elasticity tends to be at the upper bound of the range of the estimates produced by previous studies. The elasticity of fuel demand with respect to fuel price ranges between 0.48 and 0.72 in the short run and between 1.19 and 1.82 in the long run, indicating that a 10% increase in fuel price is associated with a reduction in fuel consumption between 5–7% and 12–18% in the short and long run respectively. We find that income elasticities range between 0.51 and 0.54 in the short run and between 1.26 and 1.37 in the long run. This suggests that a 10% increase in real income levels leads to an increase of fuel consumption of 5% and between 13–14% in the short and long run respectively. As for the role of vehicle stock, we find that an increase of 10% in vehicle stock (keeping population constant) is associated with an increase of fuel consumption about 6% in the short run and 1.4% in the long run. Finally, our findings provide only weak evidence in favour of a reduction of fuel consumption as a result of a reduction in public transport fares, and no effect from improved availability of urban and suburban rail networks. We also discuss how the limitation of our measure of rail availability to capture service levels may partly explain this result. We then illustrate how the fuel demand elasticities can be used to predict road transport CO2 GHG emissions using official national CO2 emission factors for gasoline and diesel. This also requires creating various scenarios making different assumptions about the growth in average real incomes, vehicle stock, and average fuel prices. Only the more pessimistic scenario (based on poor economic performance, weak growth in vehicle stock and high fuel prices) predicts that road transport CO2 emissions will meet the European target established in the New Transport White paper of reducing GHG emissions by at least 20% until 2020. In the context of medium and strong economic growth, fuel prices would need to increase annually by 7.4% and 10.8% respectively in order to meet the emissions target.

44

P.C. Melo, A.R. Ramli / Transportation Research Part A 67 (2014) 30–46

Appendix A. See Tables A.1–A.4.

Table A.1 Results of the static models of fuel demand per vehicle. Fuel demand per vehicle (Fuel/ Vstock)

Baseline model OLS

RE

FE

RE-AR(1)

FE-AR(1)

OLS

RE

FE

RE-AR(1)

FE-AR(1)

ln(Pfuel/CPI)

5.7371* (2.7388) 0.5408 (0.4168) 0.2174 (0.3274) – – – –

0.6053 (2.1523) 0.9207* (0.5127) 0.5956** (0.2575) – – – –

1.5813 (2.4909) 1.2129** (0.5689) 0.6902** (0.2872) – – – –

0.6056*** (0.1688) 0.8064*** (0.2195) 0.6001*** (0.1224) – – – –

0.5859*** (0.1726) 0.9618*** (0.2824) 0.6134*** (0.1484) – – – –

3.7644 (2.1994) 0.9133* (0.4432) 0.0904 (0.2769) 0.3459* (0.1836) 0.5386 (0.4205)

1.1924 (2.1238) 1.0986** (0.4971) 0.4940* (0.2521) 0.6375*** (0.2264) 0.2316 (0.2269)

1.9256 (2.3916) 1.3312** (0.5678) 0.5888** (0.2769) 0.7519** (0.3249) 0.1087 (0.2357)

0.7554*** (0.1830) 0.7860*** (0.2146) 0.5633*** (0.1206) 0.3753*** (0.1448) 0.2141 (0.2032)

0.7348*** (0.2031) 0.8614*** (0.2924) 0.6036*** (0.1491) 0.3546 (0.2686) 0.0170 (0.2704)

306 0.19 – – –

306 0.01 0.25 0.02 16.00***

306 0.00 0.26 0.03

306 0.01 0.17 0.00 –

289 0.01 0.12 0.00 –

306 0.32 – – –

306 0.14 0.28 0.10 11.58*

306 0.10 0.28 0.06

306 0.14 0.19 0.13 –

289 0.10 0.13 0.07 –

–

–

–

0.92

0.92

–

–

–

0.94

0.94

ln(RInc/Pop) ln(Vstock/Pop) ln(PTprice/CPI) ln(Rail/Pop) Observations R2-Overall R2-Within R2-Between Hausman test (FE vs. RE) BFN DurbinWatson test of serial correlation

Extended model

Notes: The values in parentheses are standard errors. Standard errors are robust to heteroskedasticity and adjusted for intra-municipality dependence. All regression models include a time trend. * Significance at 10%. ** Significance at 5%. *** Significance at 1%.

Table A.2 Results of the static models of fuel demand per capita. Fuel demand per capita (Fuel/Pop)

Baseline model OLS

RE

FE

RE-AR(1)

FE-AR(1)

OLS

RE

FE

RE-AR(1)

FE-AR(1)

ln(Pfuel/CPI)

5.7371* (2.7388) 0.5408 (0.4168) 0.7826** (0.3274)

0.6053 (2.1523) 0.9207* (0.5127) 0.4044 (0.2575)

1.5813 (2.4909) 1.2129** (0.5689) 0.3098 (0.2872)

0.6056*** (0.1688) 0.8064*** (0.2195) 0.3999*** (0.1224)

0.5859*** (0.1726) 0.9618*** (0.2824) 0.3866*** (0.1484)

306 0.40

306 0.16 0.24 0.13 16.00***

306 0.09 0.24 0.06

306 0.21 0.15 0.23 –

289 0.18 0.08 0.20 –

3.7644 (2.1994) 0.9133* (0.4432) 0.9096*** (0.2769) 0.3459* (0.1836) 0.5386 (0.4205) 306 0.49

1.1924 (2.1238) 1.0986** (0.4971) 0.5060** (0.2521) 0.6375*** (0.2264) 0.2316 (0.2269) 306 0.31 0.26 0.32 11.58*

1.9256 (2.3916) 1.3312** (0.5678) 0.4112 (0.2769) 0.7519** (0.3249) 0.1087 (0.2357) 306 0.23 0.27 0.22

0.7554*** (0.1830) 0.7860*** (0.2146) 0.4367*** (0.1206) 0.3753*** (0.1448) 0.2141 (0.2032) 306 0.39 0.17 0.50 –

0.7348*** (0.2031) 0.8614*** (0.2924) 0.3964*** (0.1491) 0.3546 (0.2686) 0.0170 (0.2704) 289 0.31 0.09 0.37 –

0.92

0.92

0.94

0.94

ln(RInc/Pop) ln(Vstock/Pop)

Extended model

ln(PTprice/CPI) ln(Rail/Pop) Observations R2-Overall R2-Within R2-Between Hausman test (FE vs. RE) BFN Durbin-Watson test of serial correlation

–

–

Notes: The values in parentheses are standard errors. Standard errors are robust to heteroskedasticity and adjusted for intra-municipality dependence. All regression models include a time trend. * Significance at 10%. ** Significance at 5%. *** Significance at 1%.

45

P.C. Melo, A.R. Ramli / Transportation Research Part A 67 (2014) 30–46 Table A.3 Results of the dynamic models of fuel demand per vehicle. Fuel demand per car (Fuel/ Vstock) ln(Fuel/Vstock)t1 ln(Pfuel/CPI) ln(RInc/Pop) ln(Vstock/Pop)

Baseline model Diff-GMM

Extended model

SYS-GMM

0.5659*** (0.2005) 0.6857*** (0.1683) 0.5165** (0.2083) 0.4391** (0.2141)

LSDVc-AB

0.4244* (0.2483) 0.8313*** (0.2756) 0.3189* (0.1893) 0.1138 (0.1337)

LSDVc-BB

0.5708*** (0.0559) 0.6745*** (0.2203) 0.5943*** (0.2262) 0.4679*** (0.1397)

0.6248*** (0.0539) 0.6188*** (0.2247) 0.5044** (0.2477) 0.4722*** (0.1496)

ln(PTprice/CPI) ln(Rail/Pop) Observations Hansen testa AB AR(1)a AB AR(2)a

272 0.44 0.01 0.04

289 0.59 0.03 0.16

289

289

0.01 0.04

0.03 0.16

Diff-GMM

SYS-GMM

LSDVc-AB

LSDVc-BB

0.5745*** (0.2170) 0.7050*** (0.2472) 0.7548 (0.7955) 0.3693 (0.2740) 0.3772 (0.4906) 0.0384 (0.1556) 272 0.56 0.01 0.05

0.4656* (0.2458) 0.6957** (0.3265) 0.4897** (0.2184) 0.1189 (0.0987) 0.2398 (0.1713) 0.2485 (0.1763) 289 0.60 0.02 0.13

0.5563*** (0.0561) 0.7502*** (0.2336) 0.6311*** (0.2343) 0.4350*** (0.1410) 0.3459 (0.2406) 0.1225 (0.1867) 289

0.6041*** (0.0550) 0.7215*** (0.2396) 0.5413** (0.2554) 0.4352*** (0.1497) 0.3819 (0.2636) 0.1012 (0.2024) 289

0.01 0.05

0.02 0.13

Notes: Standard errors are robust to heteroskedasticity and adjusted for intra-municipality dependence. All regression models include a time trend. The number of instruments used in the Diff-GMM and SYS-GMM estimators is 20 (22) and 22 (24) respectively for the baseline (extended) demand models. a The values reported are p-values. * Significance at 10%. ** Significance at 5%. *** Significance at 1%.

Table A.4 Results of the dynamic models of fuel demand per capita. Fuel demand per capita (Fuel/Pop)

Baseline model Diff-GMM

ln(Fuel/Pop)t1 ln(Pfuel/CPI) ln(RInc/Pop) ln(Vstock/Pop)

***

Extended model

SYS-GMM **

LSDVc-AB ***

LSDVc-BB ***

0.4325 (0.1262) 0.3443** (0.1423) 0.6335*** (0.2272) 0.1990 (0.2228)

0.4196 (0.1953) 0.5536** (0.2817) 0.2677* (0.1542) 0.5347*** (0.1937)

0.5607 (0.0518) 0.3978* (0.2308) 0.5464** (0.2344) 0.1645 (0.1404)

0.6147 (0.0500) 0.3540 (0.2229) 0.4437* (0.2433) 0.1619 (0.1449)

272 0.66 0.01 0.04

289 0.66 0.01 0.08

289

289

0.01 0.04

0.01 0.08

ln(PTprice/CPI) ln(Rail/Pop) Observations Hansen testa AB AR(1)a AB AR(2)a

Diff-GMM **

0.3906 (0.1819) 0.5190* (0.2777) 0.7195** (0.3173) 0.2840 (0.2113) 0.7807 (0.6141) 0.2072 (0.5836) 272 0.89 0.01 0.07

SYS-GMM *

0.3702 (0.1900) 0.4195 (0.2897) 0.4150 (0.2604) 0.5357* (0.2771) 0.1663 (0.1388) 0.2904 (0.1901) 289 0.65 0.02 0.10

LSDVc-AB ***

LSDVc-BB

0.5475 (0.0517) 0.5059** (0.2430) 0.6054** (0.2419) 0.2239 (0.1409) 0.4317* (0.2525) 0.0762 (0.1930) 289

0.5947*** (0.0503) 0.4838** (0.2373) 0.5101** (0.2522) 0.2247 (0.1446) 0.4695* (0.2645) 0.0510 (0.2021) 289

0.01 0.07

0.02 0.10

Notes: Standard errors are robust to heteroskedasticity and adjusted for intra-municipality dependence. All regression models include a time trend. The number of instruments used in the Diff-GMM and SYS-GMM estimators is 20 (22) and 22 (24) respectively for the baseline (extended) demand models. a The values reported are p-values. * Significance at 10%. ** Significance at 5%. *** Significance at 1%.

References Agência Portuguesa do Ambiente, 2011. Portuguese National Inventory Report on Greenhouse Gases, 1990–2009. Ministério da Agricultura, Mar, Ambiente e Ordenamento do Território. Al-Faris, A.-R.F., 1997. Demand for oil products in the GCC countries. Energy Policy 25, 55–61. Arellano, M., Bond, S., 1991. Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations. Rev. Econom. Stud. 58, 277–297. Arellano, M., Bover, O., 1995. Another look at the instrumental variable estimation of error-components models. J. Econom. 68, 29–51. Baltagi, B.H., 2008. Econometric Analysis of Panel Data. John Wiley & Sons Ltd., Chichester. Baltagi, B., Griffin, J., 1997. Pooled estimators vs their heterogeneous counterparts in the context of dynamic demand for gasoline. J. Econom. 77, 303–327.

46

P.C. Melo, A.R. Ramli / Transportation Research Part A 67 (2014) 30–46

Baltagi, B.H., Bresson, G., Griffin, J.M., Pirotte, A., 2003. Homogeneous, heterogeneous or shrinkage estimators? Some empirical evidence from French regional gasoline consumption. Empirical Econom. 28, 795–811. Basso, L.J., Oum, T.H., 2007. Automobile fuel demand: a critical assessment of empirical methodologies. Trans. Rev. 27, 449–484. Bhargava, A., Franzini, L., Narendranathan, W., 1982. Serial correlation and the fixed effects model. Rev. Econom. Stud. 49, 533–549. Blundell, R., Bond, S., 1998. Initial conditions and moment restrictions in dynamic panel data models. J. Econom. 87, 115–143. Blundell, R., Bond, S., 2000. GMM estimation with persistent panel data: an application to production functions. Econom. Rev. 3, 321–340. Brons, M., Nijkamp, P., Pels, E., Rietveld, P., 2008. A meta-analysis of the price elasticity of gasoline demand. A SUR approach. Energy Econom. 30, 2105– 2122. Bun, M.J.G., Kiviet, J.F., 2006. The effects of dynamic feedbacks on LS and MM estimator accuracy in panel data models. J. Econom. 132, 409–444. Crôtte, A., Noland, R.B., Graham, D.J., 2010. An analysis of gasoline demand elasticities at the national and local levels in Mexico. Energy Policy 38, 4445– 4456. Dahl, C.A., 2012. Measuring global gasoline and diesel price and income elasticities. Energy Policy 41, 2–13. Dargay, J., Gately, D., Sommer, M., 2007. Vehicle ownership and income growth, worldwide: 1960–2030. Energy J. 28, 143–170. De Oliveira, J.M.C., 2001. Procura de produtos petrolíferos em Portugal: uma abordagem empírica. Lisbon: DGEP, Portuguese Ministry of Finance, Documento de Trabalho Nr.24. EC, 2011. Roadmap to a Single European Transport Area. European Commission, COM(2011) 144 final, Brussels. Espey, M., 1998. Gasoline demand revisited: an international meta-analysis of elasticities. Energy Econom. 20, 273–295. Goodwin, P., Dargay, J., Hanly, M., 2004. Elasticities of road traffic and fuel consumption with respect to price and income: a review. Trans. Rev. 24, 275–292. Graham, D., Glaister, S., 2002a. Review of Income and Price Elasticities in the Demand for Road Traffic. Department for Transport, London. Graham, D.J., Glaister, S., 2002b. The demand for automobile fuel: a survey of elasticities. J. Trans. Econom. Policy 36, 1–26. Graham, D., Glaister, S., 2004. Road traffic demand elasticity estimates: a review. Trans. Rev. 24, 261–274. Hanly, M., Dargay, J., Goodwin, P., 2002. Review of Price Elasticities in the Demand for Road Traffic. ESRC TSU publication 2002/13, University of London, Centre for Transport Studies. Hansen, L.P., 1982. Large sample properties of generalized method of moments estimators. Econometrica 50, 1029–1054. Hausman, J.A., 1978. Specification tests in econometrics. Econometrica 46, 1251–1271. Hughes, J.E., Knittel, C.R., Sperling, D., 2008. Evidence of a shift in the short-run price elasticity of gasoline demand. Energy J. 29, 113–134. INE, 2012. Estatisticas dos Transportes 2011. Judson, R.A., Owen, A.L., 1999. Estimating dynamic panel data models: a guide for macroeconomists. Econom. Lett. 65, 9–15. Li, Z., Rose, J.M., Hensher, D.A., 2010. Forecasting automobile petrol demand in Australia: an evaluation of empirical models. Trans. Res. Part A: Policy Pract. 44, 16–38. Liddle, B., 2009. Long-run relationship among transport demand, income, and gasoline price for the US. Trans. Res. Part D: Trans. Environ. 14, 73–82. Park, S.Y., Zhao, G., 2010. An estimation of U.S. gasoline demand: a smooth time-varying cointegration approach. Energy Econom. 32, 110–120. Pock, M., 2010. Gasoline demand in Europe: new insights. Energy Econom. 32, 54–62. Polemis, M.L., 2006. Empirical assessment of the determinants of road energy demand in Greece. Energy Econom. 28, 385–403. Samimi, R., 1995. Road transport energy demand in Australia a cointegration approach. Energy Econom. 17, 329–339. Sterner, T., Dahl, C., Franzén, M., 1992. Gasoline tax policy: carbon emissions and the global environment. J. Trans. Econom. Policy 26, 109–119. Virley, S., 1993. The effect of fuel price increases on road transport CO2 emissions. Trans. Policy 1, 43–48.