Automobile Fuel Demand: A Critical Assessment of Empirical Methodologies

Abstract

Many surveys have attempted to convey and synthesize the information of hundreds of studies on automobile fuel demand. In most cases, the focus has been placed in giving assessments of the most likely values of various elasticities, particularly price and income, while trying to explain the differences between results. However, given the summary characteristic of these surveys, the most popular approaches and methodologies—such as dynamic reduced‐form demand models with time‐series data—have dominated the core values obtained. The present survey focuses instead on the various approaches and methods that have been used. It reviews and classifies them, showing that there are relevant findings, raised by studies using less popular approaches, which seem to challenge some of the accepted core results in the literature. These other approaches include: co‐integration techniques, use of disaggregate data at the household level and flexible functional forms, and structural models of automobile fuel consumption.

Acknowledgements

The authors wish to thank Chunyan Yu and two anonymous referees for comments and suggestions that materially improved this paper.

Notes

1. Meta‐analyses are used to determine if there are factors‐such as the type of data or the specification of the model‐that systematically affect price and income elasticities. In these analyses, elasticity estimates from over 100 studies are used as dependant variables, while the factors that possibly determine the differences are on the right‐hand side.

2. From a microeconomics point of view, demands for all commodities are determined simultaneously. Consequently, gasoline demand is potentially dependent on all prices. By assuming separability, most direct‐demand models consider gasoline demand to be a function only of its own price and income.

3. For formal descriptions of other dynamic models, see Sterner and Dahl (1992).

4. Bentzen (1994) actually does things the other way around by not including income, but only stock of vehicles.

5. Franzén and Sterner (1995) estimate ‘between’ cross‐section estimators for gasoline consumption using annual data over a period of 25 years. Pesaran and Smith (1995) do the same for energy consumption, considering a period of 18 years.

6. Interestingly, Espey (1998) found no significant difference in long‐run estimates, for both income and price, between time series and panel data studies, while Dahl and Sterner (1991) found no difference in income and price elasticities in the specific context of lagged endogenous models and for both the short‐ and long‐run, between time series and panel data studies. These results may seem to contradict the theoretical findings of Pesaran and Smith. Nevertheless, it is likely that Espey (1998) and Dahl and Sterner (1992) considered not only pooled estimators in panel data results, but also other panel estimators‐such as individual time series, aggregate time series, or between estimators.

7. Also, recall that Blum et al. (1988) obtained a smaller short‐run income elasticity by including variables reflecting the level of economy activity (such as employment) and other determinants of travel. See the second section.

8. Super‐consistent estimates approach their values at a rate proportional to n ⁻¹ instead of n ^−1/2.

9. For some seminal papers on co‐integration, see Granger (1981), Engle and Granger (1987) and Engle and Yoo (1987). In addition, many graduate books in econometrics have fairly comprehensive reviews of the techniques involved (e.g. Hayashi, 2000; Hamilton, 1994). For a complete and in depth review on unit roots and co‐integration, see Maddala and Kim (1998).

10. To be precise, Bentzen actually used stock of vehicles instead of income. He also added a time trend to capture the effect of increasing fuel efficiency.

11. Pesaran and Smith (1995) use data for nine Asian countries between 1973 and 1990 to estimate energy consumption as a function of the price of energy and income (gasoline consumption is part of energy consumption). They find consumption to be co‐integrated with price and income in six out of nine countries.

12. Since information on permanent income was unavailable, they used total household consumption expenditure as a proxy.

13. Greene and Hu (1986) found that the short‐run price elasticity of one‐vehicle households changes with prices, a result that is in line with Hausman and Newey’s finding.

14. The authors claim that, with the exception of a couple of papers, the number of licensed drivers has been virtually ignored in all previous studies. The claim appears to be correct; Archibald and Gillingham’s disaggregated study, for example, did not include the variable and it has not been considered in many reduced‐form studies with aggregate data as far as the authors know. Dahl (1986), however, claims that there seems to be no systematic significant variation in gasoline consumption when considering it per capita, per household or per licensed driver. Gallini (1983) and Eltony (1993)‐in the context of structural demand models (see the fifth section)‐did take this variable into account. Their focus, though, was more in price effects than on income effects.

15. Price–income interaction or non‐linear income terms were not tested.

16. Note that these graphics are purely qualitative and do not attempt to portray actual values. Also, only trends found empirically are depicted and not those imposed a priori by some of the authors.

17. For an example of the use of an aggregate data study to assess the impact of fuel taxes as a gasoline conservation policy, see Sterner et al. (1992).

18. Graham and Glaister (2002) hypothesized that the models used at the micro‐level tend to be much less restrictive in exogenous variable specification than the aggregate studies and, as Blum et al. (1988) suggested, this may account for the reduction of the income effect. However, as indicated, the values obtained when the number of drivers is not controlled for are still small. Moreover, Hausman and Newey, who also find an income elasticity of around 0.4, only considered price, income, and regional and time dummies in their specification, and not other exogenous variables.

19. Drollas (1984) modelled separately vehicle stock, S, and fuel utilization per car, (M/E). He use a log‐linear static function for S dependent on fuel price and income, and a partial adjustment model for fuel utilization with fuel price, income, price of other transport services, and real price of vehicles as explanatory variables. He then inserts these submodels in equation (3), obtaining a single equation‐the reduced form‐relating fuel demand to the price of fuel, price of other transport services, price of vehicles, and income. The difference between this procedure and Johansson and Shipper’s is that Drollas estimated only the reduced form, while Johansson and Shipper estimated the submodels separately, and only then constructed the reduced form.

20. Note that the fuel efficiency of new cars is modelled while the fuel elasticity of old cars is observed.

21. Price elasticity of gasoline demand will be equal to price elasticity of m minus the price elasticity of E provided that both submodels depend on price (and other variables), but not on each other. The same applies for income.

22. Gallini shows that if the yearly percentage of decay on fuel economy is constant and the same across all vintages and models, the assumption would not cause estimation problems.

23. An interesting point raised by Rowendal is that one might expect a sizeable income effect in driving efficiency working through speed, since the value of time is indeed income dependent. The absence of income elasticity was then somewhat surprising to Rowendal, although in the Dahl survey the same result was found 10 years before (Dahl hypothesized that the insignificance of income elasticities could be caused by poor data quality).

24. Note that in some studies, gross location features such as urban and non‐urban are fixed. Models that explicitly allow for changes not only in the travelling mode and the number of cars per household, but also in work or residential locations are the domain of specialized transportation models that seek to explain far more than gasoline consumption. These strategic models have been implemented in some cities in order to evaluate major transport policies or changes in private and public transportation infrastructure, but require major collections of data every 5 or 10 years.