The Rebound Effect for Passenger Vehicles.

• Author: Linn, Joshua

1. INTRODUCTION

   Motivated by a desire to reduce gasoline consumption and the associated external energy security and climate costs, the U.S. fuel economy and greenhouse gas emissions rate standards for new passenger vehicles will dramatically increase average new vehicle fuel economy. The current standards, which the U.S. Environmental Protection Agency (U.S. EPA) and the U.S. Department of Transportation set jointly, raise average fuel economy to about 35 miles per gallon (mpg) by 2016. This level represents a roughly 40 percent increase compared to the standards in the mid-2000s. Further standards to 2025 could raise fuel economy by an additional 50 percent, past 50 mpg. The U.S. policy developments are part of a larger trend in which many countries and regions are tightening standards for fuel economy or greenhouse gas emissions rates (which vary inversely with fuel economy).

   A large literature has compared the cost of reducing gasoline consumption by using fuel economy or greenhouse gas emissions rate standards with the cost of using the gasoline tax (e.g., Jacobsen 2013). A central conclusion has been that the gasoline tax is much less costly to vehicle producers and consumers per gallon of gasoline saved. An important difference between fuel economy standards and a gasoline tax is that they create different incentives for driving. A gasoline tax can be used to internalize externalities that scale with gasoline consumption, such as greenhouse gas emissions. A gasoline tax also creates incentives to reduce driving, which reduces associated congestion, accidents, and local air pollution. Fuel economy standards, on the other hand, reduce gasoline consumption by raising fuel economy, but exacerbate the gap between the private and social cost of driving. The greater is the effect of driving costs on vehicle use--i.e., the rebound effect--the smaller the fuel savings from the standards and the associated greenhouse gas emissions reductions, and the higher are the external costs from traffic congestion, accidents, and local air quality. (1) Thus, welfare analysis depends crucially on the magnitude of the rebound effect--that is, the elasticity of miles traveled to fuel economy.

   The vast rebound literature has reported a wide range of estimates of this elasticity, which imply that a 1 percent fuel economy increase raises driving 0.1 to 0.8 percent. Most estimates fall in the range of 0.1 to 0.3, and many recent estimates have fallen toward the lower end (UKERC 2007 and US EPA 2011). Based on the results of several recent studies, the U.S. government used an elasticity of 0.1 for estimating the fuel savings of the upcoming fuel economy standards.

   In the context of rising fuel economy standards, I define the fuel economy rebound effect as the percent change in miles traveled caused by a 1 percent increase in fuel economy (in this context, with fuel prices held constant, this definition is equivalent to the response to a 1 percent reduction in fuel costs; for a broader discussion of the rebound effect see Gillingham et al. 2013 and Borenstein, 2015). Much of the rebound literature has used micro data (household or vehicle-level data) to estimate the magnitude of the rebound effect and has faced several major challenges. First, because households choose the fuel economy of their vehicles, fuel economy may be correlated with other attributes of the vehicle or household that are hard to control for and which may bias econometric estimates of the rebound effect (Dubin and McFadden 1984). Second, the short-run rebound effect probably differs from the long-run rebound effect because higher fuel economy may induce gradual responses--such as carpooling, moving, or changing jobs. The long-run rebound effect, rather than the short-run effect, is relevant to welfare analysis of fuel economy standards, but estimating long-run rebound introduces the typical challenges of estimating long run responses while controlling for other factors that affect VMT, such as income.

   In this paper, I argue that, to avoid these challenges, every study in the rebound literature using micro data has made at least one of three assumptions about consumer behavior. The paper's contribution is to compare the estimates when imposing these assumptions with estimates that simultaneously relax these assumptions. The first assumption is that fuel economy is uncorrelated with vehicle or household attributes that affect a consumer's utility from driving. Studies using micro data that do not control for other vehicle characteristics, such as engine power or reliability, implicitly assume that fuel economy is uncorrelated with the other vehicle characteristics. However, Klier and Linn (2012) argue that because of the vehicle design process, fuel economy is likely to be correlated with attributes that can be measured (such as power) and attributes that are harder to measure (such as reliability). If they are correlated with miles traveled, failing to control for these attributes would bias empirical estimates of the rebound effect. Likewise, many studies using micro data fail to account for unobserved household characteristics, which would bias the results if such characteristics are correlated with miles traveled.

   The second assumption maintained in nearly all of the rebound literature is that, for multivehicle households, the VMT for one vehicle is independent of the VMT for another vehicle belonging to the same household. Or, in other words, the fuel economy of one vehicle is uncorrelated with the fuel economy and other attributes of the household's other vehicle(s). This seems unlikely, however, if the use of a vehicle for a particular purpose depends on its fuel economy. For example, a household may use a small car for a long commute and a large sport utility vehicle (SUV) for local shopping trips. With the exceptions of Greene et al. (1999), Feng et al. (2013), and Spiller (2012), econometric analysis of the demand for VMT and gasoline treats each of a household's vehicles as an independent observation; some studies, such as Frondel and Vance (2012) confine their analysis to single-vehicle households.

   The third assumption is that VMT responds similarly to gasoline prices and fuel economy. Recent and careful analysis of consumer driving behavior (e.g., Gillingham 2013 and Knittel and Sandler 2013) account for unobserved vehicle and household characteristics but their focus is on the elasticity of driving to fuel prices as opposed to fuel economy. Such an analysis only yields an accurate estimate of the fuel economy rebound effect if consumers respond by equal and opposite amounts to fuel prices and fuel economy. This assumption may not hold in practice for a variety of reasons such as differences in the persistence or uncertainty of gasoline price or fuel economy shocks. For example, if consumers expect gasoline price shocks to be temporary and changing VMT (e.g., by arranging for carpooling) has fixed costs, VMT would respond less to a gasoline price decrease than to a proportional fuel economy increase. Of the few studies that estimate the effect of fuel economy on VMT, Gillingham (2012) finds that fuel economy affects VMT less than fuel prices; Greene et al. (1999) and Frondel et al. (2012) report no difference.

   It is noteworthy that a simple approach to accounting for unobserved vehicle characteristics--including vehicle fixed effects in the estimation--cannot be used to estimate the fuel economy rebound effect unless the third assumption is valid. Many recent studies (e.g., Knittel and Sandler 2013) include vehicle fixed effects to control for unobserved vehicle characteristics. In such cases, the effect of fuel costs on driving is identified entirely by fuel price variation; this causes no problems if the objective is to estimate the effects of fuel prices on driving behavior, but it only yields unbiased estimates of the fuel economy rebound effect if consumers respond by equal and opposite amounts to fuel prices and fuel economy.

   In short, recent analysis of gasoline prices and household-level driving behavior, while addressing the first assumption and in some cases the second, does not provide an accurate estimate of the fuel economy rebound effect unless the third assumption is valid. Past studies that attempt to directly estimate the fuel economy rebound effect impose at least one of the first two assumptions. Greene et al. (1999) is perhaps the closest to relaxing all three simultaneously, but nevertheless does not account for unobserved vehicle characteristics. (2)

   This paper illustrates the empirical consequences of simultaneously relaxing these assumptions. All three assumptions introduce bias for previous estimates of the rebound effect, and the direction of the bias in each case is theoretically ambiguous. I use recent household survey data to relax the three assumptions. Using the 2009 National Household Travel Survey (NHTS), I estimate the effects on VMT of gasoline prices and fuel economy. The dependent variable is a vehicle's VMT, and the independent variables include the current gasoline price, the vehicle's fuel economy, and household and vehicle characteristics. I compare two approaches to relaxing the first assumption about the correlations between fuel economy and unobservables. First, a few studies (e.g., Gillingham 2013) control for vehicle characteristics or include vehicle model fixed effects in a linear regression. However, I report evidence that, in the NHTS sample, fuel economy is correlated with household characteristics after including such controls, which suggests that omitted household characteristics may also be correlated with fuel economy. This possibility motivates a second approach, which is to instrument for fuel economy using the gasoline price at the time the vehicle was obtained. This approach, which is similar to that of Allcott and Wozny (forthcoming), rests on the strong correlation between...