Gasoline Prices, Fuel Economy, and The Energy Paradox

GASOLINE PRICES, FUEL ECONOMY, AND THE ENERGY PARADOX
Author(s): Hunt Allcott and Nathan Wozny

Source: The Review of Economics and Statistics , December 2014, Vol. 96, No. 5
(December 2014), pp. 779-795
Published by: The MIT Press
Stable URL: https://www.jstor.org/stable/43554957
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms
The MIT Press is collaborating with JSTOR to digitize, preserve and extend access to The
Review of Economics and Statistics
This content downloaded from

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
All use subject to https://about.jstor.org/terms
The Review of Economics and Statistics
VOL. XCVI DECEMBER 2014 Number 5
GASOLINE PRICES, FUEL ECONOMY, AND THE ENERGY PARADOX

Hunt Allcott* and Nathan Wozny
Abstract - Policymakers often assert that consumers undervalue futureSimilarly, it is often asserted that gasoline costs are not
gasoline costs when they buy automobiles. We test this by measuring
fully
whether relative prices of vehicles with different fuel economy ratings fully
salient to automobile consumers when they choose
among
adjust to variation in gasoline prices. Vehicle prices move as if consumers are automobiles with different fuel economy ratings
(Greene et al., 2005). If this is true, consumers buy vehi-
indifferent between $ 1 .00 in discounted future gas cost and $0.76 in vehicle
purchase price. We show how corrections for endogenous market shares clesand with lower fuel economy and higher resulting fuel costs
utilization, measurement error, and different gasoline price forecasts affect
than they would in their private optima. In 2007, the median-
the results. We also provide unique evidence of sticky information: vehicle
income American household spent $2,400 on gasoline, and
markets respond to changes in gasoline prices with up to a six-month delay.
consumers spent $286 billion in total (U.S. BLS, 2007).
I. Introduction
Misoptimization over such a large expenditure class could
cause substantial welfare losses. The purported undervalua-
tion of future gasoline costs would also help explain what
IN among
amongsome
products
products
are situations,
less attentive
are tolessancillary
it appears
costs
attentive
than tothat to ancillary consumers costs choosing than to Jaffe and Stavins (1994) call the "energy paradox": con-
purchase prices. Consumers on eBay, for example, are less sumers and firms are puzzlingly slow to make seemingly
elastic to shipping and handling charges than to the listed pur- high-return investments in energy efficiency.
chase price (Hossain & Morgan, 2006). Mutual fund investors Externalities from energy use related to national secu-
appear to be less responsive to ongoing management fees than rity and climate change would add to these potential private
to upfront payments (Barber, Odean, & Zheng, 2005). Senior losses. Policymakers have long debated whether it is prefer-
citizens are two to five times more sensitive to a Medicare Part able to address these externalities through gasoline taxes or
D plan's premium than to its out-of-pocket costs (Abaluck &
corporate average fuel economy (CAFE) standards, which
mandate an increase in the average fuel economy of new vehi-
Gruber, 201 1). Shoppers are less elastic to sales taxes than to
cles. In the absence of other market failures or misoptimiza-
purchase prices (Chetty, Looney, & Kroft, 2009).
tion by consumers, economic analyses typically conclude that
CAFE standards are highly inefficient relative to gas taxes.1
Received for publication July 22, 201 1. Revision accepted for publication
May 15, 2013. Editor: Mark W. Watson. If consumers undervalue fuel costs when they choose among
* Allcott: New York University and NBER. vehicles, however, CAFE standards can increase welfare, as
We thank Alberto Abadie, Soren Anderson, Josh Angrist, Orley Ashen- they effectively force consumers to buy the energy-efficient
felter, David Autor, Lucas Davis, Henry Färber, Kelly Gallagher, Penny
Goldberg, Michael Greenstone, Jerry Hausman, Mark Jacobsen, James vehicles that they would want if they were optimizing. This
Kahn, Ryan Kellogg, Lutz Kilian, David Laibson, David Lee, Henry Lee, "paternalistic" argument for energy-efficiency policies has
Robin Lee, Lars Lefgren, Erin Mansur, Erich Muehlegger, Sendhil Mul- long been employed by both academic economists (Haus-
lainathan, Richard Newell, Ariel Pakes, Nancy Rose, Jesse Rothstein,
man, 1979)2 and the U.S. government.3 Put simply, while
Stephen Ryan, Jim Sallee, Jorge Silva-Risso, Rob Stavins, Chris Timmins,
Roger von Haefen, Glen Weyl, Matt White, and seminar participants at the
Air Force Academy, Analysis Group, Carnegie Mellon, Colorado School of 1 Jacobsen (2010), for example, shows that the CAFE standard has a wel-
Mines, Harvard, Kent State University, Mathematica Policy Research, MIT, fare cost of $222 per metric ton of carbon dioxide abated, compared to
NBER Summer Institute, Northwestern, Pontificia Universidad Católica de $92 per metric ton for an increase in the gasoline tax that reduces gasoline
Chile, Princeton, Society for Economic Dynamics, Stanford University, consumption by the same amount.
Stanford Institute for Theoretical Economics, Texas A&M, Universidad 2 Hausman (1979) finds that consumers implicitly use a discount rate of
de los Andes, University of California Energy Institute, University of 15% to 25% per year when they trade off purchase prices and future energy
Delaware, University of Wisconsin, University of Toronto, and Washing- costs of new air conditioners. He then argues that "this finding of a high indi-
ton University for their feedback. Funding for this project was provided vidual discount rate does not surprise most economists. At least since Pigou,
by Princeton's Industrial Relations Section, Harvard's Energy Technology many economists have commented on a 'defective telescopic faculty.' A
Innovation Policy group, the Harvard University Center for the Environ- simple fact emerges that in making decisions which involve discounting
ment, and the Sloan Foundation. We thank JD Power for providing vehicle over time, individuals behave in a manner which implies a much higher
transaction price data. Thanks to Lonnie Miller at RL Polk for helpful con- discount rate than can be explained in terms of the opportunity cost of
versations. We thank Manheim Consulting for providing auction data and funds available in credit markets. Since this individual discount rate sub-
Gary George, Nirmeet Kacheria, Pete Sauber, and Tom Webb at Manheim stantially exceeds the social discount rate used in benefit-cost calculations,
for sharing their expertise. We thank Carmen Collyns and Jordon Ricks for the divergence might be narrowed by policies which lead to purchases of
research assistance on this and related projects. We especially appreciate more energy-efficient equipment."
meticulous research assistance by Sounman Hong. Code for replicating the 3 For example, the Regulatory Impact Analysis that justifies the 201 1-
estimation is available from Hunt Allcott's website. 2015 increase in the CAFE standard argues that about $15 billion per year
A supplemental appendix is available online at http://www.mitpress in net benefits will flow to consumers who undervalue the benefits of fuel
journals.org/doi/suppl/ 10.11 62/REST_a_004 19. economy (National Highway Traffic Safety Administration, 2010).
The Review of Economics and Statistics, December 2014, 96(5): 779-795

© 2014 by the President and Fellows of Harvard College and the Massachusetts Institute of Technology
doi: 10.1 162/REST_a_00419

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
780 THE REVIEW OF ECONOMICS AND STATISTICS
that larger vehicles have

political feasibility plays an important more observed and
practical unobserved
role, pater-
amenities.
nalism is a leading economic justification for one of the most
This paper
important and costly public policies formalizes one potentially
affecting the U.S. promising
auto- alter-
motive and energy industries. native
However, one
approach, which problem
exploits with
the significant the
fluctuations
paternalistic justification for fuel economy
in gasoline price forecastsstandards is that
over the past fifteen years. The
there is not much solid evidence on exploits
approach whether automobile
the fact that a vehicle's futurebuy-
fuel costs
ers are actually misoptimizing vary
(Parry, Walls,
as a function and
of both fuel Harrington,
economy and forecast gaso-
2007). line prices, so changes in gas price forecasts should affect
The relevant null hypothesis is that consumers are will- the relative value of high- versus low-fuel-economy vehicles.
ing to pay $1.00 more to purchase a vehicle with $1.00 less Indeed, media reports and academic analyses have docu-
in total forecasted future fuel costs, discounted to present mented that as gasoline prices rise, the relative prices of low-
value at their intertemporal opportunity cost of funds. For fuel-economy vehicles drop (Busse, Knittel, & Zettelmeyer,
expositional purposes, we say that rejecting this hypothesis 2013; Langer & Miller, 2013). The above null hypothesis,
is evidence that consumers "undervalue or overvalue gasoline however, does more than predict that gasoline price fore-
costs." This hypothesis is related to a long literature, dating casts should affect vehicle demand: it predicts exactly how
at least to the energy crises of the 1970s, that estimates con- much demand should be affected. Intuitively, if relative vehi-
sumers' implied discount rates for energy-efficiency invest- cle prices are not sufficiently responsive to changes in forecast
ments and compares them to benchmark consumer discount gasoline costs, this suggests that consumers undervalue gaso-
rates.4 The typical empirical approach in this literature has line costs when they purchase vehicles. Conversely, if vehicle
been to exploit variation in the prices and energy efficiency prices respond to gas price forecasts more than theory pre-
ratings of a cross-section of energy-using durable goods. For dicts, this suggests that consumers overvalue gasoline costs.
example, our null hypothesis in a cross-sectional discrete We begin from the primitives of a discrete choice util-
choice framework would be that after conditioning on other ity function and show the empirical assumptions required to
observed product characteristics, a $1.00 increase in a prod- test the null hypothesis. We then implement this test using
uct's purchase price is associated with the same decrease in data from 86 million transactions at both auto dealerships
market share as a $1.00 increase in total discounted energy and wholesale auctions between 1999 and 2008. For each
costs. Analogous cross-sectional approaches are used in a month of this study period, these data are collapsed to the
seminal paper by Hausman ( 1 979) on air conditioners, as well average price for each new and used vehicle in consumers'
as analyses of other energy-using durables such as houses choice sets.6 Each vehicle has a different present discounted
(Dubin, 1992), water and space heating (Dubin & McFad- value (PDV) of fuel costs, depending on its fuel economy
den, 1984), and autos (e.g., Dreyfus & Viscusi, 1995; Espey rating, future survival probabilities, gasoline price forecasts,
& Nair, 2005; Goldberg, 1998). discount rates, and annual vehicle-miles traveled. We condi-
For the cross-sectional estimator to be unbiased, the function on vehicle fixed effects, which sweep out all observed and
tional form for how other observed product characteristics unobserved characteristics, and test whether relative prices
enter utility must be correctly specified, and any unob- move one-for-one with changes in the PDV of fuel costs.
served characteristics must be uncorrelated with energy Under the assumptions that consumers' gas price forecasts
efficiency. Especially for automobiles, these assumptions match those of oil futures markets and that consumers' real
appear problematic. Fuel economy is mechanically corre- intertemporal cost of funds is 6%, auto consumers appear
lated with weight and horsepower, and it has often proven to be willing to pay only $0.76 in purchase price to reduce
difficult to separately identify preferences for these different discounted future gasoline costs by $1.00. The correspond-
characteristics.5 Furthermore, fuel economy is highly nega- ing implied discount rate, the discount rate for future gas
tively correlated with price in the cross-section, suggesting costs that rationalizes market behavior, is just under 15%.
We show that the results are robust to a number of potential
confounding factors, such as endogenous changes in utiliza-
4 The implied discount rate is simply a shorthand way of capturing how
tion patterns and market shares in response to gas prices,
people appear to trade off upfront costs against an analyst's estimate of
future cash flows. Many underlying factors other than consumers' actualchanges in underlying preferences for "green vehicles," and
changes in characteristics over model years.
discount rates might affect the implied discount rate, including time hori-
zons, beliefs, and inattention. In our setting, as in many others, these factors
However, some empirical issues are extremely important.
are empirically indistinguishable. We therefore prefer the "undervaluation
or overvaluation" language, as it is more explicitly agnostic. Regardless use the grouping estimator (Wald, 1940; Angrist, 1991)
We
of the choice of language, the analyst must eventually take a standto on address potential measurement error, and we show that
consumers' actual discount rates in order to determine whether people's
failing to do so would cause significant attenuation bias. We
decisions conform to the theoretical prediction.
5 At least since Atkinson and Halvorsen (1984), it has been pointed show
out that the result that consumers undervalue gas costs is
that the high correlation between weight and fuel economy makes it difficult
6 By vehicle , we mean a model-by-age combination, where model is a
to separately identify demand for fuel economy. In fact, cross-sectional
estimation of automobile demand in characteristic space sometimes gives a
highly disaggregated definition capturing essentially all variations in make,
nameplate, trim, body type, fuel economy, engine displacement, number of
negative sign on fuel economy, which would imply that consumers dislike
fuel economy. cylinders, and design generation.

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
GASOLINE PRICES, FUEL ECONOMY, AND THE ENERGY PARADOX 78 1
functions.
largely driven by older vehicles: prices for vehicles agedSection
1 1 III presents our data, devoting particular
attention prices,
to 15 years appear to be highly insensitive to gasoline to the construction of each vehicle's total discounted
while prices for relatively new used vehicles movegasoline
much costs.
moreSection IV details our estimation strategy.
Section
closely to the theoretical prediction. In addition, V presents empirical results and a long series of
undervalua-
tion appears to be more severe when current gasrobustness checks. Section VI details the evidence of sticky
prices instead
of oil futures are used as a proxy for consumers' forecastsand
information, of section VII concludes.
future gas prices.
The data also show clear evidence of sticky information. In
A. Related Literature
models such as Mankiw and Reis (2002), Sims (2003, 2010),
Within the broader literature on consumer choice of
and Woodford (2003), information about prices, inflation,
energy-using
and other economic variables takes time to diffuse durables7 and the effects of gasoline prices on
through
the population of consumers and firms, causing vehicle markets,8
delays in our analysis is most closely related to four
other papers.
responses to news. In this context, sticky information wouldKahn (1986) tests whether relative prices of
cause vehicle prices to adjust with a delay to used vehicles
changes fully adjust to changes in the relative discounted
in gas
present value
price forecasts. Indeed, when the model is estimated of relative gas prices caused by the gasoline
using
monthly differences instead of the de-meaning priceestimator,
shocks of the 1970s and 1980s. A working paper by Kil-
vehicle markets appear to be highly insensitive ian and current
to the Sims (2006) builds on Kahn's approach with updated
month's change in gas price forecasts. Four- to data. The keylags
six-month difference between our analysis and these two
of gas price changes are conditionally associatedpapers is that
with the we use transaction prices from auctions and
cur-
rent month's change in vehicle prices, suggestingdealerships instead of data from used-car price guides such
that sticky
information keeps vehicle markets from updating as the Kelley
immedi- Blue Book or the National Auto Dealers ' Asso-
ciation Used Car Guide . It is crucial to use transaction data
ately. These results add a unique microdata-based analysis
to the limited set of empirical tests of the sticky becauseinforma-
used-car price guides reflect the opinion of a small
tion model, which includes Coibion (2010), Khan team ofandanalysts
Zhu who may or may not fully adjust their price
assessments for each vehicle to reflect current market con-
(2006), Klenow and Willis (2007), and Mankiw and Reis
(2003). Aside from being of interest in macroeconomics andditions. Data from used-car price guides could cause us to
finance, sticky information also has important implications falsely conclude that vehicle market prices do not fully adjust
for how the undervaluation hypothesis can be tested: identi- to changes in gasoline prices.
fying off of high-frequency gas price variation can cause the Sallee, West, and Fan (2009) exploit transaction data from
econometrician to falsely conclude that consumers under- used-vehicle auctions to test whether vehicle prices move
value fuel economy as a product attribute when they are one-for-one with the present discounted value of future gaso-
instead inattentive to changes in gas prices. line costs. Their identification strategy is different from and
Before continuing, it is worth doing a simple calibration complementary to ours: they exploit the fact that autos with
to demonstrate why it is difficult to overstate the importance different odometer readings have different remaining life-
of this question. Substantial volumes of academic research times, and an increase in gas prices has a larger effect on
and policy discussions have centered on the welfare losses the remaining present discounted value (PDV) of gas costs
from transport sector carbon emissions and the costs and for autos with lower current odometer readings. One ben-
benefits of different policy responses. Using the U.S. gov- efit of their approach is that it allows a more extensive set
ernment's estimated marginal damage of carbon emissions of fixed effects than we use. By construction, however, they
(Greenstone, Kopits, & Wolverton, 2011), gasoline con- are testing a different empirical hypothesis: they ask whether
sumption imposes an externality of $0. 1 8 per gallon, or about consumers correctly value differences in odometer readings
5% of the current gasoline price. Thus, if the carbon exter- within vehicle models when gas prices change, while we ask
nality is not internalized, consumers account for about 95%whether consumers correctly value differences in fuel econ-
of the total social cost of gasoline when they choose among omy across models when gas prices change. Conceptually,
vehicles with different fuel economy ratings. By comparison, our hypothesis is closer to the policy-relevant question of
an implied discount rate of 15%, well within the range of whether consumers undervalue energy efficiency.
Hausman (1979) and other estimates in the literature, would
7 This literature includes Hausman (1979), Dubin and McFadden (1984),
suggest that consumers value only 75% of the cost of gasoline
Davis and Kilian (201 la), Beresteanu and Li (201 1), and others.
when they choose vehicles. Thus, while undervaluation and 8 Other work that examines how vehicle pnces adjust m response to gaso-
uninternalized carbon externalities distort vehicle purchases line prices include Sawhill (2008), Langer and Miller (2013), and Austin
in the same direction, inducing consumers to buy vehicles (2008). Verboven (1999) estimates the discount rates implied by differ-
ences between the prices of gasoline and diesel vehicles in Europe. Ohta
that use more gas than in the first best, undervaluation could and Griliches (1986) examine whether the gasoline price shocks in the
generate distortions several times larger than the distortions 1970s affected consumers' valuations of vehicle characteristics. Edelstein
and Kilian (2009) document how gas prices affect consumer expenditures,
from climate change externalities.
including demand for domestic versus foreign automobiles. Li, Timmins,
The paper progresses as follows. In section IA, we dis- and von Haefen (2009) and Klier and Linn (2010) test how the price of
cuss related literature. Section II models consumers' utility gasoline affects sales of new vehicles with different fuel economy ratings.

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
Variable Mean SD Minimum Maximum Number
Year 2003.5 2.9 1999 2008 1,068,459

Model year 1996 5.2 1984 2007 1,068,459
Age 7.5 4.2 1 15 1,068,459
Price ($) 7,227 9,845 0 358,289 1,068,459
Quantity 18,537 27,612 0 370,512 1,068,459
PDV of gas cost ($) 7,163 4,147 686 41,576 1,068,459
Fuel economy (MPG) 18.8 4 6.9 52.9 1,068,459
Horsepower 174 61 46 700 1,024,709
Weight (pounds) 3,440 838 2.2 34,903 1,022,499
Wheelbase (inches) 110 13.9 79.9 972 1,025,023
Number of transactions 37.8 122 1 9,679 1,068,459
The number of observations differs from the base specification because this table includ
Unlike the regressions, this table is not weighted by number of transactions within an o
Busse et al. (201 3) estimate

utility for how
vehicle changes
ja across all timein gasoline
periods, and %jat
affect equilibrium prices and quantities of
'| fja is the period-specific deviation.new and
cles in different quartiles of the
Consumer fuel
/' s indirect economy
utility from vehicledistr
ja at ti
They then plug these estimates into an Excel spr
which outputs the implied discount
Mij at = "H (W/ Pjat rates
yGjat) +at
tyjawhich
~'~ %jat H"t
prices of vehicles in each MPG quartile fully adjust
The marginal
in gasoline costs. Depending utility ofassumptions,
on their money is iļ. The valuation
t
implied discount rates on fuel costs is y:
ranging if consumers
from -6.2%value purchase
to 20. p
using assumptions that discounted
correspond fuel costs equally,
most then y = 1.
closely to If o
c
undervalue
find an implied discount rate or overvalue
for usedfuel costs, then y < 1
vehicles
respectively. This framework
Our analysis differs by being somewhat more is analogous to the
fo
weight models in Chetty
example, we specify the undervaluation hypothesiset al. (2009), Della Vig
primitives of a utility and other analyses.
function, We assume helps
which that both in
r' andund
y are
ing the economic meaning In specifying
of the the identifying
basic model, we assumeass
takes the extreme
We also uncover additional empirical value distribution,
findings, giving the
su
representative-consumer
evidence of sticky information in section logit model.
VI. Integratin
distribution of € gives a market-level relationsh
prices and shares, denoted by s. Typically this i
II. Model
arranged to give the term In sjat - In sot on the left
In this section, we specify a static discrete choice model. In
the online supplement, Appendix A, we show In Sjat
that ~ In
under Sot = -TÌ Pjat - wGjat + tyja +
the
assumption of stationarity, this simple model can be derived
Of course, this identity can also be rearranged wi
from a dynamic model in which consumers buy and resell
on the left:
their vehicles. Consumers derive utility from owning a vehi-
cle and consuming a numeraire good. We define a vehicle as
a model-by-age combination, where j indexesPjat = yGjat
models and-a(In Sjat hl5of) H~ ty ja H" %
indexes age in years. Consumers also can choose an outside
option, denoted j = 0, which is to own no vehicle. The util-
This equation includes new variables 'ļ rja = ty
ity of this outside option is normalized to 0. Consumer
%jat = %jat/ti has
l, which represent the dollar value of
budget constraint ity represented by tyja and %jat. Section IV begins
The purchase price of vehicle ja at time t as pjat . Gjat
equation is the
in presenting our empirical strategy.
PDV of future gasoline costs over the vehicle's remaining
life. The average consumer's utility from owning and using III. Data
vehicle ja over its remaining life is tyjat' we call this the
average usage utility. Individual i's usage utility also set
Our data hasincludes
an average prices, quantities
acteristics
unobserved deviation from average usage utility, denoted of all used passenger vehicle mod
€^ř.
The variable tyja captures the mean valuein ofthe United
average States, in monthly cross-section
usage
ary 1999 to December 2008. Table 1 presents
statistics.
9 This 13% figure is from table 7 of Busse et al. (2013). It is the mean
Fuel economy data were obtained from the
of the implicit discount rates for the three different quartile pairs for used
mental
vehicles when using vehicle-miles traveled and survival Protection
probabilities Agency
(EPA), which meas
from
NHTSA, which uses the same underlying data asover a standardized laboratory drive cycle an
we do.

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
GASOLINE PRICES, FUEL ECONOMY, AND THE ENERGY PARADOX 783
light trucks
the results to account for the typical consumer's actual but
in- not motorcycles, cutaway motor homes,
use fuel economy. For all model years, we uselimousines,
the EPA's camper vans, chassis cab and tilt cab pickups,
most recent adjustment, which was calculated hearses,
in 2008.and Like
other unusual vehicles, where we expect the
Greene et al. (2007), we also assume that fuel economy
substitution elasticity to be very small. In our base specifi-
degrades over each vehicle's life at 0.07 MPGcation,per year.
we also exclude cargo and passenger vans as well as
Vehicle class designations - pickups, sport utility vehicles,
ultraluxury and ultrahigh performance exotic vehicles10 due
minivans, vans, two-seaters, and five classestoof cars
their low based
substitutability with the rest of the market. It turns
on interior volume - are also taken from the EPA's fuel out that including these vehicles does not change the results.
economy data set. All other vehicle characteristics, includ- We define a model j to capture all possible variation in
ing horsepower, curb weight, wheelbase, and manufacturer'seconomy ratings and much of the observable varia-
fuel
suggested retail price (MSRP), are from Ward's Automotive tion in prices. This is more disaggregated than a nameplate,
Yearbook . All dollar figures in this paper are real Julywhich 2005 refers to a colloquial name such as "Ford Taurus" or
dollars, deflated using the Bureau of Labor Statistics "Honda
Con- Civic." We define a model at the level of make, name-
sumer Price Index series for All Urban Consumers, All plate, Items trim, body type, fuel economy, engine displacement,
Less Energy. and number of cylinders. As a result, the average make and
Vehicle prices are based on microdata obtained from Man- nameplate combination in our data set has seven different
heim, a firm that intermediates approximately half of auto models. For example, there are eleven different configura-
auction transactions in the United States. The principal sellers tions of model year 2004 cars called the "Honda Civic" that
are dealerships, rental car companies, and auto manufacturers appear in our data set as separate "models," including coupe
that are reselling off-lease vehicles. Buyers are typically deal- and sedan versions of the DX, EX, and LX; the Si Hatchback;
erships, which then retail the used vehicles. We observe each the Civic Hybrid; and several others.
of Manheim's approximately 45 million transactions between Vehicles with the same model name are typically offered
1999 and 2008. While only about one in four used vehicles for several consecutive model years, although some are
traded passes through an auction (Manheim, 2009), the auc- offered for many more. As extreme examples, both the Ford
tion market is the largest source of transaction price data. F- 1 50 and Honda Civic have been offered in every model year
Furthermore, the Kelley Blue Book and other price guides, since 1973. Of course, the 1973 and 2008 versions of these
the starting point for price negotiations in many nonauction models are very different. Every several years, auto manufac-
transactions, are largely based on auction prices. To get pjaU turers redesign their models and define a new generation of a
we simply collapse the data to the level of mean price for vehicle. For example, Honda introduced new generations of
each vehicle in each month after excluding vehicles rated as the Civic for the 1980, 1984, 1988, 1992, 1996, 2001, 2006,
low quality or scrap quality. and 2012 model years. A generation is a well-defined con-
Some alternative specifications employ prices from the JD cept, and the generation definitions for each vehicle can even
Power and Associates' (JDPA) Power Information Network, be found on Wikipedia. Within each generation, models do
which collects detailed microdata on approximately 31% of not change significantly. We redefine each new generation of
U.S. retail auto transactions through a network of more than a car or truck as a separate model j in our data set.
9,500 dealers. For each vehicle ja , we obtained monthly mean
prices adjusted for customer cash rebates and, if the transac- A. Discounted Gasoline Costs
tion included a trade-in, the difference between the trade-in
vehicle's actual resale value and the negotiated trade-in price. The variable Gjat is the present discounted value of lifetime
We use the Manheim used- vehicle price data in our base spec- gasoline costs over future years y9 for the average driver of
ifications because these data include more than twice as many vehicle ja at time t:
observations at the jat level, while there are fewer than 1,000
observations in JDPA that are not in Manheim. L
We observe national-level registered quantities of each Gjat = ^ & ' g y ' mjay ' fjay ' $jay (4)
vehicle from 1999 through 2008 in the National Vehicle
Population Profile, a data set obtained from market research
firm R. L. Polk. The quantities represent all vehicles reg- L denotes the maximum possible lifetime of a vehicle,
istered to private individuals and to fleets such as taxi andwhich we take to be 25 years. The variable gy is the gaso-
rental car companies and corporate and government motor line price forecast for year y, mjay is expected vehicle-miles
traveled,^ is fuel economy in miles per gallon, ('>jay is the
pools. These quantity data are matched to the price data using
vehicle-specific serial numbers called vehicle identificationprobability that the vehicle survives to year y conditional on
numbers.
We define the choice set to include all "substitutable" 10 These excluded exotic vehicles are the Acura NSX, Audi R8 and TT,
Chrysler Prowler and TC, Cadilliac Allante and XLR Roadster, Chevro-
used gasoline-fueled light-duty vehicles that have EPA fuel
let Corvette, Dodge Viper and Stealth, Ford GT, Plymouth Prowler, and
economy ratings and are less than 25 years old. By all
substi-
vehicles made by Alfa Romeo, Bentley, Ferrari, Jaguar, Lamborghini,
tutable , we mean cars, pickups, SUVs, minivans, and other
Maserati, Maybach, Porsche, Rolls-Royce, and TVR.

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
surviving to its current age a , and 8- Calculating

Table 2. is thethe annual discount
Average Discount Rate
factor. Payment Method Share of Vehicles Discount Rate

Our parameter of interest y is the coefficient on this vari- Financed 37% 6.9%
able, and the decisions we make here mechanically affect Leased 0% NA
Cash 63% 5.8%
the parameter estimates. For example, using a lower discount
Weighted average 6.2%
rate than consumers actually face would inflate Gjat , thereby
Share of vehicles and the finance discount rate are rea
biasing y toward 0. Alternatively, understating a vehicle's of Consumer Finances. Cash discount rate is the real av
2008, inclusive.
expected lifetime or usage would deflate Gjat , biasing y away
from 0.
rate (APR). For consumers wh

Vehicle-miles traveled and survival probability. To from savings, the opportunity c
estimate vehicle-miles traveled (VMT), we use publiclybe realized on savings.
available data from the 2001 National Household Travel
Table 2 details the calculation
Survey (NHTS), a nationally representative survey of approx-
2004, and 2007 Surveys of Co
imately 25,000 households that includes the age, fuel econ-
estimate that 37% of used veh
omy, and vehicle class for each vehicle owned. As part of the
are paid for in cash. We transfo
survey, about 25,000 vehicles in the national sample had
thetheirSCF from nominal to real by
odometers read twice, with several months between readings.
tations implied by Treasury inf
These two readings were then used to estimate annualized
the SCF, the average real interes
VMT. We regress annualized VMT on class dummies
isand
6.9%. For vehicles purchase
vehicle age and use these estimates to fit mjay for all the
vehiclesopportunity cost of funds i
in our sample. When we adjust for uncertainty in the first-
the average real return on th
step estimates of Gjat (Murphy & Topel, 1985), we 2008
use the (inclusive) was 5.8%. As
Huber- White robust standard errors from this regression.
these two discount rates weight
Our base specification assumes that this fitted mjay does
tions gives a 6.2% discount rat
not depend on the gasoline price forecast in year y our
. Eitherbase specification.
of two separate arguments can justify this. First, empirical
While we think that this is the
estimates of VMT demand show that it is relatively inelastic
forward approach, it surely is
(Hughes, Knittel, & Sperling, 2008; Small & Van Dender,
real average used- vehicle loan
2007; Gillingham, 2010; Bento et al., 2009; Kilian & Murphy,
through the JD Power Informat
201 1 ; Davis & Kilian, 201 lb). Second, the envelope centage
theorem points higher than the
can be used to show that since indirect utility Uijat is a SCF.
function On the other hand, includin
of an optimized value of VMT, changes in VMT caused2008by would give lower stock m
marginal changes in gasoline prices have only second-order
if we modeled consumers wit
effects on u^t and vehicle prices. In appendix C, however,
of consumption, they would wan
we derive an alternative approach that makes mjay a function
covariance with the market. Bec
of gy and allows for nonmarginal changes. The results line areprices have very low correla
very similar to the base specification.
adjustment using the capital as
We use an analogous approach to fit vehicle survival
anproba-
interest rate close to the real risk-free rate: about 1.6%. In
bilities based on the NVPP registered quantity data. We
the use a section, we show the sensitivity of y to alternative
results
grouped data probit model, where the outcome variable is the
discount rates.
number of vehicles of a model and model year registered next
year divided by the number of vehicles registered today. We
Gasoline price forecast. We use three measures of vehi-
estimate coefficients on vehicle age dummies, model year,
cle consumers' gasoline price forecasts: seasonally adjusted
and fuel economy, with robust standard errors clustered by
retail prices, oil futures prices, and beliefs elicited via sur-
vehicle, and then we predict survival probabilities <'>jay forThe
veys. all seasonally adjusted retail price measure is equiva-
vehicles in our sample. lent to assuming that consumers forecast that gasoline prices
are a martingale with seasonal trends. To construct this mea-
Discount Rate. The discount rate r = 8_1 - 1 is the
sure, we take the U.S. city average motor gasoline retail prices
intertemporal opportunity cost of money: the rate at which
for all types of gasoline from the U.S. Energy Information
a vehicle's purchase price is amortized over future years or Administration's (2011) Monthly Energy Review and con-
future gasoline costs are discounted to the present. To imple-
vert to real July 2005 dollars.11 We then eliminate seasonal
ment this empirically, we calculate the weighted average 8trends by regressing the level of monthly gasoline prices
across used vehicle buyers. For vehicle buyers whose mar-
ginal dollar comes from a loan or lease, the opportunity cost 11 Edelstein and Kilian (2009) and Kilian (2010) also use data on the real
of paying more to purchase a vehicle is the annual percentageprice of gasoline.

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
Figure 1. - Retail Gasoline and Futures Prices

on eleven-month dummies, subtracting the fitted values,
and readjusting so that the mean monthly gasoline price is
unchanged.
Our second measure assumes that oil futures markets
reflect vehicle consumers' gasoline price forecasts. To con-
struct this measure, we use light sweet crude oil spot prices
from the U.S. Energy Information Administration's (2011)
Monthly Energy Review and futures prices from the New York
Mercantile Exchange and the Intercontinental Exchange. The
oil futures prices are transformed into current dollars using
inflation expectations implied by Treasury inflation protected
securities, then deflated into real July 2005 dollars, then
transformed to dollars per gallon of gasoline using the aver-
age historical relationship between oil and gasoline prices.12
(Appendix table Al reports annual average retail gas prices
Seasonally adjusted retail gasoline prices and the gasoline prices implied by oil futures at differen
settlement horizons.
and futures-based gas price forecasts.) In every year of our
sample, there are futures trades for settlement dates up to
six to seven years in the future, and in later years, tradesare
are substantially more volatile from month to month. Thi
observed for dates as far as ten years out.13 is not a spurious result of adjusting for seasonality: retail
Our third measure is derived from the Michigan Survey
prices are also more volatile when we do not condition on
of Consumers, which elicits beliefs about future gasoline
month-of-year effects. As we shall see, this will cause y t
prices, inflation, and other economic variables from a nation-
differ substantially between the two forecasts when using a
ally representative sample. Respondents are asked whether monthly difference estimator.
they think the price of gasoline will go up, go down, orIt is important to acknowledge several limitations of oi
stay the same during the next five years. Those who think
futures price data in light of recent work in this area. First,
it will change are asked how much the price will increase
Kilian (2010) has shown that in the short run, there can b
or decrease. Anderson, Kellogg, and Sallee (2011) analyze
important deviations between crude oil and gasoline prices
these data, and interested readers should refer to their paper
Over our sample period, monthly oil spot prices predict 94%
for more details. For our analysis, the authors provided
ofusthe variance in monthly gasoline prices, not 100%. How
with the mean real gasoline price forecast across all respon-
ever, what matters for our analysis is the expected future
dents for each month between 1999 and March 2008. Each
relationship between oil and gasoline: oil futures could be
respondent's forecast is deflated to current dollars as of thea good measure of long-run gasoline price forecasts even i
date of the survey using his or her own inflation expectation.the short-run relationship does not have an R2 of 1. Second,
Figure 1 plots seasonally adjusted retail gasoline prices
Alquist and Kilian (2010), Alquist, Kilian, and Vigfusson
along with the gasoline prices implied by oil futures at three
(2013), and Baumeister and Kilian (2011) have shown that
different horizons. The figure illustrates that futures prices
oil futures prices are typically not as good as current oil prices
differ from retail prices in two important ways. First, the at predicting actual future oil prices at both short-term hori
futures market does not believe that gasoline prices are zons
a of twelve months and longer-term horizons of up to
martingale. Especially when spot prices spiked in mid-2008,
seven years. This suggests that factors other than oil pric
the futures market expected prices to eventually decrease atexpectations can affect futures markets, which would intro-
least slightly. As a result, G averages 7% larger under the duce noise in our estimates. Third, oil futures are traded with
martingale forecast than the futures forecast. This suggestshigh liquidity only for settlement dates less than two to three
that the y" estimated under the martingale forecast will years
be in the future. While this is a concern, it is not clear if
smaller than the y with the futures-based forecast. Second, this would bias the estimates in any particular direction.
while the two series follow the same trends, retail prices
B. Vehicle Price Trends
12 Specifically, this average historical relationship is predicted with a sim-
ple linear regression of levels of the motor gasoline retail prices on the levels
Three related facts about vehicle prices have important
of light sweet crude oil prices. The reason we transform oil futures prices
implications for our empirical strategy. First, as figure 2
into gasoline prices instead of directly using gasoline futures is that gasoline
futures are typically not traded for time horizons longer than three years, illustrates, low-MPG vehicles are more expensive than high-
while oil futures are traded, albeit with limited liquidity, at time horizons
MPG vehicles. The figure shows the results of a regression of
of up to ten years.
manufacturer's suggested retail price on a set of MPG indica-
13 To model expectations for periods beyond the last settlement date
tor variables for all models of model years 1984 through 2008
observed at each time t , our base specification uses a simple model of mean-
reverting expectations, where deviations from a $ 1 .50 per gallon mean decay
in the Ward's Automotive Yearbook data. Between 15 and
exponentially using a mean reversion parameter calibrated using all futures
data since 1991. The equation fits the data very well: it explains 85% of30theMPG, which includes nearly all of our data set, MSRPs
variation in the observed futures prices over our 1999-2008 study period. decline from $30,000 to $11,000. This figure corroborates

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
Figure 2. - Manufacturer's Suggested Retail Price versus Fuel

prices within vehicles and removing seasonal trends using
Economy
month dummies, prices declined somewhat from 1999 to
2003. The decline from 1999 to 2002 appears to be driven by
a decrease in the absolute price level for low-MPG vehicles.
This may have been precipitated by an increase in the num-
ber of new low-MPG vehicles sold during the mid- and late
1990s, which appears to have caused their prices to drop as
they became available for resale as used vehicles. Manheim's
analysts attribute the nadir in 2003 to the 2001 economic
slowdown, during which both low- and high-MPG vehicles
were offered at attractive lease terms and an unusually large
share were leased instead of sold. When this larger volume
of vehicles came off lease two years later, this depressed
resale prices. We therefore document whether the estimates
are sensitive to excluding the period before 2004.
De-meaned and seasonally detrended prices remained rel-
The mean manufacturer's suggested retail price and the standard error of the estimated mean for each
integer MPG rating. Data include all new vehicles from model year 1984 through 2008.
atively steady from 2004 to 2007 before dropping sharply as
Figure 3. - Vehicle Price Trends
the 2008 recession took hold. It is possible that this reces-
sion differentially affected low- versus high-MPG vehicles,
for example, by differentially affecting people of different
income levels, who tend to buy different types of vehicles.
We therefore eliminate from our base specification all data
beginning with April 2008, when the recession began to
affect vehicle market prices. Similarly, we would certainly not
want to include data from 2009, when the Cash for Clunkers
program significantly changed used-vehicle markets, dif-
ferentially affecting prices of low-MPG versus high-MPG
vehicles.
IV. Estimation
A. Empirical Strategy
The single line shows the mean vehicle price across all transactions in our data set. The de-meaned
and detrended series is the average price conditional on vehicle fixed effects and month dummies. The
futures-based gas forecast series is the average of oil futures prices for all future years, transformed into
dollars per gallon of gasoline.
We use vehicle fixed effects and look within the same vehi-
cle over time as gasoline price forecasts change. The benefit
of a panel
concerns about the traditional cross-sectional approaches to is that the fixed effects soak up unobserved vehicle
estimating y: low-MPG and high-MPG vehicles are characteristics
sys- that may be correlated with MPG: formally,
our estimator
tematically different, and some of these differences may be is still unbiased even if £[G'|/] ^ 0.
unobservable. The figure also means that we must control Wefor
move from equation (3) to our base specification
factors such as depreciation that differentially affect vehicles equation in three steps. First, we define an econo-
estimating
with higher price levels and might be correlated with metric error term that contains % and the market share term:
gas
prices. e = (- 1/rļ) In Sjat + %jat. Second, we add time dummies x,,
Second, figure 3 shows that vehicle prices have system- which absorb the outside option share sot and any shift in the
atic seasonal patterns, and they tend to decrease within each overall market price level. Third, we define/50 as the median
year. Because this systematic depreciation is approximately of the MPG distribution, and we include vjat , eleven indicator
a percentage of price and lower-MPG vehicles have higher variables for month-of-year interacted with 1 (fja > /50). As
prices, these patterns affect the price levels of low-MPG vehi- we discuss momentarily, these control for the differential sea-
cles more than high-MPG vehicles. Since gas price forecasts sonal trends in the prices of low- versus high-MPG vehicles.
rise on average during the study period, failing to account for Our base specification is
within-year trends that affect low- versus high-MPG vehicles
differently could cause us to falsely attribute seasonality to Pjat - Y G jat H" It H" y jat ~l" ^ja ~f~ ^jat- (5)
adjustments caused by gas price forecasts.
Third, there are other underlying shifts in seasonally Equation (5) is quite intuitive: it tests whether the relative
adjusted average prices that could affect low-MPG versus vehicle prices move one-for-one with changes in the rela-
high-MPG vehicles differently. If correlated with changes in tive PDV of gasoline costs. If vehicle prices do not respond
gasoline prices, such shifts could bias our estimator. The dou- sufficiently to gasoline costs, we conclude that the market
ble line at the bottom of figure 3 shows that after de-meaning undervalues gasoline costs. If vehicle prices respond more

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
Figure 4. - Identifying Variation in Gas Cost limit our sample to used vehicles less than or equal to 15 years
old. In appendix B, we show that weakening this assumption
by allowing for endogenous market shares does not signifi-
cantly affect the estimates, largely because we have limited
our sample to the ages for which the conditional correlation
between gasoline prices and market shares is insignificant.
Assumption 2 states that changes in discounted gasoline
costs are not correlated with changes in preferences and unob-
served characteristics that vary over the different model years
within a generation of a vehicle. In section V, we discuss
and test several potential violations of this assumption and
document that they do not appear to substantially affect the
results.
The present discounted value of fuel costs under the futures-based gas price forecast for vehicles of
different MPG ratings and ages in June 2004.
B. Measurement Error and the Grouping Estimator
Jerry Hausman's (2001) "iron law of econometrics" is that

than predicted to gasoline costs, we conclude that the market
overvalues gasoline costs.14 the magnitude of a parameter estimate is usually smaller
in absolute value than expected. This is due to attenuation
Figure 4 illustrates our identifying variation by graphing
bias, which can occur even in seemingly high-quality data.
the average G for an example month within age and fuel
economy groups. Younger vehicles and lower-MPG vehicles In most applications in which the null hypothesis is that a
have higher G. Equation (5) is identified by interacting thisparameter equals 0, this makes a test more conservative. By
contrast, in our test of whether y = 1, this iron law would
cross-sectional variation by the time series variation in gaso-
make us more likely to falsely reject, meaning that we could
line price forecasts illustrated in figure 1. As gasoline price
forecasts fluctuate over time, the bars in figure 4 extend and falsely conclude that consumers undervalue gasoline costs.
contract proportionally. A given change in gas price forecasts One potential source of measurement error is the EPA fuel
has a larger effect on the level of G for younger and lower-
economy ratings, as discussed by Sallee (2010). As in any
MPG vehicles, and we test for whether relative prices move other application, the use of fixed effects exacerbates attenu-
ation bias, as even small amounts of measurement error in G
correspondingly.
Equation (5) is a consistent estimator of y ifworsen the signal-to-noise ratio in G|'|/.
E[Ge|(x, v, '|/)] = 0. Because e contains both s and Š, We address measurement error using the grouping estima-
tor, which is just Wald's (1940) binary instrumental variables
this actually contains two different economically meaningful
identifying assumptions:
(IV) estimator generalized to the case with many group
indicator variables. Wald's (1940) original objective in devel-
Assumption 1: E [Gs|(x, v, 'ļ/)] = 0. oping the estimator was to address measurement error, and
Ashenfelter (1984), Angrist (1991), and others have used the
Assumption 2: E [G% ' (x, v, 'ļ/)] = 0. grouping estimator for this same purpose in various appli-
cations. In our base specification, we place each jat- level
observation into one of 222 mutually exclusive and exhaus-
In words, assumption 1 is that gasoline costs are uncorre-
cted with market shares, conditional on time dummies and tive groups: above- and below-median MPG vehicles for
each of the 1 1 1 months in the sample. This grouping retains
vehicle fixed effects. For new vehicles, this assumption would
the fundamental identifying variation, which comes from
not hold: sales of low-fuel economy vehicles decrease rela-
how time series changes in gas price forecasts affect G for
tive to high-fuel economy vehicles when gas prices increase,
high- versus low-MPG vehicles. However, identifying y off
as documented by Busse et al. (2013), Klier and Linn (2010),
and Li et al. (2009). This assumption might also not hold of group-level aggregates means that we average over any
observation-level measurement error.
for older used vehicles, whose scrappage rates respond most
to changes in gasoline prices (Li et al. 2009). Therefore, we Formally, we define Z"at = l(fja > /50) x l(t = u) as
an indicator variable for whether observation jat occurs at
time u and has above-median MPG. We denote Z '}at as the
14 Although we derive it differently, this is fundamentally the same esti-
mating equation used by Kahn (1986) and Busse et al. (2013). The latter vector of 100 Z"at variables for every month of the sample
paper uses a slightly different configuration of fixed effects and controls. beginning with January 2000, and we instrument for Gjat with
Sallee et al. (2009) allow G to vary at the transaction level instead of at the
jat level, based on the odometer reading of the specific auto being trans- ljat. Between instruments Z jat, time controls xř, and month-
acted. They can then insert yVzMeve 1 fixed effects. This allows them to relaxby-MPG group controls Vjat, we have a full set of indicator
our two identifying assumptions and substitute an alternative assumption, variables in the first stage for each of the 222 groups; the
which is that the within-yař relationship between odometer readings and
transaction prices is not affected by unobserved factors that are correlated second-stage fitted values of G are essentially group-level
with gas prices. averages.

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
Figure 5. - Raw Data Scatter Plot Figure 6. - Conditional Variation over Time
Plots of the difference in transaction-level average prices for below-median minus above median-MPG
The double line is the difference in transaction-level average prices for above-median minus below-
vehicles against the difference in the PDV of gasoline costs, using the futures-based gas price forecast.
There is one observation for each month. median MPG vehicles, conditional on vehicle fixed effects and MPG group by month-of-year dummies
The single line is the conditional difference in the PDV of gasoline costs using the futures forecast. T
dotted line is the conditional difference in the PDV of gasoline costs using the martingale forecast.
We also present results for groups at different levels

of the futures forecast. The double solid line is the differ-
of aggregation. Aggregating observations into fewer larger
ence between average p|(v, iļr), the seasonally detrended and
groups allows us to average over measurement error that
is more severe or correlated across observations. How- de-meaned transaction price, for above- minus below-median
MPG vehicles. If y = 1, the single and double lines should
ever, using fewer groups effectively reduces the amount of
move in parallel. From this figure, it is again clear that relative
identifying variation, which increases the IV standard errors.
prices are at least somewhat responsive to relative gasoline
We need not control for any variables that vary at a more
costs. However, vehicle prices do not appear to respond very
disaggregated level than Zjat. For example, consider our con-
much to the high-frequency volatility in retail gasoline prices.
trols Vjat that soak up differential monthly price trends for
To test whether prices are as responsive as theory predicts
above- versus below-median MPG vehicles. The highest-
under the martingale and futures-based forecasts, we now
MPG quartile also has different monthly price trends from
turn to the formal estimates.
the second-highest MPG quartile. However, we do not need
to control for this because the fitted values of G do not vary
between these two quartiles, and thus y is not identified offSpecification
B. Base
these differential price trends. In alternative specifications
For our base specification, we estimate equation (5) using
where we define Z jat by more disaggregated MPG groups,
we also control for differential price trends for eachthe fixed effects (de-meaning) estimator while grouping Zjat
separate
MPG group. at the level of above- and below-median MPG by time,
assuming a 6% discount rate and using data from January
1999 through March 2008. To reflect the fact that pjat is an
V. Results
average across varying numbers of transactions, we weight
the jat- level observations by the number of transactions. We
A. Graphical
are careful to call this our base specification, indicating that
Figure 5 illustrates our identification. On the the alternative
vertical axis specifications modify individual assumptions
is the difference between the mean price acrossfrom this base, but we do not call it our "preferred specifica-
transactions
for vehicles with below-median MPG versus above-median tion." This is because different readers may prefer different
MPG. On the horizontal axis is the difference in mean G. sets of alternative assumptions.
These are raw data, unadulterated by fixed effects, time con- Table 3 reports the estimated coefficients on G from
equation (5), which correspond to -y. We denote the futures-
trols, or other manipulations. Even in the raw data, it is starkly
based and martingale gas price forecasts as Gf and Gm,
visible that relative prices and relative gas costs are negatively
correlated. The slope of a best-fit line would be -1 if respectively.
y = 1 Columns 1 and 2 show that the estimated y coef-
and if fixed effects are uncorrected with G. ficients are 0.76 and 0.55. The smaller y for the martingale
forecast is consistent with our earlier discussion of the fact
Figure 6 is a plot of the price and gas cost residuals that
identify y. The single solid line is the difference betweenthat the martingale forecast implies higher levels of G and
average G|(v, 'ļ/), the seasonally detrended and de-meaned more monthly volatility compared to the futures forecast.
G, for below- minus above-median MPG vehicles over the Using consumers' gas price forecasts elicited from the
months of the study period. The dotted line is the same as the Michigan Survey of Consumers, Anderson et al. (201 1) show
solid line, except that it uses the martingale forecast insteadthat the average consumer believes that real gasoline prices

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
Table 3. - Base Specification Results Figure 7. - Predicted Mispricing
(1) (2) (3)

Futures Martingale Both
Forecast Forecast
ď -0.76 -0.73
(0.046) (0.046)
Gm -0.55
(0.032)
(Gm-G0 -0.14
(0.037)
Time dummies Yes Yes Yes

Month-by-MPG group controls Yes Yes Yes
Vehicle fixed effects Yes Yes Yes
Observations 931,888 931,888 931,888
This table presents estimates of equation (5) assuming the futures-based and martingale gas price fore-
Thedata
casts, denoted Gf and G"' respectively. All specifications use doublefrom
line shows the price
1999 changes if
through y = 1 in2008
March response
andto a use
$1 increase in the gas price forecas
the grouping estimator with groups at the level of time byfor
MPGvehicles that would
group, be driven
with two 12,000
MPGmiles per year
groups. over a remaining lifetime of seven years, relati
Observations
are weighted by number of transactions. Robust standard to a 25 MPGclustered
errors, vehicle. The single line shows the
by vehicle, arepredicted relative price changes for the base specificati
in parentheses.
y = 0.76. The equations for the double line and the single line are, respectively, (Price Change if y = 1
(1/25 - 1//) X $1 X 12,000 X and (Predicted Price Change) = (1 - 0.76) x (1/25 - 1//) x
$1 x 12,000 x V.
in five years will be the same as they are at the time of the sur-
vey. This suggests that the martingale-based y will be very
close to the MSC-based estimate. This is indeed the case:
that amount, or $1,690. The dollar values of the
mispricing
when we use the MSC-based gas price forecast provided by are certainly not trivial.
Anderson et al., y = 0.51 (SE = 0.029), not statistically
distinguishable from the martingale-based y . C. Standard Errors
One way to test whether it is more realistic to assume the
Throughout the paper, we report robust standard errors,
futures-based or martingale gas price forecast is to test which
fits the data better in a horse race. Because Gm and clustered
Gf are at the level of model j by age a. These standard
highly correlated, we do not include both in a regression.
errors are unbiased in the presence of serial correlation over
Instead, we estimate equation (5) using Gf and add [Gm time
- Gf in
] the price of, for example, a three-year-old Honda Civic
as an explanatory variable. Column 3 presents the DX sedan. Clustering at the level of nameplate by age gives
results
a standard error of 0.054. This standard error is unbiased
of this regression. Neither the futures nor the martingale
in the presence of serial correlation over time in the price
assumption appears to fully capture consumers' forecasts:
conditional on Gf , [ Gm - Gf] is still associated with
of,vehi-
for example, all three-year-old models of Honda Civic.
cle prices. (Mechanically, this also means that conditional
Clustering at the level of model j gives a standard error of
0.086.
on Gm, [Gf - Gm ] is still associated with vehicle prices.) Of Analogously, this is unbiased in the presence of serial
course, these results are conditional on the model beingcorrelation
cor- in the price of all Honda Civic DX sedans of any
rectly specified. Because there appears to be some evidence
age.
for both forecasts under this model, we present both sets Inof
addition, all standard errors can be adjusted to account
results in most of the robustness checks that follow. for the fact that Gisa generated regressor estimated from first-
One way of putting y in context is to compare the observed step regressions that predict each observation's vehicle-miles
vehicle price adjustments to the amounts predicted by theory. traveled and survival probability. Murphy and Topel (1985)
Consider a hypothetical set of used vehicles that have differ- show that the true covariance matrix is additively separable
ent fuel economy ratings but would all be driven 12,000 milesin the usual covariance matrix and an additional matrix that
per year over a remaining lifetime of seven years. The dou- accounts for the uncertainty in the first-step parameter esti-
ble line in figure 7 reflects the change in vehicle prices that mates. We estimate this additional variance by bootstrapping
would be expected in response to a $1.00 increase in the gas draws from the estimated distribution of first-step parameters
price forecast, relative to the price of a 25 MPG vehicle. The and estimating y for each draw. Using the base specification
single "predicted price change" line illustrates the change in with Gf as an example, the clustered, robust IV standard error
relative prices predicted by y = 0.76 from the futures-based is 0.046. The standard deviation of the bootstrapped set of y
gas forecast. estimates is 0.023. Adding together these two variances and
For example, y = 0.76 predicts that the price of a 20 MPG taking the square root gives an adjusted standard error of
vehicle decreases by $506 relative to a 25 MPG vehicle. In0.052.
comparison, y = 1 predicts that the relative price should
decrease by $670, which is $164 more. As another example,D. Alternative Specifications
y = 1 predicts that the price of a 30 MPG vehicle should
increase by $2,230 relative to a 15 MPG vehicle. By contrast, Level of grouping. Correcting for measurement error
y = 0.76 suggests that the relative price adjusts by 76% ofappears to be extremely important. Rows 1 through 3 of

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
Table 4. - Alternative Assumptions

change the result mitigates our earlier concern that market-
Futures Martingale level trends over 1999-2003 could bias the estimator. In row
Row Gas Price Forecast: Specification y SE (y) y SE (y) 12, including the gas price spike of 2008 increases y more
under Gf than under Gm. This is consistent with the fact that
Base
0 2 MPG quantiles 0.76 0.046 0.55 0.032 futures markets did not expect the high spot prices in 2008
Grouping
to be sustained, which mechanically decreases G and thus
1 5 MPG quantiles 0.64 0.037 0.49 0.026 increases y .
2 10 MPG quantiles 0.64 0.034 0.49 0.024 There are two explanations for the fact that including the
3 20 MPG quantiles 0.58 0.034 0.47 0.024
4 OLS (no grouping) 0.60 0.028 0.46 0.020 latter part of 2008 moves y statistically closer to 1 under
the futures-based gas price forecast. First, changes in vehicle
Time periods
11 2004-March 2008 0.80 0.037 0.61 0.027 markets unrelated to changes in gasoline price forecasts could
12 1999-end 2008 0.85 0.035 0.57 0.025 bias the estimated y. Second, the true y may have increased
Discount rate over this period as gas prices fluctuated substantially. This
21 r = 0% 0.59 0.046 0.42 0.024
would be consistent with models of endogenous or "rational"
22 r = 3% 0.67 0.035 0.48 0.028
23 r = 10% 0.87 0.040 0.65 0.037 inattention in which consumers pay more attention to more
24 r= 11% 0.90 0.053 0.67 0.039 important attributes (Gabaix, 2012; Sallee, 2011). It would
25 r = 15% 1.01 0.060 0.77 0.045
also be consistent with models in which large changes cause
Characteristics
consumers to update beliefs between coarse categories, for
31 Sample with characteristics 0.76 0.049 0.56 0.034
32 Control for HP 0.76 0.049 0.57 0.034 example, from gas costs being "inconsequential" to gas costs
33 Control for HP, weight 0.76 0.049 0.57 0.035 being "high" (Mullainathan, 2002).
34 Control for HP, weight, wheelbase 0.76 0.049 0.56 0.034
35 Control for all characteristics 0.75 0.048 0.55 0.034
36 Control for HP, weight 0.76 0.049 0.57 0.035 Alternative discount rates. Rows 2 1 to 25 of table 4 show
37 Control for weight, wheelbase 0.76 0.049 0.56 0.034
38 Control for all characteristics 0.75 0.048 0.55 0.034 the sensitivity of y to the assumed discount rate. At a dis-
count rate just higher than 11%, we fail to reject y = 1
Preferences
42 Exclude hybrid vehicles 0.76 0.046 0.55 0.032 with 90% confidence using the Murphy-Topel standard errors
43 Exclude "green" vehicles 0.73 0.048 0.54 0.033 and the futures forecast. The implied discount rate, the dis-
44 Exclude SUVs 0.71 0.065 0.52 0.046
45 Exclude all cars 0.87 0.087 0.68 0.061
count rate that rationalizes the data by giving y = 1, is just
46 Exclude minivans 0.70 0.046 0.51 0.032 under 15% for the futures forecast and just over 24% for the
Age-by-time controls martingale.
50 One year per group 0.79 0.041 0.57 0.029
51 Two years per group 0.79 0.043 0.57 0.030
52 Four years per group 0.78 0.044 0.57 0.031 Changes in characteristics. The identifying assumption
that E[G%'(x, v, 'ļ/)] = 0 would be violated if a model's
This table presents estimates of equation (5) under alternative assumptions. Unless otherwise stated, all
specifications include 1999 through March 2008. For these specifications, N = 931,888, and there are
characteristics change in ways that are correlated with G.
37,677 ja fixed-effect groups. N = 433,815 for row ' ',N = 1,010,667 for row 12, N = 614,217 for rows
31-35, N = 931,618 for row 41, N = 886,300 for row 42, N = 826,180 for row 43, N = 319,361 for
For
row 44, and N = 874,497 for row 45. All specifications include a full set of month-by-year time dummies
example, since gasoline price forecasts rise on aver-
age during the study period, if characteristics improve more
and MPG group by month-of-year controls. Observations are weighted by number of transactions. Unless
otherwise stated, all specifications use the grouping estimator with groups at the level of time by MPG
group, with two MPG groups. Robust standard errors, clustered by vehicle, are in parentheses. over model years for high-MPG models, we would misat-
tribute these vehicles' increased desirability to changes in
gasoline price forecasts. Our policy of redefining a new gen-
table 4 present estimates of y using increasingly disaggre-
eration of a model as a different j addresses the bulk of these
gated grouping estimators, while row 4 is the ungrouped OLS
concerns. Even within a generation, however, there can be
estimator. As we disaggregate, the estimated values of y drop
some variation in observable characteristics (Knittel, 2011,
toward the OLS estimate. This suggests that measurement
figure 1).
error significantly biases the OLS estimates and even some
An additional suggestive test is to add controls for observ-
of the disaggregated grouping estimators. This pattern can
able characteristics to the estimation. Rows 31 to 35 rees-
arise if the variance of the measurement error is large or if it
timate the base specification with progressively more con-
is correlated across observations.
trols for observable characteristics: horsepower, curb weight,
wheelbase, antilock brakes, stability control, and traction
Alternative time periods. Vehicle markets changed over control. The estimated y remains very similar. Although
the study period. How sensitive is y to the choice of time whether controlling for observables affects a parameter
estimate does not directly tell us whether controlling for
periods? Rows 1 1 and 12 in table 4 repeat the base specifica-
tion for 2004 through March 2008 and for 1999 through the unobservables would affect the estimate, it is plausible that
changes in observable and unobservable characteristics are
end of 2008, respectively. In both cases, y* is closer to 1 than
the 1999-March 2008 base specification, although bothcorrelated,
are by logic similar to that of Altonji, Elder, and Taber
(2005). Therefore, this suggests that changes in unobservable
still statistically less than 1 for both Gf and Gm. In row 11,
characteristics might not bias our estimates.
the fact that excluding the early period does not significantly

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
(1) (2) (3) (4) (5)

Transaction- NHTS Survival JDPA Retail Equally Transactions
Ages Weighted y Probabilities y Prices y Weighted y per Observation
All 0.76 0.84 0.60 0.62 37.8
(0.046) (0.050) (0.041) (0.027)
I-3 0.93 1.03 0.66 0.97 89.2
(0.074) (0.082) (0.057) (0.072)
4-6 0.64 0.70 0.52 0.56 43.9
(0.054) (0.060) (0.055) (0.049)
7-10 0.45 0.50 0.37 0.48 21.3
(0.032) (0.036) (0.040) (0.030)
II-15 0.26 0.28 0.18 0.24 7.7
(0.017) (0.019) (0.032) (0.015)
This table presents estimates of equation (5) for different age groups using the
dummies and MPG group by month-of-year controls, and use the grouping est
except in column 4. Robust standard errors, clustered by vehicle, are in parenthe
Changes in preferences. Aside

E. Observation Weights and from
Heterogeneity by Vehicle Age rep
changes in vehicle characteristics, changes i
A natural question to ask is whether y might be hetero-
represent changes in preferences for a vehicle
geneous across consumer groups. For example, the types of
characteristics. Since gasoline price forecasts ri
consumers who buy newer or older vehicles might be more or
during the study period, differential trends in
less informed about or attentive to fuel economy. In fact, there
low- versus high-MPG vehicles would violate th
is significant heterogeneity inÝby vehicle age. Column 1 of
that E[G% I (x, v, 'ļ/)] = 0, which would bias y.
table 5 displays estimates of equation (5) for specific vehicle
One potential concern is that consumers bec
age groups. The y is 0.93 for vehicles aged 1 to 3, but it drops
ingly "green," or environmentally oriented,
to 0.26 for vehicles aged 11 to 15. All estimates in table 5
period, resulting in increased preference f
use the futures-based gas price forecast, but the martingale
economy vehicles independent of the financi
forecast similarly shows that y declines with vehicle age.
would bias y" upward, as it would increase
The differences in y depend mechanically on the differ-
high-MPG vehicles over the study period as
ences in G. Returning to figure 4, we can see how G varies
forecasts rose. To test this, we exclude hybrids
with age. In June 2004, the transaction- weighted average val-
most green vehicles ranked by Yahoo (2009
ues of G for above-median and below-median MPG vehicles
42 of table 4 show that this does not affect the results.
aged 11 to 15 were $2,620 and $4,150, respectively. These
An opposite concern is that preferences for particular
are about 37% of the values for vehicles aged 1 to 5 years.
classes of large vehicles, for example SUVs or pickups,
These figures depend on underlying data for survival prob-
strengthened over the study period. This would bias y down-
abilities and VMT. According to the registration data, the
ward, as it would increase the prices of low-MPG vehicles
average vehicle between 11 and 15 years old will survive
over time, attenuating the decrease in relative prices that
for 6.1 more years. According to the NHTS odometer read-
the model would expect as gasoline price forecasts rose. To
ings, the average vehicle between 1 1 and 20 years old will be
test this, rows 43 through 45 exclude SUVs, minivans, and
driven 8,980 miles per year, conditional on survival.
all cars, respectively, from the estimation. While the point
Several factors could explain the differences across ages.
estimates change somewhat, the differences are not statisti-
First, we could systematically overstate G for the nationwide
cally significant, except when excluding cars and using the
population of older vehicles by overestimating either VMT
martingale gas price forecast.
or survival probabilities. This appears to be unlikely. Because
Age-by-time controls. Rows 51 to 53 of table 4 include vehicle registrations must be renewed every year at some cost,
age-by-time controls to account for potential changes in it is unlikely that many vehicles would have their registrations
renewed if they were not in use. Thus, it seems most accurate
depreciation patterns over time. These controls address bias
that would occur if depreciation patterns changed as gasto estimate survival probabilities from vehicle registration
prices rose over the study period and the age composition data, and it is unlikely that this would overstate survival
of above- versus below-median MPG groups changed. For probabilities. However, we can also estimate survival prob-
example, older vehicles could have become systematically abilities using the cross-sectional age distribution observed
more valuable in later years, and there could be increas-in the NHTS. There are fewer older vehicles observed in the
ingly more old below-median or above-median MPG vehicles NHTS than in the registration data, suggesting that the NHTS
on the road. The table shows that identifying y only off of might undersample households that own the oldest vehicles,
within-age variation in G does not significantly change the even after weighting for national representativeness on other
estimates. observables. As a result, using the NHTS to calculate survival

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
transactions.
probabilities should make G smaller Column
for new5 of table 5 shows that most
vehicles, vehicles
thereby
increasing y. Depending on the shape
that pass throughof the
auctions survival
are less func-
than five years old, mean-
ing that the jat-
tions, this could change the estimated y level
for observations for newer vehicles
older vehicles in have
more underlying
either direction. Column 2 of table 5 presentstransactions and thus are weighted
estimates of y more
heavily. The difference
using the NHTS survival probabilities. All of in y coefficients
the point across esti-
ages means
that the
mates increase, but none of them estimated ysignificantly,
changes for all ages should be smaller
and when
the pattern of decreasing y with vehicle
weighting age
observations is instead
equally unchanged.
of by number of trans-
Second, even if our G is appropriate for
actions. The first row the 4U.S.
of column confirms pop-
this, showing
ulation of older vehicles, thethat
population
y = 0.62 for equalofweights.
vehicles
Comparing going
columns 1 and
through auctions might have4 shows
smaller G.
that when the This would
regressions happen
are restricted to particular
if auctioned vehicles had loweragefuture
groups, the VMT
y coefficient
or does
survival
not depend much
proba-
on how
observations
bilities, perhaps because they are are weighted.or being sold for
low quality
scrap. The data also do not support this. The average odometer
readings for 1 1- to 15-year-old vehicles differ by only a few
VI. Sticky Information
percent between Manheim-auctioned vehicles and vehicles in
the NHTS, so auctioned vehicles have
Mankiw not
and Reis been
(2002), Sims more heav-
(2003), Woodford (2003),
and others propose
ily used. Furthermore, in constructing models of "sticky
vehicle prices information,"
pjat ,inwe
which
exclude transactions in which the auctioneers rated the auto
macroeconomic news diffuses slowly through the popula-
tion of consumers and firms. This causes prices, savings,
as low quality or scrap quality. As a robustness check, col-
investment, and other decisions to respond to information
umn 3 presents these same specifications using JDPA retail
prices, which are for autos that are in good enough conditionwith a delay. Sticky information could explain the visual evi-
for retail sale. The significant decrease in y for older vehicles
dence in figure 6 that vehicle market prices do not respond to
remains. high-frequency fluctuations in retail gasoline prices.
Third, the identifying assumption that E [Gs|(x, v, 'ļ;)] Interestingly,
= neither the two surveys that have elicited
0 might be increasingly violated for older vehicles. As consumers'
gas beliefs about gasoline prices was structured to
provide evidence on sticky information. The Vehicle Own-
prices rise, increased scrappage of older gas guzzlers would
raise the prices of these vehicles both because demandership
is and Alternatives Survey analyzed by Allcott (2011,
downward sloping and because selective scrappage of the 201 3) asks consumers what current gasoline prices are in their
lowest-quality gas guzzlers would raise the average quality
area and shows that responses are very close to state-level
averages. However, the survey was carried out in October
of remaining ones. However, while there is a strong negative
correlation between G and market share for vehicles older 2010, when gas prices had not fluctuated by more than $0. 10
than 15 years (which are excluded from our regressions), per the gallon for eight months. If the survey were repeated at a
correlation between G and s conditional on (t, v, '|/) for 1time
1- of more volatility, the sticky information model predicts
to 15-year-old vehicles is not statistically significant, andthat
the beliefs would lag current prices.
point estimate is actually positive. The Michigan Survey of Consumers (Anderson et al.,
Fourth, buyers of older vehicles may have higher intertem- 2011) is a repeated cross-section that covers multiple peri-
ods of gas price volatility. However, it does not ask con-
poral opportunity costs of capital, that is, higher discount
rates r, compared to buyers of younger vehicles. If this were
sumers their beliefs about current gas prices. Instead, it asks
the case, our use of the average discount rate still gives
only whether consumers believe that prices will increase or
an unbiased estimate of the market average y. However, decrease
it relative to "current prices." We do not know what
would cause us to overstate y for younger vehicles and over-
consumers believe are "current prices" in months when retail
state y for older vehicles, and it would give the pattern of
gasoline prices have changed significantly.
decreasing y with age observed in table 5. The SCF dataOne prediction of a sticky information model is that y
do not support this hypothesis: auto loan interest ratesshould
are be attenuated when using the difference estimator
not statistically significantly different by vehicle age, andinstead
the of the de-meaning estimator. This is because the for-
standard errors are tight enough to rule out that the aver- mer is identified only off the variation at monthly frequency,
while the latter is identified off of de-meaned levels over
age interest rate for 11- to 15-year-old vehicles is greater
the
than 0.5 percentage points above the average rate for 1- to 5- entire period that the vehicle is observed. Table 6 tests
year-old vehicles. However, Adams, Einav, and Levin (2009)this by estimating equation (5) using the difference estima-
analyze subprime auto loans with interest rates of 25% tortoinstead of the fixed-effects (demeaning) estimator. In this
30%, and consumers buying older vehicles may have very table, unlike tables 4 and 5, we report the coefficient on the
G variable, which is - y, instead of y itself. Columns 1 to
high intertemporal opportunity costs of capital because they
cannot get any loan. 3 use the futures-based forecast Gf , while columns 4 to 6
One implication of the differences in y across agesuse is the martingale forecast Gm. Columns 1 and 4 show that
that regression weighting matters. In our base specifications,
the contemporaneous differences in G are negatively associ-
ated with contemporaneous differences in vehicle prices, as
we weight jat-leve 1 observations by the number of observed

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
(1) (2) (3) (4) (5) (6)

Futures Martingale
Forecast Forecast
AG,-0.33 -0.30 -0.27 -0.16 -0.16 -0.20

(0.036) (0.041) (0.044) (0.009) (0.011) (0.014)
AG/-1 -0.43 -0.49 -0.14 -0.19
(0.029) (0.033) (0.009) (0.012)
AG,_2 -0.40 -0.34 -0.13 -0.13
(0.030) (0.032) (0.012) (0.015)
AG,_3 -0.26 -0.13
(0.028) (0.013)
AG,_4 -0.08 -0.07
(0.037) (0.016)
AG,_5 0.03 -0.09
(0.035) (0.013)
AG,_6 -0.02 -0.03
(0.040) (0.014)
Number of observations 835,837 719,936 573,961 835,837 719,936 573,961

This table presents estimates of equation (5) with the difference estimator instead of the fixed-effects (de-meaning) estim
G jat - Gjaj-x , which corresponds to -y. All specifications use data from 1999 through March 2008, include a full set of mont
estimator with groups at the level of time by MPG group, with two MPG groups. Observations are weighted by the number of
by vehicle, are in parentheses.
time units used in the rest of the paper. For e

(1) (2) "2-month units" row presents estimates with
Futures Martingale gasoline costs aggregated to the bimonthly level. T
Forecast y Forecast y Observations
effects estimates of y in the top panel do not c
Fixed effects
this increased aggregation, as the demeaning estim
1 -month units (base) 0.76 0.55 931,888
(0.046) (0.032) tifies y off lower-frequency variation. However, th
2-month units 0.76 0.57 503,146 from the difference estimator move toward the fixed-effects
(0.045) (0.032) estimates as we aggregate over more months. This provides
3-month units 0.75 0.57 342,551
(0.045) (0.034) additional evidence that markets do not immediately adjust
4-month units 0.74 0.59 263,573 to high-frequency gasoline price variation.
(0.046) (0.034) Klenow and Willis (2007) use the data underlying the Con-
6-month units 0.72 0.58 182,300
(0.046) (0.038) sumer Price Index to document that U.S. retailers set prices
Differences based on information that is up to one year old. Our results
1 -month units (base) 0.33 0.16 835,837 complement theirs by providing additional empirical support
(0.036) (0.009)
2-month units 0.49 0.21 446,100 for the sticky information model in this particular market.
(0.035) (0.010) Table 6 suggests that changes in gasoline price forecasts take
3-month units 0.65 0.24 294,862
four to six months to be fully incorporated into vehicle market
(0.038) (0.015)
4-month units 0.82 0.39 219,650 prices. This could be either because consumers are not aware
(0.038) (0.017) that gasoline prices have changed or because they somehow
6-month units 0.86 0.57 141,295 do not incorporate this knowledge into their willingness to
(0.036) (0.026)
pay when shopping for vehicles.
This table presents estimates of equation (5) with time units at different levels of aggregation. All
These results also highlight the distinction between mod-
specifications use data from 1999 through March 2008, include a full set of time dummies and MPG group
els of rational inattention in macroeconomics and finance,

by season controls, and use the grouping estimator with groups at the level of time by MPG group, with
two MPG groups. Observations weighted by number of observed transactions. Robust standard errors,
clustered by vehicle, are in parentheses.
in which agents do not immediately update beliefs about
time- varying decision variables, and consumer choice mod-
els of static inattention, such as Gabaix and Laibson (2006)
would be expected. However, the estimates or Chetty et al. (2009). Energy-efficiency policies such as
are significantly
CAFE standards
attenuated relative to the fixed-effects estimates, are motivated by the latter form of inat-
especially
for the martingale forecast. Columns 2, tention or some
3, 4, and 6 addsimilarly
lagsstatic misoptimization, not by
inattention
of the change in G. These results show that changes to high-frequency
in vehi- fluctuations in energy prices.
While our strategy
cle prices are strongly associated with changes of exploiting time series variation in
in gasoline
prices as far as three to four months in energy costs allows
the past. The us to sweep our unobserved product
asso-
attributes using fixed
ciation becomes smaller and statistically insignificant by the effects, it is important to account for
fifth or sixth monthly lag. the fact that consumers appear to update gasoline price fore-
Table 7 builds on these results by estimating equation
casts with some (5) quarterly or higher-frequency
delay. Using
variation and
using more aggregated units of time t instead of amonthly
difference estimator could cause an analyst

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
to misattribute inattention to high-frequency fluctuations

Altonji, Joseph, Todd in
Elder, and Christoph
Instrumental Variables Strategies for E
gasoline costs to static undervaluation of fuel economy as a
Catholic Schooling," Journal of Huma
product attribute. 791-821.
Anderson, Soren, Ryan Kellogg, and James Sallee, "What Do Consumers
Believe about Future Gasoline Prices?" NBER working paper 16974
VII. Conclusion (2011).
Angrist, Joshua, "Grouped-Data Estimation and Testing in Simple Labor-
Supply Models," Journal of Econometrics 47 (1991), 243-266.
At least since the energy crises of the 1970s, economists
Ashenfelter, Orley, "Macroeconomic Analyses and Microeconomic Anal-
and policymakers have been interested in how consumers
yses of Labor Supply," Carnegie-Rochester Conference Series on
trade off future costs of energy using durable goodsPublic
withPolicy
their 21:1 (1984), 117-156.
Atkinson, Scott, and Robert Halvorsen, "A New Hedonic Technique for
purchase prices. We propose a different and relatively plau-
Estimating Attribute Demand: An Application to the Demand for
sible approach to estimating demand for energy efficiency,
Automobile Fuel Efficiency," this review 66 (1984), 417-426.
which exploits the interaction of time series variation
Austin, in
David, "Effects of Gasoline Prices on Driving Behavior and Vehicle
Markets," U.S. Congressional Budget Office working paper (2008).
gasoline price forecasts with cross-sectional variation in fuel
Barber, Brad, Terranee Odean, and Lu Zheng, "Out of Sight, Out of Mind:
economy ratings of different vehicles. In our base Thespecifica-
Effects of Expenses on Mutual Fund Flows," Journal of Business
tion assuming the futures-based gas price forecast,78consumers
(2005) 2095-2120.
Baumeister, Christiane, and Lutz Kilian, "Real-Time Forecasts of the Real
appear to value discounted future gasoline costsPrice
only 76%
of Oil," Journal of Business and Economic Statistics 30:2
as much as they value purchase prices. The parameter esti-
(2011), 326-336.
Bento,
mates are insensitive to many factors, including the Antonio, Lawrence Goulder, Mark Jacobsen, and Roger von Haefen,
potential
"Distributional and Efficiency Impacts of Increased U.S. Gasoline
endogeneity of market shares and vehicle-miles traveled
Taxes," American Economic Review 99 (2009), 667-699.
and potential changes over time in consumer preferences or
Beresteanu, Ari, and Shanjun Li, "Gasoline Prices, Government Support
vehicle characteristics. However, assuming a lower and the Demand for Hybrid Vehicles in the U.S.," International
discount
Economic Review 52 (201 1), 161-182.
rate or using the martingale gas price forecast strengthens
Berry, Steven, James Levinsohn, and Ariel Pakes, "Differentiated Products
the qualitative conclusion that y < 1. On the otherDemand hand,
Systems from a Combination of Micro and Macro Data:
other specifications move y closer to 1, including The New Car Market," Journal of Political Economy 112 (2004),
plausi-
68-105.
bly higher discount rates and adding the 2008Busse, recession
Meghan, Christopher Knittel, and Florian Zettelmeyer, "Are Con-
to the sample. Although y* is less than 1 in most sumers ofMyopic?
our Evidence from New and Used Car Purchases,"
American
alternative specifications, it is not hard to imagine Economic Review 103 (2013), 220-256.
plau-
Chetty, Raj, Adam Looney, and Kory Kroft, "Salience and Taxation: Theory
sible combinations of alternative assumptions that could American Economic Review 99 (2009), 1 145-1 177.
and Evidence,"
generate this result. Importantly, the parameter
Coibion, estimates
Olivier, "Testing the Sticky Information Phillips Curve," this
review 92 (2010), 87-101.
are very different for subsamples of younger versus older
Davis, Lucas, and Lutz Kilian, "The Allocative Cost of Price Ceilings in
vehicles.
the U.S. Residential Market for Natural Gas," Journal of Political
Given the uncertainty in these estimates, it is worth return- Economy 119 (20 1 1 a), 2 1 2-24 1 .
ing to our calculation from the introduction. Even if y is closer
Journal of Applied Econometrics 26 (201
to 1 than our base specifications suggest, misoptimization
DellaVigna, Stefano, "Psychology and Econ
distorts vehicle markets more than the failure to internal- Field," Journal of Economic Literature 1
Dreyfus, Mark, and Kip Viscusi, "Rates of Tim
ize climate change externalities. Given the amount of policy
Valuations of Automobile Safety and Fue
attention and academic interest surrounding climate change, Law and Economics 38 (1995), 79-98.
this remarkable result suggests that economists shouldDubin,
be Jeffrey, "Market Barriers to Conserv
Rates Too High?" California Institute of T
devoting significant additional effort to understanding how
working paper 802 (1992).
consumers value energy costs and understanding the policy
Dubin, Jeffrey, and Daniel McFadden, "A
consequences of potential undervaluation. Residential Electric Appliance Holdings an
metrica 52 (1984), 345-362.
Edelstein, Paul, and Lutz Kilian, "How Sensit
REFERENCES tures to Retail Energy Prices?" Journal of
(2009), 766-779.
Abaluck, Jason, and Jonathan Gruber, "Choice Inconsistencies among
Espey, Molly, and Santosh Nair, "Automobile Fuel Economy: What Is It
the Elderly: Evidence from Plan Choice in the MedicareWorth?" Part DContemporary Economic Policy 23 (2005), 317-323.
Program." American Economic Review 101 (201 1), 1 Gabaix,
180-1210.Xavier, "A Sparsity-Based Model of Bounded Rationality," New
Adams, William, Liran Einav, and Jonathan Levin, "Liquidity YorkCon-
University working paper (2012).
Gabaix,
straints and Imperfect Information in Subprime Lending," Xavier, and David Laibson, "Shrouded Attributes, Consumer
American
Economic Review 99 (2009), 49-84. Myopia, and Information Suppression in Competitive Markets,"
Allcott, Hunt, "Consumers' Perceptions and Misperceptions of Quarterly
Energy Journal of Economics 121 (2006), 505-540.
Costs," American Economic Review 101 (201 1), 98-104.Gillingham, Kenneth, Identifying the Elasticity of Driving: Evidence trom
a Gasoline Price Shock in California," Stanford University working
Calibrations from the Automobile Market," paper (2010). American Economic
Journal: Economic Policy 5:3 (2013), 30-66.
Goldberg, Pinelopi, "The Effects of the Corporate Average Fuel Economy
Alquist, Ron, and Lutz Kilian, "What Do We Standards in the US,"
Learn Journal of
from Industrial
the PriceEconomics
of 46 Crude
(1998),
Oil Futures?" Journal of Applied Econometrics 1-33. 25 (2010), 539-573.
Alquist, Ron, Lutz Kilian, and Robert Greene, David, Janet Hopson,
Vigfusson, R. Goeltz and Jia Li, "Analysis
"Forecasting the Pnce of In-Use Fuel
of Oil," in Graham Elliot and Allan Timmerman, Economy Shortfall Basededs., on Voluntarily
Handbook Reported ofMile-per-Gallon
Economic Forecasting , vol. 2 (Dordrecht: Estimates," Transportation Research Record
North-Holland, 2013).1983 (2007), 99-105.

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00
Greene, David, Philip Patterson, Margaret Singh, and Jia Li, "Feebates,
Rebates, and Gas-Guzzler Taxes: A Study of Incentives for P. Aghion, R. Frydman, J. Stiglitz, and M. Wo
Increased
Fuel Economy," Energy Policy 33 (2005), 757-775. edge, Information , and Expectations in Mod
Greenstone, Michael, Elizabeth Kopits, and Maryann Wolverton, In Honor
"Esti- of Edmund Phelps (Princeton, NJ:
Press, 2003).
mating the Social Cost of Carbon for Use in U.S. Federal Rulemak-
ings: A Summary and Interpretation," NBER working Mullainathan,
paper 16913 Sendhil, "Thinking through Categor
(2011). (2002).
Murphy,
Hausman, Jerry, "Individual Discount Rates and the Purchase Kevin M., and Robert Topel, "Estimatio
and Uti-
Step 10
lization of Energy-Using Durables," Bell Journal of Economics Econometric Models," Journal of Bu
(1979), 33-54. Statistics 3 (1985), 370-379.
National Highway Traffic Safety Administration,
Analysis:
the Right and Problems from the Left," Corporate
Journal Average Fuel Economy f
of Economic
Perspectives 15:4 (2001), 57-67. Passenger Cars and Light Trucks," Office of R
Hossain, Tanjim, and John Morgan, ". . . PlusEvaluation, National
Shipping and Center
Handling: for Statistics an
Rev-
Ohta, Makoto,
enue (Non)Equivalence in Field Experiments andAdvances
on eBay," Zvi Griliches,
in "Automobile
Economic Analysis and Policy 6 (2006), theart. Gasoline
3. Price Increases Change Consume
Hughes, Jonathan, Christopher Knittel, and Journal of Business
Daniel Sperling, and Economic
"Evidence of a Statistics
Parry,
Shift in the Short-Run Price Elasticity Ian, Margaret
of Gasoline Walls,
Demand," and Winston Harringt
Energy
Journal 29 (2008), 113-134. nalities and Policies," Journal of Economic
Jacobsen, Mark, "Evaluating U.S. Fuel Economy 374-400.Standards in a Model with
Rust, John,
Producer and Household Heterogeneity," "Stationary Equilibrium
University in a Market
of California at for Durable Assets,"
San Diego working paper (2010). Econometrica 53 (1985), 783-805.
Sallee, James,Paradox
Jaffe, Adam, and Robert Stavins, "The Energy "The Taxation of Fuel
and theEconomy," in Jeffrey Brown ed., Tax
Diffusion
of Conservation Technology," ResourcePolicyand and the Economy
Energy 25 (Cambridge, MA:
Economics 16 National Bureau of
(1994), 91-122. Economic Research, 2010).
Kahn, James, "Gasoline Prices and the Used Automobile Market: A
Rational Expectations Asset Price Approach," Quarterly Journal working paper (201 1).
of Economics 101 (1986), 323-340. Sallee, James, Sarah West, and Wei Fan, "Con
Khan, Hashmat, and Zhenhua Zhu, "Estimates of the Sticky-Information Fuel Economy: A Microdata Approach," Univ
Phillips Curve for the United States," Journal of Money, Credit, and working paper (2009). http://www.ntanet.org/
Banking 38 (2006), 195-207. /proceedings/09/032.pdf.
Sawhill, James, "Are Capital and Operating Costs
Kilian, Lutz, "Explaining Fluctuations in Gasoline Prices: A Joint Model of
the Global Crude Oil Market and the U.S. Retail Gasoline Market," Durable Goods Purchases? Evidence from the US Automobile
Energy Journal 31 (2010), 87-104 Market," University of California at Berkeley working paper (2008).
Kilian, Lutz, and Daniel Murphy, "The Role of Inventories and SpeculativeSims, Christopher, "Implications of Rational Inattention," Journal of
Trading in the Global Market for Crude Oil," University of Michigan Monetary Economics 50 (2003), 665-690.
working paper (201 1).
Kilian, Lutz, and Eric Sims, "The Effects of Real Gasoline Prices on Monetary Economics 3 (2010), 155-181.
Automobile Demand: A Structural Analysis Using Micro Data," Small, Kenneth, and Kurt Van Dender, "Fuel E
University of Michigan working paper (2006). Travel: The Declining Rebound Effect," E
Klenow, Peter, and Jonathan Willis, "Sticky Information and Sticky Prices," 25-51.
Journal of Monetary Economics 54 (2007), 79-99. Stolyarov, Dmitriy, "Turnover of Used Durables in a Stationary Equilib-
Klier, Thomas, and Joshua Linn, "The Price of Gasoline and New Vehi- rium: Are Older Goods Traded More?" Journal of Politiceli Economy
cle Fuel Economy: Evidence from Monthly Sales Data," American 110(2002), 1390-1413.
Economic Journal: Economic Policy 2 (2010), 134-153. U.S. Bureau of Labor Statistics, "Consumer Expenditure Survey" (2007).
Knittel, Christopher, "Automobiles on Steroids: Product Attribute Trade- ftp://ftp.bls.gov/pub/special.requests/ce/standard/2007/quintile.txt.
U.S. Energy Information Administration, "Monthly Energy Review"
offs and Technological Progress in the Automobile Sector," Ameri-
can Economic Review 101 (2011), 3368-3399. (201 1). http://www.eia.doe.gov/totalenergy/data/monthly/.
Langer, Ashley, and Nathan Miller, "Automakers' Short-Run Responses toVerboven, Frank, "The Markets for Gasoline and Diesel Cars in Europe,"
Changing Gasoline Prices," this review 95 (2013), 1 198-121 1. Centre for Economic Policy Research discussion paper 2069
Li, Shanjun, Christopher Timmins, and Roger von Haefen, "Do Gasoline (1999).
Prices Affect Fleet Fuel Economy?" American Economic Journal: Wald, Abraham, "The Fitting of Straight Lines if Both Variables Are
Economic Policy 1 (2009), 113-137. Subject to Error," Annals of Mathematical Statistics 11 (1940),
Manheim, Used Car Market Report , 14th ed. (Atlanta, GA: Manheim 284-300.
Consulting, 2009). Woodford, Michael, Interest and Prices : Foundations of a Theory of
Mankiw, N. Gregory, and Ricardo Reis, "Sticky Information versus Sticky Monetary Policy (Princeton, NJ: Princeton University Press, 2003).
Prices: A Proposal to Replace the New Keynesian Phillips Curve,"Yahoo Autos, "Top 100 Cars by Green Rating" (2009). http://autos.yahoo
Quarterly Journal of Economics 117 (2002), 1295-1328. .com/green_center-top 1 00/.

41.150.34.159 on Sat, 13 May 2023 21:14:26 +00:00

Gasoline Prices, Fuel Economy, and The Energy Paradox

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Gasoline Prices, Fuel Economy, and The Energy Paradox

Uploaded by

Copyright:

Available Formats

GASOLINE PRICES, FUEL ECONOMY, AND THE ENERGY PARADOX

Author(s): Hunt Allcott and Nathan Wozny

Stable URL: https://www.jstor.org/stable/43554957

This content downloaded from