How Consumers Value Fuel Economy:
A Literature Review
United States
Environmental Protection
Agency
-------
How Consumers Value Fuel Economy:
A Literature Review
Assessment and Standards Division
Office of Transportation and Air Quality
U.S. Environmental Protection Agency
Prepared for EPA by
Oak Ridge National Laboratory
EPA Contract No. DE-AC05-OOOR22725
NOTICE
This technical report does not necessarily represent final EPA decisions or
positions. It is intended to present technical analysis of issues using data
that are currently available. The purpose in the release of such reports is to
facilitate the exchange of technical information and to inform the public of
technical developments.
SER&
United States
Environmental Protection
Agency
EPA-420-R-10-008
March 2010
-------
CONTENTS
Page
LIST OF FIGURES iv
LIST OF TABLES v
EXECUTIVE SUMMARY vi
ABSTRACT xv
1. INTRODUCTION 1
2. ALTERNATIVE MODELS OF CONSUMERS' EVALUATION OF FUEL
ECONOMY 4
2.1 SUPPLY SIDE 5
2.2 DEMAND SIDE 7
3. LITERATURE REVIEW: VALUE OF FUEL ECONOMY 9
3.1 DISCRETE CHOICE MODELS 9
3.2 HEDONIC PRICE MODELS 29
3.3 OTHER METHODS 37
4. SUMMARY AND DISCUSSION 49
REFERENCES 59
Appendix: REFERENCE ASSUMPTIONS ABOUT VEHICLE USE, LIFETIME
AND DISCOUNTING 63
in
-------
LIST OF FIGURES
Figure Page
ES-1 Distribution of 25 Distinct Studies by Model Type and Value of Fuel
Economy Relative to the Reference Value ix
ES-2 Distribution of 25 Distinct Studies by Date of Study and Value of Fuel
Economy Relative to Reference Value ix
ES-3 Distribution of 25 Distinct Studies by Form of Fuel Economy Variable and
Value of Fuel Economy Relative to Reference Value x
1 Passenger Car and Light Truck Fuel Economy, Fuel Economy Standards
and the Price of Gasoline, 1978-2009 7
2 Histogram of Estimated Coefficients of Dollars per Mile from Klier and
Linn (2008) 16
3 Estimated Values of a 1 Cent per Mile Decrease in Fuel Costs Based on
Brownstone, Bunch, Golob and Ren (1996) 24
4 Estimated Willingness to Pay for a 1 MPG Increase in Fuel Economy by
Vehicle Class 30
5 Annual Estimates of the hedonic Value of a 1 MPG Increase in Fuel
Economy and the Percent of Estimated Lifetime Fuel Savings Each
Represents 36
6 Trend of Nominal Gasoline Prices over the Period of Sallee, West and Fan's
(2010) Study 47
7 Distribution of 25 Distinct Studies by Model Type and Value of Fuel
Economy Relative to the Reference Value 57
8 Distribution of 25 Distinct Studies by Date of Study and Value of Fuel
Economy Relative to Reference Value 58
9 Distribution of 25 Distinct Studies by Form of Fuel Economy Variable and
Value of Fuel Economy Relative to Reference Value 58
IV
-------
LIST OF TABLES
Table Page
ES-1 Summary of Consumers' Evaluation of Fuel Economy Improvements Based on
27 Recent Studies xi
ES-2 Summary of Key Features of 27 Econometric Studies xiii
1 Estimated Unadjusted Discount Rates 3
2 Estimated Parameters of the Demand and Pricing Equations: Berry, Levinsohn
and Fakes' Specification, 2,217 Observations 11
3 Impacts of Gas Tax Increases Calculated by Bento et al. (2005) 21
4 Estimated Value of Fuel Costs in Vehicle Ownership Choice Model of Feng,
Fullerton and Gan (2005) 22
5 Actual Versus Estimated Value of Fuel Economy 32
6 Fifer and Bunn's (2009) Hedonic Regression Results 34
7 Elasticity of Household Vehicle Choice with Respect to Fuel Cost per Mile 40
8 Manufacturer Prices and Fuel Costs 41
9 Gasoline Price Coefficient Estimates: New Car Price Equation 42
10 Gasoline Price Coefficient Estimates: Used Car Price Equation 43
11 Coefficients on the Expected Present Value of Real Remaining Fuel Costs 48
12 Summary of Consumers' Evaluation of Fuel Economy Improvements Based on
27 Recent Studies 50
13 Summary of Key Features of 27 Econometric Studies 52
Al Annual Miles and Discounted Miles for Light-Duty Vehicles 64
-------
EXECUTIVE SUMMARY
Fuel economy or CO2 emissions standards are a core component of governments' policy
strategies to address global climate change and energy security. Standards have been adopted by
the United States, the European Union, Japan and China, among others. The annual costs and
benefits of these standards easily amount to tens of billions of dollars. How consumers' value
future fuel savings in making car buying decisions has been shown to be a crucial determinant of
the consequences of such standards for economic welfare (Fischer et al., 2007). Yet surprisingly
little is known about this vitally important subject. This review examines empirical evidence
from 28 econometric studies that directly or indirectly estimated the value consumers place on
fuel economy.
The available econometric evidence is inconclusive. The 28 studies reviewed are approximately
equally divided between those that imply that consumers significantly undervalue future fuel
savings in their car buying decisions and those that find that they either approximately fully
value or significantly over-value them. The studies span a wide range of model formulations,
data sources, premises and estimation methods (Table ES-1). Yet there is no clear association
between these distinguishing features and the conclusions reached by researchers. Furthermore,
the econometric studies reviewed are, in general, technically well executed.
Because the estimates of consumers' willingness to pay for fuel economy improvements vary so
greatly it would not be meaningful to try to identify a consensus value from the literature at the
present time. Moreover, the appropriate theory of consumer decision making is also in doubt.
Many of the studies reviewed assume that consumers follow the rational economic model:
consumers make an estimate of future fuel prices, consider how long they will own a vehicle and
how much driving they will do, calculate fuel savings per mile and, applying a discount rate, add
up the present value of fuel savings over the life of the vehicle. Yet the little empirical evidence
that exists about the fuel economy decision processes of actual consumers indicates that this
model is very rarely used. The field of behavioral economics has recently provided a plausible
alternative theory. Because future fuel savings are inherently uncertain, consumers will discount
them heavily relative to certain initial costs. While this theory, referred to as loss aversion, is
firmly established in behavioral economics, its application to decisions about fuel economy has
only recently been proposed. At the present time, there is very substantial uncertainty about how
consumers make decisions about fuel economy, as well as how much they value expected future
fuel savings.
Of the 28 studies reviewed in this report, 25 can be used to derive estimates of consumers'
willingness to pay for fuel economy improvements. Two of the studies by the same author are
essentially the same, and two provide estimates of fuel price elasticities rather than the value of
fuel savings, leaving 25 distinct estimates, explicit or implicit, of the value consumers attach to
future fuel savings. The studies utilize a range of methods, including discrete choice models
based on aggregate sales or disaggregate survey data, hedonic price models, asset price models
and other methods. They are based on a wide range of data sources and cover varying time
periods from 1970 to 2010. Ten of the studies are unpublished manuscripts, 13 are peer-
reviewed journal articles and five are other published reports. In this author's opinion, there is
vi
-------
no important difference in the quality of the analyses between the manuscripts or reports and the
published journal articles.
Key findings of this literature review are:
1. Of the 25 distinctly different estimates, 12 studies indicate that consumers significantly
undervalue future fuel savings relative to a reference values based on U.S. Department of
Transportation data, 8 indicate that consumers' values are approximately equal to the
reference expected value, and 5 indicate that consumers significantly overvalue fuel
savings.
2. With a very few exceptions, there are no obvious flaws in the methods or data used by
these studies. This finding applies equally to the published and unpublished studies.
3. There does not appear to be an obvious explanation for the widely divergent results.
Neither model type, formulation of the variable representing fuel economy, data type,
time period, nor any other readily identifiable factor shows a strong association with
inferences about the values consumers place on fuel economy (Table ES-2).
4. Fifteen of the studies are based on some form of discrete choice model. These are evenly
divided between under, equal and over-valuing fuel economy; studies using hedonic
price, asset price and other models more often indicate undervaluing (Figure ES-1).
5. The studies are evenly divided between those dated 2008 or later (12) and those dated
between 1994 and 2007. Six of the earlier studies and six of the 2008-2010 studies
conclude that consumers significantly undervalue fuel economy. Seven of the earlier
studies and six of the later studies imply that consumers roughly equally value or
significantly over value fuel economy (Figure ES-2).
6. Consumers' expectations about future fuel prices are an important factor in all studies.
Almost all of the studies assume that consumers will use the current price of fuel as a best
estimate of future fuel prices, either due to static expectations or because they perceive
fuel prices will follow a random walk. Five of the studies explore alternative price
expectations models. However, none of the models allows consumers to project trends of
increasing or decreasing prices into the future. Given the importance of price
expectations to the evaluation of future fuel savings, a better understanding of how
consumers form price expectations might provide useful insights.
7. Most of the studies (15) represent fuel economy as the price of fuel divided by miles per
gallon, i.e., fuel cost per mile. These studies are evenly divided between undervaluing
(5), equally valuing (5) and overvaluing (5) (Figure ES-3). Six included fuel economy
without interaction with the price of fuel, either as miles per gallon or gallons per mile.
Of these, five found undervaluing and one equally valuing. Four used a calculated
discounted present value of future fuel costs based on assumptions about vehicle lifetime,
usage and fuel price expectations to represent fuel economy. These studies were evenly
divided between undervaluing and approximately equally valuing. These differences
suggest that there may be some insights to be gained by testing hypotheses about whether
consumers respond differently to fuel price changes as opposed to fuel economy
differences, and whether responses to rising fuel prices differ from responses to falling
fuel prices.
8. Several studies point out the empirical challenges to inferring the value of vehicle
attributes to consumers, in general: (1) vehicle attributes such as weight, size,
vii
-------
performance, luxury and fuel economy are correlated, (2) there are important difficulties
in defining and measuring the many relevant attributes of vehicles, and (3) there are
important differences (heterogeneity) in tastes among consumers. These problems can
lead to errors in variables and omitted variables and, together with correlations among
variables they can result in seriously unstable, biased parameter estimates. More recent
studies, exploiting massive data sets, have attempted to address these problems with
detailed fixed effect coefficients formulations that recognize consumer heterogeneity and
other methods. The persistent differences in results even among these studies suggest
that even these efforts may not have successfully addressed the empirical challenges.
These findings are consistent with earlier literature reviews of implicit discount rates for fuel
economy based on discrete choice models. The consistency with which the literature has yielded
widely varying, inconsistent estimates over a period of more than three decades suggests that
there is either a fundamental empirical problem in estimating the value consumers place on fuel
economy, or that the presumed theory of consumer behavior is incorrect, or both. Recent but
very limited in-depth survey evidence indicates that the rational economic model of consumer
behavior is very likely not an accurate description of consumers' decision making about fuel
economy.
Given the importance of understanding how the market values fuel economy and makes
decisions about it, it might be worthwhile to convene qualified researchers with differing results
to jointly investigate why those results differ so greatly. Such an effort would require sharing of
data sets among researchers, who would then execute a mutually agreed upon set of statistical
analyses, (1) to validate the results produced by others, and (2) to test a specified set of
alternative model formulations using the different data sets. Such a structured test of model
formulations against alternative data sets might lead to important insights about why apparently
carefully and competently done analyses can lead to widely differing results.
It is at least as important to investigate the possibility that it is the rational economic consumer
model that is incorrect. This line of inquiry might best be pursued in two steps. First, conduct
more in-depth interviews, surveys and experiments, such as reported in the seminal paper by
Turrentine and Kurani (2007), to discover what decision criteria and algorithms real consumers
actually employ when considering fuel economy and valuing fuel savings. Second, test these
alternative models using experimental methods and empirical market data.
Vll
-------
Distribution of 25 Distinct Studies by Model Type and Value
of Fuel Economy Relative to Reference
6
5
D Discrete Choice Hedonic D Other
a-
QJ
Under Equal Over
Relative Value of Fuel Economy
Figure ES-1. Distribution of 25 Distinct Studies by Model Type and Value of Fuel Economy
Relative to the Reference Value.
Distribution of 25 Distinct Studies by Date of Study and
Value of Fuel Economy Relative to Reference
0)
(U
Under Equal Over
Relative Value of Fuel Economy
Figure ES-2. Distribution of 25 Distinct Studies by Date of Study and Value of Fuel Economy
Relative to Reference Value.
IX
-------
Distribution of 25 Distinct Studies by Form of Fuel Economy
Variable and Value of Fuel Economy Relative to Reference
6
5
I w Price interaction Uw/o Price Disc. PV
Under Equal Over
Relative Value of Fuel Economy
Figure ES-3. Distribution of 25 Distinct Studies by Form of Fuel Economy Variable and Value
of Fuel Economy Relative to Reference Value.
-------
Table ES-1. Summary of Consumers' Evaluation of Fuel Economy Improvements
Based on 27 Recent Studies
Authors
Alcott & Wozny
(2009)
Gramlich (2008)
Berry, Levinsohn &
Pakes (1995)
Sawhill (2008)
Train & Winston
(2007)
Dagupta, Siddarth and
Silva-Risso (2007)
Bento, Goulder,
Henry, Jacobsen &
von Haefen (2005)
Feng, Fullerton & Gan
(2005)
Klier and Linn (2008a)
Brownstone, Bunch &
Train (2000)
Brownstone, Bunch,
Golob& Ren (1996)
Goldberg (1996, 1998)
Goldberg (1995)
Vance & Mehlin
(2009)
Cambridge
Econometrics (2008)
Eftec (2008)
Fan & Rubin (2009)
Fifer & Bunn (2009)
McManus (2007)
Espey & Nair (2005)
Arguea, Hsiao &
Taylor (1994)
Model Type
Mixed NMNL
NMNL
NMNL Ag
Mixed NMNL
Mixed NMNL
NMNLSur
NMNLSur
NMNL
Logit
Mixed NMNL
Stated &
Revealed
Preference
NMNL Stated &
Revealed
Preference
NMNL
NMNL
NMNL
Mixed logit
NMNL
Hedonic Price
Hedonic Price
Hedonic Price
Hedonic Price
Hedonic Price
Data / Time
Aggregate U.S.,
1999-2008
Aggregate U.S.,
1971-2007
gregate US,
1971-1990
Aggregate U.S.,
1971-1990
Survey, U.S.,
2000
vey, CA,
1999-2000
vey, U.S.,
2001
CES, U.S., 1996-
2000
Aggregate U.S.,
1970-2007
CA Survey, 1993
CA Survey, 1993
U.S. CES, 1984-
1990
U.S. CES, 1983-
1987
Germany,
Aggregate New
Car Sales
UK survey, 2004
to 2009
UK 2001 to 2006
State of Maine,
2007
U.S., 1996-2005
U.S., 2002
U.S., 2001
U.S., 1969 to
1986
W-T-P as % of Discounted PV Implied Annual
Discount Rate
25% > 60%
287% to 823%
<1%
Non-significant
140%, range of
-360% to 1,410%
1.3%
Non-significant
15 .2%
No direct estimate but MPG
insensitive to price of gasoline
0.03% to 1.3%
Very approximately 69%
132% to 147%
-420% to 402%
Consumers "not myopic"
Approximately 1,000%
196% but uncertain of
estimate. Authors contacted
for clarifications.
TBD - authors contacted for
clarifications.
Cars: 25% Cars: 37%
Lt. Trucks: 16% Lt. Trucks: 77%
Cars: 52%, Pickups: 283%
SUVs: 44%, Vans: 240%
90%
109%
3% to 46%
XI
-------
Bhat & Sen (2006)
Sallee, West & Fan
(2010)
Langer & Miller
(2008)
Busse, Knittel &
Zettelmeyer (2009)
Kilian and Sims
(2006)
Li, Timmins & von
Haefen (2009)
Choice model
Price Regression
Price Regression
Price Regression
Price Regression
Vehicle sales by
fuel economy
quantile
San Francisco
Bay Area, 2000
Aggregate U.S.,
Used Cars, 1978-
2009
U.S., 2003 to
2006
U.S., 1999 to
2008
Aggregate U.S.,
Used Cars, 1978-
1984
U.S. Metro
Areas 1997 to
2005
Elasticities of vehicle choice
with respect to fuel costs 2%
to 3% of purchase price
elasticities.
79%, not statistically different
from 100%
Approx. 15% of PV of fuel
cost changes reflected in
vehicle price changes.
Transaction prices adjust by
1.2 years worth of fuel savings
for new cars.
11% to 25%
Short-run price elasticity of
MPG with respect to sales mix
+0.02, long-run +0.2.
xn
-------
Table ES-2. Summary of Key Features of 27 Econometric Studies
Study
Berry, Levinsohn &
Pakesl995
Allcott & Wozny 2009
Klier & Lmn 2008
Gramlich 2008
Sawhill 2008
Publication
Status
Journal
Manuscript N]
Manuscript
Manuscript
Manuscript
Model
NMNL
/TNL
Logit
NMNL
NMNL
Dependent
Variable
Sales shaes
New & used vehicle
prices
New vehicle shares
New vehicle shares Ag
New vehicle shares
Type of
Data
Aggregate U.S.
Aggregate U.S.
Aggregate U.S.,
monthly
gregate U.S.
Aggregate U.S.
Time
Period
1971-1990
1999-2008
1970-2007
1971-2007
1971-1 990 Pg
Fuel
Economy
Measure
Miles/Pg
Disc. PV of Fuel
Cost
Disc. PV of Fuel
Cost
Pg/MPG & MPG
/MPG
Price
Expectations
Random
Walk
RW +
alternatives
Random
Walk
Random
Walk
ARIMA
Transaction
Prices?
No Yes
Yes No
n.a. Yes
No No
No
Heterogeneour
Tastes?
Yes
Simultaneous
Supply &
Demand
Yes
Yes
No
Yes
Yes
Fuel
Economy
Standards
Included?
No
n.a.
No
Yes
No
MPG
Value*
0
+
+
Tram & Winston 2007
Dasgupta, Siddarth &
Silva-Risso 2007
Bento, Goulder,
Jacobsen & von
Haefen 2005 & 2008
Feng, Fullerton & Gan
2005
Brownstone, Bunch &
Tram 2000
Brownstone, Bunch,
Golob& Ren 1996
Goldberg 1995
Goldberg 1996
Goldberg 1998
Journal
Journal
Journal
Manuscript N]
Journal
Journal
Journal
Report
Journal
Mixed Logit
Mixed Logit
Random
Coef. Logit
/TNL
Mixed Logit
NMNL
NMNL
NMNL
NMNL
Indiv. Vehicle Choices
Indiv. Vehicle Choices
Indiv. Vehicle Choices
Indiv. Vehicle Choices Cl
Indiv. Vehicle Choices
Indiv. Vehicle Choices
Indiv. Vehicle Choices C]
Indiv. Vehicle Choices Cl
Indiv. Vehicle Choices Cl
U.S. Household
Survey
So. CA Vehicle
Transactions
Nat. HH. Travel
Survey U.S.
IS
CA survey
CA survey
IS
;s
!S
2000 1
1 999-2000 Pg
2001 Pg
1996-2000
1993
1993
1983-1987
1985-1990
1984-1990
/MPG
/MPG
/MPG
Pg/MPG
Pg/MPG
Pg/MPG
Pg/MPG
Pg/MPG
Pg/MPG
Static
Static
Static
Static
Static
Static
Static
Static
Static
No
Yes
No
No
Yes, via
respondents
Yes, via
respondents
No
No
No
Yes
Yes
Yes
No
Yes No
No No
Yes
Yes
Yes
Yes
No
No
Yes
Yes
Yes
Yes
n.a.
n.a.
n.a.
n.a.
n.a.
n.a.
Yes
Yes
Yes
0
X
+
0
0
0
0
Cambridge
Econometrics 2008
eftec 2008
Vance & Mehlm 2009
Report
Report
Report
Mixed Logit
Mixed Logit
NMNL
Indiv. Vehicle Choices
Indiv. Vehicle Choices
New Vehicle Shares
UK Survey
UK Survey
Sales data,
Germany
2005-2006
2004 & 2007
1 995-2007
£/100km
1/1 00 km &
£/100km
/100km
Static
Static No
Static
No
No
Yes
Yes
No
No
No
No
No
No
No
+
+
Fan & Rubin 20 10
Espey & Nair 2005
McManus 2007
Fifer & Bunn 2009
Arguera, Hsaio &
Taylor 1994
Manuscript
Journal
Journal
Thesis
Journal
2 -stage
hedonic
1 -stage
hedonic
1 -stage
hedonic
1 -stage
hedonic
2-stage
hedonic
New vehicle prices
New vehicle prices
New vehicle prices
New vehicle prices
New vehicle prices
Maine, sales
data
U.S. vehicle
data
U.S. vehicle
data
U.S. vehicle
data
U.S. vehicle
data
20071
2001 1
2002-2005 Pg
1996-2005 1
1969-1986 M
og(MPG)
/MPG
/MPG
/MPG
'G
Static
Static
Static
Random
Walk
n.a.
No
No
Yes
No Yes
No
Yes
No
No
No
Yes
No
No
No
Yes
No
No
No
No
No
0
0
Kilhan & Sims 2006
Manuscript
Asset Price
Used Car Prices
U.S.
1978-1984
PV fuel costs
Random
Walk +
No No
No
No
Xlll
-------
Table ES-2. Summary of Key Features of 27 Econometric Studies (continued)
Study
Bhat & Sen 2005
Langer & Miller 2008
Busse, Kmttel &
Zettelmeyer 2009
Li, Timmins & von
Haefen 2009
Bailee, West & Fan
2010
Publication
Status
Journal
Manuscript A;
Manuscript
Journal
Manuscript Ai
Model
MDCEV
random
utility
set Price
Sales shares
by quartile
Vehicle
demand
set Price
Dependent
Variable
Individual vehicle
choice
New vehicle prices
New & Used vehicle
prices
Sales by Quantile
Used vehicle prices
Type of
Data
Survey: San
Francisco, CA
U.S. regions
weekly
Sample, U.S.
transactions
New & used
sales in 20 U.S.
metro areas
Sample of U.S.
auction
transactions
Time
Period
2000 Pg
2003-2006 Pg
1 999-2008 M
1997-2005
1990-2009
Fuel
Economy
Measure
/MPG
/MPG
>G Quartiles
Pg/MPG & Pg
Disc. PV of fuel
costs
Price
Expectations
Static
Random
Walk +
RW
Random
Walk
Random
Walk +
Transaction
Prices?
Price not
included
No Yes
Yes
No Yes
Yes Yes
Heterogeneour
Tastes?
Yes No
Yes
Simultaneous
Supply &
Demand
No
No
No
No
Fuel
Economy
Standards
Included?
No
No
No
No
No
MPG
Value*
X
0
* Indicates whether study generally implies that consumers undervalue (), over-value (+) or equally value (0) fuel economy, or none of the
above (X).
xiv
-------
ABSTRACT
The extent to which consumers value the expected future fuel savings from fuel economy
improvements to new passenger cars and light trucks is a key determinant of the levels of fuel
economy achieved in unregulated markets and the effects of regulatory standards on consumers'
surplus. This paper reviews 28 recent quantitative analyses of consumers' willingness to pay for
automotive fuel economy. Some of the studies estimate discrete choice models with random or
fixed coefficients, some are based on aggregate market data while others use disaggregate survey
data. Other studies make use of hedonic price analysis or other methods. Their inferences about
willingness to pay span a very broad range with roughly equal numbers finding significant
under-valuing, significant over-valuing or approximately valuing the full present value of
expected fuel savings over the lifetime of a typical vehicle. Although the methodologies or
model formulations of a few of the studies are questionable, there do not appear to be clear
associations among methods or data sources and the resulting inferences. It is suggested that
such conflicting results may be attributable to the statistical problems caused by omitted
variables, errors in variables and correlated variables, the complexity of consumers' vehicle
choice decisions, and the likelihood that the rational economic consumer model does not
adequately describe the decision-making of consumers in the real world. Additional, empirical
behavioral research appears to be needed to resolve the issue.
XV
-------
1. INTRODUCTION
Passenger cars and light trucks account for 44% of the petroleum consumed in the United States
each year and produce 16% of total U.S. greenhouse gas emissions (Davis, Diegel and Boundy,
2009, tables 1.13, 1.14, 11.3 and 11.7). Given the importance of the automotive sector to the
U.S. economy, understanding the costs and benefits of fuel economy and greenhouse gas
standards for light-duty vehicles is of great importance. The economic impacts of vehicle
standards depend on the functioning of the market for fuel economy. In particular, carefully
chosen regulatory standards have been shown to increase or decrease private welfare depending
on how consumers value future fuel savings (Fischer, Harrington and Parry, 2007). The question
has been controversial for decades, with some analysts assuming the market functions efficiently
(e.g., Kleit, 1990; Austin and Dinan, 2005) while other assert that it does not (e.g., Greene,
German and Delucchi, 2009). This paper surveys the recent literature of studies addressing
consumers' willingness to pay for fuel economy improvements, trade-offs between capital costs
and fuel costs, and related economic research. Despite a substantial body of significant new
econometric evidence, the topic remains unresolved. In short, there is evidence supporting both
sides and some in the middle, as well.
More than a quarter century ago, Greene (1983) reviewed the evidence on implicit consumer
discounting of future fuel savings arising from the burgeoning literature in a new area of research
applying discrete choice models to the problem of automobile choice.1 Based on eight studies,
Greene estimated implicit discount rates by assuming that consumers were trading off capital
cost (vehicle price) and future fuel savings (mostly represented by fuel cost per mile). For all but
one model, the implied discount rates ranged from 2% to 73% per year (Table I).2
Greene summed up his findings as follows.
"Eight recent studies are examined and estimates of asset price/operating cost
discount rates are derived for each. A critical comparison of results suggests that
most are implausible. Plausible estimates range from 4% to 40% as a function of
household income." (Greene, 1983, p. 491)
In a comment to Greene's paper, Train (1983) suggested that the assumption that consumers are
trading off capital and operating costs according to the model of rational economic behavior
might be the problem.
"There are many plausible reasons for believing that Greene's assumptions about
consumer behavior are perhaps not correct. On the one hand, consumers might
not be rational in their time allocation of money, in an economist's sense of the
1 Discrete choice models typically estimate an indirect utility function, V(p,e,x), in which p is vehicle price,
e is fuel use or fuel cost per mile. Discount rates can be estimated by noting that -(3V/3e)/(3V/Sp) is the implied
value in dollars of a 1 unit change in e. Given assumptions about vehicle use over time, vehicle lifetime and the
price of fuel, traditional discounting formulas can be used to estimate an implied discount rate.
2 The model excepted, Manski and Sherman (1980), produced implied discount rates ranging from 2840%
per year to -164% per year, depending consumers' on income and education levels.
1
-------
word "rational"... .Given the difficulty my economics students have in calculating
present values of future savings and holding costs for assets, I would not find this
at all surprising." (Train, 1983, p. 498)
In light of subsequent research on how consumers actually consider fuel economy in their car
buying decisions (Turrentine and Kurani, 2007), Train's comment appears to be prophetic.
Recent historically high fuel prices, combined with renewed interest in fuel economy and
greenhouse gas emissions standards for automobiles have engendered a number of new
assessments, many specifically aimed at understanding the effects of fuel prices and fuel
economy on consumers' vehicle choices. This paper reviews those studies, published and
unpublished, with the objectives of determining whether a consensus now exists on the value
consumers place on fuel economy, and of gleaning insights into how consumers use fuel
economy information in their car buying decisions. The intent of the review is to be
comprehensive with respect to studies of the U.S. market for fuel economy that make inferences
about consumers' willingness to pay for fuel economy improvements in both the new and used
vehicle markets. A number of recent studies from the gray (not peer reviewed) literature are
included because, in the authors judgment they are of publishable quality. There are so many of
these studies and their quality is sufficiently high that, in the author's opinion, leaving them out
would substantially reduce the value of the review. Unpublished studies are clearly indicated as
such, however.
This review attempts to compare the inferences of different studies on a consistent basis: a
typical consumer's willingness to pay for a reduction in the present value of fuel costs through
improved fuel economy. This is never more than an approximation; however, since a variety of
assumptions about vehicle use, consumer expectations and discount rates are necessary to make
the calculation. In addition, consumers' behavior and preferences are heterogeneous due to
different vehicle usage rates, and discount rates among other factors. Still, national averages
provide a useful basis for comparisons. The assumptions used here are taken from the National
Highway Traffic Safety Administration's (U.S. DOT/NHTSA, 2006) most recent assessment of
vehicle usage and life expectancy, and are explained in detail in the appendix. In brief, the
NHTSA analysis implies that passenger cars and light trucks in the United States can be
expected to last 14 years, on average. Over 14 years, passenger cars are expected to travel
168,853 miles and light trucks 188,104 miles. Most studies assume that consumers expect fuel
prices to follow a random walk, implying that the current price of fuel is the best predictor of
future prices. That assumption is used here, as well, although there is little behavioral research
to support it. If future fuel price is constant at the current level, then it is convenient to discount
future miles traveled to present value, so that different values for fuel prices and fuel economy
can be applied. Assuming a discount rate of 7% per year and discounting miles over the life of
the vehicle results in 112,600 discounted lifetime miles for passenger cars and 125,891 miles for
light trucks. These standard assumptions are used throughout the report to compare results from
different studies. At the same time, it is recognized that consumers' preferences for fuel
economy are heterogeneous, as several studies conclude.
-------
Table 1. Estimated Unadjusted Discount Rates
1. Lave and Train (1978)
Auto Price (1977$)
Income
(1977$)
10,000
20,000
25,000
30,000
50,000
2500
0.23
0.12
0.10
0.08
0.05
3500
0.21
0.12
0.10
0.08
0.05
5000
0.19
0.11
0.09
0.08
0.05
2. Cardell and Dunbar (1980)
Median = 0.43
Mean = 0.25
3. Beggs and Cardell (1980)
Base Model
Household
Income
10,000
20,000
25,000
30,000
50,000
0.59
0.35
0.31
0.29
0.24
Financial and Size Variables Only
0.73
0.35
0.31
0.28
0.23
4. Boyd and Mellman (1980)
Simple logit 0.06
Hedonic Median = 0.09
5. Manski and Sherman (1980)
a) One-vehicle households
Urban
Low I High I
College 0.10 0.06
No College 0.17 0.18
b) Two-vehicle households
Urban
Low I High I
College 0.64 0.09
No College 28.4 0.26
6. Beggs, Cardell and Hausman( 1981)
Common tastes
Mean = 0.02
Rural
Low I
0.18
0.54
High I
0.19
-0.16
Rural
Low I
-1.64
-0.61
High I
0.19
2.26
Individual tastes 0.30
Income
10,000
20,000
25,000
30,000
50,000
7. Sherman (1982)
One-vehicle households
Two-vehicle households
0.36
0.30
0.29
0.29
0.28
0.13
[dependent on hi (miles annually) here 10,000]
Income
(1978$)
10,000
20,000
30,000
Annual Miles (both
cars)
10,000 20,000
0.02
0.01 0.00
0.01 0.00
25,000
8. Train and Lohrer (1982)
One-vehicle households
0.12 if I < 12,000
Two-vehicle households
0.12 if I < 12,000
0.09 if I > 12,000
0.09 if 12,000 < I < 20,000
0.05 if I > 20,000
Source: Greene (1983, Table 3)
-------
This paper is organized as follows. Section 2 summarizes views of how the market for fuel
economy functions, including both supply and demand. The body of the report is Section 3
which reviews recent empirical estimates of the value of fuel economy based on aggregate and
disaggregate data, discrete choice models and hedonic demand analyses. Section 4 discusses the
implication of those estimates and Section 5 contains concluding observations.
2. ALTERNATIVE MODELS OF CONSUMERS' EVALUATION
OF FUEL ECONOMY
There is no doubt that consumers do care about fuel costs, do value fuel economy, and that their
interest in fuel economy increases when fuel prices increase. Past evidence of this has been
reviewed by Mahadi and Gallagher (2009) who also provide an analysis of the effects of gasoline
price increases since 2000 on consumers' interest in fuel economy. The question is not whether
consumers value fuel economy but how much? The issue is not whether the market for fuel
economy responds to higher fuel prices but whether it responds efficiently. To be more precise,
does the market value fuel economy improvements at society's discounted expected value of
future fuel savings over the lifetime of a new vehicle, or significantly less or more? Fischer,
Harrington and Parry (2007) have demonstrated that the answer to this question has profound
implications for public policy concerning automotive fuel economy and carbon dioxide
emissions. Considering only private costs and benefits, if consumers already fully value
expected lifetime fuel savings, fuel economy standards lead to private welfare losses.3 Studies
that assume efficient markets inevitably arrive at this conclusion (e.g., Austin and Dinan, 2005).
On the other hand, if consumers are myopic and consider only the first three years of fuel
savings, for example, fuel economy standards can increase welfare even based solely on private
costs and benefits.
Given the importance of fuel economy standards to manufacturers, consumers and society, it is
surprising that there has been almost no behavioral research on how consumers consider fuel
economy in their car buying decisions. Larrick and Soil (2008) found that measuring fuel
economy in miles per gallon rather than gallons per mile leads to confusion about the value of
increasing fuel economy. In general, consumers expect fuel savings to increase linearly with
miles per gallon, leading to overvaluing of fuel economy increases for high MPG vehicles
relative to lower MPG vehicles. Turrentine and Kurani (2007) conducted in-depth, semi-
structured interviews with a stratified sample of 57 California households concerning the entire
history of their car buying decisions. Their findings are well worth quoting at some length.
"We found no household that analyzed their fuel costs in a systematic way in
their automobile or gasoline purchases. Almost none of these households track
gasoline costs over time or consider them explicitly in household budgets. These
households may know the cost of their last tank of gasoline and the unit price of
gasoline on that day, but this accurate information is rapidly forgotten and
replaced by typical information. One effect of this lack of knowledge and
3 Fischer, Harrington and Parry (2007) consider external costs as well as private costs in their study. The
above comment pertains only to the private costs calculations in their report.
-------
information is that when consumers buy a vehicle, they do not have the basic
building blocks of knowledge assumed by the model of economically rational
decision-making, and they make large errors estimating gasoline costs and
savings over time.
"Moreover, we find that consumer value for fuel economy is not only about
private cost savings. Fuel economy can be a symbolic value as well, for example
among drivers who view resource conservation or thrift as important values to
communicate. Consumers also assign non-monetary meaning to fuel prices, for
example seeing rising prices as evidence of conspiracy. This research suggests
that consumer responses to fuel economy technology and changes in fuel prices
are more complex than economic assumptions suggest." (Turrentine and Kurani,
2007, p. 1213)
Turrentine and Kurani's final observation, that consumers' fuel economy decision making is
more complex than any single economic model, echoes Train's (1983) observation and is
important to keep in mind as one reviews the empirical studies of vehicle choice and fuel
economy. Very likely, there is not one correct model or one correct set of parameters waiting to
be discovered. Instead, the goal should be to find a useful generalization of a set of complex
behaviors that likely vary from one individual to another, perhaps bearing little resemblance to
the model of perfectly rational economic behavior.
2.1 SUPPLY SIDE
Although the functioning of the supply side for fuel economy is not the central focus of this
review, appropriately representing the supply of fuel economy is important both for
understanding how the market functions, as well as for estimating models of fuel economy
demand. There are three key issues: (1) the competitive structure of the industry, (2) the nature
of short- and long-run supply responses, and (3) the effect of regulatory standards versus market
decisions on fuel economy.
Most econometric studies of the U.S. market assume the industry is an oligopoly (e.g., Bento
et al., 2005; Berry, Levinsohn and Pakes, 1995; Goldberg, 1995; Austin and Dinan, 2005; Busse,
Knittel and Zettelmeyer, 2009). For example, Eftec (2008, p. vi) considered the United
Kingdom (UK) automobile market to be a non-collusive oligopoly, characterized by a relatively
small number of manufacturers all of whom take account of the actions of the others in making
product and pricing decisions. Goldberg modeled the U.S. automobile industry as an oligopoly
with multiproduct firms. Goldberg (1998) assumed that domestic and imported manufacturers
formed two groups, within which there was collusion but between which there was Bertrand
competition. However, she notes that this assumption was based on computational constraints
rather than on observation of actual business practices.
The globalization of the world economy raises doubts about the characterization of the U.S. auto
market as an oligopoly. There are now approximately 25-30 significant vehicle manufacturers in
the global marketplace, not considering partial ownership of one firm by another. Although all
-------
these firms are not now participating in the U.S. market, the barriers to entry are not great.4 With
this many firms participating and relatively low barriers to entry, it may be more appropriate to
characterize the current and future supply side as monopolistically competitive. In a
monopolistically competitive market, firms sell differentiated products and strive to capture rents
thereby, but in the long run products sell at their long-run average costs, and normal rates of
profit prevail. This assumption is consistent with profit rates observed in the industry in recent
years (Rogozhin, Gallaher and McManus, 2009).
Manufacturers' ability to respond to changes in consumer demand for fuel economy or
regulatory standards differs depending on the length of time available (NRC, 2002; Klier and
Linn, 2008). In the short run (less than two years) manufacturers can change the prices of
vehicles to induce shifts in sales and make very minor changes to vehicles themselves (e.g., tires,
lubricants, engine control algorithms). Within 2-7 years, manufacturers will have an opportunity
for major design changes to every vehicle they produce. In general, vehicles are redesigned over
a 4-5 year cycle, with 20% to 25% of any given manufacturer's product line being redesigned
each year in order to evenly distribute capital investments and use of engineering expertise.
Engine and transmission lines are typically amortized over a somewhat longer period, on the
order of 10 years. Studies vary in the degree to which they take account of these cycles of
change.
For studies that attempt to estimate the demand for fuel economy taking into consideration the
simultaneous effects of supply and demand, it is critically important to recognize the impacts of
fuel economy regulations. Although there is some disagreement about the relative effectiveness
of the Corporate Average Fuel Economy Standards versus market responses to the price of
gasoline (e.g., see NRC, 2002; Gerard and Lave, 2003; Greene, 1990), there can be no doubt that
the standards had a major impact on the level of fuel economy observed in the market (Figure 1).
Failure to recognize fuel economy constraints on manufacturers is likely to seriously bias
estimates of consumers' willingness to pay for fuel economy especially in hedonic price
regressions where identification of the demand function is a critical issue. Some studies do
explicitly incorporate fuel economy constraints while others do not.
4 Chief among these are certification to U.S. safety and environmental standards. Even in these respects
there is increasing harmonization in world markets.
-------
Passenger Car and Light Truck Fuel Economy, Fuel Economy
Standards and the Price of Gasoline, 1978-2009
-Combined (Cats)
-Car Standard
-Combined (Truck)
-TruckStandard
-Gasoline Price
-IIIIIIII-
-II
^iii-
-ii-
-iiii-
$3.00
$2.50
$2.00
$1.50
$1.00
$0.50
$0.00
1978 1981 1984 1987 1990 1993 1996 1999 2002 2005 2008
Model Year
Figure 1. Passenger Car and Light Truck Fuel Economy, Fuel Economy Standards and the Price
of Gasoline, 1978-2009.
2.2 DEMAND SIDE
While there is no doubt that consumers respond to fuel prices and value fuel economy to some
degree, there is also little doubt that very few consumers compare discounted present values of
fuel costs when making vehicle choices. Turrentine and Kurani's (2007) research demonstrates
only that the strict, rational economic model of consumer behavior probably does not apply to
consumers' decision making about fuel economy.5 By itself, this does not tell us whether
consumers under- or over-value fuel economy. There is still the possibility that consumers
approximate an efficient market response based on experience and intuition.
Some analysts contend that the market for fuel economy reasonably approximates a perfectly
competitive market from the demand side. For example, Austin and Dinan (2005) assume that
consumers fully value lifetime fuel savings when considering fuel economy in their vehicle
choices.
"We argue, though, that there is no such market failure - that the information on
new-vehicle window stickers, reporting the EPA's city and highway mileage
rating and the vehicle's estimated annual fuel cost, is sufficient to allow
consumers to make informed decisions about fuel economy." (Austin and Dinan,
2005)
5 This finding is definitive for the 57 households in their study. However, it cannot necessarily be
extrapolated to the United States as a whole because the households are from only California and were selected by a
stratified random sampling method. Turrentine and Kurani defined 10 household types of interest and then
randomly selected six households in each group. In addition, gasoline prices were relatively low when the surveys
were conducted.
7
-------
Consumers make decisions that involve present capital cost and future energy costs in many
areas and behavioral researchers have voiced doubts about the efficiency of markets for energy
efficiency for decades (e.g., Stern and Aronson, 1984). Economic analyses of consumer choices
of other types of energy using equipment have produced estimates of implicit discount rates that
are several times average rates of return on capital (Allcott and Wozny, 2009; Hausman, 1979;
Train, 1985; Howarth and Sanstad, 1995; U.S. DOE/EIA, 1996, table 3). A variety of
explanations have been proposed for the apparent undervaluing of energy savings by consumers,
ranging from irrationality and imperfect information to the rational assessment of the value of
waiting in a market in which energy prices are generally rising but highly uncertain (Hassett and
Metcalf, 1993). Energy analysts have identified several types of market failures6 that affect the
market for energy efficiency (e.g., Howarth and Sanstad, 1995; ACEEE, 2007):
Principal agent conflicts
Information asymmetry
Transaction costs
Bounded rationality
External costs and benefits
With the exception of externalities, there is little quantitative evidence of the impact of these
failures on consumers' choices of energy using durable goods (ACEEE, 2007).
Two recent analyses have quantified the potential impacts of uncertainty and risk or loss adverse
behavior on the market for fuel economy. Delucchi (2007) quantified the uncertainty in the key
elements of the fuel economy decision, and by assuming consumers would make uniformly
conservative estimates for each factor, demonstrated that the resulting decisions would appear to
reflect a high discount rate for future fuel savings. For example, consumers with a real discount
rate of 5.5% would appear to have a discount rate of 19% if they conservatively estimated factors
such as the life of the vehicle, lifetime vehicle miles, the price of fuel, and the fuel economy that
would be realized in actual use (in distinction to the official fuel economy rating).
Greene, German and Delucchi (2009) and Greene (2009) applied the theory of context dependent
loss aversion from behavioral economics to the question of valuing fuel economy and concluded
that a typically loss averse consumer would behave as if he or she required a simple 3-year
payback for increased fuel economy. Their method quantified uncertainties about realized in-use
MPG, fuel price, vehicle life expectancy, annual miles of travel, and the cost of increased fuel
economy in order to derive a probability distribution of present value rather than a single
number. Because the net present value is the difference of savings-cost, there is a chance the
consumer will lose money on the deal. By applying loss aversion functions derived by Tversky
and Kahnemann (1992) they showed that a 25% improvement in passenger car fuel economy that
had an expected present value of+$400 would be perceived by loss averse consumers to have a
present value of-$30. A potentially significant aspect of the theory of loss aversion is that it is
context dependent and allows consumers to undervalue fuel economy at the time of the purchase
6 In the author's view, the term "market failure" is unfortunate because it implies an inability to perform a
function rather than the impairment of a function. Market perfection is a high standard indeed and it is doubtful that
any market fully satisfies the criteria for rational economic decision making (Rubenstein, 1998). What matters then
is how far the market for energy efficiency is from an efficient solution. This requires quantification.
-------
decision but fully value fuel savings afterward. While these calculations are based on empirical
data, there has been no testing of the theories either via experiments with consumers or by
statistical inference from market transactions.
3. LITERATURE REVIEW: VALUE OF FUEL ECONOMY
This section reviews 27 recent empirical studies that produced quantitative inferences about the
value of fuel economy to consumers. Helfand and Wolverton (2009) provide a qualitative
analysis of the recent literature, and raise most of the issues considered here. The studies reflect
a wide range of data sources, model formulations and estimation methods. Approximately half
of the research reviewed has been published in refereed journals, however, most of the recent
research which takes advantage of the very large fuel price increases in 2008, is found in
unpublished manuscripts. In this reviewer's opinion, the quality of the unpublished research is
equal to that of the published research.
The studies have been grouped into five categories:
1. Discrete choice random utility models using aggregate data
2. Discrete choice random utility models using disaggregate survey data
3. Discrete choice models using non-U.S. data
4. Hedonic price regressions
5. Asset price models
Of these studies, twelve generally indicate that consumers significantly undervalue fuel
economy, nine conclude that consumers value fuel economy at approximately its discounted
present value of future fuel savings (three are very similar papers by the same author), and four
find that consumers significantly over-value fuel economy. There is no clear association
between a study's findings and its choice of model, data source, or estimation method. It is
suggested that this may be attributable to the complexity of vehicle choices which lead to very
difficult problems for statistical inference, combined with the likelihood that the rational
economic consumer model may be an inappropriate representation of consumers' fuel economy
decision making.
3.1 DISCRETE CHOICE MODELS
3.1.1 Aggregate Data
In a seminal paper, Berry, Levinsohn and Pakes (1995) develop a new method for estimating
both demand functions with random coefficients and cost functions, based on aggregate sales
data for over 2,000 makes and models of vehicles over the period 1971-1990. The model
assumes a utility function that is linear in the logarithm of u, in which y is the income of
consumer I, PJ the price of vehicle j, X is a vector of observed attributes of vehicle j, v a fixed
-------
effect, £ ij a random error, and the term in summation the sum of cross products of unobserved
vehicle and consumer attributes.
uy = a log(yt -pj)+ Xj/J + vj -
^m^
k
(1)
Because prices are endogenously determined in their model, Berry, Levinsohn and Pakes (1995)
estimate the consumer utility functions using instrumental variables.
Fuel economy enters the consumers' utility function not as miles per gallon but as miles per
dollar, MPG divided by price per gallon (more precisely, the variable is the number often mile
increments one could drive for $1 worth of gasoline). This formulation accounts for the very
substantial variations in the price of gasoline over the 1971 to 1990 period. However, it also
implies a linear relationship between increasing fuel economy and utility. This is less than ideal
because fuel economy is the inverse of fuel consumption and fuel expenditures are linearly
related to fuel consumption. Representing fuel economy by miles per dollar with a constant
marginal utility parameter implies that consumers purchasing vehicles with higher levels of fuel
economy place a higher value on the same quantity of fuel savings than purchasers of vehicles
with lower levels of fuel economy. This makes it all the more surprising that the authors'
analysis of elasticities shows decreasing elasticity of vehicle choice with respect to miles/$ with
increasing miles/$.
"The elasticity of demand with respect to MP$ declines almost monotonically
with the car's MP$ rating.
"Hence we conclude that consumers who purchase the high mileage cars care a
great deal about fuel economy while those who purchase cars like the BMW 73 5i
or Lexus LS400 are not concerned with fuel economy." (Berry, Levinsohn and
Pakes, p. 878)
Berry, Levinsohn and Pakes find that the average value of fuel economy (MP$) is not
significantly different from zero, regardless of whether the scale of production of each vehicle is
included in the equation or not. On the other hand, the variance of the value of fuel economy is
statistically significant. The estimated parameters of the authors' preferred model are shown in
Table 2. The average values of the coefficient estimates for MP$ are negative, implying that
more miles per dollar is undesirable, but are not close to being statistically significant. This
result, which will appear again in the random coefficient mixed logit model of Train and
Winston (2007) implies that while some car buyers value fuel economy as a positive attribute
others find it to be a negative factor. On average, the market is approximately indifferent.
10
-------
Table 2. Estimated Parameters of the Demand and Pricing Equations:
Berry, Levinsohn and Fakes' Specification, 2,217 Observations
Demand Side Parameters Variable
Means (# 's) Constant -7.061
HP/ Weight
Air
MP$
Size
Std. Deviations (op's) Constant
HP/ Weight
Air
MP$
Size
Parameter Standard Parameter Standard
Estimate Error Estimate Error
2.883
1.521
-0.122
3.460
3.612
4.628
1.818
1.050
2.056
0.941
2.091
0.891
0.320
0.610
1.485
1.88
1.695
0.272
0.585
-7.304
2.185
0.579
-0.049
2.604
2.009
1.586
1.215
0.670
1.510
0.746
0.896
0.632
0.164
0.285
1.017
1.186
1.149
0.168
0.297
Term on price (a) ln(y-p) 43.501 6.427 23.710 4.079
Cost Side Parameters
In
Air
In
In
Trend
In
Constant
(HP/Weight)
(MPG)
(Size)
(q)
0.952
0.477
0.619
-0.415
-0.046
0.019
0.194
0.056
0.038
0.055
0.081
0.002
0.726
0.313
0.290
0.293
1.499
0.026
-0.387
0.285
0.071
0.052
0.091
0.139
0.004
0.029
Source: Berry, Levinsohn and Pakes (1995, table IV)
Because of the inverse relationship between miles per dollar and fuel expenditures ($/mile) the
value of fuel economy in the Berry, Levinsohn and Fakes' model varies with both the price of
gasoline and the level of MPG. Berry, Levinsohn and Pakes report an average value of 20.86
miles per 1983 $ for their sample. The average price of gasoline from 1971 to 1990 in 1983
dollars was $1.03 per gallon (U.S. DOE/EIA, 2009, table 5.24), thus the average fuel economy of
cars in the sample was 20.2 miles per gallon. Since the average value of MP/$ is negative
(implying fuel economy is a bad) it is more interesting to calculate the value of fuel economy to
consumers whose attribute value is one standard deviation above the mean.
(-0.122+1.050)
V -=21.33
3u .0.0435 I ;MP$
(2)
Since MP$ is in units of tens of miles per dollar, the value of a 1 mile per dollar increase in fuel
economy is $2.13. Using the standard assumptions for discounting future fuel savings described
in the appendix, the present value of a 1 MPG increase from 20 MPG to 21 MPG would be $254.
Thus, even consumers who are interested in fuel economy appear to be undervaluing it by
roughly two orders of magnitude.
11
-------
Uij = utility of vehicle choice] to consumer i
Uij = log[Uij]
PJ = price of vehicle j
Xj = a vector of attributes of vehicle j
Y; = income of consumer i
Z; = a vector of characteristics of consumer i that may interact with the attributes of vehicles
P = a parameter vector of mean consumer utility values for vehicle attributes indexed k = 1, n
G = a vector of variances for consumers' values of vehicle attributes, k = 1, n
Sij = a random utility parameter varying across consumers and vehicle choices
V; = a vector of unobserved components of utility varying across vehicle choices
The stability of parameter estimates for models like the random coefficient logit model of Berry,
Levinsohn and Pakes (1995) has been studied in depth by Knittel and Metaxoglou (2008). The
objective function optimized to estimate the coefficients of such models is generally highly
nonlinear, and thus prone to multiple, local optima. One of the data sets analyzed by Knittel and
Metaxoglou was the automobile choice data base of Berry, Levinsohn and Pakes (1995). They
tested 10 different optimization algorithms, using 50 different starting values for each. Their
findings call for caution not only in estimating such models, but in interpreting any set of
estimated parameters.
"We find that convergence may occur at a number of local extrema, at saddles
and in regions of the objective function where the first-order conditions are not
satisfied. We find own-and cross-price elasticity estimates that differ by a factor
of over 100 depending on the set of candidate parameter estimates." (Knittel and
Metaxoglou, 2008)
Allcott and Wozny (2009) estimated a nested logit discrete choice model using an extensive
data set of both new and used vehicles up to 25 years old in use in the United States between
1999 and 2008. The central purpose of their analysis is to test whether the effect of a $1 change
in the price of a vehicle is the same as the effect of a $1 change in the discounted present value
of fuel costs. Their estimating equation is derived from the consumer's utility function in the
nested logit framework.
Pjat = -aGjat - Pln(sjat} + Yln(Sjat/Snt) + 5ja + 4>t + £jat
(3)
In equation (3) pjat is the price of model j of age a, in year t, G is the discounted present value
cost of future gasoline use, Sjat is the market share of model j, and snt is the market share of the
nest to which vehicle j belongs. The log ratio of market shares enters the equation due to the
nested logit structure which allows differing price effects within nests and across nests. The
constants 5 and (|> are fixed effects representing, respectively, the other attributes of model j of
age a, and shifts in the desirability of the outside good from year to year. Finally, Sjat is a random
error term. The random error term is assumed to be uncorrelated with G, while the model and
age fixed effects (5ja) can be correlated with G. Because of the simultaneity of vehicle prices and
market shares, equation (3) is estimated using instrumental variables and two-stage least squares.
12
-------
Calculating the discounted present value of fuel costs requires a number of assumptions
including expectations about future fuel prices. Allcott and Wozny's approach was to choose
assumptions that were supported by credible data and that were biased towards accepting their
null hypothesis.
"We formulate our assumptions to conservatively bias us against finding that
consumers undervalue gasoline costs. We find, however, undervaluation for any
plausible set of assumptions about gasoline cost expectations, vehicle survival
probabilities, vehicle miles traveled, and other parameters. We conservatively
estimate that U.S. auto consumers are willing to pay only twenty-five cents to
reduce expected discounted gas expenditures by one dollar." (Allcott and Wozny,
2009, p. 5)
For example, the authors assumed that consumers would discount future fuel costs at 15% per
year; if this discount rate is too high, the present value of fuel costs will be understated and the
leverage of fuel costs on market shares will be overestimated. This is consistent with their intent
to bias their analysis in favor of finding that consumers fully value or overvalue fuel economy.
Used vehicle prices came from a data base of auction prices that tracks 5 million transactions
annually. New vehicle prices came from a data base of 2.5 million new vehicle transaction
prices.
Under a wide variety of model formulations and assumptions, Allcott and Wozny found that
consumers substantially undervalue future fuel costs in their choices of new and used vehicles.
The preferred form of their nested logit model, using instrumental variables to account for
potential simultaneous equations bias, produced an estimate of 0.25 for the ratio of the
coefficient of present value fuel costs to that of vehicle price. This implies that consumers count
a present value dollar of fuel costs as only $0.25 relative to a dollar of purchase price. The
nested logit estimated by ordinary least squares (OLS) produced a ratio of 0.15, while the simple
(un-nested) logit generated an estimate of 0.33 and a simpler reduced form model yielded a ratio
of 0.23. Testing the sensitivity of the results to alternative nesting structures, Allcott and Wozny
found a range of estimates from 0.22 to 0.35.
Alternative assumptions about gas price expectations produce a range of estimates for the value
of $1 present value of fuel costs of $0.20 to $0.46. Although Alcott and Wozny's models of
consumers expectations span a range from random walk to the use of NYMEX futures prices,
there is still no consensus as to which of the alternatives, if any, accurately represents the way
consumers form their expectations about future fuel prices. At the same time, this is a critical
element in the analysis. These sensitivity tests suggest that the inference that consumers
underweight fuel costs relative to purchase price in both new and used car purchase decisions is
relatively robust. Increasing the assumed discount rate used to calculate the present value of fuel
costs (in the preferred nested simultaneous model) drove the ratio closer to one, as expected. At
an assumed 60% annual discount rate, the estimated ratio was 0.74, still lower than 1.0 implying
that consumers are using an even higher discount rate.
The undervaluing of fuel economy found by Allcott and Wozny is very large relative to vehicle
prices and the cost of fuel. It suggests a very serious departure from the rational economic
consumer model. The authors conclude that correcting this market imperfection would increase
13
-------
the privately optimal level of fuel economy by 9 miles per gallon by sales mix shifts alone. They
also conclude that the market failure of undervaluing fuel costs is even more significant than
externalities associated with fuel use.
"Comparing these figures shows that the welfare gains from reducing negative
externalities are dwarfed by the welfare gains from reducing the "internality" by
inducing consumers to make the privately-optimal choice. Perhaps the most
important take-away from this analysis then is that behavioral misoptimization
can be a more powerful justification for fuel economy policies than internalizing
environmental externalities." (Allcott and Wozny, 2009, p. 33)
Klier and Linn (2008) use a data base of new vehicle sales covering an unusually long time
period from 1970 to 2007 by detailed model and model year to estimate the effects of fuel price
changes on vehicle sales. They do not directly estimate a measure of willingness to pay, nor can
one be derived from their model without making a number of important assumptions.
Nonetheless, their model does indicate a relative insensitivity of vehicle sales and fleet average
fuel economy to fuel price. Their analysis is entirely focused on the effect of fuel economy on
sales of different makes and models and the impact of those salesmix changes on average fuel
economy. It does not address changes in vehicle technology, engineering or design made by
manufacturers to improve fuel economy. The paper also does not address the impacts of the
Corporate Average Fuel Economy (CAFE) standards on sales or fuel economy.
The form of their model is similar to that used by other studies. The utility, Uyt, of vehicle model
j, within a given model year, to individual I, in month t is assumed to be the sum of services
flows from observed, Xj, and unobserved, coj, vehicle attributes and the cost of obtaining those
services, which is the sum of the vehicle's purchase price, Pj, maintenance costs, nij, and fuel
costs, fjt.
Uijt = a(Pj + mj + fjt) + Xj/1 + a; + eijt
(4)
In equation (4) a and P are coefficients to be estimated and Sijt is a random error. The form
chosen for equation (4) allows a model-specific intercept to be defined as the sum of all the
components that vary only over j, that is everything except the cost of fuel and the error term.
Assuming the error term has the extreme value distribution; the share of model j will be a logit
function of its fuel costs, those of other models and the attractiveness of the outside good, which
is assumed to be an arbitrary used vehicle. The difference between the log market share of
model j and the log market share of the outside good is given by the following.
ln(sjt) ~ ln(.sot) = afjt + 0y
(5)
In equation (5), cj)j is a model specific constant measuring the relative value of model j in
comparison to the outside good (used car) except for the effect of fuel costs on the market share
of j. To eliminate the used car share from equation (5) the authors add dummy variables to time
periods and index 4>jy, by model year, y. Like others, they specify fuel costs as the discounted
14
-------
present value, but assume that the price of gasoline follows a random walk, so that the expected
price is equal to the current price.
ln(sjt) = a
rT+t
_1 P_t_
is=tl + rYMPGjy
MS
eJt = a1^7r + dt + (Pjy + eJt
jy
(6)
Note that the coefficient a' now includes a times discounted lifetime miles. It should therefore
be possible to approximately recover the original a by dividing by an appropriate estimate of
discounted lifetime miles. However, this would give the value of fuel costs in utils per dollar.
One would need a purchase price coefficient to convert to dollars of present value fuel cost per
dollar of purchase price, and this model does not have one. While it would be possible to assume
a purchase price coefficient based on other studies, the uncertainty would be great since price
coefficients vary substantially from study to study. However it may be useful to try and bound
the implications of Klier and Linn's analysis. In the logit model framework they use, the
coefficient of purchase price (b) is the following function of a vehicle's market share (o), its own
price (P), and the own price elasticity of its market share (P).
,= _A_
(7)
In general, with hundreds of models available in any given model year, o for any given model
will be negligible, so that the price slope is chiefly a function of the own price elasticity and own
price. For a typical car, an average price is approximately $20,000 to $25,000, and a reasonable
range for model level price elasticities is -2 to -6. Using an average price of $25,000, this
produces a range of price slope estimates from -0.00008 to -0.00024, with a midpoint value of
-0.00016. Klier and Linn's overall estimates of a' range from -10 to -15 (Klier and Linn, 2008a,
table 2). Applying our estimate of discounted miles for a passenger car of 112,600, produces a
range for a of-0.0000888 to -0.000133, with a midpoint of-0.000111. Recall that this a is the
coefficient of expected lifetime fuel costs. Dividing a by the purchase price slope gives 0.69,
implying that a new car buyer would be willing to pay $0.69 for a reduction in expected lifetime
operating costs of $1. However, given the number of assumptions required to arrive at this
estimate, all that it is meaningful to say is that its general magnitude is plausible. Klier and Linn
also estimate individual coefficients for each model in their sample. The estimates range from -
65 to +35, following a skewed distribution with a mode around -15 (Figure 2). Like other
studies that allow for heterogeneity, Klier and Linn find very substantial heterogeneity in
consumers' apparent valuation of future fuel savings.
15
-------
120
100
80
60
40
20
n _ 0 El -
-65
-55 45 -35 -25 -1S -5 5 15 25 35
Coefficient
Figure 2. Histogram of Estimated Coefficients of Dollars per Mile from Klier and Linn (2008).
Gramlich (2008) proposed a model of U.S. automobile supply and demand in which
manufacturers decided on the level of fuel economy to include in vehicle designs by trading off
increased fuel economy against increased cost and a measure of quality. New car demand is
represented by a nested logit model, estimated using U.S. sales data for 1971-2007. Consumer i
is assumed to choose the vehicle model (j) that maximizes utility. Utility is a function of vehicle
price (PJ), fuel economy (econj), other quality (qualj), which ".. .includes things such as power,
weight, acceleration, electronics, sportiness, interior room, etc. - in short, the collection of other
vehicle attributes that must be traded off with fuel efficiency." (Gramlich, 2008, p. 6). Other
variables common to all vehicles in a given year, such as macroeconomic factors, and vehicle
specific fixed effects are also included.
Gramlich makes the important observation that the correlation of fuel economy with other
vehicle attributes creates estimation biases.
"Previous models, including seminal work, have had difficulty finding parameter
estimates that show consumers care about fuel efficiency. Parameters on
preference for fuel efficiency have been biased towards zero. The reason for this
is that in automobiles, fuel efficiency (MPG) is negatively correlated with other
characteristics that provide utility. Some of these characteristics are observed and
easily controlled for, such as horsepower and weight. Others, however, are not."
(Gramlich, 2008, p. 4)
Surprisingly, the measure of quality selected by Gramlich was precisely fuel economy in miles
per gallon. The idea is that fuel economy will represent negative quality, that is, it will represent
the trade-off between fuel economy and quality. The positive value of fuel economy is
represented by the price of fuel divided by miles per gallon.
"Fuel efficiency (MPG) affects both the fuel economy (econ) and "other quality"
(qual) of a vehicle through the technological tradeoff. ... The measure I use for
16
-------
qual is MPG itself. This may seem unusual but, conditional on the economic
effects of fuel efficiency (dpm), higher MPG is strongly associated with lower
"other quality." (Gramlich, 2008, p. 7)
This formulation raises two questions. The first is whether fuel economy is an adequate
(negative) proxy for quality. The numerous excluded dimensions of quality include not only
factors that are clearly negatively related to fuel economy, such as size, power, and energy using
accessories, but also items seemingly unrelated to fuel economy, such as reliability, fit and
finish, and safety features. Indeed, the two manufacturers with the highest fuel economy given
the mass and power of their passenger cars are Toyota and Honda, manufacturers also having
among the highest reputations for quality (Knittel, 2009). Thus, using MPG as the sole proxy for
quality would seem to create problems of omitted variables and errors in variables, conditions
likely to lead to biased estimates. It is highly unlikely that MPG itself could serve solely as a
negative proxy for other quality attributes and not in any way function as a positive attribute
itself. The formulation also requires that any increase in fuel economy necessitates a reduction
in quality in other areas. That is, does not permit fuel economy to be purchased at a higher price
without degrading quality.
The second issue is that using fuel price divided by MPG does not allow for distinction between
consumers' valuation of fuel economy versus their reaction to changes in gasoline prices, since
the effects are constrained to be equal and opposite effect. Although a presumption of rational
behavior implies that the effects of price and fuel economy should be equal and opposite, this
hypothesis can be tested and, indeed, that should be one of the objectives of an analysis of the
effect of fuel economy on consumers purchase decisions. Including MPG by itself as a measure
of quality also confuses the interpretation of price/MPG.
Another difficulty with Gramlich's model formulation is that he assumes that decisions regarding
the designs of vehicles are taken the year before the current model year. This does not match the
consensus understanding of how vehicle designs are changed over time. According to the NRC
(2002) report on the Corporate Average Fuel Economy standards, manufacturers typically lock
in vehicle designs two years in advance to allow time for tooling, certification and testing. In
addition, only about 20% of a manufacturer's makes and models are substantially redesigned in
any given year, in order to allow more efficient use of engineering resources and to distribute
capital expenditures more evenly over time. Gramlich (2008) does test two alternative
formulations with 3 and 5 year lags in design, but neither reflects the continuous, gradual
redesign process manufacturers follow in reality.
Gramlich (2008) calculated willingness to pay for an increase in MPG from 25 to 30 miles per
gallon. With gasoline at $2/gal., luxury car purchasers were willing to pay $4,098, compact car
owners were willing to pay $7,377 and SUV owners $11,749. If the price of gasoline increased
to $3.50, luxury car owners would pay up to $7,172, compact car owners $12,910 and SUV
buyers, $20,560. These values are high relative to the discounted present value of fuel savings.
Using the standard assumptions of this assessment, an improvement in fuel economy from 25 to
30 MPG is worth $1,428 present value, with gasoline at $2/gal. and $2,500 with gasoline at
$3.50/gal.
17
-------
Gramlich also used his model to make predictions about the impact of the increases in gasoline
prices from 2007 to 2008 on the average fuel economy (identical to fuel efficiency in the quote
below) of new light-duty vehicles and on vehicle sales.
"To check the predictive ability of the model, I compare out-of-sample
predictions to 2008 actual figures. The model matches sales composition changes
well. Actual aggregate sales are down 12% through August from 2007 levels,
compared to a prediction of 11.9%. The model also predicts an increase in sales-
weighted fuel efficiency of 28%, a number consistent with the large 2008
reductions in purchases of SUVs and light trucks." (Gramlich, 2008, p. 4)
Indeed, according to U.S. EPA (2009) data, sales of passenger cars and light trucks declined
from 15.277 million units in model year 2007 to 13.900 million units in model year 2008, a
reduction of 9.0%, quite close to Gramlich's model's prediction. The average fuel economy of
new light-duty vehicles increased from 25.8 to 26.3 MPG (combined, unadjusted test value), a
change of only +1.9%. Using the city MPG measure Gramlich used in his model estimation, the
increase was from 21.8 to 22.1, or +1.4%. Fleet average MPG numbers reported by the NHTSA
(U.S. DOT/NHTSA, 2009) are very similar to the EPA numbers: 27.0 for 2008 versus 26.6 for
2007, an increase of 1.5%. Thus, while Gramlich's model did a reasonable job of predicting the
decline in vehicle sales it was off by more than an order of magnitude in over-predicting the
impact of higher fuel prices.
Sawhill (2008) estimated a model to quantify the trade-off between automobile operating cost
and purchase price using a virtually identical data base (1971-1990 model years) and a very
similar methodology to Berry, Levinsohn and Pakes (1995). The key differences are that
Sawhill's model includes heterogeneous consumer preferences and explicitly represents
expectations about future fuel prices. Vehicle prices are list prices, although they are assumed to
be endogenously determined along with unobserved vehicle attributes and operating costs are
represented by the price of gasoline divided by fuel economy. The authors do not state so
explicitly, but given the source and the data shown in their table 2, the fuel economy data is
almost certainly the EPA adjusted, combined MPG measure. Consumers' expectations about
future fuel prices are based on an autoregressive model in price differences fitted to annual
historical gasoline price data. This implies that consumers form an analogous model via
observation of fuel price changes. The coefficient of the lagged price difference is 0.35 in the
expectations model, indicating that the change from last year is expected to persist, with a
rapidly declining rate. The author states that fuel cost is used as a proxy for vehicle operating
cost, asserting that fuel costs are the largest share of operating costs. However, no other future
costs (e.g., insurance, maintenance and repairs, etc.) are included in the model. According to
data from the publication Motor Vehicle Facts and Figures (Davis, Diegel and Boundy, 2009,
table 10.14) depreciation and financing costs are about 56% of the total costs of owning and
operating a motor vehicle. The issue is whether fuel costs are truly a proxy for all costs paid out
over time, as opposed to vehicle purchase price, or whether the other costs should be considered
one of the many unobserved vehicle attributes. If fuel costs are a proxy for (and therefore
correlated with) other future costs, then one might expect the coefficient of fuel costs to be
biased upwards. If non-fuel future costs are more correctly interpreted as unobserved attributes,
then the coefficient of fuel costs could be unbiased.
18
-------
The key equation of Sawhill's model relates consumer i's utility for automobile j to the price of
vehicle], the expected present value operating cost of j over the car's lifetime, and vectors of
observed (xj) and unobserved (zj) attributes of vehicle j, as well as a random error term, Sy.
zj
(8)
The parameter a represents consumer i's sensitivity to price (or income) while a value of y
different from 1 indicates that consumers value fuel savings differently from purchase price.
E(cij) is a key function representing the consumer's expected present value of future operating
costs. For y to equal 1, E(cy) must be measured accurately. This requires that price expectations
be accurately represented. It also requires that fuel economy and the distribution of annual miles
be correctly estimated. Sawhill uses data from the Department of Energy's 1991 RTECS survey
to estimate the distribution of vehicle miles. Data for all model years rather than just new
vehicles were used but mileage was adjusted by a weighting factor intended to represent the
frequency of new vehicle purchases for each car as a function of its annual mileage. Vehicles
were assumed to be driven different numbers of miles per year by different households but to all
have the same total lifetime miles. This is clearly not the case, but it is also not correct to assume
that all vehicles last for the same number of years. Thus, the expected life of a vehicle is an
inverse function of annual use equal to the average lifetime miles divided by the household's
annual mileage. Based on the adjusted RTECS data an estimated average lifetime mileage of
90,000 was used (p. 15). The expected annual miles driven for a typical new car in Sawhill's
study is 13,612. At this rate the vehicle's expected life would be 6.6 years. These numbers are
far lower than the reference lifetime mileage estimates used by the Department of Transportation
(DOT): 152,137 miles for passenger cars and 179,954 for light trucks (U.S. DOT/NHTSA,
2006), based on expected lifetimes of just over 13 years for passenger cars and 14 years for light
trucks (see appendix). The fuel cost calculation also assumes an annual real discount rate of 5%
for the base model. Finally, consumers are able to adjust their annual driving rates as fuel prices
increase or decrease, although sensitivity testing showed that this factor had a minor influence on
the estimated coefficients.
Sawhill (2008) first estimates a simple model that includes neither heterogeneous consumer
preferences nor price expectations. The OLS and two-stage least squares versions of the simple
model imply that consumers undervalue the discounted present value of future fuel savings. In
contrast, versions of the full model including heterogeneity of consumers' preferences and price
expectations indicate that most consumers overvalue fuel costs (representing operating costs) by
a factor of about 1 .4. A ratio of 1 .0 would imply that consumers equally value a dollar change in
vehicle price and a dollar change in the expected present value of lifetime fuel savings.
Increasing average lifetime mileage to 1 10,000 from 90,000 miles reduced the ratio to 1.3.
Increasing the expected lifetime miles to 152,137 would presumably bring the ratio closer to 1.
The estimated variability of valuation of fuel costs is considerable; a 95% confidence interval for
the distribution of fuel costs in the population is -3.6 to 14.1. This implies that some consumers
prefer higher fuel costs but this could be a consequence of the assumed symmetrical distribution
of consumer preferences.
19
-------
The only vehicle attributes included in the base model formulation were horsepower, a measure
of car size whose average varies from 1.5 to 1.25 over the sample period, and air conditioning.
As a test of robustness to the specification of vehicle attributes, the authors also estimated a
model including a dummy variable for an automatic transmission, the number of cylinders in the
vehicle's engine and the number of doors in the vehicle. Including these variables has little
effect on the coefficient of price, which changes from -3.78 to -3.64. The coefficient of fuel
costs, however, changes from -5.26 to -3.27. The standard deviation of the fuel cost coefficient
measuring its distribution in the population also changes from 4.5 to 2.8. While these are not
dramatic changes they do indicate a nontrivial sensitivity of the estimated fuel cost coefficient to
both the included variables and the specification of the present value of fuel costs.
Because Sawhill's model is focused on fuel costs as a proxy for operating costs, it is not possible
to separate the influence of fuel economy from that of fuel price expectations on consumers' car
buying decisions. It may be that consumers respond differently to fuel price changes and
changes in fuel economy that produce equivalent changes in expected present value operating
costs. Different influences for fuel economy and fuel price have been found in econometric
analyses of vehicle miles of travel, for example (e.g., Small and Van Dender, 2007; Greene,
2010).
3.1.2 Survey Data
Train and Winston (2007) formulated a mixed logit choice model that allows the estimation of
average values, values that vary systematically with consumer attributes, and values that are
randomly distributed in the population. They estimated the model using a random sample of 458
U.S. consumers who acquired a 2000 model year vehicle. The consumer data included the make
and model of the vehicles they purchased, other vehicles they seriously considered, their vehicle
ownership histories and socioeconomic characteristics. They found that the effect of fuel
consumption (gallons per mile) in the average utility equation, was negative as expected but not
statistically significant. The estimated coefficient was reported to be -0.0032 with a standard
error of 0.0023. Fuel consumption was not included in the estimation of utilities that varied
across consumers as a function of consumers' socioeconomic attributes. There was a statistically
significant component that varied randomly across the population, comprised of the coefficient
-0.0102 (std. err. = 0.0020) times a standard normal random variable N(0,l). This implies that
about 62% of the respondents considered lower fuel consumption (higher fuel economy) to be a
positive attribute for a new vehicle while about 48% considered it to be a disadvantage. On
average, consumers considered better fuel economy to be slightly, and not statistically
significantly, advantageous. Taking the value -0.0032 at face value and using the standard
assumptions for fuel cost and vehicle use, a 0.01 gal./mi. change in fuel consumption (equivalent
to 20 MPG to 25 MPG) would be worth only $30 to the average consumer. Using the reference
assumptions about vehicle use, lifetime, a 7% real annual discount rate and gasoline at $2/gallon,
an improvement to 25 MPG should be worth $2,250 to owners of a typical passenger car, and
$2,500 to the owner of a light truck
Dasgupta, Siddarth and Silva-Risso (2007) estimated a mixed logit model of consumers'
choice of vehicle and lease-buy option using data from vehicle transactions in Southern
California for the period September 1999 to October 2000. Only 15 vehicle models were
20
-------
included in the study, all were in the entry-luxury market segment. Although the contract choice
equation included fuel cost per mile as a variable, the structure of the model is such that its
coefficient should be interpreted as the distribution of annual miles traveled. The authors report
the estimated mean of that distribution as 16,380 miles with a standard deviation of 1,914 miles.
Although this prohibits estimating an implied discount rate from the ratio of the net price (net
present value of acquisition costs) and fuel cost coefficient, consumers' implicit discount rates
were inferred from consumers' choices among purchase and leasing options. The structure of
the model implies that the discount rate for lease or loan payments should be the same as the
discounting of future fuel costs. The authors report an estimated discount rate of 15.2%, higher
than the 9% annual market rate prevailing at the time the transactions occurred. The inference
that this is the discount rate for future fuel savings also depends on the accuracy of the average
annual mileage estimate of 16,380.
Bento, Goulder, Henry, Jacobsen and von Haefen (2005) estimated a model of vehicle choice
and use using data from the 2001 National Household Travel Survey, augmented with data on
vehicle attributes and operating costs. They estimated a random coefficient model of
households' vehicle and vehicle travel choices. Although estimated coefficients and mean
values of some variables are available in an extended on-line version of the paper (Bento et al.,
2008; http://www4.ncsu.edu/~rhhaefen/auto051808.pdf) the information is not sufficient to
estimate trade-off rates between fuel economy and rental cost. The model's estimates of the
impacts of higher gasoline prices on vehicle travel and vehicle fuel efficiency shed some light on
how consumers value fuel economy. Although it is not explicitly discussed in the paper, the fuel
economy impacts presumably come about via consumers' choices among a fixed set of vehicles
and do not include manufacturers' decisions about the technological content and design of
vehicles. Their estimates of the impacts of gasoline price increases of $0.10, $0.30 and $0.50 per
gallon over a base price of $1.45 per gallon (6.9%, 20.7%, and 34.5%), in cases with and without
income-based revenue recycling are shown in Table 3. The change in MPG is very small,
ranging from 0.04% (0.01 MPG) for the $0.10 increase without revenue recycling to 0.24% (0.06
MPG) for the $0.50 per gallon increase with revenue recycling. Implied elasticities of MPG with
respect to fuel prices are shown in parentheses.
Table 3. Impacts of Gas Tax Increases Calculated by Bento et al. (2005)
Base
Fuel Price
MPG no recycling
MPG with recycling
$1.45
24.2
24.2
$0.10
+6.9%
+0.04%
(+0.006)
+0.05%
(+0.007)
$0.30
+20.7%
+0.11%
(+0.005)
+0.13%
(+0.006)
$0.50
+34.5%
+0.18%
(+0.005)
+0.24%
(+0.007)
Bento et al., 2005, table 1.
Feng, Fullerton and Gan (2005) estimated a simultaneous model of vehicle use and choice of
ownership using data from the 1996 to 2000 Consumer Expenditure Survey. The detail on
vehicle type was limited to passenger car or SUV, and only households owning 1, 2 or no
vehicles were included. A nested multinomial logit model structure was assumed, however, the
requirement for simultaneous estimation of vehicle use and ownership choice required a two-
21
-------
stage estimation method. There are six fuel cost per mile coefficients in their model: (1) single
car or first car in two-car household, (2) second car in two car household, (3) single SUV or first
SUV in two-SUV household, (4) second SUV in two SUV household, (5) car in car+SUV
household, and (6) SUV in car+SUV household. Four of the six estimated coefficients of vehicle
capital cost and fuel cost per mile were statistically significant and all had appropriate negative
signs. The coefficients, their ratio to the coefficient of capital cost, the implied value of a $0.01
per mile change in fuel cost per mile, and that value as a percent of the full lifetime discounted
present value of a $0.01 per mile change in operating costs are shown in Table 4. For estimation,
Feng et al. (2005) represented capital costs in units of thousands of dollars and they state that
they represented fuel cost per mile in units of dollars per mile (table 6 and p. 6). As seen in
Table 4, the estimated willingness to pay for reduced fuel cost per mile is on the order of 1% of
the discounted present value of lifetime value. If Feng et al. actually converted fuel cost per mile
in dollars per mile to fuel cost per mile in cents per mile but did not report that in their paper,
then the values of fuel costs implied by their model's estimates would be comparable to lifetime
discounted present value.
Table 4. Estimated Value of Fuel Costs in Vehicle Ownership Choice Model
of Feng, Fullerton and Gan (2005)
Variable
Only Car/lst Car
2nd Car
Only SUV/2nd SUV
2nd SUV
Car in Car/SUV HH
SUV in Car/SUV HH
Coefficient
Estimate
-0.433
-0.045
-0.526
-0.013
-0.399
-0.662
Implied Value
$/Unit
-1069
-111
-1299
-32
-985
-1635
Willingness to Pay as a
of Lifetime Discounted
1.0%
0.1%
1.0%
0.03%
0.9%
1.3%
%
PV
Source: Feng, Fullerton and Gan (2005, table 7). The coefficient of capital cost in dollars rather than
1,000s of dollars is -0.000405. Discounted lifetime miles are 112,600 for cars and 125,891 for SUVs.
An additional interesting finding of Feng et al.'s study was that coefficient estimates varied
considerably depending on the estimation method used.
Brownstone, Bunch and Train (2000) estimated a mixed logit choice model for automobiles
using a combination of stated and revealed preference data. Use of stated preference data was
necessary in order to include in the choice set alternative fuel vehicles, such as battery electric
and compressed natural gas vehicles, which were generally unavailable in the market at the time.
The data base was a multi-wave survey that began with a telephone survey of 7,387 California
households carried out in 1993. Households were not only asked about the vehicles they owned,
but about vehicles they intended to purchase, if any. A total of 4,747 households completed a
more detailed mailed questionnaire. Households to whom a questionnaire was mailed were also
contacted by phone and asked about their preferences for alternative fuel vehicles. More than a
year after the above surveys were completed, most of the original households were re-
interviewed. Of these, 874 had purchased a vehicle and these became the revealed preference
data set for the logit model estimation. In the end, there were two stated preference and one
revealed preference observation for each of these households.
22
-------
First, models were estimated using the stated and revealed preference data sets separately. In the
preferred stated preference mixed logit model, fuel economy entered as the denominator of fuel
cost per mile. The coefficient of fuel cost per mile was negative and statistically significant, as
expected. Vehicle price entered as price in thousands of dollars, divided by the natural logarithm
of household income in thousands. The mean of log(income) was 4. The average willingness to
pay for fuel cost reduction is therefore: -(-0.236/((-0.503/1000)74)) = $1,877 per penny per mile
change in fuel cost per mile. The normalized coefficient estimates also provided by Brownstone
et al. in their table 2 indicate a willingness to pay of $1,710 for a 1 cent decrease in fuel cost per
mile. Using the standard assumptions about lifetime miles (112,600 discounted lifetime miles
for passenger cars, 125,891 for light trucks), a one cent per mile change in fuel costs would be
worth $1,126 for passenger cars and $1,259 for light trucks, indicating that in the stated
preference survey consumers tend to overvalue lifetime fuel savings. However, the value of fuel
cost per mile has a large variance, which led Brownstone et al. to make the following
observation.
"Unfortunately, the relatively large error component for operating cost implies
that the model will generate an (implausible) positive price effect for one third of
the respondents." (Brownstone et al., 2000, p. 327)
The willingness to pay in Brownstone et al.'s (2000, tables 3 and 4) revealed preference model is
even greater, $2,240 per 1 cent per mile. The mixed logit model estimated on the combined
stated and revealed preference data indicated a willingness to pay for reduced fuel cost per mile
of $1,660.
Brownstone, Bunch, Golob and Ren (1996) estimated a nested logit choice model of household
ownership and use that included modeling of vehicle acquisition and disposal transactions. The
data came from a stated preference survey of California urban households conducted in June and
July of 1993. Households were asked to choose among hypothetical vehicles that included
alternative fuel vehicles. Choice models were estimated for households owning one and two
vehicles. Both the vehicle price (in dollars) and operating cost (in cents per mile) coefficients
were varied according to income, the presence of children in the household and ownership of a
luxury vehicle. Estimates of the willingness to pay for a reduction in fuel cost per mile for these
groups are shown in Figure 3. There is a wide range of implied willingness to pay, with only
about four out of the ten estimates being close to our reference willingness to pay estimates of
$1,126 per penny per mile for passenger cars and $1259 for light trucks. The greatest
willingness to pay ($4,718) was for households with incomes of more than $30,000 per year, no
luxury vehicle but with children in the household. The lowest value was -$4,526 for households
with incomes exceeding $75,000 per year and no children.
23
-------
$5,500
$4,500
$3,500
$2,500
$1,500
$500
-$500
-$1,500
-$2,500
-$3,500
-$4,500
-$5,500
Estimated Values of a 1 Cent per Mile Decrease in Fuel Costs
Reference Values are $1,126 for Cars, $1259 for Light Trucks
.$lv.964--
...$1,15.!..
-$4,526
er
f f f >
Figure 3. Estimated Values of a 1 Cent per Mile Decrease in Fuel Costs Based on Brownstone,
Bunch, Golob and Ren (1996, tables 3 and 4).
Goldberg (1995) estimated a nested multinomial logit model of consumers' choices among
passenger car and light truck makes and models. Three equations of make and model choice
were estimated using pooled data from the 1983 to 1987 Consumer Expenditure Surveys:
(1) small cars comprised of subcompacts and compacts, (2) luxury automobiles including sports
cars, and (3) all other vehicles. A likelihood ratio test rejected the hypothesis that the three
classes of vehicles had the same coefficient values. Fuel efficiency was represented by the
regional price of gasoline times the rate of fuel consumption in gallons per mile, i.e., fuel cost
per mile. In the small car choice equation, fuel cost per mile was decidedly statistically
significant; the coefficient of-7.143 being almost ten times its standard error. The coefficient
estimate of fuel cost per mile in the "Big Car" equation, -1.381, was only 1.86 times its standard
error, making it significant at the 10% but not the 5% level. In the luxury vehicle equation fuel
cost per mile had the wrong sign (+0.231) but was not close to being statistically significant.
Calculating the value of fuel economy using Goldberg's (1995) results is uncertain because she
is not precise about the units used when estimating her model. The variable definition for price,
for example is purchase price and vehicle specific dummy variables, and while the mean values
shown in her table A5 are in dollars, it is impossible to reconcile the estimated price coefficients
ranging from -0.5 to -4.7 with elasticities reported in table II of-1.1 to -5.2 (depending on class)
assuming price is in dollars. Goldberg (1996) use vehicle purchase price as a proxy for
annualized cost and fuel cost per mile, and reports a price coefficient of-2.991 and a fuel cost
coefficient of -0.425. Fuel cost is measured in cents per mile while purchase price is presumably
in dollars. The ratio of the fuel cost to purchase price coefficients is $0.14 per 1-cent-per-mile.
24
-------
Applying the standard estimate of discounted lifetime miles implies a present value of 1-cent-
per-mile of $1,126. Thus, it does not appear likely that purchase price is in simple dollars per
vehicle. The reported elasticities can only be reconciled with the reported parameter estimates
by assuming that vehicle price was in units of 10,000 dollars for estimation purposes. This
leaves the question of what units fuel cost is in, dollars, cents or some other units.
Goldberg's (1998) analysis is specifically focused on the short-run effects of the CAFE
standards, which limited manufacturers' strategies to pricing and domestic/import content
decisions. Changes to vehicle design and technology content were excluded. She estimated a
nested multinomial logit model of vehicle choice among nine car classes. Within each class,
consumers had a choice between a foreign or domestic car. The vehicle attributes included were
price, horsepower/weight, size, and fuel cost per mile. These were often interacted with
household characteristics such as income, household size, education level, etc. The principle
data source was the Consumer Expenditures Survey 1985 to 1989. Although Goldberg does not
report the empirical results of the estimation of her consumer choice model in her 1998 paper,
she does make specific observations about consumer valuation of fuel economy.
"A question of particular interest for environmental regulation is whether consumers
appear to be "myopic," in the sense that they respond more to current changes in
vehicle prices, than to changes in fuel costs that are felt over several periods.
"Assuming a discount rate of 5%, and an average vehicle holding period of 7 years,
produces vehicle price, and fuel costs semi-elasticity estimates that are very similar in
magnitude. Even though these calculations are sensitive to assumptions about vehicle
holding periods and discount rates, the numbers used above seem realistic enough to
safely say that we do not have any reason to believe that consumers are myopic."
(Goldberg, 1998, p. 21)
On the other hand, using our standard assumption of a 14 year lifetime, 112,600 discounted
miles, together with a 1987 fuel price of $0.96 and an average MPG of 18, produces a discounted
present value of $6,000. Goldberg reports a price elasticity of vehicle choice of-3.01 and a fuel
cost elasticity of vehicle choice of-0.5, implying that purchase price has six times the leverage of
fuel cost, in terms of a percent change. She also gives the average price of a new car as $12,000.
If a present value dollar of fuel cost were equal to a dollar of purchase price, the elasticity of fuel
cost (given our standard assumptions) should be half the purchase price elasticity. Since it is
one-sixth as large, the implication is that consumers undervalue fuel savings by 1:3.
3.1.3 Studies from the EU
Cambridge Econometrics (2008) carried out an econometric analysis of demand for cars and
their attributes for the Department for Transport, UK. The researchers estimated a mixed logit
model of vehicle choice using survey data on households in the UK who had purchased a new or
less than 12 month old car during the years 2004 and 2007. Households identified the
manufacturer, model and engine size of the vehicles they purchased. A separate data base on
vehicle attributes was mapped to the survey data to add information on vehicle attributes such as
transmission, purchase price and fuel consumption. It also provided an estimate of fuel costs for
25
-------
12,000 kms of travel. Interestingly, both fuel costs in pounds per 100 km, and fuel consumption
in liters/100km were included as vehicle attributes in the model estimation. Fuel consumption
was not statistically significant but fuel costs were. This may indicate a greater sensitivity to fuel
price than fuel economy but the authors did not explore this question.
The report estimates the value of a reduction in fuel costs by the ratio of the mean value equation
coefficients of vehicle price (-0.380) and fuel cost per 100km (-0.745), obtaining a ratio of
0.380/0.745 = 0.510. This is interpreted as a willingness to pay £510 per £l/100km reduction in
operating costs. This calculation implies that vehicle price is measured in thousands of pounds.
But this is not stated in the report, indeed, it is stated in several places that purchase price is
measured in units of pounds. Moreover, the ratio used is the marginal utility of a pound in
purchase price to the marginal utility of a pound per 100km in fuel costs. This ratio (actually the
negative of this ratio) is the marginal change in fuel costs that equates to a £1 change in purchase
price, rather than the change in purchase price that equates to a £l/100km change in fuel costs.
The latter value should be calculated as follows.
a;
8P -XF -0.745
- =_-££-=_ =-1 96
5F aj -0.380
8P
(9)
Thus, a decrease of £l/100km in fuel costs equates to a £1.96 increase in purchase price. That is,
consumers would be indifferent between no change in vehicle attributes and a decrease of
£l/100km in fuel costs coupled with a £1.96 increase in purchase price. Put another way,
consumers would be willing to pay up to £1.96 in increased purchase price for a £l/100km
reduction in fuel costs. If price is indeed measured in thousands of pounds, the willingness to pay
for a £l/100km reduction in fuel costs would be £1,960.7 An average annual vehicle use of
15,000 km is used in the study, implying a discounted lifetime kilometers traveled on the order
of 100,000 or more. This value appears to imply that consumers in the UK somewhat overvalue
fuel cost savings, since 100,000 km x £1/100 km = £1,000.
Eftec (2008) estimated an aggregate discrete choice model for purchases of automobiles by
households in the UK. Their data consisted of counts of the numbers of new vehicle
registrations of 2,190 different vehicle types, by households by region of the UK, from 2001 to
2006, a total of 847,689 records. Data on vehicle attributes were matched to the new registration
records, as were demographic and income data for the 11 Government Office Regions of Great
Britain in which the vehicles were registered.
The report states that"...we scale all cost related variables by the market average income..." and
in table 5.1 fuel cost per 100km is listed as one of the variables scaled by average income.
7 Although it does not appear to be clearly stated in the text, vehicle price was apparently in units of
£ 1,000s rather than pounds. Also, in the text of the report (p. 55) the willingness to pay for fuel economy is
calculated by dividing the price coefficient by the coefficient of fuel costs, the inverse of the correct procedure. An
e-mail request for clarification has been sent to the authors but no response has been received at time of writing.
26
-------
However, this assumption leads to the inference that households in the UK place very little value
on fuel costs, as will be shown. On the other hand, the mathematical derivation of the model
presented in chapter 3 states that fuel costs are treated differently from the fixed costs of vehicle
ownership and are included among the other physical attributes of the vehicle. This implies but
does not clearly state that fuel costs may not have been scaled by average income.
If one assumes that fuel costs have been scaled by average income, the ratio of the coefficients of
Fuel Cost/Purchase Price -((du/dFC)/(du/dPP)) yields an implied value of-£0.385 per increase of
£1 per 100km in fuel cost. Since, over the life of a car, there are likely to be on the order of
100,000 kms of travel, one would expect the present value of a change in fuel cost for a
£l/100km change in fuel costs to be on the order of £1,000, more than three orders of magnitude
larger. Average household incomes by region range from £26,921 to £40,540 in 2005, with the
median region at £30,189. Recalculating the value of a £l/100km change in fuel costs while
scaling just the purchase price coefficient by median income implies a value of
-((-0.3850)/(-1.0165))30189=- £11,434 per £l/100km. This estimate, on the other hand seems
roughly one order of magnitude too high.8 Neither result seems satisfactory.
The authors of the Eftec report are well aware of the statistical difficulties in estimating a vehicle
choice model as detailed as theirs.
"The estimation data set consisted of some 70,850 observations, where each
observation relates the market share of a particular type of vehicle in a particular
GOR in a particular year to the physical and cost attributes of that vehicle. This
represents one of the most comprehensive data sets of its kinds ever compiled.
All the same, it remains unlikely that the variables available to describe the
physical attributes are sufficient to capture the myriad details that differentiate
one vehicle from another.
"The existence of unobserved vehicle attributes is a real problem for the robust
estimation of the parameters of the utility function. Indeed, if unobserved
attributes are not controlled for in the estimation procedure, the estimated
parameters will in all likelihood be biased." (Eftec, 2008, p.xiii- xiv)
The authors attempt to control for the problem of unobserved attributes by including fixed effects
for every make, model, body type, and fuel combination. However, even this may not have been
enough.
Vance and Mehlin (2009) estimated a nested multinomial logit model of new car registrations in
Germany using the method of Berry, Levinsohn and Pakes (1995). The data were obtained from
R.L. Polk Europe and provide total sales by over 6,000 makes, models, configurations and body
styles each year, aggregated to 681 individual models spanning the years from 1995-2005 (5,007
observations). Two alternative nesting structures were estimated: (1) vehicles grouped into
eight classes by size, body style and price, and (2) a further subdivision within each class into
foreign and domestic brands.
8 A request for clarification of this point was sent to the authors via e-mail. Receipt of the e-mail was
acknowledged and a substantive response is expected soon but was not available in time for inclusion in this draft.
27
-------
Fuel economy was represented by the product of fuel price (/liter) multiplied by the New
European Drive Cycle fuel consumption rate in liter per 100 km, or /100km. Vehicle purchase
price was represented by the retail price divided by the national disposable income level per
capita. Other variables included in the model were:
1. The annual circulation tax (e.g., 5.11 to 6.75 per 0.1 liter for gasoline passenger cars)
2. Engine power in kW
3. Vehicle size, measured by length times width
4. Curb weight
5. Dummy variables for each year
The estimated coefficients of fuel costs per 100 km and car price/income were similar in the two
nested logit models. The model nesting by vehicle class only produced a fuel cost coefficient of
-0.197 and a vehicle price coefficient of-0.272, while the model nesting by vehicle class
produced corresponding estimates of-0.186 and -0.255. The ratio of the two, which determines
the willingness to pay for fuel economy, is nearly identical. Calculation of the willingness to pay
for fuel economy requires the average income per capita for the sample period ( 16,017, kindly
provided by the corresponding author) and data on vehicle use and expected lifetime. Annual
usage is assumed to be 16,000 kms per year for a new vehicle declining at 4% per year over a 15
year vehicle life. This yields 117,520 discounted lifetime kilometers or 1,175 segments of 100
kms each. The willingness to pay for a 1 reduction in fuel cost per 100 kms is given by the
following equation, in which E is fuel cost in Euros per 100 km, P is vehicle retail price in Euros
and Y is per capita income in Euros.
a;
Euros _ -0.197
1 Euro/100km~-0.272/16017 ~
(10)
This implies that a typical German new car buyer would be willing to pay 11,601 to reduce the
fuel costs of a new vehicle by 1/100 km. How reasonable this is requires knowing how many
hundreds of kilometers a new vehicle is likely to travel over its lifetime. Multiplying the
hundreds of kilometers by 1 and discounting to present value gives the present value of the
future savings in fuel produced by a 1 / 100 km reduction in fuel costs. Let K0 be the
kilometers traveled by a new vehicle and q be the annual rate of decline in kilometers traveled.
Let L be the lifetime of a vehicle in years and r be the consumer's annual discount rate. If a
vehicle in Germany travels an average of 16,000 kilometers in its first year, declining at 4% per
year over a 15 year life expectancy, and assuming a discount rate of 7% per year, a 1 change in
fuel costs per 100 km should be worth approximately 1,175 in present value.
L 15
= £L f Koe-V e-rtdt = ^- e-C04+.o7)t
0
(11)
28
-------
The willingness to pay per present value Euro in fuel savings calculated from Vance and
Mehlin's (2009) model in this way is almost ten Euros, and varies little between the vehicle class
( 9.87) and the class/origin ( 9.94) nesting strategies. It is not obvious why rational consumers
would be willing to pay an order of magnitude more for an improvement in fuel economy than it
is worth in present value.
3.2 HEDONIC PRICE MODELS
Fan and Rubin (2009) estimated a hedonic demand regression for the state of Maine in 2007,
using a two-stage method that allowed them to estimate a demand function for fuel economy.
Their first stage regression is of greatest interest here because it can be compared with the studies
by Espey and Nair (2005) and McManus (2007) discussed below. Fan and Rubin's data base
contained information on 523 passenger cars and 2,100 light trucks, with vehicle attributes
obtained from Ward's Automotive Group. A Box-Cox test for functional form indicated that the
log-log form should be used for estimation. Their first-stage model regressed log of
manufacturer's suggested retail price (MSRP) against the logs of MPG, vehicle weight, and
horsepower to weight ratio, and dummy variables representing vehicle class, transmission type
and manufacturer. A second first stage regression was run adding interactions between the
vehicle class dummy variables and the vehicle attributes, in order to explore heterogeneity in
consumer preferences across classes.
The results for passenger cars and light trucks indicated a willingness to pay for a 1 MPG
increase in fuel economy of $208 for cars and $233 for light trucks. The authors note that these
values are small in comparison to their estimated undiscounted lifetime fuel savings of $823 for
passenger cars and $1,461 for light trucks. They state that the implied discount rates are 48% for
cars and 13% for light trucks, but these estimates are not consistent with the numbers for lifetime
savings. However, recalculating the implied discount rates using national data on vehicle usage
and survival rates (U.S. DOT/NHTSA, 2006) and national average model year 2007 MPG
numbers (U.S. EPA, 2009), and assuming a national average gasoline price of $2.85 per gallon
(2007 $), this author obtains implied discount rates of 37% for cars and 77% for light trucks.
The estimates of willingness to pay for fuel economy by vehicle class, shown in Figure 4, raise
additional issues. They show negative willingness to pay for some significant vehicle classes
and substantial variability across classes that is not easy to explain. The authors' suggested
interpretation is consistent with omitted variables bias, that is, MPG is aliasing other vehicle
attributes which are not included in the regression equation but with which MPG is correlated.
"Although luxury and large car consumers may not be intentionally unwilling to
pay for higher fuel economy, luxury car consumers' preferences for luxury
features, and large car buyers' primary needs for large sizes might compromise
their value on fuel economy." (Fan and Rubin, 2009, p. 9)
29
-------
Estimated Willingness to Pay for a 1 MPG Increase in Fuel
Economy
(Fan and Rubin, 2009)
$2,000 -| $1,822
$1,500
$1,000
$0
$657
Small Car Midsize Car Large Car Luxury Car
SUV
Van
.$154
Pickup
Figure 4. Estimated Willingness to Pay for a 1 MPG Increase in Fuel Economy by Vehicle
Class. Source: Fan and Rubin (2009).
Espey and Nair (2005) used a hedonic model applied to 2001 automobiles to estimate the value
of fuel economy to consumers and concluded that consumers fully value lifetime vehicle fuel
savings, and perhaps more. The authors note the relative scarcity of studies (at that time) of the
value of fuel economy despite the importance of the issue.
"Surprisingly few economic studies have attempted to determine consumers'
willingness to pay for improvements in automobile fuel economy." (Espey and
Nair, p. 317)
The authors cite the method of Rosen (1974) as their methodological approach for regressing
vehicle price on attributes including fuel economy. However, Rosen's method explicitly
recognizes the simultaneity of the supply and demand for vehicle attributes and, therefore their
price whereas the authors use a single-stage estimation method based on least square corrected
for heteroskedasticity. A critical issue is whether the analysis is identifying the supply or
demand curve for fuel economy. The authors point out that their method depends on the market
achieving equilibrium.
"Given an equilibrium market, this value reflects both the consumer's marginal
willingness to pay for an additional unit of that attribute and the producer's marginal cost
of providing another unit of that attribute in that vehicle." (Espey and Nair, 2005, p. 3 18)
Given the existence of fuel economy standards that were a binding constraint on at least some
manufacturers during this period, it is not clear whether the analysis has identified the
30
-------
consumers' willingness to pay or the manufacturer's marginal cost. The full Rosen method, as
implemented by Fan and Rubin, does assure identification of the consumer's demand curve.
The data used in the study represented 130, 2001 automobiles whose attributes were obtained
from standard sources, such as Consumer Reports and Ward's Automotive Report web sites. In
addition to fuel economy, seven vehicle attributes were included in the regression analysis: size,
power, performance, safety, comfort, reliability and whether or not the vehicle is classified as a
luxury vehicle. In the models estimated, these factors were represented by curb weight, 0-60
mph acceleration, turning circle radius, braking distance in feet, the sum of NHTSA's front and
side crash test ratings, and a 1-5 comfort rating, a 1-5 reliability rating and a luxury designation
(presumably from Consumer Reports). Fuel economy was measured as its inverse, gallons per
mile, and city, highway and the combined average fuel economy ratings were tested separately.
In addition, four categories of the gas guzzler tax were also included.9 Four models were
estimated using least squares corrected for heteroscedasticity, each using a different fuel
economy variable: (1) city gpm, (2) highway gpm, (3) both city and highway gpm, and
(4) combined gpm. It is assumed that adjusted fuel economy numbers were used. The estimated
value of fuel economy ranged from $282/MPG for the highway MPG variable, to $613/MPG for
the combined MPG variable. The authors compute undiscounted values of fuel saving, assuming
vehicles travel 145,000 miles. This is lower than the NHTSA's estimate of undiscounted
lifetime miles (168,853) but higher than discounted lifetime miles (112,600). At $1.50 per
gallon, the undiscounted fuel savings for a vehicle driving 145,000 miles at the average
automobile MPG of 20.4 would be $561, while at $2 per gallon the undiscounted value would be
$747/MPG. Thus, Espey and Nair (Table 5) conclude,
"... model 4 estimates a higher value of an incremental change in fuel economy
than the actual fuel cost savings at a price of $1.50 per gallon.
"However, for higher fuel prices, the estimate from model 4 using average fuel
economy suggests only moderate discounting by consumers, at a rate of about 1%
for fuel prices of $1.75 per gallon and 4% for fuel prices of $2.00 per gallon."
(Espey and Nair, pp. 321-322)
9 Espey and Nair include the gas guzzler tax as if it were paid by consumers. In fact, the gas guzzler tax is
paid by manufacturers. However, the tax is required to be shown on the vehicle's window sticker.
31
-------
Table 5. Actual Versus Estimated Value of Fuel Economy
1 MPG increase in:
City Highway Average
MPG MPG MPG
Model 1 $531
Model 2 $282
Model 3 $440 $242
Model 4 $613
Actual undiscounted fuel savings assuming:
145,000 miles, $.150/gal $514 $110 $561
145,000 miles, $1.75/gal $600 $128 $654
145,000 miles, $2.00/gal $686 $146 $747
Source: Espey and Nair (2005, Table 4)
McManus (2007) also estimated a hedonic demand model including the price of gasoline
divided by fuel economy. His dependent variable was transaction prices for 445 vehicles during
the period 2002. Other explanatory variables were horsepower per ton of vehicle weight, vehicle
weight, real disposable income per capita and dummy variables for the brand of each vehicle.
The estimated coefficient of fuel price divided by MPG was -768. According to the variable
definitions, this should mean that an increase of 1 in fuel cost per mile would reduce the price of
a new vehicle by $768. Although McManus doesn't state so in his paper, it appears that the
measure of fuel cost is cents, rather than dollars per gallon. Translating to dollars per mile, the
coefficient would be 76,800. Since this coefficient translates dollars per mile into present value
dollars, its units are discounted lifetime miles. Using the standard assumptions in the appendix
to this paper, the discounted lifetime miles of a typical U.S. light-duty vehicle are 85,161, very
close to McManus' coefficient estimate.10 Like the results of Espey and Nair, McManus' results
imply that consumers would, approximately, somewhat under-value the lifetime fuel savings due
to fuel economy improvements, by about 10%. The two studies use a similar methodology. As
noted below in the discussion of Espey and Nair's work, no attempt is made to correct for the
potential endogeneity of vehicle prices and vehicle attributes, thus, there is a question as to
whether the supply or demand curve is being estimated. Also, by using fuel cost per mile as the
explanatory variable, the constraint that consumers respond equally and in opposite directions to
fuel price and fuel economy is enforced.
Fifer and Bunn (2009) estimated willingness to pay for reduced fuel consumption per mile by
means of hedonic regression analysis and compared that willingness to pay with calculated
present values of fuel consumption for passenger cars, SUVs, vans and pickup trucks. A key
feature of their analysis was accounting for the variation in annual miles traveled across
households and the implications for willingness to pay for fuel economy. Their data on 2,054
makes and models of new cars and light trucks for the period 1996 to 2005 included vehicle
price, fuel consumption (gallons per mile), weight, horsepower, engine displacement and
presence of airbags. Different coefficients for the fuel consumption variable were estimated for
10 These assumption are, 15,600 miles per year when new declining at 4.5% per year, an expected rate of
return on an investment in fuel economy of 12% and an expected 14 year lifetime (NRC, 2002, p. table 4-1).
32
-------
four broad market segments: (1) passenger cars, (2) SUVs, (3) vans, and (4) pickup trucks.
Dummy variables were also included for model year, manufacturer and "detail" market segment.
Detail market segments are subsets of the four segments for which different fuel consumption
coefficients were estimated (e.g., for passenger cars: subcompact car, compact car, midsize car,
large car, luxury car).
The authors note that 1996-2005 was a period over which there was very little variation in the
average fuel economy of new vehicles. In their sample, the average fuel economy of a new
vehicle was 21 MPG and was relatively constant over all ten years. Not only was there little
variation in new vehicle fuel economy (see Figure 1), but there was no variation in the fuel
economy standards until 2005, when the light truck standards increased from 20.7 to 21.0 MPG;
the passenger cars standards remained at 27.5 throughout the period. This is potentially
significant because if the fuel economy standards were a binding constraint on some or all
manufacturers for some or all of this period, hedonic regression estimates of consumers'
willingness to pay for fuel economy improvements could reflect the shadow price of the
standards rather than true willingness to pay.
To account for the effect of heterogeneity in vehicle use on the value of improved fuel economy,
the authors used data from the 2001 National Household Travel Survey (NHTS) to estimate
annual mileage for each of the four vehicle types. Median annual usage varied across the four
vehicle types as follows: cars, 9,185; pickup trucks, 9,500; vans, 11,400; SUVs, 11,800. The
NHTS does not measure the actual use of each vehicle over a full year. Instead, it develops a
"best guess" estimate based on a combination of odometer readings and survey responses.
Vehicles up to five years old were included in the calculation of the expected present value of
fuel savings.
Fifer and Bunn (2009) assumed consumers would perceive fuel prices to be a random walk, and
thus take the current price of fuel to be the best predictor of fuel price over the life of a new
vehicle. Indeed, fuel prices varied little over the sample period. Fuel prices increased from a
median (over states) of $1.23 in 1997 to $1.40 in 2001. The expected present value of fuel
savings was calculated as the current price of fuel time the sum over the vehicle's expected
lifetime (assumed to be 14 years for all vehicle types) of the product of miles traveled, multiplied
by the rate of fuel consumption per mile, times a discounting factor for the year in question.
Discount rates of 3% and 7% were used. Using this method and the 3% discount rate, they
calculated the expected present value of a 0.001 gallon per mile reduction in the rate of fuel
consumption to be $167.42 for cars, $193.97 for pickups, $194.67 for vans and $197.78 for
SUVs.
The hedonic regression results are shown in Table 6. The regression is linear in levels of the
variables, and all variables are statistically significant except the airbags dummy, and the
coefficient of gallons per mile for cars in Regression 1. Regression 2 combines cars and SUVs
to estimate a single coefficient for both variables. The willingness to pay for a reduction in fuel
consumption of 0.001 gallons per mile equals the coefficient estimate in Table 6 divided by
1,000 (for example, for cars & SUVs in Regression 2, that would be 87,349.1/1,000 = $87.35).
The most striking feature of the regressions is the very large differences in willingness to pay
between purchasers of cars or SUVs and vans or pickup trucks. According to Regression 2, on
33
-------
average, pickup truck buyers will pay more than six times as much to reduce their vehicle's rate
of fuel consumption (gallons per mile) as car or SUV buyers, and purchasers of new vans will
pay more than five times as much. These differences cannot be accounted for by the somewhat
higher rates of usage for pickups and vans, about 20%, on average.
Table 6. Fifer and Bunn's (2009) Hedonic Regression Results
Variable Coefficient
Regression 1
Regression 2
Std. Err.
t-stat.
Coefficient Std. Err.
t-stat
cars*GPM -
vans*GPM -
SUVs*GPM -
pickups*GPM -
(cars+SUVs)*GPM
Hp 2
Wt7
Displacement 3986.338
Airbags 13
Detail Segment Dummies
Year Dummies
Manufacturer Dummies
cons -
73022.64
471770.3
126319.9
553761.6
8.78605
.693808
0.4298
9576.885
43490.2
103601.6
60676.48
95299.28
-1.68
-4.55
-2.08
-5.81
6.204614 4.64
0.521927 14.74
408.7145 9.75
447.148 0.29
3910.963
2.45
-468153
-549569
-87349.1
103508.6
95167.13
40142.99
29.90907 6.064106
7.609789 0.5125924
3954.065 406.9469
138.2672 447.0245
-11244.6
3391.639
-4.52
-5.77
-2.18
4.93
14.85
9.72
0.31
-3.32
Number of obs
F(54, 1993)
Prob > F
R-squared 0
Adj R-squared
Root MSB
2048
216.4
0
.8543
0.8503
5697.1
2048
220.5
0
0.8542
0.8504
5696.8
On the other hand, when the authors aggregate all vehicle types, they obtain results very similar
to those of Espey and Nair (2005) and McManus (2007), namely that consumers' willingness to
pay for fuel economy improvement, on average, is approximately equal to the present value of
fuel saved.
"Combining observations from all four vehicle segment (sic), there is a net underpayment
of $4.93 per capita for the low-end discount rate and a net overpayment of $37.17 per
capita for the high-end rate. While this combination of all vehicle segments suggests a
small to negligible net overpayment, car and SUV buyers generally reap large benefits
while van and pickup buyers incur large losses from their investments in fuel economy.
Aggregating the results by summing across all vehicle segments misses the more nuanced
result that "rationality" depends greatly on the type of vehicle purchased and driven by
the consumer." (Fifer and Bunn, 2009, p. 20)
The authors, however, are not able to come up with an explanation that satisfies them as to why
apparent rationality should vary by vehicle type.
34
-------
"This raises an interesting question that is beyond the scope of our analysis: unless it is
technologically infeasible, why don't van and pickup producers introduce more fuel
efficient vehicles to capture the profits suggested by these hedonic prices?" (Fifer and
Bunn, 2009, p. 21)
In fact, analyses of the technological potential and cost of improving fuel economy have found
that it is no more, and perhaps less costly to improve the fuel economy of vans and pickup trucks
than of passenger cars and SUVs (e.g., NRC, 2002).
The data used by Fifer and Bunn (2009) presents the usual challenges for statistical inference of
the value of automobile attributes. The three included continuous variables (horsepower, weight
and engine displacement) are strongly correlated with fuel consumption and with each other. For
example, Knittel (2009) found that weight and horsepower, together with dummy variables
representing diesel engines and manual transmissions, achieved an R2 of 0.848 in a double log
regression on fuel consumption. In addition, manufacturers generally package other features
with engine options (sunroofs, alloy wheels, premium floor mats or upholstery, special paint,
etc.). The extent of packaging can vary considerably. For example, the 2010 Camry SB's
6-cylinder package adds $5,395 to the price of the vehicle while the Chevrolet Malibu's
6-cylinder option adds only $1,795. The two 4-cylinder and the two 6-cylinder engines are
almost identical in size and power. Presumably, there is more in the Camry's 6-cylinder
package. Such differences will not be captured by make and model dummy variables, let alone
vehicle class dummies. In addition, there is the question of the impacts of fuel economy
standards. Fifer and Bunn do not attempt to represent the effects of the standards in their model.
This leaves open the possibility that if the standards were a binding constraint on manufacturers,
what Fifer and Bunn have estimated is the supply function for fuel economy rather than
consumers' willingness to pay. If the standards were binding on some manufacturers and not
other, or if they were binding in some years and not other, the identification of the estimates as
consumers' willingness to pay becomes even more complicated.
Arguea, Hsiao and Taylor (1994) estimated the marginal value of fuel economy using hedonic
regression analysis and eighteen years of data from 1969 to 1986. Fuel economy was
represented by Consumer Reports magazine's open-road miles per gallon estimates. Ordinary
least squares were used to estimate a linear hedonic price function in current dollars. The
resulting year-by-year hedonic price coefficients for fuel economy represent the value of an
increase of 1 MPG, and are illustrated by the gray bars in Figure 4. The present value of lifetime
fuel savings resulting from a 1 MPG increase in the highway fuel economy of an average
passenger car (U.S. EPA, 2009, table 1, adjusted highway MPG) was calculated using the
standard vehicle usage and discounting assumptions (see appendix), and the current dollar price
of gasoline in each year (U.S. DOE/EIA, 2009, table 5.24). The ratio of the hedonic price
estimate to the estimated present value of fuel savings, expressed as a percent, is shown by the
larger unshaded bars in Figure 5. In general, the apparent willingness to pay for fuel economy is
in the vicinity of 5% to 10% of the calculated present value of fuel savings, except for the final
year, 1986, in which the estimated willingness to pay is 46% of lifetime fuel savings.
35
-------
45%
40%
35%
30%
25%
20%
15%
10%
5%
0%
Estimated Hedonic Value ($/MPG) of a 1 MPG Increase in Fuel
Economy and Percent of PresentValue Lifetime Savings
9 % °f py
$/MPG
1970
1975
1980
1985
Figure 5. Annual Estimates of the Hedonic Value of a 1 MPG Increase in Fuel Economy and
the Percent of Estimated Lifetime Fuel Savings Each Represents (Arguea, Hsiao and Taylor,
1994).
Espey and Nair (2005) use data from a single year, 2001, to estimate their model. As Arguea,
Hsiao and Taylor (1994) point out, this poses a potential problem because the estimated hedonic
prices of attributes represent an assumed equilibrium of supply and demand.
"From a single shadow price representing equilibrium of market demand and
supply, it is not possible to separately identify the structural demand and supply
functions for these characteristics." (Arguea, Hsiao and Tayler, 1009, pp. 6-7)
The potential problem arises if for some reason supply and demand are not in equilibrium. There
is a strong possibility that what McManus, Espey and Nair, and Fifer and Bunn have estimated is
the supply function for fuel economy and other vehicle attributes or a mixture of supply and
demand. Fuel economy standards for automobiles were in effect for the years covered by both
analyses. What studies finding cost-effective willingness to pay for fuel economy may be
showing is that the fuel economy standards were binding and that they required manufacturers to
provide levels of fuel economy that were approximately cost-effective based on expected
lifetime fuel savings. In the event the fuel economy standards were not binding, then market
forces may have led manufacturers to provide cost-effective levels of fuel economy. In the latter
case, one could reasonably assert that the marginal cost of fuel economy should equal the
marginal willingness to pay, and thus should be a relevant estimator of it. If the former, then
marginal cost would equal or exceed marginal willingness to pay. Without a clear identification
of the supply and demand curves, the question cannot be resolved definitively. Determining
when fuel economy standards were binding and on which manufacturers might be useful for
identifying the consumers' demand function.
Fan and Rubin's (2009) results, however, were inconsistent. They found that MPG coefficients
for luxury cars, large cars and pickup trucks all indicated a negative value for higher fuel
36
-------
economy. These results should probably be interpreted as coefficient bias as a result of fuel
economy aliasing unobserved attributes. Other interpretations seem implausible: (1) purchasers
of these vehicles are averse to saving money and prefer to make extra trips to the refueling
station, and (2) for purchasers of these vehicles, there is a snob effect associated with lower fuel
economy, everything else equal. This last interpretation seems to be favored by Fan and Rubin
but it is inconsistent with the assumption that other factors are held constant. It is difficult to
believe that even if consumers could have all the other attributes associated with luxury and size
and improved fuel economy they were willing to pay for, they would still prefer lower fuel
economy. This implies a negative utility for saving money, which seems implausible for large
market segments.
3.3 OTHER METHODS
Kilian and Sims (2006) analyzed the effect of changes in fuel cost per mile on the prices of
individual makes and models of used cars in the United States for the years 1978-1984. Their
partial equilibrium model of automobile pricing considers automobiles to be assets producing a
flow of services to households over the lifetime of the vehicle. The supply of used vehicle
services is assumed to be fixed, which is not entirely correct since vehicle lifetimes can be
extended by investment in maintenance and repair. The annual rental rate (R;jt = value of flow of
services for vehicle i of age j in year t) is a function of services (U) derived from observable and
unobservable attributes (Z) and operating costs, represented by fuel cost. Fuel costs are equal to
the price of gasoline (pt) times annual miles of travel by vehicle i (niit) divided by the fuel
economy of vehicle I (MPG;).
(12)
The current price of a used vehicle (P) is assumed to be the present, discounted value of future,
expected service flows, or rental flows over the vehicle's lifetime (L).
7 = 0
(13)
From the equation for used car prices it is clear that a decrease in the present value of expected
fuel costs of a certain amount should increase the price of a used car by the same amount. If the
other attributes of a vehicle are assumed to remain constant, then the effect of an increase in the
price of gasoline on the price of a used car is just the discounted present value of the increase in
expected fuel costs.
37
-------
7=0
(14)
In the form of the model used for estimation, the change in used car prices and all other variables
are expressed relative to the initial used car price. In the equation below, 5t represents fixed time
effects and C0yt represents the flow of vehicles services (other than fuel cost) in time period t.
t fst
= -(St + a)ijt t+1
rijt rijt
(15)
Kilian and Sims assume that all vehicles have the same lifetime (L = 10 years), the same usage
rates (mit = m = 10,000 miles per year) and that all consumers have the same discount rate (5%
per year). In general, these assumptions appear to be low relative to other data sources (see
appendix) which would tend to underestimate the present value of a change in fuel prices and
overestimate the impact of fuel price changes on used car prices. In addition to these
assumptions, the model also requires a way to represent consumers expectations about future fuel
prices. The reference assumption is static expectations, an optimal method if gasoline prices are
a random walk. However, a variety of other price expectation models were tested, including
fitted ARTMA and GARCH models, as well as lagged updating of expectations.
Data on fuel economy for 2,780 light-duty vehicles for the period 1978-1984 were obtained from
official EPA/DOE estimates. Used car price data were collected from the National Automobile
Dealers' Association Used Car Price Guides for the years 1979-1989, since used cars up to five
years of age were included in the analysis. Gasoline price data were obtained from the Energy
Information Administration's Monthly Energy Review.
The coefficient of the present value of the change in gasoline price relative to the lagged used
vehicle price (y) is predicted to be -1 by Kilian and Sims' theory. Results from the baseline
model employing static expectations finds a coefficient of-0.1096, indicating that only 11% of
the change in present value expected fuel costs translates into a change in used vehicle prices.
Interpreted as a willingness to pay for fuel economy, it would imply that in used vehicle
transactions, consumers are willing to pay for only 11% of a reduction in fuel costs per mile.
The authors provide a sample calculation for the effect of a $0.25 per gallon price change for two
vehicles costing $10,000 each, with one getting 15 MPG and the other 25 MPG. Their theory
predicts that the 15 MPG "gas guzzler" should experience a reduction in price of $1,000 while
the 25 MPG vehicle should lose only $600, for a net advantage of $400 for the more fuel
efficient car. The estimated model, however, predicts only a $50 relative price change.
Including different formulations of fixed effects did not change the resulting estimates in any
important way. Excluding all foreign autos, light trucks and diesel autos, however, produced a
coefficient estimate of-0.288, still far from -1.0. Fitting an ARIMA model to prices and using it
in place of static expectations resulted in uniformly lower impact estimates. The authors then
systematically varied their assumptions about vehicle life, usage and discount rates. Only by
38
-------
decreasing annual use to 5,000 miles per year, lowering the expected lifetime to 6 years and
assuming discount rates in excess of 10% could coefficient estimates of-1 be obtained.
"We thus conclude that the baseline model cannot be rationalized for reasonable
calibrations of the parameters or assumptions regarding expectations formation. .. .Either
individuals are extremely myopic and do not factor the near future into their decisions or
the theory and/or empirical specification is flawed in a more fundamental sense." (Kilian
and Sims, 2006, p. 16)
The authors then try several other models of gasoline price expectations, including a twelve
month average and a random walk model with a GARCH error term. This produced coefficient
estimates as high as -0.25 in one case. Finally, the authors test for asymmetric response to
gasoline price increases and decreases. In years in which gasoline prices rose, the coefficient of
the cost per mile variable was quite close to -1, -0.887, and did not differ significantly from -1.
However, in years in which gasoline prices fell the estimated y was +0.08 and was statistically
significantly different from zero. This result was generally consistent across alternative
formulations and estimation methods.
"In particular, a robust finding is that gasoline price increases have a relatively strong
effect on used automobile prices while decreases do not." (Kilian and Sims, 2006, p. 25.)
The results on asymmetry may be relevant to the willingness to pay for increased fuel economy,
since an increase in fuel economy decreases fuel cost per mile. In fuel economy terms, the
asymmetry results imply that consumers will pay nothing or close to nothing for increased fuel
economy but would pay close to expected present value to avoid decreased fuel economy. If this
result proved to be valid, it would imply that consumers respond differently to fuel price and fuel
economy.
Bhat and Sen (2006) constructed a model of household vehicle holdings and use by vehicle
type, using data from a 2000 survey of consumers in the San Francisco Bay area. Their multiple
discrete-continuous extreme value (MDCEV) model did not include the price of vehicles but did
include their fuel economy via a variable representing fuel cost per mile divided by household
income. Because vehicle price did not appear in their model, the value of fuel economy cannot
be directly calculated. However, Bhat and Sen did compute the elasticities of vehicle choice by
vehicle type with respect to fuel cost per mile by varying fuel cost from $1.40 per gallon to $2.00
per gallon (a 35% increase based on the midpoint formula). These elasticities are quite low
compared to typical estimates of the price elasticity of vehicle choice, which are on the order of
-2 to -3. Very roughly, the discounted present value of total lifetime fuel costs is on the order of
$10,000, while a typical vehicle costs on the order of $20,000. Thus, a given percent change in
discounted present value fuel costs (represented by a percent change in fuel cost per mile) should
have approximately half the impact on vehicle holdings as the same percent change in own price.
Given this, Bhat and Sen's elasticity estimates shown in table 7 suggest a significant under-
valuing of fuel economy.
39
-------
Table 7. Elasticity of Household Vehicle Choice with Respect to Fuel Cost per Mile
Pass.
% change in vehicle holdings
Fuel cost per mile elasticity
Car.
-0.1
-0.002
SUV
-5.9
-0.169
Pickup
-2.1
-0.006
Minivan
-4.9
-0.140
Van
-3.4
-0.097
In an analysis of the short-run effects of fuel prices on vehicle prices, Langer and Miller (2008)
found that when gasoline prices rise, the prices of inefficient vehicles fall while those of
especially efficient vehicles rise. Using 300,000 observations of weekly automobile price data
over the period 2003-2006, they measured the extent to which the prices of new light-duty
vehicles responded to changes in the price of gasoline. Their dependent variable was a
constructed "manufacturer price," comprised of the manufacturers' suggested retail prices minus
the average incentives offered by GM, Ford, Chrysler and Toyota in for each week in five
geographic regions during 2003-2006.
Their fundamental equation for empirical estimation expresses the price of vehicle j, in time
period t, in region r (pjtr) as a function of the price of gasoline (gptr) divided by fuel economy
(mpgj), the sales-weighted average fuel cost per mile of other vehicles made by competitors and
the weighted average fuel cost per mile of vehicles made by the same manufacturer. A vehicle is
defined as a make/model (e.g., Ford Taurus) in a specific model year. The equation also includes
fixed effects for time periods (5t) and vehicles (KJ) as well as third order polynomials of variables
representing the length of time vehicle], vehicles produced by other manufacturers and vehicle
produced by the same manufacturer other than j have been on the market, and the weighted
average lengths of time other vehicles have been on the market. The functions involving length
of time in the market are not shown in equation 5.
Patr , ». V Pa* , ». V Pa* , o , , ,
1 * x - ' * x Win? + 8<- + ki+nr
mpgj
I,
(16)
Equations were estimated with fixed values of the b's for all manufacturers, as well as allowing
the b parameters to vary across manufacturers. In the regressions with fixed b values across
manufacturers, own fuel cost, bi, was consistently statistically significant and negative in sign.
The effect of the average competitor fuel cost was similar in magnitude, statistically significant
but opposite in sign, as might would be expected (Table 8). The effect of same firm fuel costs
was generally not statistically significant. Langer and Miller point out two caveats concerning
their manufacturer price variable. First, it is not a transaction price, and the negotiations during
transactions could push the transaction price in either direction from the manufacturer price.
Second, the manufacturer price is based on the average incentive program on offer during the
period rather than the actual incentives taken by consumers. If manufacturers allow more
combining of incentives during high gasoline prices than during other periods, the impacts of
gasoline prices may be underestimated.
It is convenient to measure the magnitude of the effect of fuel costs on vehicle prices relative to
the change in expected fuel costs caused by a change in the price of gasoline. Langer and Miller
40
-------
call this the "offset percentage." In calculating expected lifetime fuel costs, they assume static
expectations in their base model but explore alternative price expectation models. The median
offset percentages in their base model are 18% for cars, 15% for SUVs, and much less for light
trucks and vans. For nearly all vehicles, higher gasoline prices reduced manufacturer prices. For
especially efficient vehicles, however, higher gasoline prices actually increased the manufacturer
prices. For example, among SUVs the Ford Escape Hybrid, Mercury Mariner Hybrid, Toyota
Highlander Hybrid and Lexus RX 400 Hybrid were the only SUVs whose prices increased when
gasoline prices increased.
Table 8. Manufacturer Prices and Fuel Costs
Incentive
Variables Regional+
Fuel cost
(7.73)
Average competitor fuel cost
(7.15)
Average same-firm fuel cost
(2.29)
R2 0.5260
# of observations
# of vehicles
National
(1)
-55.40
50.76
1.15
299,855
681
Level
Regional
Only
(2)
-56.96
(7.86)
50.16
(7.39)
2.62
(1.78)
0.6763
299,855
681
National
Only
(3)
-63.75
(8.77)
50.09
(8.12)
1.31
(2.30)
0.5289
59,971
681
Source: Longer and Miller (2008, table 3)
While Langer and Miller's results clearly indicate that manufacturers adjust vehicle prices when
fuel prices change, they do not indicate whether consumers are over- or undervaluing the
potential fuel savings. This would require an understanding of how consumers form price
expectations for the life of a vehicle, an issue that Langer and Miller do not believe could be
resolved with the data used in their study. It would also depend on the short-run price elasticities
of vehicle demand and supply. If new vehicle supply were perfectly elastic, a change in fuel
costs would affect quantities sold but not prices. If supply were perfectly inelastic, price would
change by the full present value of the change in expected lifetime fuel costs. This question is
also not addressed by Langer and Miller. However, their results suggest that either new vehicle
supply is very elastic or consumers undervalue fuel economy.
Busse, Knittel and Zettelmeyer (2009) estimated the effects of changes in gasoline prices on
the market shares and prices of new and used automobiles. Using data from a 15-20% sample of
actual vehicle transactions at new car dealers from September 1, 1999 to June 30, 2008, they
estimated regressions relating gasoline prices and a variety of control variables to market shares
and transaction prices, separately for new vehicle purchases and used vehicle purchases. Fuel
economy was brought into the analysis by ranking vehicles' fuel economy and estimating
separate equations by quartile. They found that both market shares and vehicle prices responded
to fuel price changes in ways that were predictable by economic theory.
41
-------
Busse et al.'s (2009) findings concerning the effect of gasoline prices on new and used car prices
are most relevant for this study. The estimated coefficients for gasoline price for new cars by
MPG quartile are shown in Table 9. For new cars, a $1 increase in the price of gasoline reduced
the price of the average vehicle in the least efficient quartile by $246, and increased the price of a
vehicle in the most efficient quartile by $136, for a net effect of $382. Although two of the four
coefficients are not statistically significant, they all follow a logical pattern of increasing cost
with decreasing fuel economy. According to estimates made by the authors, the difference
between the 1st and 4th quartile values is equal to 1.2 years of (undiscounted) fuel savings (based
on an average MPG of 27.9 for the 1st quartile and 16.2 MPG for the 4th, and an average annual
mileage of 12,000). These results do not necessarily imply that new car buyers undervalue fuel
economy, since the change in price also depends on the elasticities of supply and demand and on
the formulation of price expectations.
If used cars supply is perfectly inelastic, then the shift in the demand curve caused by higher fuel
prices would exactly equal the reduction in market price. If supply is highly price elastic, the
observed change in market equilibrium prices would be only a small fraction of the change in
lifetime operating costs even if consumers fully valued lifetime fuel costs because the adjustment
to a shift in the demand curve would be primarily a quantity adjustment. On the other hand, if
the supply of used vehicles is almost perfectly inelastic, then the change in price would be nearly
equal to the change in lifetime operating costs. Price expectations matter because if consumers
expect continuing increases whenever fuel prices increase, the calculated change in fuel costs
based on static expectations would greatly understate consumers' expected change in fuel costs,
and overestimate their willingness to pay for fuel economy.
Table 9. Gasoline Price Coefficient Estimates: New Car Price Equation
Variable Coefficient (std. error)
Gasoline Price * MPG Quartile 1 Dummy -246 (75)
Gasoline Price * MPG Quartile 2 Dummy -81 (40)
Gasoline Price * MPG Quartile 3 Dummy 5.2 (30)
Gasoline Price * MPG Quartile 4 Dummy 136 (43)
Source: Busse et al. (2009, p. 19)
The results for used car market shares and prices were essentially the reverse of the new car
results: market shares changed little while prices changed a great deal. Table 10 shows the
gasoline price coefficients by MPG quartile for the used car price equation. The pattern is the
same but the magnitude of price changes is almost an order of magnitude greater. For used cars,
the difference between the 1st and 4th quartiles is $2,723, equivalent to nine years of
undiscounted fuel savings. Considering that used cars have a shorter expected life than new cars,
the results for used cars suggest an overvaluing of fuel costs by used car buyers.
42
-------
Table 10. Gasoline Price Coefficient Estimates: Used Car Price Equation
Variable Coefficient (std. error)
Gasoline Price * MPG Quartile 1 Dummy -1096 (38)
Gasoline Price * MPG Quartile 2 Dummy -936 (57)
Gasoline Price * MPG Quartile 3 Dummy 76 (71)
Gasoline Price * MPG Quartile 4 Dummy 1627 (56)
Source: Busse et al. (2009, p. 22)
In general, Busse et al. (2009) found that new car market shares changed more than used car
market shares in response to fuel price changes, while used car prices changed more than used
car prices. The explanation focused on the supply side and market efficiency. The used car
market was believed to function most efficiently, on the premise that there are many used car
buyers and sellers (individuals), frequent transactions, and that buyers and sellers alike
experience the same change in operating costs when gasoline prices rise, which helps to equate
the changes in willingness to pay and willingness to accept. The authors argue that the new car
market differs primarily in that the new car manufacturers have market power, and therefore
choose not to lower prices when the demand curve shifts downward but rather to accept loss of
volume, in part due to the potential damage to their brand's reputation. An alternative
explanation could be that the supply of new cars is more price elastic than the supply of used
cars. In this case, a downward shift in the demand curve caused by an increase in operating costs
would produce a smaller change in price in the new car market but a greater change in sales, or
market shares.
The authors performed a variety of sensitivity tests as a check on the robustness of their results.
In general, these changes did not fundamentally alter the results described above. A question
that may not have been satisfactorily answered is whether price expectations play an important
role. The authors tested their model with and without a variable that interacted gasoline prices
with an indicator of whether prices had been rising, falling or constant over the past three
months. Given the historical volatility of gasoline prices, three months may not be enough time
for consumers purchasing a long-lived asset like a car to form a strong view about future price
trends. Gasoline prices were generally trending upward over the study period, but especially
during the final three years. If consumers' price expectations took longer to develop, this could
have an effect on the coefficient estimates.
Robustness tests with and without controls for seasonal and fixed effects show the greatest
differences in the estimated coefficients (Busse et al., 2009, table 2). For example, with
year*region and month-of-year*region dummies included, the coefficient of gasoline price for
the fourth MPG quartile is 136; if the year*region dummies are dropped, that coefficient changes
to -796. In the used car equation, the same coefficient switches from 1,627 to -4,053 under the
same change in dummy variables. However, in both equations the difference between the first
and fourth quartile coefficients is about the same (382 versus 386 in the new car equations and
2,723 versus 2,919 in the used car equations). Still, there is a suggestion in these results that the
differences in gasoline price coefficients across equations may be aliasing other factors that vary
by quartile.
43
-------
In the regressions for the probability of vehicle purchases within a fuel economy quartile the
authors include an extensive set of control variables for demographics, region and timing of
purchase, but they do not include explicit measures of other vehicle attributes. (In other
regressions, such as for the probability a purchase will be a new car, they do include fixed effects
for detailed vehicle types: make, model, model year, trim level, number of doors, body type,
displacement, number of cylinders and transmission type). Since a variety of other vehicle
attributes are closely correlated with fuel economy (e.g., acceleration performance, size, luxury
features) it is not clear that the fuel economy quartile-specific price coefficients are not aliasing
the effect of other vehicle attributes. In this regard, it may be significant that the new and used
car price coefficients by quartile are estimated in a single equation without quartile-specific
intercepts or other quartile specific variables (tables A-3 and A-7). On the other hand, the
authors estimate separate equations for market segments, so not only intercepts but all other
parameters are specific to the market segment. In these equations, the market segment-specific
constants are significant at the 1% level, and the coefficients of many other variables appear to
vary significantly across car classes. These results suggest that there may be unobserved
attributes that affect car prices that may be correlated with MPG, and therefore may be biasing
the estimated gasoline price coefficients in the MPG quartile equations.
The effect of gasoline prices on fuel economy via two "channels," changes in the distribution of
new vehicle purchases and in the composition of the on-road stock of vehicles, was investigated
by Li, Timmins and von Haefen (2009). In general, they found that fleet fuel economy was
relatively insensitive to the price of gasoline.
"We find that gasoline prices have statistically significant effects on both
channels, but that their combined effect results in only modest impacts on fleet
fuel economy. The short-run and long-run elasticities of fleet fuel economy with
respect to gasoline prices were estimated at 0.022 and 0.204 in 2005." (Li,
Timmins and von Haefen, 2009., p. 135)
Using model and vintage-specific data on vehicle registrations and scrappage from 20 major
metropolitan areas in the United States, they estimated models of (1) the distribution of vehicle
sales, and (2) vehicle survival rates by fuel economy level. They did not attempt to estimate the
impacts of fuel prices on manufacturers' decisions about vehicle technology and design.
The dependent variable of their vehicle sales model is the number of new vehicles registered in
fuel economy quantile q, in year t, in metropolitan area m. In the new vehicle model, vehicles
were classified into 68 fuel economy quantiles. The key explanatory variables were the price of
gasoline and gasoline cost per mile (price divided by MPG). Several other metropolitan area
attributes were included, both individually and divided by fuel economy, in order to allow
heterogeneity in parameter values across metropolitan areas. Dummies were included for
quantiles, year, and region. A lagged adjustment formulation was used and tests for serial
correlation and heteroscedasticity were conducted.
p
ln(Nqtm) =b0+bi ln(Nqt.lm) +b2 ^ +b3Ptm+Other Controls+sctm
IVLrVjqtjjj
(17)
44
-------
In models including sufficient control variables, autocorrelation in the error term was not
statistically significant. In the preferred models, the coefficient on the lagged dependent variable
was quite small (0.068 to 0.074) indicating that adjustment of new vehicle purchases to gasoline
prices was almost entirely complete within one year. In the preferred model, an increase in
gasoline prices was found to increase the sales of vehicles in the 60th percentile of the MPG
distribution above, and decrease sales of vehicles below the 60th percentile.
In the vehicle scrappage model, the survival probability of a vehicle was assumed to be a logistic
function of fuel economy, the price of gasoline and other variables controlling for metropolitan
area and vehicle attributes. Two data sets were combined using a method of moments estimation
technique: one with make/model detail covering the period 1997 to 2000 and the other with only
vehicle class detail for the period 2001 to 2005. The preferred models included only vehicles
older than 10 years, which were assumed to be more sensitive to changes in fuel prices and less
affected by migrations of vehicles into and out of metropolitan areas. In the preferred model,
higher gasoline prices increased the survival probabilities of vehicles with greater than 28.7
MPG, and reduced survival rates for vehicles with lower MPG. The average elasticity of vehicle
survival with respect to the price of gasoline was low, -0.026 in the preferred model, but
elasticities across vehicle types and areas were up to 5 times as large. Nonetheless, relative
insensitivity of survival rates to gasoline price suggests that the fuel economy of the on-road
vehicle stock will also be relatively insensitive to the price of gasoline.
Simulations using the new vehicle sales and scrappage models confirmed that gasoline price had
a very small impact on the fuel economy of the on-road vehicle stock (via its composition). The
estimated short-run elasticities of MPG with respect to the price of gasoline were 0.191 for new
vehicles, 0.006 for used vehicles and 0.022 for all vehicles. In the long run, the greatest change
in fleet MPG will come about via newer more efficient vehicles replacing the used vehicle stock.
In the long run, the simulations indicate a price elasticity of 0.204, that is, a 10% increase in the
price of gasoline would eventually result in about a 2% increase in fuel economy. These
elasticities were calculated for 2005 when the price of gasoline was $2.34 per gallon. The model
implies that elasticities will increase with increasing gasoline price. At $4 per gallon, the
elasticities increase by about 50% to 0.033 in the short run and 0.330 in the long run. These
elasticities include only the effects of fuel price on consumers' choices among existing new
vehicles and not manufacturers' decisions about the technological content and design of vehicles.
For the existing fleet, only the impact on the composition of the fleet and not its operation (e.g.,
speed, maintenance, driving style) are included. Still, this paper offers a rare quantification of
both the new vehicle sales mix and on-road fleet composition impacts of gasoline price and fuel
economy.
The most recent study of the impact of gasoline price changes on vehicle prices also uses the
most detailed data. Sallee, West and Fan (2010) estimated a model using approximately
8 million wholesale market transactions for used vehicles. Because the wholesale transactions
take place at auction houses, the market is likely to be well-informed and efficient. While the
data do not represent retail transactions with final purchasers, it seems reasonable to believe that
the wholesale market reflects the likely responses of the retail market. The model is similar to
other but includes odometer readings for each vehicle, as well as time-period fixed effects for
detailed vehicle types (make, model, model year, cylinder count, engine displacement,
45
-------
transmission and trim style). Let P;jt represent the price of the ith vehicle record of type j, in year
t, let Cjjt represent the estimated discounted present value of its remaining fuel costs, O;jt
represent the vehicle's odometer reading at time of sale, 5jt represent a fixed effect for that
vehicle type and time period and Syt represent random errors.
Pijt = PCijt + 2^ aajOfjt + 8Jt + eijt
0=1
(18)
In equation (18) the polynomial odometer terms and the time period fixed effects are intended to
represent the value of future services from the vehicle, as well as shocks not reflected in gasoline
prices, such as economic conditions.
Expected future fuel cost is the central variable in the analysis. For fuel price expectations, the
authors assume buyers consider fuel prices to follow a random walk so that the current price
becomes the best estimate of future prices. Other formulations were tested, however, with
similar results. Consumers were assumed to discount future costs at r = 5% per year, although
rates of 10% and 15% were also tested. Fuel economy is assumed to be constant over the life of
the vehicle.
r R
1 mjsps
'i)t
- s=t J-
(19)
The function Et[] denotes expected value, which could vary by time period. The function H()
represents the probability of survival, which is a function of the vehicle's odometer reading and
whether it is a passenger car or light truck.
The most important determinant of expected present value fuel costs is the expected remaining
vehicle mileage; since it identifies the cost effect versus the time period fixed effects. The key
assumption is that mileage, rather than age, is assumed to be the relevant measure of the position
of a vehicle in its lifecycle. While other studies have determined remaining expected lifetime
using a fixed number of years (e.g., 15), Sallee et al. assume vehicles will travel a fixed number
of miles but that cumulative lifetime miles vary by vehicle type (cars versus light trucks and
make). Clearly, both age and mileage matter in vehicle scrappage decisions, since used vehicle
prices depend on both and not only one or the other. Since other studies rely on age, Sallee et
al.'s estimation based on mileage provides a different perspective.
Estimating future fuel costs requires knowing how many more miles remain in a vehicle's
lifetime and their distribution over time. The approach is to estimate lifetime mileage
distributions separately for passenger cars and light trucks, by make. First, parameters
describing the relative intensity of use by vehicle class and make were estimated by regressing a
predicted odometer reading as a function of age against each vehicle's actual odometer reading
(c indexes passenger cars versus light trucks).
46
-------
-ca "cm^icm
(20)
The predicted odometer readings were obtained from a U.S. Department of Transportation
(DOT) report (U.S. DOT/NHTSA, 2006) that provides an equation for mileage by age. A value
of 0 < 1 indicates that vehicles of that make are used more than expected based on the NHTSA
function. Next the NHTSA functions that predict vehicle scrappage and annual mileage as a
function of age were transformed to be functions of odometer reading.11
Data for the study come from an enormous database of over 90 million used vehicle transactions
at wholesale auction houses from 1990 to 2009. Using truncated Vehicle Identification Numbers
(VEST) these transactions are matched to EPA fuel economy ratings using make, model, model
year, cylinder count, engine displacement, body type, transmission and trim style. The EPA
combined, adjusted fuel economy ratings (the estimates reported to the public) were used to
estimate fuel economy. Model years prior to 1978 were dropped from the sample since EPA
estimates are not generally available for these years. A 10% sample of the records was taken,
leaving still about 8 million cases. The price of gasoline was represented by the monthly
national retail price series of the Energy Information Administration (Figure 6), deflated using
the CPI-U. The undeflated price series shows a generally rising trend with considerable ups and
downs after 2001, with a sudden decline in 2009. Gasoline price varies only by month, while the
time-period fixed effects vary by month and vehicle type. What identifies the effect of cost, C;jt,
from the fixed effects is therefore the variation in odometer readings from vehicle to vehicle, the
differences in expected life implied thereby, and differences in fuel economy.
Nominal Monthly Gasoline Prices, All Grades: 8/90 to 4/09
450
_ 400
C
-------
The estimation results indicate that used vehicle purchasers come very close to fully
incorporating future fuel costs into the prices of used cars. The overall estimated beta coefficient
for the full sample is -0.788 with a standard error of 0.011, indicating that on average buyers
incorporate almost four fifths of the expected costs of future fuel expenditures into the price of
the vehicle. The null hypothesis that P = -1 could not be rejected, however. The reference
discount rate is 5%. If that is raised to 10%, the P coefficient for all vehicles becomes -1.01.
The study also found considerable variation in P values between passenger cars and light trucks
and across makes (Table 11). Purchasers of Ford passenger cars, for example, are estimated to
incorporate only about 50% of the expected future fuel costs, while purchasers of Toyota cars are
estimated to account for more than 90%.
Table 11. Coefficients on the Expected Present Value of Real Remaining Fuel Costs
Full Sample
-0.788
(0.011)
Makes
All
Chevrolet
ForH
Honda
Toyota
Passenger Cars
-0.851
(0.018)
-0.599
(0.045)
-0.513
(0.030)
-0.787
(0.034)
-0.914
(0.054)
Light Trucks
-0.736
(0.014)
-0.727
(0.033)
-0.735
(0.033)
-0.619
(0.048)
-0.792
(0.043)
There are several areas in which this analysis differs from others. Possibly a key difference is
the assumption that the expected remaining vehicle life will be determined by the odometer
reading rather than the age of the vehicle. Other studies assume that age will determine
remaining vehicle lifetime. Certainly both matter, since vehicles deteriorate with ageing of the
materials of which they are made, and their value also declines with technical obsolescence.
Another key difference is the use of variation in remaining vehicle use specific to individual
vehicles to help identify the valuation of future fuel costs. When the authors restricted their
estimation to impose the same odometer coefficients and the same time period effects (both vary
by vehicle type in the base model), they find that their beta estimates fall dramatically and are
closer to zero than one. A key question for future research is therefore to understand why
vehicle type-specific polynomial coefficients and fixed effects have such a pronounced impact
on the estimated willingness to pay. The authors also have not yet estimated a model with
asymmetric price effects. Kilian and Sims (2006) found a similar valuation of future fuel costs
during periods of rising prices (about -0.8) but no value during periods of falling prices. The
authors also intend to examine this in future work.
48
-------
4. SUMMARY AND DISCUSSION
Twenty-seven recent studies have been reviewed for evidence on the value consumers place on
fuel economy in their vehicle purchase decisions. The resulting estimates are highly variable.
They are summarized in Table 12 in terms of willingness to pay as a fraction of discounted
present value of fuel savings or in terms of implied discount rates. There are studies indicating
that consumers significantly undervalue fuel economy relative to its expected present value, and
there are others indicating that consumers value fuel economy at approximately its expected
present value, or much more. This wide disparity of results mirrors the highly variable estimates
of implicit consumer discount rates for future fuel savings found by Greene (1983) more than 25
years ago. The persistence of this lack of consensus over a period of decades, despite significant
improvements in data and advances in methodology, calls out for an explanation. First, the
evidence will be reviewed.
Among the studies using discrete choice modeling to draw inferences from aggregate vehicle
sales data, Allcott and Wozny concluded that consumers significantly undervalue future fuel
savings at the rate of about $0.25 on the expected present value dollar. Gramlich's (2008) results
contradict this finding. His results imply that consumers overvalue fuel savings by a factor of 2
to 3. However, Gramlich's model over-predicted the increase in light-duty vehicle fuel economy
from 2007 to 2008 as a result of higher fuel prices by an order of magnitude. Sawhill's (2008)
results also imply that consumers overvalue fuel economy but by a factor of 1.3 to 1.4, on
average. Berry, Levinsohn and Fakes' (1995) model found the miles per dollar was not a
significant factor in consumers' vehicle choices. Taking their estimated coefficient at face value
implies that consumers undervalue fuel economy, counting only about $0.01 on the dollar.
Among discrete choice models estimated using survey data, Train and Winston (2007) found that
fuel cost per mile was not significant on average, but that the variation across households in
preferences for fuel economy was significant. About 60 percent of the population preferred
higher fuel economy while 40% preferred lower fuel economy. Like BLP, fuel economy was
undervalued, on average, by about two orders of magnitude. Dasgupta, Siddarth and Silva-Risso
(2007) found only a slightly high implicit discount rate for fuel economy: 15.2% per annum.
Bento et al. (2005) did not directly estimate the value of fuel economy, but the very low
elasticities of vehicle choice with respect to fuel costs implied by their model indicate significant
undervaluation relative to vehicle price. Feng et al. (2005) found four out of six fuel cost per
mile coefficient were significant, and that these consumers valued fuel economy at
approximately its expected present value. Estimates made by Brownstone et al. (1996) and
Brownstone et al. (2000) using stated preference survey data in general suggest that households
overvalue fuel economy by 50% to 67%, but are highly variable across income groups and
household demographics, with certain groups preferring less fuel economy. Goldberg's (1995,
1998) estimates show statistical significance for small cars but not large cars. Overall she
concludes there is no reason to believe the consumers are myopic. Two studies from the UK
appear to show that car buyers there somewhat overvalue fuel economy, although there are
questions about the units key variables are in that remain to be resolved.
49
-------
Table 12. Summary of Consumers' Evaluation of Fuel Economy Improvements
Based on 27 Recent Studies
Authors
Alcott & Wozny
(2009)
Gramlich (2008)
Berry, Levinsohn &
Pakes (1995)
Sawhill (2008)
Train & Winston
(2007)
Dagupta, Siddarth and
Silva-Risso (2007)
Bento, Goulder,
Henry, Jacobsen &
von Haefen (2005)
Feng, Fullerton & Gan
(2005)
Klier and Linn (2008a)
Brownstone, Bunch &
Train (2000)
Brownstone, Bunch,
Golob& Ren (1996)
Goldberg (1996, 1998)
Goldberg (1995)
Vance & Mehlin
(2009)
Cambridge
Econometrics (2008)
Eftec (2008)
Fan & Rubin (2009)
Fifer & Bunn (2009)
McManus (2007)
Espey & Nair (2005)
Arguea, Hsiao &
Taylor (1994)
Bhat & Sen (2006)
Model Type
Mixed NMNL
NMNL
NMNL Ag
Mixed NMNL
Mixed NMNL
NMNLSur
NMNLSur
NMNL
Logit
Mixed NMNL
Stated &
Revealed
Preference
NMNL Stated &
Revealed
Preference
NMNL
NMNL
NMNL
Mixed logit
NMNL
Hedonic Price
Hedonic Price
Hedonic Price
Hedonic Price
Hedonic Price
Choice model
Data / Time
Aggregate U.S.,
1999-2008
Aggregate U.S.,
1971-2007
gregate US,
1971-1990
Aggregate U.S.,
1971-1990
Survey, U.S.,
2000
vey, CA,
1999-2000
vey, U.S.,
2001
CES, U.S., 1996-
2000
Aggregate U.S.,
1970-2007
CA Survey, 1993
CA Survey, 1993
U.S. CES, 1984-
1990
U.S. CES, 1983-
1987
Germany,
Aggregate New
Car Sales
UK survey, 2004
to 2009
UK 2001 to 2006
State of Maine,
2007
U.S., 1996-2005
U.S., 2002
U.S., 2001
U.S., 1969 to
1986
San Francisco
Bay Area, 2000
W-T-P as % of Discounted PV Implied Annual
Discount Rate
25% > 60%
287% to 823%
<1%
Non-significant
140%, range of
-360% to 1,410%
1.3%
Non-significant
15 .2%
No direct estimate but MPG
insensitive to price of gasoline
0.03% to 1.3%
Very approximately 69%
132% to 147%
-420% to 402%
Consumers "not myopic"
Approximately 1,000%
196% but uncertain of
estimate. Authors contacted
for clarifications.
TBD - authors contacted for
clarifications.
Cars: 25% Cars: 37%
Lt. Trucks: 16% Lt. Trucks: 77%
Cars: 52%, Pickups: 283%
SUVs: 44%, Vans: 240%
90%
109%
3% to 46%
Elasticities of vehicle choice
with respect to fuel costs 2%
to 3% of purchase price
elasticities.
50
-------
Sallee, West & Fan
(2010)
Langer & Miller
(2008)
Busse, Knittel &
Zettelmeyer (2009)
Kilian and Sims
(2006)
Li, Timmins & von
Haefen (2009)
Price Regression
Price Regression
Price Regression
Price Regression
Vehicle sales by
fuel economy
quantile
Aggregate U.S.,
Used Cars, 1978-
2009
U.S., 2003 to
2006
U.S., 1999 to
2008
Aggregate U.S.,
Used Cars, 1978-
1984
U.S. Metro
Areas 1997 to
2005
79%, not statistically different
from 100%
Approx. 1 5% of PV of fuel
cost changes reflected in
vehicle price changes.
Transaction prices adjust by
1.2 years worth of fuel savings
for new cars.
11% to 25%
Short-run price elasticity of
MPG with respect to sales mix
+0.02, long-run +0.2.
Hedonic price studies also disagree. Fan and Rubin estimate very high implicit discount rates for
consumers in the state of Maine: 37% for car buyers, 77% for purchasers of light trucks.
McManus' (2007) results indicate that consumers undervalue fuel economy by 30%, or less,
depending on assumptions. Espey and Nair (2005) find very low implied discount rates, 1% to
4% for most of their models. Arguea, Hsiao and Taylor (1994) produce estimates of willingness
to pay that range from 5% of expected present value to 46%, depending on the year.
Using other methods Bhat and Sen, 2006 found that consumers undervalue fuel economy by up
to an order of magnitude. Evidence from Langer and Miller (2008) and Busse, Knittel and
Zettelmeyer (2009) is more difficult to interpret but tends to indicate that consumers either
undervalue fuel savings or value it at close to its expected present value.
Key attributes of the 28 studies are summarized in Table 13. Less than half of the studies were
published in peer reviewed journals. Most of those in manuscript form are very recent (2008-
2010) and are likely to be published soon. Four are reports and one a thesis. There is a relatively
even distribution among the model types: 5 aggregate discrete choice, 12 disaggregate discrete
choice (only 9 are U.S. studies and three of these are slightly different versions of the same
study, leaving 7 different U.S. studies), 5 are hedonic price analyses, 3 are asset price models
with two using other modeling frameworks. Six studies use aggregate sales or sales shares as the
dependent variable, 11 use individuals' vehicle choices from surveys, 9 make new or used
vehicle prices the dependent variable, and two use sales by fuel economy quantile. There is a
similarly wide array of survey, aggregate sales, and vehicle price data sources. Some studies are
based on a single year's data, others make use of data covering decades. The time periods range
from the 1970s to 2009.
Fuel economy appears in a variety of forms, most commonly fuel cost per mile or fuel
consumption per mile. Only one study, however, tested the hypothesis that markets might
respond asymmetrically to rising and falling fuel prices. That study concluded that they do. This
subject is worthy of further attention, since there is evidence that markets in other areas respond
asymmetrically to petroleum price increases and decreases. In addition, it is possible that
consumers respond differently to fuel economy than they do to fuel prices. There is evidence
that vehicle travel responds differently to fuel costs than to miles per gallon (Small and van
51
-------
Table 13. Summary of Key Features of 27 Econometric Studies
Study
Berry, Levinsohn &
Pakes 1995
Allcott & Wozny 2009
Klier & Linn 2008
Gramlich 2008
Sawhill 2008
Publication
Status
Journal
Manuscript N]
Manuscript
Manuscript
Manuscript
Model
NMNL
/TNL
Logit
NMNL
NMNL
Dependent
Variable
Sales shaes
New & used vehicle
prices
New vehicle shares
New vehicle shares Ag
New vehicle shares
Type of
Data
Aggregate U.S.
Aggregate U.S.
Aggregate U.S.,
monthly
gregate U.S.
Aggregate U.S.
Time
Period
1971-1990
1999-2008
1970-2007
1971-2007
1971-1 990 Pg
Fuel
Economy
Measure
Miles/Pg
Disc. PV of Fuel
Cost
Disc. PV of Fuel
Cost
Pg/MPG & MPG
/MPG
Price
Expectations
Random
Walk
RW +
alternatives
Random
Walk
Random
Walk
ARIMA
Transaction
Prices?
No Yes
Yes No
n.a. Yes
No No
No
Heterogeneour
Tastes?
Yes
Simultaneous
Supply &
Demand
Yes
Yes
No
Yes
Yes
Fuel
Economy
Standards
Included?
No
n.a.
No
Yes
No
MPG
Value*
0
+
+
Tram & Winston 2007
Dasgupta, Siddarth &
Silva-Risso 2007
Bento, Goulder,
Jacobsen & von
Haefen 2005 & 2008
Feng, Fullerton & Gan
2005
Brownstone, Bunch &
Tram 2000
Brownstone, Bunch,
Golob& Ren 1996
Goldberg 1995
Goldberg 1996
Goldberg 1998
Journal
Journal
Journal
Manuscript N]
Journal
Journal
Journal
Report
Journal
Mixed Logit
Mixed Logit
Random
Coef. Logit
/TNL
Mixed Logit
NMNL
NMNL
NMNL
NMNL
Indiv. Vehicle Choices
Indiv. Vehicle Choices
Indiv. Vehicle Choices
Indiv. Vehicle Choices Cl
Indiv. Vehicle Choices
Indiv. Vehicle Choices
Indiv. Vehicle Choices C]
Indiv. Vehicle Choices Cl
Indiv. Vehicle Choices Cl
U.S. Household
Survey
So. CA Vehicle
Transactions
Nat. HH. Travel
Survey U.S.
IS
CA survey
CA survey
IS
;s
!S
2000 1
1 999-2000 Pg
2001 Pg
1996-2000
1993
1993
1983-1987
1985-1990
1984-1990
/MPG
/MPG
/MPG
Pg/MPG
Pg/MPG
Pg/MPG
Pg/MPG
Pg/MPG
Pg/MPG
Static
Static
Static
Static
Static
Static
Static
Static
Static
No
Yes
No
No
Yes, via
respondents
Yes, via
respondents
No
No
No
Yes
Yes
Yes
No
Yes No
No No
Yes
Yes
Yes
Yes
No
No
Yes
Yes
Yes
Yes
n.a.
n.a.
n.a.
n.a.
n.a.
n.a.
Yes
Yes
Yes
0
X
+
0
0
0
0
Cambridge
Econometrics 2008
eftec 2008
Vance & Mehlm 2009
Report
Report
Report
Mixed Logit
Mixed Logit
NMNL
Indiv. Vehicle Choices
Indiv. Vehicle Choices
New Vehicle Shares
UK Survey
UK Survey
Sales data,
Germany
2005-2006
2004 & 2007
1 995-2007
£/100km
1/1 00 km &
£/100km
/100km
Static
Static No
Static
No
No
Yes
Yes
No
No
No
No
No
No
No
+
+
Fan & Rubin 20 10
Espey & Nair 2005
McManus 2007
Fifer & Bunn 2009
Arguera, Hsaio &
Taylor 1994
Manuscript
Journal
Journal
Thesis
Journal
2 -stage
hedonic
1 -stage
hedonic
1 -stage
hedonic
1 -stage
hedonic
2-stage
hedonic
New vehicle prices
New vehicle prices
New vehicle prices
New vehicle prices
New vehicle prices
Maine, sales
data
U.S. vehicle
data
U.S. vehicle
data
U.S. vehicle
data
U.S. vehicle
data
20071
2001 1
2002-2005 Pg
1996-2005 1
1969-1986 M
og(MPG)
/MPG
/MPG
/MPG
'G
Static
Static
Static
Random
Walk
n.a.
No
No
Yes
No Yes
No
Yes
No
No
No
Yes
No
No
No
Yes
No
No
No
No
No
0
0
Killian & Sims 2006
Manuscript
Asset Price
Used Car Prices
U.S.
1978-1984
PV fuel costs
Random
Walk +
No No
No
No
52
-------
TablelS. Summary of Key Features of 27 Econometric Studies (continued)
Study
Bhat & Sen 2005
Langer & Miller 2008
Busse, Knittel &
Zettelmeyer 2009
Li, Timmins & von
Haefen 2009
Bailee, West & Fan
2010
Publication
Status
Journal
Manuscript A'.
Manuscript
Journal
Manuscript A;
Model
MDCEV
random
utility
set Price
Sales shares
by quartile
Vehicle
set Price
Dependent
Variable
Individual vehicle
choice
New vehicle prices
New & Used vehicle
prices
Sales by Quantile
Used vehicle prices
Type of
Data
Survey: San
Francisco, CA
U.S. regions
weekly
Sample, U.S.
transactions
New & used
sales in 20 U.S.
metro areas
Sample of U.S.
auction
transactions
Time
Period
2000 Pg
2003-2006 Pg
1 999-2008 M
1997-2005
1990-2009
Fuel
Economy
Measure
/MPG
/MPG
>G Quartiles
Pg/MPG & P8
Disc. PV of fuel
Price
Expectations
Static
Random
Walk +
RW
Random
Walk
Random
Walk +
Transaction
Prices?
Price not
included
No Yes
Yes
No Yes
Yes Yes
Heterogeneour
Tastes?
Yes No
Yes
Simultaneous
Supply &
Demand
No
No
No
No
Fuel
Economy
Standards
Included?
No
No
No
No
No
MPG
Value*
X
0
* Indicates whether study generally implies that consumers undervalue (), over-value (+) or equally value (0) fuel economy, or none of the
above (X).
53
-------
Dender, 2007; Greene 2010) and it would be interesting to see if vehicle purchases do also. The
premise that consumers respond equivalently to a change in the price of fuel or a change in fuel
use per mile is a logical consequence of the rational economic model, yet it may not actually
reflect how consumers make decisions. It is a maintained hypothesis. If it is incorrect, then
inferences about how consumers value fuel economy could be largely a reflection of a different
kind of behavior: how consumers respond to fuel prices.
Although a variety of fuel price expectation models are employed, there is a clear preference for
static expectations or random walk, both of which imply that current prices are the best predictor
of future prices. Undoubtedly the fact that, historically, oil and gasoline prices are consistent
with the random walk model (e.g., Hamilton, 2009) influenced analysts' decisions. Other
models tested were models that implied that shocks would die away over time. While these
models are reasonable from a statistical viewpoint, they may not match consumers' expectations.
A possibly useful area for future analysis would be to explore models that allowed consumers to
project trends into the future, or to base their expectations on learning from past price excursions.
Most studies did not have access to actual transaction prices and relied on manufacturer's
suggested retail prices or other standard values. Very few studies attempted to explicitly
estimate the effects of fuel economy standards on the market for fuel economy. In some cases,
such as discrete choice models for a single year, this should not be a problem. In other cases,
such as hedonic price models based on a times series, this could make it very difficult to identify
the demand function. Many studies allowed preferences to vary across individual consumers or
market segments. Methods of incorporating heterogeneity in preferences varied considerably,
however. Some models estimate distributions of preference coefficients, others use dummy
variables or interact consumer attributes with vehicle attributes. Many of the models use
estimation methods appropriate for simultaneous determination of attributes (such as fuel
economy) and their values. This is not necessary for all model formulations, however, and those
not using simultaneous equation methods generally explain the assumptions that make it
unnecessary (e.g., approximately fixed supply and attributes of used vehicles).
In the final analysis, however, none of these factors appears to be able to explain the different
conclusions reached by the 27 studies. Given the variety of models, data and estimation
methods, combined with the generally high level of technical expertise demonstrated by the
researchers, one is led to look elsewhere for an explanation. In particular, fruitful areas to
investigate appear to be the appropriateness of the classical rational economic decision making
model for explaining consumers' fuel economy choices, and the difficulties of making statistical
inferences given the complexity of the vehicle purchase decision.
Although a completely satisfactory explanation for these widely diverging results does not exist
at present, several factors can be identified that clearly play a role in creating the confusion.
First, and perhaps most importantly, Turrentine and Kurani's work has shown that real
consumers almost certainly do not make their decisions according to the strict model of rational
economic behavior. Instead, they found that households employed numerous, different decision
rules. Many of the economic models reviewed in this report allow for heterogeneity of tastes;
Turrentine and Kurani's work implies heterogeneity of decision rules, as well. This could be
what causes estimates in mixed logit models to indicate no significant value of fuel economy on
average, but significant variance across individuals. Certainly, the mere fact that vehicle use
54
-------
varies considerably across households would dictate a certain degree of variability in the value
given to fuel economy. Heterogeneity of decision rules further complicates the question.
Consumers may not be continuously trading off the multiple attributes of motor vehicles.
Estimating models that imply that they are could lead to a misspecification bias.
Second, the choice of a motor vehicle is itself a complex problem. Vehicles have multiple
attributes, many of which are difficult to define, let alone measure. Modelers cannot be expected
to include all variables that are relevant, nor to measure them precisely in all cases. Style,
reliability, prestige, and safety are examples of important vehicle attributes that themselves have
many dimensions and are difficult to define and measure. These and other omitted or
imprecisely measured attributes are likely to be correlated with fuel economy to a greater or
lesser extent. The combination of omitted variables, errors in variables and correlated variables
can easily lead to seriously biased coefficients. Most modelers do their best to deal with the
problem with liberal use of fixed effects. This should help a great deal but may not be enough to
prevent serious bias in the estimates of certain coefficients.
The literature on the economics of energy efficiency has only recently begun to appreciate that
the consumer faces substantial uncertainty in making choices about the energy efficiency of
energy-using durable goods. Uncertainty about future energy prices is the most obvious but may
not be the most important source of uncertainty. Although vehicles are clearly labeled with fuel
economy ratings, the credibility of the miles per gallon estimates is seriously questioned by
many consumers. Moreover, as the label itself states, "your mileage may vary." Driving style,
traffic conditions, trip lengths and even climate have significant impacts on realized fuel
economy. There are also non-trivial uncertainties about vehicle use, vehicle lifetime, and the
cost of increased fuel economy either in dollars or foregone other attributes. Uncertainty further
complicates both consumers' decision making and quantitative inferences about it.
Vehicle fuel economy is primarily determined by two different processes. There is the
consumer's choice among a set of vehicles with fixed attributes, and then there is the
manufacturer's decision about how to design vehicles and which technologies to use. These are
very different processes. Some studies are based solely on consumers' choices, while others
attempt to combine the two. This in itself can lead to different inferences. Potentially even more
important is the possibility that manufacturers' decision making and consumers' decision making
about fuel economy may not be as perfectly synchronized as the rational economic model
implies.
Key findings of this literature review can be summarized as follows:
1. Of the 25 distinctly different study estimates, 12 indicate that consumers significantly
undervalue future fuel savings relative to a reference expected value based on average
U.S. statistics, 8 indicate that consumers' values are approximately equal to the reference
expected value, and 5 indicate that consumers significantly overvalue fuel savings.
2. With a very few exceptions, there are no obvious flaws in the methods or data used by
these studies. This finding applies equally to the published and unpublished studies.
3. There does not appear to be an obvious explanation for the widely divergent results.
Neither model type, formulation of the variable representing fuel economy, data type,
55
-------
time period, nor any other readily identifiable factor shows a strong association with
inferences about the values consumers place on fuel economy (Table 13).
4. The fifteen studies using discrete choice models are evenly divided between under, equal
and over-valuing fuel economy; studies using hedonic price, asset price and other models
more often indicate undervaluing (Figure 7).
5. The studies are also evenly divided between those dated 2008 or later (12) and those
between 1994 and 2007. Six of the earlier studies and six of the 2008-2010 studies
conclude that consumers significantly undervalue fuel economy. Seven of the earlier
studies and six of the later studies imply that consumers roughly equally value or
significantly over value fuel economy (Figure 8).
6. Price expectations are an important factor in all studies. Almost all of the studies assume
that consumers will use the current price of fuel as a best estimate of future fuel prices,
either due to static expectations or because they perceive fuel prices will follow a random
walk. Five of the studies explore alternative price expectations models. However, none
of the models allows consumers to project trends of increasing or decreasing prices into
the future. Given the importance of price expectations to the evaluation of future fuel
savings, a better understanding of how consumers form price expectations might provide
useful insights.
7. Most of the studies (15) represent fuel economy interacted with the price of fuel,
generally as fuel cost per mile. These studies are evenly divided between undervaluing
(5), equally valuing (5) and overvaluing (5) (Figure 9). Six included fuel economy
without interaction with the price of fuel, either as miles per gallon or gallons per mile.
Of these, five found undervaluing and one equally valuing. Four calculated a discounted
present value based on assumptions about vehicle lifetime, usage and fuel price
expectations. These studies were evenly divided between undervaluing and
approximately equally valuing. These differences suggest that there may be some
insights to be gained by testing hypotheses about whether consumers respond differently
to fuel price changes as opposed to fuel economy differences, or whether responses to the
two variables are symmetric or asymmetric.
8. Several studies point out the empirical challenges to inferring the value of vehicle
attributes to consumers: (1) correlations among attributes, (2) difficulties in defining and
measuring the many relevant attributes, and (3) differences (heterogeneity) in tastes
among consumers. These problems can lead to errors in variables and omitted variables
and, together with correlations among variables they can result in seriously unstable,
biased parameter estimates. More recent studies, exploiting massive data sets, have
attempted to address these problems with detailed fixed effect coefficients and other
methods. The continued differences in results suggest that even these efforts may not
have successfully addressed the empirical challenges.
Fuel economy or CC>2 emissions standards are a core component of governments' policy
strategies to address global climate change and energy security. Standards have been adopted by
the United States, the European Union, Japan and China, among others. The annual costs and
benefits of these standards easily amount to tens of billions of dollars. How consumers' value
future fuel savings in making car buying decisions has been shown to be a crucial determinant of
the economic consequences of such standards (Fischer et al., 2007). Yet surprisingly little is
known about this vitally important subject.
56
-------
Given the importance of understanding how the market values fuel economy and makes
decisions about it, it might be worthwhile to convene top researchers with differing results to
jointly investigate why those results differ so greatly. Such an effort would require sharing of
data sets among researchers, who would then execute a mutually agreed upon set of statistical
analyses, (1) to validate the results produced by others, and (2) to test a specified set of
alternative model formulations using the different data sets. Such a structured test of model
formulations against alternative data sets might lead to important insights about why apparently
carefully and competently done analyses can lead to widely differing results.
It is at least as important to investigate the possibility that it is the rational economic consumer
model that is incorrect. This line of inquiry might best be pursued through in two steps. First,
conduct more in-depth interviews, surveys and experiments, such as reported in the seminal
paper by Turrentine and Kurani (2007), to discover what decision criteria and algorithms real
consumers actually employ when considering fuel economy and valuing fuel savings. Second,
test these alternative models using experimental methods and empirical market data.
Distribution of 25 Distinct Studies by Model Type and Value
of Fuel Economy Relative to Reference
6
5
u
QJ
Under Equal
Relative Value of Fuel Economy
Over
Figure 7. Distribution of 25 Distinct Studies by Model Type and Value of Fuel Economy
Relative to the Reference Value.
57
-------
Distribution of 25 Distinct Studies by Date of Study and
Value of Fuel Economy Relative to Reference
01
(U
Under Equal Over
Relative Value of Fuel Economy
Figure 8. Distribution of 25 Distinct Studies by Date of Study and Value of Fuel Economy
Relative to Reference Value.
Distribution of 25 Distinct Studies by Form of Fuel Economy
Variable and Value of Fuel Economy Relative to Reference
Dw Price interaction Uw/o Price Disc. PV
Under Equal Over
Relative Value of Fuel Economy
Figure 9. Distribution of 25 Distinct Studies by Form of Fuel Economy Variable and Value of
Fuel Economy Relative to Reference Value.
58
-------
REFERENCES
1. (ACEEE) American Council for an Energy Efficient Economy, 2007. "Quantifying the
Effects of Market Failures in the End-Use of Energy," Final draft report prepared for the
International Energy Agency, by ACEEE, Washington, DC, February.
2. Allcott, H. and N. Wozny, 2009. Gasoline Prices, Fuel Economy, and the Energy
Paradox, unpublished manuscript, MIT Department of Economics, Cambridge,
Massachusetts, November 16.
3. Arguea, N.M., C. Hsiao and G.A. Taylor, 1994. "Estimating Consumer Preferences
Using Market Data - An Application to U.S. Automobile Demand," Journal of Applied
Econometrics, vol. 9, no. 1, pp. 1-18.
4. Austin, D. and T. Dinan, 2005. "Clearing the Air: The Costs and Consequences of Higher
CAFE Standards and Increased Gasoline Taxes," Journal of Environmental Economics
and Management, vol. 50, pp. 562-582.
5. Bento, A.M., L.H. Goulder, E. Henry, M.R. Jacobsen and R.H. von Haefen, 2005.
"Distributional and Efficiency Impacts of Gasoline Taxes: An Econometrically Based
Multi-market Study," AEA Papers and Proceedings, vol. 95, no. 2, pp. 282-287.
6. Berry, S., J. Levinsohn and A. Pakes, 1995. "Automobile Prices in Market Equilibrium,"
Econometrica, vol. 63, no. 4, pp. 841-890.
7. Bhat, C.R. and S. Sen, 2006. "Household Vehicle Type Holdings and Usage: An
Application of the Multiple Discrete-continuous Extreme Value (MDCEV) Model,"
Transportation Research Part B, vol. 40, pp. 35-53.
8. Brownstone, D., D.S. Bunch and K. Train, 2000. "Joint Mixed Logit Models of Stated
and Revealed Preferences for Alternative-fuel Vehicles," Transportation Research Part
B, vol. 34, pp. 315-338.
9. Brownstone, D., D. S. Bunch, T.F. Golob and W. Ren, 1996. "A Transactions Choice
Model for Forecasting Demand for Alternative-Fuel Vehicles", Research in
Transportation Economics, vol. 4, pp. 87-129.
10. Busse, M.R., C.R. Knittel and F. Zettelmeyer, 2009. The Differential Effect of Usage
Cost on New and Used Durable Goods Markets: Evidence from the Automobile Industry,
manuscript, Northwestern University and NBER, September 2009, m-
busse@kellogg.northwestern.edu.
11. Cambridge Econometrics, 2008. Demand for Cars and their Attributes, a final report for
the Department for Transport, January 23, Cambridge, UK.
12. Dasgupta, S., S. Siddarth and J. Silva-Risso, 2007. "To Lease or By? A Structural Model
of a Consumer's Vehicle and Contract Choice Decisions," Journal of Marketing
Research, vol. XLIV, pp. 490-502.
13. Davis, S.C., S.W. Diegel and R.G. Boundy, 2009. Transportation Energy Data Book,
28th Edition, ORNL-6984, Oak Ridge National Laboratory, Oak Ridge, Tennessee.
14. Delucchi, M.A., 2007. Cost-Benefit Analysis of Fuel-Economy Improvement,"
Discussion Paper, Institute for Transportation Studies, University of California at Davis,
Davis, California, August.
15. (Eftec) Economics for the Environment Consultancy Ltd., 2008. Demand for Cars and
their Attributes, report prepared for the Department of Transport, UK, London, January.
16. Espey, M. and S. Nair, 2005. "Automobile Fuel Economy: What is it Worth?"
Contemporary Economic Policy, vol. 23, no. 3, July, pp. 317-323.
59
-------
17. Fan, Q. and J. Rubin, 2009. "Two-Stage Hedonic Price Model for Light-Duty Vehicles:
Consumers' Valuations of Automotive Fuel Economy in Maine," 89* Annual Meeting of
the Transportation Research Board, Washington, D.C., January, 2010.
18. Feng, Y., D. Fullerton and L. Gan, 2005. Vehicle Choices, Miles Driven and Pollution
Policies, Working Paper 11553, National Bureau of Economic Research, Cambridge,
Massachusetts, August.
19. Fifer, D.P.C. and N.P. Bunn, 2009. "Assessing Consumer Valuation of Fuel Economy in
Auto Markets", Honors Thesis, Department of Economics, Duke University, Durham,
North Carolina.
20. Fischer, C., W. Harrington and I.W.H. Parry, 2007. Should Automobile Fuel Economy
Standards Be Tightened? RFF Discussion Paper 04-53, Resources for the Future,
Washington, DC, February 2007.
21. Gerard, D. and L. Lave, 2003. The Economics of CAFE Reconsidered: A Response to
CAFE Critics and A Case for Fuel Economy Standards, Regulatory Analysis 03-10, AEI-
Brookings Joint Center for Regulatory Studies, Washington, DC, September.
22. Goldberg, P.K., 1998. "The Effects of the Corporate Average Fuel Efficiency Standards
in the U.S.," The Journal of Industrial Economics, vol. XL VI, no. 1, pp. 1-33.
23. Goldbert, P.K., 1996. "The Effects of the Corporate Average Fuel Efficiency Standards,"
Working Paper 5673, NBER, Cambridge, Massachusetts, July.
24. Goldberg, P.K., 1995. "Product Differentiation and Oligopoly in International Markets:
the Case of the U.S. Automobile Industry," Econometrica, vol. 63, no. 4, pp. 891-951.
25. Gramlich, J., 2008. Gas Prices and Endogenous Produce Selection in the U.S.
Automobile Industry, manuscript, Department of Economics, Yale University, New
Haven, Connecticut, November, 20, 2008.
26. Greene, D.L., 2010. "Rebound 2007: Analysis of U.S. Light-Duty Vehicle Travel
Statistics," forthcoming, Energy Policy.
27. Greene, D.L., 2009. "Uncertainty, Loss Aversion and Markets for Energy Efficiency,"
forthcoming, Energy Economics.
28. Greene, D.L., 1990. "CAFE or PRICE? An Analysis of the Effects of Federal Fuel
Economy Regulations and Gasoline Price on New Car MPG, 1978-89," The Energy
Journal, vol. 11, no. 3, pp. 37-57, September.
29. Greene, D.L., 1983. "A Note on Implicit Consumer Discounting of Automobile Fuel
Economy: Reviewing the Available Evidence," Transportation Research B, vol. 17B, no.
6, pp. 491-499.
30. Greene, D.L., J. German and M.A. Delucchi, 2009. "Fuel Economy: The Case for Market
Failure," in D. Sperling and J.S. Cannon, eds., Reducing Climate Impacts in the
Transportation Sector, Springer Science + Business Media.
31. Hamilton, J.D., 2009. "Understanding Crude Oil Prices," The Energy Journal, vol. 30,
no. 2, pp. 179-206.
32. Hassett, K.A. and G.E. Metcalf, 1993. "Energy Conservation Investment: Do Consumers
Discount the Future Correctly," Energy Policy, vol. 21. no. 6, pp. 710-716.
33. Hausman, J., 1979. "Individual Discount Rates and the Purchase and Utilization of
Energy -using Durables," Bell Journal of Economics, vol. 10, no. 1, pp. 33-54.
34. Helfand, G. and A. Wolverton, 2009. Evaluating the Consumer Response to Fuel
Economy: A Review of the Literature, Working Paper #09-04, National Center for
Environmental Economics, U.S. Environmental Protection Agency, Washington, DC.
60
-------
35. Howarth, R.B. and H.H. Sanstad, 1995. "Discount Rates and Energy Efficiency,"
Contemporary Economic Policy, vol. 13, no. 3, pp. 101-109.
36. Kilian, L. and E.R. Sims, 2006. "The Effects of Real Gasoline Prices on Automobile
Demand: A Structural Analysis Using Micro Data", Michigan Department of Economics,
Working Paper, University of Michigan, Ann Arbor, April 18.
37. Klier, T. and J. Linn, 2008. "The Price of Gasoline and the Demand for Fuel Efficiency:
Evidence from Monthly New Vehicle Sales Data", manuscript, Federal Reserve Bank of
Chicago, September.
38. Kleit, A.N., 1990. "The Effect of Annual Changes in Automobile Fuel Economy
Standards", Journal of Regulatory Economics, vol. 2, pp. 151-172.
39. Knittel, C.R., 2009. "Automobiles on Steroids: Product Attribute Trade-offs and
Technological Progress in the Automobile Sector", Department of Economics, University
of California at Davis, Davis, California, June 3, 2009.
40. Knittel, C.R. and K. Metaxoglou, 2008. Estimation of Random Coefficient Demand
Models: Challenges, Difficulties and Warnings, Working paper 14080, National Bureau
of Economic Research, Cambridge, Massachusetts, June.
41. Langer, A. andN. Miller, 2008. Automobile Prices, Gasoline Prices, and Consumer
Demand for Fuel Economy, Economic Analysis Group Discussion Paper EAG 08-11,
Department of Economics, University of California at Berkeley, Berkeley, California,
December.
42. Larrick, R.P. and J.B. Soil, 2008. "The MPG Illusion", Science, vol. 320, pp. 1593-1594.
43. Li, S., C. Timmins, R.G. von Haefen, 2009. "How Do Gasoline Prices Affect Fleet Fuel
Economy?" American Economic Journal: Economic Policy 2009, 1:2, pp. 113-137.
44. Mahadi, A. and K.S. Gallagher, 2009. "Fuel Prices and Consumer Preferences for
Vehicles," Harvard Kennedy School of Government, paper to be presented at the 89th
Annual Meeting of the Transportation Research Board, Washington, DC, January 2010.
45. Manski, C.F. and L. Sherman, 1980. "An Empirical Analysis of Household Choice
among Motor Vehicles," Transportation Research A, vol. 14A, pp. 349-366.
46. McManus, W., 2007. "The Link between Gasoline Prices and Vehicle Sales," Business
Economics, January, 2007, pp. 53-60.
47. (NRC) National Research Council, 2002. Effectiveness and Impact of Corporate
Average Fuel Economy (CAFE) Standards, National Academies Press, Washington, DC.
48. Rogozhin, A., M. Gallaher and W. McManus, 2009. Automobile Industry Retail Price
Equivalent and Indirect Cost Multipliers, EPA-420-R-09-003, U.S. Environmental
Protection Agency, Office of Transportation and Air Quality, Ann Arbor, Michigan,
February.
49. Rosen, S., 1974. "Hedonic Prices and Implicit Markets: Product Differentiation in Pure
Competition", Journal of Political Economy, vol. 82, pp. 34-55.
50. Rubenstein, A., 1998. Modeling Bounded Rationality, MIT Press, Cambridge,
Massachusetts.
51. Sallee, J.M., S. E. West and W. Fan, 2010. "The Effect of Gasoline Prices on the Demand
for Fuel Economy in Used Vehicles: Empirical Evidence and Policy Implications", Draft
version 1.1 of January 19, 2010, cited with permission of S.E. West, Macalester College.
52. Sawhill, J.W., 2008. Are Capital and Operating Costs Weighted Equally in Durable
Goods Purchases? Evidence from the U.S. Automobile Market, discussion paper,
Department of Economics, University of California at Berkeley, Berkeley, California,
April.
61
-------
53. Small, K. and K. Van Dender, 2007. "Fuel Efficiency and Motor Vehicle Travel: The
Declining Rebound Effect," The Energy Journal, vol. 28, no. 1, pp. 25-51.
54. Stern, P.C. and E. Aronson, eds., 1984. Energy Use The Human Dimension, Committee
on Behavioral and Social Aspects of Energy Consumption and Production, National
Research Council, W.H. Freeman and Company, New York.
55. Train, K., 1985. "Discount Rates in Consumers' Energy-related Decisions: A Review of
the Literature," Energy, vol. 10, no. 12, pp. 1243-1253.
56. Train, K., 1983. "Comment on "A Note on Implicit Consumer Discounting of
Automobile Fuel Economy: Reviewing the Available Evidence," Transportation
Research B, vol. 17B, no. 6, p. 498.
57. Train, K.E. and C. Winston, 2007. "Vehicle Choice Behavior and the Declining Market
Share of U.S. Automakers," International Economic Review, vol. 48, no. 4, pp. 1469-
1496.
58. Turrentine, T.S. and K.S. Kurani, 2007. "Car Buyers and Fuel Economy," Energy Policy,
vol. 35, pp. 1213-1223.
59. Tversky, A. and D. Kahneman, 1992. "Advances in Prospect Theory: Cumulative
Representation of Uncertainty," Journal of Risk and Uncertainty, vol. 5, pp. 297-323.
60. (U.S. DOE/EIA) U.S. Department of Energy, Energy Information Administration, 2009.
Annual Energy Review 2008, DOE/EIA-03 84(2008), Washington, DC.
61. (U.S. DOE/EIA) U.S. Department of Energy, Energy Information Administration, 1996.
Issues in Midterm Analysis and Forecasting 1996, DOE/EIA-0607(96), Washington, DC.
62. (U.S. DOT/NHTSA) U.S. Department of Transportation, National Highway Traffic
Safety Administration, 2009. Summary of Fuel Economy Performance, March 30, 2009,
available on the internet at: http://www.nhtsa.dot.gov/portal/fueleconomy.jsp .
63. (U.S. DOT/NHTSA) U.S. Department of Transportation, National Highway Traffic
Safety Administration, 2006. Vehicle Survivability and Travel Mileage Schedules, DOT
HS 809 952, Washington, D.C.
64. (U.S. EPA) U.S. Environmental Protection Agency, 2009. Light-Duty Automotive
Technology, Carbon Dioxide Emissions, and Fuel Economy Trends: 1975 through 2009,
EPA420-R-09-014, Ann Arbor, Michigan, November.
65. Vance, C. and M. Mehlin, 2009. "Tax Policy and CO2 Emissions: An Econometric
Analysis of the German Automobile Market", Ruhr Economic Papers #89, RWI Essen,
Germany, February.
62
-------
APPENDIX
REFERENCE ASSUMPTIONS ABOUT VEHICLE USE,
LIFETIME AND DISCOUNTING
Most econometric studies that estimate the value of fuel economy in consumers' automobile
choices represent vehicle price in dollars and fuel economy either as fuel cost per mile (dollars or
cents per mile) or gallons of fuel per mile. To compare estimates from different studies on a
common basis, it is necessary to have a standard set of assumptions about annual vehicle use, life
expectancy and the discount rate. This allows one to compute expected discounted lifetime
miles. Multiplying expected discounted lifetime miles times fuel cost per mile yields the
expected discounted present value of future fuel savings.
The calculations used throughout this report rely on data produced by the National Highway
Traffic Safety Administration (NHTSA, 2006). NHTSA's data, shown in Table A.I, are
estimates of the average annual miles of use for passenger cars and light trucks by age. The
same NHTSA document also provides estimated survival probabilities for the two vehicle types
as a function of age. These were used to determine expected vehicle life. The median expected
lifetime of a passenger car in the United States was estimated at just over 13 years for passenger
cars and just over 14 years for light trucks. In the calculations below, we assume an average
expected lifetime of 14 years for both vehicle types.
A real discount rate of 7% per annum is used throughout. The following discounting formula
was applied to produce the estimates in Table A. 1.
14
Expected Discounted Lifetime Miles =
(Al)
The discounted lifetime miles are 112,600 for passenger cars and 125,891 for light trucks.
When fuel economy is represented as gallons per mile in a model, it is also necessary to assume
a price of fuel. In all cases a constant fuel price has been assumed. Data for the year appropriate
for the study in question are obtained from the Energy Information Administration's Annual
Energy Review. That source provides prices in both current and constant dollars which are used
as appropriate to the study.
63
-------
Table Al. Annual Miles and Discounted Miles for Light-Duty Vehicles
Vehicle Age Discounting
(Years) Factor
1 0.96674
2 0.90349
3 0.84439
40.78914
5 0.73752
6 0.68927
70.64418
8 0.60203
9 0.56265
100.52584
11 0.49144
12 0.45929
13 0.42924
140.40116
Total
Passenger
Car Miles
14,231
13,961
13,669
13,357
13,028
12,683
12,325
11,956
11,578
11,193
10,804
10,413
10,022
9,633
168,853
Passenger
Car Disc.
Miles
13,758
12,614
11,542
10,541
9,608
8,742
7,939
7,198
6,514
5,886
5,310
4,783
4,302
3,864
112,600
Light Truck
Miles
16,085
15,782
15,442
15,069
14,667
14,239
13,790
13,323
12,844
12,356
11,863
11,369
10,879
10,396
188,104
Light Truck
Disc. Miles
15,550
14,259
13,039
11,892
10,817
9,815
8,883
8,021
7,227
6,497
5,830
5,222
4,670
4,170
125,891
Source: U.S. DOT/NHTSA (2006, tables 9 & 10).
64
------- |