Testing a Model of Consumer Vehicle
Purchases
Draft
&EPA
United States
Environmental Protection
Agency
-------
Testing a Model of Consumer Vehicle
Purchases
Draft
Gloria Helfand*3
Changzheng Liub
Marie Donahue0
Jacqueline Doremusc
Ari Kahana
Michael Shelby3
^Corresponding author:
'Office of Transportation and Air Quality
U.S. Environmental Protection Agency
2000 Traverwood Drive
Ann Arbor, Ml 48105 US
helfand.gloria@epa.gov
bOak Ridge National Laboratory
PO Box 2008 MS6472
Oak Ridge, TN 37831-6472 US
C0ak Ridge Institute for Sicience and Education Program
Office of Transportation and Air Quality
U.S. Environmental Protection Agency
2000 Traverwood Drive
Ann Arbor, Ml 48105 US
&EPA
United States
Environmental Protection
Agency
EPA-420-D-15-011
December 2015
-------
TESTING A MODEL OF CONSUMER VEHICLE PURCHASES
ABSTRACT
Consumer vehicle choice models have been estimated and used for a wide variety of policy
simulations. Infrequently, though, have predicted responses from these models been tested
against actual market outcomes. This paper presents a validation exercise for a model developed
for the U.S. Environmental Protection Agency, intended to estimate the impacts of changes in
vehicle prices and fuel economy due to changes in vehicle greenhouse gas emissions standards.
The model is a nested logit with a representative consumer and 5 levels, calibrated to vehicle
purchases in model year (MY) 2008. First, we review the model's response to a simple policy
scenario, to explore effects of different parameter values on the outcomes of that scenario; we
find that the model is not particularly sensitive to key parameters. Next, vehicle changes
between MY 2008 and 2010 are used to make predictions, and those predictions are compared to
actual outcomes in MY 2010; the model, designed to examine changes due only to price and fuel
economy, did not do as well in predicting sales impacts as assuming that market shares were the
same as in MY 2008, during this time of significant economic change. These exercises raise
questions about how to validate a model intended for comparative static analysis in a dynamic
world.
Keywords: vehicle demand; consumer vehicle choice modeling; validation; discrete choice
modeling
-------
Testing a Model of Consumer Vehicle Purchases
How well can a model predict which cars people will buy? Modeling purchase patterns
of consumer vehicles matters because of the importance of the auto sector to the U.S. economy,
and because of the contributions of vehicles to air pollution, including greenhouse gas (GHG)
emissions. Automakers have clear incentives to estimate consumer vehicle desires well,
especially when it may be difficult to change production plans quickly in response to market
signals. Many public policies, such as the federal Car Allowance Rebate System ("Cash for
Clunkers") in 2009 or California's Zero Emissions Vehicle Program, have explicit goals to affect
what vehicles people buy. Other policies, such as greenhouse gas or fuel economy standards,
may indirectly affect how many and which vehicles people buy. Measuring the effects of these
programs on vehicle sales would provide greater insights into the impacts of these programs on
the environment, auto producers, and the public.
Many researchers have developed models designed to estimate vehicle purchases. The
models, commonly econometrically estimated, are often used for prospective simulation
purposes. Almost unstudied, though, is the effectiveness of these models in predicting market
responses to changed circumstances. Researchers have rarely used their models to examine
situations where model results could be compared to actual market outcomes.
It seems evident that the utility of these models for estimating policy impacts should
depend on their effectiveness in predicting those impacts. It is also to be expected that the
success of a model depends on the purposes for which it was designed: a model well designed
-------
for examining the effects of GHG standards, for instance, may not perform well in estimating the
effects of demographic shifts on buying patterns.
The U.S. Environmental Protection Agency (EPA) has been exploring the use of vehicle
choice models in analyzing the impacts of vehicle GHG/fuel economy regulations, and recently
commissioned a consumer vehicle choice model for potential use in its analysis of the impacts of
vehicle greenhouse gas regulations on the U.S. auto market (Greene and Liu 2012). This paper
presents results of a validation exercise for this model. This validation exercise contains two
parts. In the first, we increase the fuel economy of all vehicles by 20 percent, and then examine
the effects of changing key parameters on modeling results; we find that the model's results are
not especially sensitive to changes in these parameters. Next, with the model calibrated to sales
of model year (MY) 2008 vehicles, we use the actual changes in vehicle prices and fuel economy
for MY 2010 vehicles to predict the sales of MY 2010 vehicles; we then compare the predicted
sales to actual sales. Here, we find that the model, designed to examine changes due only to
price and fuel economy, did not do well as in predicting sales impacts as assuming that market
shares were the same as in MY 2008, during this time of significant economic change. Perhaps a
key finding of this exercise is that model validation is both very important and potentially very
difficult to assess.
2
The magnitude of the auto industry in the U.S. economy and the importance of its role in
international trade and environmental protection have led to dozens of articles that analyze the
impacts of various factors and policies on consumer vehicle purchases. For instance, Goldberg
(1998), Whitefoot and Skerlos (2011), and Jacobsen (2013) examine the effects of fuel economy
-------
standards; Greene (2009) considers feebates; Train and Winston (2007) test the competitiveness
of the U.S. auto industry; and Brownstone et al. (1996) model the market acceptability of
alternative-fuel vehicles. Helfand and Wolverton (2011) review this literature, though the
literature continues to expand (e.g., Bento et al. 2012; Allcott 2013).
In most of these papers, the quality of the model is based on the econometric analysis: if
the analysis meets theoretical and statistical requirements, and the results include expected and
statistically significant coefficients on variables, then the model is considered suitable for policy
analysis. Researchers commonly use their models for simulation of counter-factual situations
based on the best estimates of the baseline situation. For instance, Goldberg (1998), Austin and
Dinan (2005), and Jacobsen (2013) all assess the relative merits of a gasoline tax vs. fuel
economy standards. Goldberg found that gasoline prices would have to double to get the same
effect on fuel consumption as fuel economy standards, due to a low estimate of responsiveness to
operating costs; both Austin and Dinan and Jacobsen, on the other hand, find a gasoline tax to be
much more efficient.1
Despite their widespread use for policy simulation, these models have typically not been
validated for their ability to predict vehicles sales in response to new circumstances. That is,
rarely have their predictions been tested against real-world outcomes, to see if they can in fact
predict out of sample. In other disciplines, this cross-checking of modeling to actual outcomes is
a significant aspect of the research agenda. For instance, experimental economists often test
hypothetical scenarios against actual market behavior (e.g., Landry and List 2007) to examine
1 These models are not directly comparable. Unlike Goldberg's model, Austin and Dinan's and Jacobsen's models
take into account the used vehicle fleet. Because a gasoline tax affects existing vehicles as well as new vehicles, it
saves fuel across the fleet. In contrast, a fuel economy standard affects only new vehicles.
-------
the validity of stated preference studies. For evaluating air quality modeling simulations, the
U.S. Environmental Protection Agency has developed the Atmospheric Model Evaluation Tool
specifically to compare predictions about meteorology and air quality against outcomes (U.S.
Environmental Protection Agency 2014).
One exception in the vehicle modeling literature is Pakes et al. (1993), who (as
summarized in Berry et al. 1995):
. . . used our model's estimates to predict the effect of the 1973 gas price hike on the
average MPG [miles per gallon] of new cars sold in subsequent years. We found that our
model predicted 1974 and 1975 average MPG almost exactly. . . . However, by 1976
new small fuel efficient models began to be introduced and our predictions, based on
fixed characteristics, became markedly worse and deteriorated further over time.
Another exception is Haaf et al. (2013), who use data from MY 2004-6 vehicles to
estimate a number of different econometric models, and test their predictions against MY 2007
and 2010 vehicle sales. The models had an average error of 0.24 percent compared to a mean
vehicle share of 0.42 percent: in other words, "the models we construct are fairly poor predictors
of future shares." They find that a "static" model - that is, one that assumes constant market
shares — outperformed their estimated models for MY 2007, while the attribute-based models
predicted better for MY 2010. Finally, they caution that "some of the models with the best
predictive accuracy have coefficients with unexpected signs - likely biased due to correlation
with unobserved attributes."
Finally, Raynaert (2014) develops a structural model of vehicle supply and demand in
Europe, using data from 1998-2007; he then compares sales-weighted aggregate predictions from
the model for MY 2011 to actual outcomes. He finds close agreement: in a period where actual
-------
emissions dropped 14 percent, his estimates for emissions differed from the observed values by
2.3 percent. Weight, footprint, and the share of diesel also had discrepancies of 3 percent or less;
price/income and horsepower differed by under 10 percent. He does not provide information
about more disaggregated results.
The paucity of research assessing the performance of vehicle choice models, along with
these papers, suggests that the predictive ability of consumer vehicle choice models is not yet
proven.2 At the same time, analysis of the impacts of policies on vehicle sales and class mix
require some prediction, whether a static approach of constant market shares, or a more
sophisticated analysis that accounts for future vehicle or market characteristics. This paper adds
to that literature by performing a validation exercise on a consumer vehicle choice model
developed for the U.S. Environmental Protection Agency (EPA). This model is designed for
very specific policy simulations: to examine the effects of changes in fuel economy and vehicle
price on U.S. vehicle purchase patterns in response to GHG standards. The fact that it was
designed for static policy simulations rather than for forecasting raises additional issues for
model validation.
3 Til
Greene and Liu (2012) developed the vehicle choice model used here for EPA
specifically to predict changes in total sales and fleet mix associated with GHG/fuel economy
standards. As will be discussed further, it is intended to compare a specified fleet with and
2 In addition, Knittel and Metaxoglou (2014) find that the estimation method Berry et al. (1995) and Raynaert (2014)
used is sensitive to start values and optimization algorithms, with results varying by substantial margins.
3 This section draws heavily from Greene and Liu (2012).
7
-------
without fuel economy standards;4 it is thus a static model, not intended to account for changes in
macroeconomic or demographic conditions.
It is a nested logit with a representative consumer and 5 layers, as described in Figure 1
and Table 1. The first layer constitutes the buy/don't buy decision. Next it distinguishes between
passenger vehicles, cargo vehicles, and ultra-prestige vehicles. In the model, sport-utility
vehicles and minivans are included as passenger vehicles, although many of these vehicles are
considered light-duty trucks for regulatory purposes. Consumers commonly consider these to be
passenger vehicles; it is more likely, for instance, that people consider an SUV to be a substitute
for a large or midsize car than for a pickup truck. Because the model is meant to reflect
consumer decision processes, it was considered appropriate to nest SUVs and minivans as
passenger vehicles rather than cargo vehicles. Ultra-prestige vehicles are defined as those with
price exceeding $75,000.
4 EPA regulates GHG emissions from vehicles; the Department of Transportation regulates vehicle fuel economy.
Because the primary way to reduce GHG emissions is to improve fuel economy, the agencies harmonized their
regulations (U.S. EPA and Department of Transportation 2010, 2012). The model uses fuel economy rather than
GHG emissions, because fuel economy is much more salient an attribute to vehicle buyers.
-------
Buy a New_
Vehicle
Cargo
Vehicle
Ultra
Prestige
Subcompact Compact Midsize Large Subcompact Compact Midsize Large
Figure 1: Nested Multinomial Logit Structure of Consumer Choice Model
Note: "Standard" is synonymous with "Non-Prestige"
Table 1: Vehicle Class Definition in the Consumer Vehicle Choice Model
Model Class
Corresponding EPA Class
1. Prestige1 Two-Seaters
2. Prestige Subcompact Cars
3. Prestige Compact Cars and Small Station
Wagons
4. Prestige Midsize Cars and Station Wagons
5. Prestige Large Cars
6. Two-Seater
7. Subcompact Cars
8. Compact Cars and Small Station Wagons
9. Midsize Cars and Station Wagons
10. Large Cars
11. Prestige SUVs
12. Small2 SUVs
13. Midsize SUVs
14. large SUVs
15. Mini Vans
16. Cargo/Large Passenger Vans
17. Small Pickup Trucks
18. Standard Pickup Trucks
Two Seaters
Subcompact Cars, Minicompact Cars
Compact cars, Small Station Wagons
Midsize Cars, Midsize Station Wagons
Large Cars
Two Seaters
Subcompact Cars, Minicompact Cars
Compact Cars, Small Station Wagons
Midsize Cars, Midsize Station Wagons
Large Cars
SUVs
SUVs
SUVs
SUVs
MiniVans
Cargo Vans, Passenger Vans
Small Pickup Trucks
Standard Pickup Trucks
-------
19. Ultra Prestige Vehicles3 See the definition (note 4) below
Notes:
(1) Prestige and non-prestige classes are defined by vehicle price: the prestige are vehicles
whose prices are higher than or equal to unweighted average price in the corresponding EPA class,
and vice versa for non-prestige vehicles. E.g., Prestige Two-Seater class is the set of relatively
expensive vehicle configurations in EPA class of two seaters with prices higher than or equal to the
unweighted average price of EPA two seaters.
(2) Non-prestige SUVs are divided into small, midsize and large SUVs by vehicle's footprint
(small: footprint <43; midsize: 43<=footprint<46; large: footprint>=46)
(3) Ultra Prestige class is defined as the set of vehicles whose prices are higher than or equal
to $75,000.
The model then separates passenger vehicles into Two Seaters, Prestige Cars, Standard
Cars, Prestige SUVs, Standard SUVs, and Minivans (with prestige also determined by price), and
cargo vehicles into Pickup Trucks and Vans. The next level continues the division into classes,
and the final level consists of individual vehicles. The model is calibrated to sales by individual
vehicle type in a base year through use of each vehicle's price and fuel economy. Fuel savings
for a vehicle are calculated as the present value, for a user-defined period of time (the "payback
period"), of fuel expenditures, based on the vehicle's mpg, vehicle miles traveled, and fuel
prices. The price and fuel savings are used to estimate an effective price; when that effective
price is combined with the price slope for that vehicle's nest, the constant term is the value that
results in matching the initial sales volume for that vehicle.
Vehicle sales are predicted to change in response to changes in net vehicle price, where
the change in net vehicle price is calculated as the increase in vehicle cost associated with
technologies to reduce GHGs, less a discounted share of the future fuel savings associated with
those technologies. Greene (2010) found highly varied estimates in the literature of consumer
willingness to pay (WTP) for additional fuel economy in the vehicle purchase decision, with a
number of studies showing WTP less than the expected value of future fuel savings, and some
others showing overvaluation. The model allows a user to choose the number of years of
10
-------
expected fuel savings that vehicle buyers are believed to consider in their purchase decisions, as
well as the future fuel prices and discount rate they might use for those calculations.
The model is designed to interact with EPA's technology-cost model, the Optimization
Model for reducing Greenhouse Gases from Automobiles (OMEGA), which seeks cost-effective
combinations of technologies to achieve GHG standards (U.S. EPA 2012). Iteration between the
model and OMEGA can be used to estimate whether sufficient technology is added to vehicles to
bring fleets into compliance with standards, after modeled consumer responses are taken into
account.
The demand elasticities in the model for each vehicle nest are not estimated from an
original data set, but rather are based on reviewing estimates in the literature (Greene and Liu
2012, Table 4). This approach has advantages and disadvantages. It allows for synthesis of the
results from multiple analyses, and professional judgment about whether the values are
appropriate. It also can be viewed as combining results from different studies, where the
differences in the studies may have implications for the value. Table 2 provides the elasticities
used in the analysis.
Table 2: Default Elasticities
Level
4
1
2
O
4
Choice of Make, Model, Engine
Transmission Configuration within a Class
Name
Prestige Two-Seater
Prestige Subcompact
Prestige Compact and
Small Station Wagon
Prestige Midsize Car
and Station Wagon
Elasti-
city
-3.8
-3.5
-3.5
-3.6
Parent
Type
Two-Seater
Prestige
Car
Prestige
Car
Prestige
Car
Level
3
1
2
O
4
Choice Among 19 Vehicle
Classes within Vehicle Type
Name
Two-
Seater
Prestige
Car
Standard
Car
Prestige
SUV
Elasti-
city
-1.3
-2.2
O
Parent
Category
Passenger
Passenger
Passenger
Passenger
11
-------
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
Prestige Large
Two- S eater
Subcompact
Compact and Small
Station Wagon
Midsize Car and
Station Wagon
Large Car
Prestige SUV
Small SUV
Midsize SUV
Large SUV
Minivan
Cargo / large
passenger van
Cargo Pickup Small
Cargo Pickup
Standard
Ultra Prestige
-3.5
-3.5
-5
-5
-5
-5
-3.7
-4.9
-5.1
-5.1
-4.9
-5.1
-5.1
-5.1
-3.9
Prestige
Car
T wo- S eater
Standard
Car
Standard
Car
Standard
Car
Standard
Car
Prestige
SUV
Standard
SUV
Standard
SUV
Standard
SUV
Minivan
Cargo Van
Cargo
Pickup
Cargo
Pickup
Ultra
Prestige
5
6
7
8
9
Level
2
1
2
O
Level
1
Standard
SUV
Minivan
Cargo
Van
Cargo
Pickup
Ultra
Prestige
-2.7
-2
Passenger
Passenger
Cargo
Cargo
Ultra
Prestige
Choice of Vehicle Type within
Passenger or Cargo Categories
Name
Passeng
er
Cargo
Ultra
Prestige
Elastic
ity
-1.1
-0.7
Parent
Node
Buy
Buy
Buy
Choice of Passenger, Cargo or
Ultra Prestige Vehicle
Name
Buy
No Buy
Elastic
ity
-0.7
Parent
Node
Root
Root
A few limitations of the model are identifiable even before any simulations are run.
Some of these limitations arise from the model being designed to be calibrated to an existing
fleet and then to estimate deviations from that initial calibration. The model thus does not
account for macroeconomic shocks that might affect either total sales or changes in fleet mix
independent of GHG standards, the introduction or departure of vehicles in the fleet, changes in
consumer preferences, or manufacturer changes in other vehicle characteristics (such as
12
-------
performance or appearance). For the purposes for which the model was built, these are not
limitations. The model was designed for static, same-year analysis of the effects on vehicle sales
of adding fuel-saving technologies and their costs; that is, it was intended to compare vehicle
sales with and without fuel-saving technologies and additional costs for a single fleet of vehicles.
In principle, then, changes in the economy, demographics, or the fleet over time should not affect
the ability of the model to predict, because it is predicting against a static counter-factual.
However, the baseline of no standards and the counter-factual of meeting the standards do not, in
reality, exist in the same year. Instead, the model will here be tested for its ability to predict
between two model years. As will be discussed further, the years for which we currently have
data involve the beginning and the depths of the Great Recession, whose effects may swamp any
predictive abilities of the model, and the Cash for Clunkers program in 2009 that may have
pulled sales forward from 2010. This limitation is therefore an issue for this method of testing
the model.
Other limitations are associated with the use of nested logit. For instance, as Train
(2009) notes, "only differences in utility matter." As a result, an equal change in prices (e.g.,
$1000) for all vehicles in the same nest would lead to no reallocation of market shares among
vehicles in that nest, although a $1000 change has a much bigger relative impact on the price of
an inexpensive car than that of a more expensive car. (The price increase would change total
sales and market shares across nests.) The nested logit also puts restrictions on demand
elasticities for the nests: responsiveness to price must be highest at the individual-vehicle level,
and decrease at each higher nest. The model includes a validation step to ensure that these
elasticity restrictions are achieved. Finally, within nests, logit exhibits "independence of
irrelevant attributes" (IIA): the ratio of probabilities (or market shares, in this model) of two
13
-------
options does not vary if a third option is added to the mix. As a result, an increase in the market
share of one alternative within a nest draws proportionately from all other alternatives (Train
2009, Chapter 3). Across nests, IIA does not hold. It is thus important for the nests within a
nested logit to contain vehicles that are close substitutes for each other, so that this substitution
pattern is a reasonable approximation.
The model testing consists of two parts. The first uses a hypothetical 20 percent increase
in fuel economy for all vehicles to examine the sensitivity of the model to changes in key
parameters, including payback period for fuel economy, discount rate, elasticities, and start
values. The second involves calibrating the model to MY 2008 vehicles and then using the
changes in fuel economy and price between 2008 and 2010 as inputs to review the ability of the
model to predict changes in vehicle sales.
4
Data requirements for the model include the vehicle's price, fuel economy, and sales, as
well as the new fuel economy and the change in price. These data come from market data
assembled by EPA and the Department of Transportation for their analysis of GHG standards for
MYs 2017-25 (U.S. Environmental Protection Agency and Department of Transportation 2012)
for both MY 2008 and MY 2010 vehicles. Both datasets contain over 1000 unique vehicles.5
The sensitivity analyses use only the MY 2008 data. Both price and fuel economy enter
into the calculation of net price that the model uses to estimate sales changes; for hypothetical
effects of the model, where the goal is to simulate a relatively arbitrary change in net price, it is
5 For example, there are 20 different versions of the Chevrolet Silverado in the 2008 data, each unique based on
engine, footprint, fuel economy, baseline sales, and other attributes.
14
-------
not necessary to change both price and fuel economy. The policy simulations therefore involve a
20 percent increase in fuel economy to all vehicles, with no increase in price.6 In essence, the
policy scenario is a reduction in the net price of all vehicles. The net price reduction is greater in
absolute terms for vehicles with lower fuel economy, because a 20 percent increase in miles per
gallon for, e.g., a vehicle that gets 10 mpg results in a much greater reduction in fuel
consumption than a 20 percent increase for a more efficient vehicle (Larrick and Sol 2008).7 The
simulation analyses use the entire MY 2008 vehicle fleet.
An additional needed set of parameters consists of fuel prices, used for the calculation of
fuel savings over the period that a vehicle buyer considers in the purchase decision (here called
the "payback period"). The sensitivity analyses use fuel prices as projected in the 2008 Annual
Energy Outlook (Energy Information Administration 2008). The calculation of fuel savings also
uses the schedule of vehicle miles traveled used in U.S. Environmental Protection Agency and
Department of Transportation, 2012.
Both the MY 2008 and 2010 datasets are needed for the prediction exercise. In this case,
the model takes as input the baseline price and fuel economy of each vehicle in MY 2008, and
then uses any change in price and fuel economy between MY 2008 and 2010 to predict sales in
MY 2010. Therefore, each MY 2008 vehicle needed to be matched with its MY 2010
counterpart. This matching is not straightforward. Vehicles enter and exit the market between
any two model years; indeed, Saab dropped out of the market entirely during this time. This
6 The model calibrates itself to the base year data, so that, if price and fuel economy do not change, the model
returns the initial sales, regardless of the values of other parameters. A policy scenario is necessary to produce
changes in vehicle sales.
7 This "mpg illusion" arises because fuel consumption is inversely related to mpg. For a vehicle that drives 15,000
miles per year, switching from a 10-mpg vehicle to a 12-mpg vehicle saves 250 gallons per year; switching from a
30-mpg vehicle to a 36-mpg vehicle saves 83 gallons per year.
15
-------
paper uses two methods to address this problem. In the first, aggregation by vehicles, multiple
trim levels (for instance, two-door vs. four-door versions of a model) of each vehicle are
combined through sales-weighting. This approach allows matching of most of the individual
vehicle models. In the second, aggregation by class, all vehicles are aggregated, by
manufacturer, to the classes of the vehicle choice model (see Table 1 for those classes). In this
study, any remaining unmatched vehicles are dropped from the analysis. Table 3 and Table 4
provide the summary statistics for these two methods compared to the whole fleets. Both cases
permit matching of over 90% of the vehicles sold in either model year, though aggregating by
class allows for representation of somewhat more vehicles.
Table 3: Summary Statistics of Baseline and Aggregated Fleets
Total
number of
unique
vehicles
Total
vehicle
sales
% Total
vehicle
sales
captured in
the final
matching
process
2008
Baseline
1302
13,851,761
—
Fleet
Aggregated
by Vehicle
524*
12,976,766
94%
Fleet
Aggregated
by Class
145**
13,573,775
98%
2010
Baseline
1171
11,190,180
—
Fleet
Aggregated
by Vehicle
524*
10,199,188
91%
Fleet
Aggregated
by Class
145**
10,648,872
95%
*108 unmatched vehicles include manufacturers or vehicles manufactured in one year but not in
the other. These are dropped in the analyses that follow.
**Two manufacturers (Spyker/Saab, Tesla) had sales in MY 2008 but not MY 2010. In 36
occasions, a manufacturer had sales in a vehicle class in one year but not in the other. These are
dropped in the analyses that follow.
16
-------
Table 4: Additional Summary Statistics
Total sales
(millions)
Weighted avg.
price
Minimum price
Maximum price
Weighted avg. fuel
economy
Min fuel economy
Max fuel economy
Share passenger
Share cargo
Share ultra-
prestige
MY
2008
Actual
13.9
$27,873
$11,783
$1.7M
26.2
12.0
65.8
86.3%
12.8%
0.9%
MY 2008
aggr. by
vehicle
13.0
$27,702
$11,783
$1.7M
26.3
12.0
65.8
85.7%
13.4%
0.9%
MY 2008
aggr. by
class
13.6
$27,850
$13,646
$254,533
25.7
15.2
49.5
86.0%
13.1%
0.9%
MY
2010
Actual
11.2
$26,767
$9,970
$1.7M
28.4
12.0
70.8
87.8%
11.6%
0.7%
MY 2010
aggr. by
vehicle
10.2
$26,624
$11,923
$1.7M
28.3
12.0
70.8
86.8%
12.7%
0.5%
MY 2010
aggr. by
class
10.6
$26,861
$12,816
$213,672
27.5
14.1
49.1
87.2%
12.1%
0.7%
Table 4 shows that the two forms of aggregation lead to small differences in fleet
characteristics. Aggregating by class matches the full fleet slightly better on weighted average
price, but aggregating by vehicle matches slightly better on average fuel economy. Differences
in shares among passenger, cargo, and ultra-prestige vehicles are less than one percent in all
cases.
17
-------
5
The baseline parameterization of the consumer choice model used the elasticities in Table
2; an assumption that consumers would consider 5 years of fuel savings in their purchases (i.e., a
five-year payback period); and a discount rate of 3 percent for calculating future fuel savings.
The policy experiment is an across-the-board 20 percent increase in fuel economy, about
equivalent in total sales impacts to a reduction in prices for all vehicles of 6.5 percent.8
As Table 5 shows, increasing the fuel economy of all vehicles by 20 percent has the
expected effect of increasing vehicle sales, from 13.85 million to 14.55 million (5 percent).
While sales increase for all vehicle classes, the largest absolute increases in sales occur for Cargo
Pickup Standard, Small SUV, and Prestige SUV, which each increase by over 100,000 vehicles,
roughly 8 percent in all three cases.9 The smallest percentage increases, 2.5% or less, were for
Subcompacts, Compacts, Small Cargo Pickups, and Two-Seaters. This pattern perhaps reflects
the model's use of expected future fuel savings in the net price calculation. The classes with the
greatest sales gains had initial fuel economy that averaged between 19 and 23 mpg; the classes
with the smallest increases had average initial fuel economy over 30 mpg, except for Small
Cargo Pickups (23.5 mpg). As discussed above, the absolute reductions in net prices for the less
efficient vehicles were greater than that for the more efficient vehicles, and the model finds
greater sales increases for those less efficient vehicles. As a result of the change in sales mix,
8 Changing price by a uniform percentage leads to different sales mix than changing fuel economy by a uniform
percentage, because price and fuel economy are not perfectly correlated.
9 Others with large percentage increases were also generally large or prestige vehicles, including Prestige Two-
Seaters, Large Cars, Minivans, Cargo and Large Passenger Vans, but had much smaller total sales.
18
-------
fleet average fuel economy is predicted to increase from 26.2 mpg to 31.2 mpg, slightly less than
the 20 percent increase applied to all vehicles.
Table 5: Effects on MY 2008 Fleet of Increasing Fuel Economy by 20 Percent
Prestige Two-Seater
Prestige Subcompact
Prestige Compact and
Small Station Wagon
Prestige Midsize Car
and Station Wagon
Prestige Large
Two-Seater
Subcompact
Compact and Small
Station Wagon
Midsize Car and
Station Wagon
Large Car
Prestige SUV
Small SUV
Midsize SUV
Large SUV
Minivan
Cargo / large
passenger van
Cargo Pickup Small
Cargo Pickup
Standard
Ultra Prestige
Total/Average
Initial
Sales
75,467
303,812
389,652
587,330
987,537
64,730
952,113
1,288,133
1,927,009
332,307
1,377,565
1,351,091
285,355
1,305,509
719,529
33,384
364,995
1,377,290
127,672
13,850,480
Initial
Fuel
Economy
24.8
26.8
27.4
26.2
27.3
34.0
34.7
33.9
33.4
29.0
21.1
23.1
25.9
27.6
24.9
19.2
23.5
19.7
20.1
26.2
Final Sales
79,806
317,930
406,431
616,536
1,030,415
66,211
970,570
1,320,900
1,984,811
355,533
1,486,070
1,467,415
298,712
1,343,232
763,159
36,071
374,321
1,494,324
136,388
14,548,836
Percent Change
in Sales
5.6%
4.5%
4.2%
4.9%
4.2%
2.3%
1.9%
2.5%
3.0%
6.8%
7.6%
8.3%
4.6%
2.8%
5.9%
7.7%
2.5%
8.2%
6.6%
4.9%
Final Fuel
Economy
29.8
32.1
32.8
31.4
32.7
40.3
41.3
40.4
39.6
34.8
25.2
27.6
30.5
32.9
29.8
22.9
28.0
23.6
24.1
31.2
For further testing of sensitivity of the results, we varied the payback period, the discount
rate, the elasticities, and the start values.
19
-------
This scenario models the effects of changing the payback period from 1 to 7 years of
future fuel savings taken into consideration by vehicle buyers, for the same scenario of a 20
percent increase in fuel economy. Greene (2010) finds a wide range of values in the literature
for the willingness of consumers to pay for fuel economy; the payback period is thus a source of
uncertainty. In estimating the effects of MY 2012-16 vehicle fuel economy/GHG standards on
vehicles, EPA used a 5-year payback period, "which is consistent with the length of a typical
new light-duty vehicle loan" (U.S. Environmental Protection Agency and Department of
Transportation 2010, p. 25517). The 5-year payback period scenario is the same as the default
scenario discussed above.
Total sales increase by approximately 100,000 vehicles, or less than 1 percent, for every
year increase in the payback period. Figure 2 shows that shorter payback periods result in less
change in market shares, because the change in net price is much smaller when the vehicle buyer
ignores more of the future fuel savings. Changes in sales mix, then, become more important as
vehicle buyers put more weight on future fuel savings in their purchase decisions.
20
-------
O.EO::
3.4C'::
0.2C'::
I
tlOK
&.QQ-*
-•PDP-IY—
• PBP=2Y
• PBP=3Y
• PBP=4'T
FEF = 5) (DEFi
PBP=6Y
• PBP=7Y
1 i • • • »
-a -o -n -a ? I ^ I F IV
I! £ a £ | I I
tH iff W tW IB" 5f •
I 1 | I f f I
_$ jf 3 g * r i
^ a P: =• R -^
"° v,
3 I
r— -a vt
* S 3
MB'
rs ^^ ^
1
Jl
Q "|
-.
2 s
K B 9
S= ii
Su
n -i
Q.
I
Figure 2: Effects on Market Shares of Changing Payback Period
Note: PBP = payback period, the number of years of fuel savings taken into consideration by
vehicle buyers.
5.2 Discount rate
The discount rate enters the model because future fuel savings come to consumers over
time; the savings in future years are discounted in the calculation, with larger discount rates
reducing the effect of future fuel savings in the net price. As a result, varying the discount rate
provides results very similar to those for varying payback period. (The 3 percent discount rate is
the same as the default scenario.) Figure 3 shows the effects of changing the discount rate on
market shares. It shows the same pattern as the payback period results, with vehicle classes
showing the same patterns of gains and losses, because it has similar effects on net price.
Because the period facing the discount rate is small (5 years), the magnitude of the effect is
small. Varying the discount rate between 2 and 10 percent led to a change in total sales of
21
-------
111,000 vehicles, less than 1 percent, and about the same amount as changing the payback period
by only 1 year.
0.40%
0.30%
I Discount = .02
Discount = .03
I Discount = .07
I discount = .10
Figure 3: Effects on Market Shares of Different Discount Rates
5.3 Elasticities
To test the effects of the chosen elasticities, we multiplied all the elasticities in Table 2
by 1.5, to see the sensitivity of the analysis to the elasticity values. In addition, we ran scenarios
where we multiplied the elasticity of only one class of vehicles by 1.5, to see the effects of
changing the elasticity for only one class. The results are in Figure 4.
22
-------
Change in Market Share
-0.50% -0.40% -0.30% -0.20% -0.10% 0.00% 0.10% 0.20% 0.30% 0.40% 0.50% 0.60%
Ultra Prestige
Cargo Pickup Standard
Cargo Pickup^wtt"
Cargo / large passenger van
Minivan
Midsize SUV
Small SUV
Prestige SUV
Large Car
Compin HIM! OiMill OUIiiH Wi|
Two-Seater •
Prestige Lw§^
Prestige Midsize & Station Wagon
Prestige Compact & Small Station Wagor*
Prestige Subcompact
Prestige Two-Seater
Figure 4: Effects on Market Shares of Multiplying All Elasticities by 1.5
The model predicts that total sales increase by about 1 million vehicles - about 7 percent
— with higher elasticities, compared to the increase of 700,000 vehicles (5 percent) with the
baseline elasticities, for the same changes in net prices. With the higher elasticities, vehicle
buyers are more responsive to the effective reduction in vehicle prices due to the improvement in
fuel economy. Changes in market shares generally exhibit the same pattern of shifts to relatively
inefficient vehicles.
When the elasticity of only one class is changed, sales increase by about 700,000
vehicles, almost exactly the amount that vehicle sales changed with the default elasticities,
regardless of which vehicle class's elasticity is changed. On average, sales in the class whose
elasticity changed increased by about 5 percent, about the same as the increase in sales for the
23
-------
other classes. Market shares also change very little: with average market share per class about
5.3 percent, the maximum change in market share for any scenario was 0.5 percent. This result
suggests that the model is not especially sensitive to elasticities in the bottom nest.
5,4
To conduct policy projections, the model would need to be calibrated to a baseline fleet
in a future year. Because these future fleets in the absence of standards are not known, they
provide an additional source of uncertainty for the model. For that reason, we experimented with
changing the initial fleet: initially, by multiplying all sales by 1.5; and next, by multiplying
initial sales of each individual class by 1.5, holding other class sales constant. We then applied
the standard policy scenario of a 20 percent improvement in fuel economy.
The sales response to a change in the fleet size is just about proportional: just as the
initial sales increased 4.9 percent in response to the changes in fuel economy, sales with the
artificially larger fleet increased 4.9 percent.10 When sales for individual classes were increased
to 150 percent, the sales of that class, and of the remaining classes, increased by about the same
proportion as in either the baseline case or when all classes had baseline sales 150 percent higher
(see Table 6).
Table 6: Effects of 150 Percent Increases in Baseline Fleet on Changes in Predicted Sales
Prestige Two-Seater
Percent
Change
from 2008
sales due to
20% fuel
economy
increase
5.59%
Percent
Change from
150% of
2008 sales
due to 20%
fuel economy
increase
5.58%
Average
Percent Change
in Own Class
Sales when
Initial Own
Class Sales =
150% of 2008
5.56%
Percent Change
in Named Class's
Sales averaged
over 1 8 cases of
another class's
initial sales 150%
of 2008 sales
5.76%
' Cutting the baseline fleet size in half also produced a 4.9 percent increase in sales due to the policy.
24
-------
Prestige Subcompact
Prestige Compact and
Small Station Wagon
Prestige Midsize Car and
Station Wagon
Prestige Large
Two- S eater
Subcompact
Compact and Small
Station Wagon
Midsize Car and Station
Wagon
Large Car
Prestige SUV
Small SUV
Midsize SUV
Large SUV
Minivan
Cargo / large passenger
van
Cargo Pickup Small
Cargo Pickup Standard
Ultra Prestige
4.54%
4.22%
4.85%
4.25%
2.26%
1.92%
2.51%
2.96%
6.75%
7.58%
8.25%
4.57%
2.85%
5.89%
7.74%
2.52%
8.15%
6.60%
4.53%
4.21%
4.84%
4.24%
2.25%
1.91%
2.50%
2.94%
6.74%
7.57%
8.24%
4.56%
2.84%
5.87%
7.73%
2.51%
8.14%
6.59%
4.63%
4.30%
4.90%
4.36%
2.28%
2.07%
2.66%
3.10%
6.87%
7.60%
8.13%
4.71%
3.23%
6.04%
8.05%
2.81%
8.18%
6.74%
4.65%
4.31%
4.98%
4.34%
2.29%
1.93%
2.54%
2.99%
7.00%
7.89%
8.64%
4.68%
2.87%
6.06%
8.05%
2.54%
8.52%
6.83%
This effect is due to the Independence of Irrelevant Alternatives (IIA) characteristic of
logit. As mentioned above, this proportionate shifting occurs because, within a nest, the ratio of
any two market shares to each other is a constant (Train 2009, Chapters 3, 4); the market shares
must change by the same proportion. Across nests, the condition does not hold, and thus the
proportions are not exactly constant. Thus, even if the baseline fleet for a policy scenario is
inaccurate, the percent change in vehicle sales that it predicts appears to be insensitive to any
errors. This finding suggests that percentage increases from the model may be a way to present
results that is less sensitive to initial values than presenting sales estimates.
In sum, the sensitivity analyses suggest that the model is not very sensitive to changes in
key parameters. Reducing the net prices of vehicles increases sales, with a tendency for sales to
25
-------
move toward those with greater reductions. In addition, changing the default parameters in the
model within a reasonable sensitivity range does not have dramatic effect on model outputs.
Finally, the percent changes in sales from the model seem to be fairly insensitive to variation in
the baseline fleet. These findings may be considered comforting, because all these parameters
are subject to a fair degree of uncertainty.
6 2010 on. 2008
The sensitivity analysis provides insights into how the model functions. Further
validation would come from testing the model's results against actual market outcomes. As
straightforward as this goal seems, this is a significant challenge. As has been discussed, this
model is meant for a comparative static analysis: if fuel economy and price change, with
everything else held constant, how do sales change? In practice, though, holding everything else
constant is impossible. In an ideal experimental scenario, a randomized part of the country
would face vehicles with the baseline prices and fuel economies, and the rest of the country
would face the new prices and fuel economies, at the same time; but that scenario cannot happen
when standards apply equally across the country. If, instead, comparisons are made across time,
then a number of factors that affect vehicle sales are unlikely to be constant: economic
conditions, demographic characteristics, and vehicle characteristics other than fuel economy and
price may change. The model is not built to address these questions. As discussed above, the
model was not built to be a forecasting model; for the analyses of the effects of potential
GHG/fuel economy standards, a comparative statics exercise is appropriate, as long as the model
results reflect the response of vehicle markets. This comparative statics framework nevertheless
raises significant questions about how to conduct model validation.
26
-------
With these misgivings as background, we nevertheless compare the model's predictions
due to changes in fuel and vehicle prices and changes in fuel economy in MY 2010, relative to
MY 2008 vehicles, to those that occurred during MY 2010. As noted above, these model years
were used as baseline datasets for the EPA/NHTSA GHG/fuel economy standards for MYs
2017-25 and were thus readily accessible.
The approach is to calibrate the model to MY 2008 vehicle sales, as was done for the
sensitivity analyses; provide the model with changes in each vehicle's fuel economy and price
between MY 2008 and 2010; and use those changes to predict sales in MY 2010. Those
predictions are then compared to actual sales in 2010. Though the Great Recession clearly had a
significant effect on the vehicle market during this time - as Table 3 shows, sales dropped by
about 20% — it is also a period of changes in vehicle characteristics that could be reflected in the
modeling results. This period thus provides an opportunity for an initial review of the model's
ability to predict changes in the vehicle fleet.
As discussed earlier, we use two datasets: sales-weighted aggregation of different trims
into a single model (aggregation by vehicle); and sales-weighted aggregation by manufacturer of
all vehicles in a class (aggregation by class). The net change in vehicle price in the model is the
change in the vehicle's purchase price, plus some part of the expected lifetime fuel consumption
of the vehicle. Expected future fuel consumption used in the model is based on a vehicle's fuel
economy, vehicle miles traveled for the specified payback period, fuel prices taken from the
Energy Information Administration's Annual Energy Outlook, a 3% discount rate, and a 5 year
payback period. As discussed above, the choice of discount rate in the model has a very small
effect on the results; the choice of payback period has a small but somewhat larger effect.
27
-------
7
Table 7 and Table 8 provide an overview of results for the two methods of aggregation.
Note that the "actual" market results in the tables omit the vehicles that were not matched
between the model years (2 - 9% of all vehicle sales), and thus were excluded from the modeling
exercise. This approach assesses the model using all vehicles included in the modeling exercise,
rather than the entire population of vehicles. Because each aggregated dataset included a slightly
different set of vehicles, the "actual" results are not the same when aggregating by vehicle
compared to aggregating by class, as shown in Table 3.
Table 7: Predicted vs. Actual Results with Aggregation by Vehicle
Total Sales
Weighted Avg.
Fuel Economy
Share passenger
Share cargo
Share ultra-prestige
Sum of Absolute
Deviations
Sum of Squared
Deviations
MY 2008
Actual
12,976,766
26.3
0.857
0.134
0.009
MY 2010
Actual
10,199,188
28.3
0.868
0.127
0.005
MY 2010
Predicted
13,470,888
27
0.849
0.141
0.010
MY 2010
Actual - MY
2008 Actual
-2,777,578
2.0
0.012
-0.008
-0.004
0.441
0.0028
MY 2010
Actual - MY
20 10 Predicted
-3,271,700
1.3
0.019
-0.014
-0.005
0.591
0.0051
Table 8: Predicted vs. Results with Aggregation by Class
Total Sales
MY 2008
Actual
13,573,775
MY 2010
Actual
10,648,872
MY 2010
Predicted
14,035,057
MY 2010
Actual -
MY 2008
Actual
-2,924,903
MY 2010
Actual - MY
2010
Predicted
-3,386,185
28
-------
Weighted Avg.
Fuel Economy
Share passenger
Share cargo
Share ultra-prestige
Sum of Absolute
Deviations
Sum of Squared
Deviations
26.3
0.860
0.131
0.009
27.5
0.872
0.121
0.007
26.9
0.862
0.129
0.009
-26.3
0.012
-0.009
-0.003
0.4354
0.0043
-26.9
0.010
-0.008
-0.002
0.5769
0.0101
Both methods do poorly in predicting total vehicle sales. This result is not a surprise,
given the model years studied and the model's function. As discussed above, the model is not
designed to predict future vehicle sales based on future changes; instead, it is intended for
comparisons within a model-year of vehicles with and without fuel-saving technologies. Sales in
MY 2010 were heavily affected by the Great Recession, which the model, calibrated to MY
2008, would not take into account. Both forms of aggregation predict increases in vehicle sales,
a result that must be due to decreases in effective prices (price plus a portion of future fuel
consumption) between the two years.
Both forms of aggregation correctly predict increases in fuel economy resulting from the
change in sales mix, though the actual increase in fuel economy is greater than that predicted in
either form of aggregation. This effect may, again, perhaps be due to the influence of the Great
Recession: people may have switched to less expensive vehicles, which may tend to be more
fuel-efficient than more expensive vehicles. There may be other explanations for this result as
well. Perhaps, for instance, people accounted more for future fuel consumption in their purchase
decisions than these model runs allowed.
29
-------
Although the model did not correctly estimate vehicle sales, perhaps it does better in
forecasting consumer shifts across vehicle classes in response to changes in price and fuel
economy. We thus calculate, for both datasets, the difference between predicted and actual
market shares at the vehicle level. In addition, as in Haaf et al., we calculate the difference
between actual market shares for MY 2010 and actual market shares for MY 2008, as a naive,
no-information alternative; one simple test of the value of the model is whether it out-performs
an alternative of assuming no change in market shares. To compare the two forecasts, we
calculate both the summed absolute deviation and the sum of squared deviations from actual MY
2010 market shares for both the modeling results and the MY 2008 actual results.
Table 7 and Table 8 provide results for shifts between passenger cars, cargo vehicles, and
ultra-prestige vehicles. At this very aggregated level, actual shifts are small - about 1 percent
from cargo vehicles to passenger vehicles. This very small shift may well be due to the way that
vehicles are classified in this model. Many vehicle classes that may be legally defined as trucks,
such as SUVs, are here considered to be passenger vehicles, because people use them that way.
With over 85 percent of vehicles in the passenger category, most shifts are likely to stay within
that category, rather than move across categories.
The two methods of aggregation produced opposite results direct!onally from the model
for the shares of passenger, cargo, and ultra-prestige vehicles: aggregation by vehicle implied a
switch from passenger vehicles to cargo vehicles, with aggregation by class showing what
actually happened, a relative increase in passenger vehicles. These shifts are small: the actual
full market share in passenger vehicles went from about 86% to 88% (see Table 4), though either
form of aggregation used a slightly smaller share of passenger vehicles. Both aggregations may
have omitted slightly more passenger vehicles than cargo or ultra-prestige vehicles, perhaps
30
-------
reflecting a greater tendency of passenger vehicles to be redesigned in ways that make them hard
to link across years.
Table 9 and Table 10 provide the results of these comparisons for the 19 vehicle classes
in the model. For both datasets, using MY 2008 market shares to forecast MY 2010 market
shares out-performs the vehicle choice model when deviations are measured at the class level.
This result is similar to those of Haaf et al., who found that using static market shares out-
performed attribute-based models when predicting one year ahead.11
11 Haaf et al. also found that attribute-based models could do better than the static market shares model for a four-
year-ahead forecast. They find generally that models with more attributes forecast better than models with fewer or
no attributes.
31
-------
Table 9: Class Shifts for Aggregation by Vehicle
Market Shares by Vehicle Class
Prestige Two-Seater
Prestige Subcompact
Prestige Compact & Small Station
Wagon
Prestige Midsize & Station Wagon
Prestige Large
Two-Seater
Subcompact
Compact and Small Station Wagon
Midsize Car and Station Wagon
Large Car
Prestige SUV
Small SUV
Midsize SUV
Large SUV
Minivan
Cargo / large passenger van
Cargo Pickup Small
Cargo Pickup Standard
Ultra Prestige
Sum of Absolute Deviations
Sum of Squared Deviations
MY
2008
Actual
0.0056
0.0200
0.0293
0.0378
0.0671
0.0029
0.0750
0.0916
0.1388
0.0314
0.0957
0.0919
0.0220
0.1004
0.0472
0.0010
0.0280
0.1053
0.0090
MY
2010
Actual
0.0031
0.0111
0.0327
0.0403
0.0552
0.0010
0.0617
0.1183
0.1830
0.0223
0.0800
0.0879
0.0171
0.1064
0.0481
0.0017
0.0244
0.1005
0.0052
MY 2010
Predicted
0.0052
0.0193
0.0321
0.0353
0.0618
0.0028
0.0858
0.0939
0.1331
0.0213
0.0932
0.0997
0.0165
0.0995
0.0497
0.0008
0.0227
0.1174
0.0099
MY 2010
Predicted High
Midsize Blast
0.0051
0.0190
0.0316
0.0347
0.0607
0.0027
0.0731
0.0799
0.1735
0.0182
0.0916
0.0980
0.0162
0.0978
0.0489
0.0008
0.0224
0.1160
0.0098
MY 2010
Actual - MY
2008 Actual
-0.0025
-0.0089
0.0034
0.0026
-0.0119
-0.0019
-0.0133
0.0267
0.0442
-0.0091
-0.0157
-0.0040
-0.0049
0.0060
0.0009
0.0008
-0.0036
-0.0049
-0.0038
0.1690
0.0036
MY 2010
Actual - MY
20 10 Predicted
-0.0021
-0.0082
0.0005
0.0050
-0.0066
-0.0018
-0.0241
0.0245
0.0499
0.0010
-0.0132
-0.0118
0.0006
0.0069
-0.0016
0.0009
0.0017
-0.0169
-0.0047
0.1820
0.0045
MY 20 10 Actual
-MY2010Pred
High Mid Blast
-0.0020
-0.0079
0.0011
0.0056
-0.0055
-0.0017
-0.0114
0.0384
0.0095
0.0041
-0.0116
-0.0101
0.0009
0.0086
-0.0008
0.0009
0.0019
-0.0155
-0.0046
0.1423
0.0024
32
-------
Table 10: Class Shifts for Aggregation by Class
Market Shares by Vehicle Class
Prestige Two-Seater
Prestige Subcompact
Prestige Compact & Small Station
Wagon
Prestige Midsize & Station Wagon
Prestige Large
Two-Seater
Subcompact
Compact and Small Station Wagon
Midsize Car and Station Wagon
Large Car
Prestige SUV
Small SUV
Midsize SUV
Large SUV
Minivan
Cargo / large passenger van
Cargo Pickup Small
Cargo Pickup Standard
Ultra Prestige
Sum of Absolute Deviations
Sum of Squared Deviations
MY
2008
Actual
0.0051
0.0220
0.0276
0.0426
0.0725
0.0028
0.0662
0.0949
0.1420
0.0245
0.0984
0.0975
0.0210
0.0962
0.0467
0.0025
0.0268
0.1015
0.0094
MY
2010
Actual
0.0026
0.0179
0.0329
0.0313
0.0288
0.0010
0.0435
0.1645
0.1790
0.0204
0.0948
0.0987
0.0111
0.0957
0.0496
0.0016
0.0233
0.0964
0.0067
MY 2010
Predicted
0.0047
0.0322
0.0293
0.0384
0.0536
0.0027
0.0868
0.1080
0.1117
0.0148
0.1033
0.1254
0.0119
0.0901
0.0490
0.0022
0.0239
0.1029
0.0090
MY 2010
Predicted High
Midsize Blast
0.0047
0.0320
0.0291
0.0382
0.0534
0.0027
0.0822
0.1023
0.1261
0.0141
0.1027
0.1247
0.0118
0.0896
0.0488
0.0022
0.0238
0.1026
0.0090
MY 2010
Actual - MY
2008 Actual
-0.0025
-0.0041
0.0053
-0.0112
-0.0437
-0.0018
-0.0226
0.0696
0.0371
-0.0041
-0.0036
0.0012
-0.0099
-0.0004
0.0029
-0.0008
-0.0034
-0.0051
-0.0027
0.2322
0.0090
MY 2010
Actual - MY
20 10 Predicted
-0.0021
-0.0143
0.0036
-0.0071
-0.0249
-0.0018
-0.0432
0.0565
0.0674
0.0055
-0.0085
-0.0267
-0.0008
0.0056
0.0006
-0.0006
-0.0005
-0.0065
-0.0023
0.2786
0.0114
MY 20 10 Actual
-MY2010Pred
High Mid Blast
-0.0021
-0.0141
0.0038
-0.0069
-0.0246
-0.0018
-0.0387
0.0622
0.0530
0.0063
-0.0079
-0.0260
-0.0007
0.0061
0.0009
-0.0006
-0.0005
-0.0062
-0.0023
0.2645
0.0099
33
-------
In predicting shares of the 19 vehicle classes included in the vehicle choice model,
aggregation by class correctly estimated the direction of shifts in more cases (14 out of 19
classes) than did aggregation by vehicle (10 out of 19 classes) (see Table 9 and Table 10). Most
of the shifts in market shares are small, though: in most cases (13 for aggregation by class, 14
for aggregation by vehicle), the predicted market share is within 1 percent of the actual market
share. With mostly small changes in market shares, it may be difficult to distinguish the quality
of modeling performance from a general tendency for market shares not to change very much.
For both aggregations, the largest class is Midsize Cars and Station Wagons. This class
experienced a relatively large shift in shares between 2008 and 2010, from about 14% to 18-
19%. Both forms of aggregation not only missed the magnitude of this shift, but even missed the
direction. It is not possible to say from which classes people switched (other than the obvious
point that people generally switched from classes where shares went down). The relatively
inaccurate performance for this large class suggests that it could be useful to experiment with
adjusting the demand elasticity for it, though results from the sensitivity analyses suggest that a
large elasticity change would be necessary to improve the results substantially.
Table 9 and Table 10 also include results from running the model with an elasticity of-15
- three times the default value - for the Midsize class. For Aggregation by Vehicle, this change
leads to an increase in market share for the Midsize class almost as large as occurred, with a
reduction in overall deviations as well; for Aggregation by Class, market share increases much
less, and using MY 2008 market shares is still the better forecast. This exercise suggests that it is
possible to improve the model's match to actual results, but it may match actual results only for
MY 2010. Additional data and further testing would be needed to evaluate whether the model's
forecasting ability is improved with the revised parameters.
34
-------
In sum, it is difficult to devise a test of the model, based on its design. As discussed
previously, the model is intended to consider the vehicle fleet with and without standards, not the
fleet response to changes in social and economic conditions. The ideal test would require having
two fleets of otherwise identical vehicles - one with fuel-saving technologies, one without -
available for sale at the same time, because that scenario is the one the model is designed to
assess. Using data from different model years clearly does not meet this ideal.
It is not a surprise that the model did not predict the reduction in sales due to the
recession. It is perhaps a bit more dismaying that it does not show a strong ability to predict
changes in market shares, perhaps due to missing changes in tastes or income effects between the
two years. Again, though, the model was not designed to consider how the recession may have
affected those factors. Given that most market shares, and changes in market shares, are small, it
may be difficult to identify those changes even under more consistent conditions. Indeed, the
results suggest that holding market shares constant from the initial year may provide better
estimates than using the model. Additional work, potentially with additional model years of
data, or development of new methods for model validation in the absence of a counter-factual,
may be needed to understand better the ability of the EPA model to estimate changes in vehicle
purchases associated with changes in vehicle fuel economy.
Consumer vehicle choice models are commonly used to simulate the effects of counter-
factual situations; they have been tested against actual market outcomes much less frequently. In
the few cases where models with forecasting ability have been tested against market outcomes,
results are still not very strong, especially for market share predictions. Perhaps innovations in
35
-------
vehicles or changes in consumer preferences in response to advertising or market conditions, or
even just the difficulties in properly specifying consumers' preferences, lead to limitations in
models' predictive abilities.
This paper adds to, and seeks to encourage, that literature by examining the performance
of a model developed for the U.S. Environmental Protection Agency to estimate responses to
changes in vehicle prices and fuel economy. The ability to adjust parameters easily allows for
sensitivity analysis with alternative assumptions of model parameters. The sensitivity analyses
suggest that moderate changes in the default parameters and baseline fleet have small effects on
the model outputs. Given the uncertainties associated with many of these parameters, this
finding suggests some robustness of modeling results to those uncertainties. The test of the
model against actual market outcomes suggests that the model is not suitable for forecasting
changes in the vehicle fleet when social and economic conditions are also changing. Because the
model was not designed to forecast such changes, this result is expected. It is nevertheless not
encouraging for model validation that assuming the market shares of the base fleet had less
forecast error than using the model.
Perhaps the major lesson is that conducting a validation exercise can be a significant
challenge, and perhaps other approaches may be needed to validate a model designed for policy
simulation rather than forecasting. First, as already mentioned, there is no obvious way to test
the model for the purpose for which it was designed, because only one vehicle fleet exists in the
U.S. in a year; no counter-factual exists. Vehicle choice models that incorporate demographic
factors and vehicle attributes may be better suited to testing across time; on the other hand, when
they are ultimately used for simulation purposes, such models require projections of those
demographic factors and vehicle attributes, which may not be of great reliability. Across time,
36
-------
any model has to face the fact that vehicles, even manufacturers, enter and exit the market.
Whitefoot et al. (2013) seek to endogenize manufacturer and consumer decisions simultaneously;
whether such efforts will reflect actual market movements is yet to be seen.
The results presented here suggest that further work is desirable. For instance, it would
be valuable to analyze additional years of data. Do predictions of responses to vehicles in future
model years follow the same pattern as in MY 2010? Or might the model predict better for non-
recession years, or worse for years further in the future? If adjustments to model parameters
improve forecasts for MY 2010 market shares, would those adjustments work as well for other
years? For other researchers, this paper aims to encourage further work on validation of other
models, both in development of methods and in application of those methods. We hope that this
paper stimulates more research on the ability of consumer vehicle choice models to predict actual
market changes.
37
-------
Allcott, H. (2013). "The Welfare Effects of Misperceived Product Costs: Data and Calibrations
from the Automobile Market." American Economic Journal: Economic Policy 5(3): 30-
66.
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 50: 562-582.
Bento, A.M., S. Li, and K. Roth (2012). "Is There an Energy Paradox in Fuel Economy? A
Note on the Role of Consumer Heterogeneity and Sorting Bias," Economics Letters 115:
44-48.
Berry, S., J. Levinsohn, and A. Pakes (1995). "Automobile Prices in Market Equilibrium,"
Econometrica 63(4): 841-940.
Brownstone, D., D. Bunch, T. Golob, and W. Ren (1996). "A Transactions Choice Model for
Forecasting Demand for Alternative-Fuel Vehicles." Research in Transportation
Economics, 4: 87-129.
Energy Information Administration (2008). "Annual Energy Outlook 2008" DOE/EIA-0383,
http://www.eia.gov/oiaf/aeo/pdf/0383%282008%29.pdf.
Goldberg, P. (1998). "The Effects of the Corporate Average Fuel Efficiency Standards in the
U.S." Journal of Industrial Economics 46(1): 1-33.
Greene, D. (2009). "Feebates, Footprints and Highway Safety," Transportation Research Part D
14: 375-384.
38
-------
Greene, D. (2010). "How Consumers Value Fuel Economy: A Literature Review." Office of
Transportation and Air Quality, U.S. Environmental Protection Agency, EPA-420-R-10-
008.
Greene, David L., and Changzheng Liu (2012). "Consumer Vehicle Choice Model
Documentation." U.S. Environmental Protection Agency EPA-420-B-12_052,
http://www.epa.gov/otaq/climate/documents/420bl2052.pdf.
Haaf, C.G., JJ. Michalek, W.R. Morrow, and Y. Liu (2014). "Sensitivity Of Vehicle Market
Share Predictions to Discrete Choice Model Specification." Journal of Mechanical
Design 136, 121402-121402-9.
Helfand, G., and A. Wolverton (2011). "Evaluating the Consumer Response to Fuel Economy:
A Review of the Literature." International Review of Environmental and Resource
Economics 5: 103-146.
Jacobsen, M. (2013). "Evaluating U.S. Fuel Economy Standards in a Model with Producer and
Household Heterogeneity." American Economic Journal: Economic Policy 5(2): 148-87.
Knittel, C.R., and K. Metaxoglou (2014). "Estimation of Random-Coefficient Demand Models:
Two Empiricists' Perspective." Review of Economics and Statistics 96(1): 34-59.
Landry, C.E., and J. List (2007). "Using Ex Ante Approaches to Obtain Credible Signals for
Value in Contingent Valuation Markets: Evidence from the Field." American Journal of
Agricultural Economics 89: 420-429.
Larrick, R.P., and J.B. Soil (2008). "The MPG Illusion." Science 320(5883): 1593-94.
Pakes, A., S. Berry, and J. Levinsohn (1993). "Applications and Limitations of Some Recent
Advances in Empirical Industrial Organization: Price Indexes and the Analysis of
39
-------
Environmental Change." American Economic Review Papers and Proceedings 83(2):
240-246.
Raynaert, Mathias (2014). "Abatement Strategies and the Cost of Environmental Regulation:
Emission Standards on the European Car Market." KU Leuven Center for Economic
Studies Discussion Paper Series DPS 14.31.
Train, K. (2009). Discrete Choice Methods with Simulation. Cambridge University Press.
http://elsa.berkeley.edu/books/choice2.html
Train, K., and Clifford Winston (2007). "Vehicle Choice Behavior and the Declining Market Share
of U.S. Automakers," International Economic Review 48: 1469-1496.
U.S. Environmental Protection Agency (2014). "Operational Performance Evaluation of Air
Quality Model Simulations."
http://www.epa.gov/AMD/Research/Air/operationalEval.html, accessed September 2,
2014.
U.S. Environmental Protection Agency (2012). "Optimization Model for reducing Emissions of
Greenhouse gases from Automobiles (OMEGA)."
http://www.epa.gov/otaq/climate/models.htm, accessed September 8, 2014.
U.S. Environmental Protection Agency and Department of Transportation (2010). "Light-Duty
Vehicle Greenhouse Gas Emissions Standards and Corporate Average Fuel Economy
Standards; Final Rule." Federal Register 75(88): 25324-25728.
U.S. Environmental Protection Agency and Department of Transportation (2012). "Joint
Technical Support Document: Final Rulemaking for 2017-25 Light-Duty Vehicle
Greenhouse Gas Emission Standards and Corporate Average Fuel Economy Standards."
40
-------
EPA-420-R-12-901, Chapter 1.
http://www.epa.gov/otaq/climate/documents/420rl2901.pdf
Whitefoot, K., M. Fowlie, and S. Skerlos (2013). "Compliance by design: Industry response to
efficiency standards." Working paper,
http://nature.berkeley.edu/~fowlie/whitefoot_fowlie_skerlos_submit.pdf
Whitefoot, K., and S. Skerlos (2012). "Design Incentives to Increase Vehicle Size Created from
the U.S. Footprint-Based Fuel Economy Standards." Energy Policy 41: 402-411.0
41
------- |