FINAL REPORT
Contract 05- 303
Follow- on Development of CARBITS:
A Response Model for the California Passenger Vehicle Market
Principal Investigator
Dr. David S. Bunch
( dsbunch@ ucdavis. edu)
Amine Mahmassani
Prepared for:
State of California Air Resources Board
Research Division
P. O. Box 2815
Sacramento, CA 95812
Prepared by:
University of California, Davis
Institute of Transportation Studies
One Shields Avenue
2028 Academic Surge
Davis, CA 95616
April 30, 2009
ii
Disclaimer
The statements and conclusions in this Report are those of the contractor and not
necessarily those of the California Air Resources Board. The mention of commercial
products, their source, or their use in connection with material reported herein is not
to be construed as actual or implied endorsement of such products.
iii
Acknowledgments
This Report was submitted in fulfillment of contract 05- 303, “ Follow- on Development of
CARBITS: A Response Model for the California Passenger Vehicle Market” by the
Institute of Transportation Studies, University of California, Davis, under sponsorship of
the California Air Resources Board. Work was completed as of April 30, 2009.
iv
Table of Contents
Title page i
Disclaimer ii
Acknowledgments iii
Table of Contents iv
Abstract v
Executive Summary vi
Background 1
Project Outcomes
1. A Generic Framework for CARBITS Models 5
2. Vehicle Class Definitions 6
3. Caltrans Travel Survey Data 20
4. Vehicle Choice Models 27
5. Department of Motor Vehicles ( DMV) Data 39
6. Vehicle Exit Modeling 41
7. Calibration 46
8. Hybrid Electric Vehicles 48
9. Concluding Remarks 51
10. Bibliography 52
Appendix on Vehicle Data 53
v
ABSTRACT
CARBITS is a market simulation model for the passenger vehicle market in California.
Professor David S. Bunch developed CARBITS for the ARB during 2003- 2004 under a
contract with the University of California, Davis. Its primary purpose is as a scenario
analysis tool to evaluate market response under alternative regulation scenarios. For
purposes of this Final Report, the version of CARBITS developed during 2003- 2004 will
be referred to as “ CARBITS 1.0.” CARBITS 1.0 was requested by the ARB to meet
specific needs for their work under AB 1493 regulating motor vehicle greenhouse gas
emissions, and was developed within a short time frame to accommodate their schedule.
The project was feasible because it was possible to base CARBITS development on pre-existing
research results developed under an earlier University of California- Institute of
Transportation Studies research program. Although time and monetary constraints
prevented development of a full range of features, ARB staff successfully used
CARBITS 1.0 in support of the climate change regulation adopted by the Board in
September 2004.
This project has produced an updated version of CARBITS (“ CARBITS 2.0”) with a
number of improvements and new features to address specific perceived “ deficiencies”
identified by ARB staff during the collaboration with Prof. Bunch. Some of these
represented desired extensions based on experience in using the model. A related area of
concern is the ever- present potential for criticism by the hired consultants of various
stakeholders. The original project proposal identifies a list of specific goals:
1. Estimate a new set of vehicle choice models using more recent datasets.
2. Specifically address the issue of vehicle market exit/ scrappage.
3. Develop re- calibration procedures to update certain model constants based on
aggregate- level vehicle counts.
4. Include the capability to address hybrid electric vehicles.
5. Address issues of statistical noise and runtimes.
These specific goals have been addressed by this project. A new set of vehicle choice
models has been estimated using data from the 2000- 2001 Caltrans Statewide Travel
Survey. This data source ( although a few years old) is attractive due to its large sample
size and high- quality sampling and weighting characteristics. In conjunction with using
these data ( which include information on vehicle holdings, but not transactions),
CARBITS was converted from a transactions microsimulation model to a vehicle
holdings model. This approach directly addresses the issue of statistical noise and run
times: holdings models use analytical computations that yield deterministic ( noise free)
results requiring relatively short run times. Substantial effort was invested in data
compilation and cleaning for this project. In particular, procedures for using DMV data
routinely accessible to ARB were developed to address needs for periodic re- calibration
using updated vehicle counts, patterns of vehicle market exit, and recent penetration of
hybrid electric vehicles. Aside from meeting specific project goals, the substantial
amount of work on data development, and the formulation of a generic vehicle market
model framework, will provide additional benefits to ARB in succeeding projects.
vi
EXECUTIVE SUMMARY
CARBITS is a market simulation model for the personal vehicle market in California.
Professor David S. Bunch developed CARBITS for the ARB during 2003- 2004 under a
contract with the University of California, Davis. Its primary purpose is as a scenario
analysis tool to evaluate market response under alternative regulation scenarios. For
purposes of this Final Report, the version of CARBITS developed during 2003- 2004 will
be referred to as “ CARBITS 1.0.” CARBITS 1.0 was commissioned by the ARB to meet
specific needs for their work under AB 1493 regulating motor vehicle greenhouse gas
emissions, and was developed under a short time frame. For practical reasons, it was
based on an existing model developed under an earlier University of California- Institute
of Transportation Studies research program. Although time and monetary constraints
prevented development of a full range of features, ARB staff successfully used
CARBITS 1.0 in support of the climate change regulation adopted by the Board in
September 2004.
Experience in working CARBITS as part of the 1493 rulemaking process led to some
ideas for potential improvements. The overall stated objective of this project is to update
and extend existing CARBITS model based on these experiences. Briefly, the stated
goals of this project are:
1. Estimate new vehicle choice models using more recently collected datasets.
2. Address issues of statistical noise and runtimes.
3. Specifically address the issue of vehicle market exit/ scrappage.
4. Develop re- calibration procedures to update certain model constants based on
aggregate- level vehicle counts.
5. Include the capability to address hybrid electric vehicles
To illuminate these goals, we first review some details about CARBITS 1.0. As noted,
CARBITS 1.0 was created using a pre- existing model. During the period 1992- 1995, a
team of Institute for Transportation ( ITS) researchers at University of California ( Davis
and Irvine campuses) pursued a multi- year research program involving data collection
and vehicle choice modeling. The California Energy Commission ( CEC) provided much
of the motivation for this work, which was targeted at exploring the future market for
alternative fuel vehicles in California, including: battery- powered electric vehicles,
compressed natural gas ( dedicated and dual fuel versions), and alcohol/ flex fuel. A major
task was fielding a panel survey of California households that included stated choice
questions on alternative fuel vehicles. One research goal was to explore household
demand models based on transaction choices ( e. g., vehicle replacement, addition, or
disposal decisions) as an alternative to vehicle holdings models ( the usual state of
practice). The results of this project were used to develop CARBITS 1.0 to meet the
needs of ARB.
The experiences and insights gained during the development and use of CARBITS 1.0
led to a number of ideas that were the motivation for this project. We briefly review
vii
these here. More details are included in the main report. First, from the very beginning
of the earlier project, concerns were raised about the dataset being “ old.” This is a
standard criticism for any model like CARBITS, given the expense and difficulty of
collecting large- scale data sets on a regular basis. Regardless of whether there are
technical merits to this narrow argument, it provides an opening to criticism by hired
consultants. Second, the transactions models adapted from the earlier research required
the use of pure microsimulation. This means that the model does not produce
deterministic, analytical results, and it also requires special expertise ( and long run times)
to produce results in the proper manner. One example of why this can be an issue
occurred during the 1493 rulemaking. Auto industry consultants ( either accidentally or
intentionally) produced results using CARBITS 1.0 that did not use enough replications
to produce stable results, and then used these in an attempt to undermine CARBITS. A
more practical concern is that using CARBITS 1.0 requires very long run times, making
analysis more burdensome to the user.
A related issue is that the original modeling approach was primarily concerned with
evaluating the entry of new types of vehicles ( none of which, by the way, were hybrid
electric— see below), with much less emphasis on vehicle exit and scrappage. CARBITS
1.0 takes an approach where vehicles exit the market “ implicitly,” based on the dynamics
of vehicle replacement. In contrast, other approaches use aggregate data to estimate
models that explicitly address vehicle exit. There are pros and cons to each method;
however, because the latter method is easier to understand, it is typically used by outside
consultants. Moreover, the AB 1493 experience suggests that a more complex model like
CARBITS is vulnerable to criticism through both misapplication of the model and
misrepresentation of results. Finally, there is the issue of hybrid electric vehicles. The
recent penetration of hybrid electric vehicles makes it obvious that future policy analyses
may need to address this new type of vehicle.
These specific goals listed above been addressed by this project. With regard to
introducing new data, various options were considered. Maintaining and updating
CARBITS 1.0 as a transactions- based model would require a new source of household
panel data that includes details on vehicle transactions. This type of data is very
expensive to collect and difficult to come by. Moreover, experience suggested that the
transactions- based approach was the common source of a number of the issues this
project was intended to address. Based on multiple factors, we decided to update
CARBITS using the 2000- 2001 Caltrans Statewide Travel Survey.
These data ( although a few years old) are attractive for a number of reasons. For a
household survey of this type it has a very large sample size ( over 17,000 households, all
from California), and uses high- quality sampling and weighting procedures. In
conjunction with using these data ( which include information on vehicle holdings, but not
transactions), CARBITS was converted from a transactions microsimulation model to a
vehicle holdings model. This approach directly addresses the issue of statistical noise
and run times, since holdings models can be implemented using analytical computations
that yield deterministic ( noise free) results requiring relatively short run times.
viii
Although it is less than obvious from the stated project goals, the decision to estimate a
completely new model for CARBITS ( regardless of which household dataset was chosen)
created a whole host of additional data requirements. Substantial effort was invested in
data compilation and cleaning for this project. One area requiring a large amount of work
was the development of a Vehicle Technology Database. Vehicle choice models have a
number of requirements for characterizing the vehicle choices faced by consumers in the
marketplace. These include such things as market prices, vehicle body types and sizes,
fuel economy, performance characteristics, and others. No one data source includes all of
these information items. This requires creating a large database by merging together data
from multiple data sources. Because each data source has its own way of defining
vehicles ( which includes character string data describing the make and model of vehicle),
cleaning and merging these data is a herculean task.
In addition to vehicle technology data, there are multiple aspects of the project that
require aggregate data on multiple aspects of the vehicle market. For example, models
like CARBITS ( which are estimated on the basis of household survey data) must
periodically be re- calibrated so that the vehicle distributions for the model base year
match the aggregated vehicle totals from an outside source ( project goal 3). In addition,
estimating a model of vehicle exit requires some type of data set that tracks the entry and
exit of vehicles from the market ( project goal 2). Finally, in recent years hybrid electric
vehicles have been entering the market. Survey data cannot possibly have the sample
size to obtain accurate measurements of this aggregate phenomenon ( project goal 4). To
address these data needs, procedures for processing Department of Motor Vehicles
( DMV) registrations data were developed.
We emphasize the data collection and cleaning aspect of this project because ( i) a
substantial amount of the contract effort was devoted to it, and ( ii) we consider the
outcome of this effort to be a major side benefit of this project that goes beyond the
narrow statement of the project goals. In a similar vein, our approach to creating the new
version of CARBITS (“ CARBITS 2.0”) incorporated system design concepts such as
object- oriented analysis and object- oriented programming. Specifically, rather than
program CARBITS 2.0 as a stand- alone one- time effort, we decided to create a generic
system framework for “ CARBITS- like models,” and then implement CARBITS 2.0 as a
specific “ instance” within this framework. The system framework and CARBITS 2.0
were implemented using the object- oriented features of MATLAB. ( In contrast,
CARBITS 1.0 was written in FORTRAN.) This approach will make any future efforts to
modify or update CARBITS much easier.
To summarize, the project outcomes include the following:
1. CARBITS was updated using a more recent data set ( 2000- 2001 Caltrans Travel
Survey)
2. CARBITS was converted to a holdings- based model from the original
transactions- based model.
1
3. Outcomes 1 and 2 directly address the issue of model runtimes and statistical
noise by using an approach that produces results based on deterministic
computations.
4. DMV data were developed as a source of data on aggregate vehicle counts,
vehicle entry and exit statistics, and penetration of hybrid electric vehicles.
5. Outcome 4 supported the development of procedures to re- calibrate model
constants to match aggregate vehicle totals, the estimation of a vehicle market
exit model, and the capability to incorporate data on hybrid electric vehicles.
6. A substantial amount of effort on compiling and cleaning data ( including many
data sets on vehicle prices and technology) yielded an additional side benefit for
future work by ARB.
7. CARBITS 2.0 was developed using object- oriented analysis and programming
methods. A generic system framework for “ CARBITS- like models” was
established, and then CARBITS 2.0 was coded as a special case.
BACKGROUND
In late 2002, ARB staff approached the Institute of Transportation Studies ( ITS) at
University of California, Davis ( UC Davis) to discuss a number of research needs related
to its charge to perform rulemaking under AB 1493 ( Pavley). One such need was for a
scenario analysis tool to provide a quantitative assessment of the effects of alternative
regulatory policies on the personal vehicle market in California over the medium and
long term. For example, manufacturers would be expected to change their vehicle
offerings in order to comply with a regulation. The operating characteristics, and new
vehicle prices would be expected to change. This, in turn, would elicit a response from
the vehicle market. Prof. David S. Bunch agreed to develop such a model under as part
of a larger research project performed during 2003- 2004. Both time and budget
requirements precluded a major research effort, e. g., fielding a household survey,
collecting data, and developing an entirely new model. The proposed solution was to
adapt models developed under an earlier research program.
The earlier research involved data collection and vehicle choice modeling for the
California market. It was performed during the mid 1990’ s by a team of ITS researchers
( including Prof. Bunch) from two University of California campuses ( Davis and Irvine).
The program was a multi- year effort with funding from multiple sources. The California
Energy Commission provided much of the motivation for this work. In addition to
funding a pilot project, they coordinated efforts for a sequence of projects funded first by
Southern California Edison, and then Pacific Gas & Electric. In addition, the research
team received pass- through federal funding from the ISTEA program.
One component of the project was a panel survey of California households. The desire
was to get observations from the same household at multiple points in time in order to
trace the transaction dynamics of their vehicle purchases. In addition, the survey
involved the application of stated preference methods to collect data on hypothetical
choice of alternative fuel vehicles, including battery- powered electric vehicles,
compressed natural gas ( dedicated and dual fuel versions), and alcohol/ flex fuel. The two
2
main goals of the research were to ( i) produce models of “ transaction choice,” based on
the argument that such models could be superior to the more traditional vehicle holdings
models that were in use at that time, and ( ii) support the analysis of policies related to the
introduction of alternative fuel vehicles into the California market. The products of this
research program were used in developing the original version of CARBITS. For
purposes of this Final Report, the original version of CARBITS developed in 2003- 2004
will be referred to as “ CARBITS 1.0.”
ARB staff used CARBITS 1.0 when developing greenhouse gas regulations to meet AB
1493 requirements. Although staff’s use of the model was considered successful, there
was also a desire to upgrade the model to address some perceived “ deficiencies.” Some
issues arose directly from the decision to rely on the earlier research results. For
example, the behavioral models used in CARBITS were based on the panel survey of
California households collected in the mid 1990’ s, so some critics considered the data to
be “ old.” However, most of the motivation for this project was based on experience and
insight gained while developing and using the model. In what follows, we give
additional background on this motivation.
As noted above, CARBITS 1.0 was developed by adaption of pre- existing behavioral
models. A key component was a transactions choice model estimated by a PhD student
at UC Irvine as a major part of her thesis ( Sheng). The original dataset used for
estimating this model was no longer available, so re- estimation or other approaches were
not possible. The most important feature of this model was that it was based on modeling
household- level vehicle transactions using observations collected from the same sample
of households at two points in time. ( In addition, responses from a stated choice
experiment were incorporated.) This model structure required that vehicle market
forecasts be computed using pure microsimulation. Specifically, the model was
populated by a large database of households. Results were obtained by repeated
simulation of individual transaction events, and taking averages. This approach required
very long computer run times. In particular, a very large number of replications are
required to produce results with the required level of smoothness.
In addition to creating something of a burden for staff, the CARBITS 1.0 approach is
vulnerable to criticism from outside consultants. The model is relatively complex and
can be readily misrepresented. For example, auto industry consultants gained access to
CARBITS 1.0 and ( either accidentally or intentionally) generated model runs without
using sufficient simulation replications. They then used the output to claim that
CARBITS 1.0 performs poorly. A related issue is that CARBITS 1.0 follows a practice
of modeling vehicle scrappage as an implicit outcome of choices made in the used
vehicle market. Alternative approaches model vehicle scrappage explicitly, giving the
modeler greater control over how model output is generated.
Other items are more practical, and support the ongoing use of CARBITS for other types
of analysis. The original CARBITS model was put together to meet the immediate needs
of ARB staff. It was calibrated “ by hand” to match vehicle count data corresponding to
the time period of the original survey data. A desirable enhancement would be to create
3
procedures for automated re- calibration of model constants when updated vehicle count
data become available. Finally, in looking ahead to future applications, it is clear that the
recent and ongoing penetration of hybrid electric vehicles could be an important factor in
formulating vehicle- related policies.
The goals for this project as based on the above discussion may be briefly summarized
as:
1. Estimate new vehicle choice models using more recently collected datasets.
2. Address issues of statistical noise and runtimes.
3. Specifically address the issue of vehicle market exit/ scrappage.
4. Develop re- calibration procedures to update certain model constants based on
aggregate- level vehicle counts.
5. Include the capability to address hybrid electric vehicles
With this as background, we give an overview of key decisions and elements of the
project, as an introduction to the remainder of the report.
1. As indicated, CARBITS 1.0
a. Is a transactions model requiring pure microsimulation.
b. Is based on a special- purpose panel survey collected in the mid 1990’ s.
2. Goals for CARBITS 2.0 include
a. Estimating models using more recent data.
b. Reducing statistical noise and run times.
3. The two previous goals can both be met by:
a. Updating CARBITS using the 2000- 2001 Caltrans Statewide Travel
Survey.
b. Converting CARBITS from a transactions model to a holdings model
In this project, various options for updating CARBITS using “ new data” were considered.
This project represented an option to directly address the issues described above.
CARBITS 1.0 was, by necessity, a transactions model. A straightforward update of
CARBITS without any changes to the modeling structure would require a panel data set
with details on vehicle transactions. Although there were some possible data sources to
support this ( i. e., the Consumer Expenditure Survey), the most attractive data set in terms
of sample size and quality is the Caltrans Travel Survey. However, this is a standard
cross- sectional data set ( not a panel data set) and can only support the estimation of a
holdings model. At the same time, the transactions model in CARBITS 1.0 requires pure
microsimulation, which is the source of the run time and statistical noise problems to be
addressed. The decision to adopt the Caltrans Travel Survey and develop holdings
models allows us to adopt the highest quality data, with the largest sample size ( all of
which comes from California), and also eliminates problems with run times and statistical
noise.
4
The other high- level goals for this project ( develop procedures for regular model
recalibration, incorporate vehicle scrappage, expand the model to address hybrid electric
vehicles) have a more general theme: Placing CARBITS on a footing whereby it can be
regularly updated and improved by incorporating new data. In conducting this project,
we strove to take a broader view to address this general theme, i. e., perform activities in
this project to enhance the ongoing viability of CARBITS. In this regard, we approached
the work to update CARBITS in two ways:
1. Designing a generic system/ framework for “ CARBITS- type” models.
2. Identifying and compiling data sources and procedures to support future
updating of CARBITS.
Regardless of the details of our approach, we remark here that the implications for the
data requirements in this project may not have been readily apparent from a discussion of
the high- level goals. The wholesale estimation of new models creates requirements for
vehicle data, not just household data. Specifically, choice models assume that
households make vehicle choices based on vehicle attributes. These include both vehicle
technology characteristics, and vehicle market prices. A substantial amount of effort in
this project was expended on the collection, cleaning, and integration of vehicle data.
Similarly, model calibration and estimation of vehicle market exit rates require data on
the vehicle population at large, at multiple points in time. In this regard, this project also
required the processing and analysis of large DMV data files.
The main body of this report provides more detailed discussion and documentation of
Project Outcomes. Project Outcomes are presented in a series of separate sections. In
accordance with the approach described here, Section 1 presents a generic framework for
what we are calling “ CARBITS- type models.” The basic framework has been
implemented in MATLAB, using principles of object- oriented analysis and
programming. One benefit of this approach is the reusability of computer code, and the
flexibility to easily alter models, update models, create multiple versions of models for
comparison and testing purposes, etc.
Specific frameworks can be defined by adopting a particular set of definitions for model
inputs and outputs. Within a given framework, many different models can be
implemented as long as they use the same inputs and outputs. CARBITS- type models
require input data related to household characteristics and vehicle classes/ attributes. A
critical requirement for this project was to adopt a specific set of Vehicle Class
definitions ( with an identified set of vehicle attributes) to provide a basis for vehicle
demand modeling. Vehicle Class definitions are discussed in Section 2. Section 3
reviews information about the Caltrans Travel Survey data that form the basis for the new
CARBITS 2.0 models. Section 4 discusses the household vehicle demand models
developed for CARBITS 2.0. It provides a review of vehicle choice models, including a
discussion of transaction versus holdings models, and then gives results for the vehicle
holdings models estimated using the Caltrans data. Section 5 gives an overview of DMV
data. Section 6 discusses a vehicle market exit model estimated using DMV data.
Section 7 discusses calibration. Section 8 contains remarks on remaining project issues.
5
Section 9 is the bibliography. Appendix A provides background on database related
issues related to vehicle technology, vehicle prices, and vehicle count data.
PROJECT OUTCOMES
1. A Generic Framework for CARBITS- type Models
The basic function of a CARBITS- type model is to simulate the behavior of the
California personal vehicle market over a specified period of time, and to do so in a way
that will support the analysis of alternative policy scenarios. There are many possible
ways to do this, and a fully documented description of any specific model’s
implementation could be rather technical, and contain a high level of detail. However, it
is also possible ( and helpful) to formulate a generic framework for modeling a “ vehicle
market system” in terms of key components and their relationships. The basic structure
would be applicable across a wide range of models, but at the same time, many of the
technical details might be different, e. g., within a given component. In our work we have
been approaching the development of CARBITS- type models using object- oriented
modeling and programming techniques. Although a full discussion of such methods is
beyond the scope of this report, the idea is that a logical system constructed of “ entities”
( e. g., households, vehicles) and “ relationships” ( vehicle ownership) can be implemented
as modules where the internal detailed workings of the various components are
“ encapsulated.” The model can be continually updated and improved in a variety of
ways with minimal changes to the system. For example, a specific behavioral model
related to household vehicle choice can be changed, improved, etc., by upgrading the
internal workings of a single module. This framework also offers the possibility of
creating multiple alternative models by substitution of modules, and comparing them on
the results they produce. These are capabilities that could be used for future
improvements or research activities. For this project, a single model (“ CARBITS 2.0”)
has been created using this framework.
In this section, we review ( informally) the generic features of what we are now calling
“ CARBITS- type” models. In addition to establishing a framework that can support
ongoing technical development, this provides useful background for later discussion.
The following is a list of basic assumptions underlying CARBITS- type models:
1. The entity that is the source of vehicle market demand is the Household.
2. Total demand in the vehicle market is the result of an aggregation of decisions
made at the individual household level.
3. In each period of a “ market simulation,” households make decisions about their
vehicle fleet. ( The details of what decisions are made, and how, can vary
depending on what type of behavioral model is used.)
6
4. In each period, both new and used vehicles are available in the market.
Manufacturers introduce new vehicle offerings in each model year. New vehicles
purchased in a model year become part of the used vehicle market in later years.
5. Households make decisions on the basis of “ utility maximization,” and have
preference functions that capture their evaluation of vehicles that are available in
the market.
6. Household preferences are formed on the basis of vehicle characteristics,
including a vehicle’s technical specifications and its market price. The fuel
operating cost of a vehicle is based on its fuel economy, but also on the price of
fuel during the period.
7. Household preferences are also a function of household demographics, such as
income, household size, age, etc.
To implement a model based on the above assumptions, the following elements are
required:
1. A Base Calendar Year ( a. k. a., “ Base Year”).
2. A database of Households that represents California for the Base Year.
3. A system for defining Vehicles that represent the unique choice “ options” in the
market. Although vehicles could be defined at the Year- Make- Model level, the
large number of such vehicles makes this impractical. The usual practice is to
define a set of Vehicle Classes to represent the types of vehicles available in the
market.
4. A Vehicle Technology Database that provides vehicle technical specifications
(“ attributes”) and new vehicle prices for Vehicle Class offerings ( typically by
model year). This requires historical data for vehicles available in the Base Year.
In addition, a forecast of available Vehicle Classes and vehicle attributes is
required for future years.
5. A Fuel Forecast specifying fuel prices for the Base Year and all future years
covered by the simulation.
6. A method for “ aging” the Household database to reflect population growth and
shifts in demographic distributions in the future.
7. Behavioral models for representing Household vehicle- related decisions.
8. A method of setting vehicle prices that “ clears the market” that balances vehicle
supply ( new and used vehicles) with Household demand. ( This also includes
scrappage of old vehicles.)
7
Vehicle Market Behavior over a multiple- year time period is “ simulated” by the
following procedure ( which assumes one- year time intervals):
1. For Base Year, initialize:
a. Households
b. Current Market Vehicles
c. Current Vehicle Counts
d. Current Year = Base Year
2. Begin Loop
a. Previous Vehicle Count = Current Vehicle Count
b. Current Year = Current Year + 1
c. Lookup Current Fuel Costs
d. Age Households
e. Update (“ age”) Current Market Vehicles
i. Introduce New Vehicles for Model Year = Current Year
ii. Update Vehicle Characteristics ( e. g., re- compute fuel operating
costs using current fuel prices)
f. Simulate Vehicle Market Behavior for Current Year
g. Summarize Current Vehicle Counts, and report results.
3. Does Current Year = Final Year?
a. If Yes, Stop
b. If No, Go To Step 2
The above procedure is generic, in that it is consistent with a wide variety of specific
model implementations. By adopting a specific set of data elements for key model inputs
and outputs, it is possible to create a well- defined “ platform” for model development and
implementation of multiple CARBITS- type models. Data elements can be selected so
that the same input and output formats can be re- used for a variety of models. For
purposes of this project, we have established conventions for inputs and outputs, and
have implemented a “ CARBITS Vehicle Market Simulation Framework.” The issue of
Model Inputs is discussed in more detail in the next sub- section. The portion of the
process denoted “ Simulate Personal Vehicle Market Behavior for Current Year”
represents a “ module” that can be implemented using, e. g., different types of household
behavioral models. This module can be further decomposed into additional sub- modules
that address such questions as how household vehicle- related decisions will be modeled
( e. g., as transaction choices, holdings choices, etc.), how the market is cleared, prices
changed, etc.
Behavioral models in this CARBITS framework are based on household- level survey
data. The availability of data and other considerations have an effect on the Base Year
and options for specific behavioral models. This is briefly addressed in sub- section 1.2,
as well as other parts of this report.
1.1 Model Inputs
8
In the current implementation of CARBITS, the main inputs that are typically used for
policy analysis are the Vehicle Technology Database ( VehTechDB) and the Fuel
Forecast. The VehTechDB includes historical data on vehicles corresponding to the
Base Year. However, scenario analysis is based on simulating how the future vehicle
market will behave in response to changes in regulations. This requires the user to
provide a forecast of vehicle technology offerings for future years. In many cases,
regulations might require vehicle manufacturers to change their offerings. If so, this must
be reflected in the model inputs provided by the user. The model then simulates how the
market would behave under this scenario.
A key design issue for a CARBITS- like model is the definition of Vehicle Classes, and
the identification of vehicle attributes to be included in the VehTechDB. Deciding on
these elements is important, because they represent the only information that can be used
as inputs to Vehicle Demand Models. Section 2 discusses the Vehicle Class definitions
adopted for CARBITS 2.0. In addition to Vehicle Class, household vehicle choice is
assumed to depend on three attributes: Market Price of the vehicle, Fuel Operating Cost
( in cents per mile), and Acceleration ( seconds for 0 to 60 miles per hour). Fuel Operating
Cost in any given year is computed from Fuel Economy and the Fuel Cost for that year
( provided in the Fuel Forecast).
Although this may seem to be a straightforward proposition, rigorously establishing
vehicle attributes for each Vehicle Class requires a procedure for aggregating data from
the large number of individual makes and models that are available in the market.
Generally speaking, weighted averages of attributes are required, which in turn requires
data on the distribution of vehicles in the market. This project required integration of
data from the following sources:
Chrome VINMatch data
Chrome New Vehicle Data ( NVD)
National Automobile Dealers Association VINPrefix Solution
California Department of Motor Vehicles ( DMV) registration data
California Bureau of Automobile Regulation ( BAR) Smog Check data
EPA Fuel Economy Guide
Wards Automotive Yearbook Vehicle Specifications
The commercially available data sets are sources of vehicle specification and market
price data. The DMV and BAR data provide weighting information to allow attributes to
be averaged over vehicle classes. In addition, the DMV data provide information on
actual vehicle counts in the California fleet, and data on the rate at which vehicles exit the
market. Appendix A provides more details about these data sources.
1.2 Base Year, Household Data, and Models
The nominal Base Year for CARBITS 2.0 is 2001, which corresponds to the household
database used for this project: The 2001 Caltrans Travel Survey. These data are used as
the Household database in the above simulation framework, and, in addition, were used
9
for estimating the household vehicle demand behavior models used in Step 6. The
database is discussed in Section 3, and details on the behavioral models are given in
Section 4.
2. Vehicle Class Definitions and Attributes
To begin, we review the vehicle classification scheme from CARBITS 1.0:
Type Size
1. Car Mini
2. Car Subcompact
3. Car Compact
4. Car Intermediate
5. Car Large
6. Car Luxury
7. Car Sports ( or, “ Sports car”)
8. Pickup Compact
9. Pickup Standard
10. Van Compact ( or, “ Minivan”)
11. Van Standard
12. Sport utility vehicle Small
13. Sport utility vehicle Large
14. Sport utility vehicle Mini
Table 2.1 CARBITS 1.0 Body Type and Size Classes
CARBITS 1.0 uses this classification scheme because it was based on a model developed
by an Irvine- Davis ITS team under a program sponsored by the California Energy
Commission ( CEC). This was the classification schemed used by the CEC at that time in
their CalCars model. At the time, a substantial amount of effort had been expended in
structuring vehicle technology data ( i. e., attributes, prices, etc.) according to this
framework, both in a historical context as well as in the form of technology forecasts. In
addition, the CEC had a substantial investment in generating DMV vehicle counts using
this framework.
One main concern with this approach is that it represents a market structure that, while
appropriate in the 70’ s and 80’ s, might no longer be an adequate representation.
Specifically, during that period in history the term “ luxury car” was generally associated
a type vehicle with a particular set of characteristics and a well- established image in the
minds of consumers. These vehicles were generally larger than other vehicles, and much
more expensive with certain types of interior features. Representative vehicles would be
the offerings from nameplates such as Cadillac, Lincoln, and Mercedes. The market is
now more differentiated so that each size class has both “ high- end” and “ low- end”
vehicles. The high- end vehicles are typically represented by a more “ prestigious” brand
name, have higher performance characteristics ( and lower fuel economy), and are more
10
expensive. In our approach, we have adopted the term “ Prestige” ( rather than “ Luxury”)
to characterize these high- end vehicles.
A similar, overlapping concern has to do with the use of the term “ sports car.” Finding
an objective standard to classify vehicles into this category is problematic, and this term
no longer means what it once did. There are also challenges associated with vehicles in
the “ Mini” ( or, “ Mini- subcompact”) category. In the range of years for which data are
currently available for updating CARBITS, there has been very low demand for these
vehicles. ( However, it seems likely that this class will be making a comeback in the near
future.)
This project represented an opportunity to re- examine these issues related to vehicle
classification, because a number of the project goals were already going to require the
type of data collection that could support the development and testing of alternative
vehicle classification schemes. Having said this, once the data had been collected and
reviewed, a greater appreciation for the practical issues associated with vehicle
classification became apparent. In what follows, we review some of the details related to
vehicle classification that were explored for this project.
2.1 Issues to Consider when Classifying Vehicles
The notion of vehicle classification can be tricky, since the concept relates both to a
consumer’s conception of what a vehicle “ is” and what it can be used for ( which drives
vehicle demand), and the physical and technological features that a vehicle may
incorporate. The latter relate to a number of issues, including the basis for how
regulations are formulated, and how vehicles can be characterized in terms of attributes in
quantitative demand models. After a detailed review, we came away with a greater
appreciation of the practical role that data availability can play in formulating vehicle
classification schemes. Briefly stated, we have adopted a scheme whereby vehicles are
characterized along three dimensions:
1. Body Type
2. Size
3. Prestige
We also consider the issue of hybrid electric vehicles, but this will be addressed
elsewhere. Because CARBITS must address both used and new vehicle markets, there
will also be a vintage/ age dimension. In what follows, Body Type and Size will be
discussed together.
2.1.1 Body- Type- Size Classes
For our purposes, “ Body Type” refers to the physical configuration of a vehicle whereby
it has a specific type of general functionality. For historical reasons, there is now a strong
bifurcation between two basic configurations: Passenger Car, and Light- Duty Truck
( LDT).
11
Passenger cars can be subdivided in a number of ways according to “ Body Style” ( e. g.,
sedan, hatchback, coupe) where the most important differences occur in the case of
station wagons, two- seaters ( roadsters), and perhaps convertibles. In our work we
collected data at the level of body style, but decided that the following three categories
represented the most fundamental distinction in terms of functionality: Car, Station
Wagon, and Two- seater. We also considered “ convertible,” but with very few exceptions
convertibles overlapped heavily with Two- seaters.
Light- Duty Trucks are now generally sub- divided into Pickups, Vans, and ( Sports) Utility
Vehicles ( SUVs). In terms of functionality, there is a clear difference between Pickups,
which have an open bed and limited seating, versus Vans and SUVs, which are enclosed
and have more seating but can also be re- configured to one degree or another for carrying
cargo. The SUV has other distinguishing features that might be more related to a type of
product image that appeals to a particular type of consumer. In considering specific
makes and models of vehicles over time, there can be some ambiguity in how to classify
certain vehicles based on their physical configurations, since many could qualify either as
a station wagon, a minivan, or an SUV. Most recently, Crossover vehicles have created
additional confusion.
It turns out that the above discussion combined with other issues ( including data
availability) has led us to a vehicle classification scheme that is not dramatically different
from CARBITS 1.0, or others used in the academic literature ( and for similar reasons).
In general, the basis for most of these is a vehicle classification scheme that has long
been used by EPA, which interacts the Body Types discussed above with some particular
definitions of Size. ( In addition, LDTs are divided into 2- wheel drive and 4- wheel drive
versions). The full EPA scheme has changed some over the years: Prior to 1998 the non-
Pickup LDTs that would generally be classified today as SUVs or Vans were
characterized as “ Special Purpose Vehicles.” The terms “ Sport Utility Vehicle” and
“ Minivan” were introduced in 1998 as a substitute.
Another factor is that a major source of vehicle attributes for this project, the Chrome
databases ( see Appendix) uses a MarketClass variable that is a slight extension of the
EPA Class ( it adds in the number of passenger doors for cars, i. e., 2 or 4), and, most
importantly, it appears to maintain complete consistency with the EPA data. In our work,
we begin with Body- Type- Size Definition 1 (“ BTS1”) classes that are based on EPA and
Chrome. See Table 2.2. Differences are: ( 1) doors and drive train information are
removed, and ( 2) Special Purpose Vehicles prior to 1998 are re- classified as Minivans or
SUVs.
12
EPA/ Chrome = BTS1 BTS2 BTS3 ( CARBITS 2.0)
Two- seater Passenger Car Two- seater 1. Two- seater
Mini- Compact Passenger Car Mini- compact Car 2. Small Car
Sub- Compact Passenger Car Subcompact Car
Compact Passenger Car Compact Car 3. Compact Car
Small Station Wagon Small SW
Midsize Passenger Car Midsize Car 4. Midsize Car
Midsize Station Wagon Midsize SW
Large Passenger Car Large Car 5. Large Car
Large Station Wagon Large SW
Small Pickup Trucks Small Pickup 6. Small Pickup
Standard Pickup Trucks Standard Pickup 7. Standard Pickup
Minivans* Minivans* 8. Minivan
Large Passenger Vans Large Passenger Vans 9. Full- size Van
Cargo Vans Cargo Vans
Sport Utility* Small SUV 10. Small SUV
Midsize SUV 11. Midsize SUV
Large SUV 12. Large SUV
Table 2.2 Development of Body- Type- Size ( BTS) Definitions for CARBITS 2.0
One major issue with EPA/ Chrome/ BTS1 classes is that SUVs are not assigned to size
classes in these published databases. However, EPA frequently must address vehicle size
issues in various publications. For example, the following definitions appear in EPA
( 2007, page 5):
Small Midsize Large
Pickup < 105” 105” to 115” > 115”
Van < 109” 109” to 124” > 124”
SUV < 100” 100” to 110” > 110”
Table 2.3 Wheelbase- based Size Definitions for Light- Duty Trucks
Note that defining the size of Pickups based on wheelbase is a different approach from
EPA’s classification system— see column 1 of Table 2.2. In BTS1, Pickups are classified
as Small and Standard Pickups based on gross vehicle weight rating ( GVWR). In the
standard EPA classification system, Vans are classified into Minivans, Large Passenger
Vans, and Cargo Vans ( based on definitions that we have as yet been unable to locate).
Also, we remark that the definitions in Table 2.3 were taken from a 2007 EPA
publication, but that these values could be different in publications from other years.
Definition BTS2 in Table 2.2 is obtained by adding SUV size classes to BTS1 based on
the definition in Table 2.3. Definition BTS3 is obtained by merging together some BTS2
classes to obtain fewer categories. BTS3 generally looks like other classifications found
in the literature, and it is based on similar concerns and considerations:
1. We have elected to merge Large Passenger Vans and Cargo Vans into the
more generic “ Full- Size Van.” Two reasons for this are: ( 1) the total demand
13
for these vehicles by households is rather small, ( 2) based on only make and
model information is very difficult to distinguish between these two when
working with most data sets.
2. In the choice modeling literature there has almost always been question of
what to do about station wagons. Although they have some functional
differences with, e. g., sedans, the sales volumes for station wagons are
relatively small. Including them increases the number of categories. We
adopted the usual practice of merging station wagons with standard cars of
similar size in order to reduce the number of categories.
3. Mini- Compact cars have been absorbed into Small cars. ( As discussed
previously, demand for minicompacts has been extremely small for many
years. Essentially all published choice models typically eliminate these as a
separate class.)
3. Two- seater has been preserved as a separate class. It is an easily
identifiable physical characteristic ( in contrast to an image- based concept) that
generally couples small size with a significant configuration feature ( limited
seating and luggage space) that is easier to identify than the less- well- defined
concept “ sports car.”
2.1.2 Prestige
For this project we elected to define Prestige on the basis of vehicle brand name,
incorporating the notion of “ brand equity” frequently used in the marketing literature.
Certain brand names are clearly associated with an image that incorporates a combination
of such things as quality, reputation, a consistently high level of amenities and features
offered as standard equipment, etc. One advantage of this approach is that it represents
an “ attribute” that is easily identifiable and readily assigned to each vehicle. Moreover,
vehicles grouped together using this dimension share a number of similarities, resulting
in more homogeneous groups ( see discussion below). Finally, it generalizes the concept
of “ luxury” that previously was assigned to a very specific type of vehicle. One
unfortunate, but unavoidable complication of this dimension is a higher degree of
correlation between purchase price and other attributes ( e. g., fuel economy and
performance), which can complicate model estimation ( see section 4).
Another dimension under consideration was “ Country/ Region of Manufacturer” ( e. g.,
“ Domestic versus Foreign,” or, “ Domestic- Asia- Europe”). There is little doubt that this
dimension can have some explanatory power. Many years ago, studies seemed to support
the idea that domestic consumers would prefer to “ buy American” all else equal.
Unfortunately, in more recent years this dimension as become convoluted with
“ reputation for quality” ( see Train and Winston 2007), with many foreign manufacturers
having a reputation for higher quality than their domestic competitors. Moreover, the
foreign- domestic distinction has become less clear, with the advent of foreign
14
manufacturers locating manufacturing plants in the U. S., and domestic manufacturers
importing some of its product lines.
Prestige Brands Region
Domestic Europe Asia Total
Acura 13.50% 13.50%
Audi 1.50% 1.50%
BMW 11.60% 11.60%
Cadillac 14.00% 14.00%
Infiniti 6.60% 6.60%
Land Rover 1.90% 1.90%
Lexus 15.10% 15.10%
Lincoln 9.80% 9.80%
Mercedes Benz 18.60% 18.60%
Saab 1.20% 1.20%
Volvo 6.20% 6.20%
23.90% 40.90% 35.20% 100.00%
Non- Prestige Brands Region
Domestic Europe Asia Total
Buick 2.80% 2.80%
Chevrolet 11.70% 11.70%
Chrysler 2.00% 2.00%
Dodge 5.90% 5.90%
Eagle 0.20% 0.20%
Ford 21.30% 21.30%
Geo 1.00% 1.00%
GMC 0.70% 0.70%
Honda 11.70% 11.70%
Hyundai 0.80% 0.80%
Isuzu 0.60% 0.60%
Jeep 2.10% 2.10%
Mazda 2.80% 2.80%
Mercury 2.40% 2.40%
Mitsubishi 1.80% 1.80%
Nissan 6.20% 6.20%
Oldsmobile 2.30% 2.30%
Plymouth 1.80% 1.80%
Pontiac 2.50% 2.50%
Saturn 2.40% 2.40%
Subaru 0.20% 0.20%
Suzuki 0.10% 0.10%
Toyota 14.80% 14.80%
Volkswagen 2.00% 2.00%
57.90% 3.00% 39.10% 100.00%
Table 2.4 Distribution of Vehicles by Manufacturer ( Classified by Prestige versus
Region) in the California Personal Vehicle Fleet ( October 2001)
15
Table 2.4 explores the two dimensions “ Prestige” and “ Region” on the basis of vehicle
count distributions in California in Fall 2001 ( October). These figures are based on
October 2001 DMV data that were assembled to match the timeframe of the most recent
Caltrans Travel Survey, and are intended to reflect the personal vehicle market— see
Section 3. In this table, we have included breakdowns by region of origin, and report the
percentage of the California vehicle fleet within each category ( Prestige versus Non-
Prestige) for model years 1989- 2002. Prestige vehicles made up about 15% of the
California fleet.
The percentages of Prestige versus Region are highly correlated. Domestic vehicles
made up 58% of the non- Prestige fleet, but only 24% of the Prestige fleet. European
vehicles had the largest share of the Prestige fleet ( 41%), and essentially none of the non-
Prestige fleet ( 3%). It is important to note that, since these figures pool together model
years 1989- 2002, they do not illustrate more recent trends in Domestic versus non-
Domestic new vehicle sales. However, even in 2001, the percentage of Lexus vehicles
on the road had reached 15%, second only to Mercedes.
2.2. CARBITS 2.0 Vehicle Classes ( Historical)
Taken together, Tables 2.2 and 2.4 illustrate some of the challenges in developing vehicle
choice models for practical use in policy analysis. BTS1 includes 17 body- type- size
classes. If one were to include ten vehicle manufacturers and 20 model years, the total
number of make/ vehicle- class/ vintage combinations would be 17 x 10 x 20 = 3,400.
( This is for gasoline vehicles only, i. e., it ignores the “ dimension” of fuel/ fuel technology
type. Moreover, using the model to evaluate the impact of policies 20 years into the
future requires forecasts of vehicle classes and attributes over this range of years.
Determining the level of detail required for policy analysis is always a difficult judgment
call.
The Vehicle Classes adopted for CARBITS 2.0 ( for the case of historical data) are
represented in Table 2.5. The table is based on scenario requirements for estimating
choice models using Caltrans Travel Survey data, where the vehicle model year window
begins in 1982 and goes through 2001. Certain Vehicle Classes do not exist over the full
range of years ( 1982- 2001). See Table 2.5. All Car types have both Non- Prestige and
Prestige versions over the entire range of years; however, there are no Prestige Pickup
Trucks or Minivans. There are no Midsize or Large SUVs included prior to 1985.
Prestige SUVs begin in 1996. There are 350 combinations in all. Note: In reality, there
are very small numbers of some vehicle types in some years that are not included in this
table. However, they have been eliminated for modeling purposes.
The main purpose of defining vehicle classes is to provide a structure for modeling
vehicle choice. Consumer choice of a vehicle class as defined in Table 2.5 is based on
preference for vehicle configuration, size, prestige level, and also vintage. However,
vehicle classes will also vary on other important attributes. Chief among these are
market price, fuel operating cost, and performance. These would be expected to vary
across vehicle class. This is illustrated next.
16
BTS3 Non- Prestige Prestige
1. Two- seater All Years* All
2. Small Car All All
3. Compact Car All All
4. Midsize Car All All
5. Large Car All All
6. Small Pickup All [ None]
7. Standard Pickup All [ None]
8. Minivan All [ None]
9. Full- size Van All [ None]
10. Small SUV All 1996- 2001
11. Midsize SUV 1985- 2001 1996- 2001
12. Large SUV 1985- 2001 1998- 2001
Table 2.5 CARBITS 2.0 Vehicle Classes (* 1982- 2001)
2.2.1. Vehicle Attributes for CARBITS 2.0 Vehicle Classes ( Historical)
This subsection reviews historical patterns of vehicle attributes for the Vehicle Classes
defined previously. As has been noted, the key attributes used for consumer choice
modeling in this project are market price, fuel operating cost, and performance. When
consumers decide to make a vehicle purchase, they take possession of a specific year-make-
model vehicle with well- defined physical characteristics. However, estimating
choice models at Vehicle Class level does not support this level of detail, and requires
representative attribute values that are typically obtained by taking averages over the
individual vehicle offerings in a class. ( Usually these are sales- weighted averages.)
There are many issues and details associated with the construction of Vehicle Technology
databases that are too numerous to discuss here. This information is included in
Appendix A. However, we provide some very brief remarks here:
1. Market price data for this study come from the National Automobile Dealer
Association ( NADA) VIN Prefix solution. These data include estimates of
market prices for both new and used vehicles for a particular month and year, at
the level of an individual VIN Prefix ( which captures information on make,
model, style, engine, and other characteristics). See Appendix A.
2. Because fuel operating cost ( measured in cents per mile) is a function of both fuel
efficiency ( mpg) and fuel price ($ per gallon), the relevant vehicle technology
variable is fuel efficiency. The original source of mpg ratings is the EPA fuel
economy guide data, which are also replicated in other vehicle specification
databases. EPA provides three ratings: city, highway, and combined. When
representative values are called for, we used the combined mpg estimate.
3. There are many possible choices for measuring vehicle performance, including:
horsepower, horsepower- to- weight ratio, top speed, etc. In this project, we use a
17
measure called “ EPA_ 0_ 60,” i. e., time ( in seconds) to accelerate from 0 to 60
miles- per- hour. However, this is not a direct measure. This measure is
computed using a formula from an EPA publication that converts horsepower- to-weight
ratio into an estimated acceleration time. The measure is computed at a
high level of detail, requiring knowledge of the transmission type. These figures
are then averaged, as discussed in Appendix.
Average market prices as a function of Model Year in December 2001 for various
combinations of Vehicle Classes are shown in Figures 2.1 and 2.2. Figure 2.1 gives
average market prices by major body type ( Car, Pickup, Van, SUV). Curves for Car and
SUV are similar to one another from 2001 to 1996, as are Pickups- Vans. For earlier
model years SUV prices drop to a point intermediate between Cars and Pickups/ Vans.
As model years get older, prices for all body types converge.
Figure 2.2 gives more detail on market prices to illustrate a point. In this figure, vehicles
are further divided into Prestige versus Non- Prestige. There are no Prestige Pickups or
Vans. The only Prestige SUVs begin in model year 1996, which explains the pattern in
Figure 2.1. With the additional level of detail in Figure 2.2, it can be seen that prices for
Non- Prestige Cars, Pickups, and Vans are similar to one another, and Non- Prestige SUVs
are priced a bit higher. There is a substantial gap between Prestige and Non- Prestige
vehicles, with Prestige Cars and SUVs having similar prices from 1996- 2001.
18
Figure 2.1 Ave. Market Prices by Body Type and Model Year ( December 2001)
Figure 2.2 Ave. Market Prices by Body Type/ Prestige Level and Model Year ( December
2001)
2.2.2 Fuel Economy
Figure 2.3 shows average fuel economy for Body Type/ Prestige level by model year. On
average, the often- stated observation that fuel economy has remained relatively flat for a
wide range of years is illustrated by this figure. The level of detail in Figure 2.3 also
illustrates some other features of fleet fuel economy. For 1985, Non- Prestige Cars have
the highest combined MPG, followed by Pickups, Prestige Cars, Vans, and Non- Prestige
SUVs, respectively. In all years, the average fuel economy for Non- Prestige Cars is
substantially higher than the light duty trucks, and also Prestige Cars. Prestige Car fuel
economy lies below Pickups and above Vans until about 1995, when the steady
downward trend in fuel economy for Pickups creates a crossover. The fuel economy of
SUVs is well below the rest of the fleet.
19
Figure 2.3 Average MPG ( Combined) by Body Type/ Prestige and Model Year
2.2.3 Performance
Average performance ( measured by EPA_ 0_ 60) for Body Type/ Prestige groupings by
Model Year are given in Figure 2.4. In contrast to fuel economy, there is a noticeable
upward trend in Performance ( downward trend in 0- 60 time) for most vehicle types, and
a clear separation between Prestige Cars and all other vehicle types. Figures 2.3 and 2.4
illustrate an often- discussed issue in policy debates: Given available fuel technology,
there is generally a tradeoff between fuel economy and performance, and in recent years
advances in fuel technology are used primarily to improve performance while leaving
fuel economy relatively flat.
20
Figure 2.4 Average Performance by Body Type/ Prestige and Model Year
3. Caltrans Travel Survey
The main household database used for updating CARBITS in this project is the 2000-
2001 California Statewide Travel Survey, which we will frequently refer to as the
“ Caltrans Travel Survey,” or the “ Caltrans Survey.” The main reference is the survey’s
Final Report— see Bibliography. For purposes of background, the following is an excerpt
from the Executive Summary of the Final Report:
The California Department of Transportation ( Caltrans) maintains a statewide
database of household socioeconomic and travel information, which is used in
regional and statewide travel demand forecasting. The most recent database, prior
to this survey, contained data from the last statewide survey that was conducted in
1991. The 2000- 2001 California Statewide Household Travel Survey was
21
conducted to update the database and will be used to help refine travel estimates,
models, and forecasts throughout the State. The resultant data set will be used to
estimate and forecast trip generation and distribution, mode choice, and
assignments, as well as for vehicle emissions analyses and estimates.
The 2000- 2001 survey was conducted between October 2000 and December
2001 among households located in each of the 58 counties throughout the State. A
total of 17,040 households participated in the survey. Household socioeconomic
data gathered in this survey includes information on household size, income,
vehicle ownership, employment status of each household member, and housing
unit type among other data. Travel information was also collected including trip
times, mode, activity at location, origin and destination, and vehicle occupancy
among other travel- related data. [ Emphasis added.]
As discussed in previous parts of this report, the Caltrans survey has a large sample size,
follows careful data collection procedures, and provides weight factors that make it an
attractive option for our purposes. The items in bold above are the main elements
required for vehicle choice modeling using “ revealed preference” data. Table 3.1
reproduces key household statistics from the survey’s final report.
Household Vehicles Available 21,448,770
Vehicles in Use on Average Weekday ( 71%) 15,252,463
Full- time Employees 10,130,359
Licensed Drivers 19,696,497
Occupied Housing Units 11,502,870
Single Housing Units 68%
Multiple and Other Housing Units 31%
Median Household Income $ 54,946
Persons Per Household 2.8
Vehicles Per Household 1.9
No Vehicles 9.3%
One Vehicle 29.7%
Two Vehicles 37.7%
Three or More Vehicles 23.4%
Licensed Drivers Per Household 1.7
Table 3.1 Key Household Statistics from
2000- 2001 California Statewide Household
The survey methodology includes the development of household weights that, when
applied, provide a way to compute statistics ( as in Table 3.1) that represent the entire
California population. In particular, the weights are chosen so that certain statistics
match those of the 2000 Census— see Chapter 6 of the Caltrans Survey Final Report.
3.1 Caltrans Survey Data Tables
Following standard database management practices, the data set is sub- divided into
separate tables that correspond to three key entities: Households, Persons, and Vehicles.
In this form, information is stored in a way that avoids inefficient replication of data
elements. The three tables are linked together through a household id number ( SAMPN).
Documentation on selected variables from the Household and Vehicle tables is replicated
22
in Tables 3.2 and 3.3, respectively. Important Household variables for choice modeling
include income ( INCOME), household size ( HHSIZE), and number of workers
( NWORK)— see Section 4. Identification of household ownership levels and
characterization of vehicle holdings on the basis of body type, year, make, and model are
also important, and present a number of practical challenges ( to be discussed). The
Persons table ( not shown here) contains details for individual household members,
including age, occupation, educational level, etc. The next sections explore data issues in
more detail.
Var Name Variable Description Data
Type Width Values
RECTYPE Record Type N 1 1= Household Data
SAMPN HH ID Number N 7 Assigned unique identifier
HHSIZE Number of persons in household N 2 Ordinal Variable
TOTVEH Number of motorized vehicles
available for use by HH members N 2 Ordinal Variable
OWN Owner/ Renter Status N 1 1= Own; 2= Rent; 7= Other, 8= DK, 9= RF
INCAT Income Category N 1 1= Above 50K; 2= Below 50K; 9= DK/ RF
INCOME Total 1999/ 2000 annual household
income N 2
1=<$ 10,000; 2=$ 10,000-$ 24,999; 3=$ 25,000-
$ 34,999; 4=$ 35,000-$ 49,999; 5=$ 50,000-
$ 74,999; 6=$ 75,000-$ 99,999; 7=$ 100,000-
$ 149,999; 8=$ 150,000+; 9= DK/ RF
NWORK Number of HH Workers N 2 Ordinal Variable
NSTUD Number of HH Students N 2 Ordinal Variable
WDWGT Weekday Weight N
Table 3.2 Selected Household Variables from Caltrans Survey
23
Var Name Variable Description Data
Type Width Values
RECTYPE Record Type N 1 3= Vehicle Data
SAMPN HH ID Number N 7 Assigned unique identifier
VEHNO Vehicle Number N 2
MAKE Vehicle X - Make C 2
1= Acura; 2= Audi; 3= BMW; 4= Buick; 5= Cadillac;
6= Chevrolet; 7= Chrysler; 8= Dodge; 9= Ford; 10= Geo;
11= GMC; 12= Harley Davidson; 13= Honda;
14= Hyundai; 15= Infiniti; 16= Isuzu; 17= Jaguar; 18= Jeep;
19= Kawasaki; 20= Kia; 21= Lexus; 22= Lincoln;
23= Mazda; 24= Mercury; 25= Mercedes- Benz;
26= Mitsubishi; 27= Nissan; 28= Oldsmobile;
29= Plymouth; 30= Pontiac; 31= Porsche; 32= Range
Rover; 33= Saab; 34= Saturn; 35= Subaru; 36= Suzuki;
37= Toyota; 38= Volkswagen; 39= Volvo; 40= Yamaha;
41= Daewoo; 42= Dotson; 43= International;
44= Winnebago; 45= MG; 97= Other, specify; 98= Don't
know; 99= Refused
O_ MAKE Other make C 60
MODEL Vehicle X- Model C 60
YEAR Vehicle X - Year F 4 8888= Don't know; 9999= Refused
BTYPE Vehicle X - Body Type N 2
1= Auto; 2= Van, 3= RV; 4= Sport utility vehicle; 5= Pick- up
truck; 6= Other truck; 7= Motorcycle/ Moped; 97= Other,
specify; 99= DK/ RF
WDWGT Weekday Weight N
Table 3.3 Selected Vehicle Variables from Caltrans Survey
3.2 Caltrans Household Income Distributions
Household income distributions from the Caltrans Survey are presented in Table 3.4.
The first columns of the table report distributions based on the un- weighted sample of
17,040 households. The final three columns show the same figures computed using the
weights developed to match Census data to represent the 11.5 million households in
California at that time. The table illustrates some common features of this type of survey
work: Households at the lowest and highest income levels are frequently under- sampled,
and many households ( 12- 13% in this case) refuse to provide income information.
Unweighted Weighted
Freq Percent
Valid
Percent Freq Percent
Valid
Percent
<$ 10,000 732 4.3 4.9 984705 8.6 9.7
$ 10,000-$ 24,999 2419 14.2 16.3 2003837 17.4 19.7
$ 25,000-$ 34,999 2244 13.2 15.1 1113007 9.7 11
$ 35,000-$ 49,999 2369 13.9 15.9 1297487 11.3 12.8
$ 50,000-$ 74,999 3389 19.9 22.8 1774103 15.4 17.5
$ 75,000-$ 99,999 1850 10.9 12.5 1103269 9.6 10.9
$ 100,000-$ 149,999 1268 7.4 8.5 1103019 9.6 10.9
$ 150,000+ 583 3.4 3.9 775768 6.7 7.6
Total 14854 87.2 100 10155194 88.3 100
Don't Know/ Refused 2186 12.8 1347671 11.7
17040 100 11502866 100
Table 3.4 Household Income Distributions in the Caltrans Travel Survey
24
3.3 Vehicle Holdings
Another distribution of interest is the level of vehicle holdings by households. Despite
the reference to “ vehicle ownership” in the Executive Summary of the Caltrans Final
Report, note that the survey generally relies a related measure termed “ vehicle
availability”, i. e. the variable TOTVEH ( Number of motorized vehicles available for use
by HH members)— see Table 3.2. Using this variable in conjunction with weights yields
the statistics in Table 3.1. An expanded distribution is given in Table 3.5. By this
measure, fewer than 10% of California households have no motorized vehicles available
( 3.5 % of the sample). About 68% of households ( 73% of the sample) hold one or two
vehicles. The mode in California is two- vehicle households.
Unweighted Weighted
No. of Vehicles Frequency Percent Percent
0 601 3.5 9.3
1 5123 30.1 29.7
2 7343 43.1 37.7
3 2742 16.1 16
4 861 5.1 4.9
5 237 1.4 1.5
6 81 0.5 0.6
7 32 0.2 0.2
8 13 0.1 0.1
9 7 0 0
Total 17040 100 100
Table 3.5 “ Vehicle Availability” Distribution for Caltrans Survey Households
( see text for definition of vehicle availability)
However, one potential issue for this project is that “ availability of motorized vehicles” is
not necessarily equivalent the choice of “ vehicle holdings” that we are concerned with,
i. e., the household’s light- duty vehicles. Specifically, in the Caltrans Survey “ motorized
vehicles” includes motorized vehicles of all types, as indicated in the text of the survey
question:
Question 19: “ How many vehicles are presently available to members of your
household? This includes all cars, vans, trucks, RVs, SUVs, motorcycles and
mopeds, whether owned or leased or provided by an employer.”
In contrast, consider the wording of the vehicle question used in the 2000 Census:
Question # 43: “ How many automobiles, vans, and trucks of one- ton capacity or
less are kept at home for use by members of your household?” There are seven
possible responses to this question ranging from “ none” to “ 6 or more.” Note
that this question does not ask about “ vehicle ownership” per se, but about
vehicles “ kept at home” whether they are owned, leased, borrowed or company
vehicles.]
25
The Census definition more closely matches the definition of vehicle holdings we are
developing choice models for. However, comparing these two definitions raises a
potential question about the validity of the weights in the Caltrans Survey, because it
appears that the weights were constructed under the assumption that the two definitions
are the same.
Another issue we faced in working with the Caltrans data was our discovery that the
vehicle data were “ dirty” in a number of ways, as can happen in surveys of this type.
Relevant vehicle variables used in this project include body type, year, make, model, and
fuel type of household vehicles— see Table 3.3. Problems we encountered included:
1. Item non- response, i. e., missing items ( Don’t Know or Refused) in variables for
Year, Make, or Model of vehicle.
2. Limited information in Model variable ( e. g., “ Car” rather than the actual model
name).
3. Errors in data entry, as evidenced by:
a. Miss- matches between Make and Model ( e. g., Nissan Camry).
b. Miss- matches between stated body type and other variables. ( For
example, the body type could be listed as “ Moped” for a 1999 Toyota
Camry.)
c. Miss- spelled model names, creating difficulties in vehicle matching.
d. Miss- matches between year and model ( e. g., a 1985 Toyota Prius does not
exist, so there is a miss- match between year and make/ model).
In addition, there were a relatively large number of very old vehicles in the data set. This
can happen in a survey of this type due to sample response bias, e. g., individuals with a
strong interest in cars might be “ collectors,” and would also be more likely to respond to
the survey. For our work, we limited the “ window” for vehicles to the 20- year period
1981- 2001 for purposes of choice modeling ( see Section 4). Constructing a data set to be
used for choice model estimation requires that vehicles in the Caltrans Survey be
‘ identified’ in enough detail to assign them to the vehicle classes discussed in Section 2.
So, even though there were problems in exactly matching vehicles at the Year- Make-
Model level, we established procedures to assign vehicle classes using available
information. This is discussed in more detail in section 3.4. For now, we summarize
some facts about the Caltrans vehicle data.
For a summary of vehicles successfully matched to vehicle technology data on the basis
of Year, Make, and Model information for model years 1981- 2002, see Table 3.6. The
table is constructed using the Caltrans survey weights, indicating that vehicles
representing 17.7M of the 21.4M ( 83%) are successfully matched. Data are presented in
cross- tab form to highlight some of the data quality issues. Specifically, the “ matched
body type” is the body type from the vehicle technology database, whereas “ btype” is the
body type recorded in the survey data. Although they are highly correlated, they
frequently disagree. In some cases the disagreements are significant, e. g., cases where
Cars are assigned a body type ( btype) of “ moped/ motorcycle” or “ RV”.
26
Matched Body Type
btype* Car Pickup Van SUV Total
Auto 10,413,401 95,413 110,645 162,297 10,781,756
Pickup 24,167 2,831,608 14,356 39,922 2,910,053
Van 62,225 24,920 1,589,550 10,270 1,686,965
SUV 194,625 61,410 10,151 1,852,022 2,118,208
Other truck 14,584 61,112 8,111 67,316 151,123
RV 4,279 474 2,801 14,194 21,748
Moped/ Motorcycle 20,740 161 1,724 1,702 24,327
Other 5,368 442 5,810
DK/ Ref 4,533 4,533
Total 10,743,922 3,075,098 1,737,338 2,148,165 17,704,523
Table 3.6 Successfully Matched Caltrans Vehicles ( 1981- 2002)
* btype variable from Caltrans Survey
Table 3.7 summarizes the status of unmatched Caltrans vehicles, and illustrates various
data issues. There are a number of ways to look at these figures. First, if we omit
concerns about the unreliability of the btype variable, this Table yields an estimate of 3M
Autos, Pickups, Vans, and SUVs that are not included in Table 3.7, for a total of 20.7M
light- duty vehicles out of the 21.4M “ available vehicles,” or about 97%. So, it may be
using “ available motorized vehicles” to represent “ vehicle holdings of light- duty
vehicles” is a reasonable approximation. About half of these 3M vehicles ( 1.5M, or 7%
of the total) are excluded from Table 3.6 because they are older vehicles ( model year <
1981). A relatively small number ( 500K, or 2%) are unmatched due to a missing model
year. In all, the light duty vehicle fleet with model years 1981- 2002 is estimated to lie in
the range 18.6- 19.1M vehicles, of which we have matched 17.7M ( approx. 95%).
YearFlag
btype DN/ REF 1981- 2002
1965-
1980 < 1965 Total
Auto 331,794 508,128 706,063 153,028 1,699,013
Pickup 98,459 283,452 416,376 84,495 882,782
Van 48,643 98,566 100,040 3,415 250,664
SUV 28,938 70,011 69,232 10,973 179,154
Other truck 10,150 57,699 40,123 4,390 112,362
RV 4,641 104,881 51,861 892 162,275
Moped/ Motorcycle 21,378 206,850 34,832 3,115 266,175
Other 2,650 4,724 2,051 1,152 10,577
DK/ Ref 89,918 86,021 4,923 186 181,048
Total 636,571 1,420,332 1,425,501 261,646 3,744,050
Table 3.7 Summary of Unmatched Caltrans Vehicles
3.4 Vehicle Matching
This section provides additional details on the problem of “ vehicle matching” using the
Year- Make- Model variables from Table 3.2. Make information is collected in the form
of a numerical code; however, the Model is typed in as a character string by an
27
interviewer collecting the information from a respondent over the phone. Cleaning these
data and performing the necessary steps to cross- reference these vehicles to entries in a
Vehicle Technology Database can be a monumental task. In addition, this illustrates an
important issue faced in vehicle choice modeling: the level of detail obtained in a
household survey like this one is relatively coarse. Information on such things as trim
levels, engine size, transmission, and drive train cannot be ascertained in a survey like
this one.
To support the requirements of this project, Caltrans Vehicles were matched to vehicle
records in the Chrome VINMatch database on the basis of Year- Make- Model ( for more
information on the Chrome database, see Appendix A). This is challenging because part
of the matching process requires comparison of character string vehicle descriptions with
no common standard. Vehicles were matched to the highest level of detail possible. In
most cases, this resulted in multiple Chrome records being matched to each Caltrans
Vehicle ( since Chrome vehicle records are relatively detailed). This approach provided
the maximum amount of flexibility for matching Caltrans Vehicles to vehicle technology
data by using the more detailed Chrome records as the potential links. Specifically, this
provided the flexibility to accommodate alternative Vehicle Class definitions should the
need arise ( now or in the future). To provide the data necessary for estimating the
models discussed in Section 4, Caltrans Vehicles were linked to the appropriate Vehicle
Classes from Section 3 to represent each household’s vehicle holdings. Although there
are usually multiple Chrome vehicles associated with each Caltrans Vehicle, the relative
lack of detail at the Vehicle Class level can help simplify the process of matching a
Caltrans Vehicle to a Vehicle Class. Specifically, in most cases all of the Chrome
vehicles matched to a Caltrans Vehicle belong to the same Vehicle Class. ( In those cases
where this is not true, the assignment is made at random using weights created from
processing the DMV data.) The next section discusses how the choice of vehicle
holdings by households is modeled.
4. CARBITS 2.0 Vehicle Market Demand Models
This section describes development of a vehicle market demand model for CARBITS 2.0.
Specifically, this is the model that performs the calculations in Step 6 (“ Simulate
Personal Vehicle Market Behavior for Current Year”) of the CARBITS Vehicle Market
Simulation Framework discussed in Section 1. CARBITS simulates the vehicle choice
behavior for households in response to current market conditions. It uses a sample of
households ( with weights) to represent California in each time period. Although there are
a number of additional details associated with simulating market behavior, the
fundamental requirement is for some type of choice model to “ simulate” each
household’s “ vehicle demand” in response to a given set of market conditions.
There are a number of options for modeling household- level vehicle purchase/ ownership
behavior. At the household level, behavior is formulated in terms of ( i) a universe of
choice options, and ( ii) choice probabilities for those options. These “ choice options”
can be characterized in various ways, e. g., the choice to purchase a vehicle, the choice to
28
hold a vehicle portfolio, or, the choice to engage in a vehicle transaction ( replacement,
addition, or disposal of currently held vehicles). This section reviews background on
vehicle choice models, describes the approach taken in CARBITS 2.0, and presents
model estimation results.
4.1 Background on Vehicle Choice Models
There are many types of vehicle choice models in the literature, and choosing which type
to use is based on a number of factors, including the purpose of the model. For example,
many models of vehicle demand are exclusively focused on the new vehicle market.
However, policy- related models like CARBITS are required to address the entire vehicle
fleet ( both new and used vehicles), which includes a much larger number of choice
options than when considering the new vehicle market alone. Moreover, the decision-making
unit in CARBITS is the Household ( not an individual making a single purchase).
In this section we briefly review some relevant background. For a more complete
introduction, see Bunch and Chen ( 2008). There are two options that are generally
available: Holdings models, and transactions models. For a holdings model, a
household’s decision- making process is described ( informally) as follows:
1. For an entire one- year period, a household will own and use a specific portfolio of
one or more vehicles ( or, the household may own no vehicles).
2. Once per year, households revisit their entire set of vehicle ownership decisions.
3. At the annual “ decision point,” household’s perform a “ complete analysis” in
which they make the following decisions for the coming year:
a. How many vehicles to own ( 0, 1, 2 or more).
b. Conditional on the number of vehicles, which vehicles to own.
4. A choice model estimates the probability of each “ holdings outcome.”
In contrast, a transactions model is described as follows:
1. A household starts in a “ base period” with a set of vehicle holdings ( including the
possibility of “ no vehicles”).
2. At certain points in time ( perhaps annually), a household makes the following
sequence of decisions:
a. Should we transact? ( Yes or No)
b. If YES, do we:
i. Replace one of our current vehicles?
1. If so, which vehicle is to be replaced?
2. What vehicle will be purchased as the replacement vehicle?
ii. Add a new vehicle to the household fleet? If so, which one?
iii. Sell one of the currently held vehicle( s)? If so, which one?
3. A choice model estimates the probability of each “ transaction outcome.”
The argument for a transaction model is that it seems like a more “ realistic” description
of household vehicle purchase behavior. In particular, a household will go along for a
period of time ( perhaps years) until some event “ triggers” the need for a transaction.
29
During this period vehicles are driven, they accumulate miles, get worn out, require
repairs, etc. In this regard, transactions models are considered to be better able to capture
“ dynamic effects” such as inertia. In contrast, a simple holdings model would seem to be
vulnerable to a much quicker market response to changes in market conditions.
Based on this discussion, a transactions model would appear to be a superior choice.
However, transactions models:
1. Require detailed household level data on such transactions in order to support
model estimation, i. e., panel data.
2. Are much more computationally intensive that holdings models ( when
implemented based on the above descriptions).
3. Have not been demonstrated to be superior in any published academic studies.
CARBITS 1.0 was implemented as a transactions model as part of a University of
California research project in the mid- 1990’ s. Choice models were estimated using a
panel data set collected on California households as part of that project. The market
simulation was implemented using a “ pure microsimulation” approach, as implied by the
above description. Specifically: In each period a household’s choice probabilities are
conditional on a specific set of vehicle holdings that a household has carried forward
from the previous period. Then, based on these probabilities, a transaction is simulated
for the current period. In most cases ( as in the real world), a household will elect to
retain its current set of vehicles for another year. A very large number of households, and
many repeated replications of the simulation, are required in order to obtain an estimate
of annual market vehicle distribution.
In contrast, a holdings model ( as described above) can be estimated using the more usual
version of household survey data in which households are interviewed at a single point in
time, and are asked to report their current vehicle holdings. Choice models are estimated
using the household sample. In the market simulation, the choice model produces a
probability for each household’s choice options. In this case, the market vehicle
distribution can be computed by taking a weighted average of the choice probabilities
over the sample of households. These numbers are deterministically computed, with no
“ simulation noise.”
This discussion provides some additional background on why CARBITS 2.0 has been
implemented as a holdings model. As noted previously, the major reason is the
availability of the Caltrans Travel Survey Data. Specifically,
1. This survey contains a very large number of California households, and also
includes weights developed by Caltrans so that the survey sample can be used to
“ represent” California.
2. This survey is a cross- sectional survey ( not a panel survey) and contains the usual
vehicle information, which is limited to vehicle holdings ( not transactions).
3. In addition to the large sample size, the data in this survey are five years more
recent the data used in CARBITS 1.0. Moreover, the panel survey data used in
30
CARBITS 1.0 was a special- purpose survey that is highly unlikely to be
replicated. In contrast, the Caltrans Survey is likely to be updated at regular
intervals. Historically, it has been replicated every ten years or so, and a certain
level of continuity and consistency in methodology has been maintained.
One final note: the above description of the two types of models is rather stylized, and
designed to illustrate certain points. In reality, the two types of models can actually be
more similar than they appear, depending on what features are included.
For example, some holdings models can be estimated with a “ transactions dummy
variable” if information on the household’s vehicle portfolio from the previous period is
available. This can be used to identify an “ inertia” effect by representing the fact that, for
a household to switch vehicle holdings requires a transaction to occur ( at some cost to the
household), so that the household’s current portfolio has a much higher probability of
being chosen than the other options. If this feature is added, the model results can be
interpreted as being “ transactions based” rather than “ holdings based,” even though the
computations are very similar.
The key question in all of this: How much information about each vehicle’s holding time
is included? If the only information carried forward in the model is whether or not a
vehicle was held during the previous period, then the two models are essentially the
same. However, in CARBITS 1.0 the model kept track of exactly how many periods each
vehicle was held by a household, and the probability of a transaction was computed
conditional on how long the household had owned the vehicle. This feature created the
requirement for a pure microsimulation approach, as indicated earlier.
4.2 Vehicle Holdings Models for Caltrans Travel Survey Data
This section summarizes vehicle holdings choice models estimated using the Caltrans
Travel Survey Data. The models are of the conditional- multinomial- logit/ nested-multinomial-
logit type similar to those that have appeared elsewhere. A full discussion is
beyond the scope of this report, but relevant references include Train ( 1986), Berkovec
( 1985), Hensher, et al. ( 1992), and Bunch and Chen ( 2008).
As discussed in the previous section, a complete vehicles holdings choice model includes
both the choice of how many vehicles to own, and which vehicle( s). One model form that
has been applied in these settings is the nested logit model. The top level has “ branches”
that correspond to the decision of how many vehicles to own ( 0, 1, 2, etc.). Under each
( non- zero) branch are the options for vehicle portfolios that a household may chose to
own. A typical nested logit model for vehicle holdings is illustrated in Figure 4.1.
One decision when developing a holdings model is how large the maximum vehicle
portfolio size should be. Most models in the literature ( e. g., Train 1986) stop with
vehicle pairs, as depicted in Figure 4.1. A few references estimate models for three-vehicle
households ( e. g., Berkovec 1985). The vehicle holdings distribution for the
Caltrans Survey households was provided in Table 3.3. Roughly 28% of households hold
31
three or more vehicles. A practical issue is that the number of possible vehicle portfolios
increases dramatically when the portfolio size increases. In Section 2 we developed 350
Vehicle Classes to represent the vehicle market in 2001. A one- vehicle household
therefore has 350 options to choose from. A two- vehicle household could theoretically
hold one of the possible pairs that can be constructed from the 350 vehicle classes,
yielding 350* 349/ 2 = 61,075 portfolio options. There are over 7 million possible vehicle
portfolios of size 3. Even if the model is limited to pairs, some type of sampling
procedure is typically employed to construct choice sets with a smaller number of
options.
0 1 2
None 2001 Two-
Seater
1982 Small
SUV
2001 Two-
Seater + 1990
Minivan
1990 Subcompact+
2001 Large SUV
Figure 4.1 Nested- logit Structure for a Vehicle Holdings Model
Our main modeling concern is capturing the interaction effects that would occur when a
household decides to hold more than one vehicle. Some combinations are more attractive
than others, e. g., households frequently hold more than one body type so that their fleet
can be used for multiple purposes. ( The three- vehicle models estimated by Berkovec
ignored such interaction effects in order to make the model estimation more tractable.)
For this project, we followed the typical practice of estimating holdings models with 0, 1,
and 2 vehicles. When simulating market behavior, a weighting procedure is employed so
that the 2- vehicle model is used to represent the vehicle choices of households with more
than two vehicles.
In a nested logit model, the “ utility” of how many vehicles to own ( one or two) is a
function of the “ expected maximum utility” conditional on the quantity choice. Consider
the case of the choice of one vehicle, conditional on the assumption that one vehicle is
being chosen. A household ( n) will choose to hold one of the J Vehicle Classes that are
available. Using a multinomial logit model ( MNL), household n’s choice probability for
Vehicle Class c is given by
􀀀
Pcn, 1 = eVcn
eVjn
j = 1
JΣ
where Vjn is household n’s preference index for Vehicle Class j. When choosing whether
to own one or two vehicles, the expected maximum utility from the decision to purchase
one of the J Vehicle Classes is given by the so- called Inclusive Value ( IV):
32
􀀀
IVn1 = ln eVjn
j = 1
JΣ
.
An analogous expression can be derived for the conditional two- vehicle choice model. If
these values were known, these and some additional factors ( e. g., household income, size,
etc.) would be expected to determine the probability of choosing one versus two vehicles.
The vehicle quantity choice model for household n can be written as
􀀀
Qnm = eWnm
eWn1 + eWn2
where Qnm is the probability that household n holds m vehicles, Wn1 and Wn2 are the
preference indexes for holding 1 and 2 vehicles, respectively, and each would include
their respective inclusive values, as well as other factors, as explanatory variables. The
full nested logit model can be directly estimated; however, a typical practice ( following
the above narrative) is to perform sequential estimation as follows:
1. Conditional one- vehicle household choice model.
2. Conditional two- vehicle household choice model.
3. Vehicle- quantity choice model.
This approach has been taken to estimate household- level vehicle holdings choice models
using the Caltrans data. Results are presented in the next sections.
4.2.1 Conditional One- Vehicle Choice Model
Consider the case of a Caltrans Household that has already decided to hold one vehicle.
A one- vehicle- household choice model can be estimated using the sample of one- vehicle
households from the survey. Based on the discussion in section 3, the household has 350
Vehicle Classes from which to choose ( summarized in Table 2.5). As noted above, the
conditional choice probability of household n choosing Vehicle Class c can be modeled
using a multinomial logit model, Vjn is household n’s preference index for Vehicle Class
j, given by the linear- in- parameters form
􀀀
Vjn = β k
k= 1
KΣ
Zk, jn .
The vector Zjn contains explanatory variables that are a function of vehicle attributes for
Vehicle Class j and household demographics from household n, and β is a K- dimensional
vector of model parameters. Household demographics used in our models are:
33
1. Household income categories
a. Income < $ 10K
b. $ 10K ≤ Income < $ 25K
c. $ 25K ≤ Income < $ 50K
d. $ 50K ≤ Income < $ 75K
e. Income ≥ $ 75K
f. Income < $ 75K
2. Household size
a. Household Size > 3
b. Household Size ≤ 3
c. Household Size > 2
d. Household Size ≤ 2
Vehicle attributes include:
1. Dummy variables for Body- Type- Size classes
a. TwoSeater [ Car]
b. Small [ Car]
c. Midsize [ Car]
d. Large [ Car]
e. Truck [ Pickup]
f. Van
g. SUV
h. LargeSUV
i. SmallSUV
2. Price ( vehicle market price, in year- 2000 $)
3. OpCost ( fuel operating cost, in cents per mile)
4. Accel ( acceleration time, seconds for 0- 60 mph)
5. LnMods ( Log of number of vehicle models in the vehicle class)
6. LnVAge ( Log of vehicle age when vehicle age is ≥ 1, 0 otherwise)
7. Prestige dummy variable
The vehicle attributes chosen for these models were based on a number of factors,
including a careful review of the literature and past experience. Price, fuel operating
cost, and acceleration cover three very important aspects of vehicle choice that are
included in essentially all ( household- level) choice models. There are a number of
possible measures of performance that could be used ( e. g., top speed, horsepower,
horsepower to weight ratio, etc.). We chose to use acceleration time because it is a
measure that consumers can relate to in terms of their direct experience ( in contrast to the
engineering characteristics). This measure is frequently used in choice experiments in
which respondent are asked to indicate their most preferred alternative. This keeps open
the possibility of, e. g., updating these choice models using stated choice data should the
need arise.
The other important dimension of vehicle functionality and size are captured relatively
well by dummy variables related to Vehicle Class. We considered using some alternative
34
measures of size such as passenger volume and luggage space ( and even did some
testing), and also vehicle footprint. However, these measures ( i) add to the vehicle data
requirements, and ( ii) are less amenable to issues related to model re- calibration. In
particular, and vehicle characteristic included in the vehicle choice model must be
forecasted for any scenario analysis being performed. The log( Number of Models)
attribute always raises concerns, but it has been shown to be important in models of this
type, i. e., those that estimate choice at the vehicle class level. ( A full discussion is
beyond the scope of this report; see, e. g., Train 1986 as a reference.) In addition to the
variables listed above, some interaction effects are also included ( e. g., interaction of
income category with Price, interaction of household- size dummy variables with different
body- type- size dummy variables).
Table 4.1 gives estimates of a multinomial logit model for 4,410 one- vehicle households.
The full choice set of 350 alternatives was used for each household ( yielding a data set
with 1,543,500 rows). The estimator is maximum likelihood, and results were obtained
using Stata ( Version 10.1).
Conditional ( fixed- effects) logistic regression Number of obs = 1543500
LR chi2( 29) = 6585.46
Prob > chi2 = 0.0000
Log likelihood = - 22540.757 Pseudo R2 = 0.1275
------------------------------------------------------------------------------
yij | Coef. Std. Err. z P>| z| [ 95% Conf. Interval]
-------------+----------------------------------------------------------------
PrLT10 | -. 0001891 .0000169 - 11.22 0.000 -. 0002222 -. 0001561
Pr10_ 25 | -. 000164 .0000121 - 13.54 0.000 -. 0001878 -. 0001403
Pr25_ 50 | -. 0000932 .0000103 - 9.02 0.000 -. 0001134 -. 0000729
Pr50_ 75 | -. 0000499 .0000103 - 4.83 0.000 -. 0000701 -. 0000296
PrGT75 | -. 000032 .0000109 - 2.94 0.003 -. 0000534 -. 0000107
PrMiss | -. 0000852 .0000118 - 7.20 0.000 -. 0001084 -. 000062
OpCost | -. 2528365 .0279414 - 9.05 0.000 -. 3076005 -. 1980724
Accel | -. 2880763 .0265671 - 10.84 0.000 -. 3401469 -. 2360057
Pres_ GT75 | -. 4308841 .213239 - 2.02 0.043 -. 8488249 -. 0129433
Pres_ LE75 | - 1.157587 .1504355 - 7.69 0.000 - 1.452436 -. 8627394
Car_ GT3 | -. 2989819 .1685597 - 1.77 0.076 -. 6293528 .031389
TwoSeat | - 2.133135 .2545379 - 8.38 0.000 - 2.63202 - 1.63425
TwoSGT2 | - 1.719884 1.016602 - 1.69 0.091 - 3.712388 .2726192
PresTS | .6419015 .616637 1.04 0.298 -. 5666849 1.850488
Subcompact | -. 5827533 .0507627 - 11.48 0.000 -. 6822463 -. 4832603
Midsize | .2291438 .0563366 4.07 0.000 .1187262 .3395615
Large | -. 6116656 .118392 - 5.17 0.000 -. 8437096 -. 3796216
PresLCar | 1.159216 .154017 7.53 0.000 .8573481 1.461084
Tr_ GT2 | -. 3725133 .1770807 - 2.10 0.035 -. 7195851 -. 0254414
Tr_ LE2 | .0407489 .1102985 0.37 0.712 -. 1754322 .2569301
Van_ GT3 | .7609437 .2287049 3.33 0.001 .3126902 1.209197
Van_ LE3 | -. 5785032 .138389 - 4.18 0.000 -. 8497406 -. 3072658
SUV_ GT75 | -. 3073329 .2273527 - 1.35 0.176 -. 752936 .1382702
SUV_ LE75 | -. 8390846 .1814757 - 4.62 0.000 - 1.19477 -. 4833989
LSUV | .4237661 .2498143 1.70 0.090 -. 065861 .9133932
SmallSUV | 1.014431 .1393791 7.28 0.000 .7412534 1.287609
New | -. 9890594 .0755862 - 13.09 0.000 - 1.137206 -. 8409132
LnVAge | -. 8244201 .0716202 - 11.51 0.000 -. 9647932 -. 684047
LnMods | .6877352 .0679447 10.12 0.000 .5545661 .8209043
------------------------------------------------------------------------------
Table 4.1 Estimates of One- Vehicle Choice Model using Caltrans Data
The coefficient estimates are highly significant, and all have interpretations that are
consistent with theory. The Price coefficients ( which are interacted with six income
35
categories) are negative, and get smaller in magnitude with increasing income category,
i. e., households become less price sensitive as income increases. Coefficients on OpCost
and Accel are both negative, and are of similar magnitudes ( similar to other models in the
literature that use these same units).
The base body- type- size category is Compact Car, with a normalized utility of zero ( not
shown). In this sample, Midsize has a positive coefficient, whereas TwoSeater,
Subcompact, and Large cars have negative coefficients. However, the PrestigeLarge- Car
interaction is strongly positive, so that the total utility of a PrestigeLarge Car is 1.16 –
0.61 = 0.55, making it the largest Car coefficient. All sizes of Cars have less utility when
households have more than 3 members, and specification testing revealed that this occurs
in about the same amount so that a single coefficient can be used.
4.2.2 Conditional Two- Vehicle Choice Model
Coefficients for two- vehicle households are given in Table 4.2. Recall that there are 350
Vehicle Classes. If one were to use all possible vehicle portfolios consisting of pairs, the
choice set size would be approximately 61,000. This model was estimated using choice
sets that were generated by a procedure designed to yield 45 vehicle pairs per household
( discussed below). Maximum likelihood estimates were obtained for a sample of 5,393
households.
In the two- vehicle model, we follow the frequently used practice of using the sum of
attributes for the two vehicles in the portfolio, e. g., Price is the sum of the two market
prices, OpCost and Accel are the sum of the values for the vehicle pair, etc. As in the
one- vehicle case, most coefficient estimates are highly significant, and have signs that
conform to theory. As before, households with progressively higher incomes become
less price sensitive. The coefficients for OpCost and Accel are similar to those in the
one- vehicle case.
This model includes many dummy variables that capture the relative desirability of
different pairs of vehicle types, e. g., Car_ Truck, Car_ Van, Car_ SUV, Truck_ Van, etc. In
addition, the sizes of Cars in the portfolio can play a role. In this specification, the “ base”
combination is a pair of Cars where one is “ Small” ( Subcompact or Compact), and the
other is “ Large” ( Midsize or Large). In addition, some of these are also interacted with
household size indicators (> 3 versus ≤ 3), income level (≥ $ 75K versus not), and
Prestige.
To illustrate, “ SmSm_ GT3” denotes two small cars, and a household with more than 3
members. Similarly, “ SmSm_ LE3” denotes two small cars, and a household with fewer
than four members. The signs of both coefficients are negative, indicating that two small
cars are less preferred than the base alternative (“ Small Car- Large Car”). Moreover, the
coefficient for SmSm_ GT3 is more negative than SmSm_ LE3, which seems logical.
36
Conditional ( fixed- effects) logistic regression Number of obs = 242685
LR chi2( 33) = 13573.04
Prob > chi2 = 0.0000
Log likelihood = - 13742.809 Pseudo R2 = 0.3306
------------------------------------------------------------------------------
yij | Coef. Std. Err. z P>| z| [ 95% Conf. Interval]
-------------+----------------------------------------------------------------
PrLT10 | -. 0002038 .0000167 - 12.20 0.000 -. 0002366 -. 0001711
Pr10_ 25 | -. 000224 8.12e- 06 - 27.60 0.000 -. 0002399 -. 0002081
Pr25_ 50 | -. 0001894 5.03e- 06 - 37.65 0.000 -. 0001993 -. 0001795
Pr50_ 75 | -. 0001559 4.87e- 06 - 32.03 0.000 -. 0001654 -. 0001463
PrGT75 | -. 0001116 4.14e- 06 - 26.96 0.000 -. 0001197 -. 0001035
PrMiss | -. 000141 5.69e- 06 - 24.77 0.000 -. 0001522 -. 0001299
OpCost | -. 3069833 .0129964 - 23.62 0.000 -. 3324558 -. 2815107
Accel | -. 3768886 .0156296 - 24.11 0.000 -. 4075221 -. 3462552
SmSm_ GT3 | -. 4179995 .1329962 - 3.14 0.002 -. 6786673 -. 1573317
SmSm_ LE3 | -. 178349 .07208 - 2.47 0.013 -. 3196233 -. 0370748
MidL_ MidL | .1055342 .0818664 1.29 0.197 -. 0549211 .2659894
HasPr_ GT75 | 1.065289 .0833516 12.78 0.000 .9019232 1.228655
HasPr_ LE75 | .0781934 .0776889 1.01 0.314 -. 0740739 .2304608
Car_ Truck | .9636605 .0651551 14.79 0.000 .8359588 1.091362
MidL_ Truck | .5590861 .0578555 9.66 0.000 .4456914 .6724808
Pr_ Tr_ GT75 | -. 5405837 .147469 - 3.67 0.000 -. 8296176 -. 2515498
Pr_ Tr_ LE75 | .2162532 .1133964 1.91 0.057 -. 0059995 .438506
Car_ Van_ GT3 | 1.589949 .0969137 16.41 0.000 1.400002 1.779896
Car_ Van_ LE3 | .3159519 .0843007 3.75 0.000 .1507256 .4811783
Car_ SUV | 1.144865 .0796171 14.38 0.000 .9888181 1.300911
Car_ SUV_ GT75 | .7329727 .0868222 8.44 0.000 .5628045 .903141
Truck_ SUV | 2.594509 .1045276 24.82 0.000 2.389638 2.799379
Van_ SUV | 1.805282 .1401018 12.89 0.000 1.530687 2.079876
TrVan_ GT3 | 2.709814 .1334341 20.31 0.000 2.448288 2.97134
TrVan_ LE3 | 1.339568 .1260801 10.62 0.000 1.092456 1.586681
Van_ Van | .5157106 .1995178 2.58 0.010 .1246629 .9067583
SUV_ SUV | 2.397272 .146996 16.31 0.000 2.109165 2.685379
Truck_ Truck | .7401467 .1232593 6.00 0.000 .4985629 .9817304
LnSMods | 2.328702 .0635653 36.63 0.000 2.204117 2.453288
numVG1 | - 2.529077 .1073028 - 23.57 0.000 - 2.739387 - 2.318768
LnTotAge | - 1.074708 .055758 - 19.27 0.000 - 1.183992 -. 9654246
numTS | -. 8002239 .1239146 - 6.46 0.000 - 1.043092 -. 5573558
nTSGT3 | - 1.014007 .3739025 - 2.71 0.007 - 1.746842 -. 2811715
------------------------------------------------------------------------------
Table 4.2 Estimates of Two- Vehicle Choice Model Using Caltrans Data
Essentially all of the other vehicle type combinations are preferred to the base alternative
( i. e., they have positive and statistically significant coefficients). Generally speaking,
most of these involve different types and sizes of vehicles, and there is a clear preference
for variety. For example, the smallest coefficients are for two “ Large” cars
( MidL_ MidL), Van_ Van, and Truck_ Truck ( an apparent exception is SUV_ SUV, with a
relatively large coefficient). Combinations such as Car_ Van, Truck_ Van are more
strongly preferred by households with more than two members, as might be expected,
due to the desirability for extra space.
There are also interactions involving Prestige and Income level. Households with more
than $ 75K in income have a higher preference for Prestige Cars. Interestingly,
households with this income have a negative coefficient for the case where a Prestige Car
is combined with a Truck. Another interaction involves Car_ SUV. High- income
households prefer this pair type more strongly.
37
As in the one- vehicle case, TwoSeaters have disutility. The coefficients here are for the
number of TwoSeaters, which are negative. In addition, there is more disutility for larger
households ( more than 3 members). Finally, as in the one- vehicle model, coefficients on
Log( Number of Models) , Log( Sum of Vehicle Ages) and number of New vehicles
( defined as model year 2000 and 2001) have the expected signs.
A final note on choice set generation: Because it is impractical to include the full choice
set of all possible pairs, subsets of alternatives are used. We elected to use an approach
with more slightly more structure than a simple random sample. We followed the
following procedure:
1. Generate all possible pairs of the 350 Vehicle Classes.
2. Randomize their ordering of the pairs.
3. Going through the list of households, one household at a time, “ deal” P ( e. g., 45)
pairs to each household from the full set. Continue until there are no more pairs
left in the “ deck”. ( In other words, pairs are randomly assigned to households
from the set of all possible pairs, without replacement).
4. If all households in the database have P pairs, stop. If there are still households in
the database without an assigned pair: Go to Step 1 and repeat the process for
those households without assigned pairs. ( If the last household in Step 3 received
a partial set of pairs, those pairs are discarded and this household becomes the
starting point for the next iteration.)
This approach ensures full coverage of the space of all possible vehicle pairs, and should
lead to more efficient estimates. This procedure is used for both estimation and
simulation. In the case of estimation, the set must include the household’s actual held
vehicles. If the randomly assigned choice set does not already include the household’s
actual holdings, one of the pairs is replaced ( at random) with the actual holdings. Note:
The results in this report are based on using choice sets with P = 45 ( 45 vehicle pairs).
However, ongoing testing could lead to variations with, e. g., larger choice set sizes.
4.2.3 Vehicle Quantity Choice Model
Inclusive values can be computed using the results of the previous sections, and used as
explanatory variables in a vehicle quantity choice model. In addition, the literature
suggests that the following factors are useful for explaining vehicle quantity choice:
1. Household size
2. Number of workers
3. Household income
4. Availability of transit.
As in the more traditional form of multinomial logit, these factors can be interacted with
the choice alternative ( one or two vehicles) as they would be expected to have different
effects. The estimated coefficients for a vehicle quantity model using Caltrans data are in
38
Table 4.3. In the current version, an index of transit availability is not available. For this
model, we used the full sample of households ( 17,040), which includes some zero-vehicle
households. The distribution of vehicle ownership was provided in Table 3.4.
The coefficients from the conditional one- and two- vehicle choice models were used to
compute inclusive values for the one- and two- vehicle choice options, respectively.
Conditional ( fixed- effects) logistic regression Number of obs = 51120
LR chi2( 11) = 18007.96
Prob > chi2 = 0.0000
Log likelihood = - 9716.3719 Pseudo R2 = 0.4810
------------------------------------------------------------------------------
v1 | Coef. Std. Err. z P>| z| [ 95% Conf. Interval]
-------------+----------------------------------------------------------------
Workers- 1v| .4175632 .0788392 5.30 0.000 .2630413 .5720851
Workers- 2v| .9057387 .0789009 11.48 0.000 .7510958 1.060382
Ln( HHSize)- 1v| -. 0896374 .1052728 - 0.85 0.395 -. 2959683 .1166935
Ln( HHSize)- 2v| 1.802626 .1068367 16.87 0.000 1.59323 2.012022
IncLT10K- 1v| - 1.566434 .1280483 - 12.23 0.000 - 1.817404 - 1.315464
IncLT10K- 2v| - 3.424094 .1464813 - 23.38 0.000 - 3.711192 - 3.136996
Inc10- 25K- 1v| -. 6500682 .1120722 - 5.80 0.000 -. 8697256 -. 4304108
Inc10- 25K- 2v| - 2.000327 .1164437 - 17.18 0.000 - 2.228553 - 1.772102
One- Veh dummy| 3.138727 .1196952 26.22 0.000 2.904129 3.373325
Two- Veh dummy| 3.738774 .2697936 13.86 0.000 3.209989 4.26756
InclValue| .2567455 .0365396 7.03 0.000 .1851292 .3283618
------------------------------------------------------------------------------
Table 4.3. Estimates of Vehicle Quantity Choice Model Using Caltrans Data
The current specification is similar to Train ( 1986). All coefficients except one are
statistically significant, and the signs are what might be expected. The alternative
specific constant for two- plus vehicles is slightly larger than for one vehicle, and both are
positive ( versus a value of 0 for the base alternative of no vehicles), indicating a
preference for more vehicles, all else equal. Coefficients for number of workers, and
natural log of household size, are estimated as interactions with the one- vehicle and two-plus-
vehicle alternatives, respectively. The coefficients for these two demographic
factors are larger for the two- vehicle alternative than the one- vehicle alternative, as
would be expected.
We also include interaction effects for the two lowest income groups. All of these
coefficients are negative. The coefficients for the lowest income group ( Less than $ 10K)
are more negative than the next- lowest group ($ 10- 25K), and the coefficients for the two-vehicle
option are more negative than for the corresponding one- vehicle option. In other
words, lower incomes result in a decrease in the expected number of vehicles per
household. The coefficient for the Inclusive Value term is positive, indicating that any
changes in vehicle features that yield increased utility will cause the probability of that
branch to increase.
The vehicle holdings models estimated here specifically model household vehicle
demand behavior, conditional on current market conditions ( whatever they may be).
These models are combined with other elements of CARBITS to simulate the total
market “ system.”
39
5. Department of Motor Vehicle ( DMV) Registrations Data
The models estimated in Section 4 are based on a specific sample of survey respondents.
These household- level data are useful for identifying important behavioral effects when
individual households make vehicle purchases. However, the sample sizes associated
with survey data are not large enough to provide an accurate measure of aggregate- level
market statistics ( e. g., new vehicle sales of various vehicle types) that can be important
when performing policy analysis. To address this issue, models estimated using survey
data are typically recalibrated so that they match aggregate level statistics from other data
sources. For example, in the case of CARBITS it would be desirable for the market
demand model to “ simulate” new vehicle sales in the base year that match actual vehicle
sales. Moreover, because CARBITS also models the used vehicle market, it would be
desirable to match vehicle count distributions by model year as well. Finally, if the
model explicitly simulates vehicle exit/ scrappage, it would be desirable to match known
vehicle exit/ scrappage rates ( if such data are available).
For this project, procedures have been developed for processing California DMV
registrations data to meet these needs. Specifically, the DMV has been producing regular
biannual data “ dumps” of all registrations for quite a number of years. Each data dump
can be thought of as a snapshot of vehicle registrations at a particular point in time. The
snapshots generally occur in October and April of each year. The practice of generating
these data sets began as the result of joint effort by the California Energy Commission
( CEC), ARB, and Caltrans to obtain data that could be used to meet needs of the various
agencies. ( A full history is beyond the scope of this project. The lead agency on this has
been the CEC, with varying levels of participation from the other two agencies.)
In what follows, we look at registrations data from October 2001. October is an
attractive month to consider because, by this time of the year, most sales of new vehicles
with the model year corresponding to the current calendar year have occurred. For
example, by October 2001 most sales of new 2001 model year vehicles have occurred. In
addition, some sales of new model year 2002 vehicles have also occurred. However, in
the DMV data there are very few of these vehicles, and our current practice is to drop
them. For an illustration using the October 2001 DMV snapshot, see Figure 5.1.
The data in Figure 5.1 are limited to light- duty vehicles. Wherever possible, vehicles that
are known to be part of government or commercial fleets have been excluded. The
vehicle total for model years 1982- 2002 is approximately 18.8 million. A few features of
this figure are noteworthy. During this period there were economic recessions in 1980-
1982, 1990- 1991, and 2001- 2003, with periods of steady growth in between. The
downturns in Figure 5.1 correspond to these periods.
As a point of comparison, recall that the Caltrans Travel Survey data were collected from
October 2000 to December 2001, and the sample is weighted so that 21.4 million vehicles
are “ available to households” ( see Section 3). The number of light- duty vehicles with
model years 1982- 2001 using this weighted sample is estimated to be 18.5 M versus the
40
18.8M in the October 2001 DMV snapshot. For a comparison of the model year
distributions from the two data sets, see Figure 5.2.
Figure 5.1 Model Year Distribution for October 2001 DMV Registrations
( Light- Duty Vehicles)
Based on our past experience in comparing such distributions across different data
sources, these are remarkably close. The DMV curve is much smoother than the Caltrans
curve, as would be expected due to the issue of sample size. The main difference is that
the vehicle counts for model year 2001 are substantially lower for the Caltrans data. This
is easily explained: The Caltrans data were collected from households over an extended
period of time starting in October 2000. Sales of model year 2001 vehicles accumulate
over the entire calendar year and beyond into the following calendar year. The earlier a
household was interviewed, the more likely it was that they could have purchased a 2001
model year vehicle after they were interviewed. More generally, this is a typical issue
faced with choice model estimation: Households interviewed early in the process could
have purchased a vehicle in the new vehicle market with model year 2000. In other
words, it can be difficult to determine “ new vehicle sales” on the bases of vehicle model
year registrations. It is these phenomena that lead to the need for re- calibration of model
constants for market simulation.
41
Figure 5.2 Model Year Distributions for DMV versus Caltrans Travel Survey
There are many details associated with processing DMV data that are not discussed in
this section— see Appendix A. An important requirement is to be able to link vehicle
counts for specific year- make- model vehicles in the California fleet to the corresponding
vehicles in other data sets ( e. g., the vehicle technology database) in order to perform
various modeling tasks.
6. Vehicle Market Exit
As discussed in Section 1, one of the stated project goals is to explicitly model the exit of
vehicles from the used vehicle fleet. In CARBITS 1.0, the exit of vehicles from the
California fleet was an implicit outcome of household vehicle transaction choices for
used vehicles over time. As vehicles continue to get older, their attractiveness diminishes
so that more used vehicles of a particular class are sold than are purchased, leading to a
net exit of vehicles from the market. An argument in favor of this approach is that the
vehicle fleet distribution is determined by an internally consistent behavioral model of
individual- level household vehicle preference and choice. A number of models
( including the CalCars model of CEC) take this approach.
42
A potential vulnerability of this approach is that, combined with microsimulation, exit
patterns of individual vehicle classes could appear noisy or inconsistent with typical
scrappage patterns when compared to smoothed, well- behaved curves generated by
models based on aggregate vehicle count data. The primary vulnerability is that it leaves
the model open to criticism by hired consultants who use aggregate- level models, which
are much simpler, and easier to both control and explain.
The literature contains examples of forecasting models in which household- level vehicle
choice models are combined in the same system with scrappage models based on
aggregate data— see, for example, Berkovec ( 1985) and Bento, at al. ( 2006). Although
this approach is not based on a theoretical framework that is completely internally
consistent, there is some behavioral theory that underlies the specification of the
scrappage models, and this approach can be considered a way of incorporating additional
information from aggregate data sources into the system. This project included a task to
add this feature to CARBITS.
Before continuing, we make a few remarks about the general issue of modeling “ vehicle
scrappage.” It will be noted that in this report we sometimes use the term “ exit,” and we
sometimes use the term “ scrappage.” The main idea is that, when modeling the behavior
of a vehicle market over time, older vehicles eventually “ disappear” from the vehicle
fleet by some process. At some point in time, most vehicles reach a state where they
cease to exist and can never be “ on the road” again. Vehicles that have been totaled in an
accident, or simply become unusable, are scrapped for raw materials and spare parts.
However, getting accurate data on this process is extremely difficult, and represents a
challenge for modelers.
Another issue is that, when modeling a vehicle market over time, the market can ideally
be treated as a “ closed system” whereby all vehicles entering the market first do so
through new vehicle sales, and they eventually exit by being scrapped. When modeling
the domestic vehicle market for the entire United States, this may be a reasonable
approximation. However, when modeling a submarket ( e. g., California), the market is
not really a “ closed system.” Vehicles of all vintages can both enter and leave the market
through migration to and from other States. In this regard, there may be a net “ exit” of
vehicle classes from the market, but this process contains a mixture of immigration and
scrappage processes. For this reason, we prefer to discuss vehicle “ exit” rather than
“ scrappage.”
Unless immigration processes are explicitly included in the model system, some
modeling assumptions are required for simulating vehicle “ exit” from the market.
However, the more immediate issues are: What data should be used for estimating such a
model, and what should a model look like? In this project, we use DMV registrations
data for two consecutive years ( October 2000 and October 2001) to estimate vehicle “ exit
rates” corresponding to the time frame of the Caltrans Travel Survey.
Our experiences mirror those reported in other research publications. Specifically, there
is little or no vehicle exit during the first few years of most vehicle types. In fact, the
43
data show a continued increase in vehicle counts for many vehicle models after the initial
year of introduction. In our case, at least part of this effect can be attributed to
immigration of vehicles into the State. However, researchers working with national- level
registrations data also observe this effect, and have attributed it to continued new vehicle
sales from the initial model year inventory for periods of up to four or more years— see
Berkovec ( 1985). There are also issues with very old vehicles, where certain types of
vehicles may be reconditioned and re- registered, leading to a net increase in vehicle
counts that should theoretically not occur.
For our analysis, we computed vehicle exit rates for vehicles at the Year- Make- Model
level. One useful piece of information contained in the DMV data is where the vehicle
was originally sold as new: either in California, or Out of State ( OS). This enabled us to
confirm that there was a substantial amount of vehicle immigration for more recent
model years, so that, on average, about 20% of the vehicle fleet will have originated from
Out of State. We estimated vehicle exit rates by first removing net increase in OS
vehicles over the period. This is a completely practical approach, and quite literally this
is a vehicle “ net exit” model, since, e. g., it is not possible for us to know if a vehicle
originating in California left the fleet by leaving the State, through scrappage, etc.
Moreover, vehicles could leave California and then return at a later time. The current
analysis cannot separately identify this effect. ( Later, we will remark on the possibility of
future work that can be done in this area.)
We estimate a model using the same approach as earlier work in the literature— see
Berkovec ( 1985). For each vehicle type n = 1,…, N, the estimated exit rate is given by
􀀀
Rn
2001 = ( Qn
2001 − Qn
2000 ) / Qn
2000
where
􀀀
Qn
y is the vehicle count of vehicle type n in year y. We use a data set with N =
2,385 vehicle types ( at the level of Year- Make- Model) using vehicles from model year
1982 to 1994. As noted earlier, it is typical in the literature to drop the first few years of
data and treat the scrappage rate as zero ( four is a typical number), for the reasons
discussed. In our case, these are State- level ( not national) data, and vehicle immigration
seems to be a major effect. The “ exit rate” figures for the first six years exhibited some
unusual patterns, so these years have been dropped ( this issue could be explored in more
detail at a later time, if it seems warranted). Figure 6.1 provides plots of average exit
rates as a function of Body Type/ Prestige and Model Year.
44
Figure 6.1 Mean Vehicle Exit Rate by Body Type/ Prestige and Model Year
There are some noticeable patterns in this plot. The exit rates for Prestige versus Non-
Prestige Cars are extremely different. Starting with the most recent model year and going
back, the curve starts out relatively flat ( below those of light duty trucks) for the first few
years. Thereafter, there is a sharp increase in exit rates for Non- Prestige Cars, reaching a
level of over 20% for the oldest vehicles. Prestige Cars have lower exit rates compared
to all other Body Types for the newest 8- 9 model years, and are comparable to the light
duty trucks for the oldest model years. Exit rate curves for light duty trucks start out
higher than cars, and have shapes that are ( i) similar to each other, but ( ii) different from
either type of car. The curves for Vans and SUVs are similar to each other, and below
the curve for Pickups.
What this plot does not include is the role that economic behavior might play in
explaining the