United States
Environmental Protection
Agency
Office of Transportation EPA420-B-05-019
and Air Quality October 2005
COMMUTER v2.0 Model
Coefficients
-------
EPA420-B-05-019
October 2005
COMMUTER v2.0 Model Coefficients
Transportation and Regional Programs Division
Office of Transportation and Air Quality
U.S. Environmental Protection Agency
NOTICE
This Technical Report does not necessarily represent final EPA decisions or positions.
It is intended to present technical analysis of issues using data that are currently available.
The purpose in the release of such reports is to facilitate an exchange of
technical information and to inform the public of technical developments.
-------
COMMUTER v2.0 Model Coefficients
The travel impacts calculated in the COMMUTER Model are based on logit mode-choice
coefficients. The COMMUTER Model impacts are highly sensitive to the values of these
coefficients, which are used to predict mode share changes in response to changes in
travel time and cost associated with transportation control programs. Given this high
degree of sensitivity, it is important that COMMUTER Model users understand the basis
for these coefficients and how they are validated and used in transportation modeling.
Potential users of the model should evaluate the influence of the coefficients, perform
sensitivity tests to understand their impacts, and verify that there is consistency among
the coefficient values used in related transportation planning models and the coefficient
values used in this model for the city or local area being analyzed. The purpose of this
discussion is to educate users about these issues. This discussion covers the following
elements relating to the use of mode-choice coefficients in the COMMUTER Model:
• Modeling Techniques outlines the various modeling processes that use
coefficients of this type and identifies how the coefficients are used in the
various models.
Review of the Coefficient Values identifies the coefficients used in the
COMMUTER Model and defines the range of values, averages, and
categories that are used in the model.
• Documentation of Sources lists the specific documentation of the
coefficient sources.
Overview
The COMMUTER Model employs a simplified logit modeling process (called pivot
point) that relies on locally derived coefficients to evaluate the influence of alternative
measures on travel behavior. The coefficients are derived from observed travel behavior
using standard survey techniques, statistical analysis, and modeling methods.
Coefficients that have not been derived from observed travel behavior (such as composite
measures or transferred coefficients) are not included as they could bias the average
values. The fact that the coefficients are all derived using similar statistical methods
explains why the coefficients are reasonably similar across the country. A review of the
coefficients indicates that while they are relatively consistent across the country, there is
enough variation in values between cities that it is essential that users understand the use
of these coefficients. This report provides the user with this background information.
The mode-choice coefficients employed in the COMMUTER Model have been
"validated" and are widely used in urban transportation modeling for a number of
reasons, which include the following:
-------
The coefficients are derived from observed travel behavior using standard
survey techniques, statistical analysis, and modeling methods.
These coefficients are not estimated separately but rather as functions of
the mode split behavior and are related to all variables included in mode-
choice model equations.
The coefficients are reasonably consistent across the country.
Many metropolitan regions have used these coefficients to "backcast"
known mode share conditions, e.g., to test the accuracy of their forecasting
models.
Metropolitan areas also have used these coefficients to verify before and
after conditions of new transit services.
Modeling Techniques
Coefficient values are used in all types of modeling techniques and represent similar
behavioral aspects of travel in each case. The primary difference in the modeling
techniques is that standard logit models estimate probabilities that a person would choose
a certain mode (e.g., driving alone, carpooling, transit, etc.) based on all variables that
impact travel decisions. Conversely, pivot-point models estimate these probabilities
based only on the changes in specific variables. Both the TDM Evaluation Model and
the COMMUTER Model are based on a pivot-point technique.
The pivot-point logit technique is a simplified version of the logit modeling process
found in most mode-choice models, which are developed at the metropolitan level by
Metropolitan Planning Organizations (MPOs). The primary difference between the
standard logit modeling process and the pivot-point technique is as follows:
Standard logit mode-choice models, as applied in regional travel models,
include many different parameters, such as transportation level of service
(e.g., travel times, transit fares, parking costs); area characteristics (e.g.,
employment density); and socioeconomic and demographic characteristics
(e.g., income, household type). In order to apply the models, baseline
levels and changes in each variable must be known for all variables.
• Pivot-point models are a simplified form of the logit mode-choice model.
Pivot-point models "pivot" off the baseline mode share, based on the
change in value for certain variables of interest (e.g., transportation LOS).
It is not necessary to know the baseline levels of any other variables, since
these baseline levels are reflected in the starting mode share. It is also not
necessary to know levels of other variables, such as demographic
characteristics, that are assumed not to change.
-2-
-------
Standard logit mode-choice models, as applied in regional travel models, can be used to
test a broader variety of impacts than pivot-point models. However, since they are
integrated with the full regional travel model, they can only be used in conjunction with
the entire set of data and modeling processes incorporated in the model. Regional travel
models are widely used by MPOs to test the impacts of changes in automobile and transit
levels of service, population, employment, demographics, and other variables such as the
pedestrian environment. Pivot-point models are based on the same behavioral
information (coefficients) and modeling methodology used in regional travel models, yet
they apply this methodology in a simplified approach that can be used in stand-alone
analysis.
The TDM Evaluation Model, developed by Comsis Corporation in 1993 and sponsored
by Federal Highway Administration, uses the same pivot-point methodology as described
in this memorandum and the COMMUTER Model User Guide, with the following
exception. The TDM Evaluation Model applies the coefficients to zone-to-zone trip
activity data (trip tables generated by regional models are input directly into the TDM
Evaluation Model); in other words, coefficients are applied separately to trip
characteristics and LOS changes between each pair of origin and destination zones. The
COMMUTER Model, in contrast, applies the coefficients to a single set of trip activity
data, whether it is aggregate metropolitan area data or individual employer data.
The coefficients used in the TDM Evaluation Model and the COMMUTER Model are
also very similar. Composite coefficients used in the TDM Evaluation Model were
derived from MPO area travel demand models, and average COMMUTER Model
coefficients for small, medium, and large size metropolitan areas were also developed
from MPO area travel demand models. The primary difference in the coefficient values
is that those in the COMMUTER Model are based on more, recent data.
Review of Coefficient Values
Review - The coefficient values used in the COMMUTER Model are defined as follows:
• In-vehicle travel time (in minutes) for transit modes.*
• Out-of-vehicle travel time (in minutes) is divided into walk and wait
parameters. The walk coefficient is used for both auto and transit modes,
and the wait coefficient is exclusive to transit modes.
Cost (in cents) is separated by auto (parking costs) and transit (fare). Auto
* The COMMUTER model was not designed to assess impacts from large changes in the
transportation system. As a result, it assumes that in-vehicle travel time remains constant for
auto modes (drive alone, carpool, and vanpool) and only allows in-vehicle travel time changes to
be applied to transit. Transportation system changes that produce measurable impacts on in-
vehicle travel time for auto modes cannot be assessed with the COMMUTER model and must be
treated with a full "four step" travel demand model.
-------
operating costs were also considered, but are typically the same as transit-
fare coefficients.
These parameters are typically established in units of minutes and cents for inclusion into
travel demand models and are set to match these units in the COMMUTER Model for
consistency. Since there is no guarantee that these units would match an individual city's
travel demand forecasting model coefficient units, this should be checked prior to use of
the COMMUTER Model. Recognizing that the typical units for cost change inputs and
outputs in travel demand forecasting models are dollars, the COMMUTER Model
internally applies a cents-to-dollars conversion factor to the cost-related coefficients that
are typically reported in cents when combining cost coefficients with cost inputs in
dollars. If the user supplies their own local coefficients (instead of using the model's
city-specific or area size defaults), the time-related coefficients must be entered in units
of minutes and the cost coefficients must be in cents.
All coefficients identified above are expected to be negative, to represent the fact that as
the value of the parameter (time or cost) increases, the probability that a person would
choose that mode (auto or transit) decreases. The larger the negative value for a
coefficient, the greater its impact on the affected mode. A review of a range of values
shows that the coefficient values change from city to city and apparently change over
time, and these changes in the coefficients can have significant impacts on the results of
modal choices.
Table 1 presents the ranges of values for all cities and shows the average coefficient
values by city size and over time. The average values in this table demonstrate some
trends that transportation planners rely on, such as the following:
• Walk time is twice as onerous as in-vehicle travel time.
• Wait time is more onerous than walk time.
• Approximately three cents of parking cost is equal to one minute of in-
vehicle travel time, which translates to an average rate of only $1.80 per
hour.
Transit fares are less onerous than parking cost.
-4-
-------
The city-specific and overall average coefficient values supplied with the COMMUTER
Model were updated for COMMUTER release version 2.0 (2005). This update included
a review of current modeling practice in various metropolitan areas to identify the most
Table 1 Range of Coefficient Values
Range of Values
Minimum
Maximum
Overall Average
In-Vehicle
(minutes)
-0.0450
-0.0113
-0.0253
Walk Time
(minutes)
-0.0931
-0.0186
-0.0473
Transit
Wait Time
(minutes)
-0.0978
-0.0155
-0.0466
Auto
Parking
Cost (cents)
-0.0173
-0.0004
-0.0056
Transit
Fare (cents)
-0.0135
-0.0004
-0.0040
recent model coefficients, and removal of all coefficients that were more than 20 years
old (i.e., developed prior to 1986). Following this review, coefficients were compared
among metropolitan area size class. Unlike in the review of coefficients for
COMMUTER version 1.0, however, no statistically significant differences by size class
were identified. Also, trends in coefficients over time were evaluated; again, however,
no statistically significant basis was identified for establishing a clear time-trend for the
coefficients.
The ranges shown in Table 1 exclude a handful of outliers that lay substantially out of the
range expected for these types of coefficients.
Examples
The values of the coefficients can best be described with examples, as shown in Table 2.
The first example shows the impacts of improved transit service by itself. The second
example shows the impacts of improved transit service combined with an additional
charge for parking. These examples present the type of sensitivity analysis that some
regional agencies conduct to compare the effects of variables both individually and in
combination. The size of the coefficient affects the impact of the variable on the mode
share, but only in context with the pivot-point logit model equation. One description of
this approach can be found in "Modeling Transport" by J. de D. Ortuzar and L.G.
Willumsen (Wiley Publishers, 1990), page 302.
-5-
-------
Table 2 Examples of Sensitivity Testing of Coefficient Values
Example #1: Reduced Transit Travel Time
Original
Improved
Difference
Travel Time (min.)
Auto
in-
vehicle
20
20
0
Transit
in-
vehicle
40
35
-5
Transit
Walk
10
10
0
Transit
Wait
10
10
0
Cost (cents)
Auto
100
100
0
Transit
150
150
0
Utility Change
Cityl
City 2
City 3
0.082
0.126
0.171
New Mode Shares:
Auto
89.2%
88.8%
88.4%
Transit
10.8%
11.2%
11.6%
Example #2: Reduced Transit Travel Time + Increased Parking Cost
Original
Improved
Difference
Travel Time (min.)
Auto
in-
vehicle
20
20
0
Transit
in-
vehicle
40
35
-5
Transit
Walk
10
10
0
Transit
Wait
10
10
0
Cost (cents)
Auto
100
150
50
Transit
150
150
0
Utility Change
Cityl
City 2
City 3
0.082
0.126
0.171
-0.029
-0.281
-0.533
New Mode Shares:
Auto
89.0%
85.7%
81.7%
Transit
11.0%
14.3%
18.3%
The calculations presented in Table 2 include default coefficients by population size
input into the COMMUTER Model, and assume current mode shares of 90% auto and
10% transit. In the first example:
• Level of service (LOS) changes include improved transit service (five
minutes faster in-vehicle travel time); and
• Computations indicate an expected increase in transit use from 10% to
11.0%, 11.1%, and 11.3% for small, medium, and large size metropolitan
areas, respectively.
In the second example:
• LOS changes include improved transit service (five minutes faster in-
vehicle travel time) combined with a parking cost increase from $1.00 to
$1.50.
-6-
-------
The computations show an expected increase in transit use from 10% to
14.6%, 15.2%, and 16.9% for small, medium, and large size metropolitan
areas, respectively. The greater shift to transit in Example #2 illustrates
the combined effect of the service changes.
Each coefficient, when multiplied by the corresponding change in LOS, indicates a
change in "utility" for the mode. Utility is a relative measure of attractiveness;
essentially, the coefficients are converting changes in different units (minutes, cents, etc.)
into similar terms so they can be directly compared. Larger coefficient values on
variables with the same units indicate a higher value on that variable. For example, as
shown in the Table 1 coefficients, out-of-vehicle travel time is valued more highly than
in-vehicle travel time (people dislike to wait). The utility changes from each LOS
component are then combined to determine an overall change in mode share using the
logit model equation.
The underlying computations are shown below for Example #1 with a small metropolitan
area.
Utility change: AU = Coefficient * Change in LOS
Transit utility change: AUTrans = (- 0.207) * (- 5) = 0. 103
New transit mode share:
P
TmnS (pAuto XeAUA«>° ) + (PTrans XeAur™ ) (0.90 x e° ) + (0.1 0 x e0103
where
P 'Tram = new transit mode share
P TranS = base (existing) transit mode share
PAuto = base auto mode share
AUTram = transit utility change
AUAuto = auto utility change
Area-Specific Coefficients
Table 3 shows the area-specific coefficients provided with the COMMUTER Model. In
most cases, these coefficients were obtained from travel model documentation or
personal communication with travel demand forecasters in each area. For COMMUTER
version 2.0, forecasting staff in the 10 largest metropolitan areas of the country were
contacted in December 2004 to identify the most recently available coefficients. In
addition, updated coefficients were obtained through available documentation from other
-7-
-------
metropolitan areas. Special effort was made to update older coefficient values; values
based on data from before 1986 were discarded in COMMUTER version 2.0.
For some areas, model coefficients were provided for specific categories (e.g., by mode
or by income level). In these cases, coefficients had to be combined for consistency with
the coefficient categories used in COMMUTER. Coefficients were combined in the
following ways:
In-Vehicle Travel Time - For areas with different auto and transit in-vehicle travel time
coefficients, a combined IVTT coefficient was developed by weighting the mode-specific
coefficients by commute mode share from the 2000 U.S. Census. This was done for the
following cities:
Dallas
Detroit
Philadelphia
Seattle
Transit Wait Time - For areas where transit time coefficients are split between less than
7 minutes and greater than 7 minutes, the transit time coefficient for greater than 7
minutes was used. This is because the transit time coefficient for less than 7 minutes is a
2 to 3 times the transit time coefficient for greater than 7 minutes. The reason for this is
that it is anticipated that for headways of over 15 minutes, a traveler will attempt to
"schedule" his or her arrival at the transit stop and therefore the wait is less onerous.
This was done for the following cities :
Atlanta
• Denver
San Diego
Tucson
Transit Fare - For areas where the transit fare was calculated based on income, a
combined transit fare coefficient was developed by weighting the income specific
coefficients by income shares reported in the household survey. This was done for the
following city:
• Denver
Also, a handful of coefficients either were unavailable for specific models, or were
removed because their values lay outside the expected range for such coefficients. In
cases in which coefficients were unavailable or removed, substitute values were inserted
based on other coefficients in the area's model, using the ratio of two different categories
based on an average ratio of coefficients across other cities. For example, in Los
Angeles, the substitute walk time coefficient shown in Table 3 (-0.1073) is equal to the
Los Angeles in-vehicle travel time coefficient times the average ratio of in-vehicle travel
time to walk time coefficients calculated across other cities in the dataset (2.38). Cost
coefficients were substitute based on an average value of time computed at $5.79 per
hour. The following substitute coefficients are included in Table 3:
-------
• Los Angeles - Walk Time
• Sacramento, Tucson - Auto Parking and Transit Fare
• San Diego - Transit Fare
Substitute coefficients were not included in the calculation of average coefficient values.
Also, the Baltimore coefficients were not included because they are taken from the travel
demand model for the Washington, D.C. area which is already included in the dataset.
Table 3 Area-Specific Coefficients
Location
Albuquerque
Atlanta
Baltimore
Boston
Chicago
Cleveland
Columbus
Dallas
Denver
Detroit
Houston
Los Angeles
Milwaukee
New York
Philadelphia
Phoenix
Portland
Reno
Sacramento
San Diego
San Francisco
San Juan
Santa Cruz
Seattle
Tucson
Washington D.C.
Year
1992
2002
1993
1991
1990
1994
1999
1996
1997
1996
1985
1996
1991
1996
1986
1991
1994
1991
2001
1995
1990
1990
1990
1990
2000
1994
In-Vehicle
Travel Time
(min)
All Modes
-0.0209
-0.0256
-0.0300
-0.0314
-0.0282
-0.0178
-0.0213
-0.0544
-0.0180
-0.0512
-0.0220
-0.0450
-0.0157
-0.0113
-0.0391
-0.0167
-0.0394
-0.0275
-0.0250
-0.0250
-0.0333
-0.0366
-0.0163
-0.0176
-0.0178
-0.0300
Out-of- Vehicle Travel
Time (min)
Walk
Time
-0.0219
-0.0639
-0.0750
-0.0330
-0.0440
-0.0444
-0.0640
-0.0640
-0.0540
-0.0186
-0.0568
-0.1073
-0.0412
-0.0380
-0.0316
-0.0206
-0.0646
-0.0550
-0.0380
-0.0500
-0.0931
-0.0717
-0.0325
-0.0206
-0.0400
-0.0750
Transit —
Wait
-0.0978
-0.0256
-0.0750
-0.0550
-0.0960
-0.0378
-0.0465
-0.0640
-0.0180
-0.0186
-0.0568
-0.0423
-0.0412
-0.0554
-0.0511
-0.0304
-0.0397
-0.0550
-0.0380
-0.0250
-0.0523
-0.0752
-0.0325
-0.0155
-0.0200
-0.0750
Out-of-Pocket Travel
Cost (cents)
Auto -
Parking
-0.0031
-0.0031
-0.0043
-0.0173
-0.0021
-0.0034
-0.0016
-0.0056
-0.0014
-0.0041
-0.0154
-0.0025
-0.0045
-0.0004
-0.0026
-0.0053
-0.0135
-0.0167
-0.0025
-0.0069
-0.0021
-0.0066
-0.0045
-0.0024
-0.0018
-0.0043
Transit -
Fare
-0.0031
-0.0013
-0.0043
-0.0083
-0.0008
-0.0024
-0.0016
-0.0055
-0.0012
-0.0041
-0.0061
-0.0025
-0.0045
-0.0004
-0.0012
-0.0053
-0.0135
-0.0067
-0.0025
-0.0025
-0.0021
-0.0066
-0.0036
-0.0024
-0.0018
-0.0043
-9-
-------
Documentation of Sources
Model documentation - Coefficients for the following cities were obtained from model
documentation (date after city indicates year of source data):
Atlanta (2002) - Atlanta Regional Commission, Mobility 2030 : Model
Documentation, December 2004.
• Baltimore (1993) - Metropolitan Washington Council of Governments, COG/TPB
Travel Forecasting Model Version 2.1 D #50 Calibration Report, November 17,
2004.
Columbus (1999) - PB Consult / Parsons Brinckerhoff for the Mid-Ohio Regional
Planning Commission, The MORPC Travel Demand Model: Validation and Final
Report, December 17, 2004.
Dallas (1996) - Cambridge Systematics for the North Central Texas Council of
Governments, NCTCOGMode Choice Model, August 2002.
• Los Angeles (1996) - Cambridge Systematics, Inc., SCAG Regional Mode Choice
Model Development Project, Final Report, October 28, 1996.
• Milwaukee (1991) - Southeastern Wisconsin Regional Planning Commission,
Travel Simulation Models for the Milwaukee East-West Corridor Transit Study,
May 1993.
• New York (1996) - Parsons Brinckerhoff Quade & Douglas with Cambridge
Systematics and others for New York Metropolitan Transportation Council,
Transportation Models and Data Initiative : Technical Memorandum Task 14.15
&16 / Milestone C Best Practice Model Development - Travel Details: Pre-Mode,
Mode Choice, and Stops Model and Travel Patterns: Journey Frequency and
Destination Choice, August 2001.
• Portland (1994) - Metro, The Phase III Travel Demand Forecasting Model: A
Summary of Inputs, Algorithms, and Coefficients, June 1, 1994.
Sacramento (2001) - DKS Associates for Sacramento Area Council of
Governments, Model Update Report, Sacramento Regional Travel Demand
Model, Version 2001 (SACMET 01), March 8, 2002.
• San Francisco (1990) - Metropolitan Transportation Commission, Travel Demand
Models for the San Francisco Bay Area (BAYCAST-90) Technical Summary,
June 1997.
• Seattle (1990) - Cambridge Systematics, Inc. with Urban Analytics for Puget
Sound Regional Council, Land Use and Travel Demand Forecasting Models :
New Model Documentation, June 30, 2001.
Tucson (2000) - Cambridge Systematics, Inc. for Pima Association of
Governments, PAG Model Evaluation and Improvement Plan : Travel Demand
Forecasting Model, August 2003.
• Washington D.C. (1994) - Metropolitan Washington Council of Governments,
COG/TPB Travel Forecasting Model Version 2.1 D #50 Calibration Report,
November 17, 2004.
Personal Communication - Many of the coefficients were obtained directly from the
consultant travel demand modelers responsible for regional model estimation. These
-10-
-------
metropolitan areas included the most recent model updates conducted in the following:
Cleveland (1994);
Denver (1997);
Detroit (1996);
Phoenix (1991);
Reno (1991).
San Diego (1995);
Santa Cruz, California (1990).
Previous Model - The following coefficients were taken from COMMUTER Model
Version 1.0:
Albuquerque (1992);
Boston (1991);
Chicago (1990);
Phoenix (1991); and
San Juan (1990).
Summary
The COMMUTER Model will be a more powerful tool to the user if the impacts of the
coefficients are understood. The best means to achieve this understanding is to use
sensitivity testing similar to that presented in this discussion. This will serve to indicate
the general impacts to changes in cost or time variables as well as to identify that the
model and data are being applied correctly.
-11-
------- |