This action might not be possible to undo. Are you sure you want to continue?
https://www.scribd.com/doc/43582938/GuidebookonStatewideTravelForecasting
11/07/2011
text
original
Department of Transportation
Federal Highway Administration
Guidebook on Statewide Travel Forecasting
Prepared for
Federal Highway Administration
Prepared by
Center for Urban Transportation Studies
University of Wisconsin – Milwaukee
In cooperation with
Wisconsin Department of Transportation
March 1999
Table of Contents i
Table of Contents
Chapter 1. Introduction................................................................................................ 1
Structure of the Guidebook.................................................................................................. 1
Reasons for Statewide Travel Forecasting ......................................................................... 1
Original ISTEA Planning Factors ..................................................................................................... 2
Overview of Approaches...................................................................................................... 2
Time Series Methods....................................................................................................................... 3
Overview of Three/Four Step Models............................................................................................... 3
Calibration and Validation................................................................................................................ 4
Specialized Methods ....................................................................................................................... 4
Introduction to Data Sources............................................................................................... 5
Key Data Sources ........................................................................................................................... 5
Key Census Databases ................................................................................................................... 6
National Personal Transportation Survey (NPTS) Overview............................................................. 6
American Travel Survey (ATS) Overview......................................................................................... 6
Important Freight Data Sources....................................................................................................... 7
Commercial Economic Forecasts..................................................................................................... 7
Tools...................................................................................................................................... 7
Role of GIS ..................................................................................................................................... 7
Software Issues............................................................................................................................... 8
Composing a Complete Forecast ........................................................................................ 8
Chapter 2. Time Series Methods ................................................................................. 9
Introduction........................................................................................................................... 9
What is Time Series Analysis?......................................................................................................... 9
Applicability of Time Series Analysis................................................................................................ 9
Important Statistical Concepts .......................................................................................... 10
Growth Factors ................................................................................................................... 11
Growth Factor Relationships.......................................................................................................... 11
Linear Trend Model............................................................................................................. 12
BoxCox Transformations .............................................................................................................. 14
Moving Averages and Seasonality .................................................................................... 15
Exponential Smoothing.................................................................................................................. 15
Central Moving Average ................................................................................................................ 15
Introduction to BoxJenkins (ARIMA) Methods ................................................................ 16
ARIMA .......................................................................................................................................... 16
Autoregressive (AR) Models.......................................................................................................... 16
Integrated (I) Models ..................................................................................................................... 17
Moving Average (MA) Models........................................................................................................ 17
Steps in Building an ARIMA Model ................................................................................................ 18
Extensions to ARIMA (Examples from WisDOT) ............................................................................ 18
Cyclic Patterns in AR Models......................................................................................................... 19
Case Studies ....................................................................................................................... 19
Air Travel I: Brown and Watkins (1968)......................................................................................... 19
Air Travel II: Oberhausen and Koppelman (1982) ......................................................................... 19
Traffic Levels I: Maine VMT.......................................................................................................... 20
Traffic Levels II: New Mexico Heavy Vehicle Traffic ...................................................................... 20
Guidebook on Statewide Travel Forecasting ii
Appendix: Some Elementary Statistical Concepts.......................................................... 21
Chapter 3. Passenger Forecasting............................................................................ 24
Introduction......................................................................................................................... 24
Urban v. Statewide Models............................................................................................................ 25
Sources of Travel Data.................................................................................................................. 25
Defining the Scope of the Model........................................................................................ 26
Statewide Network Preparation...................................................................................................... 26
Modal Categories .......................................................................................................................... 26
Purposes....................................................................................................................................... 27
Network Structure.......................................................................................................................... 27
Highway Network Attributes........................................................................................................... 28
Other Modal Networks................................................................................................................... 28
Zone Systems and Spatial Aggregation Issues .............................................................................. 29
Socioeconomic and Demographic Characteristics.......................................................................... 29
Other Data Issues.......................................................................................................................... 30
Model Steps......................................................................................................................... 30
Trip Generation Issues .................................................................................................................. 30
Trip Productions ............................................................................................................................ 30
Trip Attractions .............................................................................................................................. 32
External Stations ........................................................................................................................... 33
Balancing Productions and Attractions........................................................................................... 33
Trip Distribution: Gravity Model (Production Constraint) ................................................................ 33
DoublyConstrained Gravity Model ................................................................................................ 34
Friction Factors.............................................................................................................................. 35
Trip Distribution: Application to Statewide Forecasting.................................................................. 35
Mode Split ..................................................................................................................................... 36
Logit .............................................................................................................................................. 36
Logit Example, Forinash and Koppelman (1993)............................................................................ 37
Nested Logit .................................................................................................................................. 38
Nested Logit Example, More F&K.................................................................................................. 39
Mode Split Issues .......................................................................................................................... 40
Trip Assignment ............................................................................................................................ 40
Calibration and Validation.............................................................................................................. 41
Appendix: Overview of Four Step Models........................................................................ 42
Network Basics.............................................................................................................................. 42
Trip Generation Basics .................................................................................................................. 43
Trip Distribution Basics.................................................................................................................. 44
Mode Split Basics.......................................................................................................................... 44
Traffic Assignment Basics.............................................................................................................. 45
Chapter 4. Freight Forecasting.................................................................................. 46
Introduction......................................................................................................................... 46
Freight Model – Basic Steps.......................................................................................................... 46
Key Freight Data Sources .................................................................................................. 47
Commodity Flow Survey................................................................................................................ 47
Vehicle Inventory and Use Survey (VIUS)...................................................................................... 48
Commercial Freight Data Sources ................................................................................................. 48
Other Interesting Data Sources...................................................................................................... 48
Local Surveys of Note ................................................................................................................... 49
Defining the Scope of the Freight Model........................................................................... 49
Establish Goals for Model .............................................................................................................. 49
Table of Contents iii
Spatial Units.................................................................................................................................. 50
Selecting Modes............................................................................................................................ 50
Network Development ........................................................................................................ 51
Commodities ................................................................................................................................. 51
Developing a Flow Matrix .............................................................................................................. 51
Trip Generation for Gravity Model .................................................................................................. 52
Gravity Model for Freight ............................................................................................................... 53
Mode Split Methods....................................................................................................................... 55
Pivot Point..................................................................................................................................... 56
Elasticities..................................................................................................................................... 56
Commodity Density and Load Weights .......................................................................................... 57
Traffic Assignment......................................................................................................................... 58
Assignment Fixes.......................................................................................................................... 59
Other Freight Forecasting Issues ...................................................................................... 60
Combining Passenger and Freight Forecasts................................................................................. 60
Forecasting Inputs to a Freight Forecast ........................................................................................ 60
Chapter 5. Specialized Methods for Passenger Forecasting.................................. 62
Introduction......................................................................................................................... 62
Data Collection.................................................................................................................... 62
Validation Data Sets: HPMS Element of the Traffic Monitoring Program........................................ 62
Single Station OriginDestination Surveys...................................................................................... 64
Truck Travel Surveys..................................................................................................................... 65
Other Survey Methods................................................................................................................... 66
Validation Quality .......................................................................................................................... 67
Acceptable Error Plots................................................................................................................... 68
Stated Preference ............................................................................................................... 68
Value of Time and Value of Frequency  TriState......................................................................... 70
Single Mode Vehicle Occupancy: Michigan..................................................................... 70
Time of Day Considerations............................................................................................... 71
Interstate Trip Rate ............................................................................................................. 72
Development of Outstate Trip Generation for the Michigan Model .................................................. 72
External Travel for Other State Models .......................................................................................... 72
Composite Utilities for Trip Distribution ........................................................................... 73
Total Travel in Corridor: Wisconsin Model ...................................................................... 74
TriState Total Demand Model ....................................................................................................... 74
Hybrid Technique: Pivot Analysis .................................................................................... 75
Review of Select Link Analysis ...................................................................................................... 76
Pivot Steps.................................................................................................................................... 76
Pivot Method Refinements............................................................................................................. 77
Calibration and Validation.................................................................................................. 78
National Defaults........................................................................................................................... 78
Trip Table Estimation from Traffic Counts ...................................................................................... 79
SSOD in Validation........................................................................................................................ 79
Integration of Statewide and Urban Models.................................................................................... 80
Appendix: Review of Calibration Methods....................................................................... 80
Ordinary Least Squares (Linear Regression) ................................................................................. 80
Nonlinear Regression.................................................................................................................... 81
Guidebook on Statewide Travel Forecasting iv
Maximum Likelihood Estimation..................................................................................................... 81
Fast Trip Distribution Calibration.................................................................................................... 82
Acknowledgements..................................................................................................... 83
APPENDIX: The State of the Art in Statewide Travel Demand Forecasting .......... 84
A.1. Introduction ................................................................................................................ 84
Method of Research ...................................................................................................................... 86
Organization of Appendix .............................................................................................................. 86
A.2. Intercity Passenger Literature ................................................................................... 87
Types of Intercity Passenger Models ............................................................................................. 88
Aggregate DirectDemand Models................................................................................................. 88
Disaggregate Sequential Models ................................................................................................... 89
One Disaggregate DirectDemand Model ...................................................................................... 89
SingleMode and SinglePurpose Models ...................................................................................... 91
Discussion..................................................................................................................................... 91
A.3. Intercity Freight Literature......................................................................................... 91
Introduction ................................................................................................................................... 91
Reviews of Intercity Freight Models ............................................................................................... 92
Intercity Freight Models ................................................................................................................. 92
Intercity Freight ModeSplit Models................................................................................................ 93
Figure A.4. University of Pennsylvania ShipperCarrier Model....................................................... 93
Discussion..................................................................................................................................... 94
A.4. Statewide Passenger Forecasting Literature ........................................................... 94
Data Collection for Passenger Travel............................................................................................. 94
Data Synthesis for Passenger Travel ............................................................................................. 95
Trend Analyses of Passenger Travel ............................................................................................. 95
Statewide Models of Passenger Travel .......................................................................................... 95
Discussion..................................................................................................................................... 96
A.5. Statewide Freight Forecasting Literature ................................................................. 98
Introduction ................................................................................................................................... 98
Data for Freight Forecasting .......................................................................................................... 98
Statewide Models for Freight Forecasting...................................................................................... 99
Related Models ............................................................................................................................. 99
Discussion................................................................................................................................... 100
A.6. Two Recent Passenger Models ............................................................................... 100
The Michigan Passenger Model................................................................................................... 100
The Kentucky Passenger Model .................................................................................................. 106
Discussion................................................................................................................................... 108
A.7. Two Recent Freight Models..................................................................................... 111
The Wisconsin Freight Model ...................................................................................................... 111
The Indiana Freight Model ........................................................................................................... 112
Discussion................................................................................................................................... 116
A.8. Recommendations and Conclusions...................................................................... 117
Recommendations for Statewide Travel Demand Forecasting ..................................................... 118
Conclusions................................................................................................................................. 119
A.9. References................................................................................................................ 120
References for Section A.1.......................................................................................................... 120
References for Section A.2.......................................................................................................... 120
References for Section A.3.......................................................................................................... 123
References for Section A.4.......................................................................................................... 124
Table of Contents v
References for Section A.5.......................................................................................................... 129
References for Section A.6.......................................................................................................... 130
References for Section A.7.......................................................................................................... 131
References for Section A.8.......................................................................................................... 131
Introduction 1
Chapter 1. Introduction
This guidebook reviews the stateofthepractice of statewide travel forecasting. It focuses on
those techniques that have been considered essential to good statewide travel forecasting. In
addition, this guidebook presents specialized and advanced techniques of potential interest to
persons involved in statewide travel forecasting.
Emphasis is placed on practical methods. In some places in the guidebook, methods are
presented that have not been tried in statewide travel forecasting but show strong potential to
improve the process. This guidebook does not describe methods that have been presented in
the academic literature and are considered to be still under development. Persons interested in
recent research on this topic might want to consult the appendix, “The State of the Art in
Statewide Travel Demand Forecasting”.
There are many facets of statewide travel forecasting and it is not possible to create a onesize
fitsall model to deal with every possible situation. This guidebook advocates the need to select
the technique that is most applicable to the problem. In some cases, a simple growth factor
model may be preferable to a fullblown network analysis of the whole state.
This guidebook also makes a distinction between urban travel forecasting and statewide travel
forecasting. Although there are similarities in theory, the differences in implementation are quite
important.
This guidebook will make reference to travel forecasting software, but it does not provide
guidelines for any particular software package, nor are the recommended techniques dependent
upon the capabilities of any specific software package.
Structure of the Guidebook
After this introduction the guidebook consists of four chapters.
Time Series. Chapter 2 deals with methods for extrapolating upon existing trends in traffic
volumes. Topics range from simple growth factor methods to ARIMA (autoregressive,
integrated, movingaverage) models.
Passenger Forecasts. Chapter 3 deals primarily with means for adapting the traditional urban
“four step” model to statewide travel forecasting.
Freight Forecasts. Chapter 4 deals with the need to integrate freight forecasts with passenger
forecasts. It discusses all freight modes, but emphasizes heavy trucks. Major portions of the
chapter concern freight data sources.
Specialized Methods for Passenger Forecasting. There are many forecasting situations that are
not addressed well by a traditional “four step” model. Chapter 5 lists several methods that have
occasional value for statewide, corridor or intercity travel forecasts.
Reasons for Statewide Travel Forecasting
There are a number of reasons why a state might be interested in forecasting statewide or rural
travel.
Forecasts of Rural/Intercity Travel. Overall assessments of the adequacy of the statewide
transportation networks require forecasts of rural and intercity travel by all freight and passenger
Guidebook on Statewide Travel Forecasting 2
modes. Such forecasts can be helpful in programming the sequence of projects and their
associated costs.
Supplement Urban Forecasts. Planners in most sizable urban areas have the ability to forecast
traffic levels in their communities. A large portion of travel in most states is rural. Thus,
investigations of statewide or national transportation policies would be incomplete without
forecasts on rural highways and other intercity transportation modes. Indicators (such as VMT,
air pollution emissions and consumer surplus) require forecasts from both urban and rural
areas. In addition, statewide forecasts can be helpful to urban area forecasts by providing
information on through trips.
Satisfy Mandated Planning Requirements. TEA 21 mandates that several issues must be
considered in statewide transportation plans. The study of many of these issues can be
facilitated by a good, multimodal or intermodal travel forecasting model.
Develop ProjectLevel Forecasts in Rural Areas. The sizing of facilities in the design process
requires accurate estimates of future travel. Time series and hybrid techniques (time series
combined with conventional fourstep models) can be particularly useful for projectlevel
forecasting.
Original ISTEA Planning Factors
ISTEA was more specific than TEA 21 in its required planning factors, and the relationship
between the ISTEA factors and travel forecasting is somewhat clearer. Those factors that
closely relate to travel forecasting are:
♦ Energy use;
♦ Border crossings, major transportation facilities and military;
♦ Connectivity between metropolitan areas;
♦ Efficient use of existing facilities;
♦ Traffic congestion reduction; and
♦ Efficient movement of commercial vehicles.
These factors are consistent with TEA 21.
Overview of Approaches
There are three different approaches to travel forecasting that are of interest to planners in state
DOTs: statewide, corridor and project. All three approaches are covered by this guidebook.
Statewide. Statewide forecasts most often require a full “fourstep” simulation that considers
each of the following elements:
♦ All passenger modes, such as auto, intercity bus, conventional rail, highspeed rail and
air;
♦ All freight modes, such as truck, rail, water and air;
♦ All passenger travel purposes, such as commuter, business and recreation;
♦ All freight commodities, such as lumber, machinery and agricultural products;
♦ All or many times of day;
♦ All classes of facilities;
♦ Intermodal transfers; and
♦ Future growth or changes in industry and population.
Introduction 3
Corridor/Intercity. Often the total demand in the corridor by time of day is considered constant
or estimated externally. Thus, the forecast becomes an exercise in mode split and traffic
assignment. Mode split models of some sophistication are often chosen to give precise
estimates of modal shares.
Project. Project level forecasting is often of shorter term with few unforeseen intervening
factors. In many cases, project level forecasts can be made by extrapolating upon current
trends. Thus, time series methods are of greatest interest for project level forecasts.
Time Series Methods
A time series may consist of traffic levels, population levels, employment levels or any other
socioeconomic or demographic characteristic of interest to forecasting. A time series may be of
interest by itself or important as an input to a fourstep model.
Growth Factor Methods. Growth factor methods are simple time series techniques that assume
that the rate of growth is constant over time. Growth factor methods are recommended in
FHWA’s Traffic Monitoring Guide for interpolating traffic counts on road segments that are
missing data elements for one or two years.
Linear Regression and Extensions. There are many time series methods based on statistical
theory  most often linear regression theory. Linear regression is a welldeveloped technique
for fitting lines to XY data. The choice of formulation depends upon the nature of the data and
the eventual use of the forecasts.
♦ Simple trend models assume that the yeartoyear change is constant, differing from
growth factor models that assume that the percentage of change from year to year is
constant.
♦ Moving average models attempt to eliminate bumpiness within a data series by
averaging a few items that are close together in time. Moving average methods can be
used to eliminate seasonal, weekly and diurnal fluctuations in data. Other moving
average methods can assure that only the most recent data is used in a forecast.
♦ ARIMA techniques form a class of models for fitting complex time series, particularly
those with seasonal fluctuations. ARIMA models are sometimes referred to as Box
Jenkins models. Many transportationrelated time series can be best forecasted with a
BoxJenkins model. A full understanding of BoxJenkins models require a good
knowledge of statistics, but the underlying principles can be easily explained.
Overview of Three/Four Step Models
Many states trying to build statewide travel forecasting models are doing so using the same
theory and software used for urban models. This strategy may be appropriate or inappropriate,
depending upon the policies or projects that are being evaluated.
The four traditional steps of an urban model are trip generation, trip distribution, mode split and
trip assignment. The purposes of these steps will be explained later. There are other required
steps of less importance. A threestep model would not include mode split. Such a model is
typically used for forecasting automobile traffic on highways.
There are enough differences between statewide and urban forecasting to require changes to
most of the steps. The ability of a given software package to model statewide travel is an issue
to be considered, but the process should not be arbitrarily limited to the capabilities of any
particular commercial software product.
Guidebook on Statewide Travel Forecasting 4
The fourstep models discussed in this guidebook are macroscopic in nature. They deal with
groups of trips, travelers and vehicles rather than individual travelers and vehicles.
“Microsimulation” methods have not been tried at the statewide level and are still considered
experimental at the urban level.
Commodity Based Freight Models. This guidebook recommends that freight traffic be predicted
from commodity flows. The recommended forecasting method makes effective use of available
data sources and knowledge of freight flow processes. This method of modeling freight flow
parallels the process for modeling passenger movements.
Calibration and Validation
A major emphasis of this guidebook is on effective calibration and validation of models, once
they are created. Calibration and validation involve several distinct processes.
Adjustment of Parameters. A travel forecasting model has literally dozens of parameters that
must be individually set for any given forecast. The process of setting all these parameters is
called “calibration”. Most importance to statewide forecasting are the parameters related to trip
generation (attractions and productions), mode split and trip distribution  particularly gravity
model friction factors.
Refining the Trip Table. Techniques have been developed to estimate trip tables from traffic
counts. These techniques cannot estimate a trip table from scratch, but can be useful for
modifying an existing trip table (perhaps from the gravity model) to better match traffic counts.
Employing such a technique may be better than applying a set of ad hoc adjustments (known as
kfactors in the parlance of the gravity model).
Adjustment to Traffic Assignment Inputs. Urban models often go through an extensive network
“calibration” procedure where various link attributes, such as speed and capacity, are adjusted
to achieve better agreement with ground counts. For statewide models there is less opportunity
for adjustment, because there are fewer attributes; however, some states have found it useful to
adjust speeds by a formula to account for driver preference for certain routes, for aggregation
problems in the definition of the network and for lack of continuity of routes at state borders on
the network.
Base Case Comparisons. An important step in any model development is to compare a base
case (or year) forecast against known traffic counts. This comparison will not, by itself, assure
that future year forecasts are valid. However, the comparison will demonstrate that major
relationships have been simulated with some degree of accuracy and consistency.
Specialized Methods
There are several specialized methods to help solve particular problems related to corridor or
intercity forecasting. They can constitute a complete forecasting procedure by themselves or
enhance an existing model.
Hybrid Technique. The “pivot” method uses outputs from a travel forecasting model and from a
timeseries model to provide precise forecasts on one or a few highway segments. The forecast
is made relative to existing traffic volumes.
Nested Logit. The nested logit model is a means of simultaneously forecasting traffic and
patronage on a variety of intercity modes. The nested logit model has the ability to perform
correct forecasts when there are many modes, some of them being close substitutes for each
other.
Introduction 5
Total Corridor Demand. Demand within a corridor is often modeled as a function of
socioeconomic factors. Methods exist for forecasting total demand without needing to run a full
blown fourstep model.
Stated Preference. Stated preference techniques ask travelers about hypothetical modal
choices to determine the ridership potential of a new mode.
Introduction to Data Sources
This chapter will introduce a few data sources, but most discussion of the use of the data will be
handled in later chapters.
Alternatives to Calibration. Calibrated model steps provide the greatest policy sensitivity.
However, if the policy being evaluated does not use a calibrated model step in a meaningful
way, then the effort to calibrate is wasted and more efficient ways of forecasting may be more
appropriate. For example, a good commodity mode split model is very difficult to calibrate. If
one were available, it would be quite useful in determining how commodity mode shares are
affected by changes in shipping costs or improvements in service quality. However, it is entirely
possible that the state is uninterested in policies that might affect costs or quality. In that case,
the calibrated mode split model is not helping the forecasting process.
Alternatively, commodity mode split may be represented by a series of lookup tables developed
from historical data. Such look up tables can be tabulated from the Commodity Flow Survey
without a great deal of difficulty.
Spatial Aggregation Issues. Data is reported at different levels of spatial aggregation,
depending upon the source. Generally speaking, lower levels of spatial aggregation result in
better forecasts but cause an increase in costs and time for analysis. Possible levels of spatial
aggregations include: traffic analysis zones (TAZs), counties, municipalities, states (outside
your state), NTARs and BEA (Bureau of Economic Analysis) regions.
Moving data from one level of spatial aggregation to another is often a difficult process, so the
primary spatial unit must be chosen carefully at the beginning of model development.
Key Data Sources
Appropriate use of existing data can speed the development of statewide models. Some of the
better sources of data are public agencies, but private organizations can also provide data for
statewide travel forecasting.
Census Bureau. The Census Bureau provides complete person and household data every ten
years and data on a wide variety of socioeconomic conditions at other intervals.
Bureau of Transportation Statistics (BTS). The BTS is a source of many of the public databases
useful for statewide travel forecasting. All products are available from its web page:
www.bts.gov.
Proprietary Data Sources. Several companies provide proprietary data, demographic forecasts
and economic forecasts. Many of the proprietary databases deal with lower levels of
aggregation than public sources.
Original Data Collection. A high quality forecast will require some original data collection
beyond traffic counts. These might include:
♦ Travel surveys of households
♦ Surveys of drivers of passenger and freight vehicles at cordon stations;
Guidebook on Statewide Travel Forecasting 6
♦ Surveys of travelers within a corridor to ascertain their preference for new modes
(behavioral intention); and
♦ Origindestination information collected at a single, high demand point within the state
(single station origindestination survey).
Key Census Databases
Decennial Census. Specialized products from the decennial census have included questions
related to journey to work (JTW) and the Census Transportation Planning Package (CTPP).
The CTPP provides data aggregated to the TAZ (traffic analysis zone) level as defined by
MPOs. The CTPP also provides information on the number of employees by zone of
employment.
Economic Census. This data set includes: number of establishments (or companies); number
of employees; payroll; and measures of output (sales, receipts, revenue, value of shipments or
value of construction work done).
Census of Agriculture. All operators provide crop acreage and quantities harvested, inventories
of livestock and poultry, value of products sold, land use and ownership, irrigation activities,
amount of commodity credit loans, number of hired laborers, Federal program payments and
operator characteristics. Selected operators provide additional information on production
expenses (including interest), fertilizer and chemical use, machinery and equipment, market
value of land and buildings and income from farmrelated sources.
Commodity Flow Survey. This data set is derived from a sample of shipments from the US
covering most commodities and modes. Data are reported at the national, state and NTAR
levels.
Census of Manufacturers and Manufacturers Survey. Basic data from this data set include kind
of business, location, ownership, value of shipments, payroll and employment. Additional data
collected include cost of materials, inventories, new capital expenditures, fuel and energy costs,
hours worked and payroll supplements. Mining is included in a separate but similar database.
National Personal Transportation Survey (NPTS) Overview
The NPTS is a household travel survey last updated in 1995. The survey collects data from a
random, stratified sample of over 40 thousand households about all personal trips, by all modes,
for all purposes. Trips were included for all persons age 5 and older.
Household data in the NPTS include: household size, number of household vehicles, income
and location. Person data include: age, gender, education, relationship within the household,
driver status, annual miles driven if a worker, worker status, if drive as an essential part of work
if employed and seat belt use. Vehicle data include: annual miles driven, make, model and
model year. Trip data include: trip purpose, mode, length (in miles and minutes), time of day,
vehicle characteristics (if a household vehicle was used), number of occupants and driver
characteristics.
American Travel Survey (ATS) Overview
The ATS is a survey of long (greater than 100 miles) trips. Approximately 80,000 households
participated. Interviews were conducted approximately every three months by phone and in
person. Trip data included: the origin and destination of the trip, stops along the way and side
trips from the destination, the principal means of transportation, the access and egress modes
to airports, train and bus stations, information about the travel party, reason for the trip, number
Introduction 7
of nights spent away from home and the type of lodging. Route distances of all trips were
calculated from a network.
Important Freight Data Sources
Commodity Flow Survey (CFS). The CFS is the most complete single public source of
information on freight flows in the US. National data provides summaries by threedigit STCC,
state and NTAR data by twodigit STCC. Raw data are not available. National data will give
distance by mode by threedigit STCC. The CFS will be discussed extensively in Chapter 4.
Commercial Freight Flow Products. Reebie, for many years, has provided the TRANSEARCH
database and custom data products from TRANSEARCH. TRANSEARCH provides traffic
statistics between BEA regions by mode (water, rail, air and truck) and by commodity. It
incorporates data from a “significant number” of truckload and LTL carriers, as well as a wide
variety of publicly available databases. Reebie is currently developing a product, called the
Intermodal Freight Visual Data Base, that disaggregates the TRANSEARCH data to counties for
statelevel forecasting purposes.
Commercial Economic Forecasts
Many firms provide economic forecasts that can be useful for statewide travel forecasts.
REMI. REMI does regional forecasts with a model called Policy Insight. Simulations with the
model are used to estimate the economic and demographic effects of economic development
programs and transportation policy changes.
1
Woods & Poole. Woods & Poole’s database contains more than 550 economic and
demographic variables for every county in the United States for every year from 1970 to 2020.
This database includes population by age, sex and race; employment and earnings by major
industry; personal income by source of income; retail sales by kind of business; and data on the
number of households, their size and their income.
2
NPA. NPA’s database is also aggregated at the county level. The database contains 212
economic or demographic items for the years 1967 to 2025. Also contained in the database are
twelve categories of residential and nonresidential buildings and fifteen factors determining
demand for new construction from 1980 to 2005.
Tools
Role of GIS
A geographic information system (GIS) is a mechanism for storing, retrieving, visually
representing and analyzing spatial data. Many states (and the federal government) are creating
GIS databases containing information useful for statewide travel forecasting.
Of particular interest here are those data from a GIS that allow for rapid development of a travel
forecasting network. These data items include the location of intersections, the width of road
segments, other information related to determining capacity and the types of traffic control at
intersections and along uncontrolled road segments. Also available in GIS form are jurisdiction
1
REMI’s web site
2
Woods & Poole’s web site
Guidebook on Statewide Travel Forecasting 8
boundaries, boundaries of spatial data units, locations of bodies of water, land use inventories
and socioeconomic and demographics information. GISs can also store data useful for
validation purposes, such as traffic counts.
Even though the existence of a GIS can speed data preparation, it cannot totally automate the
process of network creation. There are substantial incompatibilities between GIS data
structures and those necessary for travel forecasting. Often the GIS provides too much detail,
fails to properly show highway connectivity or continuity, or is not digitized with sufficient
accuracy. Coding errors that can be minor for typical GIS applications (maps, statistical
summaries, etc.) can be catastrophic to a travel forecast.
By and large, travel forecasting cannot be performed on a GIS (there is one notable exception),
but many travel forecasting packages have mechanisms for transferring data to and from GISs.
Software Issues
Many organizations involved in travel forecasting have found it necessary to establish a library
of programs for simulation and calibration.
FourStep Model Packages. Several software products are available to forecast urban travel
using the standard four steps. These software products are essentially similar internally (i.e.,
they share the same algorithms for computation), but differ considerably in their user interfaces.
Models typically come with a graphics network editor to facilitate the input of nodes, links and
their attribute values. Some packages come with interfaces to GISs. The needs of statewide
travel forecasting have not been a priority in the design of these packages.
Mode Split Model Calibration. Mode split models are calibrated by applying statistical principles
to observed travel patterns and mode choices. These statistical methods are somewhat
unusual (especially when dealing with the “logit” model and its derivatives), so specialized
products (separate from the fourstep model) are often necessary to accomplish the calibration.
Statistical Packages. Calibration of other parts of the fourstep model is often best
accomplished with a standalone statistical package that contains linear regression, analysis of
variance (ANOVA), tests of significance and singlevariable descriptive statistics. A standalone
statistical package will also have good time series analysis capabilities. A subset of the
capabilities of a standalone statistical package may be found in a spreadsheet package, which
is often the most convenient and transferable method of performing a calibration.
User Knowledge. All of these packages require training on the part of the user. Training
periods can be considerable, depending upon what must be accomplished. Training consists of
learning the underlying theory and learning the user interface for the particular software
package.
Composing a Complete Forecast
Ultimately, a complete travel forecast should be able to produce estimates of link volumes for all
intercity modes. This forecast would include both passenger and freight vehicles, any relevant
urban forecasts and several measures of effectiveness (MOEs), including delay, energy
consumption and emissions. However, a complete travel forecast is not always necessary,
especially for project and corridor studies. This guidebook will not describe measures of
effectiveness (MOEs).
Time Series Methods 9
Chapter 2. Time Series Methods
Introduction
This chapter presents several methods of time series analysis that have proven useful in short
term statewide, rural or intercity forecasts. The chapter is principally concerned with:
♦ Growth factors;
♦ Trend analysis with linear regression; and
♦ BoxJenkins (ARIMA) methods.
Several enhancements to these techniques are described, including:
♦ Moving averages;
♦ Data transformations;
♦ Forecasting differences;
♦ Including economic and demographic factors;
♦ Including factors relating to the state of the system; and
♦ Handling seasonality.
Time series analysis is a branch of statistics. A good understanding of elementary statistical
concepts and the use of statistical software packages would be required for successful
application of the techniques.
What is Time Series Analysis?
Time series analysis is a means of understanding data variability over time. Because a time
series model exclusively represents past events and relationships, it can be used to forecast the
future as long as the future is expected to behave like the past.
Some of the more elementary time series methods require only readily available historical data,
so they provide quick answers. Time series analysis is particularly appropriate when the
forecast is short term and there is insufficient time and resources to build and calibrate a
behavioral model.
Given more time and a broader set of data, rather sophisticated time series models can be built.
The models can handle more than simple trends (growth and decline). They can also consider
cycles in the data (annual, weekly, daily), discrete changes to some important influential factors
and trends in important factors.
Applicability of Time Series Analysis
Modal Considerations. All modes can be analyzed with time series, but approaches may differ.
Time series analysis can be especially helpful for short term forecasts where behavioral models
have not been calibrated or input data are unavailable.
Policy Considerations. Time series can be used to forecast data needed for policy analysis,
including:
♦ Attributes of traffic, such as vehicle occupancy, vehicle weight and vehicle classes;
♦ Enforcement needs;
♦ Economic trends;
♦ Environmental conditions; and
♦ Growth in competing modes.
Guidebook on Statewide Travel Forecasting 10
Data Considerations. When doing a multimodal forecast on a network, time series are useful
for:
♦ Forecasting inputs to trip generation;
♦ Forecasting comparatively minor modes (e.g., air freight, barges) when data limitations
or time constraints prohibit application of a behavioral model;
♦ Forecasting external travel;
♦ Placing bounds on the reasonableness of forecasts; and
♦ Determining seasonal, monthly or day of week adjustment factors for postprocessing
results from a behavioral model.
Cautions. Great care must be exercised when forecasting traffic volumes. Time series analysis
has a limited ability to anticipate changes in future conditions. Events that have never before
occurred cannot be modeled. Nor is it possible to model the effects of an existing causal factor
that has not changed appreciably in the past. Like any other model, time series cannot
anticipate rare future conditions or events.
Important Statistical Concepts
This guidebook cannot provide a complete background in statistical theory. However, some
knowledge of elementary statistical concepts is required to understand most of the methods
presented in this chapter. Of particular importance is the concept of a correlation.
Correlation. A correlation coefficient ranges from 1 to +1. The sign of a correlation coefficient
is the same as the sign of the slope of the line drawn through the points on a (XY) scatter
diagram. The magnitude relates to the quality of fit to a line. A typical correlation would
describe: Traffic Volume versus Total Personal Income (either over time or over space).
The degree of agreement can be computed for unlike variables (correlation) or for data within
the same series spaced at a fixed time span apart (autocorrelation).
The correlation coefficient between two variables has only limited use in time series analysis of
traffic. Seasonal effects and other periodicity effects tend to cause lower values of correlation
coefficients. The concept of autocorrelation is often better for understanding periodicity.
It is always necessary to understand the reasons behind a large correlation coefficient. A strong
but spurious correlation can exist between two entirely unrelated variables.
Autocorrelation. An example autocorrelation might be:
Traffic in Year t versus Traffic in Year tn
For example, if n is 5, the correlation would find the agreement between:
Traffic in 1997 versus Traffic in 1992
Traffic in 1996 versus Traffic in 1991
Traffic in 1995 versus Traffic in 1990
etc.
Autocorrelations tend to diminish as n increases, unless cyclic events are present in the data.
For example, a strong autocorrelation may exist between Traffic in July of Year n and Traffic in
July of Year n1 as well as between Traffic in July of Year n and Traffic in June of Year n.
Additional elementary statistical concepts are discussed in the Appendix to this chapter.
Time Series Methods 11
Growth Factors
The growth factor method is a popular way of forecasting trends in variables that have been (by
and large) increasing in time. Growth factors work best on time series where the change from
period to period is proportional to the size of the series. A good example of a series that
behaves this way is population. With a constant fertility rate and death rate and with zero net
migration, the number of new people each year is proportional to the number of people already
alive.
Transportation data series can sometimes be accurately modeled by growth factors. Growth
factors work best when the variable to be forecasted is heavily influenced by other variables that
inherently grow proportionally. For example, recent increases in the number of licensed drivers
are heavily influenced by increases in population. Many transportation variables are heavily
influenced by the overall size of the economy, which has grown steadily over time.
A growth factor is easily computed for any data series. With only a little more work (and
spreadsheet software or a statistical package), it is possible to create a growth factor model that
incorporates external influences.
Growth Factor Relationships
For a data series Y, the next period t+1 can be forecast from current period t times a growth
factor (1 + i):
In this case, i is the growth rate, expressed as a fraction of current traffic levels. Y
t
is the traffic
in the current year, and Y
t+n
is the traffic in the nth year beyond the current year. N periods can
be forecast by repetition:
The above equations behave similarly to
compound interest. Because of the
compounding, a growth factor model (with
a positive i) will be upward bending over
time. The curve on the right illustrates a
typical pattern of growth of traffic on
highways in developing parts of a state.
Take particular note of the “S” shape of
the curve. It is important to select a
method of replicating the time series that
bends appropriately. For example, a
simple growth factor model will not
properly replicate the portion of the curve
denoted as “developed”, as it is downward
bending. In some cases it might be
necessary to gather additional information
to understand the position of the data
series on the “S” curve.
( ) i 1 Y Y
t 1 t
+ ·
+
( )
n
t n t
i 1 Y Y + ·
+
Time
T
r
a
f
f
i
c
Rural
D
e
v
e
l
o
p
i
n
g
D
e
v
e
l
o
p
e
d
Guidebook on Statewide Travel Forecasting 12
Clusters. Clusters can help span gaps in existing data. Clusters can be created to group traffic
counting stations that have behaved similarly in the past. The Traffic Monitoring Guide gives
some recommendations along these lines. The guide suggests, at a minimum, highway clusters
consisting of:
♦ Interstate rural;
♦ Other rural;
♦ Interstate urban;
♦ Other urban; and
♦ Recreational highways.
These clusters are aggregations of functional classes. Clusters created for traffic monitoring
can be used for forecasting, too. Such clusters will allow seasonal (or daily) forecasts where
seasonal data are unavailable.
Similar concepts can be applied to other modes that may have strong seasonal or daily
variations in traffic levels.
Limitations and Site Specific Uses. States are encouraged to perform their own cluster analysis
to identify reasonable clusters for their traffic data. The previously cited clusters should be
considered a rough guideline and a starting point. The Traffic Monitoring Guide provides
information on how custom clusters may be developed. When looking at data from specific
sites it is reasonable to assemble an ad hoc set of similar sites, rather than to rely on statewide
clusters. The ad hoc set, if well chosen, should give a better indication of daily and seasonal
variations in traffic.
Causality. Traffic volumes are often highly correlated with socioeconomic and demographic
variables. For example, most places have a causal relationship between traffic volume and total
personal income. Personal income forecasts are often readily available, so including personal
income into the model makes sense (both theoretically and practically).
It is important to avoid variables that are not causal, even when the correlation is high. For
example, a high correlation could exist between traffic on a state highway and the number of
admissions at a local hospital. However, a surge in hospital admissions would not imply a big
increase in traffic volume.
Additionally, it is import to understand the direction of causality. The number of speeding tickets
could also be highly correlated with traffic volume. If we were to reduce the speed limit, tickets
for speeding might increase but the traffic would probably hold steady (or maybe drop).
Linear Trend Model
Linear regression is a technique for fitting straight lines to data. Linear regression finds the best
coefficients by minimizing the sum of the squares of all residuals (misfits of the line to data
points). Linear regression can be used to extrapolate a time series into to the future.
If traffic (Y) in year t increases linearly with time, then the following equation should hold:
For the linear trend model shown above, the two coefficients, b
0
(the yintercept) and b
1
, (the
slope) are estimated from data using linear regression. Spreadsheet software packages come
with versatile linear regression modules.
t 1 0 t
t b b Y ε + + ·
Time Series Methods 13
The independent variable t can be the exact date or a sequence index from the beginning of the
series. Note that the error term ε
t
makes this equation exact for any time within the series.
Linear regression can be used to model curvilinear trends by including higher order terms. For
example:
Y
t
= b
0
+ b
1
t + b
2
t
2
+ ε
t
could be used to model a time series that bends down or up.
The output from a linear regression includes measures of goodness of fit. Standard errors and
tstatistics are provided for each coefficient, and Rsquare values are provided for the whole
equation (see the Appendix of this chapter for descriptions of these statistics).
Including Causality. A more robust form of linear regression would estimate a model consisting
of a single trend term and several terms for causal variables. Again, Y
t
is the traffic in year t.
The causal variables (x’s) may consist of economic or demographic indicators or may represent
the state of the system in any given time period.
The variable to be explained, Y
t
in this case, is called the dependent variable. The explanatory
variables, t and x’s, are called independent variables.
Choosing Variables. There is no single method for determining whether an independent
variable should be included in the equation. Often people will try every possible variable that
might be relevant, then let the software select the set that best explains the dependent variable,
as measured by Rsquare. This method ignores causality and the possibility of spurious
correlations, leading to a model with dubious forecasting validity.
Furthermore, including two independent variables that previously behaved similarly can distort
the contributions of one of them. For example, we might be able to predict the amount of traffic
on a state highway by using both population and employment in a neighboring city. However,
this city may have had an almost constant relationship between population and employment
over the years. If this relationship were to change (a recession, for instance), then the model
may lose its predictive ability. When two independent variables are strongly correlated they are
said to be collinear. When a model has more than two independent variables that are strongly
correlated, then the model is said to possess multicollinearity.
Each independent variable must make a unique, causal contribution to the model. It is
important to rely on theories of travel behavior to indicate what those variables might be.
Sometimes, the regression analysis might reject a variable that seemed at first reasonable, but
fishing expeditions are never recommended.
Dummy Variables. Dummy (0,1) independent variables are quite useful for showing effects that
occur only part of the time. For example, dummy variables can show holidays, special events
and step changes to the economy, facilities or the environment. A dummy variable is set to 0
for some of the time and 1 for the rest of the time. The coefficient of a dummy variable can be
interpreted as an additive effect. When the dummy variable is 1, the coefficient is added to the
estimate of the dependent variable. When the dummy variable is 0, the coefficient is ignored.
t n n 2 2 1 0 t
x b x b t b b Y ε + + + + + · L
Guidebook on Statewide Travel Forecasting 14
Use of Error Terms. Error terms are explicitly considered when fitting a model to data, but error
can never be predicted. Consequently, all error terms are replaced by the “expected value” of
the error, which is zero.
Difference Model. A difference model forecasts the change in the data series between two
successive periods. A simple difference model takes the following form:
In words, a difference model forecasts the next period knowing the value of the data series in
the current period. The δ is the yeartoyear difference. Of course, causal variables may be
added to a difference model.
BoxCox Transformations
A mathematical assumption of linear regression is that the standard deviation of the errors in
estimates is constant throughout the series. This assumption is not always valid for traffic data.
A BoxCox transformation is sometimes needed improve the uniformity of the standard
deviation.
Possible BoxCox transformations are given by this equation.
In BoxJenkins (ARIMA) models (to be discussed later) BoxCox transformations are often
employed to reduce the effect of the error term in larger values of the data series. Thus, β’s of 0
(giving a logarithmic transformation) or 0.5 are frequently used. The log transformation is most
logical when a constant rate of growth can be assumed and errors increase in proportion to the
size of the data item. This type of error might be found in traffic data that is affected by single
large influencing factor, such as the overall economy of the state.
When errors may be due to many random influences, then a square root (β = 0.5)
transformation is often useful. Traffic data often behaves this way, too.
When the standard deviation is constant, the errors are said to be homoscedastic. Thus, the
quality of the model can improve when the original data is transformed to make the errors
homoscedastic.
A BoxCox transformation can also be used to model curvilinear trends. The recommend
procedure for BoxCox transformations in this case is to transform the data series for several
values of β, fit the data series with linear regression, then choose the β corresponding to the
best model as indicated by its Rsquare.
Once the coefficients of the model are found for the transformed series, the relationship must be
untransformed to get estimates of the original series.
1 t t 1 t
Y Y
+ +
ε + δ + ·
¹
¹
¹
¹
¹
'
¹
· β
≠ β
β
−
·
β
β
0 Y log
0
1 Y
Y
Time Series Methods 15
Moving Averages and Seasonality
There are two forms of moving averages commonly used for simple time series forecasting:
exponential smoothing and unweighted moving averages over a cyclic time period. Both forms
of moving averages may be applied to either trend or difference models.
Exponential Smoothing
Exponential smoothing uses a declining series of weights, starting with α (0 < α < 1) and
declining with the age of data,
so that the newest values in the time series get the largest weights. The average age of an
exponentially smoothed series is (1  α)/α. Observe that the rate of decline of weights for the
moving average relates to the value of α. Values of α near 1 result in a fast decline. Values of
α near 0 result in a slow decline.
Exponential smoothing is best applied to situations where cyclic patterns are not present in the
data series.
Central Moving Average
Good time series models require many data points (50 or more), so it is often necessary to
adopt a period of less than a year. When a period length of one month has been selected,
strong seasonal fluctuations in traffic data often become apparent.
When cyclic patterns are present in the data, they are often removed by taking a central moving
average. A central moving average uses data to each side of the current time period. A moving
average based on the current period and n1 previous periods can also be used, where n is the
number of periods in a cycle.
When there is an odd number of periods in a cycle, then a central moving average is found by
averaging the current period with the (n1)/2 periods before and after. Thus, a moving average
for Friday in a weekly cycle would average all data from the Tuesday before to the Monday
after. When there is an even number of periods in a cycle (months in a year or seasons in a
year), then the central moving average for period t is found by averaging together these periods:
t, t + 1, t – 1, … t + (n – 1)/2, t – (n – 1)/2 and onehalf of t + n and t – n
This method produces a seasonally adjusted trend, where there is a seasonal adjustment factor
(a multiplicative constant) for each time period in a cycle. A monthly index (MI) can be
computed to convert the moving average (MA) back to a seasonal forecast:
MI = Average of (Data/MA) for month
In a similar fashion, seasonal adjustment factors can also be computed to convert raw monthly
data to a yearly average.
( ) ( ) L
2
1 1 α − α α − α α
Guidebook on Statewide Travel Forecasting 16
Introduction to BoxJenkins (ARIMA) Methods
BoxJenkins is a large family of time series models that are able to track very complex historical
patterns. There are many variations of BoxJenkins models. The choice depends on the data
series to be analyzed. It is possible to include causal variables in a BoxJenkins model, so this
class of models can be quite robust.
Statisticians want 50 or more data points for BoxJenkins (ARIMA) methods, so they are
probably not applicable to analysis of yearly traffic data. Monthly, daily, or (even) hourly data
should produce satisfactory results.
Major commercial statistical packages contain BoxJenkins routines. It is not possible to
perform an interesting BoxJenkins analysis on a spreadsheet. BoxJenkins models are fit to
data so as to minimize the sum of squares of the residuals, just as in linear regression.
Therefore, many of the statistical concepts of linear regression also apply to BoxJenkins
models.
ARIMA
BoxJenkins models consist of one or more of the following elements:
♦ Autoregressive (AR);
♦ Integrated (I);
♦ Moving Average (MA); and
♦ Combinations: AR, ARMA, ARIMA, MA or IMA.
Autoregressive. In an AR model, the value of the data series is estimated with one or more
earlier values of the data series.
Integrated. An I model estimates the difference in data values in the series or the difference of
differences.
Moving Average. In an MA model, the data series is estimated using knowledge of the error in
a recent estimate.
Some of the most interesting BoxJenkins models combine the three types. The examples in
the next subsections illustrate the most elementary forms of these models.
Autoregressive (AR) Models
AR (Auto Regressive) models forecast period t knowing the conditions in period t1 and,
perhaps, prior periods. The premise of an AR model is that time series data rarely takes wild
jumps or dips. The best predictor of a period is the immediate past period. For example, an
AR(1) model looks like:
when δ (delta) and ϕ (phi) are statistically estimated parameters. In this model δ is a constant
period increment over a ϕ portion of the previous value. Again, ε
t
is an error term (which is not
used in a forecast). The theory underlying BoxJenkins models tells us that the absolute value
of ϕ is always less than 1.
An AR(2) model would take portions of two previous periods and look like:
t 1 t t
Y Y ε + ϕ + δ ·
−
Time Series Methods 17
There is no theoretical limit as to the number of terms in an AR model, but the contribution of
each successive term tends to weaken with age.
Once the coefficients have been estimated, an AR model is easier to compute and understand
than a growth factor model.
Integrated (I) Models
I (Integrated) models forecast the change from one period to the next. For example, a pure I
model looks like:
The δ term is the estimated average change between periods. Pure integrated models tend not
to be very interesting. The power of an integrated model is best seen when combining it with
AR and MA models and when including causal variables.
Integrated models work best when the longterm trend is stable but the variation from period to
period contains strong random influences.
Logarithmic Transformations of Integrated Models. Economists sometimes apply a logarithmic
transformation to their independent and dependent variables within integrated models. Thus,
they forecast the change in the logarithm of the data series. The coefficients thus obtained can
be interpreted as elasticities. For example, an integrated model that predicts traffic (Y) as a
function of employment (E) might look like this.
The b coefficient is interpreted as an elasticity, that is, the percent change in traffic given a one
percent change in employment.
Moving Average (MA) Models
MA (Moving Average) models forecast period t knowing the error in the t1 period forecast and,
perhaps, prior periods. For example, an MA(1) model looks like:
where µ (mu) and θ (theta) are statistically estimated coefficients. Moving average models
include past errors to get the estimate of values in the data series. In the above formulation, the
additional error term is used to get a good estimate of µ; it cannot be known when forecasting
more than one period beyond the present. For an MA(1) model the absolute value of θ must be
less than 1. The negative sign preceding the moving average term is traditional, suggesting a
correction effect. The coefficient µ is the estimated average value of the series.
t 2 t 2 1 t 1 t
Y Y Y ε + ϕ + ϕ + δ ·
− −
t 1 t t
Y Y ε + δ · −
−
1 t t t
Y
−
θε − ε + µ ·
( ) ( ) ( ) ( ) ( )
t t t t t
E log E log b Y log Y log ε + − + δ · −
− − 1 1
Guidebook on Statewide Travel Forecasting 18
An MA(2) model would look like:
Steps in Building an ARIMA Model
The following five steps should be followed when building an ARIMA model:
♦ Plot and study data;
♦ Apply knowledge and intuition;
♦ Get indicator statistics, such as “autocorrelations”;
♦ Specify model, transform data and fit; and
♦ Analyze strength of model and individual components.
Intuition and common sense are necessary for creating a good BoxJenkins model. Usually, a
quick glance at the data series is enough to judge the best combination of elements. The major
choices are:
Should the model be integrated (i.e., deal with differences)?
How many moving average terms must be included?
How many autoregressive terms must be included and what are the lags?
Do the periodtoperiod variations in the data suggest that the data series should be
transformed?
An autocorrelation is the correlation between Y(t) and Y(tn) where n is a selected number
between 1 and the size of the time series. Any series has many autocorrelations. A partial
autocorrelation is a measurement of the improvement to the model by the next term
representing the (n1) period when period n and all later periods have already been included.
Statisticians use autocorrelations and partial autocorrelations to help determine the structure of
a BoxJenkins model. For example, an AR(1) model is characterized by slowly dying
autocorrelations and an abrupt cutoff of the partial autocorrelations after the first lag. An MA(1)
model is characterized by slowly dying partial autocorrelations and an abrupt cutoff of the
autocorrelations after the first lag. Statisticians have developed many rules of thumb to help
make their choices, but these rules are complex and beyond the scope of this guidebook.
Special software is needed to estimate an ARIMA model, because one or more of the MA
independent variables requires an estimate. Traditional linear regression methods cannot
handle this situation.
Extensions to ARIMA (Examples from WisDOT)
In a study of the use of time series models for project level forecasting, the Wisconsin
Department of Transportation (WisDOT) tried many variations on ARIMA. It added
socioeconomics, such as total personal income, added step functions, such as capacity
changes, and added impulses, such as Labor Day and deer hunting.
Here is one of many models tried.
Dependent Variable:
Log of daily traffic I94 near Johnson Creek
2 t 2 1 t 1 t t
Y
− −
ε θ − ε θ − ε + µ ·
Time Series Methods 19
Independent Variables Included:
Log daily traffic previous day
Log daily traffic previous week
Log daily traffic previous year
Impulse dummy variables for deer hunting season, July, Labor Day weekend,
Memorial Day weekend and Thanksgiving weekend
Count of days from 1/1/1983
Impulse dummy variables for 12/15/87, 12/3/90 and 12/15/90 (outliers)
Moving average from previous day (lag 1)
Autoregressive term from previous week (lag 7)
Cyclic Patterns in AR Models
It is relatively easy to model cyclic patterns (of period n) by including an autoregressive term for
n periods ago. A typical monthly forecast looks like:
The model says that the next month can be predicted from the current month and from the
same month last year. Such a model will pick up regular monthly variations in traffic.
Similar models can be built to pick up weekly, daily, or (even) hourly variations in traffic levels.
Case Studies
Air Travel I: Brown and Watkins (1968)
One model from Brown and Watkins took the form:
∆ log(T) = 0.0725  1.307 ∆ log(F) + 1.119∆ log(Y)  0.038 log t + ε
in which air passenger miles per capita T is a function of average fare F, real disposable income
per capita Y and time in years t. In this equation ∆ (delta) indicates differences between
successive periods. This is a traditional linear regression model, as it does not contain any
autoregressive or moving average terms. It is conceptually similar to an I (integrated) model.
This model contains two causal variables, fare and income, and one trend term. Brown and
Watkins looked at the differences of logs, so that the coefficients on the causal variables can be
interpreted as elasticities.
An inspection of the model reveals:
♦ Dropping fare implies increasing travel;
♦ Increasing disposable income implies increasing travel;
♦ Future years, otherwise, will have less travel (or a declining rate of increase).
Air Travel II: Oberhausen and Koppelman (1982)
Oberhausen and Koppelman developed a rather elaborate AR model of monthly air passenger
travel for a westbound trip to Hawaii.
t 12 t 12 1 t 1 t
Y Y Y ε + ϕ + ϕ + δ ·
− −
Guidebook on Statewide Travel Forecasting 20
Y
t
= 0.36Y
t1
+ 0.34Y
t2
+ 0.62Y
t12
 0.22Y
t13
 0.21Y
t14
+ 0.38Y
t24
 0.14T
t25
 0.13Y
t26
 90.7F
t1
+ 297.92 + ε
where F
t1
is the twoway fare and all other independent variables are earlier points in the time
series.
The independent variable is passenger volume (not a difference), and they did not use any
moving average terms. All autoregressive terms have coefficients less than one (in magnitude),
as is typical for AR models.
The yearly cyclic pattern in the data is evident by inspecting the various terms. The largest
coefficient is at lag 12, one full year ago. Passenger travel also seems to be influenced by a
few other months about a year earlier and two years earlier. The multiple months at yearly
intervals seems to be accomplishing a sort of moving average.
It is difficult to second guess these authors without acquiring and analyzing their original data,
but this model contains a particularly large number of autoregressive terms. Other formulations
in the ARIMA family might have produced a more elegant model. Unfortunately, BoxJenkins
software does not automatically select the best combination of terms. Judgement must be
applied, resulting in models that are sometimes less than perfect.
Traffic Levels I: Maine VMT
A pure causal variable model was created by the State of Maine to estimate total VMT in the
state.
VMT = 15L + 332G  9600
where VMT is in millions of vehicle miles, L is thousands of licensed drivers, and G is gross
state product in billions of dollars. Note the absence of any trend terms, properly reflecting the
concept of travel being a derived demand.
Maine fit its VMT model with linear regression. The model fits the data almost perfectly (R
square = 0.995). The signs of coefficients for both independent variables are intuitively correct,
even though the two independent variables are strongly correlated.
In order to use this equation, both the number of licensed drivers and the gross state product
must be forecasted. Maine could do this themselves (perhaps with separate time series
models) or by obtaining them from outside agencies or commercial services.
1
Traffic Levels II: New Mexico Heavy Vehicle Traffic
This example from New Mexico forecasts heavy commercial vehicle traffic on a single road.
The model was created with linear regression. It includes a trend term and three causal
variables.
HC = 28000 + 15Y  0.12D  0.08G + 0.078C
1
Coefficients in the above equation have been rounded from the original model.
Time Series Methods 21
where HC is heavy commercial traffic on I40; Y is year, D is US disposable income, G is US
gasoline cost, and C is New Mexico’s cost of residential construction. The overall fit was good
(Rsquare of about 0.8).
The model exhibits problems of multicollinearity, due to the strong correlations between
independent variables. The disposable income term (D) should be positively related to heavy
vehicle traffic, but has a negative sign in the model. The negative sign for gasoline costs (G)
seems reasonable. The term for cost of residential construction (C) is difficult to interpret. It
may indicate the health of the New Mexico economy, or it could account for some aspect of
consumer prices. The large negative yintercept is due to using a 4digit year in the trend term.
1
Appendix: Some Elementary Statistical Concepts
Time series analysis is inherently multivariate. The data item to be forecast (which behaves
randomly) is related to other variables, some of them behaving randomly. Time, of course, is
deterministic (i.e., not subject to random fluctuations). Random variables are expressed as a
list of numbers, with a subscript denoting the position in the list. For example, traffic volumes on
STH 43 might be given the variable X
i
where i is the position in the list. In time series work, the
lists are always ordered: for example, X
1
is the traffic in period 1 and X
n
is the traffic in period n.
A period can consist of a whole year, a month, a week, a day, an hour, etc.
Because of the randomness in the data, statistical analysis is appropriate. The statistical
analysis allows someone to forecast without further consideration of the randomness in the
data, and it allows that person to understand the accuracy of such a forecast. Typical statistics
that describe data include the mean, the standard deviation, the coefficient of variation and
correlation coefficients. Statistics that help understand accuracy include the tstatistic and R
square.
Normal Distribution. The normal
distribution underlies much of the
theory behind time series analysis.
Any event that is influenced by a
large number of random
disturbances tends to be normally
distributed.
Mean and Other Similar Statistics.
The mean is the most probable
value of a random variable, and it is
estimated by taking a simple
average of samples. The normal
distribution is symmetrical about the
mean. When data is categorized, the category with the largest number of samples is the
“mode”. The “median” value has half the samples above it and half the samples below it. The
median is especially useful in determining central tendency when there are a few really strange
samples that distort the mean.
Standard Deviation and Associated Statistics. The standard deviation is a measure of the
dispersion (or spread) of the distribution. About 68% of the area under the normal curve occurs
1
Coefficients have been rounded from the original model.
P
r
o
b
a
b
i
l
i
t
y
o
f
1
U
n
i
t
X
σ

Guidebook on Statewide Travel Forecasting 22
within one standard deviation of either side of the mean. About 95% of the area under the
normal curve occurs within 1.96 standard deviations of either side of the mean.
The square of the standard deviation in called the variance. A standard error is similar to a
standard deviation, but relates to the dispersion of parameters (e.g., a mean or a constant in a
model) that have been computed from many samples of data. The sample standard deviation,
s, can be calculated by this formula:
TTest. The ttest was developed to determine whether a statistic computed from a sample
differs from a similar statistic computed from another sample or differs from some
predetermined value. A typical use of a ttest in traffic engineering is to determine whether the
mean speed after a change in the traffic environment (enforcement, geometry, etc.) differs
significantly from the mean speed before the change. As a rule, t statistics become larger as
more samples are included and accuracy improves. It is analogous to the signaltonoise ratio
for the statistic.
A ttest is also used to interpret the quality of an individual term in a time series model. A term
consists of a model coefficient and a variable. The tstatistic is an output of regression analysis
and similar techniques. The tstatistic for a model is found by dividing the value of a model
coefficient by its standard error.
A tstatistic larger than 1.96 usually (with a sufficient number of data points) indicates that the
coefficient is significantly different from 0 with 95% confidence. That is, 19 out of 20 times the
coefficient will have the given sign (plus or minus) when a new sample is drawn each time. A
significant tstatistic is often taken as evidence that the term is useful in explaining the data
series.
A significant tstatistic does not imply that the value of the coefficient is correct. The analyst
must look at the standard error of the coefficient to determine the accuracy of the term.
Furthermore, a significant tstatistic does not by itself justify including a term in a model. There
must also be good reasons for its inclusion from knowledge of travel behavior.
The formula above for the tstatistic shows how it is computed and interpreted when estimating
coefficients of a model. The tstatistic is computed somewhat differently when comparing
means of two samples. You should refer to a good text on statistics for more information on the
ttest.
RSquare. Rsquare is the square of the correlation between the data and the estimate. It
ranges between 0 and 1. Rsquare is often expressed as a percent and called “percent of
variance explained”. It is the most often used measure of the quality
of a model. Sometimes it is useful to adjust Rsquare for the number
of coefficients in the model. An adjusted Rsquare gives a better
indication of which of several alternative models is best.
A “residual” is the vertical (parallel to the axis describing the data
series) deviation of a point in a data series from its estimate.
( )
∑
−
−
−
·
n
1 i
2
i
x x
1 n
1
s
96 . 1
s
b
t
b
> ·
{
Residual
Time Series Methods 23
Rsquare can be calculated by comparing the standard deviation of the residuals to the
standard deviation of the time series. Comparatively small residuals result in a large value of R
square.
Coefficient of Variation. The coefficient of variation reveals the compactness of a random
variable. It compares the sample standard deviation to the size of the mean, as shown in the
equation below.
x
V
σ
·
Guidebook on Statewide Travel Forecasting 24
Chapter 3. Passenger Forecasting
Introduction
Statewide passenger travel forecasts have been performed for many years in some states, but
the models are not as fully developed as those used in urban forecasts. A few states, notably
Michigan and Kentucky, have had long experience with statewide forecasts. However, the
majority of states have not maintained statewide models. In many ways, statewide travel
forecasting is more difficult than urban travel forecasting, and the need for accurate forecasts
has been less compelling. Recent Federal legislation, including ISTEA and TEA 21, has
highlighted the need for better statewide travel forecasts, so many states are now improving
their models or developing new ones from scratch.
The purpose of this chapter is to introduce the standard passenger forecasting procedure as it
relates to statewide travel forecasting. Primarily, this chapter concerns intercity forecasting.
Corridor analysis is also briefly discussed.
The chapter first describes the differences between a statewide model and an urban model.
Differences include the scale of the analysis, the types and frequencies of trips, available modes
and data sources. An appendix at the end of this chapter discusses the basics of urban
modeling.
Each step of the forecasting process, as implemented in statewide travel forecasts, is
described. In some cases, alternative methods of performing the same step are presented.
More advanced passenger travel forecasting techniques are described in Chapter 5. This later
chapter presents highly specialized and experimental methods, as well as means of collecting
and processing statewide travel data.
Borrowed Methods
Many states have implemented passenger travel models that are derived from urban models.
They perform many of the same steps with much the same mathematics. Besides the obvious
differences in scale, other differences are subtler. Thus, this chapter will review the urban
procedure and describe how it has been adapted to statewide travel forecasting. Many of the
steps described here will reappear in Chapter 4 (freight) and in Chapter 5 (advanced
passenger).
Advantages of UrbanLike Models
Building statewide models on the foundation of urban models has these advantages:
♦ Availability of commercial travel forecasting packages;
♦ Many refined algorithms;
♦ Well understood theories of travel choices; and
♦ Many active users to share knowledge.
Disadvantages of UrbanLike Models
However, building statewide models from urban principles carries some disadvantages:
♦ Convenience of the software may deter implementation of better methods;
♦ Urban models may be overly complex for many intercity applications; and
♦ Data requirements can be burdensome.
Many existing urban algorithms are deficient for statewide forecasts, for example:
Passenger Forecasting 25
♦ Traffic assignment methods rely on the notion of capacity to achieve good results;
♦ Methods for ascertaining the fraction of travel that is intrazonal trips are not good enough
for statewide travel;
♦ Size of networks may overly burden the computational process;
♦ Sizes of some zones lead to very coarse traffic assignments; and
♦ Omission of many roads leads to excessive volumes on links near urban areas.
Urban v. Statewide Models
Network Detail. Urban models have highway networks that typically contain all freeways,
freeway ramps, major arterials, minor arterials and a few collectors. Many urban models show
freeway segments as pairs of parallel oneway links. By contrast, many statewide models
greatly reduce the detail within urban areas. Included are all state and Federal highways and a
selected number of other major arterials. Statewide networks will often omit interchange ramps
and other details related to freeways.
Traffic Analysis Zone (TAZ) Size and Number. TAZ sizes vary considerably within urban areas,
being smaller near CBDs and larger in suburban areas. The number of TAZs tends to increase
with urban area size, but not proportionally. Small cities tend to have small TAZs; large cities
tend to have much larger TAZs. Statewide networks do not vary TAZ sizes as much as urban
models. Urban areas may be aggregated into just a few TAZs, so that timelines and computer
resources can be kept to a reasonable level. Obviously, the zone system must at least cover
the whole state. Michigan’s statewide model, for example, has 2300 TAZs and Kentucky’s
model has 1500 TAZs. Both Michigan’s and Kentucky’s networks and zone systems extend
considerably beyond their own borders.
Intercity v. Local Travel. Statewide models are designed to provide estimates of intercity travel.
The details of travel within urban areas are of less interest.
Intrazonal Trips. Because it is difficult to assign intrazonal trips to networks, the number of
intrazonal trips is usually kept small in urban area forecasts. Statewide models can have a
large number of intrazonal trips in TAZs within urban areas.
Mode Split. Intercity travel involves different modes than urban area travel.
Processing Time. Large networks and large numbers of TAZs cause slow calculations. Slow
calculations constrain the number of trial runs and increase the overall expense of producing a
forecast.
Sources of Travel Data
A number of data sources used in statewide forecasting are available from federal government
sources. This includes the National Personal Transportation Survey (NPTS) from FHWA and
the Census Transportation Planning Package (CTPP) and the Journey to Work (JTW) survey
from the Bureau of the Census.
The NPTS has proved to be the most valuable of these data sets for forecasting purposes. It
was last performed in 1995. It is a 42,000 sample national survey of travel behavior conducted
using travel diaries and phone interviews.
The CTPP does not provide trip data, but instead provides travelrelated socioeconomic data for
all regions of the country and has been prepared by the Bureau of the Census and AASHTO in
a format designed for use in the transportation forecasting process. The related JTW survey
provides origindestination data for the homebased work purpose only, based on a 12.5%
sample of households.
Guidebook on Statewide Travel Forecasting 26
Surveys and Counts. More traditional sources of travel data are also applicable to statewide
forecasting. These include special surveys and cordon counts. Special surveys can be
conducted in a number of ways. Two of the most common survey methods are license plate
surveys and home interviews.
License plate surveys involve observing the license plate numbers of vehicles passing or parked
at particular locations. Addresses of ownership can be easily obtained. Alternatively, home
interviews are conducted with a broad cross section of the population and involve the collection
of travel and socioeconomic data by telephone or by mail survey.
Cordon counts or roadside surveys are surveys conducted by stopping vehicles on routes
leading into or through a particular area and interviewing their occupants. This field work is
often followed up with a mailin or telephone survey. A 1995 Vermont study reported a cost of
$17.50 per home interview sample versus $12.35 for a roadside survey.
Defining the Scope of the Model
Statewide Network Preparation
Of course, the network for a statewide model should connect all areas of the state between
which travel is to be forecasted. In addition, the networks for some recent models have been
extended as much as 200 miles into adjacent states. This might be especially useful in
providing alternate paths for trips coming into the modeled state from distant regions. It is also
possible to include a skeleton national network as a means of linking TAZs in distant regions to
the statewide model. Such a national highway network would essentially consist of interstate
highways.
Any highway or railway segments that are important to the type of analysis being done should
be represented by links in the network. For a statewide passenger network this usually includes
interstate highways and all major US and state highways. Where intercity rail or bus travel is
being modeled, separate networks could also be included for these modes.
Zone Sizes. The TAZ sizes used in a statewide model should be limited by (1) the purpose of
the model and (2) the geographical levelofaggregation of the available data. For example, it
may be desirable to use census tractlevel data for corridor studies when sufficient data are
available. For evaluation of the statewide effects of a localized system change, it may only be
necessary to use data at a county level.
Different states have adopted different strategies for defining TAZs. Michigan, California and
Texas each have more than 2000 TAZs; New Hampshire has just over 400; while Rhode Island
has nearly 900. Missouri and Michigan are moving toward nested zone structures to be able to
vary the zonal detail according to the specific needs of the analysis.
Special generators are used to represent land uses that generate travel, but are not adequately
represented by conventional generation equations (i.e., equations based on population and
employment figures). Examples include airports, state parks and universities.
Modal Categories
The number of modal categories that might be included in a statewide passenger travel
forecasting model is heavily dependent on the objectives of the model itself. For instance,
airplane trips would not typically be included in a model for a short distance intercity corridor.
Similarly, it is sometimes useful to include modes that do not yet exist in a particular area (e.g.,
high speed rail) in an attempt to estimate the potential effects of future implementations.
Passenger Forecasting 27
A typical categorization of intercity passenger modes would have automobile (driver or
passenger), intercity bus, conventional rail, high speed rail and airplane.
Purposes
Statewide and intercity travel forecasts have adopted different sets of trip purposes, depending
upon the type of questions being asked. The trip purpose categories used for urban travel
forecasts are not necessarily appropriate for statewide travel forecasts.
Trip purposes should be selected to match available data and to help address policies and
access alternatives. More trip purposes may improve the precision of the model, but they cause
increases in data collection costs and computation time.
Fundamentally, a trip purpose should contain trips that have similar characteristics:
♦ Any trip produced within a purpose must be allowed to travel to any attraction zone.
♦ All trips within a purpose share many of the same decision characteristics: mode split
utility coefficients, automobile occupancy rates, time of day and direction of travel factors
and gravity model parameters.
One particular set of trip purpose categories seems to work especially well for forecasts of
intercity travel:
♦ Work related or business;
♦ Recreation/vacation; and
♦ Other nonbusiness.
Corridor studies should also include:
♦ Travel to work.
Network Structure
Sources of Networks. Statewide highway network data is often available from several sources.
In urban areas, local MPOs typically maintain network models or GIS databases that can be
modified for use in the statewide model. Network information is also available from federal
government sources including the FHWA digital network (which is based on TIGER files), the
BTS North American Transportation Atlas Data Bases CDROM and Oak Ridge National
Laboratory.
It is important to check networks that are imported from other sources, since they may contain
coding errors that went unnoticed in their original applications or that occurred as a result of
translation across different software products.
Stitching. If the statewide network is assembled from a number of sources (e.g., the various
local MPO models), the individual networks will need to be “stitched” together into a single
network. In this case differences between adjacent local models must be resolved, including
changes in network data format, renumbering of zones, transformation of coordinates, etc. A
series of criteria – based on the network sources available and the intended uses of the model –
must be developed to guide the “stitching” process.
For statewide modeling, the network’s level of detail in urban areas is generally less important
than in MPO models. It therefore may be useful to condense the TAZs from local urban models
into largerscale TAZs. For instance, a statewide model’s TAZ could consist of a group of
census tracts in an urban area and an entire county in a remote area.
Guidebook on Statewide Travel Forecasting 28
Highway Network Attributes
Speeds: As with a network of any size, while delineating the statewide network structure it is
important to include a description of the speeds (to be used in combination with link distances)
or travel times associated with all links in the network. This information is critical to determining
the shortest travel paths during the trip assignment step.
Including attributes to describe the directionality of traffic flow (i.e., “oneway” links) and turn
prohibitions is generally less important in statewide models than for smallerscale models.
Penalties. The use of penalties or delays may also be of value in representing signalized
intersections or known capacity problems at intersections or interchanges. In a statewide model
they may also be used for other purposes. For instance penalties could be assessed at state
border locations to represent the reluctance of travelers to cross the state border for work trips.
A similar situation could also exist at river crossings and toll facility locations.
Capacities. Since the travel data available for statewide modeling purposes does not usually
lend itself to consideration of peakhour conditions, the use of link capacities (often found in
smallerscale models) is also of limited value. For statewide modeling it may be more useful to
assess delays to links where congestion is expected (typically in urban areas) or to adjust link
travel times rather than to calculate speed reductions based on capacity restraints.
Other attributes that can be used to describe the highway network are of a more descriptive
nature and can be used to examine links according to other categories that might affect their
behavior. This includes classification according to count groups, functional classes, terrain
types or geographical area groups.
Other Modal Networks
Sources. Among the most important nonautomobile networks are those for rail, bus and airline
service. Nonautomobile network data is also available from a number of sources.
For rail, the Federal Railroad Administration (FRA) has developed a digital railway network for
the United States. This can be used in combination with Amtrak or commuter rail maps and
schedules to develop a statewide rail network.
Similarly, bus company schedule information can be used to determine what subset of the
highway network is included in the bus network.
Airlines tend to be a more specialized case. Airport locations and airline schedule information
could be used to develop an airline network. Much information about airline travel is national in
scale and proprietary in nature, and airline networks may only be needed for statewide modeling
to the extent that air travel is internal to the state or between the state and its neighbors.
As is the case for highway networks, any network data imported from other sources should be
carefully checked for errors and inconsistencies.
Inclusion. Historically, statewide passenger models have had limited success in accounting for
nonautomobile modes of travel. This is primarily due to the overwhelming majority of instate
trips that are made by automobile and the related difficulties in developing a statistically
meaningful mode split model at the statewide level.
Meanwhile, mode split is an increasingly important consideration in corridor studies. Due to the
more localized nature of corridor studies and to the likelihood that they include travel patterns in
major urbanized areas, data is more readily obtained (from existing sources or localized
surveys) for calibration of mode split models. Consequently, it is easier to model mode split for
corridor study applications.
Passenger Forecasting 29
Zone Systems and Spatial Aggregation Issues
The level of aggregation of the TAZs used in the model is based on two major considerations.
A first consideration is the level of aggregation of the available socioeconomic data (both current
and forecasted). Since this data feeds the model, the model’s structure must accommodate
available data. For example, if socioeconomic data is available only at a county level of
aggregation – which is not unusual for statewide modeling – there may be little value associated
with developing a model using TAZs at a census tract scale.
Assignment. A second consideration is the proposed use of the model. If the model is to be
used for small scale planning along an intercity corridor, then the TAZs should be small enough
to reflect changes in the characteristics of that corridor in a meaningful way. For analysis of
statewide trends, the TAZs can be much more coarsely modeled.
The use of smaller TAZs can have a beneficial effect on trip assignment, especially when an all
ornothing assignment is used. In this case the smaller TAZ size serves to smooth out the
assignment of traffic to the network, yielding results that may more closely resemble the gradual
changes in traffic volumes observed on actual roadway segments. This same effect could be
achieved by developing a simple disaggregation scheme for use during traffic assignment and
by keeping the TAZ sizes as needed to match the socioeconomic data and model use
considerations noted above.
The boundaries used for TAZs typically correspond to some common level of social or political
division. Some typical sizes include (in generally decreasing size) counties, townships, census
tracts and block groups.
Connectors. The characteristics of centroid connector links can be very important, especially for
the largescale TAZs. In large TAZs, the position of the centroid within the TAZ can cause its
connector links to be inordinately long and have disproportionately large travel times associated
with them. Also, with smallTAZ models, only one connector link may be needed, whereas
multiple connectors may be required for larger TAZs.
Socioeconomic and Demographic Characteristics
The same sort of socioeconomic and demographic characteristics that are important in urban
forecasting models are also important for use in statewide models. These include:
♦ Population;
♦ Household size;
♦ Employment;
♦ Income; and
♦ Auto ownership.
The data describing these characteristics are typically used as input to the trip generation step
of an urban model, and the same is true for statewide models. Base year values for these data
are used in combination with survey data to calibrate the trip generation equations. Forecast
values for these data are then used in the generation equations to predict the number of trips
that will be generated in future years.
Sources. The principal source for socioeconomic and demographic data used in travel demand
modeling is the US census. The CTPP is particularly useful, since its tables are specifically
configured to be used in transportation forecasting applications. The census, of course, only
provides base year data. Population and employment forecasts – important for predicting future
traffic – are often available from state government agencies or local universities.
Guidebook on Statewide Travel Forecasting 30
Socioeconomic and demographic data is also available from a number of private sources,
including McGrawHill, Woods & Poole and REMI. More importantly, forecasts of the same data
are readily available from these same private sources.
Other Data Issues
Other issues are mainly associated with data collected as part of a traffic count or survey, or as
part of a statewide traffic monitoring system.
With the increasing sophistication of traffic monitoring programs, timeofday effects and
directionoftravel information can now be estimated even for rural highway links. This makes it
possible to include these effects in a statewide model, just as they might be included in an
urban model.
Cyclical Patterns. In addition, the clustering of counting stations and the continuous nature of
counting programs also makes it possible to address the cyclical nature of travel on some
highways. This includes the effects of travel on different days of the week, as well as seasonal
patterns of travel (e.g., those patterns related to hunting season or summer vacations).
Trip Table Synthesis. For cases where only limited travel data is available, techniques have
been developed to synthesize trip tables. The synthetic trip tables are developed through an
optimization procedure, based on the existing count data. This will be discussed further in
Chapter 5.
Model Steps
Trip Generation Issues
For statewide modeling, an important consideration is whether to generate all trips, including
intrazonal trips, or to generate only the trips that are made between zones. Existing models
tend to generate all trips, most likely because the available techniques for developing
generation equations do not differentiate between intrazonal and interzonal trips. Instead, as
will be discussed later, intrazonal travel times are adjusted within the trip distribution step to
ascertain the correct number of intrazonal trips for individual zones.
There are also geographical variations in trip generation rates. Most importantly, small towns
and rural areas tend to generate more trips per capita than urban areas. These effects are
discussed in the next sections.
The time period of analysis must be considered when establishing trip generation rates. Urban
models are mostly concerned with forecasting the amount of traffic on a typical weekday or
within a single peak hour of a weekday. Statewide models have often been designed to
forecast the amount of travel on an average day, including both weekdays and weekends. It
should be noted that the peak hour for a rural road can occur on a Saturday or Sunday due to
recreational travel.
The next sections discuss considerations when developing production and attraction models,
deciding on how to deal with external trips and balancing the trip productions and attractions
that result from the trip generation process.
Trip Productions
The number of trips produced in any particular TAZ is usually determined using trip production
rates based on that TAZ’s socioeconomic or demographic information.
Passenger Forecasting 31
Trip production rates can be developed from federal sources such as the NPTS, CTPP, or JTW
data, or from local surveys. Different formulations may be required depending on the amount of
data available for a particular state. One state, which had little NPTS coverage, made use of
both NPTS and JTW trip production rates for the homebased work purpose in the following
fashion:
HBW Rate = (National NPTS Rate) * (Local JTW rate) ÷ (National JTW Rate)
This formulation was used to take advantage of both (what the modelers viewed as more
statistically reliable) nationallevel NPTS data and (the more locally available) JTW data.
Of course, traditional urban model production rates – such as those provided in NCHRP Report
#365 – may also be used as a last resort.
Purposes. The choice of trip purposes included in the model has an important effect on the way
trip productions are estimated. Although many trip purposes are represented in the available
data, there are generally only three purposes that must be modeled for intercity travel:
business; nonbusiness; and recreation. It is often necessary to condense the various trip
purposes found in the available data into these three categories for use in a statewide model.
Some models have included a fourth trip purpose for commute trips.
There is a comparative wealth of information available regarding business trips. For example,
the JTW was specially designed to provide this sort of information for the homebased work trip
purpose. To develop trip generation rates for other purposes, further assumptions are usually
necessary, and both the quantity and quality of available data is more limited.
In the Michigan passenger model trip production rates are developed through the following
process. First, using NCHRP Report #365, generation rates are obtained for the various cross
classification groups. There are 60 different classification groups for the Michigan model, based
on five household sizes (1, 2, 3, 4 and 5+ persons), three household income levels (low,
medium and high) and four geographical area sizes (small, medium and large cities and rural
areas). For example, in a large urban area a lowincome, singleoccupant household will
produce 3.7 trips per day.
Second, using the proportion of total trips devoted for each purpose from the NPTS, the
generated trips are divided accordingly among the purposes. Again, for a large urban area, the
following ratios are used for lowincome, singleoccupant households:
0.192 to homebased work
0.160 to homebased recreational
0.404 to homebased other
0.310 to nonhome based work
0.214 to nonhome based other
1.000 (100% of trips accounted for)
The resulting generation rates are obtained by taking the product of the two values. For
instance:
0.192 * 3.7 = 0.71 trips per day for homebased work
0.160 * 3.7 = 0.59 trips per day for homebased recreational, etc.
The process is repeated for all 60 cross classification groups.
Guidebook on Statewide Travel Forecasting 32
Trip Attractions
Determining trip attraction rates is even more troublesome than determining production rates.
The same general methods used to develop production rates are also available to develop
attraction rates. For instance, some statewide models have attraction rate equations that have
been developed using federal data sources, including NPTS and JTW. Rates from NCHRP
Report #365 may also be of value. As with urban models, production rates are usually
considered more trustworthy than attraction rates, so attractions are generally adjusted to match
productions by purpose at the end of the trip generation step.
As is also done for production rates, attraction rates are determined by trip purpose. In this
case a TAZ containing a national park would likely attract few business trips, but would attract
many recreational trips.
Special Generators. Sites like national parks or military bases are sometimes represented in
the network as special generators. Special generator sites are often modeled as trip attractors
only, depending upon their characteristics; very sitespecific information (from specialized
surveys, for example) can be used to determine their attraction rates. If special surveys have
not been undertaken, information from NCHRP Report #365 can be used to set the trip
attraction rates. For very small sites ITE’s Trip Generation could also be used, but it is unlikely
that many small sites would need to be included in a statewide model.
Michigan found that the weakest link in the whole model chain was trip attraction, so shortcuts
at this stage of the modeling effort are inadvisable.
In the Michigan passenger model special generators are modeled as attractors only. Attraction
rates used are as follows.
These rate equations are intended to show the types of ways rates can be created. They
should not be used for forecasting purposes, as new data sources are available.
Special Generator
Type
Source of Rate Rate Equation
Airports 1991 ITE Manual exp[1.368 x ln(reg. aircraft)  0.347]
or [104.74 x operations /365]
Tourist Attractions MDOT Travel &
Tourism
2 x attendance
Campgrounds 1991 ITE Manual 0.79 x campsites
State Parks 1991 ITE Manual 0.50 x acres
Golf Courses 1991 ITE Manual 37.59 x holes
Marinas 1991 ITE Manual (1.891 x berths) + 410.795
Motels 1991 ITE Manual exp[0.713 x ln(0.44 x rooms) + 3.945]
Hospitals 1991 ITE Manual exp[0.634 x ln(beds) + 4.628 ]
Shopping Centers 1991 ITE Manual exp[A x ln(ksf) + B] where
A = 0.756, B = 5.154 for > 570 ksf,
and A = 0.625, B = 5.985 otherwise
Colleges &
Universities
1987 and 1991
ITE Manuals
2.37 x students (for universities) or
x students (for community colleges)
Passenger Forecasting 33
External Stations
There are two principal ways to include the effects of external travel in the model: using
external stations or using a national highway network.
External Stations. External stations are located at the edges of the network and are used to
generate trips produced or attracted from outside the area that the network covers. For
statewide models, they are usually placed to correspond with locations where interstate
highways and other major roads enter the network. Use of external stations requires good
information about traffic entering the state at those locations, which should be obtained most
readily from survey information.
National Network. Use of a national network provides an alternative method of modeling trips
generated outside of the local network. For a statewide model, the national network would likely
consist of the interstate system and TAZs representing states or groups of states away from the
area being modeled. With a national network it is easier to make use of national data (e.g.,
NPTS) to generate trips from distant states and later to assign them into or through the local
network. The use of a national network does not entirely eliminate the need for an external
station OD survey, which is still useful for model validation.
Both states that use external stations and states that use national networks to help model
external trips have found it beneficial to extend the local network into the surrounding states. In
some cases this extended statewide network covers areas hundreds of miles into adjoining
states. The extended network is used to provide both a more detailed accounting of trips
generated in nearby states that may pass into or through the state being modeled and to
provide a way of buffering the distribution of trips generated outside the local network and being
fed into it through external stations or national network links.
Balancing Productions and Attractions
In general, the number of trips produced will not match the number of trips attracted. This is to
be expected, since the numbers of trips produced or attracted are developed by entirely
different submodels.
Since the production equations are generally considered to be more reliable than the attraction
equations, it is common to adjust the attraction values of each TAZ such that the number of trips
attracted for the whole model is equal to the number of trips produced. This process is carried
out separately for each trip purpose.
Trip Distribution: Gravity Model (Production Constraint)
Gravity models have been extensively used for statewide travel forecasting. A complete
specification for one form of the gravity model is:
where:
T
ij
is the number of trips from i to j;
P
i
is the production in zone i;
A
j
is the attraction in zone j;
t
ij
is the impedance (time) from i to j; and
( ) ( )
∑
·
k
ik k ij j i ij
t f A t f A P T
Guidebook on Statewide Travel Forecasting 34
f(t
ij
) is the friction factor from i to j.
This equation will always yield a trip matrix that is consistent with the number of productions in
each zone, as calculated in the trip generation step. The trip matrix will not be consistent with
the number of attractions. Thus, this form of the gravity model is often referred to as being
“singly constrained”.
The number of trips between any production zone i and any attraction zone j is proportional to
the number of productions in i, proportional to the number of attractions in j and proportional to
the measure of proximity, f(t
ij
). The friction factor, f(t
ij
), is larger for pairs of zones that are close
to one another and smaller for zones that are distant from one another. Friction factors are
determined empirically.
Friction Factors. Friction factors are a function of a measure of zone separation, t
ij
. Travel time
by automobile is commonly used for urban travel forecasting. For intercity travel, time and
distance are almost directly proportional to one other, so either could be used to calculate
friction factors. When there are two or more modes in the analysis, it is often desirable to use a
composite time or cost that considers all available modes. This composite time is not a true
average; it will be discussed more fully in Chapter 5.
The denominator of the expression (the sigma term) adjusts the trip table so that productions
from the trip distribution step agree with productions from the trip generation step.
Singlyconstrained gravity models are not recommended for statewide models because trip
attractions from the trip generation step will not be preserved. A doublyconstrained model
preserves both productions and attractions. Singlyconstrained models are slightly easier to
calculate than doubly constrained models. They are sometimes used for estimating the number
of trips going to or coming from an entirely new activity site.
DoublyConstrained Gravity Model
A doublyconstrained gravity model will preserve both productions and attractions, as calculated
in the trip generating step. It has the following functional form:
To conserve productions, set the X’s by
To conserve attractions, set the Y’s by
Two “balancing factors” have been introduced, X
i
and Y
j
. There are as many of each as there
are zones. These balancing factors do not provide any new information to the model; their only
purpose is to assure that the trip generation results are not changed in the trip distribution step.
( )
ij j i j i ij
t f Y X A P T ·
( )
ij j
j
j i
t f Y A 1 X
∑
·
( )
ij i
i
i j
t f X P 1 Y
∑
·
Passenger Forecasting 35
Xs and Ys must be calculated prior to calculating the number of trips between zones. An
iteration process is required because each X
i
requires knowledge of all Ys and each Y
j
requires
knowledge of all Xs. The iteration process is started by assuming that all Y’s are 1. Then Xs
are found, then Ys are found, then Xs are found again, etc. The process converges rapidly,
typically in three to five iterations.
Friction Factors
Friction factors are found as a declining function of time, distance, cost or some combination of
the three. Friction factors are commonly found from a lookup table, an exponential function, a
power function, or a combination of exponential and power functions.
For example, Michigan defined its measure of separation as:
t = 0.75 Distance + 0.5 Time + 0.1 Toll
The sum of two “gamma functions” was used for friction factors. For the homebase work
purpose, the friction factor function looks like:
They applied this function to all instate trips. Outstate trips had still another friction factor
function. The complexity of the friction factors relates to the difficulty of obtaining a single
function that properly reflects both short distance and long distance trip making.
Kentucky used tables of friction factors. They have three sets: work trips, short (≤60 minutes)
nonwork trips and long (>60 minutes) nonwork trips. Long nonwork trips constitute about 1% of
the total number of nonwork trips. Below are pieces of each table.
Trip Distribution: Application to Statewide Forecasting
Both Michigan and Kentucky had difficulty finding simple friction factor functions that match trip
making for both long and short trips. Kentucky handled the problem by designating a separate
purpose for long nonwork trips (>60 minutes). Michigan handled the problem by having a very
complex friction factor function for intrastate trips and an entirely different friction factor function
for interstate trips. Michigan also made extensive adjustments to the gravity model, tripbytrip,
t 012 . 0 441 . 0 t 123 . 0 565 . 1
e t 200 e t 1258477 ) t ( f
− − − −
+ ·
Time
(min)
Friction
Factor,
Work
1 259360
2 203007
3 159242
118 2
119 2
120 2
Time
(min)
FFactor,
Long
Nonwork
1 999999
2 367878
3 196955
58 2
59 2
60 2
Time
(min)
FFactor,
Long
Nonwork
61 200
181 50
241 6
601 1
661 1
720 1
Guidebook on Statewide Travel Forecasting 36
with K factors. The experiences in Kentucky and Michigan suggest that the gravity model is not
working well within the trip purpose categorization selected, which in both cases was similar to
urban trip purposes.
Intrazonal Trips. The fraction of intrazonal trips is very important to statewide travel forecasting.
A large error in these cells of the trip table can result in a large error in interzonal (i.e., intercity)
trips. Within the gravity model, the number of intrazonal trips is controlled entirely by the size of
the intrazonal trip time. This intrazonal trip time must be selected with great care. It should not
be calculated by a “nearest neighbor” method that is available in some travel forecasting
software packages. Still other software packages define friction factors for only whole minutes
of trip time, so rounding can cause an appreciable error in the estimates of intrazonal trips.
Perhaps a better method of handling intrazonal trips is to exclude many of them from the gravity
model. In fact, one of Michigan’s early statewide models omitted intrazonal trips at the trip
generation step. The number of intrazonal trips could be estimated for certain key zones, then
subtracted from trip generation totals. The gravity models will eliminate intrazonal trips if the
friction factor is set to zero (usually by setting intrazonal time to a very large number).
Mode Split
Mode Split Tables. The best single source of mode split information is the NPTS. However, the
NPTS does not provide the data in a convenient form. Rather, each trip is separately reported.
Thus, a considerable effort is required to develop the mode split tables. For intercity travel the
tables should break out mode by both purpose and trip length. Other break outs are may be
useful, depending upon the policies and alternatives to be tested. The NPTS modal categories
are automobile, van, sport utility vehicle, pickup truck, other truck, RV (recreational vehicle),
motorcycle, other POV (privately owned vehicle), bus, Amtrak, commuter train, streetcar/trolley,
subway/elevated rail, airplane, taxicab, bicycle, walk, school bus and other nonPOV.
A separate mode split table should be prepared for each trip purpose. The NPTS trip purpose
categories are to work, workrelated business, return to work, shopping, school, religious
activity, medical/dental, other family or personal business, take someone somewhere, pick up
someone, vacation, visit friends or relatives, went out to eat, other social/recreational, change
mode of transportation, other and home.
Other mode split techniques are variations on logit analysis. The next sections will discuss
multinomial logit and nested logit. Pivot point is illustrated in the chapter on freight forecasting.
Logit
The logit model allocates person trips to alternative modes. It does so by comparing the utilities
of all alternative modes.
The proportion of trips made by mode k can be found from:
where U
k
is the utility of mode k, z is a dummy index that ranges over all modes and e is
2.718281…
∑
·
z
U
U
k
z
k
e
e
p
Passenger Forecasting 37
Utility is a measure of the personal satisfaction of taking a trip, exclusive of the satisfaction of
reaching the destination. Because travel consumes valuable personal resources, utility is most
often a negative number. Utility is smaller (more negative) when trips are longer.
Utility is often expressed as a function of travel time, travel cost and convenience measures.
The specific functional form is often found from statistical analysis of choices made by travelers.
Logit models possess an interesting mathematical property called “Independence of Irrelevant
Alternatives” (IIA). The IIA property says that the ratio of mode shares between any two modes
is unaffected by the presence or characteristics of a third mode. The IIA property is displeasing
to many analysts because it can cause a bias toward technologies that have many distinct
modes. For example, consider a mode split between automobile and conventional rail.
Automobile has 80% of intercity traffic and conventional rail has 20% for a ratio of 4 to 1. Now
assume the addition of a third mode, high speed rail, which grabs 20% of the ridership for itself.
The new mode shares from a logit model would be 64% automobile, 16% conventional rail and
20% high speed rail (the ratio of automobile to conventional rail is still 4 to 1). It might be
argued that conventional rail should lose a higher percentage of its ridership to high speed rail
than does the automobile. After all, conventional rail and high speed rail have similar
characteristics and compete directly with one another. However, the logit model does not
recognize differences in competition across pairs of modes.
Utility is usually found from a linear equation that combines the effects of trip time, trip cost and
trip convenience.
where the a’s are empirical constants.
The utility function can be found from statistical estimation or from transferable parameters.
Because longer and more costly trips are considered bad, the coefficients on time (t) and cost
(c) are almost always negative.
The above equation defines a mode specific constant, a
k
. This constant is arbitrarily set to zero
for one mode. The value of a
k
for the remaining modes may be positive or negative, depending
upon how favorably travelers perceive that mode. A better mode would have a positive mode
specific constant. A worse mode would have a negative mode specific constant.
Logit Example, Forinash and Koppelman (1993)
Forinash and Koppelman (1993) calibrated several mode split models for intercity travel in
Canada. The data was taken from trips made between Ontario and Quebec. Some of the
models were pure logit; others were nested logit (see later discussion). The modes considered
were automobile, rail, bus and airline. One of the pure logit models is illustrated by the
coefficients in the table below.
Significant independent variables in the model were: mode specific constants; large city
dummies by mode; frequency; travel cost; invehicle time for high income; invehicle time for low
income; outvehicle time for high income; and outvehicle time form low income.
L + + + · c a t a a U
2 1 k k
Guidebook on Statewide Travel Forecasting 38
The above table lists the coefficients of the utility functions. Be aware that not all coefficients
appear in every mode’s utility function.
The breakpoint between high and low income was $30,000 Canadian. The large city variable is
1 if either end of the trip is in a large city.
Nested Logit
Nested logit models organize modes into
a hierarchy, like the one shown in the
diagram on the right. The hierarchy
implies that several decisions are made
in the process of selecting a mode. For
example, a decision is first made
between SOV and carpool and between
conventional rail and high speed rail.
Finally, another decision is made
between automobile, rail and airplane. Modes with similar characteristics are grouped together
in a nest. A nested logit model helps the analyst to avoid the IIA property of a pure logit model.
The nested logit model has the ability to differentiate between pairs of modes that are
complementary and pairs of modes that are competitive.
Nested logit models are calibrated in much the same manner as regular logit models, statistical
estimation or adoption of default parameters. The statistical estimation process can also give a
good indication of the best arrangements of modes and nests. However, professional
judgement is required to assure that the model will give dependable results. Some nesting
schemes that look good statistically can give strange forecasts.
A logit model is applied to each level in the hierarchy, but the model calculates utility somewhat
differently than discussed previously. Specifically, the utility of a nest, U
n
, is:
where U
k
is the utility of a mode one level below and φ
n
and µ
n
are constants for the current
level. The utility of modes at the lowest level are calculated in the same way as a logit model,
e.g., a linear combination of time, cost and convenience terms. The logsum term, specifically:
Variable Air Bus Auto Train All
Mode Constant 1.888 2.756 2.203 0
Large City 0.7460 0.1224 1.1306
Frequency 0.1022
Cost 0.03265
InVeh Time, High Income 0.01382
InVeh Time, Low Income 0.003797
OutVeh Time, High Income 0.04053
OutVeh Time, Low Income 0.02635
SOV
Decision
Carpool
High
Speed
Airline
Conventional
Auto
Decision
Rail
Decision
n
k
U
n n
n k
e ln U µ + φ ·
∑
φ
Passenger Forecasting 39
is simply a means of combining together the utilities of all the modes in the nest. This combined
utility takes into account the increased opportunities to travel by having several modes in the
nest. For example, to compute the utility of the automobile (refer to the previous hierarchy), the
logsum would form the combined utility of the SOV and carpool modes. The coefficient on the
logsum term φ
k
is found by matching the model to data. The coefficient, φ
k
, is usually between 0
and 1. A coefficient of 1.0 causes the nested logit model to behave as a standard logit model.
The nest bias coefficient, µ
k
, performs the same purpose as the mode specific constant in the
logit model. It raises or lowers the utility of the nest, depending upon the preference of users.
The nest bias coefficient is arbitrarily set to zero for one nest at a level; the other nests’ bias
coefficients are set relative to the arbitrarily chosen nest.
It should be noted that some applications of nested logit omit the φ
k
term where it divides utility
in the exponent. Also, some nested logit models omit the nest bias term.
Nested Logit Example, More F&K
Forinash and Koppelman also tried several nested logit
models on the travel data from Canada. One of their
nesting schemes is illustrated on the right. The exact
formulation of this nested logit model differs slightly
from the general formulation described in the previous
paragraphs. They did not include a nest bias term for
the Surface Mass utility term when splitting trips across
Airline, Auto and Surface Mass. The two formulations
are nearly equivalent, but could produce slightly
different coefficients for utility equations of the modes
within the nest.
The coefficient, φ
k
, on and within the logsum term had a value of 0.6488. The outvehicle time
term is computed as raw outvehicle time divided by the natural logarithm of the distance.
∑
φ
φ
k
U
n
n k
e ln
Auto
Decision
Rail
Airline
Bus
Surface
Mass
Variable Air Bus Auto Train All
Mode Constant 0.7359 1.3140 0.5756 0
Large City 0.7697 0.1339 1.242 0
Household Income 0.03507 0.03009 0.01303 0
Frequency 0.09843
Cost 0.02339
OutVeh Time, High Income 0.1977
OutVeh Time, Low Income 0.1830
Total Time, High Income 0.01217
Total Time, Low Income 0.00883
Guidebook on Statewide Travel Forecasting 40
Mode Split Issues
Lookup Tables. Lookup tables are particularly useful for splitting trips to minor modes or across
finer divisions of modes. For example, a state wants to split trips across SOV and carpool
categories. Rather than use a logit model, it might be more efficient to use a lookup table that
shows typical splits as a function of distance.
Mode Specific Constants. A traditional logit model is calibrated from data about actual trip
choices. When a new mode is planned, choice data for that mode are unavailable. All
constants in the logit model that are specific to that single new mode cannot be estimated. It is
sometimes possible to use psychometric techniques to estimate those constants. One such
method asks travelers about their intention to take a new mode. This method, referred to as
“stated preference”, is described in later chapters.
Collinearity. Many of the variables in the utility equation are roughly proportional to each other.
For example, travel time is almost proportional to travel cost for most intercity modes. This
proportionality is stronger for intercity trips than for urban trips. When two statistical variables
are proportional in this manner, they are said to be collinear. Collinearity can result in strange
values for estimated coefficients. If collinearity is present in the data, it is important to apply the
calibrated logit model to situations where the same type of proportionality holds.
Nesting Structure. Quite different nesting structures can have almost the same statistical
performance on data, but forecast very differently. Consequently, it is important to develop
nesting structures that are logical, particularly by keeping like modes within the same nest.
Professional judgement may be necessary.
Data Needs. Logit models require extensive data for calibration. Transferable parameters have
not been developed for intercity mode split models, and it is inappropriate to take coefficients
from urban models.
Trip Assignment
The least complicated technique for assigning trips to the links on the network is the allor
nothing method. For allornothing assignment, all of the trips between a particular origin
destination pair are assigned to the single shortest path of links between the pair. In reality
traffic does not always travel along the shortest path.
Four possible reasons that travelers do not follow the single shortest path are (1) capacity
restraints in the system, (2) random personal preference, (3) the lack of spatial detail associated
with large TAZs and (4) the omission of many smallervolume roads in a highway network.
Capacityrestrained and stochastic assignment methods have been developed to address two
of these reasons.
Capacityrestrained assignments, which include iterative and incremental techniques, involve
increasing the travel time on links where high volumes of traffic are expected. These methods
are commonly used in urban modeling situations in conjunction with the BPR equation:
t = t
o
[1 + 0.15 (q/q
max
)
4
] .
Urban models typically deal with peakhour demand situations. However, for statewide
modeling the necessary timeofday information is often not available to estimate peaking of
traffic; the data is instead available only in terms of 24hour flows. This makes application of
capacity restraint techniques difficult at a statewide level.
Passenger Forecasting 41
Stochastic assignment techniques involve assigning some of the trips between a given origin
and destination to the second or thirdshortest path, as well as to the shortest path overall.
This is done to account for the preferences for certain routes that drivers might have,
independent of travel time considerations and the imprecision associated with large zones or
arbitrarily small networks. Assignment to the various paths is done by estimating the probability
that a traveler will chose that particular path.
Unfortunately, states that have tried various traffic assignment techniques have been
disappointed with the results. It seems that allornothing assignment works as well as the
others. Allornothing traffic assignment can produce quite lumpy results when there are
comparatively few TAZs and the network omits a high percentage of minor roads.
Therefore, it is important to retain some degree of spatial detail when performing allornothing
assignments. TAZs should be kept small, and a large number of minor roads should be
included to provide a reasonable approximation of the actual path choices. However, small
zones and many links increase both data preparation and computational needs.
Computer resources are likely to be less of a problem in coming years. Currently, Michigan’s
model (at 2300 zones) is approaching the practical size limit computationally.
A large number of zones implies an increased burden of data preparation and input data
forecasting. A suggested method of reducing this burden is to adopt two levels of spatial
aggregation, districts and zones. Districts could correspond to the spatial units most often used
in socioeconomic forecasts, typically counties. Most data preparation and data forecasting
tasks could then be performed at the district level. Data from the districts could be
disaggregated to zones using a constant set of factors for each district. Special generators
could be associated with particular zones, as needed.
Integrating Freight. Freight represents a significant proportion of vehicles on roads and an even
larger percentage of passenger car equivalents (from the Highway Capacity Manual). Thus, a
full evaluation of highway performance must consider freight. Freight is discussed in the next
chapter.
Calibration and Validation
Calibration versus Validation. Calibration is the process of setting the various model
parameters of the various model steps to match existing trip making behavior. Validation is the
process of determining whether the calibrated model, as a whole, accurately forecasts current
conditions, usually assigned traffic levels. Thus, calibration is normally a lengthy process, with
extensive use of statistical estimation techniques and many trials and errors. Validation should
occur rather quickly once the validation data set is assembled.
It is important to maintain the distinction between calibration and validation. All too often the
validation data is used in the calibration process, thereby undercutting their value for validation.
Statistical methods (linear regression, maximum likelihood estimation) are preferred calibration
techniques for all steps except traffic assignment. There are no widely accepted techniques for
calibrating the traffic assignment step in a statewide model.
Sources of Validation Data. The set of validation data consists of traffic counts. Statewide
models should be validated only on links where network detail is reasonably representative of
reality, usually in rural areas.
When two or more duplicative data sets are available, it is possible to calibrate a model step
with one of them and validate with the other. For example, calibrations made with home
Guidebook on Statewide Travel Forecasting 42
interview data could be validated with CTPP or NPTS data. Cross validation data may be
reported at a different level of spatial aggregation than the original calibration data.
How good? Models need not be any better than the quality of the traffic data in the validation
data set. Since the quality of traffic counts is poor on low volume roads, all roads with less than
2000 vehicles per day (AADT) should be excluded from the validation. If a large number of
such lowvolume roads exists in a portion of a state and a validation is important, they should be
combined into a cutline.
Appendix: Overview of Four Step Models
The four major steps of the urban transportation modeling process are shown below. The steps
are executed sequentially. A feedback loop is now considered best practice; it assures that
congestion and delay are treated consistently throughout the model.
When strictly highway policies are being analyzed, the mode split step is often omitted. Thus, a
three step model consists of trip generation, trip distribution and traffic assignment.
In actuality, there are many more steps to the
modeling process. Some of these steps are quite
important. A typical modeling sequence might
involve:
♦ Find highway trip times;
♦ Find disutilities for nonhighway modes;
♦ Activity allocation;
♦ Trip generation;
♦ Trip distribution;
♦ Vehicle occupancy;
♦ Time of day and direction of travel;
♦ Highway traffic assignment;
♦ Trip assignment to nonhighway modes;
♦ Highway volume averaging;
♦ Highway delay calculation; and
♦ Feedback.
Network Basics
Networks. A network holds nearly all
data used in a travel forecast. A
network consists entirely of nodes and
links and the values of their attributes.
Networks are drawn with a specialized
program, called a network editor, or
they can be taken from a GIS.
Already prepared networks can be
obtained for many statewide and
corridor applications. Networks must
satisfy two important continuity
requirements:
♦ All links must connect to two distinct nodes (referred to as “A” and “B”);
♦ For a node to be useful, it must connect to one or more links.
Trip Generation
Trip Distribution
Mode Split
Traffic Assignment
F
e
e
d
b
a
c
k
CENTROID
ZONE
NODE
LINK
Passenger Forecasting 43
Links. A link is a straight line segment representing a portion of a travel path. Often links are
identified as being oneway or twoway. Traffic on oneway links flows in the A to B direction.
Links have attributes, including distance, travel time, speed and capacity.
Nodes. Nodes are primarily used to represent places where trips begin, end or change
direction. Most nodes represent intersections or bends in a road. In a traditional urban model,
trips begin or end at a special type of node called a centroid. Centroids usually represent areas
of the city identified by travel analysis zones (TAZs). Intersection nodes do not have many
attributes, but centroids contain attributes that describe the socioeconomic and demographic
characteristics of the zone. Many travel forecasting models require that centroids be connected
to the arterial portions of the network with a special type of link, called a centroid connector.
One type of centroid is an external station. External stations represent large areas beyond the
region of interest.
Path. A path is a sequence of nodes and links, starting at an origin centroid and ending at a
destination centroid. There are many paths between most origins and destinations. Usually,
the models identify the shortest path by some criterion, typically time or distance or some
combination of the two.
Trip Generation Basics
Overview. The trip generation step determines the total
number of person trips that begin or end in a zone,
usually over a 24hour period. Trip totals for any zone
are separately tabulated by trip purposes and by whether
they are productions or attractions. A typical set of trip
purposes for urban travel forecasting is homebased
work (HBW), homebased other (HBO), nonhomebased
(NHB) and specialty purposes. Trips identified as being
homebased have the home as either the origin or
destination. Whether a trip end is a production or an attraction depends on how the trip purpose
is satisfied. Trips are produced where the trip is conceived. Trips are attracted to where the trip
purpose is satisfied. The following rules apply within most urban models:
All homebased trips are produced at the home;
All homebased trips are attracted to the end that is not the home; and
The networkwide total of productions must equal the total of attractions for each purpose;
Nonhomebased trips are treated in a variety of ways. Many urban models generate a total
amount of nonhomebased trips, then split them evenly between productions and attractions.
Trip productions and attractions are computed with separate empirical relationships. These
relationships include the socioeconomic and demographic characteristics of zones, but rarely
include any travel characteristics of zones.
Productions. Productions are most often calculated by multiplying a trip rate for a category of
households by the number of households falling into that category. Categories are typically
organized by number of persons in the household, number of automobiles in the household or
household income.
Attractions. Attractions are most often calculated by a linear equation, where the dependent
variables are demographic characteristics, such as retail employment, nonretail employment
and households.
ZONE
TRIPS
PRODUCED
TRIPS
ATTRACTED
Guidebook on Statewide Travel Forecasting 44
Trip Distribution Basics
Overview. Trip distribution finds the number of
person trips that go between all pairs of zones.
Usually, 24hour person trips are distributed
separately for each trip purpose. The diagram
on the right shows three zones, yielding 6
interzonal trip interchanges and 3 intrazonal
trip interchanges. The direction of each arrow
is from the production end to the attraction end
of the trip. Each trip is identified by subscripts,
with the first subscript representing the
production zone and the second subscript
representing the attraction zone. The results
of the trip distribution step can also be shown in the form of a matrix, with as many rows and
columns as zones. Each row represents a production zone and each column represents an
attraction zone. In the trip table shown at the right, there
are 776 trips between production zone 2 and attraction zone
3.
Trip Distribution Technique. Trip distribution is most often
performed in urban area models by a technique called the
gravity model. The principles of the gravity model are quite
straightforward:
♦ There is a large number of trips between pairs of zones when the trip productions in the
pair are large;
♦ There is a large number of trips between pairs of zones when the trip attractions in the
pair are large; and
♦ There is a large number of trips between pairs of zones when there is little separation
between zones (as measured in time or space).
Mode Split Basics
The mode split step finds the number of trips
using each available mode between a
production/attraction pair. As with trip
distribution, mode split is typically performed
on 24hour persontrips. The figure on the
right illustrates three intercity modes:
automobile, train and airline.
The mode split step is most often found in
forecasting models for large urban areas. Small cities tend to have low transit ridership, which
is either ignored or treated as a fixed percentage of all trips.
Mode split models assume that travelers choose the best mode for themselves by weighing the
characteristics of the trips for all available modes. The measure of trip goodness is called
“utility”. Since trips consume valuable resources, the value of utility is most often a negative
number. The actual utility is unknown, as there are many personal factors and perceptions that
influence each traveler’s decision. However, major objective factors in the choice, such as
travel time and cost, can be ascertained with some degree of accuracy. The mode split model
ZONE 1
ZONE 2
ZONE 3
T
12
T
22
A’s
1 2 3
1 183 784 939
P’s 2 890 141 776
3 920 748 197
AUTOMOBILE
TRAIN
AIRLINE
Passenger Forecasting 45
takes these objective factors as inputs, but recognizes that the actual utility is known imprecisely
when calculating mode shares.
Some mode split models include socioeconomic factors in their expression of utility. Still other
mode split models take into consideration factors related to access and the lack of one or more
alternative modes for subgroups of travelers.
Most urban models containing a mode split step use a variation of the logit model for mode split.
Two of these variations are discussed in the main portion of this chapter.
Traffic Assignment Basics
The major outputs of the traffic assignment step are volumes on all links, turning movements at
selected intersections and the delays from congestion.
Trip Table Preparation for Assignment. The output of the mode split step is still in terms of 24
hour person trips in the production to attraction direction. For automobiles, the trips must be
converted to vehicles. For all modes, the trips must be factored into a specific time period (e.g.,
peak hour) and given a physical direction of travel (i.e., origin to destination). For example, a
morning homebased work trip will have its origin at home and its destination at work.
Conversely, an evening homebased
work trip will have its origin at work
and its destination at home.
AllorNothing. Allornothing
assignment loads all trips between an
origin and a destination to the
shortest path between the origin and
destination. This assignment is done
separately for each OD pair, then the
results for all OD pairs are summed.
Allornothing assignments do not
recognize the capacity of individual
links or nodes, but can account for
known delays on links, turn
restrictions and turn penalties at
intersections.
CapacityRestrained Equilibrium. Capacityrestrained equilibrium assignment recognizes
delays that may occur due to congestion. Algorithms that perform this form of assignment try to
spread trips from a single pair of origins and destinations over many paths such that all used
paths are shortest (or tied for shortest) and the delay along any link or at any node is consistent
with the amount of traffic on the link or passing through the node.
Stochastic. A stochastic assignment finds the shortest path between an origin and a destination
and also finds several alternative paths that are somewhat longer. The algorithm then assigns
the largest fraction of trips for that origindestination pair to the shortest path and lesser fractions
to the longer paths.
SHORTEST
PATH
Guidebook on Statewide Travel Forecasting 46
Chapter 4. Freight Forecasting
Introduction
The purpose of this chapter is to present a unified methodology for forecasting commodity flows
and vehicular volumes on intercity and rural portions of our transportation systems.
Freight forecasting is dependent upon knowledge of:
♦ Industrial structure of the economy;
♦ Existing and projected commodity flow;
♦ Technological characteristics of modes;
♦ Potential actions of carriers and shippers; and
♦ Modal networks.
Factors affecting freight demand are complex, involving:
♦ Presence of factors beyond the planner’s control;
♦ Variety in type and value of commodities;
♦ Multiple measuring scales for goods; and
♦ Variation in type of service (e.g., specialized services).
Forecasting requires extensive use of existing data sources (commercial or public), other local
knowledge and experience gained elsewhere.
Methods presented here have been adapted from forecasts done for Wisconsin, Michigan,
Indiana, Kentucky and Kansas. The methods follow a process that is similar to fourstep
models for passenger forecasting, but details vary considerably due to the unique aspects of
commodity movement.
Freight Model – Basic Steps
Building a freight model involves several steps that will be discussed in detail later. These steps
are as follows:
1. Obtain Freight Modal Networks. Networks are required for freight models. Networks
consisting of mathematical descriptions of routes, links and intersections are required for
each mode receiving complete analysis. Networks are drawn from scratch or modified from
existing sources.
2. Develop Commodity Groups. Vehicle traffic levels are derived from the movement of
commodities. Thus, a good understanding of commodities is necessary. Because there are
a very large number of commodity categories, commodities are grouped to facilitate the
analysis. A large number of groups gives precision, but increases the complexity of the
modeling process.
3. Relate Commodity Groups to Industrial Sectors or Economic Indicators. Separate economic
indicators should be adopted for production and consumption of each commodity. It is
easier to relate commodities to industrial indicators at the production end of the shipment.
Inputoutput (IO) tables can bridge the relationship to consumption of commodities.
4. Find Base Year Commodity Flows. Origin and destination data for commodities must be
obtained to build a factual model. These data are collected irregularly. For example, the
Commodity Flow Survey was last performed in 1997. Local data sources would most often
have other dates of completion.
Freight Forecasting 47
5. Forecast Growth in Industrial Sectors. Growth forecasts are best done by economists who
specialize in industrial forecasting. Forecasts can be obtained from a variety of
governmental, private and educational organizations. Growth forecasts are best if they are
disaggregated to the TAZ (traffic analysis zone) level.
6. Factor Commodity Flows. These industrial forecasts can be used to forecast commodities.
These forecasts can be applied to production, to consumption or to a whole commodity flow
table.
7. Develop Modal Cost for Commodities. Mode split is determined, to a large extent, by cost
considerations. The cost of any given shipment is complex, relating to the full logistics
stages of shipping. Detailed cost models (e.g., NCHRP Report #260) have been developed
for mode split purposes. A cost model is the primary way to determine the impact of public
policies on modal choice.
8. Split Commodities to Modes. There are three categories of methods for splitting
commodities to modes.
♦ Mode Split Models: Mode split models estimate mode shares by comparing the
relative costs of shipment between alternatives. The logit model, an often used
mode split model for passenger forecasting, has occasionally been used for
commodities.
♦ Tables: Historical mode split tables for individual commodities are readily available.
These tables are especially useful when cost is not a determining factor or when
testing policies without significant cost implications.
♦ Expert Opinion: Tables or models may not be available for entirely new modes or
services. In these instances, persons very familiar with goods movement can
provide information that may lead to reasonable estimates of mode shares.
9. Find Daily Vehicles from Load Weights and Days of Operation. Commodity flow data are
given in terms of tons per year. Planning studies often require model outputs in the form of
vehicles per day. Thus, it is necessary to determine the amount of goods carried in a
vehicle and the number of vehicle days in a year.
10. Assign Vehicles to Modal Networks. Flow matrices, by themselves, are of limited value for
planning. More important are the number of vehicles that use each link, intersection and
terminal. This information is provided by a traffic assignment algorithm.
Key Freight Data Sources
Commodity Flow Survey
The Commodity Flow Survey (CFS) is a sample of shipment information from 100,000
businesses. The CFS only provides summary tables of the raw data, but these tables are
extensive. The CFS can give good estimates of externaltointernal and internaltoexternal
flows for states. For the most part, it cannot provide commodity flow tables for points within a
single state, and it cannot provide through (overhead) flows. No vehicular data are provided.
The CFS excludes shipments from farms, most retail and service establishments, oil and gas
extraction, transportation and government. It does not provide data for flows passing through
the US that originated outside its borders.
Perhaps the most useful part of the CFS for statewide forecasting is its commodity mode split
tables. These tables give the percentage of weight or value for each commodity on each mode.
Guidebook on Statewide Travel Forecasting 48
Average trip lengths are also
given. The Wisconsin mode split
by weight (tons) and average trip
lengths (miles) for food and
kindred products (STCC 20) from
the 1993 CFS are shown on the
right. Modes with small or no
amounts of these commodities
are omitted from this table.
The CFS also provides for each state tables for mode by trip length and commodity by trip
length. These tables may be helpful in preparing mode split tables that vary by distance
shipped. At the national level tables of mode by commodity and trip length are available.
Commodities in the 1997 CFS are organized by SCTGs, not STCCs.
Vehicle Inventory and Use Survey (VIUS)
VIUS (formerly TIUS, Truck Inventory and Use Survey) is a large sample of trucks, their
physical characteristics, their ownership characteristics and their operations. VIUS comes with
software for comparatively easy access to data. Each data record in VIUS is for a single truck,
containing about 400 fields. Characteristics of a data record within VIUS include:
♦ Extensive description of vehicle and ownership characteristics;
♦ Empty weight, average gross weight and maximum gross weight;
♦ Weeks operated;
♦ Annual mileage and percentage of mileage by trip length (<50, 50100, 100200, 200
500, >500);
♦ Percentage of mileage by product; and
♦ Commodities reported in 33 categories (roughly twodigit STCC), including passengers
and empty.
Commercial Freight Data Sources
Commodity flow data are available from Reebie, DRI/McGraw Hill and GIS/Trans, who are co
developing the Intermodal Freight Visual Database. This database is derived from
TRANSEARCH. The database shows commodity volume by mode and lane (OD pair). Origins
and destinations are as small as counties. Assignments to modal networks are possible in
terms of vehicles, weight, VMT tonmiles and dollars. Modes include truck, rail, water and air.
Commodities are identified by fourdigit STCC. Imports and exports are also identified.
This description of this product is not intended as an endorsement, but as examples of types of
data available.
Other Interesting Data Sources
County Business Patterns. County Business Patterns reports for each county and for each four
digit SIC the number of employees, annual payroll, number of establishments and size of
establishment. The data is available online or by CDROM.
1992 Census of Agriculture. This survey helps fill in information missing from the CFS and
contains information on farmbased shipments.
Other Commercial Forecasts. Several companies provide demographic and industrial forecasts
at the county level. These forecasts are often broken down into many industrial categories,
Mode Tons Trip Length (Miles)
Parcel 19 NA
Private Truck 13509 56
For Hire Truck 11843 385
Rail 3058 901
Truck and Rail 115 1718
Other or Unknown 362 NA
Freight Forecasting 49
more than enough for the twodigit precision of STCCs (or SCTGs) and SICs that are typical for
planning models. Forecasts are sometimes available from other state agencies, banks, utility
companies and universities. Forecasts more aggregated than the county are of less value.
North American Transportation Atlas Databases (NORTAD). Some of the more aggressive
state models have national scope, but are focused on the state. Thus, national networks are
required. The NORTAD from the Bureau of Transportation Statistics contains many modal
networks, including highway, rail and water.
Other Specialized Databases. There are many specialized data sources documented in
NCHRP Report #388. These include information on:
♦ Exports/imports;
♦ Specific modes;
♦ Railroad, truck, water and air;
♦ Commodities; and
♦ Fruits and vegetables, coal, grain, petroleum and natural gas
Specialized databases have unique ways of listing or aggregating data. Consequently, a
considerable amount of analysis may be required to put the data into a usable form. For
example, rail waybill information may be assigned to a network to give traffic volumes on rail
lines (ALK Associates performs this type of analysis). Some of these data sources may help to
plug holes in the CFS.
Local Surveys of Note
Classification and Other Counts. Freight forecasting requires counts of vehicular flows on
enough facilities to validate the forecast. Publicly operated terminals or links (e.g., airports,
water ports, toll roads, toll bridges, locks and US Customs offices) should be able to readily
provide levels of traffic for their facility. Privately operated terminals may or may not be willing
to supply traffic information. The participation of carriers and shippers in counts of vehicular
movement is also uncertain. Classification counts on highways are needed to separate truck
traffic from passenger car traffic. The classification must be specific to a location and upto
date. Classification factors developed elsewhere should not be applied to ADT data.
Single Station OD Surveys. More extensive data on shipments may be obtained by surveying
trucks at weigh stations or other convenient points. Several wellchosen locations can give a
good picture of trucking in a state. Data on gross weight, payload, origin, destination,
commodity, vehicle type, time and other aspects of use can be obtained in this manner.
Michigan’s truck forecast was largely based on data of this type. It should be noted that
sampling at weigh stations is biased toward long haul shipments and appropriate corrections
are required.
Defining the Scope of the Freight Model
Establish Goals for Model
Before embarking upon a modeling effort, it is important to clearly identify the desired outputs.
♦ What policy decisions are going to be made, and should the model be sensitive to these
decisions?
♦ What level of accuracy is needed, and what is the desired level of precision for TAZs,
commodities and modes?
Guidebook on Statewide Travel Forecasting 50
♦ Can development of the model be accomplished in useful phases, when lessons can be
learned and applied in future developments?
♦ How can expert opinion be best integrated into the modeling effort?
Spatial Units
One of the most important decisions when building a model is the fundamental level of spatial
aggregation. Should zones be constructed from townships, counties, states, BEA regions,
NTARs, or some combination? Most statewide freight models have adopted TAZs based on
counties within the state, plus an additional zone for each of the remaining (contiguous 48)
states and external stations at US border crossings.
Disaggregating to smaller than county level presents methodological problems, as economic
data and forecasts are often not available for smaller areas. Thus, a series of assumptions are
required to further disaggregate the model.
Wisconsin included some counties in neighboring states. This technique is especially helpful
when there are large industrial centers just outside the state. Wisconsin traffic is heavily
influenced by the Chicago metropolitan area and the Twin Cities.
The adoption of countysized TAZs means that vehicular volumes will only be reasonable for
intercity facilities. Also, capacities may be greatly exceeded for links within major metropolitan
areas.
Michigan, which only performs forecasts of truck travel, adopted a fine grained zone system
(approximately townships in the Lower Peninsula) to be consistent with their passenger
forecasts.
Selecting Modes
Mode split models can become excessively complicated when there are many modes and even
more complicated when there are intermodal combinations. However, mode split models need
not be applied to every commodity and to every mode serving each commodity. Since the CFS
provides mode split tables by commodity group, there is some advantage in retaining a large
number of modes throughout the analysis. After traffic assignment, vehicular volumes may be
aggregated into the four major modes of truck, rail, water and air.
In the CFS not every commodity uses every mode. Here is a list of CFS modes.
♦ Parcel, US Postal Service, or courier
♦ Private truck
♦ Forhire truck
♦ Air
♦ Rail
♦ Inland water
♦ Great Lakes
♦ Deep sea water
♦ Pipeline
♦ Private truck and forhire truck
♦ Truck and air
♦ Truck and rail
♦ Truck and water
♦ Truck and pipeline
Freight Forecasting 51
♦ Rail and water
♦ Inland water and Great Lakes
♦ Inland water and deep sea
♦ Other and unknown modes
Different states have adopted different modal categories. For example, Indiana defined its
modes consistent with the CFS. Michigan, Kansas and Kentucky used only trucks. Wisconsin
adopted four modal categories: truck; rail; air cargo; and water
Network Development
The network (for all states except Alaska and Hawaii) should cover the 48 contiguous states but
focus on the state of interest. Canada and Mexico may also be included. Within the state there
would be smaller zones (see previous discussion), and the level of detail for the network will be
greater. The networks beyond state borders should consist of only those links with substantial
interstate traffic (e.g., interstate highways). Even so, networks can be quite large. For example,
the small national railroad network from the FRA contains over 16,000 links, although a
considerable amount of cleaning is possible.
Experience with these large networks indicates the possibility of numerous coding errors. The
networks need to be checked for continuity, for duplicate links, for looped links (links looping
back to the same node) and for multiple nodes at the same location. Not all elements of the
national networks can be checked, so efforts should be concentrated on parts of the network
nearest the state.
The state networks are likely to be derived from a different source than national networks, so
problems can be encountered when splicing them together. For example, a strategy must be
developed for handling loose ends of roads that pass through state borders, but do not appear
on the national network.
Commodities
For many reasons it is best to derive commodity groups from standard categories. Many
existing models have used twodigit STCCs (or SCTGs), which seems to provide enough detail
without overburdening the computations. Most data sources report commodities in ways that
can be consistent with twodigit STCCs, and twodigit SICs are roughly comparable.
Special commodity groups need to be established for waste and deadheading to account for all
types of freight vehicle movement.
Different data sources report the amount of goods by different measures. To reconcile the
various scales, conversion factors must be developed. The CFS provides data in both tons and
dollars, so it can be used to develop those conversions factors for listed commodities.
A truckonly model may bypass the notion of commodities entirely. However, this method is
best suited to urban areas.
Developing a Flow Matrix
There are a number of means available for developing a flow matrix.
1. Adjustment factors can be applied to an existing matrix or the matrix can be expanded
by splitting large zones into smaller ones.
Guidebook on Statewide Travel Forecasting 52
2. Alternatively, a gravity model can be constructed and calibrated on statetostate data,
then applied at the countytocounty level, as was done for Indiana.
Further adjustments to flow matrices may also be needed to account for information not already
included. For instance, the CFS does not include information about commodities originating
outside of the US. Furthermore, the CFS does not contain all commodities.
Import information can be separately obtained (and forecasted). Michigan used the model
INFORUM (from the University of Maryland) to forecast imports. Import data were obtained
from the Transborder Surface Transportation Project of USDOT and from data collected under
Section 6015 of ISTEA.
Trip Generation for Gravity Model
When the gravity model is used for building flow matrices, it is necessary to obtain information
on the amount of production and consumption for each TAZ and for each commodity group.
Production rates can be established in terms of tons per employee or tons per person,
depending upon the category. For some commodities, these rates can be formed by simply
dividing the tons of goods produced in a group by the number of people employed by industries
involved in this group. When there is less than a perfect match between commodity groups and
industrial categories, it is necessary to establish production rates by statistical means, such as
linear regression.
Relating commodity production to employees or population is important when forecasting future
flows, as forecasts for these particular input variables are readily available.
For example, Indiana found trip generation rates by linear regression of CFS data. Michigan
found the following relationships from state data in 1983 CTS (Commodity Transportation
Survey).
Indiana found the following statistical relationships between production in commodity groups
and industrial sectors or population:
STCC1 Farm Products SIC7
STCC11 Coal SIC11
STCC14 Nonmetallic Minerals SIC2 + SIC3
Commodity Group
Annual Tons
per Employee
1 Animals and vegetables 222.5
2 Mining 2897.7
3 Chemicals 435.3
4 Rubber and Plastics 31.0
5 Wood, pulp, paper 273.4
6 Textiles 16.3
7 Stone, glass, clay, concrete 1242.6
8 Metal products 68.6
9 Machinery and electronics 11.8
10 Transport equipment 15.3
12 Other 5.1
Freight Forecasting 53
STCC20 Food and Kindred Products SIC20
STCC22 Basic Textiles SIC22
STCC23 Apparel SIC23
STCC24 Lumber and Wood Products SIC24
STCC25 Furniture and Fixtures SIC25
STCC26 Pulp and Paper Products SIC26 + 0.54*SIC24
STCC28 Chemicals and Allied Products SIC28 + 7.8*SIC29
STCC29 Petroleum and Coal Products SIC29
STCC32 Stone, Clay and Glass Products Population
STCC33 Primary Metal Products SIC33
STCC34 Fabricated Metal Products SIC33 + 2.6*SIC34
STCC35 Machinery, except Electrical SIC35
STCC36 Electrical Machinery SIC33 + SIC34 + 0.75*SIC36
STCC37 Transportation Equipment SIC37
STCC40 Waste and Scrap Material Population
STCC50 Other Manufactured Products Population
Forecasts of consumption are also required but are more difficult to obtain. A preferred method
is to use the information in an inputoutput (IO) table.
Michigan’s IO table was organized by commodity, presumably by associating industries with
commodity groups. There were 11 commodity groups and two nonproducing sectors,
nonmanufacturing and households. The IO table originated with the Bureau of Economic
Analysis. Part of Michigan’s IO Table is shown below.
IO tables usually are organized by industrial sectors, not commodities. For the purpose of
understanding commodity consumption, the IO table should be modified so that the producing
sectors (rows) are commodity groups. Consumption (columns) can be left as industrial sectors,
consistent with economic forecasts.
The individual cells in the IO table are valueadded, not total sales. The size of shipments is
better measured by total sales. Thus, an important assumption must be made: weight of a
commodity is proportional to the valueadded, regardless of the consuming sector.
IO analysis can also provide an answer to the question: What will be the change in all sectors
due to growth in a single sector? Extensive mathematical manipulation of the IO table must be
accomplished to answer this question.
Gravity Model for Freight
The gravity model for commodities is similar to a gravity model for person travel, except that
flows are distributed from a production zone to a consumption zone and that friction factors are
a function of distance.
Commodity Group 3 4 5 6
Non
Manuf HH’s
3 Chemicals 48500 18261 11024 2351 76398 318561
4 Rubber&Plastics 6919 23183 8740 12226 23057 11824
5 Wood Products 4056 2821 88721 1064 66861 75532
6 Textiles 86 1870 3655 47109 6421 89649
Guidebook on Statewide Travel Forecasting 54
V
ij
is the commodity flow from i to j;
P
i
is the production in zone i;
A
j
is the consumption in zone j;
X
i
, Y
j
are balancing factors; and
d
ij
is the distance from i to j.
The gravity model distributes commodities in tons. The map below from the CFS showing the
50 largest flows in the US is a visual confirmation of the principles of a gravity model. Notice
that the largest flows occur between moderate sized states that are adjacent (e.g., Michigan
Ohio, GeorgiaFlorida) and between large states, regardless of the distance (e.g., California
Texas, CaliforniaNew York).
Friction Factors. As with person travel, friction factors can be set in a number of ways. Indiana
used:
although the choice was somewhat arbitrary. d
ij
was taken to be distance in miles. The single
parameter, β, varied considerably by commodity, its magnitude being roughly inversely related
to length of ton haul.
( ) [ ]
ij ij
d exp d f β ·
( )
ij j i j i ij
d f Y X A P V ·
Freight Forecasting 55
It is suggested that β be set by trial and error until average trip length from the model matches
average trip length from data. It is necessary to eliminate bias due to shipment sizes by
calculating average ton length rather than using average shipment length already provided in
the CFS. Average ton length may be found by dividing tonmiles by tons.
Application: During application of the gravity model, P
i
and A
j
are production and consumption
measured in tons. The balancing factors (X
i
and Y
j
) are computed by the software and do not
carry any useful information. The balancing factors just assure that flow is conserved at both
ends of the trip. Kansas used the gravity model for externaltointernal flows, only. The Kansas
model had external stations at state borders and forecasted agricultural commodities, only.
Mode Split Methods
Mode Split Tables. Extensive aggregated data on mode split are available from the CFS and
other sources. Mode split tables can be readily developed. For example, Indiana developed
mode splits by commodity and distance. Distance is a significant factor in mode split, as the
average length of rail hauls greatly exceeds the average length of truck hauls for most
commodities.
Diversion Models/Elasticities. Various organizations have developed diversion models for
estimating the shift of shipments between truck and rail. These include the Intermodal
Competition Model (ICM) from the AAR, the TruckRail, RailTruck Diversion Model from the
FRA and the Strategic Choice Model from Mercer Management Consulting. These models can
be used to generate elasticities that may be applied to individual OD pairs and commodities in a
statewide forecast. Example elasticities are shown later.
Expert Opinion. Expert opinion is often necessary to evaluate the effects of new technologies.
Experts can be asked to give mode splits directly or to provide reasonable assumptions to
permit forecasts by mode split models. For example, Wisconsin used an expert panel to
estimate parameters of forecasts of future truckrail intermodal traffic in the state.
Mode Split Models. Experience with mode split models in statewide forecasts has been limited.
The major factors in mode split models are intended to respond to policy issues. These factors
include:
♦ Commodity characteristics;
♦ Cost;
♦ Time, dependability and frequency of shipment;
♦ Quality; and
♦ Access.
Mode split models include aggregate demand formulations, logit, pivot point and simple
elasticities. Pivot point and elasticities will be discussed later. A typical aggregate demand
model for truck volume (VT) (due to Friedlaender and Spady) looks like:
VT = SR
a
LR
b
ST
c
LT
d
VAL
e
TR
f
TT
g
where SR is the average rail shipment size, ST is the average truck shipment size, LR is the
average length of a rail haul, LT is the average length of a truck haul, VAL is the average value
of the commodity, TR is the rail unit cost and TT is the truck unit cost. The lower case letters
are all calibrated constants. A similar equation was developed for rail shipments.
Regardless of whether the model is aggregate or disaggregate, extensive calibration is required.
Guidebook on Statewide Travel Forecasting 56
Model from NCHRP Report #260. The most complete treatment of mode split is found in
NCHRP Report #260. Because of extensive changes in technology, commodities and markets
since this report, its results are not directly applicable to forecasts done today. NCHRP Report
#260 recommended an allornothing mode split. That is, on comparison of costs all OD flow of
a particular commodity would be assigned to the least cost mode. A very detailed cost model
was provided. This report developed cost relationships for rail, truck and barge. Modal costs
are also given extensive treatment in NCHRP Report #388.
Use of Mode Split Models. Obtaining cost data for interstate OD pairs may be quite difficult,
especially considering the size of zones (states, NTARs). A solution may be to limit analytical
mode split analysis to shipments entirely within the state. External flows would then be split by
table.
Given the difficulty of calibrating models for each commodity, a comprehensive mode split
model covering each mode and each commodity is not recommended. Special analysis may be
performed for those commodities of greatest interest.
A mode split model should be selectively applied to those commodities expected to be sensitive
to the policies being tested.
Pivot Point
One technique for mode split of commodities is pivot point. Pivot point is derived from a logit
model, but requires somewhat less information to operate and is less sensitive to errors of
calibration. Pivot point models are as valid as logit models, but only allow the variation of one
term in the utility function. Since the mode split of commodities is largely determined by cost,
this limitation is not serious. The existing mode shares are required inputs.
In the above equation:
p
j
is the forecasted share for mode j, as a fraction;
p
bj
and p
bk
are the existing shares for modes j or k;
α is a calibrated coefficient which varies by commodity; and
∆c
j
and ∆c
k
are the changes in the full costs of transporting a ton of goods on mode j or k.
The single coefficient of the pivot point model, α, can be calibrated by knowing mode shares
and average modal costs at two points in time. It can be calibrated to match known cross
elasticities between modes (see later discussion). The coefficient could also be calibrated by
analyzing surveys of behavioral intention of shippers.
Because of the difficulties in calibration, it is recommended that the number of modes be kept to
a minimum: truck, rail, water and air. Tables can be used to further refine these categories, if
desired.
Elasticities
An elasticity is the fractional change in output divided by the fractional change in input.
Elasticities can be used to measure the effect on demand of any other important variable. For
( )
( )
∑
∆ α
∆ α
·
k
k bk
j bj
j
c exp p
c exp p
p
Freight Forecasting 57
transportation services it is possible to use elasticities to quantify the effect on demand of
headway, vehicular size, speed, the degree of competition, etc.
For some commodities, mode split may be best accomplished with cross elasticities. For
example, a cross elasticity between truck and rail might be the percent change in rail demand
given a 1% change in truck costs.
Detailed Elasticities. The following table was drawn from NCHRP Report #388. It contains
cross elasticities between truck and rail in Canada, as predicted by the Intermodal Competition
Model (ICM) developed by the Association of American Railroads. Each elasticity is the
estimated percent increase in rail tonmiles for each one percent increase in truck costs. The
average cross elasticity is about 0.5. These elasticities include all factors, including access.
Commodity Density and Load Weights
Converting commodity flows into vehicular
flows usually requires two sets of conversion
factors. The first factor is used to convert
annual tonnage to daily tonnage and the
second factor is used to convert tonnage to
vehicular loads (trucks, containers, cars, etc.)
Truck Days. In order to convert annual flow
to weekday flow, it is necessary to estimate
the number of equivalent weekdays in a year.
Indiana calculated 5.8 (weekdayequivalent)
flow days per week or 300 per year.
Load Weights. Michigan found that most
trucks on intercity trips carried about 4050
thousand pounds. Major exceptions were:
Commodity Low Elasticity High Elasticity
Bulk Farm Products 0.02 0.03
Finished Farm Products 3.5 3.7
Bulk Food Products 0.62 0.83
Finished Food Products 2.0 2.2
Lumber and Wood 0.57 0.73
Furniture 4.0 4.7
Pulp and Paper 0.71 0.93
Bulk Chemicals 0.49 0.67
Finished Chemicals 3.2 3.5
Primary Metals 1.2 1.5
Fabricated Metals 5.2 7.3
Machinery 3.7 4.8
Electrical Machinery 4.1 4.8
Motor Vehicles 0.21 0.28
Motor Vehicle Parts 1.1 1.4
Waste and Scrap 0.17 0.22
Bulk All Else 0.14 0.19
Finished All Else 3.9 4.5
Commodity Group
Tons per
Truck
0 Empty 5.4
1 Animals and vegetables 20.2
2 Mining 36.2
3 Chemicals 24.0
4 Rubber and Plastics 23.9
5 Wood, pulp, paper 29.4
6 Textiles 7.2
7 Stone, glass, clay, concrete 26.1
8 Metal products 27.6
9 Machinery and electronics 15.0
10 Transport equipment 15.6
12 Other 12.7
Guidebook on Statewide Travel Forecasting 58
empty; textiles, clothes, furniture and electrical machinery. Michigan allows larger trucks than
many states do, so some average payloads were considerably higher, especially bulk
commodities. Michigan eventually used the load weights on the previous page in their forecast.
Wisconsin developed load weights for commodities in twodigit STCC. These are shown in the
table below. A comparison of Wisconsin to Michigan would suggest that load weights are rather
constant for some commodities but vary considerably by state for other commodities.
Traffic Assignment
AllorNothing. Allornothing assignment (all trips between an origin and a destination are
placed on the single shortest path) has been the preferred method of assignment by those
doing statewide freight forecasts. Unfortunately, allornothing assignment is very sensitive to
speeds, tending to load most highway traffic on interstate highways. Furthermore, it is difficult
to simulate rail routing decisions with a shortest path paradigm. Extensive adjustments to
impedances (times, costs) on links is required to get acceptable assignments.
Other Methods. Other techniques that have been considered are capacity restrained
equilibrium and stochastic multipath. Capacity restraint techniques are particularly difficult
because freight vehicles, by themselves, rarely exceed the capacity of links. In addition,
highway capacity becomes useless as a modeling concept in urban areas, where only a small
percentage of roads are included in the network. Railroads experience considerable economies
of scale by routing traffic on mainlines, thus violating a major premise of capacity restrained
assignment.
STCC
Tons/
Truck STCC
Tons/
Truck
1 Farm Products 24 32 Clay, Concrete, Glass, Stone 23
8 Forest Products 13 33 Primary Metal Products 19
9 Fish or Marine Products 6 34 Fabricated Metal Products 24
10 Metallic Ores 24 35 Machinery, Not Electrical 9
11 Coal 24 36 Electrical Machinery 8
13 Crude Petro, NG, Gasoline 14 37 Transportation Equipment 12
14 Nonmetallic Minerals 19 38 Instruments, Photo or Optical 5
19 Ordinance or Accessories 24 39 Misc. Manufacturing Products 2
20 Food or Kindred Products 18 40 Waste or Scrap Materials 16
21 Tobacco Produces 5 41 Misc. Freight Shipments 23
22 Textile Mill Products 5 42 Shipping Devices Returned 4
23 Apparel, Finished Textile 3 43 Mail and Express Traffic 3
25 Lumber or Wood Products 15 44 Freight Forwarder Traffic 4
26 Furniture of Fixtures 3 45 Shipper Association Traffic 3
27 Printed Matter 9 46 Misc. Mixed Shipments 7
28 Chemicals 22 47 Small Package Freight 4
29 Petroleum or Coal Products 19 48 Hazardous Waste 16
30 Rubber or Misc. Plastic 4 49 Hazardous Materials 18
31 Leather or Leather Products 3 99 Unknown 12
Freight Forecasting 59
Stochastic multipath assignment shows promise, but there is comparatively little experience with
the technique for statewide freight forecasting. A stochastic multipath assignment sends some
trips along the shortest path and smaller proportions to other reasonable paths.
Rail Issues. The figure on
the right illustrates the
importance of mainlines of
Class 1 railroads for major
commodity movement.
Because of the method
railroads charge shippers
for freight transport,
railroads do not have a
strong incentive to either
minimize time or distance.
(Railroads are
compensated according to
fraction of distance
carried.) Still, many rail
shipments involve multiple
carriers. Thus, a shortest
path is only a rough approximation of the true path. The network must clearly designate points
where shipments change carriers. As readily seen in the figure, many counties do not have rail
access. The pure rail mode would not be a possibility for shipments produced or consumed in
these counties.
Highway Issues. Evidence suggests that for statewide models there is little need to consider
multistop tours. Michigan found that 92.5% of intercity truck trips were direct (no intermediate
destinations). 99% had 0 or 1 intermediate stops.
Assignment Fixes
In the absence of better algorithms, there are a number of fixes that can be applied to achieve
better assignments.
Highway Impedance. Kansas performed three different traffic assignments, each with a
different impedance measure, and then took a weighted average of the volumes from the
assignments. The three impedance measures were based on time, distance and cost. This
method created a smoother loading than might have been obtained otherwise. The averaging
weights were developed subjectively.
Indiana used travel time as its primary measure of impedance, but modified the speed used in
the travel time calculation to provide less of a disadvantage for noninterstate routes. The
following formula was used:
speed = (true speed) + 2 (65  true speed)
0.5
Railroad Impedance. Indiana accounted for railroads’ tendency to assign traffic to mainlines by
adopting the following measure of impedance:
I = L (1/(D + 1))
Guidebook on Statewide Travel Forecasting 60
where L is the length of the link and D is the traffic density in millions of gross tonmiles per
year.
Assignment of Intermodal Flows. Intermodal shipments in the CFS are identified by pairs (e.g.,
truckrail or truckair), without identifying the order of the modes. Thus, a truckrail intermodal
shipment may logically be by truckrail, truckrailtruck, or railtruck (ignoring drayage). Strictly
speaking, each of these three modal combinations should be given separate treatment in the
assignment process. The problem of too many intermodal combinations can be simplified by
assuming that all intermodal shipments are interstate or international. Thus, rail terminals in the
state can be either the production end or consumption end of intermodal flow. Without further
information, it is only possible to assume 50% of each. Some states, like Indiana, have so little
intermodal traffic that it makes sense to simply lump these flows into one or both of the
component modes.
Other Freight Forecasting Issues
Combining Passenger and Freight Forecasts
The Highway Capacity Manual (1997 revision) gives passenger car equivalents (PCEs) for
trucks as a function of grade. The lowest heavy truck PCE is 1.5 (level) and the largest is 15
(6% grades in excess of 1 mile when there are only 2% trucks in the traffic stream). The effect
of a single truck on delay can be considerably greater than the effect of a single passenger car.
If good estimates of delay are needed, trucks should be appropriately weighted to account for
their larger effect:
Delay = f(passenger cars + PCE*trucks)
There are two ways to account for trucks in the traffic stream: (1) apply an average PCE factor
to the whole truck trip table; or (2) apply the PCE factor to trucks on each link individually.
Whole Trip Table. If equilibrium assignment is the desired technique for traffic assignment, then
it is easiest to multiply the whole truck trip table by a single PCE factor that represents average
conditions. The algorithm would assign PCEs rather than vehicles to links, so it would be
difficult to determine how many trucks use each link.
LinkbyLink. Each link (indeed, each link direction) theoretically has a different PCE factor for
heavy trucks. Truck PCEs may be assigned to links ahead of the assignment of passenger
vehicles. In so doing, it is possible to account for the larger effect of trucks on steeper uphill
segments. An equilibrium assignment for just the passenger cars could then be pursued to
obtain final delay estimates.
Forecasting Inputs to a Freight Forecast
Industrial and Population Forecasts. Production and consumption of commodities depends on
forecasts of employment (or sales) for industries and forecasts of population. It is
recommended that these forecasts be obtained from specialists.
IO Analysis. Inputoutput analysis is a technique for forecasting the effect of industrial growth in
a single sector (or few sectors) on the overall economy. IO analysis starts with an IO table, as
illustrated previously. The IO table tells who sells to whom and how much. The analysis finds
the total value added for all industrial sectors given an increase in consumption for a single
sector. The only difficult part of IO analysis is getting the IO table. The Bureau of Economic
Freight Forecasting 61
Analysis publishes an IO table for the whole US economy. IO tables are also included with
proprietary software packages. Once an IO table has been obtained, the mathematics is simple
enough to be performed on a spreadsheet. Regional IO analysis, an extension of traditional IO
analysis, allows identification of the effects of industrial growth on a state or urban area.
Value per Ton. Value per ton may change in the future due to (1) shifts in the types of
commodities produced within a commodity group or (2) or a decrease in the actual value of a
given commodity. Forecasting these effects is difficult and should be reserved for those
commodity groups of special interest.
Productivity. If the forecast of a commodity is based on growth in employment, then an
adjustment is required for anticipated increases in worker productivity. Forecasts of productivity
increases for various economic sectors are available from proprietary data sources.
Load Weights and Shipment Sizes. There has been a recent decline in both shipment densities
and in shipment sizes. These variables affect the load weights and the choice of premium
modes.
Guidebook on Statewide Travel Forecasting 62
Chapter 5. Specialized Methods for Passenger Forecasting
Introduction
This chapter is concerned with a number of topics that enhance or replace parts of the
traditional fourstep model for passenger forecasting. These topics relate to the following.
♦ Surveys. Single station OD surveys, truck travel surveys, freight truck OD surveys,
external travel surveys and survey techniques.
♦ Stated Preference. A method to help calibrate mode split models when they must
include modes that do not yet exist in the corridor or state.
♦ Steps in between Steps. Time of day issues, external trips, vehicle occupancy and
interfacing with urban models.
♦ Total Corridor Demand. Estimating the amount of travel in a corridor, regardless of
mode, without running a complete statewide simulation.
♦ Trip Table Refinements. Using traffic counts to reestimate a trip table that might have
been created through surveys or a gravity model.
♦ Pivot Method. A method for projectlevel forecasting that combines the best features of
a simulation and timeseries analysis.
♦ Calibration. Philosophy of calibration, techniques for calibrating model steps, including
linear regression, analysis of variation (ANOVA), nonlinear regression and maximum
likelihood estimation.
♦ Validation. Understanding validation data sources, especially the Highway Performance
Monitoring System (HPMS), in developing a coherent validation database determining
the quality of a validation.
Data Collection
Validation Data Sets: HPMS Element of the Traffic Monitoring Program
HPMS constitutes a valuable source of validation data at the statewide level. HPMS is a
consolidated data set containing information on every major public road within the state, such as
the category the road belongs to (interstate, major, minor, etc.), the area where the road is
located and volume of traffic it carries. HPMS can be used to prepare traffic volume samples,
either in terms of AVMT (average vehicle miles of travel) or AADT (average annual daily traffic)
for use by the state during the validation phase. The methods adopted for the calculation of
AADT and AVMT estimates are briefly described below.
♦ AVMT Estimation: This process is based on counting periods not usually exceeding a
day. AVMT estimations are sometimes based on longer periods of 48 hours; however,
longer periods become laborious and timeconsuming. The actual calculation requires
multiplying each section’s AADT by the section length and an expansion factor. The
resulting AADTs for all expanded sections are then summed.
♦ AADT Estimation: Like the AVMT, the AADT is, for the most part, based on daily counts.
Correction factors such as seasonal adjustment factors, day of week factors and axle
correction factors are applied in the calculation of the AADT estimate.
Specialized Methods for Forecasting 63
Additional detail in the validation data may be attained by applying vehicle classification
information in HPMS volume samples. The FHWA has adopted 13 vehicle classifications based
on the physical characteristics of the vehicles: motorcycles; passenger cars; other 2axle, 4tire
single unit trucks; buses; 2axle, 6tire, single unit trucks; 3axle single unit trucks; 4 or more
axle single unit trucks; 4 or less axle single trailer trucks; 5axle single trailer trucks; 6 or more
axle single trailer trucks; 5 or less axle multitrailer trucks; 6axle multitrailer trucks; and 7 or
more axle multitrailer trucks. This vehicle classification scheme exceeds the requirements of
travel forecasting.
Traffic Monitoring Guide Recommendations for HPMS Data Collection: The Traffic Monitoring
Guide makes several recommendations about how count data should be collected and
processed for HPMS. Not all states follow these recommendations, but they give a good
indication of the quality of data within HPMS. There are basically three types of counts adopted
by states to maintain traffic volumes. They are continuous counts, conducted throughout the
year to record daily traffic volumes, seasonal counts to keep track of the seasonal factors
affecting these volumes and coverage counts to record more specific information on a particular
road or highway segment. Recommendations include the following:
♦ For coverage counts a 48hour monitoring period should be repeated once every three
years. Recording truck weight at the same time is also recommended.
♦ Both directions of travel should be monitored. Where this is not done the usefulness of
the data for validation purposes is greatly limited.
♦ Samples should be scheduled throughout the year to minimize seasonal biases.
♦ Samples should also be well distributed spatially, so that geographic biases can be
minimized. Randomness of samples should be maintained with regard to both season
and location.
♦ Growth factors may be used to interpolate counts taken every three years to yearly
estimates
♦ Dayofweek factors should be applied to the raw count data to eliminate daily variation.
Moreover, the application of axle correction factors helps to main statistical consistency.
Shortcomings of HPMS Data: The data needs of each state may be different from others.
Thus, it is important to note that although HPMS should be a very consistent data set across
states, each state varies its procedures to account for its own unique geography, such as
population, climate and size of cities.
1
Example, Traffic Counting in Wisconsin: Wisconsin has a well established counting system set
up on all the major highways in the state. Wisconsin adheres closely to the recommendations in
the Traffic Monitoring Guide. It has a total of 142 continuous counting stations located mostly
on high volume roads. Coverage counts are made on a threeyear cycle at 30,000 separate
locations. Seasonal, dayofweek and axle correction factors are accounted for in the
estimation of AADTs. Such a comprehensive database would allow validation of both statewide
and corridor models.
2
1
Traffic Monitoring Guide, FHWA, Publication No. FHWAPL92017, 1992
2
Adapted from Wisconsin Highway Traffic Volume Data, March 1996 and Wisconsin Automatic Traffic
Recorder Data, March 1997
Guidebook on Statewide Travel Forecasting 64
Single Station OriginDestination Surveys
The Michigan Department of Transportation (MDOT) is known for conducting several types of
origindestination surveys over the last 30 years, covering all major state trunklines and
interstate highways. The data set collected from single station origindestination surveys
(SSODs) is probably one of their most extensive sources of information on passenger traffic in
the state of Michigan and represents more than 800,000 “expanded” survey trips. About 300
surveys have already been conducted.
Information was mainly collected through roadside interviews conducted by MDOT personnel
throughout the state. These data included time of day, month and year that the trip was made,
the type of vehicle (or mode of travel), number of persons per vehicle and the origin, destination
and purpose of the trip. Most of the interviews were conducted during a 12hour time frame
(usually starting at 7 a.m. and ending at 8 p.m.), but the sample was eventually expanded to
represent a whole day. One or both directions of travel were recorded. This information was
further supplemented using a vehicle classification counts over a 24 hour period. Since 1985,
129 sites in 49 of Michigan’s 83 counties have been surveyed and recorded
1
.
Since most of the information for these surveys is obtained through direct personal interviews,
the data gives a very realistic picture of the travel patterns throughout the state. The main
advantage of conducting SSOD surveys is that they are capable of capturing atypical trips, such
as vacation trips during the summertime by state residents and nonresidents. In this regard,
conventional household surveys have shown poor results. SSOD surveys also pick up long
trips that are too infrequent to be seen in reasonable numbers in a household survey. Another
advantage is that the SSOD survey information can be used in conjunction with select link
analysis for validation purposes. This procedure will be dealt with in more detail later in this
chapter.
Due to several reasons, the SSOD data that was collected could not be directly applied towards
model development in Michigan. Some of these reasons are listed below:
♦ The SSOD surveys are conducted only during the summer months and on weekdays.
As a result, they may not always relate to the average annual flow.
♦ They do not distinguish between residents and nonresidents of the state.
♦ There are inconsistencies in the trip purposes, and most often conflicting with the trip
purpose breakdown used by the model. A typical example was the Mackinaw Bridge
Survey conducted for MDOT, where 61% of the trips were for vacations. In contrast, the
Michigan model was calibrated with NPTS data, which showed less than 1% of the total
trips as vacation trips. This is another troublesome outcome of conducting these types
of surveys during summer months alone.
♦ The model shows an average trip length shorter than that indicated by the SSOD data.
This is due to the influence of the trip purpose breakdown on the average trip length and
the fact that longer trips have a greater chance of being counted.
Based on this experience, Michigan recommended several possible modifications to SSOD
survey procedures.
♦ Future SSOD surveys should distinguish between residents and nonresidents of the
state.
1
KJS Associates, Inc., Statewide Travel Demand Update and Calibration: Phase II, Michigan
Department of Transportation, April 1996
Specialized Methods for Forecasting 65
♦ Additional SSOD surveys should be conducted for control purposes in other travel
months to help develop seasonal adjustment factors.
♦ All the SSOD surveys conducted in the future should incorporate commercial traffic flow
along with private vehicles.
♦ Inconsistencies with the model, such as the trip purpose definition, should be eliminated.
♦ SSOD data sets should be built in a manner that enables their effective use for the
model development phase.
1
Truck Travel Surveys
Truck travel surveys are done to record more specific information on commercial vehicles,
unlike the SSOD surveys that cater to all vehicle types. Ports of entry and truck weighing
stations on all major highways, carrying large daily volumes of commercial vehicles, in the state
are generally designated as survey stations. Survey time can extend from anywhere between
12 and 24 hours. An advantage of this survey method, unlike an SSOD study, is that it less
biased by the geographical location of a single station. An unbiased sample can be collected at
a controlled sampling rate.
Michigan. A good example is the 1994 Michigan Truck Survey conducted by the Michigan
Department of Transportation. This survey was fairly extensive and included more than 5,000
surveys. The surveys were conducted at 10 different locations within the state. MDOT
personnel used survey forms to record the information. Once the data was checked, it was
stored in database format. Several coding errors were also eliminated in the process. Details
gathered on vehicle classification were consolidated into the 13 categories recommended by the
FHWA Traffic Monitoring Guide. MDOT applied expansion factors to all the survey locations
and to the 13 vehicle classifications. The main function of the expansion factors was to weight
each sample so that the total sum of these factors would give the total volume of trucks
surveyed. For example, if the various truck classifications were to be ignored, then the
expansion factor would be given by dividing the total volume of trucks by the total number of
observations made at the survey station.
2
The commodity groups that were specified for the Michigan truck survey were based on a two
digit STCC. The data collected included the actual number of observations for each commodity
group. This information was then expanded to yield estimates of the number of trips made and
the vehicle miles of travel per group. In order to calculate these expanded numbers, the
expansion factors were multiplied by the actual trip distance.
In addition to classifying the collected information by commodity groups, to assure a statistically
valid database trips were further classified into three specific categories depending on whether
they crossed the border or not. The three categories used were:
♦ Intrastate (no border crossings);
♦ Interstate; and
♦ International (to account for the larger number of trips made across the USCanadian
border).
During the actual survey, up to three trip chains between the ultimate origin and destination
were noted. However, it was found that 92.5% of the trips did not have trip chaining at all.
1
KJS Associates, Inc.
2
Ibid.
Guidebook on Statewide Travel Forecasting 66
Hence, the idea of separately categorizing those trips with an intermediate stop was discarded.
However, for later analysis, two origindestination pairs were considered:
♦ One pair with an ultimate origin and destination;
♦ Another pair taking into consideration only that portion of the trip through the survey
station.
The final application of the model itself is the chief determinant in deciding which of these OD
pairs to use.
1
Washington. In 1993, Washington State Department of Transportation (WSDOT) conducted a
truck survey. This survey was unique as it collected statewide freight data throughout the state
through personal interviews of truck drivers. Again, as with the Michigan Truck Survey, the
main focus was on all the major commercial corridors within the state. Truck and commodity
flows were recorded over 24 hour periods keeping in consideration important seasonal factors.
A total of 25 truck weigh stations and ports of entry were designated as survey locations. The
survey procedure was similar to that followed by the Michigan Department of Transportation.
Here, the goal was to obtain at least 300 surveys during a full day at one location. The I5
corridor in Washington is particularly known for its heavy commercial traffic. In order to keep
the survey process within reasonable limits, the sample was restricted to one out of every 10
trucks on this highway. For the other highways, smaller samples were taken based on the
actual daily truck flow.
Other Survey Methods
Stateline Cordon Survey. Stateline Cordon Surveys are intended to gather information on those
trips originating or ending outside the state. Hence, these surveys should monitor traffic on all
major interstate facilities as well as intrastate highways with a fairly high traffic volume that
contribute to interstate travel.
Most stateline cordon surveys are required to monitor both directions of travel. The most
important data items to be collected include:
♦ Time of interview, such as time of day, month and year (to keep track of any daily or
seasonal variations in the travel patterns);
♦ Type and registration details of vehicle (for trucks, the classification could be based on
the guidelines given by the FHWA Traffic Monitoring Guide);
♦ Occupancy per vehicle;
♦ Details of origin and destination of the trip, both within the state and outside; and
♦ Trip purpose (whether HBW, HBO, NHB, Recreation, Vacation, Business, etc).
2
License Plate Surveys. License plate surveys are particularly useful on heavy traffic corridors
as it does not slow down the traffic. To efficiently conduct this survey, it is essential to have
access to computerized automobile registration files. However, a problem could, arise if these
files are not updated on a regular basis. Also, it is difficult to trace vehicles from other states.
Complete automation is a definite advantage of this system, making it very safe, easy to
execute and comparatively accurate.
Household and OnBoard Surveys. Some of the most practical survey techniques used at the
statewide level are mail or telephone surveys and onboard surveys. Mail or telephone surveys
1
KJS Associates, Inc.
2
Adapted from FHWA, Statewide Travel Demand Forecasting, US Department of Transportation, 1973
Specialized Methods for Forecasting 67
are more economical than conducting home interviews. The sample size can also be
expanded. Again, for efficiency, like the license plate surveys, auto registration files or
telephone directories with names and current mailing addresses can be used. Care must also
be taken to account for households without cars or telephones, although small in number. The
usual response rate for these surveys is anywhere between 20 to 50%. Telephone surveys
usually assure a better response rate, since direct contact is made and followup, when
necessary, is easier. Telephone surveys are often used as the second stage in mail surveys.
A good example is the mail survey conducted for the Kentucky Department of Transportation for
both household and truck travel data. The household survey focused on households owning an
automobile in the state. The total sample consisted of nearly 15,000 households (an average
of 60 households per county). A 45% response rate was achieved.
OnBoard Surveys. These surveys usually focus on public modes of transportation and provide
more comprehensive information than can be obtained from records of collected tickets. Care
should be taken to keep the questionnaires brief and to the point, as travelers must complete
the survey enroute. These surveys help to cover information that cannot be effectively collected
through household surveys for infrequently used modes. Typical questions include the point of
origin, point of destination, access mode, transfer points, trip purpose and traveler
characteristics. The time of day, carrier and vehicle information are usually completed by the
person administering the survey.
1
Validation Quality
Validation involves comparing a baseyear forecast to actual traffic counts. There are three
primary questions that must be answered.
♦ Are the forecasts of total traffic volumes (total volumes, VMT or VHT) sufficiently close to
those measured?
♦ Are the differences reasonable, linkbylink, between the forecast and actual volumes?
♦ Is the forecast spatially unbiased?
Good estimates of speeds are not a major consideration in statewide travel forecasts, although
it is important that inputs relating to speed are reasonable.
For many years, urban models have been validated according to the 1/2 lane rule. That is,
traffic estimates should be accurate to within + or – onehalf the amount of traffic that could be
carried by a single lane at its design capacity, usually LOS C. Roughly speaking, this rule would
mean that urban streets (which are signal controlled) should be accurate to some number
between 300 to 500 vph and freeways and multilane highways should be accurate to some
number between 600 to 700 vph. For all practical purposes, links with volumes of less than 300
vph in the peak period should be ignored. The 1/2 lane rule, if followed, assures that the model
will not cause a major blunder in sizing some future facility.
Perhaps a better means of judging validation quality is to insist that the model be about as good
as the traffic counts to which it is being compared. If the RMS error in traffic counts is known to
be e, then the maximum acceptable RMS error in a model validation should be about 1.4e.
RMS error is calculated as follows.
1
Adapted from FHWA, op. cit.
Guidebook on Statewide Travel Forecasting 68
Acceptable Error Plots
The curve in the figure on the right shows
a typical relationship between link volume
(24 hours) and the RMS error between a
baseyear forecast and ground counts.
The points on the chart are hypothetical,
but illustrate the important statistical
principle that about 68% of all errors
should fall below the line and about 32%
of the points should fall above the line.
Road segments with less than 2000
vehicles per day in a single direction should be ignored, as they are highly unlikely to require
more than one lane per direction under any standard design criteria. (Design criteria include
peak hour in week, 30th highest hour, or 100th highest hour.)
Validation in Low Volume Areas. If validation is required in a portion of the state where there
are many lowvolume roads, it is acceptable to validate the total of these roads, as if they
formed a single link.
Spatial Validation. Validation should be achieved across the state to assure that forecasts will
not be spatially biased. There are no well established procedures for performing a spatial
validation. Some considerations are:
♦ Good agreement with VMT within several big districts that cover the state;
♦ Good agreement with screenline counts across major travel corridors; and
♦ Good agreement with VMT in counties constituting major activity centers.
Stated Preference
Stated Preference is a survey technique that can overcome problems with using only revealed
preference (actual behavior) in calibrating models. For example, Wisconsin DOT wanted to
investigate the demand for a high speed rail alternative. Existing modes in the corridor
consisted of conventional passenger rail, automobile and airline. WisDOT believed that the
characteristics of high speed rail could be significantly different than conventional rail. Thus, it
was necessary to investigate users’ perceptions of the new mode and to approximate its
coefficients within a logit model. As an end product they would need to ascertain a mode bias
coefficient for high speed rail and to determine whether the coefficients on travel attributes
(time, cost, etc.) needed to be modified. There are three general methods of obtaining stated
preference information.
Method A, Ask about Choices. This method describes two alternative modes for a trip already
experienced by the respondent. The respondent is asked whether the new mode would be
chosen, if it is implemented. The respondent’s current mode need not be included among the
choice set. A logit model can be calibrated to the data, giving, among other things, the mode
bias of the new mode relative to existing modes.
P
e
r
c
e
n
t
E
r
r
o
r
Volume
5
10
15
30
25
20
10K 20K 30K 40K 50K 60K
( )
∑
·
− ·
N
1 i
2
estimated counted
V V
N
1
RMSE
Specialized Methods for Forecasting 69
Method B, Ask about Tradeoffs. Tradeoff methods ask respondents to judge the value of one
attribute against another. For example, a popular tradeoff method rates attributes in terms of
their monetary (dollar) value. A typical question would be: “How much more fare (toll) would
you be willing to pay to avoid 10 minutes of travel time on this trip?” This method can yield a
value of time for the new mode.
Method C, Direct Disutility Estimation. Respondents can be shown characteristics of trips and
asked to rate the disutility of the trip on an openended scale. Each respondent can be asked
about many trips, so the sample size can be small. The ratings, along with the trip data, can be
statistically analyzed to obtain most of the coefficients of a utility function. To obtain stable
ratings, the respondents should initially be told the ratings of two arbitrary trips. One trip is
described as having no length and no disutility (and thus a 0 rating). The other trip is a typical
trip and is given an arbitrary rating (e.g., 100).
Wisconsin’s High Speed Rail Study illustrates method A. Respondents were given four sets of
pairwise choices under various automobile LOS conditions, purposes and end points:
♦ automobile v. conventional rail;
♦ automobile v. bus;
♦ automobile v. air; and
♦ automobile v. high speed rail.
These stated preferences were included into a nested logit model calibration with revealed
preferences (actual choices). There were far more stated preference data points than revealed
preference data points.
The survey form did not force a definite response to the stated preference questions. The
respondent could indicate one of five possible answers to any comparison (definitely choose
mode A, probably choose mode A,
indifferent between A and B, probably
choose mode B, definitely choose mode
B). Data points were weighted,
depending upon the strength of the
response (0.9 for definitely, 0.7 for
probably and 0.5 for indifferent).
The final nesting structure is shown on
the right.
Service quality was measured primarily in terms of invehicle time, outvehicle time, cost and
frequency of service. At first the calibration did not result in good estimates of all parameters for
servicerelated variables. Subsequent calibrations held constant the ratios of parameters for the
servicerelated variables, based on earlier studies. Thus, the model estimated parameters for
this combined service quality index, a Chicago Loop (downtown) dummy variable, an income
variable for rail and bus modes, mode specific constants and the logsum constant for the
surfacepublic nest.
The model was further adjusted by hand to replicate existing mode shares and to deal with
aggregate data (e.g., average zonal income) instead of household data. The final set of
coefficients is shown in the table on the next page.
The nest utility is calculated by the equation below. The nest bias constant (seen in a previous
chapter but eliminated here) is redundant, given that each of the modes in the nest have their
own mode specific constant.
Automobile
Decision
Intercity
Bus
High Speed
Rail
Airline
Conventional
Rail
Surface
Public
Guidebook on Statewide Travel Forecasting 70
Value of Time and Value of Frequency  TriState
Wisconsin participated in another high speed rail study with Illinois and Minnesota. Again,
stated preference techniques were used. The TriState study conducted a stratified random
sample of travelers by all modes in the corridor. Respondents were given a series of trip
choices, systematically
varying time, frequency and
cost. There were separate
questions for each mode and
purpose. The data was
analyzed by binary logit.
The table on the right
illustrates a comparison
between the TriState study
and three earlier rail corridor
studies for values of time
(dollars per hour).
Single Mode Vehicle Occupancy: Michigan
Michigan did not use a formal mode split model for statewide travel forecasts. Rather, it
converted person trips to vehicle trips (all relevant modes) by a vehicle occupancy factor. NTPS
(1990) data were used to ascertain vehicle occupancy rates as a function of household
characteristics and trip length. Michigan used vehicles/person (VP) as its primary measure of
occupancy. Michigan’s relationship between VP and distance was determined by linear
regression for both work and nonwork purposes:
∑
φ
φ ·
k
U
n n
n k
e ln U
Mode and
Purpose TriState
New
York
Ontario
Quebec Illinois
Air Business 65 51 58 54
Air Other 34 32 32 19
Rail Business 40 26 25 28
Rail Other 28 21 19 13
Auto Business 43 26 25 23
Auto Other 26 26 18 13
Bus Business 25  17 
Bus Other 22 32 12 
Variable Commute Business Recreation Other
Total Travel Time 0.0126 0.0097 0.0103 0.0103
(OutVehicle AET)/Distance 1.2557 0.9648 1.0292 1.0292
(Travel Cost)/(HH Income) 1.5672 1.0034 1.6055 1.6055
Inverse Frequency 6.67810 2.6051 5.5576 5.5576
Chicago Loop – Bus 1.2729 0.2583 2.2886 2.2886
Chicago Loop – Rail, HSR 0.9523 1.0450 0.3603 0.3603
Income – Rail, Bus 0.0161 0.0122 0.0290 0.0290
Distance < 75 mi  Bus 0.0335 0.0218 0.0397 0.0397
Distance < 75 mi – nonBus 0.4488 0.1725 0.4046 0.4046
Air Constant 14.1117 0.3723 2.1906 0.8380
Rail Constant 2.9360 0.5231 1.8201 1.6875
Bus Constant 0.0452 1.5935 0.9536 0.8126
HSR Constant 1.4067 0.7545 1.3868 1.2501
Logsum – Surface/Public Nest 0.8769 0.8769 0.8769 0.8769
Specialized Methods for Forecasting 71
Work:
adjustment = 0.8679 D
0.05561
exp(0.002045 D)
Nonwork:
adjustment = 1.100 D
0.052769
exp(0.001346 D)
These equations suggest that VP declines with distance (or vehicle occupancy increases with
distance). The distance adjustment is applied to a table of vehicletoperson ratios, one ratio for
each combination of income and household size categories. Michigan found the OD pair VP
ratio by averaging the ratios for the two involved zones. The income and household size
categories are consistent with the trip generation step.
Time of Day Considerations
If there is a need for an hourly forecast, then provisions must be made for the unique nature of
intercity travel. A distinction should be made between statewide and urban models. Time of
day is handled in different ways in different urban models. There are two principle techniques.
First and easiest, trips are generated and distributed for a specific hour in the day. Assignments
are made hour by hour. Second, trips are generated and distributed for a full 24hour period.
The 24hour assigned volumes are allocated to hours and directions by a table of factors for
each link.
These urban methods will not work in statewide models (unless the state is quite small),
because trips found on rural roads are quite long. Long trips (greater than two hours in length)
will likely impact a link one or more hours after its generation. The peak hour of any given link
will not necessarily correspond to the peak hour of zonal trip generation. Tables of factors for
each link are unavailable, difficult to acquire and difficult to estimate. Under HPMS guidelines
states are likely to have time of day information for only a small fraction of their counting
stations.
Urban models do not attempt to distinguish between long and short trips when applying timeof
day factors, because average trip lengths are considerably shorter than the smallest time period
of analysis (typically, one hour or greater). However, those trips on the rural portions of a
statewide network that are quite long in duration are more likely to start in the early parts of the
day. Thus, a statewide model may find it beneficial to develop separate timeofday factors for
trips of different durations.
A Solution, Dynamic AllorNothing Traffic Assignment. Dynamic assignment takes into
consideration the length of trip when assigning it to a network. Dynamic assignment can be
quite difficult to implement when analyzing congested urban networks, but it is conceptually
much easier with less congested statewide networks that would otherwise be analyzed with all
ornothing traffic assignment.
Only a modest change is necessary to existing traffic assignment algorithms in order to
implement a dynamic allornothing (AON) assignment. The following procedure should be
performed.
1. Ascertain the maximum length trip in the network, N hours.
2. Run a special AON assignment for the current hour and each of the N1 previous hours.
The special AON assignment only records link volumes that are between J and J+1 hours in
Guidebook on Statewide Travel Forecasting 72
length, where J is the number of hours prior to the current hour. In essence, only a fragment
of each trip will be assigned to the network.
3. Sum the AON assignments.
4. Validate the baseyear assignment against traffic counts for several hours in the day and for
both directions of travel.
Many existing travel forecasting packages are now incapable of performing this type of
assignment.
Interstate Trip Rate
Regression analysis conducted by the Michigan Department of Transportation (MDOT), with
regard to their statewide travel demand model, has indicated that the trip rate for trips with at
least one trip end outside the state is inversely proportional to the area of the state in square
miles. Or in other words, the proportion of “interstate” trips will be smaller if the state itself is
large. The MDOT analyzed interstate trip rates with regard to total household trip rates. The
main conclusion from this analysis was that the outstate trips were far less relevant than the
trips made within the state, especially when considering the major highways within the state.
Another interesting feature of the outstate trip generation approach in the Michigan model is that
trip productions and trip attractions are computed by the same model, thereby assuming that the
TAZs outside the state produce and attract nearly the same number of trips. However, this
assumption is not valid for the smaller TAZs within the state itself.
1
Development of Outstate Trip Generation for the Michigan Model
The 1990 NPTS data were used by Michigan for calculating outstate trip generation rates for the
Michigan model. These data were drawn from respondents who were asked to report only
those trips with destinations at least 75 miles from home during a stipulated 14day period. The
process was based on the following steps:
♦ Identification of intrastate versus interstate trips;
♦ Segregation of trips by trip purpose;
♦ Calculation of interstate trip rates for households based on both the NPTS and the 1990
census;
♦ Calculation of trip rates for states not found in the NPTS data, a total of 16 states plus
Canada and Mexico; and
♦ Generation of trip ends by applying household trip rates to households.
2
External Travel for Other State Models
Calculation of Through Nonwork Auto Travel for KySTM (Kentucky Statewide Traffic Model). In
1995, Kentucky calculated the total number of productions and attractions for through “nonwork”
trips as a percentage of the total traffic entering and exiting the area under consideration. One
of the disadvantages of using a gravity model to estimate the number and distribution of through
trips is that it is based on accessibility, which has little influence over through trip travel patterns.
For the preliminary distribution of these trips, it was assumed that those through trips were less
likely to change their direction of travel. A spreadsheetbased method was developed following
1
KJS Associates, Inc.
2
Ibid
Specialized Methods for Forecasting 73
this assumption. The productions and attractions at all the external stations were then balanced
using the Fratar method. This method produced a reasonable number and distribution of
through nonwork trips on all the major corridors at the edge of the modeling area.
1
External Trip Generation for Florida Statewide Model. The external trip generation model for the
Florida Statewide model was based on external trip tables developed from roadside interviews
conducted as part of the Florida Department of Transportation (FDOT) Statewide Origin
Destination Survey. The Statewide model considers all trips with an origin or destination
outside Florida to be external. Future year projections are made by applying Fratar factoring to
this trip table.
The Florida Statewide Model has assumed a separate trip purpose for outstate trips with one
end in Florida. The attractions for this separate trip purpose were set proportionally to the
attractions for other trip purposes. The outstate trip attractions at statewide zones were also
assumed to be roughly inversely proportional to the distance from the nearest border.
2
Composite Utilities for Trip Distribution
There does not seem to be a truly satisfying method for adjusting trip distribution to account for
the availability of alternative modes. The need for such adjustment should be obvious when the
different distributions for airline and automobile trips are considered. Airline trips tend to be
long; automobile trips tend to be short. A larger mode split to airlines would imply a lengthening
of trips or a spreading of the distribution. Because typical fourstep models calculate distribution
ahead of the mode split step, it is not possible to cleanly account for multiple modes in trip
distribution. There are two approaches that may be tried.
Approach #1. It may be possible with custom software to perform trip distribution and mode
split simultaneously. In essence, a multinomial logit model is calibrated to estimate shares
going to each destination and each mode. Such a model involves many choices, n*m, where n
is the number of zones and m is the number of modes. It is also possible to use a nested logit
model, where the trip distribution decision is made after the mode split decision.
Approach #2. It is possible to include a composite utility term in the trip distribution model to
better account of the availability of alternative modes. The theory underlying composite utilities
suggests that destinations are more attractive when there are many available modes between
the origin and destination.
The use of composite utilities in urban models is considered to be only a small refinement.
Because of the widely varying modal technologies for intercity travel, the use of composite
utilities in statewide models is considerably more important.
A two mode form of a composite utility function is shown below.
1
Wilbur Smith Associates, Kentucky Statewide Traffic Model Final Calibration Report, Kentucky
Transportation Cabinel, April 1997
2
The Corradino Group, Statewide Highway Traffic Forecasting Model, Technical Report 3, Florida DOT,
August 1990; Post, Buckly, Schuh and Jernigan, Final Technical Report  Validation and Refinement of
1990 Statewide Model, Florida DOT, August 1994
[ ] ) t exp( ) t exp( ln
1
t
2 ij 1 ij ij
α + α
α
·
Guidebook on Statewide Travel Forecasting 74
where t
ij1
is the utility of mode 1 between i and j, t
ij2
is the utility of mode 2, and α is a calibrated
coefficient.
This equation can be extended to more than two available modes by adding more “exp” terms
inside the square brackets.
The value of t
ij
is used in the gravity model as the measure of spatial separation. The coefficient
α is calibrated, but many planners have found that a good value of α can be taken from the
coefficient on invehicle time from the mode split model.
The composite utility is always less than or equal to the smallest utility among the available
modes.
Total Travel in Corridor: Wisconsin Model
Wisconsin used the following equation to forecast volume increases for all traffic in a corridor,
regardless of mode.
where V is volume, P is population, E is employment, I is income, L is level of service, f denotes
future year and b denotes base year. Greek letters are calibrated parameters. The level of
service variable, L, is computed as the composite utility of all modes in the corridor.
This equation was used by the Wisconsin Department of Transportation to estimate demand for
a high speed rail alternative in the ChicagoMilwaukee corridor. WisDOT calibrated this model
for each of four trip purposes, as illustrated in the table below.
TriState Total Demand Model
The TriState study also created a total demand model for a corridor. It had the form:
where
T
ijp
is the OD volume for a purpose;
E
ijp
is a socioeconomic variable, dependent on trip purpose;
C
ijp
is the generalized cost of travel; and
B
0p
, B
1p
, B
2p
are calibrated coefficients.
ε γ β α
,
`
.

,
`
.

,
`
.

,
`
.

·
b
f
b
f
b
f
b
f
b
f
L
L
I
I
E
E
P
P
V
V
Variable Commute Business Recreation Other
Population 0.5621 0.3662 0.5081 0.4371
Employment 0.9046 0.5894 0.5478 0.2878
Per Capita Income 0.3007 0 0.3711 0.0104
Level of Service 0.1865 0.1898 0.3499 0.3606
( ) ( )
p 2 p 1 p 0
B
ijp
B
ijp
B
ijp
C E e T ·
Specialized Methods for Forecasting 75
The three trip purposes were business, commuting and other. The socioeconomic variables
chosen for the trip purpose was:
♦ business  E = (Employment)(Annual Household Income)
♦ commuting, other  E = (Population)(Annual Household Income)
Generalized cost of travel is found from the following equation.
where:
IT
ijm
= in vehicle time between zones i and j and mode m (hours);
OT
ijm
= out vehicle time between zones i and j and mode m (hours);
P
ijm
= interchange penalty in units of time (hours);
TC
ijmp
= travel cost between zones i and j, mode m and purpose p (1990$);
VOT
mp
= value of time for mode m and purpose p (1990$);
VOF
mp
= value of frequency for mode m and purpose p ($ per hour between departures);
OH = operating hours per week (hours); and
F
ijm
= frequency of departures per week between zones i and j and mode m;
Value of time (VOT) and value of frequency (VOF) were found using stated preference
techniques.
The model was calibrated using linear regression on travel survey data, resulting in the
calibrated coefficients shown in the following table.
Hybrid Technique: Pivot Analysis
Four step models are often unsuitable for project level analysis. Errors in link volumes can be
large, both in the base year and for alternatives. However, it is possible to use results from a
fourstep model to “pivot” about known levels of traffic. Select link analysis can be used to
obtain a relationship between zonal activity and traffic levels. Consequently, forecasts of zonal
activity from time series methods can be directly related to traffic levels for a chosen facility.
This method is particularly well suited for project level analysis, where only a few links are being
analyzed and where highly accurate forecasts are essential for each facility. The method
applies to any situation where traffic will not be redistributed due to major network changes or
( )
( )( )
( )( )
ijm mp
mp
ijmp
ijmp
ijm ijm ijm ijmp
F VOT
OH VOF
VOT
TC
P OT 2 IT C + + + + ·
Purpose B0 B1 B2
Business 0.709 1.169 2.750
Commuting 0.241 1.080 2.814
Other 1.186 1.136 2.596
Guidebook on Statewide Travel Forecasting 76
capacity restraint. An advantage of this technique is that is requires only one basecase run of
a fourstep model. There is no need to adjust and rerun the model for future scenarios.
Review of Select Link Analysis
The conventional form of select link
analysis reports the OD flow matrix
for a single direction on a single link.
The figure on the right illustrates a
select link analysis. There are four
zones for a possible 12 interzonal OD
pairs. Two of the OD pairs (AB and
AC) send trips along paths that use
the selected link. These OD pairs
and their associated volumes are
listed in a select link analysis. In this
case, the AB flow is 60 trips and the
AC flow is 40 trips. In a large network
with many zones, a huge amount of data may be generated by a select link analysis. To
achieve an understanding of the results, it is usually necessary to restrain the number of
selected links and to filter out OD flows that are trivially small. A select link analysis, properly
executed, will tell the analyst which zones and OD pairs contribute most to a link’s volume.
Some software packages also implement a form of select link analysis where it is possible to
determine the number of vehicles using a pair of links. For example, an analyst might want to
determine the number of vehicles traveling on First Street (northbound, between Oak and
Maple) and later on Ninth Street (northbound, between Elm and Spruce).
A related technique is select zone analysis. This method finds the number of vehicles using
each link that have either their origin or destination in a given zone. Select zone analysis is
particularly helpful when trying to determine the impact of site developments on surrounding
streets.
It is important that the select link analysis be compatible with the method of traffic assignment.
Some forecasting packages compute their selected link flows from an allornothing assignment,
which is of limited value when equilibrium assignment is the chosen method.
Pivot Steps
The pivot method may be implemented in seven steps.
Step 1. Obtain good average volumes for the link of interest. Counts should be repeated to
assure that statistical variation, dayofweek and seasonal factors have been eliminated.
Ideally, stock adjustment factors should not be used. Counts should be taken for both directions
of travel and by hour of the day.
Step 2. A reasonably well calibrated fourstep model should be available. A “reasonable” level
of calibration means that the model provides good estimates of overall system performance, has
an RMS error in link volumes no more than twice what would be anticipated from traffic counting
variation alone and includes all links in the project. In addition, the links in the project show
good agreement between actual volumes and the basecase forecast. “Good agreement” is
subjective, but the pivot technique would be difficult to defend when the basecase forecast is in
error by more than 50% (high or low) on a given link. In the network shown earlier, the volume
LINK
A
B
D
C
60
40
Specialized Methods for Forecasting 77
on the project link is 130 vehicles, which is considered acceptably close to the 100 vehicles from
the basecase forecast.
Step 3. The fraction contribution of all OD pairs should be obtained from the select link
analysis. From the drawing, there are two OD pairs contributing volumes: AC contributes 40
vehicles, or 40%, and BC contributes 60 vehicles, or 60%. All other OD pairs are ignored.
Step 4: Assuming that the fraction contribution of OD pair to the link’s volume is accurate, it is
possible to estimate the amount of actual traffic associated with each OD pair. In this example,
the 130 vehicles can be split 60/40 across the two OD pairs. Thus, 52 vehicles come from AC
and 78 vehicles come from BC.
Step 5. The growth in activity in an individual zone is forecasted by a time series. As indicated
previously, a measure of zonal activity closely related to intercity travel is total personal income.
Forecasting the growth in an OD pair requires an assumption about the interaction between two
zones. The gravity model would suggest that the growth factor for an OD pair is related
multiplicatively to the growth factor for each zone. For example, if zone B is expected to grow
by 55% and zone C is expected to grow by 38%, then the growth factor for the OD pair can be
found from:
BC growth factor = (1+0.55)*(1+0.38) = 2.139 or a 113.9% growth
The growth in zone A is 46%, so the AC growth factor is 201.5%.
Step 6. The forecasted vehicles for AC is 105 (52*2.015) and the forecasted vehicles for BC is
165 (78*2.139).
Step 7. The forecasted link volume can found from summing the contributions from all OD
pairs. In this example, the forecast is 270 vehicles (105 + 165).
These steps should then be repeated for the other links in the project.
Pivot Method Refinements
Use Larger OD Pairs. A selected link can have contributions from many OD pairs, especially
when an equilibrium assignment or a stochastic multipath assignment has been performed.
Many OD pairs contribute very small amounts of traffic to a link. It is possible to eliminate many
of those smaller contributions, concentrating on a few of the largest OD pairs.
Forecasting OD Flows, Revised. The multiplicative assumption of the previous example would
not hold when large portions of the state are growing at about the same rate. In this case,
growth in traffic for the OD pair should be roughly related to the average growth rate of zones.
A workable compromise is to assume that an OD pair growth equals the regional average
growth times a factor calculated from the degree to which the OD pair exceeds the regional
growth. The following calculations illustrate this concept.
In the previous example, the regional average growth rate is 41%. The growth factors for each
zone in the example can be easily split into regional and local components:
A growth factor = 1.41*1.035 = 1.46
B growth factor = 1.41*1.10 = 1.55
C growth factor = 1.41*0.98 = 1.38
Consequently the OD factors can be easily found:
BC growth factor = 1.41*1.10*0.98 = 1.52
AC growth factor = 1.41*1.035*0.98 = 1.43
Guidebook on Statewide Travel Forecasting 78
With these revised growth factors, the forecasted volume on the project link is:
volume = 52*1.43 + 78*1.52 = 193
which is considerably less than the 270 vehicles calculated earlier.
Calibration and Validation
Calibration and validation are separate tasks, although many transportation planners try to do
both at the same time. Calibration applies to each step in the modeling process, while
validation applies to the model as a whole.
Calibration. Each model step has one or more parameters that can be adjusted to assure that
the step is replicating known travel behavior. Calibration is the process of performing that
adjustment. Very often calibration is performed by statistical methods, such as linear
regression or maximum likelihood estimation. In some cases, it is possible to refine by hand
one or two parameters in a step by adjusting them to match known aggregate measures of
travel. To perform such a hand refinement, it is usually necessary to adopt the remaining
parameters from an earlier study in the same city or from a similar study in another city.
Validation. Validation primarily involves comparing a baseyear forecast to known traffic levels.
A poor quality validation would indicate the need for additional calibration. There is no formal
provision in the validation process to improve the accuracy of the model.
A Muddle. Many transportation planning agencies have adopted a joint validation/calibration
strategy to increase the perceived accuracy of the model. These agencies use the validation
data in the following ways:
♦ Use measures of total travel on a network to adjust the trip production equations;
♦ Use link volumes near special generators to establish trip generation relationships for
those generators;
♦ Use link volumes on major roads to introduce Kfactors into a gravity model; and
♦ Use individual link volumes to adjust free speed on links.
The desire to meld the validation and calibration process is understandable, but the results are
less valid than if a good calibration had first been achieved.
National Defaults
The calibration process can be greatly accelerated if parameters are adopted from national
databases, such as NPTS or ATS, or studies performed elsewhere. Defaults can give good
starting points for calibration exercises or can eliminate the need for some locallycollected data.
At this writing, a dependable source of default parameters for statewide travel forecasting does
not exist. Some states have tried to use the parameters of NCHRP #187 (superceded by
NCHRP #365), but these parameters have been created for urban applications.
Some states have applied results from national databases (e.g., NPTS or CFS) to their
statewide model with good success. The use of the NPTS is facilitated because the raw data is
available on CDROM. Unfortunately, the CFS data is aggregated by state or NTAR.
At this time, only few states have calibrated statewide models. However, as more models
become operational, the potential to share parameters increases.
Specialized Methods for Forecasting 79
Trip Table Estimation from Traffic Counts
A good statewide trip table can be difficult to approximate from a gravity model. For instance,
(1) friction factors may vary considerably across the state; (2) errors in estimating the number of
intrazonal trips may unduly influence the accuracy in interzonal trips; and large biases can be
introduced by barriers to travel, such as bridges and state borders. Common practice has been
to adjust the gravity model with “K” factors to account for sizable discrepancies in the gravity
model. An alternative method worth investigating is to statistically adjust an estimated trip table
to match traffic counts. A smallscale test of this concept has been tried in Wyoming.
Several methods of estimating OD tables have been described in the literature; the details of
these methods are beyond the scope of this guidebook. One method, entropy maximizing, can
be adequately described by its inputs:
V
a
= volume of traffic on link a;
p
aij
= proportion of trips from origin i to destination j carried by link a; and
t
ij
= prior trip table of trips from origin i to destination j
The method finds the “maximum entropy” trip table given these input data. The prior trip table
could be taken from surveys or estimated by a gravity model or Fratar model, and the volumes
are taken directly from traffic counts.
The proportion of trips between i and j on link a, p
aij
, needs further explanation. They must
satisfy the requirement:
where T
ij
is the final trip table. These proportions are most easily found by traffic assignment.
When allornothing assignment is performed, the p’s are Boolean (0 if the link is not used, 1 if it
is used). With any multipath techniques (including equilibrium), fractional values are possible.
SSOD in Validation
SSOD data is an important input in the validation phase of the model. This is done by means of
select link analysis. The selected link assignment output from the model can be directly
compared with the SSOD survey data, either as individual links or as groups of links. The
Michigan Department of Transportation has used recent SSOD weekday surveys in the
validation phase of their model, as SSOD data was not directly used in the model’s
development. To do so, Michigan assigned the SSOD trip table back to the network, then
compared the resulting assignment with a select link analysis. A major shortcoming in the
MDOT model’s selected link analysis was that only one link could be selected per run of the
model, whereas the complete SSOD data set involved many “stations” and, thus, several links.
1
Michigan’s application of this procedure would be better classified as a calibration tool, because
it focuses entirely on the assignment step. For validation, it would have been better to have
assigned the full trip table as calculated in the trip distribution step, thereby testing all steps in
the model together.
1
KJS Associates, Inc.
ij
i j
aij a
T p V
∑∑
·
Guidebook on Statewide Travel Forecasting 80
Integration of Statewide and Urban Models
Statewide and urban models use different types of data and different algorithms. Generally
speaking, urban models show greater geographic and geometric detail and should be far more
accurate within the urban area. Consequently, urban model results should be used for links
within urban areas, with the results from the statewide models used elsewhere.
In order to assure that the urban models are properly using statewide model results, the
statewide model must be capable of performing linktolink select link analysis. This type of
select link analysis gives the OD flows between many pairs of links. If the selected links in the
statewide model correspond exactly to the external stations in the urban model, the select link
OD flows are theoretically equivalent to the EE (externaltoexternal) trip table in the urban
model.
Some fixing is usually necessary to account for errors in the statewide model. A common
problem is that the forecasted traffic from the statewide model in the base year fails to match
the known traffic counts. Errors of this sort can be overcome by Fratar factoring the trip table.
Another problem is numerous zerovalued cells in the OD matrix. This problem arises when the
statewide model uses allornothing assignment. To the extent possible using local knowledge
of existing travel pattern, the number of zerovalued cells should be reduced by manual
adjustments of the trip table.
Appendix: Review of Calibration Methods
Ordinary Least Squares (Linear Regression)
Ordinary least squares, or linear regression, analysis is the primary method for calibrating
statewide travel demand models.
Trip Attractions. The estimation of trip attractions requires little more than an elementary
knowledge of linear regression analysis and can be easily performed on a spreadsheet. Trip
attraction equations are typically of the form:
where the x’s are measures of zonal activity (population, employment, etc.) and the a’s are
calibrated coefficients. It is important to tell the statistical software to set the constant term (“a
0
”
in the above equation) to zero in order to force the line through the origin. Otherwise, zero
activity in a zone would not be associated with a zero amount of trip making.
Trip Productions. The estimation of trip productions is most often accomplished by category
analysis or by cross classification. In both methods the parameters can be estimated using a
dummy variable to represent each possible category or cell. A related method, analysis of
variance (ANOVA) can used to determine whether a categorization or a classification scheme is
appropriate. The ANOVA routines found on spreadsheets are inappropriate, because they do
not allow an unequal number of observations in cells.
Trip Productions with a Covariate. When the parameters of a trip production model are
estimated with linear regression analysis, it is possible to include continuous variables into the
model. For example, a trip production model might consist primarily of a crossclassification
procedure with cells representing all combinations of household size and automobile ownership.
n n 2 2 1 1 0
x a x a x a a A + + + + · L
Specialized Methods for Forecasting 81
The model might be enhanced by including an accessibility term, such as distance to the
nearest regional shopping mall. All parameters would be estimated at the same time.
Linearization. Acceptable fits of nonlinear relationships can often be achieved by transforming
them so that their parameter can be estimated with linear regression analysis.
Consider the total demand model from the TriState High Speed Rail Study:
The parameters of this equation were estimated by first taking the natural logarithm of both
sides of the equation:
This equation is now linear in the logarithmically transformed variables. Logarithmic
transformations disturb the distribution of errors of the dependent variable, so linearization may
not always be the best strategy for obtaining unbiased parameters. As discussed in an earlier
chapter, logarithmic transformations are entirely appropriate when the size of the error is
proportional to the size of the dependent variable.
Nonlinear Regression
Nonlinear regression is implementable on a spreadsheet using a “solver” feature. In nonlinear
regression, the following expression is minimized by varying the parameters of the model,
Y
i,estimate
. That is,
where each i represents a data point.
Maximum Likelihood Estimation
Certain model steps that involve estimating probabilities are often calibrated with maximum
likelihood estimation. These steps include mode split (logit and nested logit) and trip distribution
(gravity model with complex friction factor functions and attractiveness factors). Like regression
analysis, maximum likelihood estimation finds the parameters of a model. However, maximum
likelihood estimation is designed to find predictive equations for choices and does not attempt to
find the best line to fit a set of data. Maximum likelihood estimation sets parameters of the
model so that the probability of exactly replicating by chance the observed pattern of choices is
the highest. Usually, there are many possible choices in a data set and the probability of an
exact replication is extremely low, regardless of the parameters.
Maximum likelihood estimation begins by specifying a “likelihood function”, L, which computes
the probability of seeing an exact replication of observed choice patterns.
( ) ( )
p 2 p 1 p 0
B
ijp
B
ijp
B
ijp
C E e T ·
( ) ( ) ( )
ijp p 2 ijp p 1 p 0 ijp
C ln B E ln B B T ln + + ·
[ ]
2
i
estimate , i actual , i
Y Y min
∑
−
Guidebook on Statewide Travel Forecasting 82
Here P
jn
is the estimated probability of any given traveler n choosing mode j, as given by the
choice model (containing assorted variables and parameters to be estimated). Also, δ
jn
is a
Boolean (0,1) variable that is set equal to 1 if the traveler n had chosen mode j and set to zero
otherwise. This equation is difficult to work with, because it usually evaluates to a very small
number. Thus, statisticians commonly try to maximize the logarithm of the likelihood function 
an equivalent and much easier task.
It is readily seen that all terms in the log likelihood function that represent unchosen modes are
zero, and all other terms are negative numbers (i.e., the logarithm of a number less than 1 is
negative). Thus, maximizing the log likelihood function involves making it less negative by
reducing its magnitude. The means of performing the maximization vary, depending upon the
software. The “solver” capability of a spreadsheet is one possible method.
Fast Trip Distribution Calibration
When there is a single parameter in a gravity model friction factor function, that parameter can
be set so that the model yields the same average trip length as observed in reality. That is,
where t “hat” represents the model’s average trip length and t “star” represents the actual
average trip length. Single parameter friction factor functions include the power function:
and the exponential function:
The parameter is estimated by repeatedly running the model and adjusting the parameter up or
down. If the estimated average trip length is too small, then the magnitude of the parameter
(exponential or power) should be made smaller. If the estimated average trip length is too large,
then the magnitude of the parameter (exponential or power) should be made larger.
The choice between the power function or the exponential function (or any other convenient
oneparameter function) must be made by comparing the differences between the estimated
and observed trip length frequency distributions.
The true average trip length can be ascertained by questionnaire or by analysis of NPTS.
∏∏
· ·
δ
·
N
1 n
J
1 j
jn
jn
P L
jn
N
1 n
J
1 j
jn
*
P ln L max
∑∑
· ·
δ ·
α
·
ij
ij
t
1
F
ij
t
ij
e F
β −
·
*
t t ·
Specialized Methods for Forecasting 83
Acknowledgements
The project was conducted by the Center for Urban Transportation Studies, University of
Wisconsin  Milwaukee. The guidebook was written by Alan J. Horowitz, Professor of Civil
Engineering, with major contributions by David Farmer and Smitha Vijayan. The appendix was
adapted from David Farmer’s MS thesis. Editorial contributions were made by Linda Rupp.
The project was coordinated with the Wisconsin Department of Transportation; Randall Wade
served as WisDOT’s main contact.
FHWA’s staff contributing to this guidebook include Leroy Chimini, Robert Gorman and Stefan
Natzke. Additional guidance was provided by Tony Esteve and Phil Hazen (retired) of FHWA.
Guidebook on Statewide Travel Forecasting 84
APPENDIX: The State of the Art in Statewide Travel Demand
Forecasting
A.1. Introduction
The belief has long been held that the characteristics of travel on a statewide scale are in many
ways different than those of travel within an urban area. Thus, encouraged by early progress in
the development of urbanarea models, attempts have been made over the last 35 years to
formulate models and techniques to forecast transportation activities on a statewide scale.
Many states expended a significant amount of time and expense – especially in the late 1970s
and early 1980s – to develop statewide models. Some succeeded in developing a working
model, but most did not.
At present, most states do not have a forecasting process in place at a statewide level –
whether due to past difficulties in developing and maintaining a statewide model, or due to a
cynicism bred by the often highly political process of transportation planning. However, a
formalized statewide forecasting process offers a rational basis for making planning decisions.
The use of travel forecasting models can assist with decisions regarding future facility needs,
budget projections and the assessment of the large scale effects of alternative projects.
Statewide models help to tie the decisionmaking process to a knowledge of the interaction
between transportation systems and socioeconomic structures at a statewide scale.
Most recently, in reaction to the requirements of the 1990 Clean Air Act Amendments (CAAA)
and the 1991 Intermodal Surface Transportation Efficiency Act (ISTEA), there has been
renewed activity in travel forecasting at a statewide level. In fact, implementation of most of the
23 “planning factors” (see Figure A.1) outlined in the ISTEA legislation [1.1] can be assisted
greatly by a statewide forecasting effort.
Statewide forecasting can have a significant impact on at least six of these “factors” including:
1. Transportation needs for nonmetropolitan areas;
2. Connectivity between metropolitan areas;
3. Recreational travel and tourism;
4. Preservation of rightsofway for future projects;
5. Longrange needs of the state transportation system; and
6. Methods to enhance the efficient movement of commercial goods.
These planning factors have been consolidated in TEA 21, but the need for a statewide model
remains. More specifically a statewide forecasting effort can, as one report noted [1.2]:
1. Assess demand by specific customer/market segments;
2. Forecast passenger and commodity flows for a 20year horizon;
3. Provide an improved tool for trunkline planning and analysis;
4. Enable multimodal analysis along major intercity corridors;
5. Provide quantitative data input to management systems; and
6. Integrate with and provide input to urban models.
The purpose of this appendix is to present a review of the state of the art travel forecasting at
the statewide level and to identify possible avenues for improving the forecasting process.
The State of the Art in Statewide Travel Demand Forecasting 85
1. The results of required management systems.
2. Any federal, state, or local energy use goals, objectives, programs or requirements.
3. Strategies for incorporating bicycle transportation facilities and pedestrian walkways.
4. International border crossings and access to ports, airports, intermodal transportation
facilities, major freight distribution routes, national parks, recreation and scenic areas,
monuments and historic sites and military installations.
5. The transportation needs of nonmetropolitan areas.
6. Any metropolitan area plan.
7. The connectivity between metropolitan areas within a state or with metropolitan areas in
other states.
8. Recreational travel and tourism.
9. Any state plan developed pursuant to the federal Water Pollution Control Act.
10. Transportation system management and investment strategies to make the most efficient
use of existing facilities.
11. The overall social, economic, energy and environmental effects of transportation decisions.
12. Methods to reduce traffic congestion and to prevent congestion from developing.
13. Methods to expand and enhance transit services and to increase the use of such services.
14. The effect of transportation decisions on land use and land development.
15. Strategies for identifying and implementing transportation enhancements.
16. The use of innovative mechanisms for financing projects.
17. Preservation of rightsofway for construction of future transportation projects.
18. Longrange needs of the state transportation system for movement of persons and goods.
19. Methods to enhance the efficient movement of commercial motor vehicles.
20. The use of lifecycle costs in the design and engineering of bridges, tunnels or pavement.
21. The coordination of transportation plans and programs developed by MPOs.
22. Investment strategies to improve adjoining state and local roads that support rural
economic growth and tourism development, federal agency renewable resources
management and multipurpose land management.
23. Concerns of Indian tribal governments.
Figure A.1. Summary of ISTEA Statewide “Planning Factors”
It should be noted that the title of this appendix deals with the “state of the art” in travel
forecasting. Some previous research [1.3] has attempted to differentiate between the “state of
the art” and the “state of the practice”. This appendix, however, makes no such distinction. The
approach followed here assumes that the knowledge that is elsewhere divided into “art” and
“practice” is, in fact, merely part of the same continuum.
Guidebook on Statewide Travel Forecasting 86
Method of Research
In order to assemble information for this appendix, two distinct research methods were applied.
First, a search was made through the nearlycomplete set of Transportation Research Board
(TRB) bulletins and reports, National Cooperative Highway Research Program (NCHRP) reports
and syntheses and other published materials available in the collection of the Center for Urban
Transportation Studies at the University of WisconsinMilwaukee. This search was
supplemented by an examination of the references found in applicable TRB and NCHRP
literature and through online searches of the Northwestern University library. Surprisingly little
information that applies specifically to statewide travel forecasting was discovered in these
searches. Instead, much of the available information describes “intercity” models, which (as
discussed in section A.2) bear many relationships to statewide models.
The second method of research consisted of direct contact with various state department of
transportation (DOT) officials. For as many states as possible, contacts were initiated using the
Internet. Many DOT officials were directly accessible via email or could be easily reached
through a general information email addresses at their respective DOT web sites. Other
contacts could be made through access to DOT phone lists that are also available online. As
shown in Table A.1, attempts at contact were made with 45 states. Alaska and Hawaii were not
contacted because of their geographical separation from the lower 48 states, Rhode Island and
Delaware were not contacted because of their small size, and Mississippi was not contacted
because its DOT did not support a home page on the internet. As also shown in Table A.1, an
overwhelming majority of the states contacted replied via email or telephone, and many sent
documentation of their individual statewide forecasting models and procedures. A similar,
electronicallybased procedure was followed in contacting several commercial sources that
provide the population and economic forecasts that are often used as inputs to the
transportation forecasting process.
Organization of Appendix
The following sections of this appendix present a review of the information gathered during the
research process described above and some observations and recommendations about the
state of the art as portrayed by that information. Section A.2 and Section A.3 provide a review
of the literature of intercity models for passenger travel and freight transportation, respectively.
The formulation of intercity models, in general, historically preceded the development of
statewide models. Thus, the intercity models are presented first. The information in these
sections is drawn principally from TRB and NCHRP publications. Similarly, Section A.4 and
Section A.5 provide a review of statewide forecasting methods for passenger travel and freight
transportation, respectively. Much of the information reviewed in these sections was collected
from the DOT sources. Section A.6 presents a review of two statewide passenger models now
under development: the Michigan Statewide Travel Demand Model and the Kentucky Statewide
Traffic Model. Section A.7, likewise, presents a review of two statewide freight models now
under development: Multimodal Freight Forecasts for Wisconsin and Transport Flows in the
State of Indiana. Finally, Section A.8 provides some recommendations for “best practice” based
on the observations made in the preceding sections.
The State of the Art in Statewide Travel Demand Forecasting 87
Table A.1: Contact with State Departments of Transportation
DOT Contact
Made
Reply
Rec’d
Items
Sent
DOT Contact
Made
Reply
Rec’d
Items
Sent
AL Yes No  MT Yes No 
AK No   NE Yes Yes No
AZ Yes Yes No NV Yes Yes No
AR Yes No  NH Yes Yes Yes
CA Yes Yes Yes NJ Yes Yes Yes
CO Yes Yes  NM Yes Yes Yes
CT Yes Yes Yes NY Yes Yes Yes
DE No   NC Yes Yes No
FL Yes Yes Yes ND Yes Yes No
GA Yes Yes No OH Yes Yes No
HI No   OK Yes Yes No
ID Yes Yes No OR Yes Yes No
IL Yes Yes No PA Yes Yes No
IN Yes Yes Yes RI No  
IA Yes Yes No SC Yes Yes No
KS Yes Yes Yes SD Yes Yes No
KY Yes Yes Yes TN Yes No 
LA Yes Yes Yes TX Yes Yes Yes
ME Yes Yes Yes UT Yes No 
MD Yes Yes No VT No  Yes
MA Yes Yes No VA Yes Yes No
MI Yes Yes Yes WA Yes Yes No
MN Yes Yes Yes WV Yes Yes Yes
MS No   WI Yes Yes Yes
MO Yes Yes No WY No  Yes
A.2. Intercity Passenger Literature
Intercity travel is a broad heading that includes statewide travel. As used here, the term
“intercity” forecasting involves the prediction and assignment of traffic volumes between cities or
other points of interest that are separated by some significant distance. The term “intercity” is
also used to distinguish these models from “urban” models, which typically involve travel
between more closely spaced points of interest within a localized area. Intercity models include
corridor, statewide, regional and national models. Statewide models are therefore a subset of
intercity models. The main point, first expressed as early as 1960 [2.1, 2.2], is that the
characteristics of intercity travel are inherently different from those of travel within an urban
area. It is assumed that people travel according to a somewhat different set of rules over longer
distances and between metropolitan areas. The intercity models encountered in the literature
are often associated with an academic exercise and therefore make use of fewer, more carefully
chosen origindestination (OD) pairs than would normally be included in a meaningful statewide
model. Consequently, they generally present situations that are a little more abstract in nature.
The similarities to statewide models are many.
Guidebook on Statewide Travel Forecasting 88
Types of Intercity Passenger Models
A number of reviews have been made of the early history of intercity modeling [2.3, 2.4, 2.5, 2.6,
2.7] and most include some discussion of the taxonomy of intercity models. Intercity models
can essentially be divided into four types on the basis of two categories: data and structure.
The models can make use of either aggregate or disaggregate data and can be of a direct
demand or sequential structure. The four resulting combinations are (1) aggregate direct
demand models, (2) aggregate sequential models, (3) disaggregate directdemand models and
(4) disaggregate sequential models. Intercity travel demand models can be further classified by
whether they encompass only a single mode (modespecific) or multiple modes (total demand)
and by which trip purposes they include.
Aggregate data makes use of the socioeconomic data for the OD pairs in the model and can
also include the service characteristics of the modes of travel between them. Disaggregate data
goes further to examine the motives and characteristics of the trip makers at an individual or
household level and are typically used to generate the probability that a particular trip is taken or
mode is used. In terms of model structure, a direct demand model is one that calculates all of
the desired travel information in one, singly calibrated step. (Direct demand models are
sometimes called econometric models because of their resemblance to statistical models of
economic demand.) A sequential model, on the other hand, divides the modeling process into
several individually calibrated steps. The urban “fourstep” modeling process, which many
DOTs have adopted for the statewide modeling purposes, presents the quintessential example
of a sequential model.
Aggregate DirectDemand Models
The earliest intercity models were of the direct demand type and were developed in the 1960s
as part of an examination of the Northeast Corridor [2.6]. The most famous of these was
Quandt and Balmol’s abstract mode model [2.8]. The reader is referred to the reviews
referenced in the previous section (especially Koppelman et al [2.6]) for a more complete
historical perspective of significant intercity modeling efforts. The following directdemand
models – some of which are not mentioned in those references – are noted here because they
possess features that might prove useful to modeling at the statewide level.
A notable early innovation was attempted by Yu [2.9]. Yu took the standard directdemand
formulation – regressed from crosssectional data – and recognized that the elasticities present
in the crosssectional data would not necessarily remain constant over time. His paper presents
two singlepurpose (one for business travel and one for personal travel) directdemand models
in which the regression coefficients each include a timeseries component. It is a novel idea
that does not appear to have been picked up by succeeding authors. Another innovative idea is
found in Cohen et al [2.10]. Here, as part of two singlepurpose (business and nonbusiness)
directdemand models, the authors propose to use a pivotpoint procedure. The procedure is
intended to eliminate the effects (on the traffic volumes to be forecasted) that result from
variables that have been excluded from the models. Description of the pivotpoint procedure is
brief, however, and use of this procedure does not seem to have been adopted by other
researchers.
By the late 1970s directdemand models were being constructed to include an increasingly
wider range of variables to account for the enormous variety of factors that influence travel
behavior. Models presented by Peers and Bevilacqua [2.11] and Kaplan et al [2.12] give some
sense of this trend. Peers and Bevilacqua describe a model that includes a long list of policy
sensitive variables, arranged into three groups: (1) extensive variables, including population and
employment; (2) intensive variables, including persons per household, income per household
The State of the Art in Statewide Travel Demand Forecasting 89
and employment per acre; and (3) system variables, including travel speeds and costs.
Meanwhile, Kaplan et al describe their Passenger Oriented Intercity Network Travel Simulation
(POINTS) model, a multimodal model that explicitly includes consideration of accessibility to the
transportation system. Both of these models provide a bridge from an earlier emphasis on
aggregate modeling to the growth in disaggregate modeling research by the early 1980s.
Disaggregate Sequential Models
One of the first applications of disaggregate (or behavioral) modeling was for the modechoice
step of sequential models. It is possible to develop a modechoice model without disaggregate
data, as DiRenzo and Rossi did, using a “reasoned’ diversion model [2.13]. Disaggregate
models, however, typically use a logit formulation to provide a convenient way of including a
number of modeabstract, transportation accessibility, policy related and behaviorallybased
variables in the modeling process. Due to parallel research in urbanarea forecasting in the
early 1980s, these models became more attractive. They were thought to be especially useful
in the effort to estimate the shifts in mode share that were expected from deregulation in the air
and intercity bus industries and from the anticipated implementation of highspeed rail
transportation [2.14, 2.15]. Again, Koppelman et al [2.6] provides a review of many of the earlier
disaggregate mode choice models. In addition, Miller [2.16], Forinash [2.17], and Forinash and
Koppelman [2.18] provide studies of the various structures (binomial, multinomial and nested
multinomial) available to more realistically represent the crosselasticities between modes and
to eliminate irrelevant alternatives in the logit modesplit formulation.
Armed with an increasing understanding about the implementation of disaggregate modeling
techniques and fueled by the increasing availability of disaggregate data, several researchers
have developed complete travel demand models based on the analysis of disaggregate data in
a number of discrete, nested steps. Morrison and Winston, for instance, present multimodal
models (one for vacation travel and one for business) with the hierarchical structure shown in
Figure A.2 [2.19]. Similarly, Koppelman [2.20] and Koppelman and Hirsh [2.21, 2.22] present a
multimodal model with a structure shown in Figure A.3. Morrison and Winston make use of the
1977 National Travel Survey (NTS) data, while Koppelman and Hirsh use both the NTS and the
1977 National Personal Transportation Survey (NPTS) data. Both pairs of researchers seek to
use this disaggregate data in a model structure that mimics the behavioral logic of trip making.
One Disaggregate DirectDemand Model
Another model of interest is the disaggregate directdemand model developed in the 1980s by
the Egypt National Transportation Study [2.23, 2.24, 2.25]. The Egyptian Intercity
Transportation Planning Model estimates travel on seven modes for travelers in three income
levels. It is unusual in its use of disaggregate data in a single equation (directdemand) format.
Also, unlike many intercity passenger models, it includes capacity restraints on the network,
most notably for the shortage of passenger rail cars. Because it deals with a very practical
situation, the Egyptian model could reasonably be noted in the section of this appendix
describing statewide forecasting techniques, but since the transportation situation in Egypt is
sufficiently an abstraction of the situation in the United States, it seems fitting to include it with
the intercity models. It might also be noted that, in its treatment of rail car capacity restraints, it
resembles some freight models, as well.
Guidebook on Statewide Travel Forecasting 90
Figure A.2. Structure of Morrison and Winston’s Model
Trip Frequency
One None
Destination 1 Destination 2 Destination i
Mode 1 Mode 2 Mode j
Service Class 1 Service Class 2 Service Class k
. . .
. . .
. . .
Figure A.3. Structure of Koppelman and Hirsh’s Model
Make Trip
Destination 1 Destination 2 Destination i
Mode 1 Mode 2 Mode j
Rent a car Don't rent a car
. . .
. . .
The State of the Art in Statewide Travel Demand Forecasting 91
SingleMode and SinglePurpose Models
Besides the ubiquitous singlemode automobile models, there are two other types of single
mode models of interest: bus and air. (Most passenger rail models are a part of a multimodal
model.) Modeling of intercity bus travel has proved to be difficult [2.26] and examples of
intercity bus models are rare. One interesting bus model is presented by Neumann [2.27]. His
model describes a probabilistic (disaggregate) model based on a Poisson distribution of
ridership, as opposed to a regression model. He concludes that this formulation provides a
simpler and more reasonable estimate of ridership on rural bus routes.
Several air travel models are also of interest. As early as the 1960s it was recognized that the
yeartoyear growth in air travel makes the use of timeseries techniques valuable, and a 1968
paper by Brown and Watkins [2.28] addresses this issue with simple linear regression
techniques. A later paper by Oberhausen and Koppelman [2.29] also looks at timeseries
analysis of air traffic patterns using a BoxJenkins procedure to account for cyclical (seasonal
and yearly) variations in travel behavior. In another study, Pickrell [2.30] uses a combination of
techniques to assess future trends in intercity air travel. Pickrell uses a singlemode direct
demand model to estimate the total demand for air travel. At the same time, he uses an
aggregate modechoice model to predict the percentage of market share that the air mode could
generate under several alternative futures. Other air travel models of interest include a
regression analysis of travel between small cities in Iowa by Thorson and Brewer [2.31] and an
elaborate directdemand model of intercity air travel based on qualityofservice measures by
Ghobrial and Kanafani [2.32].
Finally, the one other singlepurpose intercity model worth noting is the disaggregate model of
recreational travel presented by Gilbert [2.33]. Gilbert’s model is sufficiently abstract to be
included here with the other intercity models, but more will be said about recreational travel
models in Section A.4. It should be sufficient to state here that Gilbert’s paper, published in
1974, is one of the latest papers found to specifically address the recreational trip purpose.
Discussion
As will be seen in the following sections, the intercity forecasting techniques employed in most
existing statewide models are principally those of the aggregate sequential type. This is partly
due to the strong traditions of and training in the fourstep modeling process, but it is also due to
the general failure of disaggregate techniques at a statewide scale. Although disaggregate
models are attractive because of their ability to include the behavioral aspects of travel, their
principal drawback is the lack of sufficient disaggregate data for calibration of statistically
meaningful statewide models. Until further data is available their use will remain limited.
It should also be noted that there is a place for aggregate directdemand models at a statewide
scale. This econometric type of model can be especially useful in tying the forecast of single
quantity (annual VMT or emissions, for instance) to forecasts of socioeconomic data.
A.3. Intercity Freight Literature
Introduction
Intercity freight models are similar to intercity passenger models in their attempt to model traffic
– in this case, freight traffic – between spatially distant locations. Compared to the amount of
literature available about intercity passenger forecasting, the amount of intercity freight
forecasting literature is rather small. The history of interregional inputoutput analysis – which
Guidebook on Statewide Travel Forecasting 92
is closely related to some models of freight transportation – dates back to the 1930s, but serious
attempts to forecast freight movements at a national or regional level were generally begun only
after the first intercity passenger models were developed in the middle 1960s.
Reviews of Intercity Freight Models
As in the case for passenger models, several reviews have been published about the history of
intercity freight modeling. A paper by Smith [3.1] divides previous models into six categories
based on their structure: (1) market share, (2) inputoutput, (3) inventory theoretic, (4) gravity,
(5) abstract mode and (6) linear programming. Smith concludes that either a gravity model or
an abstractmode model would be best for use in the situation where the available data are
limited. In a 1983 paper, Winston [3.2] divides freight models into two categories familiar to
passenger modelers: aggregate and disaggregate. He observes that models based on
aggregate data might be better for regional level freight flows. He also recognizes the
difficulties inherent in collecting the immense amounts of data required to calibrate disaggregate
models. Friesz et al [3.3] also present an overview of early freight modeling efforts, with
conclusions concentrating on work performed at the University of Pennsylvania. That work is
discussed briefly below. Meanwhile, Bronzini [3.4] traces the development of various
multimodal freight transportation models at the national level. The models described include
those developed by the Inland Navigation Systems Analysis project, the Transportation Systems
Center and the National Energy Transportation Study. An important conclusion of Bronzini’s
review is the need for a “comprehensive interregional commodity flow data base” – something
that has been under development in the intervening years by many sources and will be
discussed further in Section A.5. Another important observation by Bronzini regards the
benefits of using an equilibrium assignment for nonhighway modes. His investigations indicate
that using an equilibrium assignment tends to redistribute traffic to cause more efficient use of a
particular modal network, rather than cause a switch of traffic to a competing mode.
Intercity Freight Models
One of the earliest intercity freight studies was published by Morton in 1969 [3.5]. In it he
applies a linear regression analysis to develop equations for national rail freight volume and
truck freight volume based on GNP, rail shipping rates and truck rates. Apart from nationallevel
models, however, little appears to have been written regarding workable intercity freight
transportation models until the late 1970s. In 1977 Jones and Sharp [3.6] published research
on demand modeling for undeveloped rural regions of the United States. Their model explicitly
did not have any predictive function, but it did recognize that:
In underdeveloped regions… the patterns of economic development
cannot be predicted by past trends because past trends lead nowhere.
(p. 523)
Like others in the 1970s, Jones and Sharp noted the lack of available commodity flow data that
could make their model useful.
Throughout the 1980s researchers at the University of Pennsylvania concentrated on the
differences between freight modeling and passenger modeling [3.7, 3.8]. They formulated a
model that recognizes the special situation of shippers and carriers and attempts to account for
the different information available to each group. Their model takes OD pairs determined at the
shippers’ level and optimizes the flow for the carriers’ network, in a model structure shown in
Figure A.4. Other intercity freight modeling efforts, including Canadian research in the late
1980s and early 1990s, bears a closer resemblance to statewide modeling practice and will be
discussed in Section A.5.
The State of the Art in Statewide Travel Demand Forecasting 93
Intercity Freight ModeSplit Models
In addition to the models that address the full extent of the freight transportation process,
several address only modesplit. Some attempts have been made to examine statewide freight
flows for single modes; recent studies have included air freight [3.9] and specialuse trucks
[3.10]. However, a primary concern of freight modeling is the division of the freight flow
between competing modes, usually between truck and rail. A clear example of the importance
of modesplit models for intercity freight transportation is provided in Lindesmeyer’s paper [3.11]
about the drastic effects on rural freight trucking in Nebraska that were brought about by
changes rail freight practices.
A number of methods have been employed in the effort to understand how freight traffic
becomes divided among the available modes. The methods used have included discriminant
analysis [3.12] and a diversion matrix method [3.13], as well as the probabilistic methods more
familiar to urban modesplit modelers [3.14, 3.15, 3.16]. Of these studies, the most interesting
is the diversion matrix study. Although it was written before the recent explosion of truck/rail
intermodal business and is based on an uncomplicated analysis, the study suggests that only
about one quarter of total manufacturedgoods cargo is really subject to competition between
modes. The choice of modes for the vast majority of manufacturedgoods, the study concludes,
is instead determined by the weight of the shipment and the length of its haul, or by other
factors, such as shipper prejudices toward one mode or another.
Consumption
Amounts
SHIPPERS' SUBMODEL

User Optimized
Elastic Demand Assumed
Aggregate Network Used
CARRIERS SUBMODEL

System Optimized
Fixed Demand Assumed
Detailed Network Used
Production
Amounts
Demand
Function
OD Demands by Path
from Production Site
to Consumption Site
DECOMPOSITION ALGORITHM

Path Construction
Path Decomposition
OD Demands
by Carriert
Arc and Path Flows Arc and Path Costs
Figure A.4. University of Pennsylvania ShipperCarrier Model
Guidebook on Statewide Travel Forecasting 94
Discussion
Research on intercity freight transportation appears to have been minimal, apart from the
specifically statewide research that will be discussed in the Section A.5. Research into shipper
carrier problems clearly addresses an important behavioral link in the freight transportation
process, but there may be other, easier ways to address this complicated interface. One way
might be to employ expert panels (as will be discussed in Section A.7) that include shipper and
carrier representatives to examine traffic forecasts from their individual perspectives.
A.4. Statewide Passenger Forecasting Literature
In spite of the amount of research involving the characteristics of intercity travel and its
concentrations on econometric models and probabilitybased models, passenger travel
forecasting, as practiced by the various state DOTs, has remained much more basic. In most of
the states contacted as part of the research for this appendix (see Table A.1) no travel modeling
is done on a statewide level. At the majority of state DOTs, forecasting is done for specific
projects only, and forecasts are made based on historic trends, rather than on some formal
model.
For the states that are engaged in some type of modeling process, the models used are all
“fourstep” models, with a modeling procedure borrowed almost wholly from the urban
transportation planning (UTP) process. This is likely a function of the ready availability of urban
modeling software and personnel trained to use it. As early as 1967 Arizona and Illinois had
developed UTPstyle models [4.1], and by 1972 at least 19 different states were using or
preparing statewide models [4.2]. Modeling activities were evidently so popular that in 1973 the
Federal Highway Administration (FHWA) perceived the need to standardize the thinking about
statewide modeling and issued a guidebook on the subject [4.3] – effectively institutionalizing
the UTPstyle model for statewide use. The enthusiasm for developing statewide models that
was present in the late 1960s and early 1970s soon waned, however, whether due to funding
cuts or to frustration with the model results, and little activity seems to have taken place (studies
in Florida and Kansas are an exception [4.4, 4.5, 4.6]) until very recently. Apparently, only
Connecticut, Kentucky and Michigan have been continuously developing models from the
earlier period.
By the early 1990s, prompted by new federal legislation (CAAA and ISTEA), several states were
rethinking their strategies. New Mexico [4.7] and Texas [4.8] produced interesting reports that
outline this renewed focus on statewide modeling. The New Mexico report addresses both
passenger and goods movement models within the broader context of statewide transportation
planning. The Texas report, which includes reviews of circa1990 models from Florida,
Kentucky and Michigan, concentrates more on the details of statewide modeling, especially the
difficulties in isolating interzonal trips and the proliferation of “Kfactors” in recent models.
Despite this promising trend, neither New Mexico nor Texas is currently involved in statewide
modeling. (Texas is, however, scheduled to issue a request for proposals (RFP) for a model
development contract in the Fall of 1997). A list of states contacted that sent information about
their current passenger modeling efforts is presented in Table A.2, and these are discussed
below.
Data Collection for Passenger Travel
Ideally, travel forecasts are based on some sort of travel data. One obvious source of travel
data is the survey. Surveys have been conducted at the statewide level since the earliest days
of highway modeling [4.9] and continue to be conducted at the statewide level [4.10, 4.11].
The State of the Art in Statewide Travel Demand Forecasting 95
However, they are relatively expensive to conduct and must be supplemented by other data.
Two other options make use of data that is already available: Federal survey data and
statewide traffic counts. US census data have always been valuable as inputs to travel
modeling. The 1990 census improved upon this by including a journeytowork (JTW) survey
and by introducing the Census Transportation Planning Package (CTPP) [4.12]. The JTW has
proved especially useful in estimating homebased work trips on a statewide level, but has been
criticized for its lack of information about other purposes [4.13]. The CTPP provides
transportationrelated information at a transportation analysis zone (TAZ) level, which can be
readily aggregated into township or county level data for statewide modeling. Another federal
data source is provided by the US Department of Transportation, which conducted its most
recent National Personal Transportation Survey (NPTS) in 1995. The NPTS data, which
measures some intercity travel, have been used in the development of a number of statewide
models. In addition to the aforementioned federal government sources, it should also be noted
that estimated and forecasted data is also available from a wide variety of state, academic and
commercial sources.
Of course, for many years state DOTs have had in place systems of traffic counting equipment
operating at a statewide scale. Research in the early 1980s [4.14, 4.15, 4.16] developed
statistical methods of clustering together traffic counts on different roads based on their similar
functional and geographical characteristics. In association with the introduction of the FHWA’s
Traffic Monitoring Guide in 1985 [4.17], Pennsylvania [4.18], Washington [4.19, 4.20, 4.21] and
New Mexico [4.22, 4.23] began to reevaluate their traffic monitoring systems to take advantage
of clustering. The result is a larger and more statistically valid collection of traffic count data
available for use in travel forecasting.
Data Synthesis for Passenger Travel
Even with advanced systems for traffic data collection, it is difficult for a state DOT to collect
enough data to account for all of the likely paths between OD pairs being examined. To get
around this difficulty, optimization methods have been developed to synthesize trip tables from
available traffic count information [4.24, 4.25, 4.26]. These methods have subsequently been
applied to statewide analyses in Wyoming [4.27, 4.28]. Attempts have also been made to
synthesize trip tables from census data at a substate level in New Jersey [4.29].
Trend Analyses of Passenger Travel
As noted above, many of the DOT officials contacted for this appendix indicated that the only
forecasts they make are not based on models, but are instead based on the extrapolation of
trends observed in historical data. The Minnesota DOT has formalized this process as it applies
to forecasting traffic for their state trunk highways [4.30], but such documentation seems to be
the exception. Some indication of the possibilities of trendline analysis is given in a paper by
Harmatuck [4.31] for the Wisconsin DOT. In it he provides further insight into the particular
ways of dealing with traffic data as a time series. In addition, at least one state contacted for
this appendix indicated that a growth factor method, similar to the method outlined for updating
coverage counts in the FHWA’s 1992 Traffic Monitoring Guide [4.32], is used for forecasting
purposes. Otherwise little information is available on travel forecasting techniques in the
absence of a statewide model.
Statewide Models of Passenger Travel
Of the states contacted as part of the research for this appendix, those having ongoing
modeling efforts sent documentation of their progress. A summary of the passenger models in
Guidebook on Statewide Travel Forecasting 96
existence or underdevelopment is presented in Table A.2. This includes work done in
Connecticut [4.33, 4.34], Florida [4.35, 4.36], Indiana [4.37], Kentucky [4.38], Michigan [4.39],
New Hampshire [4.40], New Jersey [4.41, 4.42], Vermont [4.43], Wisconsin [4.44] and Wyoming
[4.27, 4.28]. In addition to the states shown in Table A.2, California has a statewide model, but
it is being redesigned, so documentation is currently unavailable for it. Oregon is also in the
early stages of developing a comprehensive forecasting model that will include a land use
element [4.45]. Several other states are currently in the very beginning stages of modeling
projects – issuing RFPs to interested consultants.
As can be seen from Table A.2, most of the models consider a large number of trip types (as
many as 5 or 6), but only a few modes. All of the models are of the “fourstep” style. All use
fairly standard UTP procedures, except for the model under development for New Hampshire.
New Hampshire proposes to use logit formulations for trip generation and distribution. The
Wisconsin model is unique in that it is essentially an intercounty model, with comparatively few
TAZs. The Florida and New Jersey models are also interesting in the degree to which they
have attempted to incorporate existing MPO models into the statewide modeling effort. The
Kentucky and Michigan models are two of the more recent useable models from states with long
histories of model development and are representative of the current state of the practice.
Section A.6 examines both of these models in greater detail. Recreational Travel Models
As early as 1963 recreational trips were considered an important enough purpose to warrant
separate study [4.46]. In fact, in the late 1960s and early 1970s the NCHRP [4.47], Indiana
[4.48, 4.49], Kentucky [4.50, 4.51] and other states [4.52, 4.53] conducted studies of the special
characteristics of recreational travel. Strangely, although Americans seem to have dedicated an
increasing amount of time to pursuing recreational activities, the last of these studies was
published more than twenty years ago. Since many state economies depend heavily upon
recreational activities, it would seem that this trip type might be important enough to require a
closer examination than it has received in the past two decades.
Discussion
Using trendline procedures in statewide forecasting is probably better than not forecasting at all,
especially for short term planning horizons, where large variations from recent trends are less
likely. The use of travel forecasting models, however, grounds the forecast in the underlying
statewide and national socioeconomic trends. Although these socioeconomic trends are
themselves forecasts, it is hoped that they broaden the basis of the transportation model
sufficiently to provide a more reasonable forecast of future travel. Further discussion of the
structure of typical statewide passenger models is presented in Section A.6.
The State of the Art in Statewide Travel Demand Forecasting 97
Table A.2. Current Statewide Passenger Models
STATE TAZs MODES PURPOSES COMMENTS
Connecticut 1300 total 1. SOV
2. HOV
3. Bus
4. Rail
1. HBW
2. HBNW
3. NHB
• Mode split based on LOS
information.
• Iterativeequilibrium assignment
for highways.
Florida 440
internal
32
external
Highway
vehicles
only
1. HBW
2. HB Shop
3. HB Soc./Rec.
4. HB Misc.
5. NHB
6. Truck/Taxi
• All trips are modeled to maximize
use of MPO models.
• Gravity friction factors based on
MPO urban models.
• Mode split is auto occupancy only
based on production zone.
• Extensive use of Kfactors.
Indiana 500
internal
5060
external
1. Auto
2. Truck
3. Transit
1. HBW
2. Other
Business
3. HB other
4. NHB
5. Recreational
6. Truck
• Under development.
• Internal TAZs at the township
level.
• Aggregate mode choice.
Kentucky 756
internal
706
external
Auto only 1. HBW
2. HBO
3. NHB
• Model includes a large portion of
surrounding states.
• NPTS national average data used
for trip generation
Michigan 2392 total Auto only 1. HB Work/Biz.
2. HB
Soc./Rec./Vac.
3. HB Other
4. NHB
Work/Biz.
5. NHB Other
• All trips modeled – previous
models did not consider local
trips.
• Two possible mode split models:
(1) simple crossclassification and
(2) LOSbased.
• LOSbased mode split model still
under development.
• NTPS data used for calibration;
CTPP data used for validation.
• Extensive use of Kfactors.
New
Hampshire
1 per
5000 pop.
1. SOV
2. HOV2
3. HOV3+
4. Bus
5. 5. Rail
1. HBW
2. Business
related
3. Personal
4. Shopping
5. Recreational
6. Other
• Under development.
• Logit trip generation and
distribution.
• Time of day and seasonal factors.
New Jersey 2762
internal
51
external
  • Model created by merging 5 MPO
models.
Vermont 622
internal
70
external
Highway
vehicles
only
1. HBW
2. HB Shop
3. HB School
4. HB Other
5. NHB
6. Truck
• Based on extensive statewide
survey.
Guidebook on Statewide Travel Forecasting 98
Table A.2. Current Statewide Passenger Models (continued)
STATE TAZs MODES PURPOSES COMMENTS
Wisconsin 112
internal
45
external
1. Auto
2. Air
3. Rail
4. Bus
1. Business
2. Other
• Under development.
• No external trips considered.
• Network used only to develop
impedances for mode share
calculations.
Wyoming 5 internal
5 external
1. Auto
2. Truck
 • Model created mostly to
demonstrate techniques.
• Summer weekend travel is
modeled.
• Full trip tables estimated using
entropy maximization technique
A.5. Statewide Freight Forecasting Literature
Introduction
For various reasons, it has been suggested that forecasting freight transportation flows is more
complex than modeling passenger travel volumes [5.1]. This is partly because of the numerous
parties involved in shipping the large variety of commodities that are regularly moved by the
several modes available. The development of freight forecasting techniques, therefore, has
historically lagged behind the development of passenger techniques. At the same time, the
methods of analyzing freight traffic at a statewide level have remained similar in form to those
used in predicting passenger travel. There are essentially two ways that state DOTs forecast
freight traffic: (1) by analyzing truck traffic or (2) by using a commodity flow model [5.2].
There are two techniques generally applied to truck traffic analysis. The first technique is a
simple trendline analysis similar to that described in Section A.4 for passenger travel
forecasting. The other technique is to include truck trips as a separate trip purpose in the
passenger model, based on survey data or on counts of truck traffic on various links in the
highway network. In either case the similarity to passenger forecasting is obvious.
Forecasts that are based on commodity flows bear a resemblance to passenger models in the
way they are structured. They are typically employed in a “fourstep” sequential process that
employs a gravitymodel distribution, a cursory modesplit step and some sort of simple
assignment. The only significant difference is that the trip generation step is often based on
freight flow data (usually classified by industry groups), instead of regression equations for
employment and population, as with passenger models.
Data for Freight Forecasting
Freight data, especially for truck analysis techniques, can be collected by survey methods, as
has been done recently by the Washington DOT [5.3], but increasingly models are using
commodity flow data as their basis. NCHRP released two reports in the late 1970s [5.4, 5.5]
that began to address the data requirements of statewide freight modeling. These two reports
present 228 different sources of data that could be used for freight forecasting. More recently,
NCHRP Report 388 [5.1] has provided an update to the list of data sources. Meanwhile, the
Bureau of the Census’s 1993 Commodity Flow Survey has been used by several states to
develop their own commodity flow interactions. A number of private firms also offer (for a fee)
access to their collections of historical and forecast data, not only for population and general
The State of the Art in Statewide Travel Demand Forecasting 99
employment data, but for employment and commodity flows by industry. Reebie’s
TRANSEARCH database has been a popular source, but many others are available [5.1] or are
under development. Much of the currently available data is unfortunately provided only at a
Bureau of Economic Analysis (BEA) region level, which is generally too big for statewide
analysis. The data must therefore be disaggregated to at least a county level before use.
Some databases under development by the FHWA (with Reebie and Colography, another
private firm, for instance) are aimed at directly providing countylevel data.
Statewide Models for Freight Forecasting
Likely influenced by deregulation in the railroad industry, there seems to have been a flurry of
interest in statewide freight forecasting in the early 1980s. At that time NCHRP Report 260 [5.6]
and papers in Transportation Research Record 889 [5.7, 5.8, 5.9] presented organized methods
of modeling freight traffic at a statewide level. The modeling techniques described are of two
types: (1) calculation of growth factors (by commodity) to be applied to existing traffic patterns
[5.8] or (2) use of forecasts of future freight flows (by commodity) distributed by a gravity model.
Aside from isolated examinations of specific issues [5.10], most proposals for statewide freight
modeling have closely followed the pattern of these early 1980s papers.
Documentation for a few statewide freight models was obtained as part of the research for this
appendix, and their features are summarized in Table A.3. Of these models, only Louisiana’s
[5.11] is of the growth factor type noted above. It is a relatively simple example, but it is also the
most explicitly multimodal model provided. The report of a recent FHWA project [5.12] offers
recommendations on how to apply similar growth factor methods with increasing degrees of
sophistication. The Indiana [5.13] and the Wisconsin [5.14] models are both of the commodity
forecast type noted above, and the comparatively coarse structure of their models – an order of
magnitude fewer TAZs than for a typical passenger model – is evidence of the reduced ability to
disaggregate freight data to a lessthancounty level. Both the Indiana and Wisconsin models
are reviewed in greater detail in Section A.7. The Michigan [5.15] and New Jersey [5.16]
models are, as can be seen from Table A.3, truck models and share an identical network
structure with their respective statewide passenger models (see Table A.2).
Related Models
Some work done in the late 1980s with models for Alberta and Brazil by Canadian researchers
is also applicable to statewide freight modeling in the United States. In papers describing their
model for the Province of Alberta, Ashtakala and Murthy [5.17, 5.18] present a sequential model
with a gravitystyle trip distribution based on commodity flows. Its useful features include a Box
Cox procedure to calibrate the friction factors for the gravity model and a “Commodity Haul
Frequency Diagram” to visually assist in calibration of the gravity model. Their related work with
a logit modesplit model has already been noted in Section A.3.
Another notable effort from Canada is the STAN system developed by the Centre de Recherche
sur le Transports at the Universite of Montreal [5.19, 5.20, 5.21]. The STAN model was
developed from the EMME/2 passenger modeling set of programs and includes a sophisticated
“multimode multiproduct” model [5.21], which assigns traffic to the various modes and links
according to the solution to an optimization problem. Application of STAN was made to a large
region of Brazil, and ample evidence is provided of the graphical output made available by
STAN.
Guidebook on Statewide Travel Forecasting 100
Table A.3: Current Statewide Freight Models
STATE TAZs MODES PURPOSES COMMENTS
Indiana 145 total 1. Truck
2. Rail
21 Commodity
Groups
• Based on 1993 Commodity
Flow Survey.
• Distribution by fullyconstrained
gravity model.
Louisiana  1. Truck
2. Rail
3. Water
4. Air
11 Commodity
Groups
• Based on commodityspecific
growth factors applied to
existing traffic volumes.
Michigan 2392 total Truck only 11 Commodity
Groups
• Trip generation regressed by
commodity using (1)
employment and (2) tons
shipped.
• Michigan truck survey, BEA
commodity flows and US
Canada trade flows (FHWA)
used.
New Jersey 2762 internal
51 external
Truck only  • Based on fourdigit STCC code
commodity flow analysis.
Wisconsin 106 internal
34 external
1. Truck
2. Rail
39 Commodity
Groups
• Internal TAZs at a county level.
• Consideration of railtotruck
intermodal diversion scenarios.
Discussion
Statewide models for freight transportation appear to be structured with a strong sense of the
limitations inherent in forecasting at a large geographical scale. This is reflected in their
documentation which is generally more straightforward than that for passenger modeling. This
may be because freight forecasting models have developed independently enough from the
fourstep urban modeling process that they can begin to address problems (such as data
availability) that are unique to their statewide nature. Further discussion of the characteristics of
statewide freight forecasting models is presented in Section A.7.
A.6. Two Recent Passenger Models
This section provides a closer examination of the recent statewide passenger modeling efforts
for the states of Michigan and Kentucky. As discussed in Section A.4, both Michigan and
Kentucky have comparatively long histories of statewide modeling, and both have produced
documentation describing their most recent modeling efforts. This documentation provides
some indication of the difficulties involved in producing a workable statewide model and the
significant number of assumptions and adjustments that often must be made in the statewide
modeling process.
The Michigan Passenger Model
The Michigan model [6.1] is a traditional statewide model, in the sense that it retains the
characteristics of an urban model (many TAZs, many purposes), but on a much larger scale.
The Michigan model is a “fourstep” model that, as can be seen in Table A.2, forecasts travel for
five trip purposes between 2392 TAZs. A graphical depiction of the Michigan network can be
seen in Figures A.5, A.6 and A.7. Most of the TAZs appear to be at the townshiplevel, with
smaller TAZs in large urban areas and larger TAZs in rural areas and the Upper Peninsula. The
The State of the Art in Statewide Travel Demand Forecasting 101
vast majority (2307) of the TAZs are within Michigan while the remaining 85 represent the other
47 contiguous United States, Canada and Mexico. In addition, the model includes thousands of
special generator sites, divided into 10 general categories by type of facility (airports, tourist
attractions, campgrounds, state parks, golf courses, marinas, motels, hospitals, shopping
centers and colleges).
Michigan’s modeling procedure starts by using socioeconomic data developed at a county level
from a Regional Economic Model, Inc. (REMI) model. This information is then disaggregated to
a TAZlevel using data from the Michigan Employment Security Commission. Census data from
the Public Use Microdata Sample (PUMS) and the CTPP are then used to develop cross
classification tables based on five different household sizes (1, 2, 3, 4 and 5+ persons per
household) and three income groups (low, medium and high) for a total of 15 categories. It
should be noted that the PUMS data is available at the Public Use Microdata Area (PUMA)
level. Since there are 67 PUMAs in Michigan, a PUMA is essentially equivalent to a county in
its level of aggregation. For forecasts of future travel activities, REMIgenerated countylevel
growth factors are applied to all TAZs in a particular county.
The trip production step for internal trips involves the use of equations for trip production from
the asyetunpublished update to NCHRP Report 187. These production rates are distributed to
the five trip purposes according to proportions based on NPTS data and are applied to the 15
crossclassification categories noted above, further classified by five geographical categories
(four for different city sizes and one for rural areas). Figure A.8 shows some example
calculations for the production rates that are used. Equations for internal trip attractions are
based on an evaluation of alternative rates generated from NPTS data, metropolitan area
studies in Michigan and available data from the San Francisco area. Table A.4 shows the final
attraction rates. Total productions and attractions are then the result of entering TAZlevel
household socioeconomic data into these equations.
The numerous special generator sites are evaluated as having attraction value only. The
Institute of Transportation Engineers’ Trip Generation was used (along with local surveys) to
develop the attraction equations. Table A.5 shows the information used in developing these
attraction rates. The trips attracted to the special generator sites are only for the homebased
social/recreational and homebased other purposes, since a preliminary analysis revealed that
attractions for the other purposes were inconsequential. For trips with ends outside of Michigan,
production and attraction equations were developed using NPTS data for trips greater than 75
miles in length. Trips from any state not represented in the NPTS data are estimated as a
function of the state’s area.
Trip distribution is accomplished using a gravity model. Friction factors for the gravity model are
calculated using a gamma function of the generalized cost of travel (see Figure A.9). The
gamma function was chosen to provide maximum flexibility in accounting for both very short and
very long trips that are possible in a statewide model. The gravity model was calibrated using
NPTS data and validated using CTPP and traffic count data. Three types of “geographic
adjustment” factors (Kfactors) are also used in calibrating the gravity model. These Kfactors
are based on information regarding existing (1) countytocounty, (2) major citytocity and (2)
outstate/instate traffic flows and the values of the Kfactors listed in the documentation range
from as low as 0.10 to as high as 9.67. Unlike previous statewide models for Michigan, this
most recent model includes intrazonal trips. The impedances for intrazonal trips are calculated
separately from the network, however, to avoid complications due to the arbitrary placement of
TAZ centroids with respect to highway links.
Guidebook on Statewide Travel Forecasting 102
Figure A.5. Instate TAZs from Michigan Model
Figure A.6. Instate Highway Network from Michigan Model
The State of the Art in Statewide Travel Demand Forecasting 103
Figure A.7. Outstate Highway Network for Michigan Model
Table A.4a. Attraction Rates from Michigan Model, Urban Areas
Employment Category
Purpose Total Retail Whole
sale
Service Mfg. Other House
holds
HBW 1.486      
HB Rec  1.300  0.260   0.522
HB Other  6.360 2.650 1.802  0.530 
NHB Work  0.797  0.232 0.097  
NHB Other  4.123  1.207  0.583 0.350
Table A.4b. Attraction Rates from Michigan Model, Rural Areas
Employment Category
Purpose Total Retail Wholes
ale
Service Mfg. Other House
holds
HBW 1.486      
HB Rec  0.522  0.087   0.522
HB Other  10.07 2.650 1.802  0.530 
NHB Work  0.728  0.212 0.088  
NHB Other  3.748  1.097  0.530 0.318
Guidebook on Statewide Travel Forecasting 104
First, using the NCHRP 187 update, obtain generation rates for the various cross
classification categories. For instance, in a large urban area (population greater than 1
million), a lowincome, singleoccupant household will produce 3.7 trips per day.
Second, from the NPTS, apportion the trips produced to the Michigan model’s purposes
according fixed ratios. Again, for large urban areas the following ratios are used for low
income, singleoccupant households:
0.192 to homebased work (HBW),
0.160 to homebased recreational,
0.404 to homebased other,
0.310 to nonhome based work and
0.214 to nonhome based other.
1.00 (100% of trips accounted for)
Next, multiply these factors. For instance, (3.7 x 0.192) = 0.71 HBW trips from each low
income, single occupant household. Similarly, (3.7 x 0.160) = 0.59 homebased
recreational trips, etc.
Finally, repeat the process all combinations of urban (and rural) areas, income groups,
household sizes and trip purposes.
Figure A.8. Sample Production Rates Calculations from Michigan Model
The modesplit step is really a vehicle occupancy step, since a more sophisticated intermodal
model has proved difficult to develop. The vehicle occupancy step consists of three substeps.
First, a countylevel transit share is developed for work trips based on average CTPP shares
and extended to other purposes by using work/nonwork ratios from the NPTS. The resulting
shares are applied to every TAZ in the county. Second, person trip and vehicle trip data from
the NPTS are used to develop average occupancy rates by trip purpose for each of the 15
crossclassification categories noted earlier. Representative minimum and maximum values are
shown in Table A.6. Finally, adjustment factors are developed to account for trip length, based
again on NPTS data. A graph of the resulting factors is shown in Figure A.10. After applying
these occupancy factors, the trips are then assigned to the network using an allornothing
procedure.
It should be noted that the model documentation describes experiments made during model
development with proprietary software for synthesizing trip tables based on singlestation origin
anddestination surveys and for employing a “stochastic user equilibrium” algorithm for trip
assignment. Both experiments proved unsatisfactory and are not included in normal operation
of the model.
The State of the Art in Statewide Travel Demand Forecasting 105
Table A.5. Special Generator Attractions from Michigan Model
Special Generator
Type
Source of Rate Rate Equation
Airports 1991 ITE Manual exp[1.368 x ln(reg. aircraft)  0.347]
or
[104.73 x operations /365]
Tourist Attractions MDOT Travel &
Tourism
2 x attendance
Campgrounds 1991 ITE Manual 0.79 x campsites
State Parks 1991 ITE Manual 0.50 x acres
Golf Courses 1991 ITE Manual 37.59 x holes
Marinas 1991 ITE Manual (1.891 x berths) + 410.795
Motels 1991 ITE Manual exp[0.713 x ln(0.44 x rooms) + 3.945]
Hospitals 1991 ITE Manual exp[0.634 x ln(beds) + 4.628 ]
Shopping Centers 1991 ITE Manual exp[A x ln(ksf) + B]
where
A = 0.756, B = 5.154 for > 570 ksf and A =
0.625, B = 5.985 otherwise
Colleges &
Universities
1987 and 1991 ITE
Manuals
2.37 x students (for universities)
or
1.55 x students (for community colleges)
Gamma function formulation for friction factors:
f = a T
b
x exp(c T)
where:
f = friction factor
T = travel time (or generalized cost of travel)
a, b and c = constants of calibration
Figure A.9. Gamma Function for Friction Factors from Michigan Model
Table A.6. Maximum and Minimum Occupancy Rates from Michigan Model
Trip Purposes
Occupancy Rates HBW HB Rec HB Other NHB
Work
NHB
Other
Minimum 1.07 1.20 1.13 1.12 1.22
Maximum 1.13 1.76 1.57 1.19 1.55
Guidebook on Statewide Travel Forecasting 106
Figure A.10. Trip Length Occupancy Factors from Michigan Model
The Kentucky Passenger Model
The most recent Kentucky model [6.2] forecasts travel for three trip purposes (plus trucks)
among nearly 1500 TAZs, almost half of which are outside of the state. The large number of
TAZs outside of the state borders is a distinctive feature of the Kentucky model. The model’s
network actually extends to include almost all of Tennessee, more than half of West Virginia,
Ohio and Indiana, as well as significant proportions of the other surrounding states. The TAZs
inside the state consist of individual census tracts in rural areas and groups of census tracts in
urban areas. In addition to the censusbased TAZs, special generator TAZs are included at 40
significant recreational areas and military bases and 29 external station TAZs are included to
model travel into the network from the remainder of the United States (mostly via the interstate
highway system). A graphical depiction of the Kentucky network is shown in Figure A.11.
Except for special generator and external station TAZs, the trip generation step uses projections
of population and employment provided by a commercial source, Woods & Poole. Trip
generation rates for homebased work (HBW) trips are developed in a spreadsheet that
calculates county level trip generation rates by multiplying the national NTPS production rates
(for large urban areas and for other areas) by a ratio of countyspecific to national production
and attraction rates based on JTW data. The national rates used from the 1990 NPTS are
shown in Table A.7. The Kentucky model’s documentation indicates that the JTW rates vary
between 0.246 and 0.435 trips produced per person and between 0.532 and 0.744 trips
attracted per employee, depending upon the county examined. Total HBW attractions are then
adjusted to match productions.
The State of the Art in Statewide Travel Demand Forecasting 107
Figure A.11. Network Structure from Kentucky Model
Table A.7. NPTS Trip Rates Used in Kentucky Model
NPTS Trip Purpose Rate for Large Urban Areas
(trips per person)
Rate for Other Areas
(trips per person)
HB Work 0.5390 0.5236
HB Business 0.0364 0.0414
HB Shop 0.3261 0.3346
HB SocialRec 0.4420 0.4282
HB Other 0.5969 0.6313
Non Home Based 0.6041 0.6122
Total 2.5445 2.5714
Less guidance is available from census sources regarding nonwork trips, which in the case of
the Kentucky model include homebased other (HBO) and nonhome based (NHB) purposes.
Instead, nonwork trips are divided into three categories: (1) “short” trips – those less than 60
minutes in length; (2) “long” trips – those longer than 60 minutes in length; and (3) through trips.
Starting from nationallevel NPTS generation rates, it is assumed that 99% of nonwork trips are
short trips and that 70% of all nonwork productions are HBO. Attraction rates are taken from
NCHRP Report 187, assuming that retail employment in all areas is 20% of total employment.
Again, for nonwork trips, attractions are adjusted to match productions.
Recreational special generator TAZs are classified as either “local” or “national”, with the
assumption that 67% of the nonwork trips generated are “short” trips and 33% are “long” trips
for “local” areas and vice versa for “national” recreation areas. For military special generators
only work trips are considered, and it is assumed that 20% of the work trips generated are
Guidebook on Statewide Travel Forecasting 108
intrazonal trips and are not assigned to the network. A Fratar model is used to predict nonwork
trips passing through the network. There is no modesplit or autooccupancy step, and traffic
assignment is evidently done using an allornothing technique, since no other details are
provided in the documentation.
The trip distribution step is accomplished using a gravity model. However, limitations in
Kentucky’s software used to calibrate the friction factors, require that HBW trips over 120
minutes in length be eliminated from the process. Kentucky judged this to have a beneficial
effect on model performance, since a few very long trips can have a disproportionately large
effect on the average trip length calculated for calibration. Friction factors for HBW trips are
developed from JTW survey information, but intrazonal travel times are deliberately kept small
(a maximum of 15 minutes) in order to keep intrazonal trips off the statewide network. For
nonwork trips, friction factors are developed from the previously noted unpublished update to
NCHRP Report 187, but friction factors for “long” trips are synthesized to give a “reasonable”
trip frequency distribution. A summary of the various friction factors in their final form is
presented in Figures A.8 and A.9. Kfactors are also used at a county level to adjust the
distribution of HBW trips, but the documentation does not indicate a range of values used. The
friction factors and Kfactors are subsequently adjusted to provide a good fit with known
screenline counts and average trip length frequency data from the JTW survey.
A further procedure described as “trip table calibration” is mentioned in the documentation, but
described only as “similar to the Fratar process”. An appendix to the documentation further
describes it as an iterative adjustment procedure where the modeled volumes on network links
are matched to corresponding traffic counts on the roads they represent. Of course, the end
result is a better fit between the model and the base year traffic. Unfortunately, since
accessibility effects are not considered in this data synthesis procedure, the gravity model
cannot be used to forecast future volumes when the “trip table calibration” procedure is used.
The Kentucky model’s documentation instead recommends that if accessibility changes are
expected in the future, the following threepart procedure be followed: (1) a “fratared” forecast
should be prepared using the “trip table calibration” results, (2) a gravity model forecast should
be produced that includes accessibility changes and (3) a final forecast should be manually
generated to resolve any inconsistencies.
Discussion
As noted above, one similarity between the Michigan and Kentucky passenger models is their
fine geographical level of detail in comparison with statewide freight models and with several
other statewide passenger models (see Tables A.2 and A.3). The 756 instate TAZs for
Kentucky are, on average, twice as big as the 2307 instate TAZs for Michigan (52 sq. mi. per
TAZ versus 25 sq. mi. per TAZ, respectively), but are relatively close in terms of average
population per TAZ. Based on 1990 population figures, the Kentucky TAZs average slightly
more than 5000 residents each, while the Michigan TAZs average 4100 residents each.
Selection of the “grainsize” for the model should be made considering its intended use. If
forecasts of general trends are being sought, perhaps fewer, larger TAZs will suffice. Similarly,
if projectspecific planning capabilities are desired using the statewide model, a more “fine
grain” model may be useful. In either case it would appear that using a large number of TAZs
precludes the use of any but the most basic data from any individual TAZ. The likelihood that
meaningful adjustments could be made to the model based on detailed knowledge of activities
in any particular TAZ is therefore diminished. It should also be noted that many calculations
(involving generation rates, growth factors, Kfactors, transit shares, etc.) are already made
using only countylevel data in both of the models examined.
The State of the Art in Statewide Travel Demand Forecasting 109
Table A.8. HBW Friction Factors from Kentucky Model
Travel
Time
(min.)
Friction
Factor
Travel
Time
(min.)
Friction
Factor
Travel
Time
(min.)
Friction
Factor
Travel
Time
(min.)
Friction
Factor
1 259,360 31 4675 61 257 91 24
2 203,007 32 4177 62 236 92 22
3 159,242 33 3737 63 218 93 20
4 125,178 34 3348 64 200 94 19
5 98,609 35 3003 65 184 95 17
6 77,840 36 2697 66 170 96 16
7 61,572 37 2426 67 157 97 15
8 61,572 38 2184 68 144 98 13
9 61,572 39 1969 69 133 99 12
10 61,572 40 1777 70 123 100 11
11 61,572 41 1605 71 114 101 11
12 53,244 42 1452 72 105 102 10
13 46,129 43 1315 73 97 103 9
14 40,040 44 1192 74 90 104 8
15 34,818 45 1082 75 83 105 8
16 30,333 46 982 76 76 106 7
17 26,472 47 893 77 71 107 6
18 23,143 48 813 78 65 108 6
19 20,268 49 740 79 60 109 5
20 17,780 50 675 80 56 110 5
21 15,624 51 616 81 52 111 4
22 13,752 52 563 82 48 112 4
23 12,123 53 514 83 44 113 4
24 10,705 54 470 84 41 114 3
25 9468 55 431 85 38 115 3
26 8386 56 394 86 35 116 3
27 7439 57 362 87 32 117 3
28 6609 58 332 88 30 118 2
29 5880 59 304 89 28 119 2
30 5239 60 280 90 26 120 2
The most immediately visible difference between the Michigan and Kentucky models is their
treatment of the geographical areas outside of their respective state boundaries. The Michigan
model’s roadway network is built with connections to outstate areas at relatively few discrete
points (see Figure A.6). In contrast, the Kentucky model’s highway network extends more than
200 miles into the surrounding states (see Figure A.11). While limiting border crossings to a few
very select locations may be acceptable for a peninsular state like Michigan, it is a bit more
troublesome for a landlocked state like Kentucky. It should also be noted that, since the
Kentucky model does not include a national network (as Michigan’s does), it must include
external stations to generate travel between the network and external areas. The cushioning
effect of the 200mile wide outstate network makes the precise trip generation values for the
external stations less important. The Kentucky modeling strategy, to include TAZs and network
links well outside its state boundaries, poses a similar problem to that posed by using a large
number of instate TAZs. That problem is the need to know detailed information about hundreds
of TAZs and road segments outside of Kentucky.
Guidebook on Statewide Travel Forecasting 110
Table A.9. Nonwork Friction Factors from Kentucky Model
“Short” NonWork Trips “Long” NonWork Trips
Travel
Time
(min.)
Friction
Factor
Travel
Time
(min.)
Friction
Factor
Travel
Time
(min.)
Friction
Factor
1 999,999 31 639 1 1
2 367,878 32 557 60 1
3 196,955 33 486 61 200
4 122,954 34 425 181 50
5 83,497 35 372 241 6
6 59,802 36 326 301 3
7 44,435 37 285 361 1
8 33,916 38 250 421 1
9 26,425 39 220 481 1
10 20,925 40 193 541 1
11 16,788 41 170 601 1
12 13,616 42 150 661 1
13 11,144 43 132 720 1
14 9192 44 116
15 7632 45 103
16 6374 46 91
17 5351 47 80
18 4512 48 71
19 3821 49 63
20 3247 50 56
21 2768 51 49
22 2367 52 44
23 2029 53 39
24 1744 54 34
25 1503 55 30
26 1297 56 27
27 1122 57 24
28 972 58 21
29 844 59 19
30 734 60 17
Travel data used to construct both the Michigan and Kentucky models comes primarily from
three US government sources, specifically the NPTS, the CTPP and the JTW survey. Perhaps
the most troubling aspect of the Kentucky model is its extensive use of national average values
derived from these sources in combination with assumed ratios of travel characteristics (99% of
nonwork trips are “short”, 70% of nonwork trips are HBO, etc.) that are also based on national
averages. According to the Kentucky documentation, this dependence on national average
figures is due to the dearth of readily available data for Kentucky.
A similarly troubling feature of the Michigan model is its extensive use of Kfactors to adjust the
gravity model results. As noted above, the Michigan model makes use of three overlapping sets
of Kfactors – one set at a county level and two other sets for specific destination TAZs – to
modify the distribution of travel predicted by the model. The Kfactors appear to be produced in
a mechanical fashion, without consideration of any behavioral basis for factoring (e.g.,
reluctance to cross a state line for work or social differences between regions). The very large
and very small values of the Kfactors developed (see earlier discussion) seem to indicate that
some simplification of this process would not compromise the modeled results. The Kentucky
model also makes use of Kfactors, but in a less aggressive fashion. It is also important to note
The State of the Art in Statewide Travel Demand Forecasting 111
that Kentucky attempted to develop rationallybased Kfactors, but those efforts proved
problematic and were abandoned.
One final contrasting element of the Michigan and Kentucky models is their treatment of special
generator sites. Michigan’s lengthy list of special generator sites continues for 51 pages of
small print type, including individual hotels, shopping centers and golf courses. It would appear
that merely maintaining the database for these thousands of locations would be a significant
task in itself. It is unclear what advantage this provides for statewide modeling. Kentucky’s
approach, using only 40 of the most significant special generator sites, is more limited and
presumably more in keeping with modeling at a statewide level.
A.7. Two Recent Freight Models
As a complement to the passenger models discussed in the previous section, this section
provides a closer examination of two statewide models for freight traffic forecasting. The
models examined – Wisconsin and Indiana – represent typical examples of recent thinking in
freight forecasting. Both models are constructed to operate in a sequential fashion, similar to
passenger models, but their TAZ structures are much more coarse. As with the statewide
passenger models, these freight models show the number of simplifying assumptions must be
made to facilitate model development.
The Wisconsin Freight Model
The Wisconsin freight model [7.1] estimates the
freight traffic carrying the products of 39 important
commodity groups between 140 TAZs by four
modes: air, water, truck and rail. Of the 140 TAZs,
106 are counties. Each of Wisconsin’s 72 counties
is represented by a separate TAZ, while the
remaining 34 countylevel TAZs represent counties
in adjacent states. The other 34 TAZs are
composed of multiple BEA regions that represent
other states. Figures A.12 and A.13 show the TAZ
structure for the Wisconsin model. Table A.10 lists
the 39 commodity groups that are considered
important for Wisconsin and are used in the
modeling process.
The principal data on which the Wisconsin model is
taken from Reebie’s TRANSEARCH. The
TRANSEARCH data, which is provided at a BEA
region level of aggregation, is supplemented by
some commodity flow data specific to Wisconsin.
No generation equations for freight flow are developed,
instead the TRANSEARCH flows are simply distributed
at the county level in a four step process as follows. First, the total flows are determined from
the baseyear TRANSEARCH data for each important commodity group. Second, freight
origins are identified and are assigned to the countylevel TAZs based on county employment
data. Third, based on a national inputoutput table, it is determined which proportion of each
commodity group’s flow is destined for industrial consumption and which is destined for
household consumption. Finally, countylevel destinations are allocated based on employment
(for industrial consumption) and population (for household consumption). Factors are also
Figure A.12. Local TAZs from
Wisconsin Model
Guidebook on Statewide Travel Forecasting 112
calculated for “secondary” trucking volumes as a function of the primary freight flows. These
factors represent the additional movements for freight distribution and drayage from intermodal
yards.
Forecasts of future freight flows are
made using econometric models that
include employment forecasts obtained
under a contract with WEFA (another
private firm) and productivity forecasts
made using information from REMI.
This constitutes what is called the
“trendline” forecast. Adjustments are
made to the “trendline” forecast by
enlisting the services of various expert
panels, who add a “market driven”
element to the forecast values. Since
the forecast flows would project the
current modal shares into the future,
Wisconsin also developed an approach
which uses another panel of experts to
identify alternative railtruck modal splits
based on shipment distance and
frequency of rail service. This approach
is less formalized than that used in the
“trendline” model, but is important to
making policy decisions based on the
model results.
Once the freight flows (by weight) are determined, they are converted into an equivalent number
of vehicles. Initially weightpervehicle ratios of 100 tons per railcar and 24 tons per truck were
used for all commodities. Tonspertruck values were subsequently modified by commodity
group to the values shown in Table A.4. Daily flows are determined by dividing the total number
of vehicles by a value of 312 working days per year, based on a sixday work week. The
resulting daily flows are then assigned to the appropriate modal networks. The truck
assignment is done allornothing. For rail each shipment is designated as having a “most likely
carrier”, and the shortest path using that carrier alone is assigned the shipment. Air and water
traffic are not assigned.
The Indiana Freight Model
The TAZ structure for the Indiana freight model [7.2] is very similar to that used in the Wisconsin
model. The Indiana model predicts both truck and rail traffic volumes for a network that includes
a TAZ for each of Indiana’s 92 counties and 53 more TAZs that represent the remaining 47
contiguous states and the District of Columbia. (There are three TAZs for Ohio and two each
for Illinois, Kentucky and Michigan. All other states and the District have one TAZ each.) Both
the truck and rail networks were developed from US DOT sources. Figures A.14 and A.15
graphically depict the Indiana freight network. It should be noted that, as for the Kentucky
passenger model, the detailed roadway network for the Indiana freight model extents to about
200 miles beyond the state’s border.
Figure A.13. Outstate TAZs from Wisconsin
Model
The State of the Art in Statewide Travel Demand Forecasting 113
Table A.10. Important Commodity Groups and Traffic Densities
Wisconsin
Model Indiana Model
STCC Description
Tons per
Truck
Tons per
Truck
Tons per
Railcar
1 Farm Products 24 38 96
8 Forest Products 13  
9 Fish or Marine Products 6  
10 Metallic Ores 24  
11 Coal 24 40 100
13 Crude Petroleum, Nat. Gas, Gasoline 14  
14 Nonmetallic Ores 19 39 97
19 Ordinance or Accessories 24  
20 Food and Kindred Products 18 32 80
21 Tobacco Products 5 See note 1 See note 1
22 Textile Mill Products 5 7 18
23 Apparel or Finished Textile Products 3 4 10
24 Lumber or Wood Products 15 29 72
25 Furniture or Fixtures 3 6 15
26 Pulp, Paper or Allied Products 16 25 62
27 Printed Matter 9 See note 1 See note 1
28 Chemicals 22 35 88
29 Petroleum or Coal Products 19 26 66
30 Rubber or Misc. Plastics Products 4 See note 1 See note 1
31 Leather or Leather Products 3 See note 1 See note 1
32 Clay, Concrete, Glass or Stone Products 23 32 81
33 Primary Metal Products 19 34 86
34 Fabricated Metal Products 24 8 20
35 Machinery  Other than Electrical 9 11 28
36 Electrical Machinery, Equip., Supplies 8 7 17
37 Transportation Equipment 12 9 23
38 Instruments  Photo. or Optical Goods 5 See note 1 See note 1
39 Misc. Manufacturing Products 2 See note 1 See note 1
40 Waste or Scrap Metals 16 31 78
41 Misc. Freight Shipments 23  
42 Shipping Devices Returned Empty 4  
43 Mail and Express Traffic 3 See note 2 See note 2
44 Freight Forwarder Traffic 4  
45 Shipper Association Traffic 3  
46 Misc. Mixed Shipments 7  
47 Small Packaged Freight Shipments 4  
48 Hazardous Waste 16  
49 Hazardous Materials 18  
99 Unknown 12  
Other  35 87
Notes: 1. For Indiana model, “Other” includes STCC groups 21, 27, 30, 31, 38, 39.
2. For Indiana model, US mail and express mail groups are analyzed separately.
The actual workings of the model are very similar to a UTP model. For each of 21 commodity
groups that are considered important to Indiana (see Table A.10), trip generation equations
were developed based on a regression of data available from the 1993 CFS. Forecasts for
Guidebook on Statewide Travel Forecasting 114
Indiana county productions and attractions are then
based on countylevel employment and population
projections commercially available from Woods &
Poole. Table A.11 shows the trip generation
equations that were developed. For areas outside of
Indiana, forecasts are based on national growth
factors.
Following trip generation, freight shipments are
distributed by a gravity model that is also calibrated
using the CFS data. Special care is taken to match
the average shipping distance per ton for each
commodity group. This prevents an inappropriate
weighting for many shortdistance lightweight
deliveries versus a few longdistance heavyweight
shipments that might be included in the same
commodity group. The mode split step also utilizes
the 1993 CFS, projecting the 1993 national shares
into the future.
Before assigning traffic to the network, the Indiana
model (like the Wisconsin model) divides the freight
tonnages into an equivalent number of vehicles, with
tonspervehicle rates determined separately for each
commodity group. The rates are based on values (by
commodity group) from the ICC Rail Waybill sample
and the assumption that each truckload carries 40%
of the load carried by a railcar. The tonspertruck
and tonsperrailcar values used are shown in Table
A.10. A daily traffic conversion is
also made for the Indiana model,
assuming 5 working weekdays and
(from the Highway Capacity Manual)
0.44 working days for each
weekend day. This results in a 5.88
day work week or a 306 day
shipping year.
Finally, the traffic is assigned to the
network using an allornothing
process. Since a straight allor
nothing assignment typically loads
too many trips onto the interstate
highways, a procedure to adjust the
link speeds for noninterstate
highway segments is provided.
The adjustment involves
calculating new speeds for noninterstate links using the equation shown in Figure A.16. This
serves to draw more trips from the interstate roads to the competing US and state highways that
run parallel to them.
Figure A.14. Instate TAZs from
Indiana Model
Figure A.15. OutofState TAZs for Indiana Model
The State of the Art in Statewide Travel Demand Forecasting 115
Figure A.16. Link Speed Adjustments from Indiana Model
Table A.11. Trip Generation Equations from Indiana Model
Commodity Group
(by STCC)
1
Trip Generation Equations
2
(tons produced or attracted)
1 P = 1445  0.523(ag services) + 0.0048(ag cash)
A = 0.819 P
11 P = 7.6 (coal)
A = 3.1(coal) + 5.3(mining)
14 P = 0.078(manufacturing)
A = 0.997 P
20 P = 0.282(food)
A = 0.832(population) + 0.162(food)
22 P = 0.016(textiles)
A = 0.003(apparel) + 0.0001(all)
23 P = 0.004(apparel)
A = 0.002(apparel) + 0.011(population)
24 P = 0.668(lumber)
A = 0.728 P
25 P = 0.017(furniture)
A = 0.033(population) + 0.002(furniture)
26 P = 0.103(pulp) + 0.056(lumber)
A = 0.085(pulp) + 0.259(population)
28 P = 0.150(chemicals) + 1.164(petroleum)
A = 0.077(chemicals) + 0.455(petroleum) + 0.683(population)
29 P = 6.857(petroleum)
A = 4.007(petroleum) + 1.881(population)
32 P = 2.882(population)
A = 2.914(population)
33 P = 0.085(metals)
A = 0.093(metals) + 0.061(fabrication)
S
adj
= S
prev
+ [2 x (65mph  S
prev
)
0.5
]
where:
S
adj
= the adjusted link speed
S
prev
= the previously used link speed (usually the speed
limit on the link)
Guidebook on Statewide Travel Forecasting 116
Table A.11. Trip Generation Equations from Indiana Model (continued)
Commodity Group
(by STCC)
1
Trip Generation Equations
2
(tons produced or attracted)
34 P = 0.013(metals) + 0.034(fabrication)
A = 0.035(fabrication)
35 P = 0.013(machinery)
A = 0.010(machinery)
36 P = 0.004(metals) + 0.004(fabrication) + 0.003(electrical)
A = 0.005(fabrication) + 0.034(population)
37 P = 0.040(transportation)
A = 0.027(transportation)
40 P = 0.00048(population)
A = 0.0067(manufacturing)
Other P = 1.097 A
A = 0.254 (population)
Notes: 1.
2.
See Table A.10 for description of groups.
population = total population
all = total employment
ag services = employment is SIC 7
ag cash = gross receipts (in $1000) from farming
coal = employment in SIC 11
mining = employment in SIC 14
manufacturing = employment in SIC 2 and SIC 3
food = employment in SIC 20
textiles = employment in SIC 22
apparel = employment in SIC 23
lumber = employment in SIC 24
furniture = employment in SIC 25
pulp = employment in SIC 26
chemicals = employment in SIC 28
petroleum = employment in SIC 29
metal = employment in SIC 33
fabrication = employment in SIC 34
machinery = employment in SIC 35
electrical = employment in SIC 36
transportation = employment in SIC 37
Discussion
As noted above, both the Wisconsin freight model and the Indiana freight model use a county
level TAZ structure. This appears to be the accepted level of geographical detail for freight
modeling for two reasons. First, essential commodity flow information has been widely available
only at the BEA region level of detail, which is too large a geographical area for meaningful use
in a statewide model. Use of this information at a county level involves only one step of
disaggregation. In contrast, using smaller scale TAZs (e.g., at the census tract or township
level) typically involves additional disaggregation steps and additional layers of assumptions
about the characteristics of the commodity flows. Second, a large amount of economic
information is available (e.g., County Business Patterns data) or will soon be available (e.g., the
The State of the Art in Statewide Travel Demand Forecasting 117
Reebie and Colography databases mentioned in Section A.5) at a county level for use as inputs
to a statewide freight model.
As it is, the Indiana model depends heavily upon CFS data, which still must be disaggregated
once to get to the county level. The CFS is then supplemented with information that includes
important commodity flows that are not covered by the CFS, such as certain agricultural
products or solid waste. On the other hand, Wisconsin’s model makes use of a commercially
available commodity flow database. In doing so, it loses some sense of where the data is
coming from and what assumptions have been made in supplementing it.
The most obvious difference between the two models is the techniques each uses to generate
and distribute the predicted freight flows. The Wisconsin model does not have a conventional
trip generation step. Instead, future freight flows are forecast by a growth factor procedure,
where the growth factors (one for employment growth and one for productivity growth) are
calculated by econometric models developed by specialized consultants. These factors are
then applied to the base year flows. The Indiana model, in more a typical fourstep fashion,
uses regression equations developed from the CFS to generate its freight flows and a gravity
model to distribute them. The method used in the Wisconsin model seems appropriate for use
in a state where little change is expected in the transportation network over the forecast period.
However, the Indiana model offers greater flexibility for considering the effects of changes in the
network (e.g., a new highway corridor with reduced travel times).
In the end, both models use very similar processes to divide the predicted freight flows into an
equivalent number of vehicles and assign them to the network. Many of the tonspertruck
values are higher for the Indiana model, but both models use comparable daysperyear factors.
The Wisconsin model gives less attention to the tonspervehicle considerations for railcars, but
that is likely a function of the Wisconsin DOT’s primary goal of estimating future highway
demand. An unfortunate part of the assignment step for both models is the failure to address
the possibility of congestion due to the presence of a large number of passenger vehicles
sharing the road. Attention to congestion considerations (and related issues such as the
development of factors for seasonal or even hourly flows) should not be difficult given available
forecasting software.
A.8. Recommendations and Conclusions
Some general advice on constructing statewide travel forecasting models (or models for any
geographical area) is contained in the documentation for the Indiana freight model [8.1], which
states:
There is a temptation to evaluate the … forecasts. It should be obvious
that this is not possible until the forecast dates have been reached …
One’s acceptance of the forecasts should be based on the quality of the
methods used in the analysis of the [current] flows and the accuracy of
the methods in replicating existing conditions. (p. 141)
Keep the process sound and rational, the Indiana report seems to say. To this might be added
two more common sense suggestions: use methods of modeling that are appropriate to the
results desired and keep the process simple. Any recommendations for “best practice” in travel
forecasting should be based on these principles. Thus, a good model should be easy to explain
to an informed audience and easy to justify to an interested public (who will likely be funding the
modeling efforts). The following eight recommendations are directed toward developing just
such models for use in statewide travel forecasting.
Guidebook on Statewide Travel Forecasting 118
Recommendations for Statewide Travel Demand Forecasting
1. Use a statewide model for analysis of statewide effects of system or socioeconomic
changes or for statewide corridor planning only. This is a philosophical decision that drives
the rest of the modeling process. The statewide passenger models examined for this
appendix appear (because of their continued manipulation of information at a small scale) to
be aimed at examining flows at a project level. Use of a statewide model for smallscale or
projectspecific purposes is dubious, due to the increased necessity for adjustments (K
factors) and the reduced ability of a statewide modeler to be aware of localized conditions
across the state. Separate methods should be used for forecasts at the smaller scale,
where more specialized and local knowledge can be applied.
2. Build the statewide model in a form consistent with available data. This second
recommendation logically follows from the first. As demonstrated in the passenger and
freight models reviewed in this appendix, the majority of transportation (as opposed to
merely demographic) information is generally available only at the county level. Further
steps to artificially disaggregate the information involve assumptions that are probably not
necessary for modeling at a statewide level. These assumptions are too easy to forget
when adjustments are made to try to match the base year traffic flows. With county level
modeling, there is less temptation to match flows at a small (census tract) level, when the
model is really based on information from a much larger (countylevel) scale. At the county
level of aggregation it is also more likely that a sufficient number of data samples can be
found to calibrate logitstyle models and other more “advanced” modeling structures similar
to those discussed in reference to intercity modeling.
3. Examine, simultaneously, alternative methods of modeling. Concurrent use of alternative
modeling techniques is, as noted in Section A.6, the procedure suggested at the close of the
Kentucky model’s documentation. There are obviously a large number of techniques
available for use in developing a statewide forecast, including various disaggregate
behavioral models. Examinations could also be made to determine whether it is worthwhile
to extend the network into other states to buffer external station effects or to include a
national network or to experiment with the number of purposes. If several different model
structures are available, then a better sense can be gained of the sensitivities of the various
techniques and the possible range of forecasts for future travel.
4. Reexamine the structure of any “traditional” UTP methods used. If “advanced” techniques
are not warranted or do not prove fruitful, then “traditional” techniques may require some
overhaul. A principal example is the use of Kfactors for adjusting the gravity model. The K
factors used in the models examined appear to have been developed mostly in an effort to
adjust the gravity model results to match the base year traffic, rather than through
behavioral principles. Behavioral methods might be better for increasing or decreasing trips
between particular TAZs. For instance, if it is desired to represent the reluctance of
residents of one state to work in another, adding a short slowspeed (highdisutility) link at a
border crossing might work just as well.
5. Make use of existing government and commercial databases. Most of the data used to
develop the models reviewed in Section A.6 and Section A.7 came from a few government
(NPTS, CTPP, CFS) sources or similarly few commercial (Reebie, REMI, WEFA, Woods &
Poole) commercial sources. This data, judiciously supplemented and modified by local
information, proved sufficient to develop workable statewide models. New database
products under development (at the only the county level for freight) will make the process
still easier.
The State of the Art in Statewide Travel Demand Forecasting 119
6. Make use of existing statewide traffic monitoring programs. Aside from the obvious value in
assessing the reasonableness of model results, the information gathered by statewide
monitoring programs can be useful in developing daily and seasonal factors for various
roadway types and geographical areas in the state, thereby contributing to the accuracy of
the model. Monitoring data could also be used to assist in making modifications to national
rates used for special generators and external stations. The use of traffic monitoring data
would be greatly assisted by the use of time series techniques of data analysis, such as the
BoxJenkins methods noted in Section A.2.
7. Plan for future data collections that will enhance an existing model. Although statewide
forecasting model may have been built primarily from existing data sources, it could be
improved with additional, locallycollected data. Examples are data that may allow spatial
disaggregation below the county level and data on recreational travel, expecially on the
weekend.
8. Make use of expert panels in the modeling process. Expert panels can be very useful in
filling gaps in existing socioeconomic data, assisting with model assumptions and
disaggregating model results to smaller divisions of zones. The Wisconsin freight model is
notable for its use of expert panels, and it shows how assembling groups of people with
direct experience in freight can be used to take some of the burden away from individual
modelers (or small groups of modelers) and open up the forecasting process to
knowledgeable people outside of the DOT.
9. At future dates, assess the performance of the model(s) used. In all of the material
reviewed for this appendix, there is one subject that is glaringly absent: comparison of
model results with subsequent traffic demands. At the time a model is developed it is
obviously impossible to guess how effective it will be in predicting future traffic. Without
some effort at future dates to assess how well previous models have worked, modelers may
be doomed to repeating the same mistakes. It might also be possible to determine how
much modeling effort is sufficient to generate forecasts for the particular transportation
decisions being made.
Conclusions
The full arsenal of available techniques discussed in the literature of intercity and statewide
travel forecasting is not being brought to bear in existing statewide models. One of the primary
reasons is the lack of data available in sufficient quantities to build and calibrate “fine grain”
models at a statewide level. Another reason is a general lack of confidence in the value of
developing a statewide model in the first place. This is attested to by the vast majority of states
that do not perform any travel forecasting at the statewide level and by the states that built
models in the past, but have not continued to use them. Early efforts in statewide model
building were handicapped by rapidly changing political and economic environments.
The increasing availability of more userfriendly computer programs for travel forecasting and
the increasing availability of travel and socioeconomic data to feed these programs should begin
to alleviate some of the problems that caused states to abandon their statewide forecasting
efforts. The cardpunch technologies that existed during the first wave of statewide modeling in
the early 1970s are long gone. In addition, a continued federal government focus on planning
issues that can be addressed by statewide modeling – as begun with ISTEA and the CAAA –
should provide an impetus for state DOTs to once again explore the possibilities of forecasting
travel demand at a statewide level. A continuing development of databases useful for
calibration of models (especially behavioral models) at a statewide scale would also be helpful.
Guidebook on Statewide Travel Forecasting 120
As an additional impetus to model building, state DOTs are now being pressured to address
traffic congestion by programming transportation improvements. Improvements in intercity and
interregional corridors are best analyzed with statewide models.
Future research might include an investigation of the proprietary forecasting techniques
employed by private transportation firms that operate over large geographical areas. This
includes railroads, airlines, trucking companies and express delivery services. Other future
research might be aimed at understanding the underlying causes in the socioeconomic trends
that drive the forecasting models.
As this appendix has sought to point out, there are a number of techniques available to
prospective modelers – from the familiar fourstep models and growth factor techniques to more
advanced probabilitybased or optimization methods. There is an enormous opportunity to
combine and overlap the available techniques and to explore the benefits of travel demand
modeling at a statewide level. The challenge is now to put these techniques to the test of use.
A.9. References
References for Section A.1
1.1 Pub. L. 102240, Title I, Sec. 1025 Statewide Planning, 105 Stat. 1914.
1.2 KJS Associates, Inc. Statewide Travel Demand Model Update and
Calibration: Phase II, Michigan Department of Transportation, April
1996.
1.3 Cambridge Systematics. New Hampshire Statewide and Subarea Travel
Models Plan, New Hampshire Department of Transportation, March
1995.
References for Section A.2
2.1 Church, D. E. Outlook for Better Regional and National Forecasts of
Highway Traffic and Finance. In HRB Bulletin 257, HRB, National
Research Council, Washington, D.C., 1960, pp. 3638.
2.2 Burch, J. S. Traffic Interactance Between Cities. In HRB Bulletin 297,
HRB, National Research Council, Washington, D.C., 1961, pp. 1417.
2.3 Vogt, Ivers and Associates. NCHRP Report 70: Social and Economic
Factors Affecting Intercity Travel, HRB, National Research Council,
Washington, D.C., 1969.
2.4 Watson, P. L. Comparison of the Model Structure and Predictive Power
of Aggregate and Disaggregate Models of Intercity Mode Choice. In
Transportation Research Record 527, TRB, National Research Council,
Washington, D.C., 1974 pp. 5965.
The State of the Art in Statewide Travel Demand Forecasting 121
2.5 Rallis, T. Intercity Transport: Engineering and Planning. Wiley, New
York, 1977.
2.6 Koppelman F.S., GK. Kuah, and M. Hirsh. Review of Intercity
Passenger Demand Modeling: Mid60’s to the Mid80’s, The
Transportation Center, Northwestern University, Evanston, Illinois, 1984.
2.7 Khan, A. M. Towards the Development of Innovative Models of Intercity
Travel Demand. Transportation Quarterly, Vol. 39, No. 2, April 1985, pp.
297316.
2.8 Quandt, R. E. and W. J. Baumol. The Abstract Mode Model: Theory and
Measurement. Journal of Regional Science, Vol. 6, No. 2, 1966, pp. 13
26.
2.9 Yu, J. C. Demand Model for Intercity Multimode Travel. Transportation
Engineering Journal of ASCE, Vol. 96, No. TE2, May, 1970, pp. 203218.
2.10 Cohen, G. S., N. S. Erlbaum, and D. T. Hartgen. Intercity Rail Travel
Models. In Transportation Research Record 673, TRB, National
Research Council, Washington, D.C., 1978, pp. 2125.
2.11 Peers, J. B., and M. Bevilacqua. Structural Travel Demand Models: An
Intercity Application. In Transportation Research Record 569, TRB,
National Research Council, Washington, D.C., 1976 pp. 124135.
2.12 Kaplan, M. P., A. D. Vyas, M. Millar, and Y. Gur. Forecasts of Intercity
Passenger Demand and Energy Use Through 2000. In Transportation
Research Record 870, TRB, National Research Council, Washington,
D.C., 1982, pp. 8390.
2.13 DiRenzo, J. F. and L. P. Rossi. Diversion Model for Estimating High
Speed Rail Use. In Highway Research Record 369, HRB, National
Research Council, Washington, D.C., 1971, pp. 1525.
2.14 Brand, D., T. E. Parody, P. S. Hsu, and K. F. Tierney. Forecasting High
Speed Rail Ridership. In Transportation Research Record 1342, TRB,
National Research Council, Washington, D.C., 1992, pp. 1218.
2.15 Buckeye, K. R. Implications of HighSpeed Rail on Air Traffic. In
Transportation Research Record 1341, TRB, National Research Council,
Washington, D.C., 1992, pp. 1927.
2.16 Miller, E. J. Intercity Passenger Travel Demand Modelling: Present
State and Future Possibilities. In Proceedings of the 27
th
Annual Meeting
of the Canadian Transportation Research Forum, University of
Saskatchewan, Saskatoon, Saskatchewan,1992, pp. 378389.
2.17 Forinash, C. V. A Comparison of Model Structures for Intercity Travel
Mode Choice, M. S. Thesis, Northwestern University, 1992.
Guidebook on Statewide Travel Forecasting 122
2.18 Forinash, C. V. and F. S. Koppelman. Application and Interpretation of
Nested Logit Models of Intercity Mode Choice. In Transportation
Research Record 1413, TRB, National Research Council, Washington,
D.C., 1993, pp. 98106.
2.19 Morrison, S. A., and C. Winston. An Econometric Analysis of the
Demand for Passenger Transportation. Research in Transportation
Economics, Vol. 2, 1985, pp. 213237.
2.20 Koppelman, F. S. Multidimensional Model System for Intercity Travel
Choice Behavior. In Transportation Research Record 1241, TRB,
National Research Council, Washington, D.C., 1989, pp. 18.
2.21 Koppelman, F. S., and M. Hirsh. Intercity Passenger Decision Making:
Conceptual Structure Data. In Transportation Research Record 1085,
TRB, National Research Council, Washington, D.C., 1986, pp. 7075.
2.22 Koppelman, F. S. and M. Hirsh. Intercity Passenger Travel Demand
Analysis and Forecasting, The Transportation Center, Northwestern
University, Evanston, Illinois, 1987.
2.23 Moavenzadeh, F., M. J. Markow, B. D. Brademeyer, and K. N. A. Safwat.
A Methodology for Intercity Transportation Planning in Egypt.
Transportation ResearchA, Vol. 17A, No. 6, 1983, pp. 481491.
2.24 ElHawary, M. and A. Gadallah. Egypt Intercity Transportation Model. In
Research for Tomorrow’s Transportation Requirements, The Centre for
Transportation Studies, University of British Columbia, Vancouver, 1986,
pp. 9971009.
2.25 Safwat, K. N. A. Application of Simultaneous Transportation Equilibrium
Model to Intercity Passenger Travel in Egypt. In Transportation
Research Record 1120, TRB, National Research Council, Washington,
D.C., 1987, pp. 5259.
2.26 Dean, D. L. Modeling Dilemma of Intercity Bus Transportation. In
Transportation Research Record 887, TRB, National Research Council,
Washington, D.C., 1982, pp. 3537.
2.27 Neumann, E. S. and B. F. Byrne. A Poisson Model of Rural Transit
Ridership. In Transportation Research Record 661, TRB, National
Research Council, Washington, D.C., 1978, pp. 2127.
2.28 Brown, S. L. and W. S. Watkins. The Demand for Air Travel: A
Regression Study of TimeSeries and Cross Sectional Data in the US
Domestic Market. In Highway Research Record 213, HRB, National
Research Council, Washington, D.C., 1968, pp. 2134.
2.29 Oberhausen, P. J. and F. S. Koppelman. TimeSeries Analysis of
Intercity Air Travel Volume. In Transportation Research Record 840,
TRB, National Research Council, Washington, D.C., 1982, pp. 1521.
The State of the Art in Statewide Travel Demand Forecasting 123
2.30 Pickrell, D. H. Model of Intercity Travel Demand. In Deregulation and
the Future of Intercity Travel (J. R. Meyer and C. V. Oster, eds.), MIT
Press, Cambridge, Mass., 1987, pp. 249260.
2.31 Thorson, B. A. and K. A. Brewer. Model to Estimate Commuter Airline
Demand in Small Cities. In Transportation Research Record 673, TRB,
National Research Council, Washington, D.C., 1978, pp. 187193.
2.32 Ghobrial, A. and A. Kanafani. QualityofService Model of Intercity Air
Travel Demand. Journal of Transportation Engineering, Vol. 121, No. 2,
March/April 1995, pp. 135140.
2.33 Gilbert, G. Recreational Travel Behavior: The Case for Disaggregate,
Probabilistic Models. In Special Report 149, TRB, National Research
Council, Washington, D.C., 1974, pp. 223230.
References for Section A.3
3.1 Smith, P. L. Forecasting Freight Transport Demand – The State of the
Art. The Logistics and Transportation Review, Vol. 10, No. 4, 1974, pp.
311326.
3.2 Winston, C. The Demand for Freight Transportation: Models and
Applications. Transportation ResearchA, Vol. 17A, No. 6, 1983, pp.
419427.
3.3 Friesz, T. L., R. L. Tobin, and P. T. Harker. Predictive Interstate Freight
Network Models: The State of the Art. Transportation ResearchA, Vol.
17A, No. 6, 1983, pp. 409417.
3.4 Bronzini, M. S. Evolution of a Multimodal Freight Transportation System
Network Model. In Transportation Research Forum Proceedings: 20
th
Annual Meeting, Vol. XXI, No. 1, 1980, pp. 475485.
3.5 Morton, A. L. A Statistical Sketch of Intercity Freight Demand. In
Highway Research Record 296, HRB, National Research Council,
Washington, D.C., 1969, pp. 4765.
3.6 Jones, P. S., and G. P. Sharp. MultiMode Intercity Freight
Transportation Planning for Underdeveloped Regions. In Transportation
Research Forum Proceedings: 18
th
Annual Meeting, Vol. XVIII, No. 1,
1977, pp. 523531.
3.7 Friesz, T. L. and E. K. Morlok. Recent Advances in Network Modelling
and Their Implication for Freight System Planning. In Transportation
Research Forum Proceedings: 21
st
Annual Meeting, Vol. XXI, No. 1,
1980, pp. 513520.
Guidebook on Statewide Travel Forecasting 124
3.8 Friesz, T. L., J. A. Gottfried, and E. K. Morlok. A Sequential Shipper
Carrier Network Model for Predicting Freight Flows. Transportation
Science, Vol. 20, No. 2, May 1986, pp. 8091.
3.9 Carey, D. E., H. S. Mahmassani, and G. S. Toft. Air Freight Usage
Patterns of Technology Based Industries. In Transportation Research
Record 1179, TRB, National Research Council, Washington, D.C., 1988,
pp. 3339.
3.10 Middleton, D. R., J. M. Mason, and T. ChiraChavala. Trip Generation for
SpecialUse Truck Traffic. In Transportation Research Record 1090,
TRB, National Research Council, Washington, D.C., 1986, pp. 813.
3.11 Linsenmeyer, D. Effect of UnitTrain Grain Shipments on Rural
Nebraska Roads. In Transportation Research Record 875, TRB,
National Research Council, Washington, D.C., 1982, pp. 6064.
3.12 Sasaki, B. R. Predicting Transporter’s Choice of Mode. In
Transportation Research Record 577, TRB, National Research Council,
Washington, D.C., 1976, pp. 1318.
3.13 Roth, R. D. Approach to Measurement of Modal Advantage. In
Transportation Research Record 637, TRB, National Research Council,
Washington, D.C., 1977, pp. 6270.
3.14 Jelavich, M. S. Determinants of Freight Modal Choice. In Transportation
Research Record 668, TRB, National Research Council, Washington,
D.C., 1978, pp. 1417.
3.15 Wilson, F. R., B. G. Bisson, and K. B. Kobia. Factors that Determine
Mode Choice in the Transportation of General Freight. In Transportation
Research Record 1061, TRB, National Research Council, Washington,
D.C., 1986, pp. 2631.
3.16 Murthy, A. S. and B. Ashtakala. Modal Split Analysis Using Logit Models.
Journal of Transportation Engineering, Vol. 113, No. 5, September 1987,
pp. 502519.
References for Section A.4
4.1 Whiteside, R.E., C. L. Cothran, and W. M. Kean. Intercity Traffic
Projections. In Highway Research Record 205, HRB, National Research
Council, Washington, D.C., 1967, pp. 110135.
4.2 Hazen, P. I. A Comparative Analysis of Statewide Transportation
Studies. In Highway Research Record 401, HRB, National Research
Council, Washington, D.C., 1972, pp. 3954.
4.3 Statewide Travel Demand Forecasting, US Department of
The State of the Art in Statewide Travel Demand Forecasting 125
Transportation, FHWA, Washington, D.C., 1973.
4.4 Assessment of Modeling Techniques and Data Sources for Multimodal
Statewide Transportation Planning, Peat, Marwick, Mitchell & Co.,
October, 1977.
4.5 Person Travel Forecasting Procedures for Multimodal Statewide
Transportation Planning, Peat, Marwick, Mitchell & Co., February, 1979.
4.6 Russell, E. R., J. H. Jatko, and M. Ahsan. Investigation and
Development of a Statewide Traffic Assignment Model for the State of
Kansas, Kansas Department of Transportation, Topeka, Kansas, 1977.
4.7 BartonAschman Associates. A Research Process for Developing a
Statewide Multimodal Transport Forecasting Model, New Mexico State
Highway and Transportation Department, Santa Fe, N. M., 1991.
4.8 Benson, J. D., J. A. Mullins, and G. B. Dresser. Feasibility of Developing
a Statewide Modeling System for Forecasting Intercity Highway Volumes
in Texas, Texas Transportation Institute, College Station, Texas, 1991.
4.9 Green, M. K. Multiple Screenline Study to Determine Statewide Traffic
Patterns. In HRB Bulletin 253, HRB, National Research Council,
Washington, D.C., 1960, pp. 139144.
4.10 Virkud, U. and C. S. Keyes. Design and Implementation of a Statewide
Roadside Origin Destination Survey in Vermont. In Transportation
Research Record 1477, TRB, National Research Council, Washington,
D.C., 1995, pp. 1525.
4.11 Crevo, C. C., R. S. Niedowski, and D. J. Scott. Design and Conduct of a
Statewide Household Travel Survey in Vermont. In Transportation
Research Record 1477, TRB, National Research Council, Washington,
D.C., 1995, pp. 2630.
4.12 Glass, D., R. W. Tweedie, and R. J. Zabinski. Transportation Planning
Data for the 1990’s. In Transportation Research Record 1283, TRB,
National Research Council, Washington, D.C., 1990, pp. 4550.
4.13 Meyer, M. D. and G. Mazur. Census Data Use for Statewide
Transportation Planning and Transportation Planning in Smaller
Metropolitan Areas. In Conference Proceedings 4: Decennial Census
Data for Transportation Planning, TRB, National Research Council,
Washington, D.C., 1995, pp. 7582.
4.14 Sharma, S. C. and A. Werner. Improved Method of Grouping
Provincewide Permanent Traffic Counters. In Transportation Research
Record 815, TRB, National Research Council, Washington, D.C., 1981,
pp. 1217.
4.15 Bester, C. J. and J. D. De Joubert. Classifying a Rural Road Network for
Guidebook on Statewide Travel Forecasting 126
Traffic Counting. In Transportation Research Record 1090, TRB,
National Research Council, Washington, D.C., 1986, pp. 2226.
4.16 Allbright, D. A Quick Cluster Control Method: Permanent Control Station
Cluster Analysis in Average Daily Traffic Calculations. In Transportation
Research Record 1134, TRB, National Research Council, Washington,
D.C., 1987, pp. 5764.
4.17 Traffic Monitoring Guide, U. S. Department of Transportation, FHWA,
Washington, D. C., 1985.
4.18 DiRenzo, J., N. Serianni, and L. King. Improving the Productivity of the
Pennsylvania Transportation Department’s Traffic Count Program. In
Transportation Research Record 1050, TRB, National Research Council,
Washington, D.C., 1985, pp. 1319.
4.19 Hallenbeck, M. E. Development of an Integrated Statewide Traffic
Monitoring System. In Transportation Research Record 1050, TRB,
National Research Council, Washington, D.C., 1985, pp. 512.
4.20 Ritchie, S. G. A Statistical Approach to Statewide Traffic Counting. In
Transportation Research Record 1090, TRB, National Research Council,
Washington, D.C., 1986, pp. 1421.
4.21 Ritchie, S. G. and M. E. Hallenbeck. Evaluation of a Statewide Highway
Data Collection Program. In Transportation Research Record 1090,
TRB, National Research Council, Washington, D.C., 1986, pp. 2735.
4.22 Albright, D., C. Blewett, and R. Goldstine. Development of State Traffic
Monitoring Standards: Implementing Traffic Monitoring Guide in New
Mexico. In Transportation Research Record 1236, TRB, National
Research Council, Washington, D.C., 1989, pp. 913.
4.23 Albright, D. P. Revision of Statewide Traffic Data Standards Indicated
During Implementation of a Traffic Monitoring System. In Transportation
Research Record 1271, TRB, National Research Council, Washington,
D.C., 1990, pp. 4854.
4.24 Van Zuylen, H. J. and L. G. Willumsen. The Most Likely Trip Matrix
Estimated from Traffic Counts. Transportation ResearchB, Vol 14B,
1980, pp. 281293.
4.25 Bell, M. G. H. The Estimation of an OriginDestination Matrix from Traffic
Counts. Transportation Science, Vol. 17, No. 2, May 1983, pp. 198217.
4.26 Hendrickson, C. and S. McNeil. Estimation of OriginDestination
Matrices with Constrained Regression. In Transportation Research
Record 976, TRB, National Research Council, Washington, D.C., 1984,
pp. 2532.
4.27 Wilson, E. M. and J. Wang. Statewide Transportation Planning – An
The State of the Art in Statewide Travel Demand Forecasting 127
Interactive Modeling Process, Wyoming Department of Transportation,
May 1995.
4.28 Wilson, E. M. and J. Wang. Interactive Statewide Transportation
Planning Modeling Process. In Transportation Research Record 1499,
TRB, National Research Council, Washington, D.C., 1995, pp. 16.
4.29 Walker, W. T. Method to Synthesize a Full Matrix of Interdistrict Highway
Travel Times from Census JourneytoWork Data. In Transportation
Research Record 1236, TRB, National Research Council, Washington,
D.C., 1989, pp. 5058.
4.30 Procedures Manual for Forecasting Traffic on the Rural Trunk Highway
System, Minnesota Department of Transportation, Traffic Forecasting
Unit, Minneapolis, Minnesota, April, 1985.
4.31 Harmatuck, D. J. Improving Traffic Forecasting with Time Series Models,
Wisconsin Department of Transportation, Madison, January, 1997.
4.32 Traffic Monitoring Guide, U. S. Department of Transportation, FHWA,
Washington, D. C., 1992.
4.33 ConnDOT Modeling Procedure, Connecticut Department of
Transportation, February 1997.
4.34 PERFORM Person Forecasting Model: Trip Generation, Connecticut
Department of Transportation, Staff Paper 9310, 1993.
4.35 Shimpeler, Corridino Associates. Florida Highway Traffic Forecasting
Models Technical Report 1: Statewide Model Development, Florida
Department of Transportation, September 1990.
4.36 Shimpeler, Corridino Associates. Statewide Highway Traffic Forecasting
Model Technical Report 2: Collection and Assembly of Data, Florida
Department of Transportation, October 1990.
4.37 Cambridge Systematics. Major Corridor InvestmentBenefit Analysis
System, Appendix A: Scope of Work, Indiana Department of
Transportation, Undated.
4.38 Wilbur Smith Associates. Kentucky Statewide Traffic Model Final
Calibration Report, Kentucky Transportation Cabinet, April 1997.
4.39 KJS Associates, Inc. Statewide Travel Demand Model Update and
Calibration: Phase II, Michigan Department of Transportation, April
1996.
4.40 Cambridge Systematics. New Hampshire Statewide and Subarea Travel
Models Plan, New Hampshire Department of Transportation, March
1995.
Guidebook on Statewide Travel Forecasting 128
4.41 URS Consultants, Inc. Effects of Interstate Completion and Other Major
Highway Improvements on Regional Trip Making and Goods Movement:
Network Development, New Jersey Department of Transportation, June
1995.
4.42 URS Consultants, Inc. Effects of Interstate Completion and Other Major
Highway Improvements on Regional Trip Making and Goods Movement:
Auto Trip Table, New Jersey Department of Transportation, March 1995.
4.43 Vanasse, Hangen Brustlin, Inc. Statewide Travel Demand Model
Development, Vermont Agency of Transportation, April 1996.
4.44 KPMG Peat Marwick. Translinks 21: MultiModal Intercity Passenger
Analysis, Wisconsin Department of Transportation, June 1995.
4.45 Parsons Brinkerhoff Quade & Douglas, Inc. Consultant
Recommendations for ths Development of Phase II Databases, Models,
and Forecasting Methods: Transportation and Land Use Model
Integration Program Phase I, Task 1.6, Oregon Department of
Transportation, Salem, September 13, 1996.
4.46 Crevo, C. C. Characteristics of Summer Weekend Recreational Travel.
In Highway Research Record 41, HRB, National Research Council,
Washington, D.C., 1963, pp. 5160.
4.47 Ungar, A. NCHRP Report 44: Traffic Attraction of Rural Outdoor
Recreational Areas, HRB, National Research Council, Washington, D.C.,
1967.
4.48 Matthias, J. S. and W. L. Grecco. Simplified Procedure for Estimating
Recreational Travel to MultiPurpose Reservoirs. In Highway Research
Record 250, HRB, National Research Council, Washington, D.C., 1968,
pp. 5469.
4.49 Robinson, D. C. and W. L. Grecco. Stability of Recreational Demand
Model. In Highway Research Record 392, HRB, National Research
Council, Washington, D.C., 1972, pp. 147156.
4.50 Deacon, J. A., J. G. Pigman, and R. C. Deen. Travel to Outdoor
Recreation Areas in Kentucky. In Highway Research Record 392, HRB,
National Research Council, Washington, D.C., 1972, pp. 134135.
4.51 Deacon, J. A., J. C. Pigman, K. D. Kaltenbach, and R.C. Deen. Models
of Recreational Travel. In Highway Research Record 472, HRB, National
Research Council, Washington, D.C., 1973, pp. 4562.
4.52 Gyamfi, P. A Model for Allocating Recreational Travel Demand to
National Forests. In Highway Research Record 408, HRB, National
Research Council, Washington, D.C., 1972, pp. 5061.
4.53 Berg, W. D., P. A. Koushki, C. L. Krueger, and W. L. Bittner.
The State of the Art in Statewide Travel Demand Forecasting 129
Development of a Simulation Model for Regional Travel. In
Transportation Research Record 569, TRB, National Research Council,
Washington, D.C., 1976, pp. 96106.
References for Section A.5
5.1 Cambridge Systematics. NCHRP Report 388: A Guidebook for
Forecasting Freight Transportation Demand, TRB, National Research
Council, Washington, D.C., 1997.
5.2 Parsons Brinkerhoff Quade & Douglas, Inc. Consultant
Recommendations for the Development of Phase II Databases, Models,
and Forecasting Methods: Transportation and Land Use Model
Integration Program Phase I, Task 1.6, Oregon Department of
Transportation, Salem, September 13, 1996.
5.3 Casavant, K. L., W. R. Gillis, D. Blankenship, and C. Howard. Survey
Methodology for Collecting Freight Truck and Destination Data. In
Transportation Research Record 1477, TRB, National Research Council,
Washington, D.C., 1995, pp. 714.
5.4 Roger Creighton Associates. NCHRP Report 177: Freight Data
Requirements for Statewide Systems Planning, Research Report. TRB,
National Research Council, Washington, D.C., 1977.
5.5 Roger Creighton Associates. NCHRP Report 178: Freight Data
Requirements for Statewide Systems Planning, Users Manual. TRB,
National Research Council, Washington, D.C., 1977.
5.6 Memmott, F. W. NCHRP Report 260: Application of Statewide Freight
Demand Forecasting Techniques, TRB, National Research Council,
Washington, D.C., 1983.
5.7 Memmott, F. W. and R. H. Boekenkroeger. Practical Methodology for
Freight Forecasting. In Transportation Research Record 889, TRB,
National Research Council, Washington, D.C., 1982, pp. 17.
5.8 Middendorf, D. P., M. Jelavich, and R. H. Ellis. Development and
Application of Statewide, Multimodal Freight Forecasting Procedures for
Florida. In Transportation Research Record 889, TRB, National
Research Council, Washington, D.C., 1978, pp. 714.
5.9 Kim T. J. and J. J. Hinkle. Model for Statewide Freight Transportation
Planning. In Transportation Research Record 889, TRB, National
Research Council, Washington, D.C., 1982, pp. 1519.
5.10 Linsenmeyer, D., A. Azzam, and D. Olsen. Planning Rural
Transportation Systems: Applications of a Statewide Network Model. In
Profitability and Mobility in Rural America (W. Gillis, ed.), Pennsylvania
Guidebook on Statewide Travel Forecasting 130
State University Press, 1989, pp. 171191.
5.11 Statewide Intermodal Transportation Plan, State of Louisiana,
Department of Transportation and Development, March 1996.
5.12 Cambridge Systematics. Quick Response Freight Manual, FHWA,
September 1996.
5.13 Black, W. R. Transport Flows in the State of Indiana: Commodity
Database Development and Traffic Assignment, Phase 2, Transportation
Research Center, Indiana University, Bloomington, July 15, 1997.
5.14 Wilbur Smith Associates. Multimodal Freight Forecasts for Wisconsin,
Wisconsin Department of Transportation, July 1996.
5.15 KJS Associates, Inc. Statewide Travel Demand Model Update and
Calibration: Phase II, Michigan Department of Transportation, April
1996.
5.16 URS Consultants, Inc. Effects of Interstate Completion and Other Major
Highway Improvements on Regional Trip Making and Goods Movement:
Truck Trip Table, New Jersey Department of Transportation, August
1995.
5.17 Ashtakala, B. and A. S. Murthy. Optimized Gravity Models for
Commodity Transportation. Journal of Transportation Engineering, Vol.
114, No. 4, July 1988, pp. 393408.
5.18 Ashtakala, B. and A. S. N. Murthy. Sequential Models to Determine
Intercity Commodity Transportation Demand. Transportation Research
A, Vol. 27A, No. 5, 1993, pp. 373382.
5.19 Crainic, T.G., M. Florian, J. Guelat, and H. Spiess. Strategi Planning of
Freight Transportation: STAN, An InteractiveGraphic System. In
Transportation Research Record 1283, TRB, National Research Council,
Washington, D.C., 1990, pp. 97124.
5.20 Crainic, T.G., M. Florian, and J.E. Leal. A Model for the Strategic
Planning of National Freight Transportation by Rail. Transportation
Science, Vol. 24, No. 1, February 1990, pp. 124.
5.21 Guelat, J., M. Florian, and T. G. Crainic. A Multimode Multiproduct
Network Assignment Model for Strategic Planning of Freight Flows.
Transportation Science, Vol. 24, No. 1, February 1990, pp. 2539.
References for Section A.6
6.1 KJS Associates, Inc. Statewide Travel Demand Model Update and
Calibration: Phase II, Michigan Department of Transportation, April
1996.
The State of the Art in Statewide Travel Demand Forecasting 131
6.2 Wilbur Smith Associates. Kentucky Statewide Traffic Model Final
Calibration Report, Kentucky Transportation Cabinet, April 1997.
References for Section A.7
7.1 Wilbur Smith Associates. Multimodal Freight Forecasts for Wisconsin,
Wisconsin Department of Transportation, July 1996.
7.2 Black, W. R. Transport Flows in the State of Indiana: Commodity
Database Development and Traffic Assignment, Phase 2, Transportation
Research Center, Indiana University, Bloomington, July 15, 1997.
References for Section A.8
8.1 Black, W. R. Transport Flows in the State of Indiana: Commodity
Database Development and Traffic Assignment, Phase 2, Transportation
Research Center, Indiana University, Bloomington, July 15, 1997.
Guidebook on Statewide Travel Forecasting
Prepared for Federal Highway Administration
Prepared by Center for Urban Transportation Studies University of Wisconsin – Milwaukee In cooperation with Wisconsin Department of Transportation
March 1999
Table of Contents
Chapter 1. Introduction ................................................................................................ 1
Structure of the Guidebook..................................................................................................1 Reasons for Statewide Travel Forecasting .........................................................................1
Original ISTEA Planning Factors ..................................................................................................... 2
Overview of Approaches ......................................................................................................2
Time Series Methods....................................................................................................................... 3 Overview of Three/Four Step Models............................................................................................... 3 Calibration and Validation................................................................................................................ 4 Specialized Methods ....................................................................................................................... 4
Introduction to Data Sources ...............................................................................................5
Key Data Sources ........................................................................................................................... 5 Key Census Databases ................................................................................................................... 6 National Personal Transportation Survey (NPTS) Overview ............................................................. 6 American Travel Survey (ATS) Overview ......................................................................................... 6 Important Freight Data Sources ....................................................................................................... 7 Commercial Economic Forecasts..................................................................................................... 7
Tools ......................................................................................................................................7
Role of GIS ..................................................................................................................................... 7 Software Issues............................................................................................................................... 8
Composing a Complete Forecast ........................................................................................8
Chapter 2. Time Series Methods ................................................................................. 9
Introduction...........................................................................................................................9
What is Time Series Analysis?......................................................................................................... 9 Applicability of Time Series Analysis................................................................................................ 9
Important Statistical Concepts ..........................................................................................10 Growth Factors ...................................................................................................................11
Growth Factor Relationships.......................................................................................................... 11
Linear Trend Model.............................................................................................................12
BoxCox Transformations .............................................................................................................. 14
Moving Averages and Seasonality ....................................................................................15
Exponential Smoothing.................................................................................................................. 15 Central Moving Average ................................................................................................................ 15
Introduction to BoxJenkins (ARIMA) Methods ................................................................16
ARIMA .......................................................................................................................................... 16 Autoregressive (AR) Models .......................................................................................................... 16 Integrated (I) Models ..................................................................................................................... 17 Moving Average (MA) Models........................................................................................................ 17 Steps in Building an ARIMA Model ................................................................................................ 18 Extensions to ARIMA (Examples from WisDOT) ............................................................................ 18 Cyclic Patterns in AR Models......................................................................................................... 19
Case Studies .......................................................................................................................19
Air Travel I: Brown and Watkins (1968)......................................................................................... 19 Air Travel II: Oberhausen and Koppelman (1982) ......................................................................... 19 Traffic Levels I: Maine VMT .......................................................................................................... 20 Traffic Levels II: New Mexico Heavy Vehicle Traffic ...................................................................... 20
Table of Contents
i
................................................. 41 Appendix: Overview of Four Step Models................................... 34 Friction Factors...................... 38 Nested Logit Example................................................................................................................................................... 32 External Stations .................................................................................. Freight Forecasting........... 35 Trip Distribution: Application to Statewide Forecasting ...............................................................................Appendix: Some Elementary Statistical Concepts .... 40 Trip Assignment ...........................................................................42 Network Basics........................ 29 Other Data Issues..............................49 Establish Goals for Model..............................................................................................................................................................................................................................47 Commodity Flow Survey...................................................................................................................................... 28 Other Modal Networks ............................................................................. 35 Mode Split .................................................................. Passenger Forecasting................................................................................................................................................................................................................... 29 Socioeconomic and Demographic Characteristics.............................................................................................................................................................................................................................................................................................................................................. 33 Balancing Productions and Attractions............................ 30 Trip Attractions ........................................................... 27 Network Structure........................................................................................................................................................................................................................................................................................................... 42 Trip Generation Basics ................................................................................................................................................................................................................................................................................................................ 30 Model Steps................................................................................................................................................... 33 DoublyConstrained Gravity Model .............................................................................................................. 25 Defining the Scope of the Model.................................................................................. 36 Logit Example....... Forinash and Koppelman (1993)................................................................................................................................................................ 49 Guidebook on Statewide Travel Forecasting ii .................................................................................................................. 24 Introduction. 28 Zone Systems and Spatial Aggregation Issues ................................................... 44 Traffic Assignment Basics........................................................................... 30 Trip Productions ...................... 48 Commercial Freight Data Sources ....................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................... 26 Purposes............................................................... 46 Introduction.................................................................................................................................................................................................. 39 Mode Split Issues ........................... 43 Trip Distribution Basics ..................................................................................................................................................... 33 Trip Distribution: Gravity Model (Production Constraint) ................................................................................... 45 Chapter 4.................................................................................................................................................... 26 Modal Categories ................................................................................................................ 46 Key Freight Data Sources ............................ 48 Other Interesting Data Sources.........................................24 Urban v......................... More F&K...... 47 Vehicle Inventory and Use Survey (VIUS)........................................................................... 36 Logit........................................................... Statewide Models .............................................................................................. 44 Mode Split Basics...................46 Freight Model – Basic Steps ................. 49 Defining the Scope of the Freight Model..................................................................................... 27 Highway Network Attributes.............................................................30 Trip Generation Issues ........................................................................................................................................ 25 Sources of Travel Data............................................................................................21 Chapter 3................. 48 Local Surveys of Note .........26 Statewide Network Preparation........................................................................................................................................ 37 Nested Logit .............................. 40 Calibration and Validation..........
.........................................................................................................................................................................................................................................................................................................................................................................................71 Interstate Trip Rate .................................................................................................................................. 77 Calibration and Validation................................................................. 67 Acceptable Error Plots............. 59 Other Freight Forecasting Issues ........................................................... 60 Chapter 5........................................................................................................................................................................... 72 Composite Utilities for Trip Distribution ....................................80 Ordinary Least Squares (Linear Regression) ...............................................................................75 Review of Select Link Analysis ................................................................................................................................. 68 Stated Preference ......................................................... 56 Commodity Density and Load Weights ..............................................................................................................................................................................................TriState...................................................................................................................................................................................................................................................................................................................................................................................................................................................................78 National Defaults ................................................ 79 SSOD in Validation........................................................... 72 External Travel for Other State Models ........................................................................................51 Commodities ................................................................................................68 Value of Time and Value of Frequency .. 76 Pivot Steps........................... 80 Nonlinear Regression ............................................................ 79 Integration of Statewide and Urban Models........................................................................ Specialized Methods for Passenger Forecasting...................................................... 81 Table of Contents iii ......................... 57 Traffic Assignment.................................................................................................................................. 58 Assignment Fixes ............................ 50 Network Development ...........................................................70 Time of Day Considerations.......................................................................................................................................................................................................72 Development of Outstate Trip Generation for the Michigan Model.................................. 62 Single Station OriginDestination Surveys.... 52 Gravity Model for Freight ....................................................................................................................................................................................................................................... 60 Forecasting Inputs to a Freight Forecast.......................................... 51 Developing a Flow Matrix ...................................................................................................................................... 64 Truck Travel Surveys................................................................. 53 Mode Split Methods..................................................... 50 Selecting Modes............................................ 56 Elasticities ............................................................................................................................................................................................................................................................ 62 Introduction...........................62 Validation Data Sets: HPMS Element of the Traffic Monitoring Program.......................60 Combining Passenger and Freight Forecasts............................................................................................................................................................ 74 Hybrid Technique: Pivot Analysis ...................................................... 65 Other Survey Methods............................................................. 76 Pivot Method Refinements...........................................................................................................................................................74 TriState Total Demand Model............................................................................................................ 70 Single Mode Vehicle Occupancy: Michigan...........................................................73 Total Travel in Corridor: Wisconsin Model ................................. 78 Trip Table Estimation from Traffic Counts ..........Spatial Units .................62 Data Collection.................................................................................................. 66 Validation Quality ............................. 80 Appendix: Review of Calibration Methods................................................................ 55 Pivot Point................................................................................................................................................................................................................................................. 51 Trip Generation for Gravity Model.
.....................................................................1................................. 118 Conclusions......................................... 120 References for Section A........................................... 111 The Indiana Freight Model....................... 89 One Disaggregate DirectDemand Model .......................... 99 Related Models ............... 120 References for Section A...5...............4.................................................................................... 99 Discussion..................................................................................................4...........................................................................................7............................................................................................................................................................................8........................................ 98 Data for Freight Forecasting ............................................................................................. Introduction ..................................................................................................................................2... 100 The Michigan Passenger Model.......... 93 Discussion.............................................................................................................4.................... 84 A....................... 91 Discussion............................................................................................. 91 A.............................................................................................................84 Method of Research ..................................................................................................................................... 100 The Kentucky Passenger Model ........................................ Statewide Passenger Forecasting Literature ............................................................... 100 A..............................................................................................................................94 Data Collection for Passenger Travel.................................................................................................................................. 88 Aggregate DirectDemand Models .............................. 123 References for Section A........................... 95 Statewide Models of Passenger Travel ............................... 117 Recommendations for Statewide Travel Demand Forecasting .............................. 119 A........ Statewide Freight Forecasting Literature ........................................................................................................ 81 Fast Trip Distribution Calibration............................................................................................................................................................... 106 Discussion.................................... 92 Intercity Freight ModeSplit Models...................................................................................................................3............... Two Recent Passenger Models .................................................................................................................................... 94 A................................................... 92 Intercity Freight Models ..................................... 98 Statewide Models for Freight Forecasting ...................................2................................................ 91 Reviews of Intercity Freight Models .................................................................91 Introduction .........87 Types of Intercity Passenger Models .................................................................................................................... Intercity Freight Literature ................................................................................................................................................ 86 Organization of Appendix ............................... 120 References for Section A......................................................................................................................................................................................................................................................................... 124 Guidebook on Statewide Travel Forecasting iv ....................................................... 93 Figure A..........................................................................................................................................................9........................................................................................ 95 Trend Analyses of Passenger Travel ......................................................................... 88 Disaggregate Sequential Models ........................................................................ Recommendations and Conclusions .................................6................................................................................................... 116 A.......... References ......................................... Two Recent Freight Models ..................... 89 SingleMode and SinglePurpose Models .... Intercity Passenger Literature .................... 82 Acknowledgements ................................................................. 108 A...................................... 111 The Wisconsin Freight Model ...................................3................................................................... 83 APPENDIX: The State of the Art in Statewide Travel Demand Forecasting ..............................98 Introduction ...................................................................... 96 A........................................................................................ 94 Data Synthesis for Passenger Travel................................................................ 112 Discussion......................................................................................................1.................... 95 Discussion............................................................................................................................................................................. University of Pennsylvania ShipperCarrier Model.................................................................................................... 86 A............Maximum Likelihood Estimation..................................
............................. 131 Table of Contents v ........... 129 References for Section A............................................................................... 130 References for Section A.....................................................8................................................................ 131 References for Section A.................................................7....................References for Section A...............................6..............................................................................5...............
but it does not provide guidelines for any particular software package. Freight Forecasts. Topics range from simple growth factor methods to ARIMA (autoregressive. Chapter 3 deals primarily with means for adapting the traditional urban “four step” model to statewide travel forecasting. Time Series. Forecasts of Rural/Intercity Travel. This guidebook will make reference to travel forecasting software. Chapter 5 lists several methods that have occasional value for statewide. Introduction This guidebook reviews the stateofthepractice of statewide travel forecasting. methods are presented that have not been tried in statewide travel forecasting but show strong potential to improve the process. Specialized Methods for Passenger Forecasting.Chapter 1. this guidebook presents specialized and advanced techniques of potential interest to persons involved in statewide travel forecasting. but emphasizes heavy trucks. Chapter 2 deals with methods for extrapolating upon existing trends in traffic volumes. Reasons for Statewide Travel Forecasting There are a number of reasons why a state might be interested in forecasting statewide or rural travel. nor are the recommended techniques dependent upon the capabilities of any specific software package. In addition. It discusses all freight modes. Major portions of the chapter concern freight data sources. It focuses on those techniques that have been considered essential to good statewide travel forecasting. the differences in implementation are quite important. This guidebook does not describe methods that have been presented in the academic literature and are considered to be still under development. Overall assessments of the adequacy of the statewide transportation networks require forecasts of rural and intercity travel by all freight and passenger Introduction 1 . Chapter 4 deals with the need to integrate freight forecasts with passenger forecasts. Although there are similarities in theory. There are many forecasting situations that are not addressed well by a traditional “four step” model. Emphasis is placed on practical methods. This guidebook also makes a distinction between urban travel forecasting and statewide travel forecasting. integrated. corridor or intercity travel forecasts. “The State of the Art in Statewide Travel Demand Forecasting”. In some cases. Passenger Forecasts. In some places in the guidebook. Persons interested in recent research on this topic might want to consult the appendix. There are many facets of statewide travel forecasting and it is not possible to create a onesizefitsall model to deal with every possible situation. This guidebook advocates the need to select the technique that is most applicable to the problem. a simple growth factor model may be preferable to a fullblown network analysis of the whole state. Structure of the Guidebook After this introduction the guidebook consists of four chapters. movingaverage) models.
Border crossings. air pollution emissions and consumer surplus) require forecasts from both urban and rural areas. Satisfy Mandated Planning Requirements. such as truck. In addition. Develop ProjectLevel Forecasts in Rural Areas. A large portion of travel in most states is rural. major transportation facilities and military. All three approaches are covered by this guidebook. Efficient use of existing facilities. Supplement Urban Forecasts. ♦ All freight modes. Thus. multimodal or intermodal travel forecasting model. Time series and hybrid techniques (time series combined with conventional fourstep models) can be particularly useful for projectlevel forecasting. The study of many of these issues can be facilitated by a good. Connectivity between metropolitan areas. ♦ All passenger travel purposes. and the relationship between the ISTEA factors and travel forecasting is somewhat clearer. statewide forecasts can be helpful to urban area forecasts by providing information on through trips. Those factors that closely relate to travel forecasting are: ♦ ♦ ♦ ♦ ♦ ♦ Energy use. and ♦ Future growth or changes in industry and population. ♦ All freight commodities. such as lumber. ♦ All or many times of day. Such forecasts can be helpful in programming the sequence of projects and their associated costs. The sizing of facilities in the design process requires accurate estimates of future travel. Statewide forecasts most often require a full “fourstep” simulation that considers each of the following elements: ♦ All passenger modes. water and air. TEA 21 mandates that several issues must be considered in statewide transportation plans.modes. Guidebook on Statewide Travel Forecasting 2 . machinery and agricultural products. These factors are consistent with TEA 21. such as commuter. such as auto. business and recreation. Indicators (such as VMT. highspeed rail and air. investigations of statewide or national transportation policies would be incomplete without forecasts on rural highways and other intercity transportation modes. corridor and project. ♦ All classes of facilities. Planners in most sizable urban areas have the ability to forecast traffic levels in their communities. conventional rail. Overview of Approaches There are three different approaches to travel forecasting that are of interest to planners in state DOTs: statewide. and Efficient movement of commercial vehicles. Statewide. intercity bus. rail. Traffic congestion reduction. ♦ Intermodal transfers. Original ISTEA Planning Factors ISTEA was more specific than TEA 21 in its required planning factors.
but the process should not be arbitrarily limited to the capabilities of any particular commercial software product. Overview of Three/Four Step Models Many states trying to build statewide travel forecasting models are doing so using the same theory and software used for urban models. the forecast becomes an exercise in mode split and traffic assignment. The ability of a given software package to model statewide travel is an issue to be considered. There are enough differences between statewide and urban forecasting to require changes to most of the steps. Thus. Often the total demand in the corridor by time of day is considered constant or estimated externally. Time Series Methods A time series may consist of traffic levels. A time series may be of interest by itself or important as an input to a fourstep model. depending upon the policies or projects that are being evaluated.most often linear regression theory. Thus. weekly and diurnal fluctuations in data. time series methods are of greatest interest for project level forecasts. Mode split models of some sophistication are often chosen to give precise estimates of modal shares. A threestep model would not include mode split. The four traditional steps of an urban model are trip generation. Growth factor methods are recommended in FHWA’s Traffic Monitoring Guide for interpolating traffic counts on road segments that are missing data elements for one or two years. A full understanding of BoxJenkins models require a good knowledge of statistics. project level forecasts can be made by extrapolating upon current trends. Moving average methods can be used to eliminate seasonal. but the underlying principles can be easily explained. particularly those with seasonal fluctuations. employment levels or any other socioeconomic or demographic characteristic of interest to forecasting. Other moving average methods can assure that only the most recent data is used in a forecast. ARIMA models are sometimes referred to as BoxJenkins models. ♦ Simple trend models assume that the yeartoyear change is constant. Project level forecasting is often of shorter term with few unforeseen intervening factors. ♦ ARIMA techniques form a class of models for fitting complex time series. Growth factor methods are simple time series techniques that assume that the rate of growth is constant over time. trip distribution. There are other required steps of less importance. ♦ Moving average models attempt to eliminate bumpiness within a data series by averaging a few items that are close together in time. Introduction 3 . differing from growth factor models that assume that the percentage of change from year to year is constant.Corridor/Intercity. Linear regression is a welldeveloped technique for fitting lines to XY data. The choice of formulation depends upon the nature of the data and the eventual use of the forecasts. The purposes of these steps will be explained later. In many cases. This strategy may be appropriate or inappropriate. Project. population levels. Many transportationrelated time series can be best forecasted with a BoxJenkins model. There are many time series methods based on statistical theory . Such a model is typically used for forecasting automobile traffic on highways. Linear Regression and Extensions. Growth Factor Methods. mode split and trip assignment.
some of them being close substitutes for each other. This comparison will not. Adjustment of Parameters. The nested logit model is a means of simultaneously forecasting traffic and patronage on a variety of intercity modes. They can constitute a complete forecasting procedure by themselves or enhance an existing model. Employing such a technique may be better than applying a set of ad hoc adjustments (known as kfactors in the parlance of the gravity model). but can be useful for modifying an existing trip table (perhaps from the gravity model) to better match traffic counts. such as speed and capacity. This method of modeling freight flow parallels the process for modeling passenger movements.particularly gravity model friction factors. Specialized Methods There are several specialized methods to help solve particular problems related to corridor or intercity forecasting. Techniques have been developed to estimate trip tables from traffic counts. because there are fewer attributes. travelers and vehicles rather than individual travelers and vehicles. The “pivot” method uses outputs from a travel forecasting model and from a timeseries model to provide precise forecasts on one or a few highway segments. However. Guidebook on Statewide Travel Forecasting 4 . some states have found it useful to adjust speeds by a formula to account for driver preference for certain routes. For statewide models there is less opportunity for adjustment. by itself. are adjusted to achieve better agreement with ground counts. Nested Logit. Hybrid Technique. The recommended forecasting method makes effective use of available data sources and knowledge of freight flow processes. Adjustment to Traffic Assignment Inputs. The nested logit model has the ability to perform correct forecasts when there are many modes. Commodity Based Freight Models. These techniques cannot estimate a trip table from scratch. Urban models often go through an extensive network “calibration” procedure where various link attributes. Calibration and Validation A major emphasis of this guidebook is on effective calibration and validation of models. “Microsimulation” methods have not been tried at the statewide level and are still considered experimental at the urban level. Base Case Comparisons. the comparison will demonstrate that major relationships have been simulated with some degree of accuracy and consistency. They deal with groups of trips. The forecast is made relative to existing traffic volumes. Refining the Trip Table. assure that future year forecasts are valid. Calibration and validation involve several distinct processes. This guidebook recommends that freight traffic be predicted from commodity flows. mode split and trip distribution . A travel forecasting model has literally dozens of parameters that must be individually set for any given forecast. however.The fourstep models discussed in this guidebook are macroscopic in nature. The process of setting all these parameters is called “calibration”. Most importance to statewide forecasting are the parameters related to trip generation (attractions and productions). for aggregation problems in the definition of the network and for lack of continuity of routes at state borders on the network. once they are created. An important step in any model development is to compare a base case (or year) forecast against known traffic counts.
A high quality forecast will require some original data collection beyond traffic counts.Total Corridor Demand. Data is reported at different levels of spatial aggregation. Introduction to Data Sources This chapter will introduce a few data sources. Stated preference techniques ask travelers about hypothetical modal choices to determine the ridership potential of a new mode. Calibrated model steps provide the greatest policy sensitivity. Spatial Aggregation Issues. Proprietary Data Sources. However. Alternatively. the calibrated mode split model is not helping the forecasting process. Demand within a corridor is often modeled as a function of socioeconomic factors.gov.bts. Original Data Collection. Generally speaking. In that case. Some of the better sources of data are public agencies. lower levels of spatial aggregation result in better forecasts but cause an increase in costs and time for analysis. NTARs and BEA (Bureau of Economic Analysis) regions. depending upon the source. Several companies provide proprietary data. Alternatives to Calibration. it is entirely possible that the state is uninterested in policies that might affect costs or quality. These might include: ♦ Travel surveys of households ♦ Surveys of drivers of passenger and freight vehicles at cordon stations. but most discussion of the use of the data will be handled in later chapters. commodity mode split may be represented by a series of lookup tables developed from historical data. The Census Bureau provides complete person and household data every ten years and data on a wide variety of socioeconomic conditions at other intervals. The BTS is a source of many of the public databases useful for statewide travel forecasting. demographic forecasts and economic forecasts. but private organizations can also provide data for statewide travel forecasting. Introduction 5 . then the effort to calibrate is wasted and more efficient ways of forecasting may be more appropriate. Methods exist for forecasting total demand without needing to run a fullblown fourstep model. Bureau of Transportation Statistics (BTS). Census Bureau. However. Key Data Sources Appropriate use of existing data can speed the development of statewide models. If one were available. Moving data from one level of spatial aggregation to another is often a difficult process. states (outside your state). a good commodity mode split model is very difficult to calibrate. For example. Possible levels of spatial aggregations include: traffic analysis zones (TAZs). counties. municipalities. so the primary spatial unit must be chosen carefully at the beginning of model development. Such look up tables can be tabulated from the Commodity Flow Survey without a great deal of difficulty. Stated Preference. if the policy being evaluated does not use a calibrated model step in a meaningful way. it would be quite useful in determining how commodity mode shares are affected by changes in shipping costs or improvements in service quality. Many of the proprietary databases deal with lower levels of aggregation than public sources. All products are available from its web page: www.
income and location. irrigation activities. land use and ownership. Trips were included for all persons age 5 and older. for all purposes. Additional data collected include cost of materials. mode. Commodity Flow Survey. fuel and energy costs. Basic data from this data set include kind of business. This data set includes: number of establishments (or companies). if drive as an essential part of work if employed and seat belt use. Economic Census. Selected operators provide additional information on production expenses (including interest). market value of land and buildings and income from farmrelated sources. high demand point within the state (single station origindestination survey). reason for the trip. education. Data are reported at the national. revenue. inventories. relationship within the household. Approximately 80. Mining is included in a separate but similar database. ownership. state and NTAR levels. payroll. National Personal Transportation Survey (NPTS) Overview The NPTS is a household travel survey last updated in 1995. gender. Person data include: age. number of employees. annual miles driven if a worker. and ♦ Origindestination information collected at a single. Interviews were conducted approximately every three months by phone and inperson. time of day. Key Census Databases Decennial Census. value of shipments. new capital expenditures. train and bus stations. The CTPP also provides information on the number of employees by zone of employment. payroll and employment. receipts. Federal program payments and operator characteristics. and measures of output (sales. The CTPP provides data aggregated to the TAZ (traffic analysis zone) level as defined by MPOs. worker status. amount of commodity credit loans. value of products sold. length (in miles and minutes). American Travel Survey (ATS) Overview The ATS is a survey of long (greater than 100 miles) trips. stops along the way and side trips from the destination. the access and egress modes to airports. Vehicle data include: annual miles driven. The survey collects data from a random. number of occupants and driver characteristics. Census of Manufacturers and Manufacturers Survey. information about the travel party. location. the principal means of transportation. Trip data include: trip purpose. number of hired laborers.000 households participated. inventories of livestock and poultry. Household data in the NPTS include: household size.♦ Surveys of travelers within a corridor to ascertain their preference for new modes (behavioral intention). fertilizer and chemical use. This data set is derived from a sample of shipments from the US covering most commodities and modes. driver status. hours worked and payroll supplements. make. All operators provide crop acreage and quantities harvested. number Guidebook on Statewide Travel Forecasting 6 . Specialized products from the decennial census have included questions related to journey to work (JTW) and the Census Transportation Planning Package (CTPP). Census of Agriculture. value of shipments or value of construction work done). stratified sample of over 40 thousand households about all personal trips. vehicle characteristics (if a household vehicle was used). number of household vehicles. model and model year. Trip data included: the origin and destination of the trip. machinery and equipment. by all modes.
for many years. Reebie. retrieving. It incorporates data from a “significant number” of truckload and LTL carriers. has provided the TRANSEARCH database and custom data products from TRANSEARCH. Also available in GIS form are jurisdiction 1 2 REMI’s web site Woods & Poole’s web site Introduction 7 . personal income by source of income. Reebie is currently developing a product.1 Woods & Poole. Woods & Poole’s database contains more than 550 economic and demographic variables for every county in the United States for every year from 1970 to 2020. Important Freight Data Sources Commodity Flow Survey (CFS). TRANSEARCH provides traffic statistics between BEA regions by mode (water. Tools Role of GIS A geographic information system (GIS) is a mechanism for storing. rail. employment and earnings by major industry. their size and their income. as well as a wide variety of publicly available databases. visually representing and analyzing spatial data. that disaggregates the TRANSEARCH data to counties for statelevel forecasting purposes. called the Intermodal Freight Visual Data Base. Simulations with the model are used to estimate the economic and demographic effects of economic development programs and transportation policy changes. Also contained in the database are twelve categories of residential and nonresidential buildings and fifteen factors determining demand for new construction from 1980 to 2005. other information related to determining capacity and the types of traffic control at intersections and along uncontrolled road segments. air and truck) and by commodity. These data items include the location of intersections. Many states (and the federal government) are creating GIS databases containing information useful for statewide travel forecasting. Of particular interest here are those data from a GIS that allow for rapid development of a travel forecasting network. and data on the number of households. The CFS will be discussed extensively in Chapter 4. National data provides summaries by threedigit STCC. This database includes population by age. NPA’s database is also aggregated at the county level. The database contains 212 economic or demographic items for the years 1967 to 2025.2 NPA.of nights spent away from home and the type of lodging. the width of road segments. REMI does regional forecasts with a model called Policy Insight. retail sales by kind of business. Raw data are not available. The CFS is the most complete single public source of information on freight flows in the US. sex and race. Route distances of all trips were calculated from a network. state and NTAR data by twodigit STCC. REMI. National data will give distance by mode by threedigit STCC. Commercial Freight Flow Products. Commercial Economic Forecasts Many firms provide economic forecasts that can be useful for statewide travel forecasts.
land use inventories and socioeconomic and demographics information. but many travel forecasting packages have mechanisms for transferring data to and from GISs. This guidebook will not describe measures of effectiveness (MOEs). etc. statistical summaries. they share the same algorithms for computation). A standalone statistical package will also have good time series analysis capabilities. FourStep Model Packages. fails to properly show highway connectivity or continuity. However. Software Issues Many organizations involved in travel forecasting have found it necessary to establish a library of programs for simulation and calibration. Statistical Packages. so specialized products (separate from the fourstep model) are often necessary to accomplish the calibration. especially for project and corridor studies. Mode Split Model Calibration. User Knowledge. which is often the most convenient and transferable method of performing a calibration. or is not digitized with sufficient accuracy. locations of bodies of water. The needs of statewide travel forecasting have not been a priority in the design of these packages. Guidebook on Statewide Travel Forecasting 8 . Models typically come with a graphics network editor to facilitate the input of nodes. Training consists of learning the underlying theory and learning the user interface for the particular software package. These software products are essentially similar internally (i. By and large. tests of significance and singlevariable descriptive statistics. any relevant urban forecasts and several measures of effectiveness (MOEs). Mode split models are calibrated by applying statistical principles to observed travel patterns and mode choices.boundaries. A subset of the capabilities of a standalone statistical package may be found in a spreadsheet package. it cannot totally automate the process of network creation.e. but differ considerably in their user interfaces. This forecast would include both passenger and freight vehicles. depending upon what must be accomplished. a complete travel forecast is not always necessary. Often the GIS provides too much detail.. energy consumption and emissions. There are substantial incompatibilities between GIS data structures and those necessary for travel forecasting. such as traffic counts. Training periods can be considerable. All of these packages require training on the part of the user. links and their attribute values. Some packages come with interfaces to GISs. boundaries of spatial data units. Calibration of other parts of the fourstep model is often best accomplished with a standalone statistical package that contains linear regression. Coding errors that can be minor for typical GIS applications (maps. Several software products are available to forecast urban travel using the standard four steps. travel forecasting cannot be performed on a GIS (there is one notable exception). Even though the existence of a GIS can speed data preparation. including delay. GISs can also store data useful for validation purposes. Composing a Complete Forecast Ultimately. analysis of variance (ANOVA). These statistical methods are somewhat unusual (especially when dealing with the “logit” model and its derivatives). a complete travel forecast should be able to produce estimates of link volumes for all intercity modes.) can be catastrophic to a travel forecast.
♦ Trend analysis with linear regression. Because a time series model exclusively represents past events and relationships. Enforcement needs. A good understanding of elementary statistical concepts and the use of statistical software packages would be required for successful application of the techniques. The chapter is principally concerned with: ♦ Growth factors. Economic trends. All modes can be analyzed with time series. The models can handle more than simple trends (growth and decline). Including factors relating to the state of the system. and Growth in competing modes. Given more time and a broader set of data. Some of the more elementary time series methods require only readily available historical data. Time series analysis is a branch of statistics. including: ♦ ♦ ♦ ♦ ♦ Attributes of traffic. Time Series Methods Introduction This chapter presents several methods of time series analysis that have proven useful in shortterm statewide. such as vehicle occupancy. Time series analysis can be especially helpful for short term forecasts where behavioral models have not been calibrated or input data are unavailable. discrete changes to some important influential factors and trends in important factors. rural or intercity forecasts. They can also consider cycles in the data (annual. and Handling seasonality. vehicle weight and vehicle classes. Policy Considerations. rather sophisticated time series models can be built. Applicability of Time Series Analysis Modal Considerations. What is Time Series Analysis? Time series analysis is a means of understanding data variability over time. so they provide quick answers. daily). but approaches may differ. Environmental conditions. 9 Time Series Methods . Data transformations. and ♦ BoxJenkins (ARIMA) methods. including: ♦ ♦ ♦ ♦ ♦ ♦ Moving averages. it can be used to forecast the future as long as the future is expected to behave like the past. Several enhancements to these techniques are described. Time series analysis is particularly appropriate when the forecast is short term and there is insufficient time and resources to build and calibrate a behavioral model. Time series can be used to forecast data needed for policy analysis. Forecasting differences. Including economic and demographic factors.Chapter 2. weekly.
A strong but spurious correlation can exist between two entirely unrelated variables. Correlation. Time series analysis has a limited ability to anticipate changes in future conditions. barges) when data limitations or time constraints prohibit application of a behavioral model. A typical correlation would describe: Traffic Volume versus Total Personal Income (either over time or over space).g. ♦ Placing bounds on the reasonableness of forecasts. An example autocorrelation might be: Traffic in Year t versus Traffic in Year tn For example. Seasonal effects and other periodicity effects tend to cause lower values of correlation coefficients. Autocorrelations tend to diminish as n increases. monthly or day of week adjustment factors for postprocessing results from a behavioral model. Additional elementary statistical concepts are discussed in the Appendix to this chapter. When doing a multimodal forecast on a network. Of particular importance is the concept of a correlation. the correlation would find the agreement between: Traffic in 1997 versus Traffic in 1992 Traffic in 1996 versus Traffic in 1991 Traffic in 1995 versus Traffic in 1990 etc. some knowledge of elementary statistical concepts is required to understand most of the methods presented in this chapter.. air freight. The sign of a correlation coefficient is the same as the sign of the slope of the line drawn through the points on a (XY) scatter diagram. The correlation coefficient between two variables has only limited use in time series analysis of traffic. and ♦ Determining seasonal. The degree of agreement can be computed for unlike variables (correlation) or for data within the same series spaced at a fixed time span apart (autocorrelation). ♦ Forecasting external travel. A correlation coefficient ranges from 1 to +1. Nor is it possible to model the effects of an existing causal factor that has not changed appreciably in the past. a strong autocorrelation may exist between Traffic in July of Year n and Traffic in July of Year n1 as well as between Traffic in July of Year n and Traffic in June of Year n. ♦ Forecasting comparatively minor modes (e. Events that have never before occurred cannot be modeled. It is always necessary to understand the reasons behind a large correlation coefficient. Great care must be exercised when forecasting traffic volumes. unless cyclic events are present in the data. The concept of autocorrelation is often better for understanding periodicity. time series are useful for: ♦ Forecasting inputs to trip generation. Autocorrelation. Cautions. Like any other model. time series cannot anticipate rare future conditions or events. For example. Important Statistical Concepts This guidebook cannot provide a complete background in statistical theory.Data Considerations. The magnitude relates to the quality of fit to a line. Guidebook on Statewide Travel Forecasting 10 . if n is 5. However.
expressed as a fraction of current traffic levels. as it is downward bending. Growth Factor Relationships For a data series Y. The curve on the right illustrates a typical pattern of growth of traffic on highways in developing parts of a state. Transportation data series can sometimes be accurately modeled by growth factors. the next period t+1 can be forecast from current period t times a growth factor (1 + i): Yt +1 = Yt (1 + i) In this case. Take particular note of the “S” shape of the curve. a growth factor model (with a positive i) will be upward bending over time. which has grown steadily over time. A growth factor is easily computed for any data series. Yt is the traffic in the current year. recent increases in the number of licensed drivers are heavily influenced by increases in population. With only a little more work (and spreadsheet software or a statistical package). Growth factors work best when the variable to be forecasted is heavily influenced by other variables that inherently grow proportionally. For example. Because of the compounding. it is possible to create a growth factor model that incorporates external influences. Time Series Methods Traffic d pe lo ve De De ve lop ing Rural Time 11 . Many transportation variables are heavily influenced by the overall size of the economy. a simple growth factor model will not properly replicate the portion of the curve denoted as “developed”. With a constant fertility rate and death rate and with zero net migration. It is important to select a method of replicating the time series that bends appropriately. A good example of a series that behaves this way is population. In some cases it might be necessary to gather additional information to understand the position of the data series on the “S” curve. N periods can be forecast by repetition: Yt +n = Yt (1 + i) n The above equations behave similarly to compound interest. For example. Growth factors work best on time series where the change from period to period is proportional to the size of the series. i is the growth rate. the number of new people each year is proportional to the number of people already alive. and Yt+n is the traffic in the nth year beyond the current year.Growth Factors The growth factor method is a popular way of forecasting trends in variables that have been (by and large) increasing in time.
Interstate urban. States are encouraged to perform their own cluster analysis to identify reasonable clusters for their traffic data. The previously cited clusters should be considered a rough guideline and a starting point. Linear regression finds the best coefficients by minimizing the sum of the squares of all residuals (misfits of the line to data points). rather than to rely on statewide clusters. If traffic (Y) in year t increases linearly with time. if well chosen. For example. Similar concepts can be applied to other modes that may have strong seasonal or daily variations in traffic levels. These clusters are aggregations of functional classes. Limitations and Site Specific Uses. For example. highway clusters consisting of: ♦ ♦ ♦ ♦ ♦ Interstate rural. Traffic volumes are often highly correlated with socioeconomic and demographic variables. then the following equation should hold: Yt = b0 + b1t + ε t For the linear trend model shown above. b0 (the yintercept) and b1. Clusters can be created to group traffic counting stations that have behaved similarly in the past. tickets for speeding might increase but the traffic would probably hold steady (or maybe drop). should give a better indication of daily and seasonal variations in traffic. at a minimum. even when the correlation is high. most places have a causal relationship between traffic volume and total personal income. It is important to avoid variables that are not causal. Causality. too. Additionally. a surge in hospital admissions would not imply a big increase in traffic volume. a high correlation could exist between traffic on a state highway and the number of admissions at a local hospital. The Traffic Monitoring Guide gives some recommendations along these lines. Such clusters will allow seasonal (or daily) forecasts where seasonal data are unavailable. Other urban. The guide suggests. and Recreational highways. Guidebook on Statewide Travel Forecasting 12 . However. Other rural. The number of speeding tickets could also be highly correlated with traffic volume. Linear Trend Model Linear regression is a technique for fitting straight lines to data. The Traffic Monitoring Guide provides information on how custom clusters may be developed. Linear regression can be used to extrapolate a time series into to the future. Clusters can help span gaps in existing data. so including personal income into the model makes sense (both theoretically and practically). If we were to reduce the speed limit.Clusters. it is import to understand the direction of causality. Personal income forecasts are often readily available. (the slope) are estimated from data using linear regression. the two coefficients. Clusters created for traffic monitoring can be used for forecasting. Spreadsheet software packages come with versatile linear regression modules. The ad hoc set. When looking at data from specific sites it is reasonable to assemble an ad hoc set of similar sites.
this city may have had an almost constant relationship between population and employment over the years. and Rsquare values are provided for the whole equation (see the Appendix of this chapter for descriptions of these statistics). Each independent variable must make a unique. A dummy variable is set to 0 for some of the time and 1 for the rest of the time. but fishing expeditions are never recommended. The explanatory variables. Including Causality. Linear regression can be used to model curvilinear trends by including higher order terms. If this relationship were to change (a recession. Dummy (0. When two independent variables are strongly correlated they are said to be collinear. When the dummy variable is 1. A more robust form of linear regression would estimate a model consisting of a single trend term and several terms for causal variables. for instance). Note that the error term εt makes this equation exact for any time within the series. Again. we might be able to predict the amount of traffic on a state highway by using both population and employment in a neighboring city. Time Series Methods 13 . including two independent variables that previously behaved similarly can distort the contributions of one of them. Yt is the traffic in year t. However. facilities or the environment. Standard errors and tstatistics are provided for each coefficient. special events and step changes to the economy. Choosing Variables. The coefficient of a dummy variable can be interpreted as an additive effect. When a model has more than two independent variables that are strongly correlated. t and x’s. Sometimes. Often people will try every possible variable that might be relevant. When the dummy variable is 0. leading to a model with dubious forecasting validity. then the model is said to possess multicollinearity. Dummy Variables. the coefficient is added to the estimate of the dependent variable. This method ignores causality and the possibility of spurious correlations. There is no single method for determining whether an independent variable should be included in the equation.1) independent variables are quite useful for showing effects that occur only part of the time.The independent variable t can be the exact date or a sequence index from the beginning of the series. then let the software select the set that best explains the dependent variable. then the model may lose its predictive ability. as measured by Rsquare. Furthermore. dummy variables can show holidays. causal contribution to the model. are called independent variables. the coefficient is ignored. the regression analysis might reject a variable that seemed at first reasonable. For example: Yt = b0 + b1t + b2t2 + εt could be used to model a time series that bends down or up. is called the dependent variable. For example. Yt = b0 + b1t + b 2 x 2 + L + bn x n + ε t The variable to be explained. Yt in this case. For example. The causal variables (x’s) may consist of economic or demographic indicators or may represent the state of the system in any given time period. It is important to rely on theories of travel behavior to indicate what those variables might be. The output from a linear regression includes measures of goodness of fit.
then a square root (β = 0. but error can never be predicted. causal variables may be added to a difference model. The recommend procedure for BoxCox transformations in this case is to transform the data series for several values of β. When the standard deviation is constant. the relationship must be untransformed to get estimates of the original series. A BoxCox transformation is sometimes needed improve the uniformity of the standard deviation. Traffic data often behaves this way. which is zero. Thus.Use of Error Terms. Guidebook on Statewide Travel Forecasting 14 . Consequently. fit the data series with linear regression. Thus.5 are frequently used. such as the overall economy of the state. Of course. too. Possible BoxCox transformations are given by this equation. The δ is the yeartoyear difference. then choose the β corresponding to the best model as indicated by its Rsquare. β’s of 0 (giving a logarithmic transformation) or 0. BoxCox Transformations A mathematical assumption of linear regression is that the standard deviation of the errors in estimates is constant throughout the series. A difference model forecasts the change in the data series between two successive periods. A simple difference model takes the following form: Yt +1 = Yt + δ + ε t+1 In words. the errors are said to be homoscedastic. A BoxCox transformation can also be used to model curvilinear trends. all error terms are replaced by the “expected value” of the error. the quality of the model can improve when the original data is transformed to make the errors homoscedastic. This assumption is not always valid for traffic data. Difference Model. a difference model forecasts the next period knowing the value of the data series in the current period. The log transformation is most logical when a constant rate of growth can be assumed and errors increase in proportion to the size of the data item. Error terms are explicitly considered when fitting a model to data. This type of error might be found in traffic data that is affected by single large influencing factor. Once the coefficients of the model are found for the transformed series. When errors may be due to many random influences.5) transformation is often useful. Yβ − 1 β≠0 β Yβ = log Y β = 0 In BoxJenkins (ARIMA) models (to be discussed later) BoxCox transformations are often employed to reduce the effect of the error term in larger values of the data series.
Moving Averages and Seasonality There are two forms of moving averages commonly used for simple time series forecasting: exponential smoothing and unweighted moving averages over a cyclic time period. a moving average for Friday in a weekly cycle would average all data from the Tuesday before to the Monday after. strong seasonal fluctuations in traffic data often become apparent. Exponential Smoothing Exponential smoothing uses a declining series of weights. α α(1 − α) α(1 − α) L 2 so that the newest values in the time series get the largest weights. Thus. then a central moving average is found by averaging the current period with the (n1)/2 periods before and after.α)/α. When a period length of one month has been selected. Observe that the rate of decline of weights for the moving average relates to the value of α. When cyclic patterns are present in the data. Exponential smoothing is best applied to situations where cyclic patterns are not present in the data series. starting with α (0 < α < 1) and declining with the age of data. where n is the number of periods in a cycle. A moving average based on the current period and n1 previous periods can also be used. t – 1. they are often removed by taking a central moving average. Values of α near 1 result in a fast decline. A central moving average uses data to each side of the current time period. The average age of an exponentially smoothed series is (1 . Time Series Methods 15 . then the central moving average for period t is found by averaging together these periods: t. Central Moving Average Good time series models require many data points (50 or more). … t + (n – 1)/2. where there is a seasonal adjustment factor (a multiplicative constant) for each time period in a cycle. Values of α near 0 result in a slow decline. so it is often necessary to adopt a period of less than a year. A monthly index (MI) can be computed to convert the moving average (MA) back to a seasonal forecast: MI = Average of (Data/MA) for month In a similar fashion. When there is an odd number of periods in a cycle. When there is an even number of periods in a cycle (months in a year or seasons in a year). Both forms of moving averages may be applied to either trend or difference models. t – (n – 1)/2 and onehalf of t + n and t – n This method produces a seasonally adjusted trend. t + 1. seasonal adjustment factors can also be computed to convert raw monthly data to a yearly average.
Major commercial statistical packages contain BoxJenkins routines. Moving Average (MA). The choice depends on the data series to be analyzed. In an AR model. Integrated. Therefore. prior periods. In an MA model. an AR(1) model looks like: Yt = δ + ϕYt −1 + ε t when δ (delta) and ϕ (phi) are statistically estimated parameters. An AR(2) model would take portions of two previous periods and look like: Guidebook on Statewide Travel Forecasting 16 . It is not possible to perform an interesting BoxJenkins analysis on a spreadsheet. perhaps. ARIMA BoxJenkins models consist of one or more of the following elements: ♦ ♦ ♦ ♦ Autoregressive (AR). For example. the value of the data series is estimated with one or more earlier values of the data series. Statisticians want 50 or more data points for BoxJenkins (ARIMA) methods. Some of the most interesting BoxJenkins models combine the three types. Autoregressive. An I model estimates the difference in data values in the series or the difference of differences. There are many variations of BoxJenkins models. so this class of models can be quite robust. ARIMA. Monthly. ARMA. Integrated (I). The best predictor of a period is the immediate past period. BoxJenkins models are fit to data so as to minimize the sum of squares of the residuals. Autoregressive (AR) Models AR (Auto Regressive) models forecast period t knowing the conditions in period t1 and. Moving Average. Again. MA or IMA. or (even) hourly data should produce satisfactory results. the data series is estimated using knowledge of the error in a recent estimate. just as in linear regression. daily. The theory underlying BoxJenkins models tells us that the absolute value of ϕ is always less than 1. The premise of an AR model is that time series data rarely takes wild jumps or dips.Introduction to BoxJenkins (ARIMA) Methods BoxJenkins is a large family of time series models that are able to track very complex historical patterns. and Combinations: AR. In this model δ is a constant period increment over a ϕ portion of the previous value. many of the statistical concepts of linear regression also apply to BoxJenkins models. It is possible to include causal variables in a BoxJenkins model. so they are probably not applicable to analysis of yearly traffic data. εt is an error term (which is not used in a forecast). The examples in the next subsections illustrate the most elementary forms of these models.
For an MA(1) model the absolute value of θ must be less than 1. that is. it cannot be known when forecasting more than one period beyond the present. suggesting a correction effect. perhaps. The coefficient µ is the estimated average value of the series. an MA(1) model looks like: Yt = µ + ε t − θε t−1 where µ (mu) and θ (theta) are statistically estimated coefficients. Thus. For example. Time Series Methods 17 . an AR model is easier to compute and understand than a growth factor model. the percent change in traffic given a one percent change in employment. prior periods. Moving Average (MA) Models MA (Moving Average) models forecast period t knowing the error in the t1 period forecast and. Economists sometimes apply a logarithmic transformation to their independent and dependent variables within integrated models. log(Yt ) − log(Yt −1 ) = δ + b(log(E t ) − log(E t −1 )) + ε t The b coefficient is interpreted as an elasticity. Integrated models work best when the longterm trend is stable but the variation from period to period contains strong random influences. The power of an integrated model is best seen when combining it with AR and MA models and when including causal variables. Logarithmic Transformations of Integrated Models. a pure I model looks like: Yt − Yt −1 = δ + ε t The δ term is the estimated average change between periods. For example. Once the coefficients have been estimated. The coefficients thus obtained can be interpreted as elasticities. For example. an integrated model that predicts traffic (Y) as a function of employment (E) might look like this. they forecast the change in the logarithm of the data series. but the contribution of each successive term tends to weaken with age. Moving average models include past errors to get the estimate of values in the data series. Pure integrated models tend not to be very interesting. In the above formulation.Yt = δ + ϕ1Yt −1 + ϕ 2 Yt −2 + ε t There is no theoretical limit as to the number of terms in an AR model. The negative sign preceding the moving average term is traditional. the additional error term is used to get a good estimate of µ. Integrated (I) Models I (Integrated) models forecast the change from one period to the next.
An MA(1) model is characterized by slowly dying partial autocorrelations and an abrupt cutoff of the autocorrelations after the first lag. because one or more of the MA independent variables requires an estimate. For example. such as capacity changes. Here is one of many models tried. such as Labor Day and deer hunting.e. such as “autocorrelations”. Apply knowledge and intuition. Specify model. A partial autocorrelation is a measurement of the improvement to the model by the next term representing the (n1) period when period n and all later periods have already been included.. Extensions to ARIMA (Examples from WisDOT) In a study of the use of time series models for project level forecasting. Special software is needed to estimate an ARIMA model. deal with differences)? How many moving average terms must be included? How many autoregressive terms must be included and what are the lags? Do the periodtoperiod variations in the data suggest that the data series should be transformed? An autocorrelation is the correlation between Y(t) and Y(tn) where n is a selected number between 1 and the size of the time series. the Wisconsin Department of Transportation (WisDOT) tried many variations on ARIMA. Traditional linear regression methods cannot handle this situation. and Analyze strength of model and individual components. Get indicator statistics. added step functions. but these rules are complex and beyond the scope of this guidebook. an AR(1) model is characterized by slowly dying autocorrelations and an abrupt cutoff of the partial autocorrelations after the first lag. The major choices are: Should the model be integrated (i. a quick glance at the data series is enough to judge the best combination of elements. Dependent Variable: Log of daily traffic I94 near Johnson Creek Guidebook on Statewide Travel Forecasting 18 .An MA(2) model would look like: Yt = µ + ε t − θ1ε t −1 − θ2 ε t −2 Steps in Building an ARIMA Model The following five steps should be followed when building an ARIMA model: ♦ ♦ ♦ ♦ ♦ Plot and study data. transform data and fit. and added impulses. Intuition and common sense are necessary for creating a good BoxJenkins model. such as total personal income. Any series has many autocorrelations. It added socioeconomics. Usually. Statisticians have developed many rules of thumb to help make their choices. Statisticians use autocorrelations and partial autocorrelations to help determine the structure of a BoxJenkins model.
Time Series Methods 19 . Brown and Watkins looked at the differences of logs. Similar models can be built to pick up weekly. An inspection of the model reveals: ♦ Dropping fare implies increasing travel.119∆ log(Y) . This model contains two causal variables. A typical monthly forecast looks like: Yt = δ + ϕ1Yt−1 + ϕ12 Yt −12 + ε t The model says that the next month can be predicted from the current month and from the same month last year. will have less travel (or a declining rate of increase).Independent Variables Included: Log daily traffic previous day Log daily traffic previous week Log daily traffic previous year Impulse dummy variables for deer hunting season. Air Travel II: Oberhausen and Koppelman (1982) Oberhausen and Koppelman developed a rather elaborate AR model of monthly air passenger travel for a westbound trip to Hawaii. 12/3/90 and 12/15/90 (outliers) Moving average from previous day (lag 1) Autoregressive term from previous week (lag 7) Cyclic Patterns in AR Models It is relatively easy to model cyclic patterns (of period n) by including an autoregressive term for n periods ago. fare and income.0. This is a traditional linear regression model. otherwise. ♦ Increasing disposable income implies increasing travel. Labor Day weekend. In this equation ∆ (delta) indicates differences between successive periods. ♦ Future years. or (even) hourly variations in traffic levels. and one trend term. daily. real disposable income per capita Y and time in years t. Such a model will pick up regular monthly variations in traffic.038 log t + ε in which air passenger miles per capita T is a function of average fare F. July.1.307 ∆ log(F) + 1. Memorial Day weekend and Thanksgiving weekend Count of days from 1/1/1983 Impulse dummy variables for 12/15/87.0725 . It is conceptually similar to an I (integrated) model. so that the coefficients on the causal variables can be interpreted as elasticities. as it does not contain any autoregressive or moving average terms. Case Studies Air Travel I: Brown and Watkins (1968) One model from Brown and Watkins took the form: ∆ log(T) = 0.
0.0. BoxJenkins software does not automatically select the best combination of terms. but this model contains a particularly large number of autoregressive terms. resulting in models that are sometimes less than perfect. The multiple months at yearly intervals seems to be accomplishing a sort of moving average. The yearly cyclic pattern in the data is evident by inspecting the various terms. The largest coefficient is at lag 12. L is thousands of licensed drivers. The independent variable is passenger volume (not a difference).36Yt1 + 0.0.7Ft1+ 297.12D .62Yt12 . even though the two independent variables are strongly correlated. Other formulations in the ARIMA family might have produced a more elegant model. both the number of licensed drivers and the gross state product must be forecasted. Passenger travel also seems to be influenced by a few other months about a year earlier and two years earlier. Note the absence of any trend terms.0. The signs of coefficients for both independent variables are intuitively correct. one full year ago. All autoregressive terms have coefficients less than one (in magnitude).0.34Yt2 + 0. Maine fit its VMT model with linear regression.14Tt25 .995).22Yt13 .078C 1 Coefficients in the above equation have been rounded from the original model. and G is gross state product in billions of dollars. It includes a trend term and three causal variables. Judgement must be applied. It is difficult to second guess these authors without acquiring and analyzing their original data. properly reflecting the concept of travel being a derived demand.38Yt24 .92 + ε where Ft1 is the twoway fare and all other independent variables are earlier points in the time series.08G + 0. as is typical for AR models. HC = 28000 + 15Y . The model was created with linear regression. Traffic Levels I: Maine VMT A pure causal variable model was created by the State of Maine to estimate total VMT in the state. and they did not use any moving average terms. Guidebook on Statewide Travel Forecasting 20 . VMT = 15L + 332G .Yt = 0. Unfortunately. The model fits the data almost perfectly (Rsquare = 0.0. In order to use this equation.1 Traffic Levels II: New Mexico Heavy Vehicle Traffic This example from New Mexico forecasts heavy commercial vehicle traffic on a single road.90.9600 where VMT is in millions of vehicle miles.13Yt26 .21Yt14 + 0. Maine could do this themselves (perhaps with separate time series models) or by obtaining them from outside agencies or commercial services.
of course. The overall fit was good (Rsquare of about 0. The normal distribution underlies much of the theory behind time series analysis. due to the strong correlations between independent variables. In time series work. The large negative yintercept is due to using a 4digit year in the trend term.. Standard Deviation and Associated Statistics. A period can consist of a whole year. is deterministic (i. Statistics that help understand accuracy include the tstatistic and Rsquare. Random variables are expressed as a list of numbers. traffic volumes on STH 43 might be given the variable Xi where i is the position in the list. D is US disposable income. The mean is the most probable value of a random variable. the category with the largest number of samples is the “mode”. X1 is the traffic in period 1 and Xn is the traffic in period n. The model exhibits problems of multicollinearity. The statistical analysis allows someone to forecast without further consideration of the randomness in the data. a month. the standard deviation. When data is categorized. Mean and Other Similar Statistics.8). a day. etc. The “median” value has half the samples above it and half the samples below it. The normal distribution is symmetrical about the mean. statistical analysis is appropriate. Time. Time Series Methods 21 . or it could account for some aspect of consumer prices. the coefficient of variation and correlation coefficients. and it is estimated by taking a simple X average of samples.where HC is heavy commercial traffic on I40. The negative sign for gasoline costs (G) seems reasonable.1 Appendix: Some Elementary Statistical Concepts Time series analysis is inherently multivariate. G is US gasoline cost. Typical statistics that describe data include the mean. The disposable income term (D) should be positively related to heavy vehicle traffic. For example. and it allows that person to understand the accuracy of such a forecast. a week. Y is year. Because of the randomness in the data. with a subscript denoting the position in the list. The standard deviation is a measure of the dispersion (or spread) of the distribution. Probability of 1 Unit Normal Distribution. The data item to be forecast (which behaves randomly) is related to other variables. It may indicate the health of the New Mexico economy. but has a negative sign in the model.e. not subject to random fluctuations). About 68% of the area under the normal curve occurs σ 1 Coefficients have been rounded from the original model. the lists are always ordered: for example. The median is especially useful in determining central tendency when there are a few really strange samples that distort the mean. The term for cost of residential construction (C) is difficult to interpret. an hour. and C is New Mexico’s cost of residential construction. Any event that is influenced by a large number of random disturbances tends to be normally distributed. some of them behaving randomly.
A standard error is similar to a standard deviation. t statistics become larger as more samples are included and accuracy improves. The tstatistic for a model is found by dividing the value of a model coefficient by its standard error.96 A tstatistic larger than 1. 19 out of 20 times the coefficient will have the given sign (plus or minus) when a new sample is drawn each time. a mean or a constant in a model) that have been computed from many samples of data. The tstatistic is an output of regression analysis and similar techniques. Furthermore. There must also be good reasons for its inclusion from knowledge of travel behavior. It is the most often used measure of the quality of a model. An adjusted Rsquare gives a better indication of which of several alternative models is best.g. You should refer to a good text on statistics for more information on the ttest. About 95% of the area under the normal curve occurs within 1. A ttest is also used to interpret the quality of an individual term in a time series model. A significant tstatistic does not imply that the value of the coefficient is correct. A typical use of a ttest in traffic engineering is to determine whether the mean speed after a change in the traffic environment (enforcement. The ttest was developed to determine whether a statistic computed from a sample differs from a similar statistic computed from another sample or differs from some predetermined value. s. As a rule. Rsquare is the square of the correlation between the data and the estimate. The formula above for the tstatistic shows how it is computed and interpreted when estimating coefficients of a model. Sometimes it is useful to adjust Rsquare for the number of coefficients in the model. can be calculated by this formula: s= 1 n 2 ∑ (x i − x ) − 1 i−1 n TTest. RSquare. The sample standard deviation. The analyst must look at the standard error of the coefficient to determine the accuracy of the term.) differs significantly from the mean speed before the change.96 standard deviations of either side of the mean. The tstatistic is computed somewhat differently when comparing means of two samples. geometry. a significant tstatistic does not by itself justify including a term in a model. That is. { Guidebook on Statewide Travel Forecasting 22 .96 usually (with a sufficient number of data points) indicates that the coefficient is significantly different from 0 with 95% confidence. Rsquare is often expressed as a percent and called “percent of variance explained”. Residual A “residual” is the vertical (parallel to the axis describing the data series) deviation of a point in a data series from its estimate. etc. It ranges between 0 and 1. A term consists of a model coefficient and a variable. but relates to the dispersion of parameters (e.within one standard deviation of either side of the mean. A significant tstatistic is often taken as evidence that the term is useful in explaining the data series.. The square of the standard deviation in called the variance. It is analogous to the signaltonoise ratio for the statistic. t= b sb > 1.
The coefficient of variation reveals the compactness of a random variable. Coefficient of Variation.Rsquare can be calculated by comparing the standard deviation of the residuals to the standard deviation of the time series. V=σ x Time Series Methods 23 . Comparatively small residuals result in a large value of Rsquare. It compares the sample standard deviation to the size of the mean. as shown in the equation below.
this chapter will review the urban procedure and describe how it has been adapted to statewide travel forecasting. In many ways. notably Michigan and Kentucky.Chapter 3. Recent Federal legislation. More advanced passenger travel forecasting techniques are described in Chapter 5. Many of the steps described here will reappear in Chapter 4 (freight) and in Chapter 5 (advanced passenger). Passenger Forecasting Introduction Statewide passenger travel forecasts have been performed for many years in some states. Primarily. Borrowed Methods Many states have implemented passenger travel models that are derived from urban models. has highlighted the need for better statewide travel forecasts. but the models are not as fully developed as those used in urban forecasts. Thus. They perform many of the same steps with much the same mathematics. A few states. the types and frequencies of trips. Well understood theories of travel choices. Disadvantages of UrbanLike Models However. for example: Guidebook on Statewide Travel Forecasting 24 . alternative methods of performing the same step are presented. Many refined algorithms. The chapter first describes the differences between a statewide model and an urban model. and Many active users to share knowledge. other differences are subtler. Advantages of UrbanLike Models Building statewide models on the foundation of urban models has these advantages: ♦ ♦ ♦ ♦ Availability of commercial travel forecasting packages. An appendix at the end of this chapter discusses the basics of urban modeling. However. ♦ Urban models may be overly complex for many intercity applications. available modes and data sources. Besides the obvious differences in scale. statewide travel forecasting is more difficult than urban travel forecasting. as implemented in statewide travel forecasts. In some cases. Corridor analysis is also briefly discussed. Many existing urban algorithms are deficient for statewide forecasts. have had long experience with statewide forecasts. Each step of the forecasting process. The purpose of this chapter is to introduce the standard passenger forecasting procedure as it relates to statewide travel forecasting. so many states are now improving their models or developing new ones from scratch. is described. This later chapter presents highly specialized and experimental methods. the majority of states have not maintained statewide models. including ISTEA and TEA 21. this chapter concerns intercity forecasting. and the need for accurate forecasts has been less compelling. building statewide models from urban principles carries some disadvantages: ♦ Convenience of the software may deter implementation of better methods. and ♦ Data requirements can be burdensome. as well as means of collecting and processing statewide travel data. Differences include the scale of the analysis.
minor arterials and a few collectors. By contrast. the number of intrazonal trips is usually kept small in urban area forecasts. Statewide networks do not vary TAZ sizes as much as urban models. Michigan’s statewide model. being smaller near CBDs and larger in suburban areas. Large networks and large numbers of TAZs cause slow calculations. Processing Time. ♦ Sizes of some zones lead to very coarse traffic assignments. The number of TAZs tends to increase with urban area size. Statewide models can have a large number of intrazonal trips in TAZs within urban areas. Urban models have highway networks that typically contain all freeways. Statewide Models Network Detail. Passenger Forecasting 25 . has 2300 TAZs and Kentucky’s model has 1500 TAZs. so that timelines and computer resources can be kept to a reasonable level. Urban v. for example. Many urban models show freeway segments as pairs of parallel oneway links. It was last performed in 1995. Traffic Analysis Zone (TAZ) Size and Number. This includes the National Personal Transportation Survey (NPTS) from FHWA and the Census Transportation Planning Package (CTPP) and the Journey to Work (JTW) survey from the Bureau of the Census. Both Michigan’s and Kentucky’s networks and zone systems extend considerably beyond their own borders. Small cities tend to have small TAZs. It is a 42. Urban areas may be aggregated into just a few TAZs.000 sample national survey of travel behavior conducted using travel diaries and phone interviews. Intrazonal Trips. TAZ sizes vary considerably within urban areas. freeway ramps. The details of travel within urban areas are of less interest. but instead provides travelrelated socioeconomic data for all regions of the country and has been prepared by the Bureau of the Census and AASHTO in a format designed for use in the transportation forecasting process. many statewide models greatly reduce the detail within urban areas. Sources of Travel Data A number of data sources used in statewide forecasting are available from federal government sources. large cities tend to have much larger TAZs. but not proportionally. Intercity travel involves different modes than urban area travel.5% sample of households. The CTPP does not provide trip data. major arterials. Statewide networks will often omit interchange ramps and other details related to freeways. Slow calculations constrain the number of trial runs and increase the overall expense of producing a forecast. the zone system must at least cover the whole state. and ♦ Omission of many roads leads to excessive volumes on links near urban areas. The NPTS has proved to be the most valuable of these data sets for forecasting purposes. The related JTW survey provides origindestination data for the homebased work purpose only. Intercity v. Because it is difficult to assign intrazonal trips to networks. Obviously. ♦ Methods for ascertaining the fraction of travel that is intrazonal trips are not good enough for statewide travel. Mode Split.♦ Traffic assignment methods rely on the notion of capacity to achieve good results. based on a 12. Local Travel. Included are all state and Federal highways and a selected number of other major arterials. Statewide models are designed to provide estimates of intercity travel. ♦ Size of networks may overly burden the computational process.
Where intercity rail or bus travel is being modeled. Alternatively. equations based on population and employment figures). Such a national highway network would essentially consist of interstate highways. For evaluation of the statewide effects of a localized system change. These include special surveys and cordon counts. high speed rail) in an attempt to estimate the potential effects of future implementations. Special surveys can be conducted in a number of ways. This field work is often followed up with a mailin or telephone survey.g. Cordon counts or roadside surveys are surveys conducted by stopping vehicles on routes leading into or through a particular area and interviewing their occupants.35 for a roadside survey. It is also possible to include a skeleton national network as a means of linking TAZs in distant regions to the statewide model. it may be desirable to use census tractlevel data for corridor studies when sufficient data are available.50 per home interview sample versus $12. Defining the Scope of the Model Statewide Network Preparation Of course. Two of the most common survey methods are license plate surveys and home interviews. Missouri and Michigan are moving toward nested zone structures to be able to vary the zonal detail according to the specific needs of the analysis. More traditional sources of travel data are also applicable to statewide forecasting. Special generators are used to represent land uses that generate travel. In addition. Guidebook on Statewide Travel Forecasting 26 .Surveys and Counts. the network for a statewide model should connect all areas of the state between which travel is to be forecasted. it is sometimes useful to include modes that do not yet exist in a particular area (e. airplane trips would not typically be included in a model for a short distance intercity corridor. For instance. Any highway or railway segments that are important to the type of analysis being done should be represented by links in the network. California and Texas each have more than 2000 TAZs. it may only be necessary to use data at a county level. but are not adequately represented by conventional generation equations (i. For a statewide passenger network this usually includes interstate highways and all major US and state highways. the networks for some recent models have been extended as much as 200 miles into adjacent states. For example. Addresses of ownership can be easily obtained. This might be especially useful in providing alternate paths for trips coming into the modeled state from distant regions.e. home interviews are conducted with a broad cross section of the population and involve the collection of travel and socioeconomic data by telephone or by mail survey. separate networks could also be included for these modes. Different states have adopted different strategies for defining TAZs. License plate surveys involve observing the license plate numbers of vehicles passing or parked at particular locations. Zone Sizes. Examples include airports. New Hampshire has just over 400. A 1995 Vermont study reported a cost of $17. Similarly. while Rhode Island has nearly 900. Modal Categories The number of modal categories that might be included in a statewide passenger travel forecasting model is heavily dependent on the objectives of the model itself. state parks and universities. Michigan.. The TAZ sizes used in a statewide model should be limited by (1) the purpose of the model and (2) the geographical levelofaggregation of the available data..
automobile occupancy rates.g. the individual networks will need to be “stitched” together into a single network. It is important to check networks that are imported from other sources. transformation of coordinates. Corridor studies should also include: ♦ Travel to work. Network information is also available from federal government sources including the FHWA digital network (which is based on TIGER files). For statewide modeling. high speed rail and airplane. the BTS North American Transportation Atlas Data Bases CDROM and Oak Ridge National Laboratory. and ♦ Other nonbusiness.. For instance. a trip purpose should contain trips that have similar characteristics: ♦ Any trip produced within a purpose must be allowed to travel to any attraction zone. More trip purposes may improve the precision of the model. Trip purposes should be selected to match available data and to help address policies and access alternatives. One particular set of trip purpose categories seems to work especially well for forecasts of intercity travel: ♦ Work related or business. Stitching. Network Structure Sources of Networks. In urban areas. ♦ All trips within a purpose share many of the same decision characteristics: mode split utility coefficients. the various local MPO models). Purposes Statewide and intercity travel forecasts have adopted different sets of trip purposes. a statewide model’s TAZ could consist of a group of census tracts in an urban area and an entire county in a remote area. It therefore may be useful to condense the TAZs from local urban models into largerscale TAZs. depending upon the type of questions being asked. intercity bus. time of day and direction of travel factors and gravity model parameters. renumbering of zones. etc. the network’s level of detail in urban areas is generally less important than in MPO models. since they may contain coding errors that went unnoticed in their original applications or that occurred as a result of translation across different software products. ♦ Recreation/vacation. Statewide highway network data is often available from several sources. Passenger Forecasting 27 . including changes in network data format. Fundamentally. The trip purpose categories used for urban travel forecasts are not necessarily appropriate for statewide travel forecasts. In this case differences between adjacent local models must be resolved. A series of criteria – based on the network sources available and the intended uses of the model – must be developed to guide the “stitching” process. conventional rail. If the statewide network is assembled from a number of sources (e. local MPOs typically maintain network models or GIS databases that can be modified for use in the statewide model. but they cause increases in data collection costs and computation time.A typical categorization of intercity passenger modes would have automobile (driver or passenger).
bus and airline service. mode split is an increasingly important consideration in corridor studies. “oneway” links) and turn prohibitions is generally less important in statewide models than for smallerscale models. while delineating the statewide network structure it is important to include a description of the speeds (to be used in combination with link distances) or travel times associated with all links in the network.Highway Network Attributes Speeds: As with a network of any size. bus company schedule information can be used to determine what subset of the highway network is included in the bus network. Including attributes to describe the directionality of traffic flow (i. it is easier to model mode split for corridor study applications. and airline networks may only be needed for statewide modeling to the extent that air travel is internal to the state or between the state and its neighbors. Due to the more localized nature of corridor studies and to the likelihood that they include travel patterns in major urbanized areas. For rail. This is primarily due to the overwhelming majority of instate trips that are made by automobile and the related difficulties in developing a statistically meaningful mode split model at the statewide level. Other attributes that can be used to describe the highway network are of a more descriptive nature and can be used to examine links according to other categories that might affect their behavior. the use of link capacities (often found in smallerscale models) is also of limited value. Inclusion. For statewide modeling it may be more useful to assess delays to links where congestion is expected (typically in urban areas) or to adjust link travel times rather than to calculate speed reductions based on capacity restraints. any network data imported from other sources should be carefully checked for errors and inconsistencies. Historically. Airport locations and airline schedule information could be used to develop an airline network.e. the Federal Railroad Administration (FRA) has developed a digital railway network for the United States. Meanwhile. Guidebook on Statewide Travel Forecasting 28 . Penalties. This information is critical to determining the shortest travel paths during the trip assignment step. statewide passenger models have had limited success in accounting for nonautomobile modes of travel.. Nonautomobile network data is also available from a number of sources. The use of penalties or delays may also be of value in representing signalized intersections or known capacity problems at intersections or interchanges. This can be used in combination with Amtrak or commuter rail maps and schedules to develop a statewide rail network. Consequently. Since the travel data available for statewide modeling purposes does not usually lend itself to consideration of peakhour conditions. Among the most important nonautomobile networks are those for rail. Airlines tend to be a more specialized case. Similarly. Capacities. As is the case for highway networks. For instance penalties could be assessed at state border locations to represent the reluctance of travelers to cross the state border for work trips. Other Modal Networks Sources. Much information about airline travel is national in scale and proprietary in nature. terrain types or geographical area groups. This includes classification according to count groups. A similar situation could also exist at river crossings and toll facility locations. functional classes. data is more readily obtained (from existing sources or localized surveys) for calibration of mode split models. In a statewide model they may also be used for other purposes.
Employment. Assignment. Household size. the TAZs can be much more coarsely modeled.Zone Systems and Spatial Aggregation Issues The level of aggregation of the TAZs used in the model is based on two major considerations. and the same is true for statewide models. Income. census tracts and block groups. A second consideration is the proposed use of the model. If the model is to be used for small scale planning along an intercity corridor. the model’s structure must accommodate available data. This same effect could be achieved by developing a simple disaggregation scheme for use during traffic assignment and by keeping the TAZ sizes as needed to match the socioeconomic data and model use considerations noted above. yielding results that may more closely resemble the gradual changes in traffic volumes observed on actual roadway segments. Sources. if socioeconomic data is available only at a county level of aggregation – which is not unusual for statewide modeling – there may be little value associated with developing a model using TAZs at a census tract scale. especially for the largescale TAZs. The use of smaller TAZs can have a beneficial effect on trip assignment. Connectors. since its tables are specifically configured to be used in transportation forecasting applications. with smallTAZ models. For example. especially when an allornothing assignment is used. Forecast values for these data are then used in the generation equations to predict the number of trips that will be generated in future years. townships. These include: ♦ ♦ ♦ ♦ ♦ Population. The principal source for socioeconomic and demographic data used in travel demand modeling is the US census. only one connector link may be needed. Socioeconomic and Demographic Characteristics The same sort of socioeconomic and demographic characteristics that are important in urban forecasting models are also important for use in statewide models. In large TAZs. Base year values for these data are used in combination with survey data to calibrate the trip generation equations. the position of the centroid within the TAZ can cause its connector links to be inordinately long and have disproportionately large travel times associated with them. Some typical sizes include (in generally decreasing size) counties. The data describing these characteristics are typically used as input to the trip generation step of an urban model. and Auto ownership. The CTPP is particularly useful. then the TAZs should be small enough to reflect changes in the characteristics of that corridor in a meaningful way. The characteristics of centroid connector links can be very important. Passenger Forecasting 29 . The census. In this case the smaller TAZ size serves to smooth out the assignment of traffic to the network. For analysis of statewide trends. Also. Population and employment forecasts – important for predicting future traffic – are often available from state government agencies or local universities. A first consideration is the level of aggregation of the available socioeconomic data (both current and forecasted). only provides base year data. The boundaries used for TAZs typically correspond to some common level of social or political division. of course. Since this data feeds the model. whereas multiple connectors may be required for larger TAZs.
Trip Productions The number of trips produced in any particular TAZ is usually determined using trip production rates based on that TAZ’s socioeconomic or demographic information. just as they might be included in an urban model. including intrazonal trips. Instead. The time period of analysis must be considered when establishing trip generation rates. the clustering of counting stations and the continuous nature of counting programs also makes it possible to address the cyclical nature of travel on some highways. Guidebook on Statewide Travel Forecasting 30 . More importantly. an important consideration is whether to generate all trips. This will be discussed further in Chapter 5. techniques have been developed to synthesize trip tables.g. Urban models are mostly concerned with forecasting the amount of traffic on a typical weekday or within a single peak hour of a weekday. Cyclical Patterns. as well as seasonal patterns of travel (e. or to generate only the trips that are made between zones. Most importantly. forecasts of the same data are readily available from these same private sources. These effects are discussed in the next sections. based on the existing count data. Model Steps Trip Generation Issues For statewide modeling. It should be noted that the peak hour for a rural road can occur on a Saturday or Sunday due to recreational travel. timeofday effects and directionoftravel information can now be estimated even for rural highway links. In addition. deciding on how to deal with external trips and balancing the trip productions and attractions that result from the trip generation process. small towns and rural areas tend to generate more trips per capita than urban areas. The synthetic trip tables are developed through an optimization procedure. as will be discussed later. or as part of a statewide traffic monitoring system. including both weekdays and weekends. Woods & Poole and REMI. With the increasing sophistication of traffic monitoring programs. For cases where only limited travel data is available. most likely because the available techniques for developing generation equations do not differentiate between intrazonal and interzonal trips. This includes the effects of travel on different days of the week. Existing models tend to generate all trips. Trip Table Synthesis.Socioeconomic and demographic data is also available from a number of private sources. The next sections discuss considerations when developing production and attraction models. Other Data Issues Other issues are mainly associated with data collected as part of a traffic count or survey. Statewide models have often been designed to forecast the amount of travel on an average day. intrazonal travel times are adjusted within the trip distribution step to ascertain the correct number of intrazonal trips for individual zones.. There are also geographical variations in trip generation rates. those patterns related to hunting season or summer vacations). This makes it possible to include these effects in a statewide model. including McGrawHill.
Trip production rates can be developed from federal sources such as the NPTS, CTPP, or JTW data, or from local surveys. Different formulations may be required depending on the amount of data available for a particular state. One state, which had little NPTS coverage, made use of both NPTS and JTW trip production rates for the homebased work purpose in the following fashion:
HBW Rate = (National NPTS Rate) * (Local JTW rate) ÷ (National JTW Rate)
This formulation was used to take advantage of both (what the modelers viewed as more statistically reliable) nationallevel NPTS data and (the more locally available) JTW data. Of course, traditional urban model production rates – such as those provided in NCHRP Report #365 – may also be used as a last resort. Purposes. The choice of trip purposes included in the model has an important effect on the way trip productions are estimated. Although many trip purposes are represented in the available data, there are generally only three purposes that must be modeled for intercity travel: business; nonbusiness; and recreation. It is often necessary to condense the various trip purposes found in the available data into these three categories for use in a statewide model. Some models have included a fourth trip purpose for commute trips. There is a comparative wealth of information available regarding business trips. For example, the JTW was specially designed to provide this sort of information for the homebased work trip purpose. To develop trip generation rates for other purposes, further assumptions are usually necessary, and both the quantity and quality of available data is more limited. In the Michigan passenger model trip production rates are developed through the following process. First, using NCHRP Report #365, generation rates are obtained for the various crossclassification groups. There are 60 different classification groups for the Michigan model, based on five household sizes (1, 2, 3, 4 and 5+ persons), three household income levels (low, medium and high) and four geographical area sizes (small, medium and large cities and rural areas). For example, in a large urban area a lowincome, singleoccupant household will produce 3.7 trips per day. Second, using the proportion of total trips devoted for each purpose from the NPTS, the generated trips are divided accordingly among the purposes. Again, for a large urban area, the following ratios are used for lowincome, singleoccupant households: 0.192 to homebased work 0.160 to homebased recreational 0.404 to homebased other 0.310 to nonhome based work 0.214 to nonhome based other 1.000 (100% of trips accounted for) The resulting generation rates are obtained by taking the product of the two values. For instance:
0.192 * 3.7 = 0.71 trips per day for homebased work 0.160 * 3.7 = 0.59 trips per day for homebased recreational, etc.
The process is repeated for all 60 cross classification groups.
Passenger Forecasting
31
Trip Attractions
Determining trip attraction rates is even more troublesome than determining production rates. The same general methods used to develop production rates are also available to develop attraction rates. For instance, some statewide models have attraction rate equations that have been developed using federal data sources, including NPTS and JTW. Rates from NCHRP Report #365 may also be of value. As with urban models, production rates are usually considered more trustworthy than attraction rates, so attractions are generally adjusted to match productions by purpose at the end of the trip generation step. As is also done for production rates, attraction rates are determined by trip purpose. In this case a TAZ containing a national park would likely attract few business trips, but would attract many recreational trips. Special Generators. Sites like national parks or military bases are sometimes represented in the network as special generators. Special generator sites are often modeled as trip attractors only, depending upon their characteristics; very sitespecific information (from specialized surveys, for example) can be used to determine their attraction rates. If special surveys have not been undertaken, information from NCHRP Report #365 can be used to set the trip attraction rates. For very small sites ITE’s Trip Generation could also be used, but it is unlikely that many small sites would need to be included in a statewide model. Michigan found that the weakest link in the whole model chain was trip attraction, so shortcuts at this stage of the modeling effort are inadvisable. In the Michigan passenger model special generators are modeled as attractors only. Attraction rates used are as follows. Special Generator Type Airports Tourist Attractions Campgrounds State Parks Golf Courses Marinas Motels Hospitals Shopping Centers Source of Rate 1991 ITE Manual MDOT Travel & Tourism 1991 ITE Manual 1991 ITE Manual 1991 ITE Manual 1991 ITE Manual 1991 ITE Manual 1991 ITE Manual 1991 ITE Manual Rate Equation exp[1.368 x ln(reg. aircraft)  0.347] or [104.74 x operations /365] 2 x attendance 0.79 x campsites 0.50 x acres 37.59 x holes (1.891 x berths) + 410.795 exp[0.713 x ln(0.44 x rooms) + 3.945] exp[0.634 x ln(beds) + 4.628 ] exp[A x ln(ksf) + B] where A = 0.756, B = 5.154 for > 570 ksf, and A = 0.625, B = 5.985 otherwise 2.37 x students (for universities) or x students (for community colleges)
Colleges & Universities
1987 and 1991 ITE Manuals
These rate equations are intended to show the types of ways rates can be created. They should not be used for forecasting purposes, as new data sources are available.
Guidebook on Statewide Travel Forecasting
32
External Stations
There are two principal ways to include the effects of external travel in the model: using external stations or using a national highway network. External Stations. External stations are located at the edges of the network and are used to generate trips produced or attracted from outside the area that the network covers. For statewide models, they are usually placed to correspond with locations where interstate highways and other major roads enter the network. Use of external stations requires good information about traffic entering the state at those locations, which should be obtained most readily from survey information. National Network. Use of a national network provides an alternative method of modeling trips generated outside of the local network. For a statewide model, the national network would likely consist of the interstate system and TAZs representing states or groups of states away from the area being modeled. With a national network it is easier to make use of national data (e.g., NPTS) to generate trips from distant states and later to assign them into or through the local network. The use of a national network does not entirely eliminate the need for an external station OD survey, which is still useful for model validation. Both states that use external stations and states that use national networks to help model external trips have found it beneficial to extend the local network into the surrounding states. In some cases this extended statewide network covers areas hundreds of miles into adjoining states. The extended network is used to provide both a more detailed accounting of trips generated in nearby states that may pass into or through the state being modeled and to provide a way of buffering the distribution of trips generated outside the local network and being fed into it through external stations or national network links.
Balancing Productions and Attractions
In general, the number of trips produced will not match the number of trips attracted. This is to be expected, since the numbers of trips produced or attracted are developed by entirely different submodels. Since the production equations are generally considered to be more reliable than the attraction equations, it is common to adjust the attraction values of each TAZ such that the number of trips attracted for the whole model is equal to the number of trips produced. This process is carried out separately for each trip purpose.
Trip Distribution: Gravity Model (Production Constraint)
Gravity models have been extensively used for statewide travel forecasting. A complete specification for one form of the gravity model is:
Tij = Pi A j f (t ij )
where:
∑ A k f (t ik )
k
Tij is the number of trips from i to j; Pi is the production in zone i; Aj is the attraction in zone j; tij is the impedance (time) from i to j; and
Passenger Forecasting 33
time and distance are almost directly proportional to one other. For intercity travel. These balancing factors do not provide any new information to the model. it will be discussed more fully in Chapter 5. Xi and Yj. Friction Factors. f(tij). When there are two or more modes in the analysis. it is often desirable to use a composite time or cost that considers all available modes. so either could be used to calculate friction factors. The denominator of the expression (the sigma term) adjusts the trip table so that productions from the trip distribution step agree with productions from the trip generation step. They are sometimes used for estimating the number of trips going to or coming from an entirely new activity site. as calculated in the trip generation step. Travel time by automobile is commonly used for urban travel forecasting. set the Y’s by Yj = 1 ∑ Pi X i f (t ij ) i Two “balancing factors” have been introduced. tij. set the X’s by X i = 1 ∑ A j Yj f (t ij ) j To conserve attractions. Singlyconstrained models are slightly easier to calculate than doubly constrained models. Singlyconstrained gravity models are not recommended for statewide models because trip attractions from the trip generation step will not be preserved. The number of trips between any production zone i and any attraction zone j is proportional to the number of productions in i. It has the following functional form: Tij = Pi A j X i Yj f (t ij ) To conserve productions. proportional to the number of attractions in j and proportional to the measure of proximity. The trip matrix will not be consistent with the number of attractions. This composite time is not a true average. DoublyConstrained Gravity Model A doublyconstrained gravity model will preserve both productions and attractions. This equation will always yield a trip matrix that is consistent with the number of productions in each zone. A doublyconstrained model preserves both productions and attractions. Friction factors are a function of a measure of zone separation. this form of the gravity model is often referred to as being “singly constrained”. is larger for pairs of zones that are close to one another and smaller for zones that are distant from one another. Thus. their only purpose is to assure that the trip generation results are not changed in the trip distribution step. as calculated in the trip generating step. There are as many of each as there are zones. Friction factors are determined empirically. Guidebook on Statewide Travel Forecasting 34 .f(tij) is the friction factor from i to j. f(tij). The friction factor.
an exponential function.5 Time + 0. or a combination of exponential and power functions. then Xs are found again. a power function. The iteration process is started by assuming that all Y’s are 1.75 Distance + 0. They have three sets: work trips. Friction Factors Friction factors are found as a declining function of time. The process converges rapidly. Michigan handled the problem by having a very complex friction factor function for intrastate trips and an entirely different friction factor function for interstate trips. typically in three to five iterations.441e −0. etc. Outstate trips had still another friction factor function. Friction factors are commonly found from a lookup table. short (≤60 minutes) nonwork trips and long (>60 minutes) nonwork trips. Long nonwork trips constitute about 1% of the total number of nonwork trips. distance. An iteration process is required because each Xi requires knowledge of all Ys and each Yj requires knowledge of all Xs. Long Nonwork 999999 367878 196955 2 2 2 Time (min) 61 181 241 601 661 720 FFactor. tripbytrip.Xs and Ys must be calculated prior to calculating the number of trips between zones. Passenger Forecasting 35 . then Ys are found.123 t + 200t −0. Kentucky used tables of friction factors. Time (min) 1 2 3 118 119 120 Friction Factor. Below are pieces of each table. Kentucky handled the problem by designating a separate purpose for long nonwork trips (>60 minutes). the friction factor function looks like: f (t ) = 1258477t −1. For the homebase work purpose. Long Nonwork 200 50 6 1 1 1 Trip Distribution: Application to Statewide Forecasting Both Michigan and Kentucky had difficulty finding simple friction factor functions that match trip making for both long and short trips. Michigan defined its measure of separation as: t = 0. Michigan also made extensive adjustments to the gravity model. For example. The complexity of the friction factors relates to the difficulty of obtaining a single function that properly reflects both short distance and long distance trip making.565 e −0. cost or some combination of the three. Work 259360 203007 159242 2 2 2 Time (min) 1 2 3 58 59 60 FFactor. Then Xs are found.1 Toll The sum of two “gamma functions” was used for friction factors.012 t They applied this function to all instate trips.
airplane. a considerable effort is required to develop the mode split tables. van. other family or personal business. The gravity models will eliminate intrazonal trips if the friction factor is set to zero (usually by setting intrazonal time to a very large number).with K factors. The best single source of mode split information is the NPTS. which in both cases was similar to urban trip purposes. In fact. the number of intrazonal trips is controlled entirely by the size of the intrazonal trip time. bicycle. went out to eat. change mode of transportation. taxicab. bus. The NPTS trip purpose categories are to work. other social/recreational. other truck.e. return to work. The next sections will discuss multinomial logit and nested logit. z is a dummy index that ranges over all modes and e is 2. Within the gravity model. vacation. so rounding can cause an appreciable error in the estimates of intrazonal trips. The fraction of intrazonal trips is very important to statewide travel forecasting. school bus and other nonPOV. motorcycle. Perhaps a better method of handling intrazonal trips is to exclude many of them from the gravity model.. The number of intrazonal trips could be estimated for certain key zones. Logit The logit model allocates person trips to alternative modes. For intercity travel the tables should break out mode by both purpose and trip length. It does so by comparing the utilities of all alternative modes. medical/dental. Other break outs are may be useful. take someone somewhere. The NPTS modal categories are automobile. other POV (privately owned vehicle). Other mode split techniques are variations on logit analysis. However. sport utility vehicle. each trip is separately reported. Mode Split Mode Split Tables. subway/elevated rail. streetcar/trolley. Pivot point is illustrated in the chapter on freight forecasting. It should not be calculated by a “nearest neighbor” method that is available in some travel forecasting software packages. intercity) trips. Rather. A separate mode split table should be prepared for each trip purpose. walk. depending upon the policies and alternatives to be tested. visit friends or relatives. Still other software packages define friction factors for only whole minutes of trip time. Thus. The proportion of trips made by mode k can be found from: pk = e Uk ∑ eU z z where Uk is the utility of mode k. other and home. pick up someone. commuter train. the NPTS does not provide the data in a convenient form. A large error in these cells of the trip table can result in a large error in interzonal (i. This intrazonal trip time must be selected with great care. pickup truck. shopping. Intrazonal Trips. workrelated business. religious activity. RV (recreational vehicle). Amtrak. The experiences in Kentucky and Michigan suggest that the gravity model is not working well within the trip purpose categorization selected.718281… Guidebook on Statewide Travel Forecasting 36 . school. one of Michigan’s early statewide models omitted intrazonal trips at the trip generation step. then subtracted from trip generation totals.
bus and airline. However. The value of ak for the remaining modes may be positive or negative. This constant is arbitrarily set to zero for one mode. trip cost and trip convenience. Now assume the addition of a third mode. rail. and outvehicle time form low income. the coefficients on time (t) and cost (c) are almost always negative. The utility function can be found from statistical estimation or from transferable parameters. Uk = a k + a1t + a 2 c + L where the a’s are empirical constants. Utility is smaller (more negative) when trips are longer. Utility is usually found from a linear equation that combines the effects of trip time. The modes considered were automobile. The specific functional form is often found from statistical analysis of choices made by travelers. Because longer and more costly trips are considered bad. One of the pure logit models is illustrated by the coefficients in the table below. high speed rail. depending upon how favorably travelers perceive that mode. travel cost and convenience measures. conventional rail and high speed rail have similar characteristics and compete directly with one another. utility is most often a negative number. Logit Example. invehicle time for low income. ak. The IIA property says that the ratio of mode shares between any two modes is unaffected by the presence or characteristics of a third mode. Significant independent variables in the model were: mode specific constants. invehicle time for high income. consider a mode split between automobile and conventional rail. Forinash and Koppelman (1993) Forinash and Koppelman (1993) calibrated several mode split models for intercity travel in Canada. The above equation defines a mode specific constant. It might be argued that conventional rail should lose a higher percentage of its ridership to high speed rail than does the automobile.Utility is a measure of the personal satisfaction of taking a trip. Logit models possess an interesting mathematical property called “Independence of Irrelevant Alternatives” (IIA). exclusive of the satisfaction of reaching the destination. outvehicle time for high income. Utility is often expressed as a function of travel time. others were nested logit (see later discussion). 16% conventional rail and 20% high speed rail (the ratio of automobile to conventional rail is still 4 to 1). which grabs 20% of the ridership for itself. Because travel consumes valuable personal resources. The IIA property is displeasing to many analysts because it can cause a bias toward technologies that have many distinct modes. Passenger Forecasting 37 . The new mode shares from a logit model would be 64% automobile. The data was taken from trips made between Ontario and Quebec. Some of the models were pure logit. frequency. large city dummies by mode. A worse mode would have a negative mode specific constant. After all. the logit model does not recognize differences in competition across pairs of modes. Automobile has 80% of intercity traffic and conventional rail has 20% for a ratio of 4 to 1. For example. A better mode would have a positive mode specific constant. travel cost.
Speed Finally. Low Income Air 1. Low Income OutVeh Time.04053 0. but the model calculates utility somewhat differently than discussed previously.Variable Mode Constant Large City Frequency Cost InVeh Time.. is: Un = φn ln ∑ eUk k φn + µn where Uk is the utility of a mode one level below and φn and µn are constants for the current level. the utility of a nest. The logsum term. a linear combination of time.01382 0. High Income OutVeh Time. specifically: Guidebook on Statewide Travel Forecasting 38 .1022 0. rail and airplane. For Airline Decision Decision example. The statistical estimation process can also give a good indication of the best arrangements of modes and nests. another decision is made between automobile. Nested Logit Nested logit models organize modes into Decision a hierarchy. cost and convenience terms. statistical estimation or adoption of default parameters.1224 Auto 2.000 Canadian. A logit model is applied to each level in the hierarchy. Specifically.02635 The above table lists the coefficients of the utility functions.03265 0. professional judgement is required to assure that the model will give dependable results.7460 Bus 2. The nested logit model has the ability to differentiate between pairs of modes that are complementary and pairs of modes that are competitive. The large city variable is 1 if either end of the trip is in a large city. The hierarchy implies that several decisions are made Auto Rail in the process of selecting a mode.203 1. Modes with similar characteristics are grouped together in a nest. Un. Be aware that not all coefficients appear in every mode’s utility function.g. Some nesting schemes that look good statistically can give strange forecasts. Nested logit models are calibrated in much the same manner as regular logit models. The utility of modes at the lowest level are calculated in the same way as a logit model. like the one shown in the diagram on the right. e.756 0. a decision is first made between SOV and carpool and between High SOV Carpool Conventional conventional rail and high speed rail.888 0. A nested logit model helps the analyst to avoid the IIA property of a pure logit model. High Income InVeh Time. However.1306 Train 0 All 0.003797 0. The breakpoint between high and low income was $30.
Passenger Forecasting 39 .6488. the logsum would form the combined utility of the SOV and carpool modes. the other nests’ bias coefficients are set relative to the arbitrarily chosen nest. Low Income Air Bus 0.02339 0.03009 Decision Airline Auto Surface Mass Bus Rail Auto 0.242 0.3140 0. Nested Logit Example. High Income Total Time. The nest bias coefficient. It raises or lowers the utility of the nest. For example. The outvehicle time term is computed as raw outvehicle time divided by the natural logarithm of the distance.7697 0. µk. is usually between 0 and 1.01217 0. The two formulations are nearly equivalent. They did not include a nest bias term for the Surface Mass utility term when splitting trips across Airline. performs the same purpose as the mode specific constant in the logit model. The coefficient on the logsum term φk is found by matching the model to data. some nested logit models omit the nest bias term.09843 0. to compute the utility of the automobile (refer to the previous hierarchy).03507 0. More F&K Forinash and Koppelman also tried several nested logit models on the travel data from Canada.01303 Train 0 0 0 All 0. The nest bias coefficient is arbitrarily set to zero for one nest at a level. Variable Mode Constant Large City Household Income Frequency Cost OutVeh Time. A coefficient of 1. This combined utility takes into account the increased opportunities to travel by having several modes in the nest.1977 0. One of their nesting schemes is illustrated on the right. High Income OutVeh Time. Auto and Surface Mass. It should be noted that some applications of nested logit omit the φk term where it divides utility in the exponent.φn ln ∑ e Uk k φn is simply a means of combining together the utilities of all the modes in the nest. depending upon the preference of users. on and within the logsum term had a value of 0. The coefficient.7359 1.0 causes the nested logit model to behave as a standard logit model.5756 1. but could produce slightly different coefficients for utility equations of the modes within the nest.1830 0. The exact formulation of this nested logit model differs slightly from the general formulation described in the previous paragraphs. Low Income Total Time. Also.00883 The coefficient. φk.1339 0. φk.
When two statistical variables are proportional in this manner. One such method asks travelers about their intention to take a new mode. travel time is almost proportional to travel cost for most intercity modes. Quite different nesting structures can have almost the same statistical performance on data.Mode Split Issues Lookup Tables. If collinearity is present in the data. particularly by keeping like modes within the same nest. is described in later chapters. the data is instead available only in terms of 24hour flows. Capacityrestrained and stochastic assignment methods have been developed to address two of these reasons. For example. This proportionality is stronger for intercity trips than for urban trips. A traditional logit model is calibrated from data about actual trip choices. and it is inappropriate to take coefficients from urban models. Collinearity can result in strange values for estimated coefficients. all of the trips between a particular origindestination pair are assigned to the single shortest path of links between the pair. For allornothing assignment. When a new mode is planned. Collinearity. Many of the variables in the utility equation are roughly proportional to each other. for statewide modeling the necessary timeofday information is often not available to estimate peaking of traffic. These methods are commonly used in urban modeling situations in conjunction with the BPR equation: t = to [1 + 0.15 (q/qmax)4] . Four possible reasons that travelers do not follow the single shortest path are (1) capacity restraints in the system. choice data for that mode are unavailable. which include iterative and incremental techniques. This makes application of capacity restraint techniques difficult at a statewide level. Urban models typically deal with peakhour demand situations. Transferable parameters have not been developed for intercity mode split models. it is important to apply the calibrated logit model to situations where the same type of proportionality holds. Data Needs. This method. but forecast very differently. For example. Nesting Structure. involve increasing the travel time on links where high volumes of traffic are expected. Guidebook on Statewide Travel Forecasting 40 . All constants in the logit model that are specific to that single new mode cannot be estimated. Lookup tables are particularly useful for splitting trips to minor modes or across finer divisions of modes. It is sometimes possible to use psychometric techniques to estimate those constants. Rather than use a logit model. However. Professional judgement may be necessary. Logit models require extensive data for calibration. referred to as “stated preference”. Mode Specific Constants. a state wants to split trips across SOV and carpool categories. they are said to be collinear. In reality traffic does not always travel along the shortest path. (3) the lack of spatial detail associated with large TAZs and (4) the omission of many smallervolume roads in a highway network. it is important to develop nesting structures that are logical. Trip Assignment The least complicated technique for assigning trips to the links on the network is the allornothing method. Consequently. Capacityrestrained assignments. (2) random personal preference. it might be more efficient to use a lookup table that shows typical splits as a function of distance.
or thirdshortest path. Special generators could be associated with particular zones. Unfortunately. There are no widely accepted techniques for calibrating the traffic assignment step in a statewide model. typically counties. Freight represents a significant proportion of vehicles on roads and an even larger percentage of passenger car equivalents (from the Highway Capacity Manual). A large number of zones implies an increased burden of data preparation and input data forecasting. All too often the validation data is used in the calibration process. A suggested method of reducing this burden is to adopt two levels of spatial aggregation. and a large number of minor roads should be included to provide a reasonable approximation of the actual path choices. small zones and many links increase both data preparation and computational needs. Computer resources are likely to be less of a problem in coming years. Integrating Freight. Thus. Freight is discussed in the next chapter. Data from the districts could be disaggregated to zones using a constant set of factors for each district. Most data preparation and data forecasting tasks could then be performed at the district level. Statewide models should be validated only on links where network detail is reasonably representative of reality. with extensive use of statistical estimation techniques and many trials and errors.Stochastic assignment techniques involve assigning some of the trips between a given origin and destination to the second. It is important to maintain the distinction between calibration and validation. usually assigned traffic levels. Validation should occur rather quickly once the validation data set is assembled. This is done to account for the preferences for certain routes that drivers might have. accurately forecasts current conditions. as needed. it is possible to calibrate a model step with one of them and validate with the other. Assignment to the various paths is done by estimating the probability that a traveler will chose that particular path. The set of validation data consists of traffic counts. Thus. Currently. For example. calibrations made with home Passenger Forecasting 41 . However. independent of travel time considerations and the imprecision associated with large zones or arbitrarily small networks. TAZs should be kept small. Districts could correspond to the spatial units most often used in socioeconomic forecasts. as well as to the shortest path overall. Calibration and Validation Calibration versus Validation. It seems that allornothing assignment works as well as the others. states that have tried various traffic assignment techniques have been disappointed with the results. a full evaluation of highway performance must consider freight. usually in rural areas. Michigan’s model (at 2300 zones) is approaching the practical size limit computationally. Allornothing traffic assignment can produce quite lumpy results when there are comparatively few TAZs and the network omits a high percentage of minor roads. Therefore. Sources of Validation Data. as a whole. it is important to retain some degree of spatial detail when performing allornothing assignments. Statistical methods (linear regression. Calibration is the process of setting the various model parameters of the various model steps to match existing trip making behavior. When two or more duplicative data sets are available. Validation is the process of determining whether the calibrated model. districts and zones. thereby undercutting their value for validation. calibration is normally a lengthy process. maximum likelihood estimation) are preferred calibration techniques for all steps except traffic assignment.
Find disutilities for nonhighway modes. all roads with less than 2000 vehicles per day (AADT) should be excluded from the validation. a three step model consists of trip generation. Already prepared networks can be obtained for many statewide and corridor applications. Activity allocation. When strictly highway policies are being analyzed. Trip generation. it must connect to one or more links. Vehicle occupancy. trip distribution and traffic assignment. A typical modeling sequence might involve: ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ Find highway trip times. they should be combined into a cutline. Networks are drawn with a specialized program. Cross validation data may be reported at a different level of spatial aggregation than the original calibration data. Highway traffic assignment. ♦ For a node to be useful. A network holds nearly all data used in a travel forecast. Highway volume averaging. Guidebook on Statewide Travel Forecasting 42 . the mode split step is often omitted. Since the quality of traffic counts is poor on low volume roads. Trip assignment to nonhighway modes. Highway delay calculation. Trip distribution. In actuality. Trip Generation Trip Distribution Mode Split Traffic Assignment Feedback NODE LINK Network Basics Networks. Appendix: Overview of Four Step Models The four major steps of the urban transportation modeling process are shown below. there are many more steps to the modeling process. called a network editor. If a large number of such lowvolume roads exists in a portion of a state and a validation is important. Some of these steps are quite important.interview data could be validated with CTPP or NPTS data. Networks must satisfy two important continuity requirements: CENTROID ZONE ♦ All links must connect to two distinct nodes (referred to as “A” and “B”). Thus. or they can be taken from a GIS. and Feedback. The steps are executed sequentially. it assures that congestion and delay are treated consistently throughout the model. A network consists entirely of nodes and links and the values of their attributes. How good? Models need not be any better than the quality of the traffic data in the validation data set. Time of day and direction of travel. A feedback loop is now considered best practice.
where the dependent variables are demographic characteristics. A link is a straight line segment representing a portion of a travel path. end or change direction. External stations represent large areas beyond the region of interest. Often links are identified as being oneway or twoway. called a centroid connector. Trips are attracted to where the trip purpose is satisfied. All homebased trips are attracted to the end that is not the home. travel time. A path is a sequence of nodes and links. TRIPS ZONE PRODUCED usually over a 24hour period. nonretail employment and households. nonhomebased TRIPS (NHB) and specialty purposes. Path. Productions. Trip totals for any zone are separately tabulated by trip purposes and by whether they are productions or attractions. trips begin or end at a special type of node called a centroid. Nodes. including distance. These relationships include the socioeconomic and demographic characteristics of zones. Many travel forecasting models require that centroids be connected to the arterial portions of the network with a special type of link. number of automobiles in the household or household income. the models identify the shortest path by some criterion. Trip Generation Basics Overview. Nonhomebased trips are treated in a variety of ways. The trip generation step determines the total number of person trips that begin or end in a zone. Trips identified as being ATTRACTED homebased have the home as either the origin or destination. but rarely include any travel characteristics of zones. Links have attributes. such as retail employment. Traffic on oneway links flows in the A to B direction. homebased other (HBO). Passenger Forecasting 43 . typically time or distance or some combination of the two. Many urban models generate a total amount of nonhomebased trips. Attractions. The following rules apply within most urban models: All homebased trips are produced at the home.Links. but centroids contain attributes that describe the socioeconomic and demographic characteristics of the zone. Most nodes represent intersections or bends in a road. A typical set of trip purposes for urban travel forecasting is homebased work (HBW). Attractions are most often calculated by a linear equation. Centroids usually represent areas of the city identified by travel analysis zones (TAZs). In a traditional urban model. One type of centroid is an external station. starting at an origin centroid and ending at a destination centroid. Whether a trip end is a production or an attraction depends on how the trip purpose is satisfied. Trip productions and attractions are computed with separate empirical relationships. Usually. then split them evenly between productions and attractions. speed and capacity. Productions are most often calculated by multiplying a trip rate for a category of households by the number of households falling into that category. Nodes are primarily used to represent places where trips begin. Categories are typically organized by number of persons in the household. Intersection nodes do not have many attributes. and The networkwide total of productions must equal the total of attractions for each purpose. Trips are produced where the trip is conceived. There are many paths between most origins and destinations.
The measure of trip goodness is called “utility”. which is either ignored or treated as a fixed percentage of all trips. 1 Trip Distribution Technique. Usually. with the first subscript representing the production zone and the second subscript representing the attraction zone. train and airline. In the trip table shown at the right. P’s 2 3 183 890 920 784 141 748 939 776 197 ♦ There is a large number of trips between pairs of zones when the trip productions in the ♦ There is a large number of trips between pairs of zones when the trip attractions in the pair are large. Each trip is identified by subscripts. as there are many personal factors and perceptions that influence each traveler’s decision. there A’s are 776 trips between production zone 2 and attraction zone 1 2 3 3. and ♦ There is a large number of trips between pairs of zones when there is little separation between zones (as measured in time or space). major objective factors in the choice. Trip distribution is most often performed in urban area models by a technique called the gravity model. Mode split models assume that travelers choose the best mode for themselves by weighing the characteristics of the trips for all available modes. The principles of the gravity model are quite straightforward: pair are large. Each row represents a production zone and each column represents an attraction zone. yielding 6 ZONE 1 interzonal trip interchanges and 3 intrazonal ZONE 3 trip interchanges. Trip distribution finds the number of T12 person trips that go between all pairs of zones. The results of the trip distribution step can also be shown in the form of a matrix. the value of utility is most often a negative number. Since trips consume valuable resources.Trip Distribution Basics T22 Overview. 24hour person trips are distributed ZONE 2 separately for each trip purpose. The diagram on the right shows three zones. However. mode split is typically performed on 24hour persontrips. Mode Split Basics The mode split step finds the number of trips using each available mode between a production/attraction pair. can be ascertained with some degree of accuracy. The direction of each arrow is from the production end to the attraction end of the trip. such as travel time and cost. The actual utility is unknown. AUTOMOBILE TRAIN AIRLINE The mode split step is most often found in forecasting models for large urban areas. The mode split model Guidebook on Statewide Travel Forecasting 44 . The figure on the right illustrates three intercity modes: automobile. As with trip distribution. Small cities tend to have low transit ridership. with as many rows and columns as zones.
an evening homebased work trip will have its origin at work and its destination at home. the trips must be converted to vehicles.. This assignment is done separately for each OD pair. Passenger Forecasting 45 . peak hour) and given a physical direction of travel (i.takes these objective factors as inputs. Some mode split models include socioeconomic factors in their expression of utility. Traffic Assignment Basics The major outputs of the traffic assignment step are volumes on all links.e. For all modes. the trips must be factored into a specific time period (e. but recognizes that the actual utility is known imprecisely when calculating mode shares. origin to destination). Still other mode split models take into consideration factors related to access and the lack of one or more alternative modes for subgroups of travelers. then the results for all OD pairs are summed. a morning homebased work trip will have its origin at home and its destination at work. Trip Table Preparation for Assignment. Two of these variations are discussed in the main portion of this chapter. Allornothing assignments do not recognize the capacity of individual links or nodes. Capacityrestrained equilibrium assignment recognizes delays that may occur due to congestion. Algorithms that perform this form of assignment try to spread trips from a single pair of origins and destinations over many paths such that all used paths are shortest (or tied for shortest) and the delay along any link or at any node is consistent with the amount of traffic on the link or passing through the node. SHORTEST PATH CapacityRestrained Equilibrium. turn restrictions and turn penalties at intersections. A stochastic assignment finds the shortest path between an origin and a destination and also finds several alternative paths that are somewhat longer.g. Stochastic. but can account for known delays on links.. Conversely. For automobiles. Most urban models containing a mode split step use a variation of the logit model for mode split. The output of the mode split step is still in terms of 24hour person trips in the production to attraction direction. For example. Allornothing assignment loads all trips between an origin and a destination to the shortest path between the origin and destination. AllorNothing. The algorithm then assigns the largest fraction of trips for that origindestination pair to the shortest path and lesser fractions to the longer paths. turning movements at selected intersections and the delays from congestion.
specialized services). Existing and projected commodity flow. and Variation in type of service (e. Because there are a very large number of commodity categories. Factors affecting freight demand are complex. other local knowledge and experience gained elsewhere.g. Technological characteristics of modes. 2. Inputoutput (IO) tables can bridge the relationship to consumption of commodities. Networks are required for freight models. Origin and destination data for commodities must be obtained to build a factual model. Find Base Year Commodity Flows. Thus. Networks are drawn from scratch or modified from existing sources. Develop Commodity Groups. Kentucky and Kansas. A large number of groups gives precision. Networks consisting of mathematical descriptions of routes. Potential actions of carriers and shippers. Vehicle traffic levels are derived from the movement of commodities. Local data sources would most often have other dates of completion. the Commodity Flow Survey was last performed in 1997. Multiple measuring scales for goods. Michigan. 3. Obtain Freight Modal Networks. involving: Presence of factors beyond the planner’s control. commodities are grouped to facilitate the analysis. links and intersections are required for each mode receiving complete analysis. a good understanding of commodities is necessary. Forecasting requires extensive use of existing data sources (commercial or public). and Modal networks. The methods follow a process that is similar to fourstep models for passenger forecasting. 4. Indiana. Freight Forecasting Introduction The purpose of this chapter is to present a unified methodology for forecasting commodity flows and vehicular volumes on intercity and rural portions of our transportation systems.. For example. It is easier to relate commodities to industrial indicators at the production end of the shipment. Methods presented here have been adapted from forecasts done for Wisconsin. Separate economic indicators should be adopted for production and consumption of each commodity. These steps are as follows: 1. Guidebook on Statewide Travel Forecasting 46 . Relate Commodity Groups to Industrial Sectors or Economic Indicators. Variety in type and value of commodities. These data are collected irregularly. Freight forecasting is dependent upon knowledge of: ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ Industrial structure of the economy. Freight Model – Basic Steps Building a freight model involves several steps that will be discussed in detail later. but details vary considerably due to the unique aspects of commodity movement. but increases the complexity of the modeling process.Chapter 4.
♦ Expert Opinion: Tables or models may not be available for entirely new modes or services. The cost of any given shipment is complex. intersection and terminal. This information is provided by a traffic assignment algorithm. Planning studies often require model outputs in the form of vehicles per day. by themselves. by cost considerations. oil and gas extraction. and it cannot provide through (overhead) flows. Perhaps the most useful part of the CFS for statewide forecasting is its commodity mode split tables. The CFS only provides summary tables of the raw data. Key Freight Data Sources Commodity Flow Survey The Commodity Flow Survey (CFS) is a sample of shipment information from 100. Mode split is determined. to consumption or to a whole commodity flow table. The CFS excludes shipments from farms. Assign Vehicles to Modal Networks. No vehicular data are provided. Split Commodities to Modes. Factor Commodity Flows. it is necessary to determine the amount of goods carried in a vehicle and the number of vehicle days in a year. 9. Flow matrices. Forecasts can be obtained from a variety of governmental. Forecast Growth in Industrial Sectors. Detailed cost models (e.000 businesses. Growth forecasts are best done by economists who specialize in industrial forecasting.. Thus. 6. 10. A cost model is the primary way to determine the impact of public policies on modal choice. These tables give the percentage of weight or value for each commodity on each mode. persons very familiar with goods movement can provide information that may lead to reasonable estimates of mode shares. most retail and service establishments. Freight Forecasting 47 . It does not provide data for flows passing through the US that originated outside its borders. These industrial forecasts can be used to forecast commodities. Commodity flow data are given in terms of tons per year. transportation and government. has occasionally been used for commodities.g. an often used mode split model for passenger forecasting. to a large extent. relating to the full logistics stages of shipping. Growth forecasts are best if they are disaggregated to the TAZ (traffic analysis zone) level. but these tables are extensive. ♦ Mode Split Models: Mode split models estimate mode shares by comparing the relative costs of shipment between alternatives. NCHRP Report #260) have been developed for mode split purposes. ♦ Tables: Historical mode split tables for individual commodities are readily available. 7. it cannot provide commodity flow tables for points within a single state. Develop Modal Cost for Commodities.5. These forecasts can be applied to production. The CFS can give good estimates of externaltointernal and internaltoexternal flows for states. More important are the number of vehicles that use each link. There are three categories of methods for splitting commodities to modes. In these instances. private and educational organizations. are of limited value for planning. Find Daily Vehicles from Load Weights and Days of Operation. For the most part. 8. These tables are especially useful when cost is not a determining factor or when testing policies without significant cost implications. The logit model.
These forecasts are often broken down into many industrial categories. DRI/McGraw Hill and GIS/Trans. ♦ Percentage of mileage by product. Other Commercial Forecasts. 50100. 1992 Census of Agriculture. The Wisconsin mode split by weight (tons) and average trip lengths (miles) for food and kindred products (STCC 20) from the 1993 CFS are shown on the right. weight. At the national level tables of mode by commodity and trip length are available. VIUS comes with software for comparatively easy access to data. Other Interesting Data Sources County Business Patterns. Empty weight. This database is derived from TRANSEARCH. 200500. Truck Inventory and Use Survey) is a large sample of trucks. These tables may be helpful in preparing mode split tables that vary by distance shipped. Assignments to modal networks are possible in terms of vehicles. Commodities are identified by fourdigit STCC. Guidebook on Statewide Travel Forecasting 48 .Average trip lengths are also given. their physical characteristics. This description of this product is not intended as an endorsement. including passengers and empty. and ♦ Commodities reported in 33 categories (roughly twodigit STCC). Origins and destinations are as small as counties. Annual mileage and percentage of mileage by trip length (<50. annual payroll. The data is available online or by CDROM. The database shows commodity volume by mode and lane (OD pair). VMT tonmiles and dollars. water and air. Vehicle Inventory and Use Survey (VIUS) VIUS (formerly TIUS. but as examples of types of data available. their ownership characteristics and their operations. Imports and exports are also identified. Each data record in VIUS is for a single truck. Commercial Freight Data Sources Commodity flow data are available from Reebie. This survey helps fill in information missing from the CFS and contains information on farmbased shipments. Weeks operated. Commodities in the 1997 CFS are organized by SCTGs. containing about 400 fields. average gross weight and maximum gross weight. Mode Parcel Private Truck For Hire Truck Rail Truck and Rail Other or Unknown Tons Trip Length (Miles) 19 NA 13509 56 11843 385 3058 901 115 1718 362 NA The CFS also provides for each state tables for mode by trip length and commodity by trip length. not STCCs. Modes with small or no amounts of these commodities are omitted from this table. Modes include truck. >500). who are codeveloping the Intermodal Freight Visual Database. 100200. rail. number of establishments and size of establishment. County Business Patterns reports for each county and for each fourdigit SIC the number of employees. Several companies provide demographic and industrial forecasts at the county level. Characteristics of a data record within VIUS include: ♦ ♦ ♦ ♦ Extensive description of vehicle and ownership characteristics.
locks and US Customs offices) should be able to readily provide levels of traffic for their facility. Commodities. Publicly operated terminals or links (e. including highway. It should be noted that sampling at weigh stations is biased toward long haul shipments and appropriate corrections are required. and should the model be sensitive to these decisions? ♦ What level of accuracy is needed. North American Transportation Atlas Databases (NORTAD). vehicle type. petroleum and natural gas Specialized databases have unique ways of listing or aggregating data. truck.g. Several wellchosen locations can give a good picture of trucking in a state. For example. The participation of carriers and shippers in counts of vehicular movement is also uncertain. Classification factors developed elsewhere should not be applied to ADT data. The classification must be specific to a location and uptodate.. Forecasts more aggregated than the county are of less value. The NORTAD from the Bureau of Transportation Statistics contains many modal networks. water ports. it is important to clearly identify the desired outputs. Railroad. Defining the Scope of the Freight Model Establish Goals for Model Before embarking upon a modeling effort. rail and water. There are many specialized data sources documented in NCHRP Report #388. rail waybill information may be assigned to a network to give traffic volumes on rail lines (ALK Associates performs this type of analysis). airports. coal. Thus. Forecasts are sometimes available from other state agencies. time and other aspects of use can be obtained in this manner. national networks are required. Other Specialized Databases. a considerable amount of analysis may be required to put the data into a usable form.more than enough for the twodigit precision of STCCs (or SCTGs) and SICs that are typical for planning models. toll bridges. grain. commodity. and what is the desired level of precision for TAZs. These include information on: ♦ ♦ ♦ ♦ ♦ Exports/imports. commodities and modes? Freight Forecasting 49 . origin. Michigan’s truck forecast was largely based on data of this type. ♦ What policy decisions are going to be made. Single Station OD Surveys. banks. and Fruits and vegetables. toll roads. Privately operated terminals may or may not be willing to supply traffic information. utility companies and universities. Freight forecasting requires counts of vehicular flows on enough facilities to validate the forecast. Local Surveys of Note Classification and Other Counts. payload. More extensive data on shipments may be obtained by surveying trucks at weigh stations or other convenient points. Some of these data sources may help to plug holes in the CFS. Classification counts on highways are needed to separate truck traffic from passenger car traffic. destination. Consequently. water and air. but are focused on the state. Specific modes. Some of the more aggressive state models have national scope. Data on gross weight.
plus an additional zone for each of the remaining (contiguous 48) states and external stations at US border crossings. Wisconsin included some counties in neighboring states. there is some advantage in retaining a large number of modes throughout the analysis. Selecting Modes Mode split models can become excessively complicated when there are many modes and even more complicated when there are intermodal combinations. counties. The adoption of countysized TAZs means that vehicular volumes will only be reasonable for intercity facilities. as economic data and forecasts are often not available for smaller areas. However. states.♦ Can development of the model be accomplished in useful phases. capacities may be greatly exceeded for links within major metropolitan areas. Here is a list of CFS modes. adopted a fine grained zone system (approximately townships in the Lower Peninsula) to be consistent with their passenger forecasts. After traffic assignment. BEA regions. Since the CFS provides mode split tables by commodity group. or courier Private truck Forhire truck Air Rail Inland water Great Lakes Deep sea water Pipeline Private truck and forhire truck Truck and air Truck and rail Truck and water Truck and pipeline 50 Guidebook on Statewide Travel Forecasting . when lessons can be learned and applied in future developments? ♦ How can expert opinion be best integrated into the modeling effort? Spatial Units One of the most important decisions when building a model is the fundamental level of spatial aggregation. water and air. a series of assumptions are required to further disaggregate the model. Wisconsin traffic is heavily influenced by the Chicago metropolitan area and the Twin Cities. NTARs. Thus. In the CFS not every commodity uses every mode. vehicular volumes may be aggregated into the four major modes of truck. ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ ♦ Parcel. Should zones be constructed from townships. Michigan. This technique is especially helpful when there are large industrial centers just outside the state. Also. mode split models need not be applied to every commodity and to every mode serving each commodity. rail. US Postal Service. which only performs forecasts of truck travel. or some combination? Most statewide freight models have adopted TAZs based on counties within the state. Disaggregating to smaller than county level presents methodological problems.
Different data sources report the amount of goods by different measures. networks can be quite large. Michigan. Special commodity groups need to be established for waste and deadheading to account for all types of freight vehicle movement. For example. Experience with these large networks indicates the possibility of numerous coding errors. air cargo. conversion factors must be developed. for duplicate links.. and water Network Development The network (for all states except Alaska and Hawaii) should cover the 48 contiguous states but focus on the state of interest. The networks need to be checked for continuity. interstate highways). Even so. Canada and Mexico may also be included. To reconcile the various scales. For example. Adjustment factors can be applied to an existing matrix or the matrix can be expanded by splitting large zones into smaller ones. for looped links (links looping back to the same node) and for multiple nodes at the same location. and the level of detail for the network will be greater. Freight Forecasting 51 . Kansas and Kentucky used only trucks. so efforts should be concentrated on parts of the network nearest the state.♦ ♦ ♦ ♦ Rail and water Inland water and Great Lakes Inland water and deep sea Other and unknown modes Different states have adopted different modal categories. 1. and twodigit SICs are roughly comparable. so it can be used to develop those conversions factors for listed commodities. but do not appear on the national network. rail. Wisconsin adopted four modal categories: truck.000 links. However. Commodities For many reasons it is best to derive commodity groups from standard categories. this method is best suited to urban areas. The networks beyond state borders should consist of only those links with substantial interstate traffic (e.g. the small national railroad network from the FRA contains over 16. A truckonly model may bypass the notion of commodities entirely. Many existing models have used twodigit STCCs (or SCTGs). Within the state there would be smaller zones (see previous discussion). so problems can be encountered when splicing them together. a strategy must be developed for handling loose ends of roads that pass through state borders. Not all elements of the national networks can be checked. Most data sources report commodities in ways that can be consistent with twodigit STCCs. For example. The CFS provides data in both tons and dollars. The state networks are likely to be derived from a different source than national networks. although a considerable amount of cleaning is possible. Indiana defined its modes consistent with the CFS. Developing a Flow Matrix There are a number of means available for developing a flow matrix. which seems to provide enough detail without overburdening the computations.
2.3 31.7 435. Michigan used the model INFORUM (from the University of Maryland) to forecast imports. clay.1 1 2 3 4 5 6 7 8 9 10 12 Commodity Group Animals and vegetables Mining Chemicals Rubber and Plastics Wood. Trip Generation for Gravity Model When the gravity model is used for building flow matrices. Production rates can be established in terms of tons per employee or tons per person. it is necessary to obtain information on the amount of production and consumption for each TAZ and for each commodity group. paper Textiles Stone. For some commodities. Alternatively. depending upon the category.8 15. a gravity model can be constructed and calibrated on statetostate data. Further adjustments to flow matrices may also be needed to account for information not already included.4 16. such as linear regression.3 5. Relating commodity production to employees or population is important when forecasting future flows. pulp.6 11. then applied at the countytocounty level. For example. the CFS does not contain all commodities.3 1242.0 273. Indiana found trip generation rates by linear regression of CFS data. Michigan found the following relationships from state data in 1983 CTS (Commodity Transportation Survey). these rates can be formed by simply dividing the tons of goods produced in a group by the number of people employed by industries involved in this group. Furthermore. Import information can be separately obtained (and forecasted). it is necessary to establish production rates by statistical means. glass. Import data were obtained from the Transborder Surface Transportation Project of USDOT and from data collected under Section 6015 of ISTEA. the CFS does not include information about commodities originating outside of the US. For instance. When there is less than a perfect match between commodity groups and industrial categories.6 68. concrete Metal products Machinery and electronics Transport equipment Other Indiana found the following statistical relationships between production in commodity groups and industrial sectors or population: STCC1 Farm Products STCC11 Coal STCC14 Nonmetallic Minerals SIC7 SIC11 SIC2 + SIC3 Guidebook on Statewide Travel Forecasting 52 . Annual Tons per Employee 222. as forecasts for these particular input variables are readily available. as was done for Indiana.5 2897.
The individual cells in the IO table are valueadded. A preferred method is to use the information in an inputoutput (IO) table. nonmanufacturing and households. The size of shipments is better measured by total sales. Thus. except Electrical Electrical Machinery Transportation Equipment Waste and Scrap Material Other Manufactured Products SIC20 SIC22 SIC23 SIC24 SIC25 SIC26 + 0.8*SIC29 SIC29 Population SIC33 SIC33 + 2. Part of Michigan’s IO Table is shown below. not total sales.STCC20 STCC22 STCC23 STCC24 STCC25 STCC26 STCC28 STCC29 STCC32 STCC33 STCC34 STCC35 STCC36 STCC37 STCC40 STCC50 Food and Kindred Products Basic Textiles Apparel Lumber and Wood Products Furniture and Fixtures Pulp and Paper Products Chemicals and Allied Products Petroleum and Coal Products Stone. There were 11 commodity groups and two nonproducing sectors. The IO table originated with the Bureau of Economic Analysis. an important assumption must be made: weight of a commodity is proportional to the valueadded. Freight Forecasting 53 .6*SIC34 SIC35 SIC33 + SIC34 + 0. Consumption (columns) can be left as industrial sectors. the IO table should be modified so that the producing sectors (rows) are commodity groups. NonManuf 76398 23057 66861 6421 Commodity Group 3 Chemicals 4 Rubber&Plastics 5 Wood Products 6 Textiles 3 48500 6919 4056 86 4 18261 23183 2821 1870 5 11024 8740 88721 3655 6 2351 12226 1064 47109 HH’s 318561 11824 75532 89649 IO tables usually are organized by industrial sectors. IO analysis can also provide an answer to the question: What will be the change in all sectors due to growth in a single sector? Extensive mathematical manipulation of the IO table must be accomplished to answer this question.54*SIC24 SIC28 + 7. regardless of the consuming sector. except that flows are distributed from a production zone to a consumption zone and that friction factors are a function of distance. not commodities.75*SIC36 SIC37 Population Population Forecasts of consumption are also required but are more difficult to obtain. For the purpose of understanding commodity consumption. Clay and Glass Products Primary Metal Products Fabricated Metal Products Machinery. consistent with economic forecasts. presumably by associating industries with commodity groups. Gravity Model for Freight The gravity model for commodities is similar to a gravity model for person travel. Michigan’s IO table was organized by commodity.
g. Indiana used: f (dij ) = exp[βdij ] although the choice was somewhat arbitrary. MichiganOhio. Yj are balancing factors. Pi is the production in zone i. GeorgiaFlorida) and between large states. Guidebook on Statewide Travel Forecasting 54 . regardless of the distance (e. Notice that the largest flows occur between moderate sized states that are adjacent (e.Vij = Pi A j X i Yj f (dij ) Vij is the commodity flow from i to j. Friction Factors. As with person travel. friction factors can be set in a number of ways.g. Aj is the consumption in zone j. β. and dij is the distance from i to j. CaliforniaNew York). The gravity model distributes commodities in tons. varied considerably by commodity. its magnitude being roughly inversely related to length of ton haul... Xi. The single parameter. The map below from the CFS showing the 50 largest flows in the US is a visual confirmation of the principles of a gravity model. CaliforniaTexas. dij was taken to be distance in miles.
Various organizations have developed diversion models for estimating the shift of shipments between truck and rail. These models can be used to generate elasticities that may be applied to individual OD pairs and commodities in a statewide forecast. The balancing factors (Xi and Yj) are computed by the software and do not carry any useful information. Average ton length may be found by dividing tonmiles by tons. Freight Forecasting 55 . Pivot point and elasticities will be discussed later. The lower case letters are all calibrated constants. extensive calibration is required. as the average length of rail hauls greatly exceeds the average length of truck hauls for most commodities. These include the Intermodal Competition Model (ICM) from the AAR. For example. the TruckRail. It is necessary to eliminate bias due to shipment sizes by calculating average ton length rather than using average shipment length already provided in the CFS. Mode Split Methods Mode Split Tables. A typical aggregate demand model for truck volume (VT) (due to Friedlaender and Spady) looks like: VT = SRaLRbSTcLTdVALeTRfTTg where SR is the average rail shipment size. Distance is a significant factor in mode split. For example. The balancing factors just assure that flow is conserved at both ends of the trip. Kansas used the gravity model for externaltointernal flows. Application: During application of the gravity model. Experts can be asked to give mode splits directly or to provide reasonable assumptions to permit forecasts by mode split models. only. Mode Split Models. Expert Opinion. logit. A similar equation was developed for rail shipments.It is suggested that β be set by trial and error until average trip length from the model matches average trip length from data. Extensive aggregated data on mode split are available from the CFS and other sources. LR is the average length of a rail haul. Time. Expert opinion is often necessary to evaluate the effects of new technologies. Regardless of whether the model is aggregate or disaggregate. and Access. TR is the rail unit cost and TT is the truck unit cost. LT is the average length of a truck haul. dependability and frequency of shipment. These factors include: ♦ ♦ ♦ ♦ ♦ Commodity characteristics. VAL is the average value of the commodity. RailTruck Diversion Model from the FRA and the Strategic Choice Model from Mercer Management Consulting. ST is the average truck shipment size. Mode split tables can be readily developed. The major factors in mode split models are intended to respond to policy issues. The Kansas model had external stations at state borders and forecasted agricultural commodities. Indiana developed mode splits by commodity and distance. pivot point and simple elasticities. only. Cost. Quality. Example elasticities are shown later. Mode split models include aggregate demand formulations. Wisconsin used an expert panel to estimate parameters of forecasts of future truckrail intermodal traffic in the state. Diversion Models/Elasticities. Experience with mode split models in statewide forecasts has been limited. Pi and Aj are production and consumption measured in tons.
Pivot Point One technique for mode split of commodities is pivot point. Given the difficulty of calibrating models for each commodity. pbj and pbk are the existing shares for modes j or k. and ∆cj and ∆ck are the changes in the full costs of transporting a ton of goods on mode j or k. This report developed cost relationships for rail. It can be calibrated to match known cross elasticities between modes (see later discussion). Use of Mode Split Models. Because of extensive changes in technology. it is recommended that the number of modes be kept to a minimum: truck. That is. commodities and markets since this report. water and air. if desired. For Guidebook on Statewide Travel Forecasting 56 . but only allow the variation of one term in the utility function. its results are not directly applicable to forecasts done today. A solution may be to limit analytical mode split analysis to shipments entirely within the state. on comparison of costs all OD flow of a particular commodity would be assigned to the least cost mode. can be calibrated by knowing mode shares and average modal costs at two points in time. as a fraction. NTARs). Elasticities An elasticity is the fractional change in output divided by the fractional change in input. The existing mode shares are required inputs. Pivot point is derived from a logit model. α. this limitation is not serious. Elasticities can be used to measure the effect on demand of any other important variable. especially considering the size of zones (states. NCHRP Report #260 recommended an allornothing mode split. Special analysis may be performed for those commodities of greatest interest. pj = ∑ pbk exp(α∆c k ) k pbj exp(α∆c j ) In the above equation: pj is the forecasted share for mode j. Since the mode split of commodities is largely determined by cost. Modal costs are also given extensive treatment in NCHRP Report #388. External flows would then be split by table. α is a calibrated coefficient which varies by commodity. Because of the difficulties in calibration. truck and barge. but requires somewhat less information to operate and is less sensitive to errors of calibration. The most complete treatment of mode split is found in NCHRP Report #260. The single coefficient of the pivot point model.Model from NCHRP Report #260. The coefficient could also be calibrated by analyzing surveys of behavioral intention of shippers. rail. Tables can be used to further refine these categories. A mode split model should be selectively applied to those commodities expected to be sensitive to the policies being tested. Pivot point models are as valid as logit models. a comprehensive mode split model covering each mode and each commodity is not recommended. A very detailed cost model was provided. Obtaining cost data for interstate OD pairs may be quite difficult.
57 4. Major exceptions were: Freight Forecasting 0 1 2 3 4 5 6 7 8 9 10 12 Commodity Group Empty Animals and vegetables Mining Chemicals Rubber and Plastics Wood. containers. speed.73 4.4 20. the degree of competition.2 5.62 2.5. For some commodities.7 57 . Indiana calculated 5.5 7. it is necessary to estimate the number of equivalent weekdays in a year.) Truck Days.0 23. These elasticities include all factors.7 0. as predicted by the Intermodal Competition Model (ICM) developed by the Association of American Railroads.03 3. vehicular size. including access.7 0.1 0.28 1.14 3. Commodity Bulk Farm Products Finished Farm Products Bulk Food Products Finished Food Products Lumber and Wood Furniture Pulp and Paper Bulk Chemicals Finished Chemicals Primary Metals Fabricated Metals Machinery Electrical Machinery Motor Vehicles Motor Vehicle Parts Waste and Scrap Bulk All Else Finished All Else Low Elasticity 0. The average cross elasticity is about 0.71 0. Detailed Elasticities.8 (weekdayequivalent) flow days per week or 300 per year. In order to convert annual flow to weekday flow.transportation services it is possible to use elasticities to quantify the effect on demand of headway.8 4.2 0.3 4. etc.5 1.93 0.9 29.0 15.8 0. It contains cross elasticities between truck and rail in Canada.67 3.6 15.17 0.5 Commodity Density and Load Weights Converting commodity flows into vehicular flows usually requires two sets of conversion factors.0 0.49 3.21 1. Michigan found that most trucks on intercity trips carried about 4050 thousand pounds.5 0. etc.2 3. clay.19 4. glass.2 26.02 3. The first factor is used to convert annual tonnage to daily tonnage and the second factor is used to convert tonnage to vehicular loads (trucks.6 12. paper Textiles Stone.2 36.4 0.22 0.2 1. mode split may be best accomplished with cross elasticities. For example.0 0. cars.1 0.1 27.4 7.2 24. The following table was drawn from NCHRP Report #388. concrete Metal products Machinery and electronics Transport equipment Other Tons per Truck 5.9 High Elasticity 0.83 2. Load Weights. Each elasticity is the estimated percent increase in rail tonmiles for each one percent increase in truck costs. a cross elasticity between truck and rail might be the percent change in rail demand given a 1% change in truck costs. pulp.7 4.
Unfortunately. Mixed Shipments Small Package Freight Hazardous Waste Hazardous Materials Unknown Tons/ Truck 23 19 24 9 8 12 5 2 16 23 4 3 4 3 7 4 16 18 12 Traffic Assignment AllorNothing. Other techniques that have been considered are capacity restrained equilibrium and stochastic multipath. clothes. Allornothing assignment (all trips between an origin and a destination are placed on the single shortest path) has been the preferred method of assignment by those doing statewide freight forecasts. Furthermore. Gasoline Nonmetallic Minerals Ordinance or Accessories Food or Kindred Products Tobacco Produces Textile Mill Products Apparel. Plastic Leather or Leather Products Tons/ Truck 24 13 6 24 24 14 19 24 18 5 5 3 15 3 9 22 19 4 3 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 99 STCC Clay. especially bulk commodities. Michigan allows larger trucks than many states do. Photo or Optical Misc. NG. Extensive adjustments to impedances (times. costs) on links is required to get acceptable assignments.empty. Wisconsin developed load weights for commodities in twodigit STCC. Concrete. tending to load most highway traffic on interstate highways. allornothing assignment is very sensitive to speeds. A comparison of Wisconsin to Michigan would suggest that load weights are rather constant for some commodities but vary considerably by state for other commodities. Not Electrical Electrical Machinery Transportation Equipment Instruments. Capacity restraint techniques are particularly difficult because freight vehicles. 1 8 9 10 11 13 14 19 20 21 22 23 25 26 27 28 29 30 31 STCC Farm Products Forest Products Fish or Marine Products Metallic Ores Coal Crude Petro. Glass. In addition. Finished Textile Lumber or Wood Products Furniture of Fixtures Printed Matter Chemicals Petroleum or Coal Products Rubber or Misc. Guidebook on Statewide Travel Forecasting 58 . textiles. where only a small percentage of roads are included in the network. rarely exceed the capacity of links. Stone Primary Metal Products Fabricated Metal Products Machinery. thus violating a major premise of capacity restrained assignment. so some average payloads were considerably higher. Railroads experience considerable economies of scale by routing traffic on mainlines. highway capacity becomes useless as a modeling concept in urban areas. Other Methods. These are shown in the table below. furniture and electrical machinery. it is difficult to simulate rail routing decisions with a shortest path paradigm. Michigan eventually used the load weights on the previous page in their forecast. Manufacturing Products Waste or Scrap Materials Misc. Freight Shipments Shipping Devices Returned Mail and Express Traffic Freight Forwarder Traffic Shipper Association Traffic Misc. by themselves.
5% of intercity truck trips were direct (no intermediate destinations). Thus. The network must clearly designate points where shipments change carriers. The pure rail mode would not be a possibility for shipments produced or consumed in these counties. a shortest path is only a rough approximation of the true path.true speed)0. and then took a weighted average of the volumes from the assignments. Highway Impedance. distance and cost. The averaging weights were developed subjectively. each with a different impedance measure. but modified the speed used in the travel time calculation to provide less of a disadvantage for noninterstate routes. there are a number of fixes that can be applied to achieve better assignments. Assignment Fixes In the absence of better algorithms. Rail Issues. railroads do not have a strong incentive to either minimize time or distance. but there is comparatively little experience with the technique for statewide freight forecasting.5 Railroad Impedance. As readily seen in the figure. Indiana used travel time as its primary measure of impedance. A stochastic multipath assignment sends some trips along the shortest path and smaller proportions to other reasonable paths. (Railroads are compensated according to fraction of distance carried. The three impedance measures were based on time. The following formula was used: speed = (true speed) + 2 (65 . many rail shipments involve multiple carriers.) Still. Highway Issues.Stochastic multipath assignment shows promise. 99% had 0 or 1 intermediate stops. Because of the method railroads charge shippers for freight transport. This method created a smoother loading than might have been obtained otherwise. Kansas performed three different traffic assignments. Indiana accounted for railroads’ tendency to assign traffic to mainlines by adopting the following measure of impedance: I = L (1/(D + 1)) Freight Forecasting 59 . many counties do not have rail access. Michigan found that 92. The figure on the right illustrates the importance of mainlines of Class 1 railroads for major commodity movement. Evidence suggests that for statewide models there is little need to consider multistop tours.
without identifying the order of the modes.. Inputoutput analysis is a technique for forecasting the effect of industrial growth in a single sector (or few sectors) on the overall economy. The algorithm would assign PCEs rather than vehicles to links. Truck PCEs may be assigned to links ahead of the assignment of passenger vehicles. Each link (indeed. so it would be difficult to determine how many trucks use each link. it is only possible to assume 50% of each. Some states. In so doing. like Indiana. Thus. LinkbyLink.g. The only difficult part of IO analysis is getting the IO table. Forecasting Inputs to a Freight Forecast Industrial and Population Forecasts.where L is the length of the link and D is the traffic density in millions of gross tonmiles per year. truckrail or truckair). It is recommended that these forecasts be obtained from specialists. The problem of too many intermodal combinations can be simplified by assuming that all intermodal shipments are interstate or international. as illustrated previously. If good estimates of delay are needed. a truckrail intermodal shipment may logically be by truckrail. IO analysis starts with an IO table. Other Freight Forecasting Issues Combining Passenger and Freight Forecasts The Highway Capacity Manual (1997 revision) gives passenger car equivalents (PCEs) for trucks as a function of grade. Without further information. Strictly speaking. Assignment of Intermodal Flows. rail terminals in the state can be either the production end or consumption end of intermodal flow. The lowest heavy truck PCE is 1. each of these three modal combinations should be given separate treatment in the assignment process. The analysis finds the total value added for all industrial sectors given an increase in consumption for a single sector. The effect of a single truck on delay can be considerably greater than the effect of a single passenger car. truckrailtruck. The Bureau of Economic Guidebook on Statewide Travel Forecasting 60 .5 (level) and the largest is 15 (6% grades in excess of 1 mile when there are only 2% trucks in the traffic stream). Whole Trip Table. Intermodal shipments in the CFS are identified by pairs (e. or (2) apply the PCE factor to trucks on each link individually. have so little intermodal traffic that it makes sense to simply lump these flows into one or both of the component modes. or railtruck (ignoring drayage). then it is easiest to multiply the whole truck trip table by a single PCE factor that represents average conditions. trucks should be appropriately weighted to account for their larger effect: Delay = f(passenger cars + PCE*trucks) There are two ways to account for trucks in the traffic stream: (1) apply an average PCE factor to the whole truck trip table. Thus. If equilibrium assignment is the desired technique for traffic assignment. The IO table tells who sells to whom and how much. An equilibrium assignment for just the passenger cars could then be pursued to obtain final delay estimates. Production and consumption of commodities depends on forecasts of employment (or sales) for industries and forecasts of population. it is possible to account for the larger effect of trucks on steeper uphill segments. IO Analysis. each link direction) theoretically has a different PCE factor for heavy trucks.
Productivity. If the forecast of a commodity is based on growth in employment. the mathematics is simple enough to be performed on a spreadsheet. Once an IO table has been obtained. Forecasts of productivity increases for various economic sectors are available from proprietary data sources. Freight Forecasting 61 . These variables affect the load weights and the choice of premium modes. Regional IO analysis. Value per ton may change in the future due to (1) shifts in the types of commodities produced within a commodity group or (2) or a decrease in the actual value of a given commodity.Analysis publishes an IO table for the whole US economy. IO tables are also included with proprietary software packages. allows identification of the effects of industrial growth on a state or urban area. Forecasting these effects is difficult and should be reserved for those commodity groups of special interest. an extension of traditional IO analysis. then an adjustment is required for anticipated increases in worker productivity. Value per Ton. Load Weights and Shipment Sizes. There has been a recent decline in both shipment densities and in shipment sizes.
external trips. ♦ Pivot Method. Philosophy of calibration. Using traffic counts to reestimate a trip table that might have been created through surveys or a gravity model. in developing a coherent validation database determining the quality of a validation. The actual calculation requires multiplying each section’s AADT by the section length and an expansion factor. especially the Highway Performance Monitoring System (HPMS). freight truck OD surveys. including linear regression. Specialized Methods for Passenger Forecasting Introduction This chapter is concerned with a number of topics that enhance or replace parts of the traditional fourstep model for passenger forecasting. ♦ Steps in between Steps. either in terms of AVMT (average vehicle miles of travel) or AADT (average annual daily traffic) for use by the state during the validation phase. ♦ Validation. HPMS is a consolidated data set containing information on every major public road within the state.Chapter 5. the area where the road is located and volume of traffic it carries. without running a complete statewide simulation. etc. A method for projectlevel forecasting that combines the best features of a simulation and timeseries analysis. A method to help calibrate mode split models when they must include modes that do not yet exist in the corridor or state. ♦ AADT Estimation: Like the AVMT. Time of day issues. Data Collection Validation Data Sets: HPMS Element of the Traffic Monitoring Program HPMS constitutes a valuable source of validation data at the statewide level. such as the category the road belongs to (interstate. ♦ Stated Preference. based on daily counts. HPMS can be used to prepare traffic volume samples. day of week factors and axle correction factors are applied in the calculation of the AADT estimate. minor. regardless of mode. Guidebook on Statewide Travel Forecasting 62 . The resulting AADTs for all expanded sections are then summed. vehicle occupancy and interfacing with urban models. external travel surveys and survey techniques. longer periods become laborious and timeconsuming. analysis of variation (ANOVA). Correction factors such as seasonal adjustment factors. truck travel surveys. These topics relate to the following. AVMT estimations are sometimes based on longer periods of 48 hours. nonlinear regression and maximum likelihood estimation. for the most part. ♦ Trip Table Refinements. The methods adopted for the calculation of AADT and AVMT estimates are briefly described below. major.). ♦ Total Corridor Demand. however. ♦ Surveys. techniques for calibrating model steps. Understanding validation data sources. the AADT is. ♦ AVMT Estimation: This process is based on counting periods not usually exceeding a day. Single station OD surveys. ♦ Calibration. Estimating the amount of travel in a corridor.
March 1997 Specialized Methods for Forecasting 63 . Traffic Counting in Wisconsin: Wisconsin has a well established counting system set up on all the major highways in the state.or more axle single trailer trucks. Coverage counts are made on a threeyear cycle at 30. 6axle multitrailer trucks. each state varies its procedures to account for its own unique geography. other 2axle. 1992 Adapted from Wisconsin Highway Traffic Volume Data. March 1996 and Wisconsin Automatic Traffic Recorder Data. but they give a good indication of the quality of data within HPMS. 4. ♦ Samples should be scheduled throughout the year to minimize seasonal biases. The FHWA has adopted 13 vehicle classifications based on the physical characteristics of the vehicles: motorcycles. 2axle. 5. conducted throughout the year to record daily traffic volumes. 4. ♦ Both directions of travel should be monitored. Seasonal. 6.2 1 2 Traffic Monitoring Guide. Shortcomings of HPMS Data: The data needs of each state may be different from others. Thus. passenger cars. Randomness of samples should be maintained with regard to both season and location. Moreover. climate and size of cities. Where this is not done the usefulness of the data for validation purposes is greatly limited. This vehicle classification scheme exceeds the requirements of travel forecasting. Wisconsin adheres closely to the recommendations in the Traffic Monitoring Guide. buses. ♦ Growth factors may be used to interpolate counts taken every three years to yearly estimates ♦ Dayofweek factors should be applied to the raw count data to eliminate daily variation. 5axle single trailer trucks. Recommendations include the following: ♦ For coverage counts a 48hour monitoring period should be repeated once every three years.or more axle single unit trucks. Recording truck weight at the same time is also recommended. It has a total of 142 continuous counting stations located mostly on high volume roads. Not all states follow these recommendations. seasonal counts to keep track of the seasonal factors affecting these volumes and coverage counts to record more specific information on a particular road or highway segment.1 Example. dayofweek and axle correction factors are accounted for in the estimation of AADTs. FHWAPL92017. 6tire. such as population. it is important to note that although HPMS should be a very consistent data set across states.or less axle multitrailer trucks. There are basically three types of counts adopted by states to maintain traffic volumes.Additional detail in the validation data may be attained by applying vehicle classification information in HPMS volume samples. single unit trucks. and 7. ♦ Samples should also be well distributed spatially. Traffic Monitoring Guide Recommendations for HPMS Data Collection: The Traffic Monitoring Guide makes several recommendations about how count data should be collected and processed for HPMS. 3axle single unit trucks. Such a comprehensive database would allow validation of both statewide and corridor models. the application of axle correction factors helps to main statistical consistency. They are continuous counts. so that geographic biases can be minimized. FHWA.or less axle single trailer trucks. 4tire single unit trucks. Publication No.000 separate locations.or more axle multitrailer trucks.
Another advantage is that the SSOD survey information can be used in conjunction with select link analysis for validation purposes. such as vacation trips during the summertime by state residents and nonresidents. Michigan recommended several possible modifications to SSOD survey procedures. SSOD surveys also pick up long trips that are too infrequent to be seen in reasonable numbers in a household survey. In contrast.). which showed less than 1% of the total trips as vacation trips.Single Station OriginDestination Surveys The Michigan Department of Transportation (MDOT) is known for conducting several types of origindestination surveys over the last 30 years. and most often conflicting with the trip purpose breakdown used by the model. they may not always relate to the average annual flow. The main advantage of conducting SSOD surveys is that they are capable of capturing atypical trips. Information was mainly collected through roadside interviews conducted by MDOT personnel throughout the state. the type of vehicle (or mode of travel). One or both directions of travel were recorded. Inc. These data included time of day. This is another troublesome outcome of conducting these types of surveys during summer months alone. Since 1985. and ending at 8 p. ♦ There are inconsistencies in the trip purposes. April 1996 Guidebook on Statewide Travel Forecasting 64 . ♦ The model shows an average trip length shorter than that indicated by the SSOD data. but the sample was eventually expanded to represent a whole day. the SSOD data that was collected could not be directly applied towards model development in Michigan. This is due to the influence of the trip purpose breakdown on the average trip length and the fact that longer trips have a greater chance of being counted. ♦ Future SSOD surveys should distinguish between residents and nonresidents of the state. In this regard. Since most of the information for these surveys is obtained through direct personal interviews. About 300 surveys have already been conducted. Most of the interviews were conducted during a 12hour time frame (usually starting at 7 a. Michigan Department of Transportation. Statewide Travel Demand Update and Calibration: Phase II. 129 sites in 49 of Michigan’s 83 counties have been surveyed and recorded1. month and year that the trip was made. Based on this experience. This information was further supplemented using a vehicle classification counts over a 24.. ♦ They do not distinguish between residents and nonresidents of the state. conventional household surveys have shown poor results. Some of these reasons are listed below: ♦ The SSOD surveys are conducted only during the summer months and on weekdays. This procedure will be dealt with in more detail later in this chapter. Due to several reasons. destination and purpose of the trip.hour period.m. number of persons per vehicle and the origin. where 61% of the trips were for vacations. the data gives a very realistic picture of the travel patterns throughout the state.m. covering all major state trunklines and interstate highways. A typical example was the Mackinaw Bridge Survey conducted for MDOT. The data set collected from single station origindestination surveys (SSODs) is probably one of their most extensive sources of information on passenger traffic in the state of Michigan and represents more than 800.000 “expanded” survey trips. 1 KJS Associates. As a result. the Michigan model was calibrated with NPTS data.
5% of the trips did not have trip chaining at all. Ibid. For example. MDOT personnel used survey forms to record the information. This survey was fairly extensive and included more than 5. Details gathered on vehicle classification were consolidated into the 13 categories recommended by the FHWA Traffic Monitoring Guide. it was stored in database format. the expansion factors were multiplied by the actual trip distance. Ports of entry and truck weighing stations on all major highways.000 surveys. MDOT applied expansion factors to all the survey locations and to the 13 vehicle classifications. then the expansion factor would be given by dividing the total volume of trucks by the total number of observations made at the survey station.1 Truck Travel Surveys Truck travel surveys are done to record more specific information on commercial vehicles. Specialized Methods for Forecasting 65 . Several coding errors were also eliminated in the process. ♦ Inconsistencies with the model. should be eliminated. In order to calculate these expanded numbers.♦ Additional SSOD surveys should be conducted for control purposes in other travel months to help develop seasonal adjustment factors. A good example is the 1994 Michigan Truck Survey conducted by the Michigan Department of Transportation. However. The three categories used were: ♦ Intrastate (no border crossings). Inc. up to three trip chains between the ultimate origin and destination were noted. unlike the SSOD surveys that cater to all vehicle types. such as the trip purpose definition. ♦ Interstate. An advantage of this survey method.2 The commodity groups that were specified for the Michigan truck survey were based on a twodigit STCC. carrying large daily volumes of commercial vehicles. The data collected included the actual number of observations for each commodity group. 1 2 KJS Associates. Survey time can extend from anywhere between 12 and 24 hours. Michigan. in the state are generally designated as survey stations. to assure a statistically valid database trips were further classified into three specific categories depending on whether they crossed the border or not. The main function of the expansion factors was to weight each sample so that the total sum of these factors would give the total volume of trucks surveyed. unlike an SSOD study. During the actual survey. ♦ All the SSOD surveys conducted in the future should incorporate commercial traffic flow along with private vehicles. An unbiased sample can be collected at a controlled sampling rate. Once the data was checked. it was found that 92. is that it less biased by the geographical location of a single station. if the various truck classifications were to be ignored. The surveys were conducted at 10 different locations within the state. ♦ SSOD data sets should be built in a manner that enables their effective use for the model development phase. In addition to classifying the collected information by commodity groups. and ♦ International (to account for the larger number of trips made across the USCanadian border). This information was then expanded to yield estimates of the number of trips made and the vehicle miles of travel per group.
US Department of Transportation. Statewide Travel Demand Forecasting. Inc. such as time of day. Most stateline cordon surveys are required to monitor both directions of travel. arise if these files are not updated on a regular basis. Here. The final application of the model itself is the chief determinant in deciding which of these OD pairs to use. it is difficult to trace vehicles from other states. two origindestination pairs were considered: ♦ One pair with an ultimate origin and destination. For the other highways. as with the Michigan Truck Survey. Also. these surveys should monitor traffic on all major interstate facilities as well as intrastate highways with a fairly high traffic volume that contribute to interstate travel. Truck and commodity flows were recorded over 24 hour periods keeping in consideration important seasonal factors. the main focus was on all the major commercial corridors within the state.Hence. NHB. the classification could be based on the guidelines given by the FHWA Traffic Monitoring Guide). HBO. Adapted from FHWA. and ♦ Trip purpose (whether HBW. Mail or telephone surveys 1 2 KJS Associates. However. Business. This survey was unique as it collected statewide freight data throughout the state through personal interviews of truck drivers. Household and OnBoard Surveys. Other Survey Methods Stateline Cordon Survey. However. the idea of separately categorizing those trips with an intermediate stop was discarded. The most important data items to be collected include: ♦ Time of interview. Some of the most practical survey techniques used at the statewide level are mail or telephone surveys and onboard surveys. In 1993. License plate surveys are particularly useful on heavy traffic corridors as it does not slow down the traffic. ♦ Type and registration details of vehicle (for trucks. A total of 25 truck weigh stations and ports of entry were designated as survey locations. for later analysis. To efficiently conduct this survey. smaller samples were taken based on the actual daily truck flow. month and year (to keep track of any daily or seasonal variations in the travel patterns). Again.2 License Plate Surveys. The I5 corridor in Washington is particularly known for its heavy commercial traffic. Vacation. the sample was restricted to one out of every 10 trucks on this highway.1 Washington. Washington State Department of Transportation (WSDOT) conducted a truck survey. the goal was to obtain at least 300 surveys during a full day at one location. ♦ Occupancy per vehicle. a problem could. Complete automation is a definite advantage of this system. The survey procedure was similar to that followed by the Michigan Department of Transportation. it is essential to have access to computerized automobile registration files. making it very safe. 1973 Guidebook on Statewide Travel Forecasting 66 . Hence. ♦ Details of origin and destination of the trip. Stateline Cordon Surveys are intended to gather information on those trips originating or ending outside the state. both within the state and outside. ♦ Another pair taking into consideration only that portion of the trip through the survey station. In order to keep the survey process within reasonable limits. easy to execute and comparatively accurate. etc). Recreation.
point of destination. These surveys usually focus on public modes of transportation and provide more comprehensive information than can be obtained from records of collected tickets. although it is important that inputs relating to speed are reasonable. assures that the model will not cause a major blunder in sizing some future facility. since direct contact is made and followup. The time of day. then the maximum acceptable RMS error in a model validation should be about 1. Care should be taken to keep the questionnaires brief and to the point. Perhaps a better means of judging validation quality is to insist that the model be about as good as the traffic counts to which it is being compared. transfer points. The household survey focused on households owning an automobile in the state.are more economical than conducting home interviews. although small in number.000 households (an average of 60 households per county). op. OnBoard Surveys. The usual response rate for these surveys is anywhere between 20 to 50%. if followed. There are three primary questions that must be answered. usually LOS C. when necessary.4e. as travelers must complete the survey enroute. Typical questions include the point of origin. Care must also be taken to account for households without cars or telephones. carrier and vehicle information are usually completed by the person administering the survey. A 45% response rate was achieved. trip purpose and traveler characteristics. links with volumes of less than 300 vph in the peak period should be ignored. These surveys help to cover information that cannot be effectively collected through household surveys for infrequently used modes. If the RMS error in traffic counts is known to be e. The total sample consisted of nearly 15. VMT or VHT) sufficiently close to those measured? ♦ Are the differences reasonable. For all practical purposes. this rule would mean that urban streets (which are signal controlled) should be accurate to some number between 300 to 500 vph and freeways and multilane highways should be accurate to some number between 600 to 700 vph. That is. Telephone surveys are often used as the second stage in mail surveys. Specialized Methods for Forecasting 67 . is easier. auto registration files or telephone directories with names and current mailing addresses can be used. cit. ♦ Are the forecasts of total traffic volumes (total volumes. Telephone surveys usually assure a better response rate. The sample size can also be expanded.1 Validation Quality Validation involves comparing a baseyear forecast to actual traffic counts. traffic estimates should be accurate to within + or – onehalf the amount of traffic that could be carried by a single lane at its design capacity. linkbylink. A good example is the mail survey conducted for the Kentucky Department of Transportation for both household and truck travel data. Roughly speaking. for efficiency. access mode. Again. RMS error is calculated as follows. between the forecast and actual volumes? ♦ Is the forecast spatially unbiased? Good estimates of speeds are not a major consideration in statewide travel forecasts. urban models have been validated according to the 1/2 lane rule. like the license plate surveys. The 1/2 lane rule. For many years. 1 Adapted from FHWA.
RMSE = 1 N 2 ∑ (Vcounted − Vestimated ) N i=1 Acceptable Error Plots Percent Error 30 The curve in the figure on the right shows 25 a typical relationship between link volume 20 (24 hours) and the RMS error between a 15 baseyear forecast and ground counts. 30th highest hour. As an end product they would need to ascertain a mode bias coefficient for high speed rail and to determine whether the coefficients on travel attributes (time. ♦ Good agreement with screenline counts across major travel corridors. automobile and airline. it is acceptable to validate the total of these roads. it was necessary to investigate users’ perceptions of the new mode and to approximate its coefficients within a logit model.) needed to be modified. the mode bias of the new mode relative to existing modes. Spatial Validation. etc. For example. If validation is required in a portion of the state where there are many lowvolume roads. Thus. Stated Preference Stated Preference is a survey technique that can overcome problems with using only revealed preference (actual behavior) in calibrating models. The respondent’s current mode need not be included among the choice set. among other things. This method describes two alternative modes for a trip already experienced by the respondent. Existing modes in the corridor consisted of conventional passenger rail. as they are highly unlikely to require more than one lane per direction under any standard design criteria. as if they formed a single link. Some considerations are: ♦ Good agreement with VMT within several big districts that cover the state. There are three general methods of obtaining stated preference information. 10 but illustrate the important statistical 5 principle that about 68% of all errors should fall below the line and about 32% 10K 20K 30K 50K 60K 40K of the points should fall above the line. A logit model can be calibrated to the data. The respondent is asked whether the new mode would be chosen. or 100th highest hour. Wisconsin DOT wanted to investigate the demand for a high speed rail alternative. WisDOT believed that the characteristics of high speed rail could be significantly different than conventional rail. There are no well established procedures for performing a spatial validation. The points on the chart are hypothetical. Ask about Choices. cost. and ♦ Good agreement with VMT in counties constituting major activity centers. Guidebook on Statewide Travel Forecasting 68 . (Design criteria include peak hour in week. Validation should be achieved across the state to assure that forecasts will not be spatially biased. Volume Road segments with less than 2000 vehicles per day in a single direction should be ignored.) Validation in Low Volume Areas. giving. Method A. if it is implemented.
Respondents were given four sets of pairwise choices under various automobile LOS conditions. the respondents should initially be told the ratings of two arbitrary trips. Wisconsin’s High Speed Rail Study illustrates method A. cost and frequency of service. The respondent could indicate one of five possible answers to any comparison (definitely choose mode A. can be statistically analyzed to obtain most of the coefficients of a utility function. The other trip is a typical trip and is given an arbitrary rating (e. a Chicago Loop (downtown) dummy variable. To obtain stable ratings. outvehicle time. Each respondent can be asked about many trips. Surface depending upon the strength of the Automobile Airline Public response (0. Data points were weighted. definitely choose mode B). mode specific constants and the logsum constant for the surfacepublic nest. automobile v. Direct Disutility Estimation. These stated preferences were included into a nested logit model calibration with revealed preferences (actual choices).g. Thus. The survey form did not force a definite response to the stated preference questions.Method B. automobile v. Specialized Methods for Forecasting 69 . a popular tradeoff method rates attributes in terms of their monetary (dollar) value. and automobile v. The nest utility is calculated by the equation below. 0. There were far more stated preference data points than revealed preference data points.7 for probably and 0. an income variable for rail and bus modes. The final nesting structure is shown on the right. Method C. indifferent between A and B. probably choose mode A..g. air. Conventional Rail High Speed Rail Intercity Bus Service quality was measured primarily in terms of invehicle time.9 for definitely. along with the trip data.5 for indifferent). Respondents can be shown characteristics of trips and asked to rate the disutility of the trip on an openended scale. given that each of the modes in the nest have their own mode specific constant. Tradeoff methods ask respondents to judge the value of one attribute against another. 100). probably Decision choose mode B. For example. The final set of coefficients is shown in the table on the next page. Subsequent calibrations held constant the ratios of parameters for the servicerelated variables. bus. based on earlier studies. One trip is described as having no length and no disutility (and thus a 0 rating). The model was further adjusted by hand to replicate existing mode shares and to deal with aggregate data (e.. A typical question would be: “How much more fare (toll) would you be willing to pay to avoid 10 minutes of travel time on this trip?” This method can yield a value of time for the new mode. the model estimated parameters for this combined service quality index. At first the calibration did not result in good estimates of all parameters for servicerelated variables. average zonal income) instead of household data. so the sample size can be small. conventional rail. purposes and end points: ♦ ♦ ♦ ♦ automobile v. Ask about Tradeoffs. high speed rail. The ratings. The nest bias constant (seen in a previous chapter but eliminated here) is redundant.
8769 Other 0.3723 0.8380 1.0292 1.5935 0. The TriState study conducted a stratified random sample of travelers by all modes in the corridor. Michigan used vehicles/person (VP) as its primary measure of occupancy.Bus Distance < 75 mi – nonBus Air Constant Rail Constant Bus Constant HSR Constant Logsum – Surface/Public Nest Commute 0.1117 2.0161 0.9523 0.8769 Recreation 0. Rather. Again. The data was Air Other 34 32 32 19 analyzed by binary logit. NTPS (1990) data were used to ascertain vehicle occupancy rates as a function of household characteristics and trip length.2557 1.0450 0.2583 1.6875 0.4046 0.4488 14.0097 0.0335 0.6055 5. Single Mode Vehicle Occupancy: Michigan Michigan did not use a formal mode split model for statewide travel forecasts.1906 1. frequency and Mode and New Ontariocost.0290 0. Bus Distance < 75 mi .8126 1.3868 0.9648 1.4046 2.7545 0.2729 0.0122 0.2501 0. There were separate Purpose TriState York Quebec Illinois questions for each mode and Air Business 65 51 58 54 purpose. Respondents were given a series of trip choices.Un = φn ln ∑ eUk k φn Variable Total Travel Time (OutVehicle AET)/Distance (Travel Cost)/(HH Income) Inverse Frequency Chicago Loop – Bus Chicago Loop – Rail.0397 0. it converted person trips to vehicle trips (all relevant modes) by a vehicle occupancy factor.0218 0.9536 1.0397 0. stated preference techniques were used.4067 0.2886 0.5231 1.5576 2.0452 1. Michigan’s relationship between VP and distance was determined by linear regression for both work and nonwork purposes: Guidebook on Statewide Travel Forecasting 70 .3603 0.0126 1.0292 1.8769 Business 0. HSR Income – Rail. systematically varying time.5672 6. Rail Business 40 26 25 28 The table on the right Rail Other 28 21 19 13 illustrates a comparison Auto Business 43 26 25 23 between the TriState study Auto Other 26 26 18 13 and three earlier rail corridor Bus Business 25 17 studies for values of time Bus Other 22 32 12 (dollars per hour).0290 0.2886 0.5576 2.0103 1.67810 1.0034 2.8201 0.TriState Wisconsin participated in another high speed rail study with Illinois and Minnesota.1725 0.0103 1.9360 0.8769 Value of Time and Value of Frequency .6051 0.3603 0.6055 5.
A distinction should be made between statewide and urban models. Time of Day Considerations If there is a need for an hourly forecast.05561 exp(0. difficult to acquire and difficult to estimate.002045 D) Nonwork: adjustment = 1. Thus. The distance adjustment is applied to a table of vehicletoperson ratios. The income and household size categories are consistent with the trip generation step. one hour or greater). then provisions must be made for the unique nature of intercity travel. Dynamic AllorNothing Traffic Assignment. because average trip lengths are considerably shorter than the smallest time period of analysis (typically. The special AON assignment only records link volumes that are between J and J+1 hours in Specialized Methods for Forecasting 71 . trips are generated and distributed for a specific hour in the day. The 24hour assigned volumes are allocated to hours and directions by a table of factors for each link. The following procedure should be performed. These urban methods will not work in statewide models (unless the state is quite small). There are two principle techniques. 2. Ascertain the maximum length trip in the network. 1. Only a modest change is necessary to existing traffic assignment algorithms in order to implement a dynamic allornothing (AON) assignment. because trips found on rural roads are quite long. Urban models do not attempt to distinguish between long and short trips when applying timeofday factors. Assignments are made hour by hour. those trips on the rural portions of a statewide network that are quite long in duration are more likely to start in the early parts of the day. but it is conceptually much easier with less congested statewide networks that would otherwise be analyzed with allornothing traffic assignment. Second. Under HPMS guidelines states are likely to have time of day information for only a small fraction of their counting stations.Work: adjustment = 0. The peak hour of any given link will not necessarily correspond to the peak hour of zonal trip generation. Long trips (greater than two hours in length) will likely impact a link one or more hours after its generation. A Solution. trips are generated and distributed for a full 24hour period.052769 exp(0.001346 D) These equations suggest that VP declines with distance (or vehicle occupancy increases with distance). However. Dynamic assignment takes into consideration the length of trip when assigning it to a network. First and easiest. Run a special AON assignment for the current hour and each of the N1 previous hours. a statewide model may find it beneficial to develop separate timeofday factors for trips of different durations. one ratio for each combination of income and household size categories. Time of day is handled in different ways in different urban models. Tables of factors for each link are unavailable.100 D0. N hours.8679 D0. Dynamic assignment can be quite difficult to implement when analyzing congested urban networks. Michigan found the OD pair VP ratio by averaging the ratios for the two involved zones.
The main conclusion from this analysis was that the outstate trips were far less relevant than the trips made within the state. One of the disadvantages of using a gravity model to estimate the number and distribution of through trips is that it is based on accessibility. 4. The process was based on the following steps: ♦ Identification of intrastate versus interstate trips. Inc. has indicated that the trip rate for trips with at least one trip end outside the state is inversely proportional to the area of the state in square miles. Another interesting feature of the outstate trip generation approach in the Michigan model is that trip productions and trip attractions are computed by the same model. In essence. Validate the baseyear assignment against traffic counts for several hours in the day and for both directions of travel. In 1995. the proportion of “interstate” trips will be smaller if the state itself is large. Or in other words. and ♦ Generation of trip ends by applying household trip rates to households. a total of 16 states plus Canada and Mexico. this assumption is not valid for the smaller TAZs within the state itself. A spreadsheetbased method was developed following 1 2 KJS Associates. where J is the number of hours prior to the current hour. Sum the AON assignments. These data were drawn from respondents who were asked to report only those trips with destinations at least 75 miles from home during a stipulated 14day period. which has little influence over through trip travel patterns. 3. only a fragment of each trip will be assigned to the network. with regard to their statewide travel demand model. it was assumed that those through trips were less likely to change their direction of travel.2 External Travel for Other State Models Calculation of Through Nonwork Auto Travel for KySTM (Kentucky Statewide Traffic Model). However. especially when considering the major highways within the state. Many existing travel forecasting packages are now incapable of performing this type of assignment. ♦ Segregation of trips by trip purpose. For the preliminary distribution of these trips. Ibid Guidebook on Statewide Travel Forecasting 72 . Interstate Trip Rate Regression analysis conducted by the Michigan Department of Transportation (MDOT). ♦ Calculation of interstate trip rates for households based on both the NPTS and the 1990 census. Kentucky calculated the total number of productions and attractions for through “nonwork” trips as a percentage of the total traffic entering and exiting the area under consideration. The MDOT analyzed interstate trip rates with regard to total household trip rates.length. thereby assuming that the TAZs outside the state produce and attract nearly the same number of trips.1 Development of Outstate Trip Generation for the Michigan Model The 1990 NPTS data were used by Michigan for calculating outstate trip generation rates for the Michigan model. ♦ Calculation of trip rates for states not found in the NPTS data.
It may be possible with custom software to perform trip distribution and mode split simultaneously. where the trip distribution decision is made after the mode split decision. Approach #1. automobile trips tend to be short. Airline trips tend to be long. The outstate trip attractions at statewide zones were also assumed to be roughly inversely proportional to the distance from the nearest border. t ij = 1 ln[exp(αt ij1 ) + exp( αt ij2 )] α 1 Wilbur Smith Associates. n*m. It is possible to include a composite utility term in the trip distribution model to better account of the availability of alternative modes. the use of composite utilities in statewide models is considerably more important. A larger mode split to airlines would imply a lengthening of trips or a spreading of the distribution. Kentucky Transportation Cabinel. where n is the number of zones and m is the number of modes. Because typical fourstep models calculate distribution ahead of the mode split step. Florida DOT. There are two approaches that may be tried. The attractions for this separate trip purpose were set proportionally to the attractions for other trip purposes. Statewide Highway Traffic Forecasting Model. It is also possible to use a nested logit model. This method produced a reasonable number and distribution of through nonwork trips on all the major corridors at the edge of the modeling area. Future year projections are made by applying Fratar factoring to this trip table. April 1997 The Corradino Group. The need for such adjustment should be obvious when the different distributions for airline and automobile trips are considered.Validation and Refinement of 1990 Statewide Model. Final Technical Report . The Florida Statewide Model has assumed a separate trip purpose for outstate trips with one end in Florida. a multinomial logit model is calibrated to estimate shares going to each destination and each mode. Approach #2.1 External Trip Generation for Florida Statewide Model. The use of composite utilities in urban models is considered to be only a small refinement.this assumption. The productions and attractions at all the external stations were then balanced using the Fratar method. A two mode form of a composite utility function is shown below. it is not possible to cleanly account for multiple modes in trip distribution. The Statewide model considers all trips with an origin or destination outside Florida to be external. August 1990. Kentucky Statewide Traffic Model Final Calibration Report. Technical Report 3. Florida DOT. August 1994 2 Specialized Methods for Forecasting 73 . Post. In essence. The theory underlying composite utilities suggests that destinations are more attractive when there are many available modes between the origin and destination. Because of the widely varying modal technologies for intercity travel.2 Composite Utilities for Trip Distribution There does not seem to be a truly satisfying method for adjusting trip distribution to account for the availability of alternative modes. Buckly. Schuh and Jernigan. The external trip generation model for the Florida Statewide model was based on external trip tables developed from roadside interviews conducted as part of the Florida Department of Transportation (FDOT) Statewide OriginDestination Survey. Such a model involves many choices.
The coefficient α is calibrated. as illustrated in the table below. is computed as the composite utility of all modes in the corridor. E is employment. The composite utility is always less than or equal to the smallest utility among the available modes.5621 0. Eijp is a socioeconomic variable. P is population.4371 0. It had the form: Tijp = e where B0 p (E ) (C ) B1p ijp ijp B2 p Tijp is the OD volume for a purpose.0104 0.3499 Other 0. WisDOT calibrated this model for each of four trip purposes. Total Travel in Corridor: Wisconsin Model Wisconsin used the following equation to forecast volume increases for all traffic in a corridor.9046 0. and α is a calibrated coefficient. The level of service variable.5478 0. B2p are calibrated coefficients.5894 0 0. regardless of mode. Guidebook on Statewide Travel Forecasting 74 . Cijp is the generalized cost of travel. and B0p. Variable Population Employment Per Capita Income Level of Service Commute 0. This equation was used by the Wisconsin Department of Transportation to estimate demand for a high speed rail alternative in the ChicagoMilwaukee corridor.2878 0.3007 0.3711 0. f denotes future year and b denotes base year. L. but many planners have found that a good value of α can be taken from the coefficient on invehicle time from the mode split model.3662 0. B1p. L is level of service. dependent on trip purpose. The value of tij is used in the gravity model as the measure of spatial separation.where tij1 is the utility of mode 1 between i and j. tij2 is the utility of mode 2.5081 0. This equation can be extended to more than two available modes by adding more “exp” terms inside the square brackets.3606 TriState Total Demand Model The TriState study also created a total demand model for a corridor.1865 Business 0. Vf Pf = Vb Pb α Ef E b β If I b γ Lf L b ε where V is volume.1898 Recreation 0. I is income. Greek letters are calibrated parameters.
This method is particularly well suited for project level analysis. VOFmp = value of frequency for mode m and purpose p ($ per hour between departures).750 2. resulting in the calibrated coefficients shown in the following table.241 1. commuting and other. forecasts of zonal activity from time series methods can be directly related to traffic levels for a chosen facility.E = (Population)(Annual Household Income) Generalized cost of travel is found from the following equation.709 0. Cijmp = ITijm + 2(OTijm ) + Pijm + where: TCijmp VOTijmp + (VOF )(OH) mp (VOT )(F ) mp ijm ITijm = in vehicle time between zones i and j and mode m (hours).814 2. TCijmp = travel cost between zones i and j. it is possible to use results from a fourstep model to “pivot” about known levels of traffic. Pijm = interchange penalty in units of time (hours). OTijm = out vehicle time between zones i and j and mode m (hours). other . The socioeconomic variables chosen for the trip purpose was: ♦ business . VOTmp = value of time for mode m and purpose p (1990$). Value of time (VOT) and value of frequency (VOF) were found using stated preference techniques. mode m and purpose p (1990$). Purpose Business Commuting Other B0 0. Errors in link volumes can be large. Consequently.136 B2 2.186 B1 1.E = (Employment)(Annual Household Income) ♦ commuting. both in the base year and for alternatives. Select link analysis can be used to obtain a relationship between zonal activity and traffic levels. where only a few links are being analyzed and where highly accurate forecasts are essential for each facility. The method applies to any situation where traffic will not be redistributed due to major network changes or Specialized Methods for Forecasting 75 . However. OH = operating hours per week (hours). The model was calibrated using linear regression on travel survey data.The three trip purposes were business.596 Hybrid Technique: Pivot Analysis Four step models are often unsuitable for project level analysis. and Fijm = frequency of departures per week between zones i and j and mode m.169 1.080 1.
A “reasonable” level of calibration means that the model provides good estimates of overall system performance. the AB flow is 60 trips and the AC flow is 40 trips. which is of limited value when equilibrium assignment is the chosen method. Step 2. For example. it is usually necessary to restrain the number of selected links and to filter out OD flows that are trivially small. A related technique is select zone analysis. Obtain good average volumes for the link of interest. Some software packages also implement a form of select link analysis where it is possible to determine the number of vehicles using a pair of links. In this case. but the pivot technique would be difficult to defend when the basecase forecast is in error by more than 50% (high or low) on a given link. A reasonably well calibrated fourstep model should be available. Review of Select Link Analysis The conventional form of select link B analysis reports the OD flow matrix LINK for a single direction on a single link. Select zone analysis is particularly helpful when trying to determine the impact of site developments on surrounding streets. Counts should be taken for both directions of travel and by hour of the day. the volume Guidebook on Statewide Travel Forecasting 76 . There is no need to adjust and rerun the model for future scenarios. Some forecasting packages compute their selected link flows from an allornothing assignment. A select link analysis. Step 1. It is important that the select link analysis be compatible with the method of traffic assignment. There are four A zones for a possible 12 interzonal OD pairs. In a large network with many zones. dayofweek and seasonal factors have been eliminated. a huge amount of data may be generated by a select link analysis. Pivot Steps The pivot method may be implemented in seven steps. In the network shown earlier. 60 The figure on the right illustrates a select link analysis.capacity restraint. between Oak and Maple) and later on Ninth Street (northbound. “Good agreement” is subjective. These OD pairs and their associated volumes are D listed in a select link analysis. properly executed. An advantage of this technique is that is requires only one basecase run of a fourstep model. will tell the analyst which zones and OD pairs contribute most to a link’s volume. To achieve an understanding of the results. This method finds the number of vehicles using each link that have either their origin or destination in a given zone. has an RMS error in link volumes no more than twice what would be anticipated from traffic counting variation alone and includes all links in the project. Counts should be repeated to assure that statistical variation. an analyst might want to determine the number of vehicles traveling on First Street (northbound. the links in the project show good agreement between actual volumes and the basecase forecast. In addition. stock adjustment factors should not be used. Ideally. between Elm and Spruce). Two of the OD pairs (AB and 40 C AC) send trips along paths that use the selected link.
The multiplicative assumption of the previous example would not hold when large portions of the state are growing at about the same rate. which is considered acceptably close to the 100 vehicles from the basecase forecast.139).on the project link is 130 vehicles.035 = 1.46 B growth factor = 1. The fraction contribution of all OD pairs should be obtained from the select link analysis. These steps should then be repeated for the other links in the project. In this example. it is possible to estimate the amount of actual traffic associated with each OD pair. Step 3. a measure of zonal activity closely related to intercity travel is total personal income. In the previous example. there are two OD pairs contributing volumes: AC contributes 40 vehicles.55)*(1+0. concentrating on a few of the largest OD pairs. growth in traffic for the OD pair should be roughly related to the average growth rate of zones. In this case.035*0.43 Specialized Methods for Forecasting 77 . the forecast is 270 vehicles (105 + 165).5%.52 AC growth factor = 1. or 60%. the 130 vehicles can be split 60/40 across the two OD pairs.41*1. As indicated previously. The forecasted vehicles for AC is 105 (52*2. Step 5.98 = 1. if zone B is expected to grow by 55% and zone C is expected to grow by 38%.015) and the forecasted vehicles for BC is 165 (78*2.38 Consequently the OD factors can be easily found: BC growth factor = 1.41*1. 52 vehicles come from AC and 78 vehicles come from BC. so the AC growth factor is 201. The forecasted link volume can found from summing the contributions from all OD pairs. Many OD pairs contribute very small amounts of traffic to a link. Revised. In this example. the regional average growth rate is 41%. The growth in activity in an individual zone is forecasted by a time series. All other OD pairs are ignored. Step 7. Thus. Step 6. Pivot Method Refinements Use Larger OD Pairs. A selected link can have contributions from many OD pairs. From the drawing. The gravity model would suggest that the growth factor for an OD pair is related multiplicatively to the growth factor for each zone. It is possible to eliminate many of those smaller contributions.9% growth The growth in zone A is 46%. then the growth factor for the OD pair can be found from: BC growth factor = (1+0. A workable compromise is to assume that an OD pair growth equals the regional average growth times a factor calculated from the degree to which the OD pair exceeds the regional growth.98 = 1. especially when an equilibrium assignment or a stochastic multipath assignment has been performed.41*0.41*1.10*0. The growth factors for each zone in the example can be easily split into regional and local components: A growth factor = 1.55 C growth factor = 1. Forecasting OD Flows. The following calculations illustrate this concept. For example.139 or a 113.10 = 1.41*1. or 40%.38) = 2. Forecasting the growth in an OD pair requires an assumption about the interaction between two zones. Step 4: Assuming that the fraction contribution of OD pair to the link’s volume is accurate. and BC contributes 60 vehicles.98 = 1.
Defaults can give good starting points for calibration exercises or can eliminate the need for some locallycollected data. Some states have applied results from national databases (e. such as linear regression or maximum likelihood estimation. only few states have calibrated statewide models. ♦ Use link volumes near special generators to establish trip generation relationships for those generators. Many transportation planning agencies have adopted a joint validation/calibration strategy to increase the perceived accuracy of the model. Each model step has one or more parameters that can be adjusted to assure that the step is replicating known travel behavior. National Defaults The calibration process can be greatly accelerated if parameters are adopted from national databases. Calibration applies to each step in the modeling process. Validation. or studies performed elsewhere. NPTS or CFS) to their statewide model with good success. Calibration is the process of performing that adjustment. At this writing. However. but the results are less valid than if a good calibration had first been achieved. The use of the NPTS is facilitated because the raw data is available on CDROM. Guidebook on Statewide Travel Forecasting 78 . Unfortunately. the potential to share parameters increases. Calibration and Validation Calibration and validation are separate tasks. To perform such a hand refinement.With these revised growth factors. The desire to meld the validation and calibration process is understandable. There is no formal provision in the validation process to improve the accuracy of the model. A poor quality validation would indicate the need for additional calibration. it is usually necessary to adopt the remaining parameters from an earlier study in the same city or from a similar study in another city. Validation primarily involves comparing a baseyear forecast to known traffic levels. A Muddle. it is possible to refine by hand one or two parameters in a step by adjusting them to match known aggregate measures of travel. as more models become operational. the CFS data is aggregated by state or NTAR. such as NPTS or ATS. Very often calibration is performed by statistical methods.52 = 193 which is considerably less than the 270 vehicles calculated earlier. while validation applies to the model as a whole. although many transportation planners try to do both at the same time. In some cases. Some states have tried to use the parameters of NCHRP #187 (superceded by NCHRP #365). ♦ Use link volumes on major roads to introduce Kfactors into a gravity model. but these parameters have been created for urban applications. and ♦ Use individual link volumes to adjust free speed on links. Calibration..g. the forecasted volume on the project link is: volume = 52*1. a dependable source of default parameters for statewide travel forecasting does not exist. These agencies use the validation data in the following ways: ♦ Use measures of total travel on a network to adjust the trip production equations. At this time.43 + 78*1.
can be adequately described by its inputs: Va = volume of traffic on link a. thus. These proportions are most easily found by traffic assignment. whereas the complete SSOD data set involved many “stations” and. several links. the details of these methods are beyond the scope of this guidebook. The proportion of trips between i and j on link a. then compared the resulting assignment with a select link analysis. An alternative method worth investigating is to statistically adjust an estimated trip table to match traffic counts. needs further explanation. Michigan assigned the SSOD trip table back to the network. They must satisfy the requirement: Va = ∑ ∑ paijTij i j where Tij is the final trip table. thereby testing all steps in the model together. and tij = prior trip table of trips from origin i to destination j The method finds the “maximum entropy” trip table given these input data. as SSOD data was not directly used in the model’s development.1 Michigan’s application of this procedure would be better classified as a calibration tool. paij. One method. Common practice has been to adjust the gravity model with “K” factors to account for sizable discrepancies in the gravity model. 1 if it is used). such as bridges and state borders. To do so. entropy maximizing. 1 KJS Associates. The Michigan Department of Transportation has used recent SSOD weekday surveys in the validation phase of their model. A smallscale test of this concept has been tried in Wyoming. The selected link assignment output from the model can be directly compared with the SSOD survey data. For validation. A major shortcoming in the MDOT model’s selected link analysis was that only one link could be selected per run of the model. Inc. This is done by means of select link analysis. When allornothing assignment is performed.Trip Table Estimation from Traffic Counts A good statewide trip table can be difficult to approximate from a gravity model. The prior trip table could be taken from surveys or estimated by a gravity model or Fratar model. because it focuses entirely on the assignment step. the p’s are Boolean (0 if the link is not used. and large biases can be introduced by barriers to travel. and the volumes are taken directly from traffic counts. Several methods of estimating OD tables have been described in the literature. With any multipath techniques (including equilibrium). (1) friction factors may vary considerably across the state. Specialized Methods for Forecasting 79 . (2) errors in estimating the number of intrazonal trips may unduly influence the accuracy in interzonal trips. fractional values are possible. paij = proportion of trips from origin i to destination j carried by link a. it would have been better to have assigned the full trip table as calculated in the trip distribution step. For instance. SSOD in Validation SSOD data is an important input in the validation phase of the model. either as individual links or as groups of links.
because they do not allow an unequal number of observations in cells. If the selected links in the statewide model correspond exactly to the external stations in the urban model. the number of zerovalued cells should be reduced by manual adjustments of the trip table. the select link OD flows are theoretically equivalent to the EE (externaltoexternal) trip table in the urban model. In order to assure that the urban models are properly using statewide model results. it is possible to include continuous variables into the model. analysis of variance (ANOVA) can used to determine whether a categorization or a classification scheme is appropriate. The estimation of trip productions is most often accomplished by category analysis or by cross classification. A common problem is that the forecasted traffic from the statewide model in the base year fails to match the known traffic counts. Trip Attractions. analysis is the primary method for calibrating statewide travel demand models. Trip attraction equations are typically of the form: A = a 0 + a1x 1 + a 2 x 2 + L + an x n where the x’s are measures of zonal activity (population. Guidebook on Statewide Travel Forecasting 80 . Appendix: Review of Calibration Methods Ordinary Least Squares (Linear Regression) Ordinary least squares. Some fixing is usually necessary to account for errors in the statewide model. This type of select link analysis gives the OD flows between many pairs of links. In both methods the parameters can be estimated using a dummy variable to represent each possible category or cell. with the results from the statewide models used elsewhere. Trip Productions. the statewide model must be capable of performing linktolink select link analysis. employment. The ANOVA routines found on spreadsheets are inappropriate. Otherwise. etc. Generally speaking. Trip Productions with a Covariate. a trip production model might consist primarily of a crossclassification procedure with cells representing all combinations of household size and automobile ownership. urban models show greater geographic and geometric detail and should be far more accurate within the urban area. To the extent possible using local knowledge of existing travel pattern. urban model results should be used for links within urban areas.Integration of Statewide and Urban Models Statewide and urban models use different types of data and different algorithms.) and the a’s are calibrated coefficients. It is important to tell the statistical software to set the constant term (“a0” in the above equation) to zero in order to force the line through the origin. For example. When the parameters of a trip production model are estimated with linear regression analysis. The estimation of trip attractions requires little more than an elementary knowledge of linear regression analysis and can be easily performed on a spreadsheet. zero activity in a zone would not be associated with a zero amount of trip making. Another problem is numerous zerovalued cells in the OD matrix. or linear regression. Errors of this sort can be overcome by Fratar factoring the trip table. Consequently. This problem arises when the statewide model uses allornothing assignment. A related method.
Nonlinear Regression Nonlinear regression is implementable on a spreadsheet using a “solver” feature. Maximum likelihood estimation begins by specifying a “likelihood function”. min ∑ [Yi. logarithmic transformations are entirely appropriate when the size of the error is proportional to the size of the dependent variable.actual − Yi. Linearization.estimate ] i 2 where each i represents a data point. Logarithmic transformations disturb the distribution of errors of the dependent variable. maximum likelihood estimation finds the parameters of a model. However. In nonlinear regression.The model might be enhanced by including an accessibility term. All parameters would be estimated at the same time. which computes the probability of seeing an exact replication of observed choice patterns. there are many possible choices in a data set and the probability of an exact replication is extremely low. Maximum Likelihood Estimation Certain model steps that involve estimating probabilities are often calibrated with maximum likelihood estimation. regardless of the parameters. As discussed in an earlier chapter. Consider the total demand model from the TriState High Speed Rail Study: Tijp = e B0 p (E ) (C ) B1p ijp ijp B2 p The parameters of this equation were estimated by first taking the natural logarithm of both sides of the equation: ln(Tijp ) = B0p + B1p ln(Eijp ) + B2p ln(Cijp ) This equation is now linear in the logarithmically transformed variables.estimate. Acceptable fits of nonlinear relationships can often be achieved by transforming them so that their parameter can be estimated with linear regression analysis. These steps include mode split (logit and nested logit) and trip distribution (gravity model with complex friction factor functions and attractiveness factors). Like regression analysis. such as distance to the nearest regional shopping mall. Specialized Methods for Forecasting 81 . L. Yi. Usually. the following expression is minimized by varying the parameters of the model. Maximum likelihood estimation sets parameters of the model so that the probability of exactly replicating by chance the observed pattern of choices is the highest. That is. maximum likelihood estimation is designed to find predictive equations for choices and does not attempt to find the best line to fit a set of data. so linearization may not always be the best strategy for obtaining unbiased parameters.
then the magnitude of the parameter (exponential or power) should be made smaller. This equation is difficult to work with. Single parameter friction factor functions include the power function: Fij = 1 α t ij and the exponential function: Fij = e − β t ij The parameter is estimated by repeatedly running the model and adjusting the parameter up or down. as given by the choice model (containing assorted variables and parameters to be estimated). If the estimated average trip length is too large. maximizing the log likelihood function involves making it less negative by reducing its magnitude. The means of performing the maximization vary. Fast Trip Distribution Calibration When there is a single parameter in a gravity model friction factor function. If the estimated average trip length is too small. because it usually evaluates to a very small number. Thus. statisticians commonly try to maximize the logarithm of the likelihood function an equivalent and much easier task. depending upon the software. The “solver” capability of a spreadsheet is one possible method. Also. the logarithm of a number less than 1 is negative). and all other terms are negative numbers (i. t = t* where t “hat” represents the model’s average trip length and t “star” represents the actual average trip length.e. That is.1) variable that is set equal to 1 if the traveler n had chosen mode j and set to zero otherwise. that parameter can be set so that the model yields the same average trip length as observed in reality. δjn is a Boolean (0. Thus. The choice between the power function or the exponential function (or any other convenient oneparameter function) must be made by comparing the differences between the estimated and observed trip length frequency distributions.. then the magnitude of the parameter (exponential or power) should be made larger.L = ∏ ∏ Pjnjn δ n =1 j=1 N J Here Pjn is the estimated probability of any given traveler n choosing mode j. Guidebook on Statewide Travel Forecasting 82 . max L* = ∑ ∑ δ jn ln Pjn n=1 j=1 N J It is readily seen that all terms in the log likelihood function that represent unchosen modes are zero. The true average trip length can be ascertained by questionnaire or by analysis of NPTS.
Randall Wade served as WisDOT’s main contact. with major contributions by David Farmer and Smitha Vijayan. The guidebook was written by Alan J. FHWA’s staff contributing to this guidebook include Leroy Chimini. Robert Gorman and Stefan Natzke. Additional guidance was provided by Tony Esteve and Phil Hazen (retired) of FHWA. The appendix was adapted from David Farmer’s MS thesis. Specialized Methods for Forecasting 83 .Milwaukee. University of Wisconsin . Professor of Civil Engineering. Horowitz.Acknowledgements The project was conducted by the Center for Urban Transportation Studies. The project was coordinated with the Wisconsin Department of Transportation. Editorial contributions were made by Linda Rupp.
5. Guidebook on Statewide Travel Forecasting 84 . Assess demand by specific customer/market segments. Statewide models help to tie the decisionmaking process to a knowledge of the interaction between transportation systems and socioeconomic structures at a statewide scale. Connectivity between metropolitan areas.2]: 1. More specifically a statewide forecasting effort can. Many states expended a significant amount of time and expense – especially in the late 1970s and early 1980s – to develop statewide models. These planning factors have been consolidated in TEA 21. Most recently. 6. 5. The use of travel forecasting models can assist with decisions regarding future facility needs. Some succeeded in developing a working model.1. but most did not. Longrange needs of the state transportation system. Recreational travel and tourism. 3. and Integrate with and provide input to urban models. but the need for a statewide model remains. Preservation of rightsofway for future projects. there has been renewed activity in travel forecasting at a statewide level.1) outlined in the ISTEA legislation [1. 6. Provide quantitative data input to management systems. Transportation needs for nonmetropolitan areas. In fact. Enable multimodal analysis along major intercity corridors. most states do not have a forecasting process in place at a statewide level – whether due to past difficulties in developing and maintaining a statewide model. attempts have been made over the last 35 years to formulate models and techniques to forecast transportation activities on a statewide scale. and Methods to enhance the efficient movement of commercial goods. The purpose of this appendix is to present a review of the state of the art travel forecasting at the statewide level and to identify possible avenues for improving the forecasting process. implementation of most of the 23 “planning factors” (see Figure A. Forecast passenger and commodity flows for a 20year horizon. However. as one report noted [1. or due to a cynicism bred by the often highly political process of transportation planning. 3. encouraged by early progress in the development of urbanarea models.1] can be assisted greatly by a statewide forecasting effort. Provide an improved tool for trunkline planning and analysis. Introduction The belief has long been held that the characteristics of travel on a statewide scale are in many ways different than those of travel within an urban area. budget projections and the assessment of the large scale effects of alternative projects. At present. 2. in reaction to the requirements of the 1990 Clean Air Act Amendments (CAAA) and the 1991 Intermodal Surface Transportation Efficiency Act (ISTEA). 4. Statewide forecasting can have a significant impact on at least six of these “factors” including: 1. Thus. a formalized statewide forecasting process offers a rational basis for making planning decisions. 2.APPENDIX: The State of the Art in Statewide Travel Demand Forecasting A. 4.
or local energy use goals. 3. 16. national parks. The State of the Art in Statewide Travel Demand Forecasting 85 . The results of required management systems. however. tunnels or pavement.3] has attempted to differentiate between the “state of the art” and the “state of the practice”. The use of innovative mechanisms for financing projects. 22. 13. The coordination of transportation plans and programs developed by MPOs. 5. 6. Summary of ISTEA Statewide “Planning Factors” It should be noted that the title of this appendix deals with the “state of the art” in travel forecasting. objectives. The transportation needs of nonmetropolitan areas. Transportation system management and investment strategies to make the most efficient use of existing facilities. monuments and historic sites and military installations. state. in fact. 10. Methods to expand and enhance transit services and to increase the use of such services. The connectivity between metropolitan areas within a state or with metropolitan areas in other states. 7. 20. 2. This appendix. Methods to enhance the efficient movement of commercial motor vehicles. Any state plan developed pursuant to the federal Water Pollution Control Act. 8. 23. 12. 11. Concerns of Indian tribal governments. Any federal. merely part of the same continuum. The overall social. recreation and scenic areas. Some previous research [1.1. Any metropolitan area plan. 19. Recreational travel and tourism. Figure A. Investment strategies to improve adjoining state and local roads that support rural economic growth and tourism development. airports. The use of lifecycle costs in the design and engineering of bridges. The effect of transportation decisions on land use and land development. intermodal transportation facilities. federal agency renewable resources management and multipurpose land management. Strategies for incorporating bicycle transportation facilities and pedestrian walkways. The approach followed here assumes that the knowledge that is elsewhere divided into “art” and “practice” is. programs or requirements. energy and environmental effects of transportation decisions. International border crossings and access to ports. 21. Strategies for identifying and implementing transportation enhancements. 9. major freight distribution routes. makes no such distinction. 14. Longrange needs of the state transportation system for movement of persons and goods.1. 18. Preservation of rightsofway for construction of future transportation projects. 15. 4. economic. 17. Methods to reduce traffic congestion and to prevent congestion from developing.
As also shown in Table A. Section A. in general.1.2 and Section A. Many DOT officials were directly accessible via email or could be easily reached through a general information email addresses at their respective DOT web sites. respectively. Alaska and Hawaii were not contacted because of their geographical separation from the lower 48 states. As shown in Table A.8 provides some recommendations for “best practice” based on the observations made in the preceding sections. presents a review of two statewide freight models now under development: Multimodal Freight Forecasts for Wisconsin and Transport Flows in the State of Indiana. Section A.4 and Section A. which (as discussed in section A.6 presents a review of two statewide passenger models now under development: the Michigan Statewide Travel Demand Model and the Kentucky Statewide Traffic Model. A similar.5 provide a review of statewide forecasting methods for passenger travel and freight transportation. This search was supplemented by an examination of the references found in applicable TRB and NCHRP literature and through online searches of the Northwestern University library. Other contacts could be made through access to DOT phone lists that are also available online. Finally.Method of Research In order to assemble information for this appendix. National Cooperative Highway Research Program (NCHRP) reports and syntheses and other published materials available in the collection of the Center for Urban Transportation Studies at the University of WisconsinMilwaukee. likewise.7. an overwhelming majority of the states contacted replied via email or telephone. respectively. and Mississippi was not contacted because its DOT did not support a home page on the internet. Section A. and many sent documentation of their individual statewide forecasting models and procedures. attempts at contact were made with 45 states. Section A. The information in these sections is drawn principally from TRB and NCHRP publications. The formulation of intercity models. two distinct research methods were applied. Instead. Organization of Appendix The following sections of this appendix present a review of the information gathered during the research process described above and some observations and recommendations about the state of the art as portrayed by that information. historically preceded the development of statewide models. Much of the information reviewed in these sections was collected from the DOT sources. Similarly. Section A.2) bear many relationships to statewide models. electronicallybased procedure was followed in contacting several commercial sources that provide the population and economic forecasts that are often used as inputs to the transportation forecasting process. a search was made through the nearlycomplete set of Transportation Research Board (TRB) bulletins and reports.1. much of the available information describes “intercity” models. Thus. Guidebook on Statewide Travel Forecasting 86 . Rhode Island and Delaware were not contacted because of their small size. contacts were initiated using the Internet. Surprisingly little information that applies specifically to statewide travel forecasting was discovered in these searches. the intercity models are presented first. For as many states as possible.3 provide a review of the literature of intercity models for passenger travel and freight transportation. First. The second method of research consisted of direct contact with various state department of transportation (DOT) officials.
regional and national models. 2. Statewide models are therefore a subset of intercity models. As used here. The term “intercity” is also used to distinguish these models from “urban” models. they generally present situations that are a little more abstract in nature. The similarities to statewide models are many. Consequently. The State of the Art in Statewide Travel Demand Forecasting 87 . The intercity models encountered in the literature are often associated with an academic exercise and therefore make use of fewer. The main point.1: Contact with State Departments of Transportation DOT AL AK AZ AR CA CO CT DE FL GA HI ID IL IN IA KS KY LA ME MD MA MI MN MS MO Contact Made Yes No Yes Yes Yes Yes Yes No Yes Yes No Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes No Yes Reply Rec’d No Yes No Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Items Sent No Yes Yes Yes No No No Yes No Yes Yes Yes Yes No No Yes Yes No DOT MT NE NV NH NJ NM NY NC ND OH OK OR PA RI SC SD TN TX UT VT VA WA WV WI WY Contact Made Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes No Yes Yes Yes Yes Yes No Yes Yes Yes Yes No Reply Rec’d No Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes No Yes No Yes Yes Yes Yes Items Sent No No Yes Yes Yes Yes No No No No No No No No Yes Yes No No Yes Yes Yes A. statewide. It is assumed that people travel according to a somewhat different set of rules over longer distances and between metropolitan areas. is that the characteristics of intercity travel are inherently different from those of travel within an urban area. more carefully chosen origindestination (OD) pairs than would normally be included in a meaningful statewide model.2].Table A.1. which typically involve travel between more closely spaced points of interest within a localized area. the term “intercity” forecasting involves the prediction and assignment of traffic volumes between cities or other points of interest that are separated by some significant distance. first expressed as early as 1960 [2. Intercity Passenger Literature Intercity travel is a broad heading that includes statewide travel. Intercity models include corridor.2.
Peers and Bevilacqua describe a model that includes a long list of policysensitive variables. Intercity travel demand models can be further classified by whether they encompass only a single mode (modespecific) or multiple modes (total demand) and by which trip purposes they include. including population and employment. The models can make use of either aggregate or disaggregate data and can be of a directdemand or sequential structure. The following directdemand models – some of which are not mentioned in those references – are noted here because they possess features that might prove useful to modeling at the statewide level. Models presented by Peers and Bevilacqua [2. The four resulting combinations are (1) aggregate directdemand models. income per household Guidebook on Statewide Travel Forecasting 88 . Yu took the standard directdemand formulation – regressed from crosssectional data – and recognized that the elasticities present in the crosssectional data would not necessarily remain constant over time. Disaggregate data goes further to examine the motives and characteristics of the trip makers at an individual or household level and are typically used to generate the probability that a particular trip is taken or mode is used. including persons per household. which many DOTs have adopted for the statewide modeling purposes. Aggregate data makes use of the socioeconomic data for the OD pairs in the model and can also include the service characteristics of the modes of travel between them. however.Types of Intercity Passenger Models A number of reviews have been made of the early history of intercity modeling [2. It is a novel idea that does not appear to have been picked up by succeeding authors.6]. 2. The reader is referred to the reviews referenced in the previous section (especially Koppelman et al [2. (3) disaggregate directdemand models and (4) disaggregate sequential models. arranged into three groups: (1) extensive variables.6]) for a more complete historical perspective of significant intercity modeling efforts. Here. The procedure is intended to eliminate the effects (on the traffic volumes to be forecasted) that result from variables that have been excluded from the models. and use of this procedure does not seem to have been adopted by other researchers. (Direct demand models are sometimes called econometric models because of their resemblance to statistical models of economic demand. Another innovative idea is found in Cohen et al [2.4.12] give some sense of this trend. 2. Intercity models can essentially be divided into four types on the basis of two categories: data and structure.) A sequential model. 2. as part of two singlepurpose (business and nonbusiness) directdemand models.5. on the other hand. 2.9]. The most famous of these was Quandt and Balmol’s abstract mode model [2.3. the authors propose to use a pivotpoint procedure.10]. (2) aggregate sequential models.8]. singly calibrated step. By the late 1970s directdemand models were being constructed to include an increasingly wider range of variables to account for the enormous variety of factors that influence travel behavior. Description of the pivotpoint procedure is brief. The urban “fourstep” modeling process. His paper presents two singlepurpose (one for business travel and one for personal travel) directdemand models in which the regression coefficients each include a timeseries component.7] and most include some discussion of the taxonomy of intercity models. presents the quintessential example of a sequential model. A notable early innovation was attempted by Yu [2. divides the modeling process into several individually calibrated steps. (2) intensive variables. In terms of model structure. Aggregate DirectDemand Models The earliest intercity models were of the direct demand type and were developed in the 1960s as part of an examination of the Northeast Corridor [2.6. a direct demand model is one that calculates all of the desired travel information in one.11] and Kaplan et al [2.
as well.6] provides a review of many of the earlier disaggregate mode choice models. most notably for the shortage of passenger rail cars. They were thought to be especially useful in the effort to estimate the shifts in mode share that were expected from deregulation in the air and intercity bus industries and from the anticipated implementation of highspeed rail transportation [2. policy related and behaviorallybased variables in the modeling process. including travel speeds and costs. 2. 2. It is possible to develop a modechoice model without disaggregate data.2 [2. Disaggregate models.23. The Egyptian Intercity Transportation Planning Model estimates travel on seven modes for travelers in three income levels. as DiRenzo and Rossi did. Miller [2. 2.25]. the Egyptian model could reasonably be noted in the section of this appendix describing statewide forecasting techniques. it resembles some freight models. It might also be noted that.17].24. and (3) system variables. Disaggregate Sequential Models One of the first applications of disaggregate (or behavioral) modeling was for the modechoice step of sequential models. a multimodal model that explicitly includes consideration of accessibility to the transportation system. several researchers have developed complete travel demand models based on the analysis of disaggregate data in a number of discrete. Due to parallel research in urbanarea forecasting in the early 1980s. transportation accessibility. unlike many intercity passenger models. Koppelman [2. for instance. nested steps. Morrison and Winston. using a “reasoned’ diversion model [2.18] provide studies of the various structures (binomial. In addition. Similarly. Both of these models provide a bridge from an earlier emphasis on aggregate modeling to the growth in disaggregate modeling research by the early 1980s.and employment per acre.20] and Koppelman and Hirsh [2.21.16]. Kaplan et al describe their Passenger Oriented Intercity Network Travel Simulation (POINTS) model.22] present a multimodal model with a structure shown in Figure A. multinomial and nestedmultinomial) available to more realistically represent the crosselasticities between modes and to eliminate irrelevant alternatives in the logit modesplit formulation.3. these models became more attractive.14. while Koppelman and Hirsh use both the NTS and the 1977 National Personal Transportation Survey (NPTS) data. Armed with an increasing understanding about the implementation of disaggregate modeling techniques and fueled by the increasing availability of disaggregate data. Koppelman et al [2. present multimodal models (one for vacation travel and one for business) with the hierarchical structure shown in Figure A. Again. it includes capacity restraints on the network. Both pairs of researchers seek to use this disaggregate data in a model structure that mimics the behavioral logic of trip making. in its treatment of rail car capacity restraints. Morrison and Winston make use of the 1977 National Travel Survey (NTS) data. 2.15]. but since the transportation situation in Egypt is sufficiently an abstraction of the situation in the United States. The State of the Art in Statewide Travel Demand Forecasting 89 .13]. Forinash [2. Also. It is unusual in its use of disaggregate data in a single equation (directdemand) format. and Forinash and Koppelman [2. Because it deals with a very practical situation. however. One Disaggregate DirectDemand Model Another model of interest is the disaggregate directdemand model developed in the 1980s by the Egypt National Transportation Study [2.19]. Meanwhile. typically use a logit formulation to provide a convenient way of including a number of modeabstract. it seems fitting to include it with the intercity models.
Mode j Rent a car Don't rent a car Figure A. Structure of Morrison and Winston’s Model Trip Frequency One None Destination 1 Destination 2 . . . . . Destination i Mode 1 Mode 2 . . . Structure of Koppelman and Hirsh’s Model Guidebook on Statewide Travel Forecasting 90 .Make Trip Destination 1 Destination 2 . . .3. . Service Class k Figure A. Destination i Mode 1 Mode 2 . .2. Mode j Service Class 1 Service Class 2 .
SingleMode and SinglePurpose Models Besides the ubiquitous singlemode automobile models.4. but it is also due to the general failure of disaggregate techniques at a statewide scale. This is partly due to the strong traditions of and training in the fourstep modeling process.31] and an elaborate directdemand model of intercity air travel based on qualityofservice measures by Ghobrial and Kanafani [2. He concludes that this formulation provides a simpler and more reasonable estimate of ridership on rural bus routes. Pickrell [2. His model describes a probabilistic (disaggregate) model based on a Poisson distribution of ridership. The history of interregional inputoutput analysis – which The State of the Art in Statewide Travel Demand Forecasting 91 . their principal drawback is the lack of sufficient disaggregate data for calibration of statistically meaningful statewide models. the intercity forecasting techniques employed in most existing statewide models are principally those of the aggregate sequential type. published in 1974.26] and examples of intercity bus models are rare. for instance) to forecasts of socioeconomic data. Other air travel models of interest include a regression analysis of travel between small cities in Iowa by Thorson and Brewer [2. Several air travel models are also of interest. In another study. as opposed to a regression model.30] uses a combination of techniques to assess future trends in intercity air travel. is one of the latest papers found to specifically address the recreational trip purpose.28] addresses this issue with simple linear regression techniques. Intercity Freight Literature Introduction Intercity freight models are similar to intercity passenger models in their attempt to model traffic – in this case. Finally. Until further data is available their use will remain limited. One interesting bus model is presented by Neumann [2. the amount of intercity freight forecasting literature is rather small. he uses an aggregate modechoice model to predict the percentage of market share that the air mode could generate under several alternative futures. and a 1968 paper by Brown and Watkins [2.3.33]. Compared to the amount of literature available about intercity passenger forecasting. Discussion As will be seen in the following sections. (Most passenger rail models are a part of a multimodal model. the one other singlepurpose intercity model worth noting is the disaggregate model of recreational travel presented by Gilbert [2. Although disaggregate models are attractive because of their ability to include the behavioral aspects of travel. but more will be said about recreational travel models in Section A.29] also looks at timeseries analysis of air traffic patterns using a BoxJenkins procedure to account for cyclical (seasonal and yearly) variations in travel behavior. there are two other types of singlemode models of interest: bus and air. freight traffic – between spatially distant locations. This econometric type of model can be especially useful in tying the forecast of single quantity (annual VMT or emissions. It should also be noted that there is a place for aggregate directdemand models at a statewide scale. It should be sufficient to state here that Gilbert’s paper. A. As early as the 1960s it was recognized that the yeartoyear growth in air travel makes the use of timeseries techniques valuable.) Modeling of intercity bus travel has proved to be difficult [2.32]. Pickrell uses a singlemode directdemand model to estimate the total demand for air travel. At the same time.27]. Gilbert’s model is sufficiently abstract to be included here with the other intercity models. A later paper by Oberhausen and Koppelman [2.
They formulated a model that recognizes the special situation of shippers and carriers and attempts to account for the different information available to each group. Reviews of Intercity Freight Models As in the case for passenger models. In it he applies a linear regression analysis to develop equations for national rail freight volume and truck freight volume based on GNP. Bronzini [3. but it did recognize that: In underdeveloped regions… the patterns of economic development cannot be predicted by past trends because past trends lead nowhere. That work is discussed briefly below. including Canadian research in the late 1980s and early 1990s. A paper by Smith [3.2] divides freight models into two categories familiar to passenger modelers: aggregate and disaggregate. Smith concludes that either a gravity model or an abstractmode model would be best for use in the situation where the available data are limited. He also recognizes the difficulties inherent in collecting the immense amounts of data required to calibrate disaggregate models. An important conclusion of Bronzini’s review is the need for a “comprehensive interregional commodity flow data base” – something that has been under development in the intervening years by many sources and will be discussed further in Section A. the Transportation Systems Center and the National Energy Transportation Study.3] also present an overview of early freight modeling efforts. Intercity Freight Models One of the earliest intercity freight studies was published by Morton in 1969 [3.4] traces the development of various multimodal freight transportation models at the national level. rail shipping rates and truck rates. Meanwhile. several reviews have been published about the history of intercity freight modeling. Another important observation by Bronzini regards the benefits of using an equilibrium assignment for nonhighway modes. Other intercity freight modeling efforts. Friesz et al [3.8]. with conclusions concentrating on work performed at the University of Pennsylvania. Jones and Sharp noted the lack of available commodity flow data that could make their model useful. but serious attempts to forecast freight movements at a national or regional level were generally begun only after the first intercity passenger models were developed in the middle 1960s. (2) inputoutput. rather than cause a switch of traffic to a competing mode. Their model takes OD pairs determined at the shippers’ level and optimizes the flow for the carriers’ network. (p.5. Guidebook on Statewide Travel Forecasting 92 . however.7. (4) gravity. little appears to have been written regarding workable intercity freight transportation models until the late 1970s. The models described include those developed by the Inland Navigation Systems Analysis project.4. In a 1983 paper. In 1977 Jones and Sharp [3.5.5]. in a model structure shown in Figure A. He observes that models based on aggregate data might be better for regional level freight flows. (3) inventory theoretic. 3.is closely related to some models of freight transportation – dates back to the 1930s. Throughout the 1980s researchers at the University of Pennsylvania concentrated on the differences between freight modeling and passenger modeling [3.6] published research on demand modeling for undeveloped rural regions of the United States. bears a closer resemblance to statewide modeling practice and will be discussed in Section A. 523) Like others in the 1970s. (5) abstract mode and (6) linear programming. Winston [3. His investigations indicate that using an equilibrium assignment tends to redistribute traffic to cause more efficient use of a particular modal network. Apart from nationallevel models. Their model explicitly did not have any predictive function.1] divides previous models into six categories based on their structure: (1) market share.
Intercity Freight ModeSplit Models In addition to the models that address the full extent of the freight transportation process. Of these studies. A number of methods have been employed in the effort to understand how freight traffic becomes divided among the available modes. 3.10]. as well as the probabilistic methods more familiar to urban modesplit modelers [3. is instead determined by the weight of the shipment and the length of its haul. recent studies have included air freight [3. the study concludes. The choice of modes for the vast majority of manufacturedgoods. several address only modesplit. a primary concern of freight modeling is the division of the freight flow between competing modes.9] and specialuse trucks [3. the study suggests that only about one quarter of total manufacturedgoods cargo is really subject to competition between modes. such as shipper prejudices toward one mode or another.15. 3.11] about the drastic effects on rural freight trucking in Nebraska that were brought about by changes rail freight practices. University of Pennsylvania ShipperCarrier Model The State of the Art in Statewide Travel Demand Forecasting 93 . Some attempts have been made to examine statewide freight flows for single modes. Production Amounts Consumption Amounts Demand Function SHIPPERS' SUBMODEL User Optimized Elastic Demand Assumed Aggregate Network Used OD Demands by Path from Production Site to Consumption Site DECOMPOSITION ALGORITHM Path Construction Path Decomposition OD Demands by Carriert CARRIERS SUBMODEL System Optimized Fixed Demand Assumed Detailed Network Used Arc and Path Flows Arc and Path Costs Figure A. or by other factors.12] and a diversion matrix method [3. usually between truck and rail.14. A clear example of the importance of modesplit models for intercity freight transportation is provided in Lindesmeyer’s paper [3. The methods used have included discriminant analysis [3.4. However. the most interesting is the diversion matrix study.16].13]. Although it was written before the recent explosion of truck/rail intermodal business and is based on an uncomplicated analysis.
The New Mexico report addresses both passenger and goods movement models within the broader context of statewide transportation planning.Discussion Research on intercity freight transportation appears to have been minimal.3] – effectively institutionalizing the UTPstyle model for statewide use. travel forecasts are based on some sort of travel data. only Connecticut. rather than on some formal model. New Mexico [4. A. Guidebook on Statewide Travel Forecasting 94 . 4. and by 1972 at least 19 different states were using or preparing statewide models [4. prompted by new federal legislation (CAAA and ISTEA). and little activity seems to have taken place (studies in Florida and Kansas are an exception [4. For the states that are engaged in some type of modeling process. apart from the specifically statewide research that will be discussed in the Section A. as practiced by the various state DOTs. has remained much more basic. The enthusiasm for developing statewide models that was present in the late 1960s and early 1970s soon waned. scheduled to issue a request for proposals (RFP) for a model development contract in the Fall of 1997).4. but there may be other. however. Surveys have been conducted at the statewide level since the earliest days of highway modeling [4. easier ways to address this complicated interface. 4. Statewide Passenger Forecasting Literature In spite of the amount of research involving the characteristics of intercity travel and its concentrations on econometric models and probabilitybased models. neither New Mexico nor Texas is currently involved in statewide modeling. As early as 1967 Arizona and Illinois had developed UTPstyle models [4. 4.7] and Texas [4.9] and continue to be conducted at the statewide level [4.7) that include shipper and carrier representatives to examine traffic forecasts from their individual perspectives. especially the difficulties in isolating interzonal trips and the proliferation of “Kfactors” in recent models.6]) until very recently. This is likely a function of the ready availability of urban modeling software and personnel trained to use it. forecasting is done for specific projects only.1].2. One obvious source of travel data is the survey. Research into shippercarrier problems clearly addresses an important behavioral link in the freight transportation process. the models used are all “fourstep” models. with a modeling procedure borrowed almost wholly from the urban transportation planning (UTP) process.1) no travel modeling is done on a statewide level. Modeling activities were evidently so popular that in 1973 the Federal Highway Administration (FHWA) perceived the need to standardize the thinking about statewide modeling and issued a guidebook on the subject [4. At the majority of state DOTs.2]. Data Collection for Passenger Travel Ideally. and forecasts are made based on historic trends. several states were rethinking their strategies.10. In most of the states contacted as part of the research for this appendix (see Table A. concentrates more on the details of statewide modeling. One way might be to employ expert panels (as will be discussed in Section A. By the early 1990s.11]. Apparently. (Texas is.5. which includes reviews of circa1990 models from Florida. The Texas report. and these are discussed below. however.4. Kentucky and Michigan. whether due to funding cuts or to frustration with the model results. A list of states contacted that sent information about their current passenger modeling efforts is presented in Table A. Despite this promising trend.5. Kentucky and Michigan have been continuously developing models from the earlier period. passenger travel forecasting.8] produced interesting reports that outline this renewed focus on statewide modeling.
4.16] developed statistical methods of clustering together traffic counts on different roads based on their similar functional and geographical characteristics.20.22. 4. Pennsylvania [4.18].27. US census data have always been valuable as inputs to travel modeling. The Minnesota DOT has formalized this process as it applies to forecasting traffic for their state trunk highways [4.29]. The JTW has proved especially useful in estimating homebased work trips on a statewide level. Statewide Models of Passenger Travel Of the states contacted as part of the research for this appendix. Attempts have also been made to synthesize trip tables from census data at a substate level in New Jersey [4. In addition.13]. Research in the early 1980s [4. The CTPP provides transportationrelated information at a transportation analysis zone (TAZ) level. These methods have subsequently been applied to statewide analyses in Wyoming [4.19.32]. have been used in the development of a number of statewide models.24. is used for forecasting purposes. which measures some intercity travel. Trend Analyses of Passenger Travel As noted above. for many years state DOTs have had in place systems of traffic counting equipment operating at a statewide scale.12]. Data Synthesis for Passenger Travel Even with advanced systems for traffic data collection. but are instead based on the extrapolation of trends observed in historical data. The result is a larger and more statistically valid collection of traffic count data available for use in travel forecasting.23] began to reevaluate their traffic monitoring systems to take advantage of clustering. similar to the method outlined for updating coverage counts in the FHWA’s 1992 Traffic Monitoring Guide [4. In association with the introduction of the FHWA’s Traffic Monitoring Guide in 1985 [4. Two other options make use of data that is already available: Federal survey data and statewide traffic counts. To get around this difficulty. optimization methods have been developed to synthesize trip tables from available traffic count information [4. Some indication of the possibilities of trendline analysis is given in a paper by Harmatuck [4.17]. Of course. 4.26]. but such documentation seems to be the exception.21] and New Mexico [4. those having ongoing modeling efforts sent documentation of their progress. but has been criticized for its lack of information about other purposes [4. 4. The 1990 census improved upon this by including a journeytowork (JTW) survey and by introducing the Census Transportation Planning Package (CTPP) [4. 4.However.14. In addition to the aforementioned federal government sources. it is difficult for a state DOT to collect enough data to account for all of the likely paths between OD pairs being examined.15.28]. it should also be noted that estimated and forecasted data is also available from a wide variety of state. at least one state contacted for this appendix indicated that a growth factor method. 4. 4. Otherwise little information is available on travel forecasting techniques in the absence of a statewide model. which conducted its most recent National Personal Transportation Survey (NPTS) in 1995. which can be readily aggregated into township or county level data for statewide modeling. Washington [4. 4. many of the DOT officials contacted for this appendix indicated that the only forecasts they make are not based on models. academic and commercial sources. In it he provides further insight into the particular ways of dealing with traffic data as a time series.25.31] for the Wisconsin DOT.30]. The NPTS data. Another federal data source is provided by the US Department of Transportation. A summary of the passenger models in The State of the Art in Statewide Travel Demand Forecasting 95 . they are relatively expensive to conduct and must be supplemented by other data.
49].52. in the late 1960s and early 1970s the NCHRP [4.53] conducted studies of the special characteristics of recreational travel.36]. New Hampshire [4.2. Section A. Since many state economies depend heavily upon recreational activities.45]. Florida [4.43].33. Recreational Travel Models As early as 1963 recreational trips were considered an important enough purpose to warrant separate study [4. New Jersey [4. 4. The Florida and New Jersey models are also interesting in the degree to which they have attempted to incorporate existing MPO models into the statewide modeling effort. Several other states are currently in the very beginning stages of modeling projects – issuing RFPs to interested consultants. As can be seen from Table A. In addition to the states shown in Table A. so documentation is currently unavailable for it. Oregon is also in the early stages of developing a comprehensive forecasting model that will include a land use element [4.37]. Further discussion of the structure of typical statewide passenger models is presented in Section A.28]. 4. 4.46].51] and other states [4.41. The Wisconsin model is unique in that it is essentially an intercounty model. All use fairly standard UTP procedures. but only a few modes. where large variations from recent trends are less likely. most of the models consider a large number of trip types (as many as 5 or 6). 4.2. Kentucky [4. Vermont [4. Although these socioeconomic trends are themselves forecasts.existence or underdevelopment is presented in Table A. Strangely.48.34]. 4.50. In fact. Discussion Using trendline procedures in statewide forecasting is probably better than not forecasting at all. Kentucky [4. The use of travel forecasting models. it is hoped that they broaden the basis of the transportation model sufficiently to provide a more reasonable forecast of future travel. Michigan [4. California has a statewide model.6 examines both of these models in greater detail.27.44] and Wyoming [4. with comparatively few TAZs. 4.35. New Hampshire proposes to use logit formulations for trip generation and distribution. especially for short term planning horizons. but it is being redesigned. Indiana [4. Guidebook on Statewide Travel Forecasting 96 . Indiana [4.42]. The Kentucky and Michigan models are two of the more recent useable models from states with long histories of model development and are representative of the current state of the practice. it would seem that this trip type might be important enough to require a closer examination than it has received in the past two decades. except for the model under development for New Hampshire.6.2. however. grounds the forecast in the underlying statewide and national socioeconomic trends. the last of these studies was published more than twenty years ago.40]. Wisconsin [4.38].39]. 4. This includes work done in Connecticut [4. All of the models are of the “fourstep” style.47]. although Americans seem to have dedicated an increasing amount of time to pursuing recreational activities.
NHB 5. • All trips are modeled to maximize use of MPO models. NPTS national average data used for trip generation All trips modeled – previous models did not consider local trips. 3. • Aggregate mode choice. 1. • Internal TAZs at the township level. The State of the Art in Statewide Travel Demand Forecasting 97 . HBW HB Shop HB School HB Other NHB Truck • Based on extensive statewide survey. Bus 4. • Gravity friction factors based on MPO urban models. Shopping 5. 2. 3. Time of day and seasonal factors. 2. HB Work/Biz. 6. 4. Other  • Model created by merging 5 MPO models. 5. HBW 2. HB Misc. 2. SOV 2. Under development. 4. 2. Truck 3. HOV 3. Florida 440 internal 32 external Indiana 500 internal 5060 external 1. • Iterativeequilibrium assignment for highways. HB Other 4. NHB 1. Rail Highway vehicles only PURPOSES 1. LOSbased mode split model still under development. NTPS data used for calibration. 5. Current Statewide Passenger Models STATE Connecticut TAZs 1300 total MODES 1. Truck 1. Logit trip generation and distribution. SOV HOV2 HOV3+ Bus 5.Table A. 5. Recreational 6. Michigan Auto only 1. HBO 3. 3. Transit Kentucky 756 internal 706 external 2392 total Auto only 1. • Mode split is auto occupancy only based on production zone. Other Business 3. • Under development. HBNW 3. HBW HB Shop HB Soc. Business related 3./Vac.2. HB other 4. HBW 2. • Extensive use of Kfactors. 4. Recreational 6. HBW 2. Two possible mode split models: (1) simple crossclassification and (2) LOSbased. HBW 2. CTPP data used for validation. NHB • • • • • • • • • • Model includes a large portion of surrounding states. Vermont Highway vehicles only 1. 5. Personal 4. NHB Work/Biz./Rec. NHB Truck/Taxi COMMENTS • Mode split based on LOS information. HB Soc. Auto 2. Rail New Jersey 2762 internal 51 external 622 internal 70 external  1. NHB Other New Hampshire 1 per 5000 pop. 6. Extensive use of Kfactors./Rec. 3.
as with passenger models. NCHRP Report 388 [5. At the same time. but increasingly models are using commodity flow data as their basis.1]. The other technique is to include truck trips as a separate trip purpose in the passenger model. based on survey data or on counts of truck traffic on various links in the highway network. • No external trips considered. the Bureau of the Census’s 1993 Commodity Flow Survey has been used by several states to develop their own commodity flow interactions. Current Statewide Passenger Models (continued) STATE Wisconsin TAZs 112 internal 45 external 5 internal 5 external MODES 1. can be collected by survey methods.2]. not only for population and general Guidebook on Statewide Travel Forecasting 98 .5] that began to address the data requirements of statewide freight modeling. Statewide Freight Forecasting Literature Introduction For various reasons. has historically lagged behind the development of passenger techniques.3]. There are two techniques generally applied to truck traffic analysis. NCHRP released two reports in the late 1970s [5. • Network used only to develop impedances for mode share calculations. A number of private firms also offer (for a fee) access to their collections of historical and forecast data.5. More recently. The development of freight forecasting techniques. Truck PURPOSES 1.Table A. Other COMMENTS • Under development. Bus 1. the methods of analyzing freight traffic at a statewide level have remained similar in form to those used in predicting passenger travel. Air 3.2.4. a cursory modesplit step and some sort of simple assignment. Meanwhile. These two reports present 228 different sources of data that could be used for freight forecasting. This is partly because of the numerous parties involved in shipping the large variety of commodities that are regularly moved by the several modes available. • Full trip tables estimated using entropy maximization technique Wyoming  A. 5.1] has provided an update to the list of data sources. • Model created mostly to demonstrate techniques. as has been done recently by the Washington DOT [5. instead of regression equations for employment and population.4 for passenger travel forecasting. The only significant difference is that the trip generation step is often based on freight flow data (usually classified by industry groups). Auto 2. especially for truck analysis techniques. Data for Freight Forecasting Freight data. The first technique is a simple trendline analysis similar to that described in Section A. Business 2. • Summer weekend travel is modeled. Rail 4. Forecasts that are based on commodity flows bear a resemblance to passenger models in the way they are structured. it has been suggested that forecasting freight transportation flows is more complex than modeling passenger travel volumes [5. They are typically employed in a “fourstep” sequential process that employs a gravitymodel distribution. therefore. In either case the similarity to passenger forecasting is obvious. Auto 2. There are essentially two ways that state DOTs forecast freight traffic: (1) by analyzing truck traffic or (2) by using a commodity flow model [5.
21].11] is of the growth factor type noted above.2).8] or (2) use of forecasts of future freight flows (by commodity) distributed by a gravity model.1] or are under development. Much of the currently available data is unfortunately provided only at a Bureau of Economic Analysis (BEA) region level.14] models are both of the commodity forecast type noted above. It is a relatively simple example. 5.20.3.18] present a sequential model with a gravitystyle trip distribution based on commodity flows. The State of the Art in Statewide Travel Demand Forecasting 99 . 5. Another notable effort from Canada is the STAN system developed by the Centre de Recherche sur le Transports at the Universite of Montreal [5.3.7.9] presented organized methods of modeling freight traffic at a statewide level. another private firm. as can be seen from Table A. The STAN model was developed from the EMME/2 passenger modeling set of programs and includes a sophisticated “multimode multiproduct” model [5. Their related work with a logit modesplit model has already been noted in Section A. for instance) are aimed at directly providing countylevel data. Application of STAN was made to a large region of Brazil. Aside from isolated examinations of specific issues [5.13] and the Wisconsin [5. 5. which assigns traffic to the various modes and links according to the solution to an optimization problem. most proposals for statewide freight modeling have closely followed the pattern of these early 1980s papers. but many others are available [5. In papers describing their model for the Province of Alberta.8. but it is also the most explicitly multimodal model provided. The modeling techniques described are of two types: (1) calculation of growth factors (by commodity) to be applied to existing traffic patterns [5.3. Statewide Models for Freight Forecasting Likely influenced by deregulation in the railroad industry. Related Models Some work done in the late 1980s with models for Alberta and Brazil by Canadian researchers is also applicable to statewide freight modeling in the United States. there seems to have been a flurry of interest in statewide freight forecasting in the early 1980s.19. truck models and share an identical network structure with their respective statewide passenger models (see Table A. Ashtakala and Murthy [5. Both the Indiana and Wisconsin models are reviewed in greater detail in Section A.7. only Louisiana’s [5. The data must therefore be disaggregated to at least a county level before use. Some databases under development by the FHWA (with Reebie and Colography.15] and New Jersey [5.17. Of these models.21]. and the comparatively coarse structure of their models – an order of magnitude fewer TAZs than for a typical passenger model – is evidence of the reduced ability to disaggregate freight data to a lessthancounty level. The report of a recent FHWA project [5.6] and papers in Transportation Research Record 889 [5. Documentation for a few statewide freight models was obtained as part of the research for this appendix. 5. Reebie’s TRANSEARCH database has been a popular source.12] offers recommendations on how to apply similar growth factor methods with increasing degrees of sophistication.16] models are. The Indiana [5. Its useful features include a BoxCox procedure to calibrate the friction factors for the gravity model and a “Commodity Haul Frequency Diagram” to visually assist in calibration of the gravity model.employment data. 5.10]. At that time NCHRP Report 260 [5. but for employment and commodity flows by industry. which is generally too big for statewide analysis. and their features are summarized in Table A. The Michigan [5. and ample evidence is provided of the graphical output made available by STAN.
Truck 2. • Based on commodityspecific growth factors applied to existing traffic volumes.2. • Distribution by fullyconstrained gravity model. Rail 39 Commodity Groups Discussion Statewide models for freight transportation appear to be structured with a strong sense of the limitations inherent in forecasting at a large geographical scale. A. As discussed in Section A. This documentation provides some indication of the difficulties involved in producing a workable statewide model and the significant number of assumptions and adjustments that often must be made in the statewide modeling process.5. BEA commodity flows and USCanada trade flows (FHWA) used.7. Further discussion of the characteristics of statewide freight forecasting models is presented in Section A. and both have produced documentation describing their most recent modeling efforts. This is reflected in their documentation which is generally more straightforward than that for passenger modeling. Internal TAZs at a county level. with smaller TAZs in large urban areas and larger TAZs in rural areas and the Upper Peninsula. A graphical depiction of the Michigan network can be seen in Figures A. many purposes). A.3: Current Statewide Freight Models STATE Indiana TAZs 145 total MODES 1. Two Recent Passenger Models This section provides a closer examination of the recent statewide passenger modeling efforts for the states of Michigan and Kentucky. as can be seen in Table A. Air Truck only 11 Commodity Groups 11 Commodity Groups • New Jersey 2762 internal 51 external 106 internal 34 external Truck only  • • • Wisconsin 1. both Michigan and Kentucky have comparatively long histories of statewide modeling.6. The Guidebook on Statewide Travel Forecasting 100 . • Trip generation regressed by commodity using (1) employment and (2) tons shipped.7. Based on fourdigit STCC code commodity flow analysis. Consideration of railtotruck intermodal diversion scenarios.4. forecasts travel for five trip purposes between 2392 TAZs. Rail PURPOSES 21 Commodity Groups COMMENTS • Based on 1993 Commodity Flow Survey. but on a much larger scale. Most of the TAZs appear to be at the townshiplevel. in the sense that it retains the characteristics of an urban model (many TAZs.1] is a traditional statewide model. Rail 3. Michigan truck survey.6 and A. The Michigan model is a “fourstep” model that. Truck 2. Truck 2. Louisiana  Michigan 2392 total 1. The Michigan Passenger Model The Michigan model [6. Water 4.Table A. This may be because freight forecasting models have developed independently enough from the fourstep urban modeling process that they can begin to address problems (such as data availability) that are unique to their statewide nature.
The trips attracted to the special generator sites are only for the homebased social/recreational and homebased other purposes. 2. Unlike previous statewide models for Michigan. The Institute of Transportation Engineers’ Trip Generation was used (along with local surveys) to develop the attraction equations. further classified by five geographical categories (four for different city sizes and one for rural areas). Three types of “geographic adjustment” factors (Kfactors) are also used in calibrating the gravity model. REMIgenerated countylevel growth factors are applied to all TAZs in a particular county. Trips from any state not represented in the NPTS data are estimated as a function of the state’s area.5 shows the information used in developing these attraction rates. Since there are 67 PUMAs in Michigan. 4 and 5+ persons per household) and three income groups (low. For trips with ends outside of Michigan. These production rates are distributed to the five trip purposes according to proportions based on NPTS data and are applied to the 15 crossclassification categories noted above. Friction factors for the gravity model are calculated using a gamma function of the generalized cost of travel (see Figure A. since a preliminary analysis revealed that attractions for the other purposes were inconsequential. Table A. this most recent model includes intrazonal trips. The gamma function was chosen to provide maximum flexibility in accounting for both very short and very long trips that are possible in a statewide model. state parks. shopping centers and colleges). the model includes thousands of special generator sites.67.vast majority (2307) of the TAZs are within Michigan while the remaining 85 represent the other 47 contiguous United States. This information is then disaggregated to a TAZlevel using data from the Michigan Employment Security Commission. The impedances for intrazonal trips are calculated separately from the network. hospitals. Census data from the Public Use Microdata Sample (PUMS) and the CTPP are then used to develop cross classification tables based on five different household sizes (1. Figure A. Canada and Mexico.8 shows some example calculations for the production rates that are used. Trip distribution is accomplished using a gravity model. It should be noted that the PUMS data is available at the Public Use Microdata Area (PUMA) level. For forecasts of future travel activities. In addition.10 to as high as 9. however. motels. The trip production step for internal trips involves the use of equations for trip production from the asyetunpublished update to NCHRP Report 187. Inc. (REMI) model.4 shows the final attraction rates. Equations for internal trip attractions are based on an evaluation of alternative rates generated from NPTS data. The State of the Art in Statewide Travel Demand Forecasting 101 . The numerous special generator sites are evaluated as having attraction value only. Total productions and attractions are then the result of entering TAZlevel household socioeconomic data into these equations. tourist attractions. The gravity model was calibrated using NPTS data and validated using CTPP and traffic count data. metropolitan area studies in Michigan and available data from the San Francisco area. campgrounds. golf courses. These Kfactors are based on information regarding existing (1) countytocounty. Table A. (2) major citytocity and (2) outstate/instate traffic flows and the values of the Kfactors listed in the documentation range from as low as 0. to avoid complications due to the arbitrary placement of TAZ centroids with respect to highway links. marinas. 3. medium and high) for a total of 15 categories. divided into 10 general categories by type of facility (airports. Michigan’s modeling procedure starts by using socioeconomic data developed at a county level from a Regional Economic Model. production and attraction equations were developed using NPTS data for trips greater than 75 miles in length.9). a PUMA is essentially equivalent to a county in its level of aggregation.
Figure A. Instate Highway Network from Michigan Model Guidebook on Statewide Travel Forecasting 102 .5.6. Instate TAZs from Michigan Model Figure A.
Figure A.650 1.530 0.4a. Attraction Rates from Michigan Model.350 Table A.530 House holds 0.318 The State of the Art in Statewide Travel Demand Forecasting 103 .4b.486 Retail 0.522 0.097 1.522 0. Outstate Highway Network for Michigan Model Table A. Attraction Rates from Michigan Model.300 6.797 4. sale 0. ale 0.802 0.7.123 Employment Category WholeService Mfg.650 1.207 Other 0.748 Employment Category Wholes Service Mfg.530 0. Rural Areas Purpose HBW HB Rec HB Other NHB Work NHB Other Total 1.232 0.728 3.802 0.087 2.486 Retail 1.260 2. Urban Areas Purpose HBW HB Rec HB Other NHB Work NHB Other Total 1.212 0.097 Other 0.360 0.583 Households 0.07 0.088 1.522 10.
310 to nonhome based work and 0. singleoccupant household will produce 3. For instance. apportion the trips produced to the Michigan model’s purposes according fixed ratios. a lowincome. Again. a countylevel transit share is developed for work trips based on average CTPP shares and extended to other purposes by using work/nonwork ratios from the NPTS. adjustment factors are developed to account for trip length. from the NPTS. since a more sophisticated intermodal model has proved difficult to develop. 0.160 to homebased recreational. Representative minimum and maximum values are shown in Table A. obtain generation rates for the various crossclassification categories. for large urban areas the following ratios are used for lowincome. income groups. Guidebook on Statewide Travel Forecasting 104 . based again on NPTS data. First. singleoccupant households: 0. Both experiments proved unsatisfactory and are not included in normal operation of the model. the trips are then assigned to the network using an allornothing procedure. 0.192) = 0. The vehicle occupancy step consists of three substeps. person trip and vehicle trip data from the NPTS are used to develop average occupancy rates by trip purpose for each of the 15 crossclassification categories noted earlier.214 to nonhome based other.7 x 0.59 homebased recreational trips.6. in a large urban area (population greater than 1 million). using the NCHRP 187 update. Similarly. The resulting shares are applied to every TAZ in the county.7 x 0. Sample Production Rates Calculations from Michigan Model The modesplit step is really a vehicle occupancy step. household sizes and trip purposes. Second. Finally. single occupant household.8. A graph of the resulting factors is shown in Figure A. repeat the process all combinations of urban (and rural) areas. multiply these factors. (3. etc.First.404 to homebased other. It should be noted that the model documentation describes experiments made during model development with proprietary software for synthesizing trip tables based on singlestation originanddestination surveys and for employing a “stochastic user equilibrium” algorithm for trip assignment.00 (100% of trips accounted for) Next.160) = 0. Second. (3.7 trips per day.192 to homebased work (HBW).71 HBW trips from each low income. After applying these occupancy factors. 0. 1. Finally.10. For instance. Figure A.
154 for > 570 ksf and A = 0.76 NHB Other 1.22 1. B = 5.12 1.55 x students (for community colleges) Tourist Attractions Campgrounds State Parks Golf Courses Marinas Motels Hospitals Shopping Centers MDOT Travel & Tourism 1991 ITE Manual 1991 ITE Manual 1991 ITE Manual 1991 ITE Manual 1991 ITE Manual 1991 ITE Manual 1991 ITE Manual Colleges & Universities 1987 and 1991 ITE Manuals Gamma function formulation for friction factors: b f = a T x exp(c T) where: f = friction factor T = travel time (or generalized cost of travel) a.368 x ln(reg.628 ] exp[A x ln(ksf) + B] where A = 0.Table A. Gamma Function for Friction Factors from Michigan Model Table A.713 x ln(0.625.57 1.20 1.634 x ln(beds) + 4.756. b and c = constants of calibration Figure A. B = 5.5. Special Generator Attractions from Michigan Model Special Generator Type Airports Source of Rate 1991 ITE Manual Rate Equation exp[1. Maximum and Minimum Occupancy Rates from Michigan Model Trip Purposes HB Other NHB Work 1.13 1.795 exp[0.985 otherwise 2.50 x acres 37.07 1.59 x holes (1.79 x campsites 0. aircraft) .37 x students (for universities) or 1.6.945] exp[0.44 x rooms) + 3.13 HB Rec 1.347] or [104.0.73 x operations /365] 2 x attendance 0.19 Occupancy Rates Minimum Maximum HBW 1.891 x berths) + 410.9.55 The State of the Art in Statewide Travel Demand Forecasting 105 .
Trip generation rates for homebased work (HBW) trips are developed in a spreadsheet that calculates county level trip generation rates by multiplying the national NTPS production rates (for large urban areas and for other areas) by a ratio of countyspecific to national production and attraction rates based on JTW data. more than half of West Virginia. Except for special generator and external station TAZs. almost half of which are outside of the state.2] forecasts travel for three trip purposes (plus trucks) among nearly 1500 TAZs. The Kentucky model’s documentation indicates that the JTW rates vary between 0. Guidebook on Statewide Travel Forecasting 106 . Trip Length Occupancy Factors from Michigan Model The Kentucky Passenger Model The most recent Kentucky model [6. A graphical depiction of the Kentucky network is shown in Figure A.Figure A.744 trips attracted per employee. depending upon the county examined. The TAZs inside the state consist of individual census tracts in rural areas and groups of census tracts in urban areas. as well as significant proportions of the other surrounding states.246 and 0. Woods & Poole.7.532 and 0. Total HBW attractions are then adjusted to match productions. In addition to the censusbased TAZs. The model’s network actually extends to include almost all of Tennessee. the trip generation step uses projections of population and employment provided by a commercial source. special generator TAZs are included at 40 significant recreational areas and military bases and 29 external station TAZs are included to model travel into the network from the remainder of the United States (mostly via the interstate highway system). The large number of TAZs outside of the state borders is a distinctive feature of the Kentucky model. The national rates used from the 1990 NPTS are shown in Table A.11. Ohio and Indiana.10.435 trips produced per person and between 0.
and (3) through trips.6041 2.4282 0. (2) “long” trips – those longer than 60 minutes in length. it is assumed that 99% of nonwork trips are short trips and that 70% of all nonwork productions are HBO. Instead.0364 0. For military special generators only work trips are considered.5390 0. which in the case of the Kentucky model include homebased other (HBO) and nonhome based (NHB) purposes. assuming that retail employment in all areas is 20% of total employment. NPTS Trip Rates Used in Kentucky Model NPTS Trip Purpose Rate for Large Urban Areas (trips per person) 0.5714 HB Work HB Business HB Shop HB SocialRec HB Other Non Home Based Total Less guidance is available from census sources regarding nonwork trips.5969 0.Figure A. with the assumption that 67% of the nonwork trips generated are “short” trips and 33% are “long” trips for “local” areas and vice versa for “national” recreation areas.7. Network Structure from Kentucky Model Table A.3346 0. for nonwork trips.5236 0. Attraction rates are taken from NCHRP Report 187.3261 0.11. Again.4420 0. Recreational special generator TAZs are classified as either “local” or “national”.5445 Rate for Other Areas (trips per person) 0. and it is assumed that 20% of the work trips generated are The State of the Art in Statewide Travel Demand Forecasting 107 . nonwork trips are divided into three categories: (1) “short” trips – those less than 60 minutes in length.6313 0. attractions are adjusted to match productions.0414 0.6122 2. Starting from nationallevel NPTS generation rates.
A Fratar model is used to predict nonwork trips passing through the network. A summary of the various friction factors in their final form is presented in Figures A. etc. transit shares. while the Michigan TAZs average 4100 residents each. The Kentucky model’s documentation instead recommends that if accessibility changes are expected in the future. The friction factors and Kfactors are subsequently adjusted to provide a good fit with known screenline counts and average trip length frequency data from the JTW survey. require that HBW trips over 120 minutes in length be eliminated from the process. but described only as “similar to the Fratar process”. the gravity model cannot be used to forecast future volumes when the “trip table calibration” procedure is used. per TAZ.3).2 and A. It should also be noted that many calculations (involving generation rates. Selection of the “grainsize” for the model should be made considering its intended use. perhaps fewer. If forecasts of general trends are being sought. if projectspecific planning capabilities are desired using the statewide model. The 756 instate TAZs for Kentucky are. The likelihood that meaningful adjustments could be made to the model based on detailed knowledge of activities in any particular TAZ is therefore diminished. since a few very long trips can have a disproportionately large effect on the average trip length calculated for calibration. the Kentucky TAZs average slightly more than 5000 residents each. growth factors. There is no modesplit or autooccupancy step. friction factors are developed from the previously noted unpublished update to NCHRP Report 187. but intrazonal travel times are deliberately kept small (a maximum of 15 minutes) in order to keep intrazonal trips off the statewide network. Based on 1990 population figures.9. mi. Discussion As noted above. one similarity between the Michigan and Kentucky passenger models is their fine geographical level of detail in comparison with statewide freight models and with several other statewide passenger models (see Tables A. Kfactors. respectively). larger TAZs will suffice. but friction factors for “long” trips are synthesized to give a “reasonable” trip frequency distribution. mi. the following threepart procedure be followed: (1) a “fratared” forecast should be prepared using the “trip table calibration” results. on average. twice as big as the 2307 instate TAZs for Michigan (52 sq. a more “finegrain” model may be useful. but the documentation does not indicate a range of values used. limitations in Kentucky’s software used to calibrate the friction factors.intrazonal trips and are not assigned to the network. Of course. since accessibility effects are not considered in this data synthesis procedure. However. An appendix to the documentation further describes it as an iterative adjustment procedure where the modeled volumes on network links are matched to corresponding traffic counts on the roads they represent. The trip distribution step is accomplished using a gravity model. the end result is a better fit between the model and the base year traffic.) are already made using only countylevel data in both of the models examined. Similarly. per TAZ versus 25 sq. Unfortunately. Kentucky judged this to have a beneficial effect on model performance. (2) a gravity model forecast should be produced that includes accessibility changes and (3) a final forecast should be manually generated to resolve any inconsistencies. and traffic assignment is evidently done using an allornothing technique. since no other details are provided in the documentation. Guidebook on Statewide Travel Forecasting 108 . For nonwork trips. A further procedure described as “trip table calibration” is mentioned in the documentation. Kfactors are also used at a county level to adjust the distribution of HBW trips. In either case it would appear that using a large number of TAZs precludes the use of any but the most basic data from any individual TAZ.8 and A. but are relatively close in terms of average population per TAZ. Friction factors for HBW trips are developed from JTW survey information.
572 61. While limiting border crossings to a few very select locations may be acceptable for a peninsular state like Michigan.572 61.268 17. poses a similar problem to that posed by using a large number of instate TAZs.8. The Kentucky modeling strategy.333 26. In contrast.705 9468 8386 7439 6609 5880 5239 Travel Time (min. since the Kentucky model does not include a national network (as Michigan’s does). The Michigan model’s roadway network is built with connections to outstate areas at relatively few discrete points (see Figure A.840 61.572 61.Table A.242 125.143 20.) 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 Friction Factor 24 22 20 19 17 16 15 13 12 11 11 10 9 8 8 7 6 6 5 5 4 4 4 3 3 3 3 2 2 2 The most immediately visible difference between the Michigan and Kentucky models is their treatment of the geographical areas outside of their respective state boundaries. the Kentucky model’s highway network extends more than 200 miles into the surrounding states (see Figure A. it is a bit more troublesome for a landlocked state like Kentucky.360 203.) 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 Friction Factor 257 236 218 200 184 170 157 144 133 123 114 105 97 90 83 76 71 65 60 56 52 48 44 41 38 35 32 30 28 26 Travel Time (min. The State of the Art in Statewide Travel Demand Forecasting 109 . it must include external stations to generate travel between the network and external areas.609 77.752 12.) 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 Friction Factor 4675 4177 3737 3348 3003 2697 2426 2184 1969 1777 1605 1452 1315 1192 1082 982 893 813 740 675 616 563 514 470 431 394 362 332 304 280 Travel Time (min. The cushioning effect of the 200mile wide outstate network makes the precise trip generation values for the external stations less important.624 13.244 46.572 53.472 23.007 159.572 61.040 34. It should also be noted that.123 10.178 98.818 30.6).780 15. to include TAZs and network links well outside its state boundaries.129 40. HBW Friction Factors from Kentucky Model Travel Time (min.11).) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 Friction Factor 259. That problem is the need to know detailed information about hundreds of TAZs and road segments outside of Kentucky.
According to the Kentucky documentation.9.916 9 39 26.999 2 32 367. A similarly troubling feature of the Michigan model is its extensive use of Kfactors to adjust the gravity model results.144 14 44 9192 15 45 7632 16 46 6374 17 47 5351 18 48 4512 19 49 3821 20 50 3247 21 51 2768 22 52 2367 23 53 2029 24 54 1744 25 55 1503 26 56 1297 27 57 1122 28 58 972 29 59 844 30 60 734 Friction Factor 639 557 486 425 372 326 285 250 220 193 170 150 132 116 103 91 80 71 63 56 49 44 39 34 30 27 24 21 19 17 “Long” NonWork Trips Travel Friction Time Factor (min. Nonwork Friction Factors from Kentucky Model “Short” NonWork Trips Travel Friction Travel Time Factor Time (min.497 6 36 59. As noted above. The very large and very small values of the Kfactors developed (see earlier discussion) seem to indicate that some simplification of this process would not compromise the modeled results.788 12 42 13.616 13 43 11.878 3 33 196. The Kentucky model also makes use of Kfactors. Perhaps the most troubling aspect of the Kentucky model is its extensive use of national average values derived from these sources in combination with assumed ratios of travel characteristics (99% of nonwork trips are “short”.g. specifically the NPTS.435 8 38 33. etc.Table A.) 1 1 60 1 61 200 181 50 241 6 301 3 361 1 421 1 481 1 541 1 601 1 661 1 720 1 Travel data used to construct both the Michigan and Kentucky models comes primarily from three US government sources.) that are also based on national averages.802 7 37 44.) 1 31 999.425 10 40 20. reluctance to cross a state line for work or social differences between regions). but in a less aggressive fashion. It is also important to note Guidebook on Statewide Travel Forecasting 110 .955 4 34 122.925 11 41 16. 70% of nonwork trips are HBO. the Michigan model makes use of three overlapping sets of Kfactors – one set at a county level and two other sets for specific destination TAZs – to modify the distribution of travel predicted by the model. this dependence on national average figures is due to the dearth of readily available data for Kentucky. the CTPP and the JTW survey. The Kfactors appear to be produced in a mechanical fashion.954 5 35 83.) (min. without consideration of any behavioral basis for factoring (e..
The TRANSEARCH data. Finally. The principal data on which the Wisconsin model is taken from Reebie’s TRANSEARCH. this section provides a closer examination of two statewide models for freight traffic forecasting. Michigan’s lengthy list of special generator sites continues for 51 pages of small print type.that Kentucky attempted to develop rationallybased Kfactors. using only 40 of the most significant special generator sites. Both models are constructed to operate in a sequential fashion. the total flows are determined from the baseyear TRANSEARCH data for each important commodity group. Two Recent Freight Models As a complement to the passenger models discussed in the previous section. It is unclear what advantage this provides for statewide modeling. As with the statewide passenger models. Table A. Each of Wisconsin’s 72 counties is represented by a separate TAZ. Figure A. It would appear that merely maintaining the database for these thousands of locations would be a significant task in itself. it is determined which proportion of each commodity group’s flow is destined for industrial consumption and which is destined for household consumption. these freight models show the number of simplifying assumptions must be made to facilitate model development. First. The other 34 TAZs are composed of multiple BEA regions that represent other states. Of the 140 TAZs. The models examined – Wisconsin and Indiana – represent typical examples of recent thinking in freight forecasting. is supplemented by some commodity flow data specific to Wisconsin. A.13 show the TAZ structure for the Wisconsin model. Second. similar to passenger models. truck and rail. Factors are also The State of the Art in Statewide Travel Demand Forecasting 111 .1] estimates the freight traffic carrying the products of 39 important commodity groups between 140 TAZs by four modes: air. shopping centers and golf courses. Local TAZs from No generation equations for freight flow are developed. The Wisconsin Freight Model The Wisconsin freight model [7. Wisconsin Model instead the TRANSEARCH flows are simply distributed at the county level in a four step process as follows. One final contrasting element of the Michigan and Kentucky models is their treatment of special generator sites. but their TAZ structures are much more coarse. is more limited and presumably more in keeping with modeling at a statewide level. Kentucky’s approach.10 lists the 39 commodity groups that are considered important for Wisconsin and are used in the modeling process. countylevel destinations are allocated based on employment (for industrial consumption) and population (for household consumption). including individual hotels. based on a national inputoutput table.12.7. water. Third. 106 are counties. but those efforts proved problematic and were abandoned. freight origins are identified and are assigned to the countylevel TAZs based on county employment data.12 and A. which is provided at a BEA region level of aggregation. Figures A. while the remaining 34 countylevel TAZs represent counties in adjacent states.
The Indiana Freight Model The TAZ structure for the Indiana freight model [7. Air and water traffic are not assigned. Since the forecast flows would project the current modal shares into the future. Adjustments are made to the “trendline” forecast by enlisting the services of various expert panels.) Both the truck and rail networks were developed from US DOT sources.13. they are converted into an equivalent number of vehicles. Forecasts of future freight flows are made using econometric models that include employment forecasts obtained under a contract with WEFA (another private firm) and productivity forecasts made using information from REMI. Initially weightpervehicle ratios of 100 tons per railcar and 24 tons per truck were used for all commodities. The Indiana model predicts both truck and rail traffic volumes for a network that includes a TAZ for each of Indiana’s 92 counties and 53 more TAZs that represent the remaining 47 contiguous states and the District of Columbia. Figure A. The truck assignment is done allornothing. who add a “market driven” element to the forecast values.14 and A.15 graphically depict the Indiana freight network. Guidebook on Statewide Travel Forecasting 112 . Outstate TAZs from Wisconsin Model Once the freight flows (by weight) are determined. as for the Kentucky passenger model. Tonspertruck values were subsequently modified by commodity group to the values shown in Table A. All other states and the District have one TAZ each. For rail each shipment is designated as having a “most likely carrier”. Kentucky and Michigan. The resulting daily flows are then assigned to the appropriate modal networks.4. Figures A. based on a sixday work week. and the shortest path using that carrier alone is assigned the shipment. (There are three TAZs for Ohio and two each for Illinois. These factors represent the additional movements for freight distribution and drayage from intermodal yards.2] is very similar to that used in the Wisconsin model. It should be noted that. This constitutes what is called the “trendline” forecast. This approach is less formalized than that used in the “trendline” model. Daily flows are determined by dividing the total number of vehicles by a value of 312 working days per year. but is important to making policy decisions based on the model results. Wisconsin also developed an approach which uses another panel of experts to identify alternative railtruck modal splits based on shipment distance and frequency of rail service. the detailed roadway network for the Indiana freight model extents to about 200 miles beyond the state’s border.calculated for “secondary” trucking volumes as a function of the primary freight flows.
Table A.10. Important Commodity Groups and Traffic Densities
Wisconsin Model Indiana Model Tons per Tons per Tons per STCC Description Truck Truck Railcar 1 Farm Products 24 38 96 8 Forest Products 13 9 Fish or Marine Products 6 10 Metallic Ores 24 11 Coal 24 40 100 13 Crude Petroleum, Nat. Gas, Gasoline 14 14 Nonmetallic Ores 19 39 97 19 Ordinance or Accessories 24 20 Food and Kindred Products 18 32 80 21 Tobacco Products 5 See note 1 See note 1 22 Textile Mill Products 5 7 18 23 Apparel or Finished Textile Products 3 4 10 24 Lumber or Wood Products 15 29 72 25 Furniture or Fixtures 3 6 15 26 Pulp, Paper or Allied Products 16 25 62 27 Printed Matter 9 See note 1 See note 1 28 Chemicals 22 35 88 29 Petroleum or Coal Products 19 26 66 30 Rubber or Misc. Plastics Products 4 See note 1 See note 1 31 Leather or Leather Products 3 See note 1 See note 1 32 Clay, Concrete, Glass or Stone Products 23 32 81 33 Primary Metal Products 19 34 86 34 Fabricated Metal Products 24 8 20 35 Machinery  Other than Electrical 9 11 28 36 Electrical Machinery, Equip., Supplies 8 7 17 37 Transportation Equipment 12 9 23 38 Instruments  Photo. or Optical Goods 5 See note 1 See note 1 39 Misc. Manufacturing Products 2 See note 1 See note 1 40 Waste or Scrap Metals 16 31 78 41 Misc. Freight Shipments 23 42 Shipping Devices Returned Empty 4 43 Mail and Express Traffic 3 See note 2 See note 2 44 Freight Forwarder Traffic 4 45 Shipper Association Traffic 3 46 Misc. Mixed Shipments 7 47 Small Packaged Freight Shipments 4 48 Hazardous Waste 16 49 Hazardous Materials 18 99 Unknown 12 Other 35 87 Notes: 1. For Indiana model, “Other” includes STCC groups 21, 27, 30, 31, 38, 39. 2. For Indiana model, US mail and express mail groups are analyzed separately.
The actual workings of the model are very similar to a UTP model. For each of 21 commodity groups that are considered important to Indiana (see Table A.10), trip generation equations were developed based on a regression of data available from the 1993 CFS. Forecasts for
The State of the Art in Statewide Travel Demand Forecasting
113
Indiana county productions and attractions are then based on countylevel employment and population projections commercially available from Woods & Poole. Table A.11 shows the trip generation equations that were developed. For areas outside of Indiana, forecasts are based on national growth factors. Following trip generation, freight shipments are distributed by a gravity model that is also calibrated using the CFS data. Special care is taken to match the average shipping distance per ton for each commodity group. This prevents an inappropriate weighting for many shortdistance lightweight deliveries versus a few longdistance heavyweight shipments that might be included in the same commodity group. The mode split step also utilizes the 1993 CFS, projecting the 1993 national shares into the future. Before assigning traffic to the network, the Indiana model (like the Wisconsin model) divides the freight tonnages into an equivalent number of vehicles, with tonspervehicle rates determined separately for each commodity group. The rates are based on values (by commodity group) from the ICC Rail Waybill sample and the assumption that each truckload carries 40% of the load carried by a railcar. The tonspertruck and tonsperrailcar values used are shown in Table A.10. A daily traffic conversion is also made for the Indiana model, assuming 5 working weekdays and (from the Highway Capacity Manual) 0.44 working days for each weekend day. This results in a 5.88 day work week or a 306 day shipping year.
Figure A.14. Instate TAZs from Indiana Model
Finally, the traffic is assigned to the network using an allornothing process. Since a straight allornothing assignment typically loads too many trips onto the interstate highways, a procedure to adjust the link speeds for noninterstate Figure A.15. OutofState TAZs for Indiana Model highway segments is provided. The adjustment involves calculating new speeds for noninterstate links using the equation shown in Figure A.16. This serves to draw more trips from the interstate roads to the competing US and state highways that run parallel to them.
Guidebook on Statewide Travel Forecasting
114
Sadj = Sprev + [2 x (65mph  Sprev) where:
0.5
]
Sadj = the adjusted link speed Sprev = the previously used link speed (usually the speed limit on the link) Figure A.16. Link Speed Adjustments from Indiana Model
Table A.11. Trip Generation Equations from Indiana Model Commodity Group (by STCC)1 1 11 14 20 22 23 24 25 26 28 29 32 33 Trip Generation Equations2 (tons produced or attracted) P = 1445  0.523(ag services) + 0.0048(ag cash) A = 0.819 P P = 7.6 (coal) A = 3.1(coal) + 5.3(mining) P = 0.078(manufacturing) A = 0.997 P P = 0.282(food) A = 0.832(population) + 0.162(food) P = 0.016(textiles) A = 0.003(apparel) + 0.0001(all) P = 0.004(apparel) A = 0.002(apparel) + 0.011(population) P = 0.668(lumber) A = 0.728 P P = 0.017(furniture) A = 0.033(population) + 0.002(furniture) P = 0.103(pulp) + 0.056(lumber) A = 0.085(pulp) + 0.259(population) P = 0.150(chemicals) + 1.164(petroleum) A = 0.077(chemicals) + 0.455(petroleum) + 0.683(population) P = 6.857(petroleum) A = 4.007(petroleum) + 1.881(population) P = 2.882(population) A = 2.914(population) P = 0.085(metals) A = 0.093(metals) + 0.061(fabrication)
The State of the Art in Statewide Travel Demand Forecasting
115
Table A.11. Trip Generation Equations from Indiana Model (continued) Commodity Group (by STCC)1 34 35 36 37 40 Other Trip Generation Equations2 (tons produced or attracted)
P = 0.013(metals) + 0.034(fabrication) A = 0.035(fabrication) P = 0.013(machinery) A = 0.010(machinery) P = 0.004(metals) + 0.004(fabrication) + 0.003(electrical) A = 0.005(fabrication) + 0.034(population) P = 0.040(transportation) A = 0.027(transportation) P = 0.00048(population) A = 0.0067(manufacturing) P = 1.097 A A = 0.254 (population) Notes: 1. See Table A.10 for description of groups. 2. population = total population all = total employment ag services = employment is SIC 7 ag cash = gross receipts (in $1000) from farming coal = employment in SIC 11 mining = employment in SIC 14 manufacturing = employment in SIC 2 and SIC 3 food = employment in SIC 20 textiles = employment in SIC 22 apparel = employment in SIC 23 lumber = employment in SIC 24 furniture = employment in SIC 25 pulp = employment in SIC 26 chemicals = employment in SIC 28 petroleum = employment in SIC 29 metal = employment in SIC 33 fabrication = employment in SIC 34 machinery = employment in SIC 35 electrical = employment in SIC 36 transportation = employment in SIC 37
Discussion
As noted above, both the Wisconsin freight model and the Indiana freight model use a countylevel TAZ structure. This appears to be the accepted level of geographical detail for freight modeling for two reasons. First, essential commodity flow information has been widely available only at the BEA region level of detail, which is too large a geographical area for meaningful use in a statewide model. Use of this information at a county level involves only one step of disaggregation. In contrast, using smaller scale TAZs (e.g., at the census tract or township level) typically involves additional disaggregation steps and additional layers of assumptions about the characteristics of the commodity flows. Second, a large amount of economic information is available (e.g., County Business Patterns data) or will soon be available (e.g., the
Guidebook on Statewide Travel Forecasting
116
which states: There is a temptation to evaluate the … forecasts. a good model should be easy to explain to an informed audience and easy to justify to an interested public (who will likely be funding the modeling efforts). Wisconsin’s model makes use of a commercially available commodity flow database. The State of the Art in Statewide Travel Demand Forecasting 117 . In the end.. Recommendations and Conclusions Some general advice on constructing statewide travel forecasting models (or models for any geographical area) is contained in the documentation for the Indiana freight model [8. the Indiana model offers greater flexibility for considering the effects of changes in the network (e. Many of the tonspertruck values are higher for the Indiana model. It should be obvious that this is not possible until the forecast dates have been reached … One’s acceptance of the forecasts should be based on the quality of the methods used in the analysis of the [current] flows and the accuracy of the methods in replicating existing conditions. The Wisconsin model does not have a conventional trip generation step.g. the Indiana model depends heavily upon CFS data. Any recommendations for “best practice” in travel forecasting should be based on these principles. Thus. future freight flows are forecast by a growth factor procedure.5) at a county level for use as inputs to a statewide freight model. in more a typical fourstep fashion. The most obvious difference between the two models is the techniques each uses to generate and distribute the predicted freight flows. In doing so. both models use very similar processes to divide the predicted freight flows into an equivalent number of vehicles and assign them to the network. As it is. (p. The CFS is then supplemented with information that includes important commodity flows that are not covered by the CFS. a new highway corridor with reduced travel times). but both models use comparable daysperyear factors. Attention to congestion considerations (and related issues such as the development of factors for seasonal or even hourly flows) should not be difficult given available forecasting software.Reebie and Colography databases mentioned in Section A. it loses some sense of where the data is coming from and what assumptions have been made in supplementing it. Instead. The Indiana model. but that is likely a function of the Wisconsin DOT’s primary goal of estimating future highway demand. The Wisconsin model gives less attention to the tonspervehicle considerations for railcars. To this might be added two more common sense suggestions: use methods of modeling that are appropriate to the results desired and keep the process simple. These factors are then applied to the base year flows. A. On the other hand. The following eight recommendations are directed toward developing just such models for use in statewide travel forecasting. An unfortunate part of the assignment step for both models is the failure to address the possibility of congestion due to the presence of a large number of passenger vehicles sharing the road. However. where the growth factors (one for employment growth and one for productivity growth) are calculated by econometric models developed by specialized consultants. the Indiana report seems to say. 141) Keep the process sound and rational. which still must be disaggregated once to get to the county level. uses regression equations developed from the CFS to generate its freight flows and a gravity model to distribute them.1]. The method used in the Wisconsin model seems appropriate for use in a state where little change is expected in the transportation network over the forecast period. such as certain agricultural products or solid waste.8.
WEFA. then “traditional” techniques may require some overhaul. rather than through behavioral principles. adding a short slowspeed (highdisutility) link at a border crossing might work just as well. due to the increased necessity for adjustments (Kfactors) and the reduced ability of a statewide modeler to be aware of localized conditions across the state. At the county level of aggregation it is also more likely that a sufficient number of data samples can be found to calibrate logitstyle models and other more “advanced” modeling structures similar to those discussed in reference to intercity modeling. New database products under development (at the only the county level for freight) will make the process still easier. This data. As demonstrated in the passenger and freight models reviewed in this appendix. With county level modeling. as noted in Section A. The Kfactors used in the models examined appear to have been developed mostly in an effort to adjust the gravity model results to match the base year traffic. Reexamine the structure of any “traditional” UTP methods used.Recommendations for Statewide Travel Demand Forecasting 1. 3. Woods & Poole) commercial sources. if it is desired to represent the reluctance of residents of one state to work in another. alternative methods of modeling. simultaneously. Behavioral methods might be better for increasing or decreasing trips between particular TAZs. Further steps to artificially disaggregate the information involve assumptions that are probably not necessary for modeling at a statewide level. Make use of existing government and commercial databases. there is less temptation to match flows at a small (census tract) level. Guidebook on Statewide Travel Forecasting 118 . CTPP. then a better sense can be gained of the sensitivities of the various techniques and the possible range of forecasts for future travel. the majority of transportation (as opposed to merely demographic) information is generally available only at the county level. the procedure suggested at the close of the Kentucky model’s documentation. These assumptions are too easy to forget when adjustments are made to try to match the base year traffic flows. Most of the data used to develop the models reviewed in Section A. Build the statewide model in a form consistent with available data. Use a statewide model for analysis of statewide effects of system or socioeconomic changes or for statewide corridor planning only. Use of a statewide model for smallscale or projectspecific purposes is dubious. when the model is really based on information from a much larger (countylevel) scale. If “advanced” techniques are not warranted or do not prove fruitful.6 and Section A. Examine. 4. 5. including various disaggregate behavioral models. Examinations could also be made to determine whether it is worthwhile to extend the network into other states to buffer external station effects or to include a national network or to experiment with the number of purposes. REMI. Concurrent use of alternative modeling techniques is. For instance.6. If several different model structures are available. This is a philosophical decision that drives the rest of the modeling process. The statewide passenger models examined for this appendix appear (because of their continued manipulation of information at a small scale) to be aimed at examining flows at a project level. where more specialized and local knowledge can be applied. Separate methods should be used for forecasts at the smaller scale. There are obviously a large number of techniques available for use in developing a statewide forecast. judiciously supplemented and modified by local information. CFS) sources or similarly few commercial (Reebie. This second recommendation logically follows from the first. proved sufficient to develop workable statewide models. A principal example is the use of Kfactors for adjusting the gravity model. 2.7 came from a few government (NPTS.
In addition. Aside from the obvious value in assessing the reasonableness of model results. Examples are data that may allow spatial disaggregation below the county level and data on recreational travel. and it shows how assembling groups of people with direct experience in freight can be used to take some of the burden away from individual modelers (or small groups of modelers) and open up the forecasting process to knowledgeable people outside of the DOT. a continued federal government focus on planning issues that can be addressed by statewide modeling – as begun with ISTEA and the CAAA – should provide an impetus for state DOTs to once again explore the possibilities of forecasting travel demand at a statewide level. The cardpunch technologies that existed during the first wave of statewide modeling in the early 1970s are long gone. Although statewide forecasting model may have been built primarily from existing data sources. such as the BoxJenkins methods noted in Section A. Without some effort at future dates to assess how well previous models have worked. assess the performance of the model(s) used. The State of the Art in Statewide Travel Demand Forecasting 119 . Make use of expert panels in the modeling process. thereby contributing to the accuracy of the model. 8. In all of the material reviewed for this appendix. Plan for future data collections that will enhance an existing model. it could be improved with additional. Expert panels can be very useful in filling gaps in existing socioeconomic data. Another reason is a general lack of confidence in the value of developing a statewide model in the first place. modelers may be doomed to repeating the same mistakes. At future dates. but have not continued to use them. there is one subject that is glaringly absent: comparison of model results with subsequent traffic demands. The Wisconsin freight model is notable for its use of expert panels. Make use of existing statewide traffic monitoring programs. Monitoring data could also be used to assist in making modifications to national rates used for special generators and external stations. At the time a model is developed it is obviously impossible to guess how effective it will be in predicting future traffic.6. It might also be possible to determine how much modeling effort is sufficient to generate forecasts for the particular transportation decisions being made. This is attested to by the vast majority of states that do not perform any travel forecasting at the statewide level and by the states that built models in the past. The use of traffic monitoring data would be greatly assisted by the use of time series techniques of data analysis. The increasing availability of more userfriendly computer programs for travel forecasting and the increasing availability of travel and socioeconomic data to feed these programs should begin to alleviate some of the problems that caused states to abandon their statewide forecasting efforts. One of the primary reasons is the lack of data available in sufficient quantities to build and calibrate “fine grain” models at a statewide level.2. the information gathered by statewide monitoring programs can be useful in developing daily and seasonal factors for various roadway types and geographical areas in the state. 9. Conclusions The full arsenal of available techniques discussed in the literature of intercity and statewide travel forecasting is not being brought to bear in existing statewide models. assisting with model assumptions and disaggregating model results to smaller divisions of zones. A continuing development of databases useful for calibration of models (especially behavioral models) at a statewide scale would also be helpful. 7. locallycollected data. expecially on the weekend. Early efforts in statewide model building were handicapped by rapidly changing political and economic environments.
Washington.4 Guidebook on Statewide Travel Forecasting 120 . D. Traffic Interactance Between Cities. HRB. Sec. The challenge is now to put these techniques to the test of use. airlines. D. Improvements in intercity and interregional corridors are best analyzed with statewide models.3 References for Section A. National Research Council.C... NCHRP Report 70: Social and Economic Factors Affecting Intercity Travel.C. In HRB Bulletin 257. Washington. Outlook for Better Regional and National Forecasts of Highway Traffic and Finance.1 Church. April 1996.C. P. New Hampshire Department of Transportation.2 Pub. 3638. Ivers and Associates. Statewide Travel Demand Model Update and Calibration: Phase II. HRB. 2. In HRB Bulletin 297. A. Inc.C.. D.3 2. 5965. 1. New Hampshire Statewide and Subarea Travel Models Plan. S.1 1. 1914. As this appendix has sought to point out. pp. D.2 2. Cambridge Systematics. 1960. 1961. This includes railroads.1 1. 102240.As an additional impetus to model building. TRB. Future research might include an investigation of the proprietary forecasting techniques employed by private transportation firms that operate over large geographical areas. state DOTs are now being pressured to address traffic congestion by programming transportation improvements. Michigan Department of Transportation. Burch. 105 Stat.2 2. Washington. In Transportation Research Record 527. March 1995. Other future research might be aimed at understanding the underlying causes in the socioeconomic trends that drive the forecasting models. trucking companies and express delivery services. 1974 pp. KJS Associates. pp.. J. Title I. National Research Council. There is an enormous opportunity to combine and overlap the available techniques and to explore the benefits of travel demand modeling at a statewide level. there are a number of techniques available to prospective modelers – from the familiar fourstep models and growth factor techniques to more advanced probabilitybased or optimization methods. Comparison of the Model Structure and Predictive Power of Aggregate and Disaggregate Models of Intercity Mode Choice. HRB. Vogt. E. National Research Council. D. L. 1417. L. National Research Council.9. 1969. 1025 Statewide Planning. References References for Section A. Watson. Washington.
pp. In Highway Research Record 369. S. Rossi. J. National Research Council. 96. Northwestern University. Demand Model for Intercity Multimode Travel.. 1976 pp. 1977. E. 2. The Transportation Center. Cohen. C. 1984. Hirsh. 1992. Towards the Development of Innovative Models of Intercity Travel Demand.. 203218. M. Erlbaum. No. Thesis. Washington. Journal of Regional Science. D. GK. DiRenzo. Hartgen.10 2. Washington. J. Vol. No. R. and W. Hsu. P. Washington.C. TE2. Brand. 1978. Miller. A Comparison of Model Structures for Intercity Travel Mode Choice. D. F. 1970. pp. Intercity Rail Travel Models.7 2. In Transportation Research Record 673. 39. 1992. Vol. 1326. pp.1992. Intercity Transport: Engineering and Planning. 8390. and M.C. Khan. Forinash. K. 1992.. and L. Forecasting HighSpeed Rail Ridership. Transportation Engineering Journal of ASCE. D.15 2. N. 297316.5 Rallis.16 2. May. Baumol.2. pp. T.. F. 1525. Diversion Model for Estimating HighSpeed Rail Use. J. New York. V. Illinois. C.9 2. Structural Travel Demand Models: An Intercity Application. M. Yu. Vyas. P.S. 124135. The Abstract Mode Model: Theory and Measurement.C. Wiley. D. Forecasts of Intercity Passenger Demand and Energy Use Through 2000. 2. 6.8 2. National Research Council. 1218. D. Quandt. Vol. Implications of HighSpeed Rail on Air Traffic. Saskatoon. Washington. G. and M. Tierney. pp. R. TRB. In Transportation Research Record 870. J. J. University of Saskatchewan.. Review of Intercity Passenger Demand Modeling: Mid60’s to the Mid80’s. No. Saskatchewan.. TRB. A. pp. T. Millar. D. National Research Council.6 2.. E. S. 1927. P. pp. Parody.C. In Transportation Research Record 1341. Intercity Passenger Travel Demand Modelling: Present State and Future Possibilities. HRB. April 1985. 1982. Kuah. Peers. 2125. M. Washington. E. and D.. B.14 2. S. TRB. National Research Council. and K. S. Evanston.13 2. TRB. Bevilacqua. T. In Proceedings of the 27th Annual Meeting of the Canadian Transportation Research Forum. In Transportation Research Record 569. Buckeye. Transportation Quarterly. Gur.11 2.. Kaplan.17 The State of the Art in Statewide Travel Demand Forecasting 121 . Washington.C. 2. Koppelman F. D. 1966. pp. In Transportation Research Record 1342. pp. 1971. Northwestern University. and Y.. National Research Council. M.. A. National Research Council.12 2.C. 378389. TRB. D.
1989. D. D.24 2.29 Guidebook on Statewide Travel Forecasting .25 2.27 2. Winston. Safwat. F. Washington. 2. Research in Transportation Economics. D.C. D. S.. Intercity Passenger Decision Making: Conceptual Structure Data. and W.26 2. Vol. 122 2. F. Northwestern University. ElHawary. 1993. Multidimensional Model System for Intercity Travel Choice Behavior. B. N. Vancouver. S. S. P. Gadallah. 17A. pp.C. S.18 Forinash. A Poisson Model of Rural Transit Ridership. Brademeyer. Dean. pp. Washington.. and F.. J. F. and A.20 2. Hirsh. A. Brown. S.C. 1521.C. D. 5259. Morrison. Markow. A Methodology for Intercity Transportation Planning in Egypt. Washington. pp. pp. L.. University of British Columbia. 1985. In Transportation Research Record 1120. Illinois. 3537. 1968. Koppelman. Vol. Egypt Intercity Transportation Model. Koppelman. Washington. The Centre for Transportation Studies. Neumann. Koppelman. National Research Council. M. TRB. 7075. pp. In Transportation Research Record 887. Washington..22 2. N. 1982. TRB. and M. TRB. Koppelman. 1987.C. The Demand for Air Travel: A Regression Study of TimeSeries and Cross Sectional Data in the US Domestic Market. 1978.. In Transportation Research Record 1241. A.19 2. TRB.21 2. National Research Council. TimeSeries Analysis of Intercity Air Travel Volume. L. HRB. 98106. Watkins. 1983.23 2. In Highway Research Record 213. D. D. and M. S. National Research Council. 1986. pp. 213237. Washington. National Research Council. E. pp. F. Hirsh. V. 1986. 481491. In Transportation Research Record 1413. pp. pp. and B.C... Safwat. K. Koppelman.. In Transportation Research Record 1085. Application of Simultaneous Transportation Equilibrium Model to Intercity Passenger Travel in Egypt. F. In Research for Tomorrow’s Transportation Requirements. TRB. An Econometric Analysis of the Demand for Passenger Transportation. D. Byrne. Application and Interpretation of Nested Logit Models of Intercity Mode Choice.C... C.28 2. Washington. National Research Council. and F. 2134. 6. 1987. S. In Transportation Research Record 661.2. J. D. M. In Transportation Research Record 840. TRB. 2127. TRB.C. Intercity Passenger Travel Demand Analysis and Forecasting. No. National Research Council. S. A. 18. National Research Council. 9971009. 1982. Oberhausen. Modeling Dilemma of Intercity Bus Transportation. Moavenzadeh. The Transportation Center. National Research Council. Evanston. D. Transportation ResearchA. and K. Washington. S. and C. pp. pp.
4 3. Vol.5 3. 1977. Kanafani. Vol. Transportation ResearchA. pp. In Special Report 149. Tobin. 1. P. pp. National Research Council. 223230. Washington. 1.. Morlok. 513520. Predictive Interstate Freight Network Models: The State of the Art. Washington.C. The Logistics and Transportation Review.2 3. In Transportation Research Forum Proceedings: 21st Annual Meeting.. T. A. 3. 1978.3 3. 1. XVIII. T. 4765. Mass.1 Smith. XXI. H. QualityofService Model of Intercity AirTravel Demand. Bronzini. R. In Deregulation and the Future of Intercity Travel (J. National Research Council. HRB. pp. Journal of Transportation Engineering. 419427. 475485. L. TRB. A.32 2.6 3. 523531. D. L.3 3. 2. Ghobrial.C.. Vol. pp. In Transportation Research Forum Proceedings: 18th Annual Meeting. MIT Press. pp. Vol. No. TRB. Meyer and C. pp. P. Recent Advances in Network Modelling and Their Implication for Freight System Planning. P.. and E. and G. D. L.7 The State of the Art in Statewide Travel Demand Forecasting 123 . D. National Research Council. and P. S. March/April 1995. Vol. No. L. XXI. 1983. Recreational Travel Behavior: The Case for Disaggregate. pp. 17A. 121. and A. Vol. M. A. A. 2. Vol. Evolution of a Multimodal Freight Transportation System Network Model. B. 1980.30 Pickrell. G. No. 17A. C. Model of Intercity Travel Demand. pp.2.. K. Morton. Friesz. Gilbert. V. 1980. 6. Winston. 1969. 249260. 1974. No. S. Model to Estimate Commuter Airline Demand in Small Cities. The Demand for Freight Transportation: Models and Applications. No. pp. 1983. Jones. Thorson. In Transportation Research Forum Proceedings: 20th Annual Meeting.C. Harker. Washington.31 2. No. Probabilistic Models. pp. 135140. Friesz. No.33 References for Section A. 1987. T. D. Cambridge. Forecasting Freight Transport Demand – The State of the Art. Oster.. Brewer. 187193. L. eds. 409417. MultiMode Intercity Freight Transportation Planning for Underdeveloped Regions. 311326. 10. Sharp. pp. In Transportation Research Record 673. In Highway Research Record 296. and K. 1974. 4. A Statistical Sketch of Intercity Freight Demand. R. Transportation ResearchA. 6.).
M.1 Whiteside. pp. D. ChiraChavala. Washington.. 110135. 1976. 1417. Kobia. 6064.C. D. National Research Council.. National Research Council. National Research Council. S. TRB. US Department of 124 4. Roth. Gottfried. Washington. Morlok. pp.E. B. 1988. Mason.13 3. A Sequential ShipperCarrier Network Model for Predicting Freight Flows. TRB. Washington... and E. G. May 1986. R. and W. M. Washington. 1967. HRB. D. 8091. F.. 1982.. R. In Transportation Research Record 1090. Air Freight Usage Patterns of Technology Based Industries. Vol. S. Trip Generation for SpecialUse Truck Traffic. Mahmassani. Effect of UnitTrain Grain Shipments on Rural Nebraska Roads. and T.9 3. pp. P. 3954. HRB. and G. In Highway Research Record 401. 1972. Transportation Science. 6270. M. D.C. Predicting Transporter’s Choice of Mode. In Transportation Research Record 1061. Kean. pp.3 Guidebook on Statewide Travel Forecasting . B. In Transportation Research Record 875.2 4. D. TRB.. 3339. and K. pp. Determinants of Freight Modal Choice. National Research Council. R. D. TRB.. A Comparative Analysis of Statewide Transportation Studies. Journal of Transportation Engineering. 113. In Highway Research Record 205. Linsenmeyer. National Research Council. Bisson. L. R. pp. Jelavich. Washington. 1986. 813. Ashtakala. National Research Council. 1318. Factors that Determine Mode Choice in the Transportation of General Freight. D.4 4. J. In Transportation Research Record 637.C. pp..C.12 3. C. 502519. In Transportation Research Record 1179. 3. 5. National Research Council. pp. Washington.C. R. T. D. pp. Hazen.. Washington. Approach to Measurement of Modal Advantage. National Research Council.. D.. E. TRB. No.C. A. H. 2631. I. Middleton.11 3. L.. S.3. Washington. 1986.10 3.C. September 1987. A. and B. Intercity Traffic Projections.8 Friesz. D. Sasaki. D. In Transportation Research Record 577. Carey. TRB. pp. Wilson. K. pp.C.14 3. Vol. Washington. D. 20..16 References for Section A. Cothran.15 3. National Research Council. Statewide Travel Demand Forecasting. Murthy. In Transportation Research Record 668.C. 2. Toft. S. 1977. Modal Split Analysis Using Logit Models. J. TRB. 1978. D. B. No.
. Tweedie. A. De Joubert. H. M. Werner. Washington. D. National Research Council.C.C.. R. TRB. National Research Council. Mitchell & Co. Texas. Washington..Transportation. D. M. and J. D. 2630. October. National Research Council. A Research Process for Developing a Statewide Multimodal Transport Forecasting Model. Zabinski. S. 1991. Meyer. 1960. Washington. Peat. National Research Council..11 4. and R. Mitchell & Co.12 4..C. In Conference Proceedings 4: Decennial Census Data for Transportation Planning. 4. D.. D. 1991.7 4. February. C.6 4. College Station. and C.C. pp. C. U.. J.. 1979. W. E. Russell. pp. S. D. Peat. and A. Improved Method of Grouping Provincewide Permanent Traffic Counters.C.13 4. Design and Conduct of a Statewide Household Travel Survey in Vermont. In HRB Bulletin 253. TRB. 1995. FHWA. Glass. Benson. Classifying a Rural Road Network for 125 4. Ahsan.10 4.. R. Santa Fe. N. Kansas Department of Transportation. pp. TRB.. Jatko. 7582. Census Data Use for Statewide Transportation Planning and Transportation Planning in Smaller Metropolitan Areas. New Mexico State Highway and Transportation Department. Marwick. 1525. and D. C. TRB. Niedowski.. 139144. Crevo. and M. Person Travel Forecasting Procedures for Multimodal Statewide Transportation Planning.14 4. 1973. K. National Research Council. In Transportation Research Record 1477. D. J. Washington. 1977. 1990.8 4. Investigation and Development of a Statewide Traffic Assignment Model for the State of Kansas. Bester. M. Kansas. HRB. 1995. In Transportation Research Record 815. pp. Feasibility of Developing a Statewide Modeling System for Forecasting Intercity Highway Volumes in Texas. Scott. 1217. National Research Council. Mazur. In Transportation Research Record 1283. Mullins. C.. pp. In Transportation Research Record 1477.15 The State of the Art in Statewide Travel Demand Forecasting .5 4. 4550. J. J. Texas Transportation Institute. Transportation Planning Data for the 1990’s. Topeka. Dresser. Design and Implementation of a Statewide Roadside Origin Destination Survey in Vermont.C. Washington.. D. and G. S. Sharma. Multiple Screenline Study to Determine Statewide Traffic Patterns.9 4. TRB.4 Assessment of Modeling Techniques and Data Sources for Multimodal Statewide Transportation Planning. 1981. Marwick. Keyes. D. B. 1995. Virkud. BartonAschman Associates. J..C. D. R. Washington. and G. J. Washington. D. Green. 1977. pp.
2735.. D. FHWA.20 4.27 Guidebook on Statewide Travel Forecasting . 281293. The Estimation of an OriginDestination Matrix from Traffic Counts. 1985. In Transportation Research Record 1090. In Transportation Research Record 1090. pp. Washington. TRB. Revision of Statewide Traffic Data Standards Indicated During Implementation of a Traffic Monitoring System. D. pp. 4.23 4. and L. Hendrickson. and S. E. pp.26 4. D.. C. National Research Council. DiRenzo. Hallenbeck.. G. pp. TRB.24 4.C. In Transportation Research Record 976. Albright. pp. pp. TRB. pp. Bell. 1990. National Research Council. Washington. N... National Research Council. and L.19 4. TRB.25 4.17 4. Goldstine. 4854. Washington. D. Willumsen. pp.C. 1986. Washington. National Research Council. Vol. 17. G. D.. Washington. Development of State Traffic Monitoring Standards: Implementing Traffic Monitoring Guide in New Mexico. No. S.C. Statewide Transportation Planning – An 126 4. 1985. S. TRB. S. P. Vol 14B.16 Allbright. 1319. 1984. 1980. National Research Council. G. Estimation of OriginDestination Matrices with Constrained Regression. In Transportation Research Record 1050.21 4. In Transportation Research Record 1050. National Research Council. and J. McNeil. and R. C. M. Washington. D. King. and M.C. Washington. In Transportation Research Record 1271. Washington. 2532. H. Serianni. Department of Transportation. H. 1987. TRB. 1421. May 1983. TRB.C. M. Washington. National Research Council. D. Improving the Productivity of the Pennsylvania Transportation Department’s Traffic Count Program. pp.22 4. Hallenbeck. D. E. 1986. pp. Albright. Transportation Science. Blewett. 5764. Wilson. Transportation ResearchB.C. D. National Research Council. Development of an Integrated Statewide TrafficMonitoring System. Washington. 512. J. Van Zuylen. Traffic Monitoring Guide. 913.C. A Quick Cluster Control Method: Permanent Control Station Cluster Analysis in Average Daily Traffic Calculations.18 4.. 2226.. D. E. D. D. National Research Council. M.C.. In Transportation Research Record 1236. J. In Transportation Research Record 1090. TRB.C. 2. In Transportation Research Record 1134. A Statistical Approach to Statewide Traffic Counting. C. 1985. D. Ritchie. 1986. TRB. 198217. Ritchie...Traffic Counting. pp. 1989. Wang. Evaluation of a Statewide Highway Data Collection Program. The Most Likely Trip Matrix Estimated from Traffic Counts. G.. U.
New Hampshire Statewide and Subarea Travel Models Plan. April 1996. U.38 4. Walker. D. Statewide Travel Demand Model Update and Calibration: Phase II.Interactive Modeling Process. National Research Council. J. Improving Traffic Forecasting with Time Series Models. May 1995.29 4. 5058. E. W. 16. T. 1997. PERFORM Person Forecasting Model: Trip Generation.32 4. and J. D. Florida Department of Transportation. Washington. October 1990. 4. Shimpeler. Harmatuck. TRB. February 1997. D.37 4. April 1997. Statewide Highway Traffic Forecasting Model Technical Report 2: Collection and Assembly of Data.. 1995. pp. ConnDOT Modeling Procedure. Florida Department of Transportation.30 4. Minnesota Department of Transportation.33 4. Shimpeler. M. Kentucky Statewide Traffic Model Final Calibration Report.C.34 4. Connecticut Department of Transportation. March 1995. 4. Corridino Associates. New Hampshire Department of Transportation. Cambridge Systematics. TRB. Traffic Monitoring Guide.. In Transportation Research Record 1236. Major Corridor InvestmentBenefit Analysis System. Minnesota.28 Wilson. Traffic Forecasting Unit. Kentucky Transportation Cabinet. Indiana Department of Transportation. In Transportation Research Record 1499. Minneapolis. Corridino Associates. Michigan Department of Transportation. 1989. Appendix A: Scope of Work. D. April.39 4. S. Inc. Washington. September 1990. Wang. Staff Paper 9310. pp. Wisconsin Department of Transportation. Procedures Manual for Forecasting Traffic on the Rural Trunk Highway System. C. Department of Transportation.31 4.40 The State of the Art in Statewide Travel Demand Forecasting 127 . Madison. Wilbur Smith Associates.C. 1993. Wyoming Department of Transportation. Cambridge Systematics. Method to Synthesize a Full Matrix of Interdistrict Highway Travel Times from Census JourneytoWork Data. 1992. KJS Associates. Interactive Statewide Transportation Planning Modeling Process. 1985.. FHWA.36 4. Washington. January. Florida Highway Traffic Forecasting Models Technical Report 1: Statewide Model Development. Undated. National Research Council. Connecticut Department of Transportation.35 4.
1972. Inc.41 URS Consultants. 1963. National Research Council. 4562.43 4. D.. National Research Council. National Research Council. D. S. C.C.. Models of Recreational Travel. 1968.44 4. HRB. Crevo. Washington. L. Berg. Inc. and W. D. June 1995. Effects of Interstate Completion and Other Major Highway Improvements on Regional Trip Making and Goods Movement: Network Development. Characteristics of Summer Weekend Recreational Travel.C. HRB. D.4. 147156. Gyamfi. In Highway Research Record 392. D. Matthias. Effects of Interstate Completion and Other Major Highway Improvements on Regional Trip Making and Goods Movement: Auto Trip Table. April 1996. G. and W. D. pp. Pigman. L. Deen. National Research Council.46 4. D. Hangen Brustlin. C.47 4. P. URS Consultants. In Highway Research Record 392. J. Washington. C.C. Krueger. September 13. J. Koushki.. Travel to Outdoor Recreation Areas in Kentucky. A.. Washington. 5160. Pigman. D.. and Forecasting Methods: Transportation and Land Use Model Integration Program Phase I. In Highway Research Record 472.53 Guidebook on Statewide Travel Forecasting . 128 4. New Jersey Department of Transportation. National Research Council. Washington.C.. 1972. C. J. 1996. J.. A Model for Allocating Recreational Travel Demand to National Forests. Grecco. Models. In Highway Research Record 408. Vanasse. W. pp. 1972. NCHRP Report 44: Traffic Attraction of Rural Outdoor Recreational Areas. C. Kaltenbach.48 4.51 4. 1967. D. Translinks 21: MultiModal Intercity Passenger Analysis. Inc. Bittner. and R.C. Deen. Ungar. Wisconsin Department of Transportation. June 1995. HRB.6. K.52 4. and R.C. L. Task 1. pp. March 1995. In Highway Research Record 250. Deacon. Consultant Recommendations for ths Development of Phase II Databases. National Research Council. Statewide Travel Demand Model Development. Parsons Brinkerhoff Quade & Douglas. Vermont Agency of Transportation. Deacon. Washington. HRB. Stability of Recreational Demand Model. 134135. P. HRB.45 4. 5469. Grecco. L.C. Salem. Simplified Procedure for Estimating Recreational Travel to MultiPurpose Reservoirs. Washington..42 4. J.50 4. New Jersey Department of Transportation. C..49 4. In Highway Research Record 41. Robinson.C. 1973. 5061. pp. Inc. A. pp. D. Washington. Oregon Department of Transportation.. and W. HRB. A. pp. KPMG Peat Marwick. National Research Council. HRB. A.
Washington. In Profitability and Mobility in Rural America (W. 1519. Washington. Development and Application of Statewide. 1977. NCHRP Report 177: Freight Data Requirements for Statewide Systems Planning. Middendorf. and D. D.C. and R. Users Manual. TRB. Roger Creighton Associates. TRB. 1995. National Research Council. Salem. Casavant. R.6 5. 1997. W. pp. W.8 5. NCHRP Report 178: Freight Data Requirements for Statewide Systems Planning.2 5. Washington. National Research Council. Model for Statewide Freight Transportation Planning. Parsons Brinkerhoff Quade & Douglas. W.. D.3 5. In Transportation Research Record 569. Inc. D. F. Gillis. Blankenship. D.9 5. NCHRP Report 260: Application of Statewide Freight Demand Forecasting Techniques.5 5. In Transportation Research Record 1477. Consultant Recommendations for the Development of Phase II Databases. Practical Methodology for Freight Forecasting.C. Washington. Memmott. Jelavich. Washington. 1977. National Research Council. Linsenmeyer.). In Transportation Research Record 889. 17. K.5 5. D.6. Azzam.. Survey Methodology for Collecting Freight Truck and Destination Data. H..C. Task 1. Roger Creighton Associates. National Research Council. Memmott. 1983. TRB... 1978.C. TRB. Washington. TRB.7 5.1 Cambridge Systematics. TRB. 1976. A. and J.. Washington. In Transportation Research Record 889. pp. J. Olsen. Kim T.. Howard. Research Report. pp. In Transportation Research Record 889. D. M. D. National Research Council. TRB.C. Pennsylvania 129 5. TRB. Hinkle. J. and Forecasting Methods: Transportation and Land Use Model Integration Program Phase I. D.. National Research Council. D.C. 1982. ed. D. 96106. 714. pp. Ellis. Multimodal Freight Forecasting Procedures for Florida. D.. Boekenkroeger. F. D.. Oregon Department of Transportation. P. National Research Council. Planning Rural Transportation Systems: Applications of a Statewide Network Model. NCHRP Report 388: A Guidebook for Forecasting Freight Transportation Demand..C. pp. and C. 1996. Models. H.4 5. References for Section A. September 13. 1982.C. Washington.. National Research Council. National Research Council. L. TRB. Washington. and R.C. Gillis.Development of a Simulation Model for Regional Travel.10 The State of the Art in Statewide Travel Demand Forecasting . 714.
July 1996. September 1996. 1993. Inc. FHWA. Vol. Department of Transportation and Development.. T. Transportation Science. Indiana University. Vol. Murthy. July 15. 1990. Multimodal Freight Forecasts for Wisconsin. S. 4. State of Louisiana. and J. pp.17 5. Inc.G. D. Michigan Department of Transportation. July 1988.14 5.19 5.State University Press. 1989. No. Murthy. Statewide Travel Demand Model Update and Calibration: Phase II. 5.1 KJS Associates.6 6. Wilbur Smith Associates. pp. Quick Response Freight Manual. No. Inc. M. 24. Strategi Planning of Freight Transportation: STAN. pp. 1. Florian. Ashtakala. URS Consultants. Crainic. No. 5. 130 Guidebook on Statewide Travel Forecasting .13 5. Transportation Research Center. Florian. TRB. Vol.G. J. Bloomington. Journal of Transportation Engineering. No. 1997. Cambridge Systematics. Michigan Department of Transportation. February 1990. August 1995. 5. S.11 Statewide Intermodal Transportation Plan. Crainic. 27A.. Phase 2. 2539. 171191. and A. Florian.15 5. A Multimode Multiproduct Network Assignment Model for Strategic Planning of Freight Flows. T. pp. 114. An InteractiveGraphic System. R. 97124. Optimized Gravity Models for Commodity Transportation. Leal. and T. Effects of Interstate Completion and Other Major Highway Improvements on Regional Trip Making and Goods Movement: Truck Trip Table. W. B.. N. Spiess. Guelat. 373382. Transport Flows in the State of Indiana: Commodity Database Development and Traffic Assignment. A Model for the Strategic Planning of National Freight Transportation by Rail. KJS Associates. G. pp. March 1996.21 References for Section A.18 5. 1. Vol. M. Black. Transportation Science. 124. In Transportation Research Record 1283. and A.16 5. pp. Crainic. M. Transportation ResearchA.C. Wisconsin Department of Transportation. New Jersey Department of Transportation.12 5. April 1996. Washington. Statewide Travel Demand Model Update and Calibration: Phase II.E. J. Sequential Models to Determine Intercity Commodity Transportation Demand. February 1990. and H. Ashtakala. Guelat. National Research Council. 24..20 5. 393408. B. April 1996.
Indiana University. Wisconsin Department of Transportation. 7. July 15.6. Transportation Research Center. 1997. R. July 15. W.1 Black. April 1997. Bloomington.8 8. Indiana University.2 Wilbur Smith Associates. Kentucky Transportation Cabinet. R.2 References for Section A. Kentucky Statewide Traffic Model Final Calibration Report. Transport Flows in the State of Indiana: Commodity Database Development and Traffic Assignment. Transport Flows in the State of Indiana: Commodity Database Development and Traffic Assignment. Bloomington. 1997. Multimodal Freight Forecasts for Wisconsin. July 1996. The State of the Art in Statewide Travel Demand Forecasting 131 . Black. W.7 7. Transportation Research Center.1 Wilbur Smith Associates. References for Section A. Phase 2. Phase 2.
This action might not be possible to undo. Are you sure you want to continue?