Travel Model Development and Refinement - Trip Generation - Final Report
Click HERE for graphic.
DKS Associates
Contents
EXECUTIVE SUMMARY
1. INTRODUCTION. . . . . . . . . . . . . . . . . . . . . . . . . 7
2. LITERATURE REVIEW/STATE OF THE PRACTICE . . . . . . . . . . . 9
3. GENERAL STRATEGY FOR THE REGIONAL COUNCIL'S TRIP GENERATION
MODEL . . . . . . . . . . . . . . . . . . . . . . . . . . . .27
4. ADAPTATION OF LOCAL SURVEY DATA FOR TRIP GENERATION ANALYSIS.30
5. TRIP PRODUCTION ANALYSIS. . . . . . . . . . . . . . . . . . .37
6. RECOMMENDED TRIP PRODUCTION MODEL . . . . . . . . . . . . . .66
7. TRIP ATTRACTION ANALYSIS. . . . . . . . . . . . . . . . . . .77
8. SPECIAL GENERATORS AND COMMERCIAL VEHICLES. . . . . . . . . .88
9. NEXT STEPS. . . . . . . . . . . . . . . . . . . . . . . . . .98
BIBLIOGRAPHY
DKS Associates
Tables
Table 5
Cross-Sectional Travel Survey Validations. . . . . . . . . . . . .32
Table 5 (continued)
Cross-Sectional Travel Survey Validations. . . . . . . . . . . . .33
Table 6
Activity Codes in the Cross-Sectional Travel Survey. . . . . . . .36
Table 7
Conversion of Activity Codes to Model Trip Purpose . . . . . . . .36
Table 8
Estimated Home-Based-Work Trip Productions
Core Variables . . . . . . . . . . . . . . . . . . . . . . . . . .41
Table 9
Comparison of Core Variables (Analysis of Variance)
Home-Based-Work Trips. . . . . . . . . . . . . . . . . . . . . . .42
Table 10
Estimated Home-Based Work Productions
By Number of Workers . . . . . . . . . . . . . . . . . . . . . . .43
Table 11
Estimated Trip Productions for Home-Based Work
Workers + Household Income
(Main Effects, adjusted for interaction) . . . . . . . . . . . . .43
Table 12
Estimated Trip Productions for Home-Based Work
Workers x Household Income
(Full Cross-Classification). . . . . . . . . . . . . . . . . . . .44
Table 13
Estimated Trip Productions for Home-Based Work
Workers + Vehicles
(Main Effects, adjusted for interaction) . . . . . . . . . . . . .45
ii
DKS Associates
Table 14
Estimated Trip Productions for Home-Based Work
Workers x Vehicles
(Full Cross-Classification). . . . . . . . . . . . . . . . . . . .45
Table 15
Comparison of Combined Variables (Analysis of Variance)
Home-Based Work Trips. . . . . . . . . . . . . . . . . . . . . . .46
Table 16
Estimated Home-Based-Shop Trip Productions
Core Variables . . . . . . . . . . . . . . . . . . . . . . . . . .47
Table 17
Comparison of Core Variables (Analysis of Variance)
Home-Based Shop Productions. . . . . . . . . . . . . . . . . . . .48
Table 18
Estimated Home-Based Shop Trip Production
By Number of Persons . . . . . . . . . . . . . . . . . . . . . . .48
Table 19
Estimated Home-Based Shop Trip Productions
Persons + Workers
(Main Effects, adjusted for interaction) . . . . . . . . . . . . .49
Table 20
Estimated Home-Based Shop Trip Productions
Persons x Workers
(Full Cross-Classification). . . . . . . . . . . . . . . . . . . .49
Table 21
Estimated Home-Based Shop Trip Productions
Persons + Household Income
(Main Effects, adjusted for interaction) . . . . . . . . . . . . .50
Table 22
Estimated Home-Based Shop Trip Productions
Persons x Household Income
(Full Cross-Classification). . . . . . . . . . . . . . . . . . . .50
iii
DKS Associates
Table 23
Estimated Home-Based Shop Trip Productions
Persons + Vehicles
(Main Effects, adjusted for interaction) . . . . . . . . . . . . .51
Table 24
Estimated Home-Based Shop Trip Productions
Persons x Vehicles
(Full Cross-Classification). . . . . . . . . . . . . . . . . . . .51
Table 25
Comparison of Combined Variables (Analysis of Variance)
Home-Based Shop Productions. . . . . . . . . . . . . . . . . . . .52
Table 26
Estimated Home-Based-School Trip Productions
Core Variables . . . . . . . . . . . . . . . . . . . . . . . . . .53
Table 27
Estimated Home-Based School Trip Productions
Persons x Workers
(Full Cross-Classification). . . . . . . . . . . . . . . . . . . .54
Table 28
Comparison of Cross-Classification Schemes (Analysis of Variance)
Home-Based-School Productions. . . . . . . . . . . . . . . . . . .55
Table 29
Comparison of Core Variables (Analysis of Variance)
Estimated Home-Based-Other Trip Productions. . . . . . . . . . . .56
Table 30
Comparison of Core Variables (Analysis of Variance)
Home-Based-Other Productions . . . . . . . . . . . . . . . . . . .56
Table 31
Estimated Home-Based Other Trip Productions
By Number of Persons . . . . . . . . . . . . . . . . . . . . . . .57
Table 32
Estimated Home-Based Other Trip Productions
Persons + Workers
(Main Effects, adjusted for interaction) . . . . . . . . . . . . .57
iv
DKS Associates
Table 33
Estimated Home-Based Other Trip Productions
Persons x Workers
(Full Cross-Classification). . . . . . . . . . . . . . . . . . . .58
Table 34
Estimated Home Based Other Trip Productions
Persons + Household Income
(Main Effects, adjusted for interaction) . . . . . . . . . . . . .58
Table 35
Estimated Home-Based Other Trip Productions
Persons x Household Income
(Full Cross-Classification). . . . . . . . . . . . . . . . . . . .59
Table 36
Estimated Home-Based Other Trip Productions
Persons + Vehicles
(Main Effects, adjusted for interaction) . . . . . . . . . . . . .59
Table 37
Estimated Home-Based Other Trip Productions
Persons x Vehicles
(Full Cross-Classification). . . . . . . . . . . . . . . . . . . .60
Table 38
Estimated Home-Based Other Trip Productions
Workers + Presence of Children Under 5
(Main Effects, adjusted for interaction) . . . . . . . . . . . . .61
Table 39
Estimated Home-Based Other Trip Productions
Workers x Presence of Children Under 5
(Full Cross-Classification). . . . . . . . . . . . . . . . . . . .61
Table 40
Comparison of Combined Variables (Analysis of Variance)
Home-Based-Other Productions . . . . . . . . . . . . . . . . . . .62
Table 41
Estimated Home-Based College Trip Productions
Core Variables . . . . . . . . . . . . . . . . . . . . . . . . . .64
vi
DKS Associates
Table 42
Estimated Home-Based College Trip Productions
Persons x Workers
(Full Cross-Classification). . . . . . . . . . . . . . . . . . . .65
Table 43
Comparison of Cross-Classification Schemes (Analysis of Variance)
Home-Based College Productions . . . . . . . . . . . . . . . . . .65
Table 44
Interim Home-Based Work Trip Production Rates. . . . . . . . . . .69
Table 45
Interim Home-Based Shop Trip Production Rates. . . . . . . . . . .70
Table 46
Interim Home-Based School Trip Production Rates. . . . . . . . . .70
Table 47
Interim Home-Based Other Trip Production Rates . . . . . . . . . .71
Table 48
Interim Home-Based College Trip Production Rates . . . . . . . . .71
Table 49
Comparison of Cross-Sectional and Panel Surveys
Home-Based Work Trip Productions . . . . . . . . . . . . . . . . .72
Table 50
Comparison of Cross-Sectional and Panel Survey
Home-Based Shop Trip Productions . . . . . . . . . . . . . . . . .72
Table 51
Comparison of Cross-Sectional and Panel Survey
Home-Based Other Trip Productions. . . . . . . . . . . . . . . . .73
Table 52
Comparison of Cross-Sectional and Panel Surveys
Home-Based Work, Shop and Other Trip Productions . . . . . . . . .73
Table 53
Comparison of Cross-Sectional and Panel Survey
Household Sample Sizes . . . . . . . . . . . . . . . . . . . . . .74
vi
DKS Associates
Table 54
Trip Purposes Used by Other MPO's. . . . . . . . . . . . . . . . .79
Table 55
Trip Attractions Estimated for PSRC Region Using Rates and
Equations from Other MPOs. . . . . . . . . . . . . . . . . . . . .81
Table 56
Comparison of Attraction Rates of PSRC to Synthesis from Other MPOs83
Table 57
Comparison of Attraction Trip Estimations of PSRC to Synthesis from
Other MPOs . . . . . . . . . . . . . . . . . . . . . . . . . . . .84
Table 58
Comparison of Trip Attraction Rates
Current Model versus Revised Trip Rates for Model
Calibration. . . . . . . . . . . . . . . . . . . . . . . . . . . .86
Table 59
Comparison of 1990 Trip Productions and Attraction
Current Model versus Revised Trip Estimates for Model Calibration.87
Table 60
Commercial Vehicle Trip Rates
Phoenix Urban Truck Travel Model Project . . . . . . . . . . . . .90
Table 61
Trip Rates for Truck and Taxi Trips
Florida. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .90
Table 62
Truck Trip Rates
San Francisco Bay Area . . . . . . . . . . . . . . . . . . . . . .91
Table 63
Comparison of Commercial Vehicle Trip Rates. . . . . . . . . . . .93
Table 64
Internal - External Person Trips
University of Washington . . . . . . . . . . . . . . . . . . . . .95
Table 65
Comparison of Trip Generation Rates for the
University of Washington . . . . . . . . . . . . . . . . . . . . .95
vii
DKS Associates
Table 66
University/College Trip Generation Rates . . . . . . . . . . . . .96
Table 67
Community College Trip Rates . . . . . . . . . . . . . . . . . . .96
Table 68
Faculty, Staff, and Students at Universities and Colleges in the
PSRC Model Region. . . . . . . . . . . . . . . . . . . . . . . . .97
viii
Executive Summary
DKS Associates
Executive Summary
The Puget Sound Regional Council has undertaken an effort to
enhance and refine its travel demand modeling capabilities. This
project, which is part of the Regional Council's long range plan
for model development, focusses on the trip generation component of
the modeling chain. The intent of the trip generation model
development effort is to redefine the structure of the current
model and meet the requirements of the Intermodal Surface
Transportation Efficiency Act (ISTEA) and federal legislation as
described in the 1991 Clean Air Act Amendments (CAAA). The project
will include examination of the current practices in trip
generation modeling, data needs for model development,
redevelopment of model structure, and the implications for further
model development.
Literature Review and Model Strategy
A literature review was conducted on the latest efforts in trip
generation modeling throughout the U.S. and overseas. Based on
this review, and an analysis of the Regional Council's current
model and data availability, a general model strategy was
developed.
Any model improvement strategy should focus both on making
immediate adjustments to enhance model validity and credibility and
on ensuring a smooth path to more extensive but longer-range and
more fundamental changes. Any strategy that uses carefully checked
data from the recent household travel diary surveys to produce new
trip generation estimates should improve the travel models
performance. Other key considerations for the new trip generation
models are:
- It includes travel by all modes including non-motorized modes.
- It uses cross-classification as its basic methodology.
- ANOVA methods should be used to measure and compare model fit.
- The partitioning of trips into purposes should preserve as
much information about the activity pattern as possible.
- Predictions for new trip purposes may be combined to match the
trip purposes in the remainder of the current model system.
- The model development should explore a basic core of household
variables including household size, number of workers, income,
auto availability and the age of household members.
- Additional variables relate to life-cycle status should be
considered.
P93168-14/SA 1
DKS Associates
Trip Production Analysis
The trip production analysis involved developing and testing a
number of alternative cross classification schemes for each trip
purpose. The 1985-1988 Puget Sound cross-sectional survey was the
primary source for the household trip production analysis. A
number of "filters" were applied to eliminate households associated
with missing, invalid, or inconsistent household and/or travel
data. These filters excluded 24 percent of the surveyed
households. Compared to other travel surveys, this rejection rate
is actually quite low and is evidence of a good level of integrity
of the survey data. The remaining 3507 households were still quite
adequate for the trip generation analysis. The Regional Council
may do well to review and correct some of the excluded observations
in preparation for its future work on a mode choice model, due to
mode choice models' sensitivity to and dependence on "unusual"
observations (i.e. "outliers").
The significance of various household characteristics, alone and in
combinations, in explaining home-based trip productions in certain
trip purposes was analyzed, using the 'filtered' 1985-88 Puget
Sound cross-sectional survey. The household characteristics
studied were:
- Number of persons in household,
- Number of employed persons in household,
- Household income,
- Number of vehicles available to household members,
- Persons in various age groups, applicable to the trip purpose,
in household.
Trips included all modes including walk and bicycle. The trip
purposes analyzed were:
- Home-based work
- Home-based shop
- Home-based school
- Home-based college
- Home-based other
Major conclusions found in the trip production analyses include:
- Most of the household variables have statistically significant
relationships to each trip purpose, apparently due to multiple
colinearities between a household's numbers of persons,
workers, vehicles, and income.
- In general, the strongest single predictor of trips in a given
purpose is the number of persons eligible or most expected to
make such trips. For work trips, this is the number of
employed persons; for school trips, it is the number of
school-age children; for college trips, it is the number of
persons in a college age group. (The college trip analysis
checked persons age 18 to 24. Measures of proximity to a
college were unavailable.) For shop and other trips, this is
simply the number of persons counted as trip makers,
P93168-14/SA 2
DKS Associates
in this case, persons age 5 and older.
- Other household variables can be used as less-direct
predictors. Counts of persons approximating the most direct
measures also yield strong models (such as total persons in
lieu of persons age 5 and up). Other measures, such as income
and vehicles, give statistically significant but less strong
models, because they are only indirectly related to numbers of
particular groups of persons. (Some such variables should be
more strongly related to mode choice than to trip generation.)
- Several two-variable models demonstrated statistical
significance over the respective one variable models, but the
greatest part of variance is still explained by the strongest
one variable model. In particular, Persons x Workers models
show home-based shop and home-based other trips tend to be
made more by non-workers, and offer a strong substitute for
school-age children in predicting home-based school travel
(should direct predictions of school-age children be
unavailable). Some of the other two-way models may be used to
predict secondary effects (say, of income) in conjunction with
principal predictor variables, should this be deemed
necessary.
Implementation of the strongest model in each trip purpose requires
development of both current and future year dataset of household in
a cross-classified form. The future year household forecasts could
initially be based on the current year stratification of
households. A true forecast of household cross-classified by
workers and persons will require a land use model that can predict
changes in household demographics.
The Regional Council would need to develop the following data for
each traffic zone to implement the recommended trip generation
model:
- Number of households in 16 cross-classified categories using
persons (1, 2, 3, 4+) x workers (0, 1, 2, 3, 3+)
- Persons 5-17 in household (with households in a zone
classified into 0, 1, 2, 3+)
- Persons 18-24 in household (with households in a zone
classified into 0, 1, 2+)
The above classification schemes could be developed using the 1990
Census Public Use Microdata Sample (PUMS) and techniques developed
by other MPO's (i.e., Portland and Sacramento). Due to the level
of effort needed to prepare this 1990 dataset, and anticipating the
eventual need to update the mode choice model, it was recommended
that the Regional Council also include income as a cross-classified
variable.
Trip rates were estimated for both total person trips (with walk
and bike modes) and for motorized person trips, both using the
recommended cross-classification schemes. The use of motorized
trip rates is equivalent to using percentages for each household
classification as a "pre-
P93168-14/SA 3
DKS Associates
mode choice" model. A better estimate of walk and bike mode choice
could be made but only as part of an overall mode choice model
update that includes key variables that are important to the walk
and bike mode choice (variables not used in the Regional Council's
current model structure). Until the mode choice model is updated,
the motorized trip rates for the new cross-classified household
schemes should be used.
Trip Attraction Analysis
Non-residential trip generation has received considerably less
attention than home-based trip generation, and the techniques that
have been used are generally less sophisticated. The Regional
Council currently uses a two-way cross-classification scheme using
6 employment categories and 3 activity density categories.
Attempts were made to estimate "relative attraction rates for each
employment category using both regression analysis and "aggregate
maximum likelihood' techniques. These methods were not successful
in developing reasonable attraction rates. Therefore, the analysis
focussed on comparisons between the Regional Councils current
attraction rates and those from a number of other MPO'S. This
analysis resulted in some minor adjustment to the Regional
Council's attraction rates. These rates may be adjusted further in
a re-calibration of the regional model using the new trip
production rates.
Commercial Vehicle Trips
Commercial vehicle trips presented a special difficulty in the trip
generation study because these trips are not reported in the
household surveys. Their estimation would require separate surveys
of commercial vehicle owners or other special data. For these same
reasons, few regional travel models have estimated separate trip
rates for commercial vehicles. Many modelers use the approach
taken by the Regional Council; they estimate commercial vehicles as
a percentage of the estimated person or vehicle trip attractions
for some or all trip purposes. The Regional Council currently
estimates commercial vehicle trips based on the number of non-home-
based trips produced and attracted in a zone.
No recent data is available on commercial vehicle trips in the
Puget Sound region. Therefore, improving the estimation of
commercial vehicle trips requires either special surveys of
commercial vehicle owners, or transferring trip generation rates
(and possibly other model parameters) from other regional models
where special surveys were conducted. Although strategies used to
survey and estimate commercial vehicle trips are well documented
and straight forward to apply, the costs of doing the data
collection is significant. For this trip generation analysis, the
transferability of commercial vehicle trip generation rates was
explored.
Commercial vehicle trip generation data from Phoenix, Chicago,
Florida and the San Francisco Bay Area was reviewed. It was
decided that the Phoenix commercial vehicle trip rates were the
most appropriate to transfer to the Regional Council's model since
they:
P93168-14/SA 4
DKS Associates
- included all commercial vehicle trips not accounted for in the
household surveys.
- provide employee-based trip rates for most of the non-
residential categories used by the Regional Council.
- were based on the most recent surveys of commercial vehicle
owners in the U.S.
Special Generators
Discussions with the staff of the Regional Council indicated that
they are comfortable with the trip generation for the major
generators in their current model. DKS Associates and Rao
Associates, however, have identified the following additional
categories that should be considered as special generators in the
revised model:
- Resident Colleges, including the University of Washington,
have significant percent of students living on or near these
campuses.
- Community Colleges which are purely "commuter" colleges.
- Major military bases which have a large number of employees
but may also have a large number of on-base military housing.
Trip rates for these additional special generators have been
recommended.
Next Steps
This trip generation project is only a first step in updating the
Regional Council's full model chain based on data from the recent
household surveys, and to fully incorporate walk and bike trips
into the model. Some important next steps that will lead to the
eventual full model update include the following:
- recalibration of the current model structure using "motorized"
trip rates developed in this project
- developing some basic strategies for the updates of the model
choice and distribution models
- developing the data needed for the mode choice and
distribution models
A recalibration of the Regional Council's current model structure
using the motorized person trip rates outlined in this report will
be more limited in scope than a "full" recalibration effort
involving updates to the mode choice and distribution models and
other structural changes. Given that other elements of the model
will remain relatively unchanged, the recalibration effort will
P93168-14/SA 5
DKS Associates
principally focus on comparing traffic assignments to traffic count
data using a regional and systematic approach.
The Regional Council has a large number of "screenlines" that it
uses for both model calibration/validation and for analysis. It is
likely that initial 1990 model assignments using the new trip
production rates will generally be lower than 1990 traffic count
data when all screenlines are viewed in aggregate. This will
likely be due to some under-reporting of trips in the household
survey data.
It is generally believed that work and school trips are not subject
to significant under-reporting in a household survey since these
are regular trips known by all household members (except in cases
where a person has a very irregular work schedule). It is also
generally believed that short distance non-work trips are most
affected by under-reporting. A trip made across town for any
purpose is seldom not reported while a quick trip to the store may
not be reported.
If the model assignments indicate that short distance are a bigger
problem than longer distance trips then the trip distribution
should also be changed to truly reflect the impact of missing short
distance trips in the survey data.
The screening process of the household survey data eliminated about
24 percent of the households due to missing or invalid data. The
screened database of 3,507 households provides an adequate survey
size for development of the trip generation model. Estimation of a
sophisticated mode choice model, however, would benefit from a
larger sample size. Therefore, the Regional Council should have
trained analysts systematically reviewing the rejected
observations, and attempt to correct as many as possible, using the
original data sheets and judgement.
Other key data issues for the mode choice and destination model
updates involve development of variables that are important to walk
and bike travel. These include the following:
- Development of "pedestrian environment factors" - A variety of
factors have been used in other models including ones in
Maryland, Portland and Sacramento. These factors have
included the availability of sidewalks, continuity of streets,
topography and other barriers to walking. A GIS system could
help provide these factors, but a careful review of maps and
aerials and consistent judgement is an adequate method.
- Estimation of household auto ownership (or auto availability).
This will likely require a new submodel, but its estimation is
important to walk, bike, transit and HOV modes.
- Improving the walk distance measurements in the network that
do not merely rely on auto centroid connectors and auto
related distance.
- Development of accessibility measures for walking, such as the
number of total jobs and retail jobs within one mile of the
household.
P93168-14/SA 6
1. Introduction
DKS Associates
1. Introduction
The Puget Sound Regional Council has undertaken an effort to
enhance and refine its travel demand modeling capabilities. This
project, which is part of the Regional Council's long range plan
for model development, focuses on the trip generation component of
the modeling chain. The intent of the trip generation model
development effort is to redefine the structure of the current
model and meet the requirements of the Intermodal Surface
Transportation Efficiency Act (ISTEA) and federal legislation as
described in the 1991 Clean Air Act Amendments (CAAA). The project
will include examination of the current practices in trip
generation modeling, data needs for model development,
redevelopment of model structure, and the implications for further
model development.
The travel model refinements will be based on two household travel
survey resources: 1) the "cross-sectional" travel survey of about
4,500 household collected in the four-county region between 1985
and 1988, and 2) multi-year travel diary data of about 1,600
households from the Puget Sound Transportation Panel. The travel
behavior information from these surveys enables the Regional
Council travel model to be based on recent local data, and the
survey's detailed demographic information allows the development of
a more sophisticated model structure.
Regional Council's travel model is an important tool for the
Seattle region. The model fulfills a very important role in the
transportation, land use, and air quality planning of the region.
Typical applications of the model include:
- Regional Transportation Plan Update
- ISTEA Congestion Management System
- Washington State SIP
- Emission estimates for air quality conformity analysis of
transportation plans and programs
- Transit systems planning and alternatives analysis
- Corridor or sub-area analysis
- Market based transportation control measures
- County and City General Plans
In addition to Regional Council, other "users' of the model include
counties and cities in the region, WSDOT, Metro, the University of
Washington, and transportation consultants. Regional Council has
organized a Technical Advisory Committee (TAC) that includes a
number of the public users. There are also many "non-users"
directly affected by the model, who either use the model's output
or who have significant interest in its process and results. For
all these reasons, the model's structure must be sound, and its
estimation must be based on travel data that has been carefully
checked.
The trip generation model development effort involved the following
tasks:
P93168-14/SA 7
DKS Associates
1) Literature review on the latest efforts in trip generation
modeling.
2) Development of a general strategy for the trip generation
model that identifies the trip purposes that will be used and
the demographics and accessibility variables that will be
tested for potential use in the new trip generation model.
3) Testing of alternative cross-classification schemes and
selection of a trip production model that includes non-
motorized person trips.
4) Refinement of the trip generation attraction model.
5) Enhancement of external trip making characteristics and trip
balancing.
6) Enhancement of special generator modeling capabilities.
7) Examining the feasibility of including the effects of
congestion and/or the effects of lifecycle into the
traditional trip generation model.
8) Development of computer code to implement the trip generation
model and preparation of a final report on the model
development
Interim reports were prepared on the literature review, development
of a general model strategy, and the trip production and attraction
analyses. Those interim reports have been incorporated into this
Draft Final Report, along with comments from the Technical Advisory
Committee. This report also contains new documentation on the
analysis of special generators, commercial vehicles and trip
balancing.
P93168-14/SA 8
2. Literature Review/State of the Practice
DKS Associates
2. Literature Review/State of the Practice
Overview
Trip generation models are used to predict the number of trips
originating from and attracted to a zone ("trip ends"), usually on
a daily basis and for several trip purposes. These purposes vary
from region to region, primarily based on the sophistication and
complexity of the model system. In the simplest cases, home-based
work, home-based non-work, and non-home-based trips have trip
generation models. More complex model systems may split home-based
non-work trip origins into shop, school, and other; non-home-based
trip origins into work-related and other; and zonal attractions
into a variety of employment and land use categories.
Trip ends are considered to be either productions (typically
emanating from the home or from the beginning point of a non-home-
based trip) or attractions (to the point where an out-of-home
activity will be undertaken). Separate models are used to predict
productions and attractions. Simple logic requires that the
numbers of productions and attractions be equal. However, since
trip productions and trip attractions are calculated independently,
total productions do not necessarily equal total attractions. This
discrepancy is usually handled mechanically, by multiplying each
zone's trip attraction by the ratio of total productions to total
attractions (i.e., demonstrating greater faith in the production
models). More elaborate balancing algorithms are sometimes used.
Variables commonly used to estimate trip productions include
household size, number of workers, income, and auto ownership; land
use factors such as residential density, and accessibility factors
such as distance of the zone from the central business district,
(CBD) are less frequently included. Trip attraction variables
include employment levels disaggregated by occupation type and
floor space disaggregated by business type; accessibility to the
work force, represented by travel times, is rare in US applications
but found fairly frequently in applications overseas.
Most trip generation models have considered only trips made by
vehicle (often called "vehicle trips", but more accurately called
'person trips by vehicle"), although some MPOs have developed
models which estimate total trips regardless of mode ("total person
trips"). In the most common approach to producing estimates of
person trips by vehicle, trips on foot or by bicycle are excluded
from the data sets used to estimate trip generation rates. An
alternate approach, mostly used by smaller MPO,s and local
jurisdictions, directly estimates auto trips, i.e., transit trips
also are excluded from the estimation data sets. The latter models
are typically used in highway capacity and level of service
studies, or where transit is virtually nonexistent.
The distinction between modeling person trips by vehicle and
modeling total person trips, rather than being purely semantic, is
a fundamental issue in model development. Modelers have preferred
to exclude walk trips as early in the modeling sequence as
possible, because doing so avoids complexity in mode choice (i.e.,
the need to introduce walk as an explicit mode, and to develop
metrics for variables that determine the propensity to walk) and in
trip distribution (i.e.,
P93168-14/SA 9
DKS Associates
the need to develop an accessibility measure that covers both walk
trips and trips by vehicle). However, recent work has shown that
accessibility and land use conditions are powerful determinants of
the decision to walk (and to link trips into complex chains), and
thus strongly influence the number of person trips by vehicle (see,
eg., Portland METRO [1991] and Replogle [1991]). Such strong
correlations are not apparent in the total number of person trips.
Hence, there is a clear tradeoff between introducing complexity in
trip generation and introducing complexity in later model stages.
Given the desirability of accounting for pricing and land use
options, and of properly representing the effects of congestion,
some model developers are opting for greater complexity in later
model steps, in return for a simplification of trip generation.
Basic Practice
Two general approaches to trip generation are in common use: cross-
classification analysis and regression models (see Dickey [1983],
FHWA [19751, and Stopher and Meyburg [1975] for elaboration on
traditional methods of trip generation analysis).
Cross-classification models group individual households together
according to common socioeconomic characteristics (auto ownership
level, income, household size, etc.) to create relatively
homogeneous groups. Average trip production rates are then
computed for each group from observed data. Cross-classification
analysis similarly can be performed for trip attraction
calculations. Classification is generally by land use or
employment (e.g., manufacturing, retail, office; number of
employees per acre).
Among the advantages of cross-classification are that it is simple
to apply, it captures correlations among the independent variables
well, and it imposes no a priori assumptions about functional
relationships among the variables. But the method also has a
number of drawbacks:
- in typical applications within-cell variances are ignored,
even though the vast majority of variation arises within
rather than between cells;
- large numbers of categories or dimensions lead quickly to
empty or sparsely populated cells;
- because of this sample size problem, it typically is necessary
to minimize the number of cells either by limiting the number
of variables or by aggregating the values for each variable
into a few ranges;
- also because of the sample size problem, confidence intervals
on cell mean values may exhibit wide variation among cells;
- the method is sensitive to the grouping applied in defining
ranges for each variable;
- when the dependent variable is a zonal average, the cross-
classification method is
P93168-14/SA 10
DKS Associates
sensitive to the zone system used; and
- it is particularly difficult to account for land use and
accessibility factors in a cross classification methodology,
both because the number of cells quickly becomes too large and
because these variables are particularly difficult to divide
into meaningful ranges.
Nevertheless, cross-classification is the most common method in
current practice, and is a reliable method when a small number of
variables is thought to be sufficient for a good trip generation
model
Regression Models were once a common technique for trip generation
analysis, though today they are used less frequently than cross-
classification. Linear regression models are the most common; they
are simple and inexpensive to estimate from data typically
available to MPOS. However, the imposition of linearity introduces
a number of problems in modeling. For example, most surveys have
shown that trip-making is not linearly related to auto ownership,
but increases dramatically with the first car and to a declining
extent as the number of cars increases. The use of a linear form
in such circumstances represents a basic misspecification.
Transforms of variables (e.g., exponential forms, Box-Cox
transforms, Box-Tukey transforms) provide a way of overcoming some
of these difficulties while retaining the use of linear regression
estimation software (see, e.g., Gaudry and Wills [1978]).
Nonlinear regression techniques allow more modeling flexibility but
are less frequently available in basic statistical software
packages and hence are less commonly applied. Nevertheless, these
techniques are finding their way into use in some regional
agencies' practice, primarily because nonlinear models allow both a
high degree of flexibility in functional form (much like cross--
classification) and a large number of explanatory variables (e.g.,
Cambridge Systematics [1980]).
Examples: Home-Based Trip Generation models represent the
propensity of households to make trips, as functions of
socioeconomic and, sometimes, locational characteristics.
Models have been developed for all household tripmaking together
(person trips by all modes), for all household trips by motorized
modes (person trips by vehicle), for all household vehicle trips,
for the minimal breakdown of person trips by purpose that was
mentioned earlier (homebased work, home-based other, and non-home-
based), and for a variety of more detailed trip purpose
disaggregations dictated by local analysis needs.
Models have been developed using variables such as:
- Household size
- Number of workers
- Household income
- Auto ownership
- Number of licensed drivers
P93168-14/SA 11
DKS Associates
- Number of household members under rive years old
- Number of household members over five but under 16
- Age of head of household
- Occupation of head of household
- Occupations of other workers
- Marital status
- Housing type
- Home ownership status
- Length of residence
- Distance from the central business district (CBD)
Note that characteristics of the transportation system are not
included in the list of variables. This reflects an implicit
assumption that transportation level of service is not an important
influence on trip rates. As discussed above, this assumption is
much more likely to be warranted in the case of total person trips,
where the choice between motorized and non-motorized modes is not
subsumed in the trip generation model.
There is extensive literature on both cross-classification and
regression approaches to trip generation (see, e.g., the references
cited at the beginning of this section). Tables 1 through 4 and
Figures 1 through 4 show simple cross-classification schemes of
trip generation by motorized means , spanning nearly three decades
and four different metropolitan areas (including the Puget Sound).
It is notable that these tables show strong similarities, despite
the differences in data sources, treatments, and interpretations.
Such consistency has led many to argue that cross-classification is
a robust, perhaps transferrable technique, and has led to
widespread acceptance of the method. However, as the gradual trend
toward higher cell values suggests, other factors may be at work as
well, particularly the effects of changes in highway accessibility
and/or land use patterns.
Regression models are less and less used in basic practice,
primarily because in their simple functional forms they are more
likely than cross-classification to introduce errors into
forecasts. It also is true that the development of more credible
regression equations (e.g., with polynomial terms to capture non-
linearities and cross terms to capture correlations) requires a
substantially higher knowledge of statistics than is the case for a
cross-classification model that implicitly reflects the same
features. Nevertheless, many regression models remain in use for
basic practice, and for this reason it is worth discussing their
general characteristics.
Two basic types of equations have been used: one estimated on zonal
averages for independent and dependent variables, and one estimated
on the values of variables for a sample of individual households.
When data are averaged at the zonal level, as much as 80 percent of
the sample
1. The source for Madison and Miami is FHWA, 1975; the source for
the Bay Area is a tabulation of the 1981 home interview
survey; the source for Puget Sound is a tabulation of the
1985-88 home interview survey.
P93168-14/SA 12
DKS Associates
Table 1
Sample Cross-classification
Madison, Wisconsin
Total Household Person Trips (Motorized Modes
Cars Owned
Family Size 0 1 2+
1 1.0 2.7 4.4
2 1.5 5.1 7.0
3 3.1 7.2 9.4
4 3.2 8.0 11.7
5 5.2 9.2 13.4
Table 2
Sample Cross-classification
Miami, Florida
Total Household Person Trips (Motorized Modes
Cars Owned
Family Size 0 1 2+
1 1.0 2.9 5.6
2 1.9 4.5 7.0
3 2.9 6.2 9.4
4 4.1 8.5 11.7
5 5.8 10.2 13.4
P93168-14/SA 13
DKS Associates
Table 3
Sample Cross-classification
San Francisco Bay Area
Total Household Person Trips (Motorized Modes
Cars Owned
Family Size 0 1 2+
1 1.1 3.2 6.2
2 2.1 5.3 7.9
3 3.2 7.4 9.6
4 4.3 8.9 11.9
5 5.9 11.3 14.1
Table 4
Sample Cross-classification
Puget Sound Region
Total Household Person Trips (Motorized Modes
Cars Owned
Family Size 0 1 2+
1 2.2 2.7 2.8
2 4.7 4.9 5.4
3 4.9 6.7 7.6
4 4.5 8.8 9.8
5+ 5.8 11.4 12.3
P93168-14SA 14
Click HERE for graphic.
15
Click HERE for graphic.
16
Click HERE for graphic.
17
Click HERE for graphic.
18
DKS Associates
variability (in, say, trips per household) is removed in the
averaging process. The resulting regression equations give the
superficial appearance of fitting the data much better, simply
because the regression goodness-of-fit statistics seem higher. Two
models estimated on the same data, drawn from McCarthy [1969],
illustrate this well:
A zonal average model for home-based trips per household:
Ti = -1.09 + 1.66HHSi - 1.83CH5i + 1.44Ai + 0.611i=.61li R2=.61
(1)
where: Ti = the average number of person trips per
household in zone i
HHSi = the average household size in zone i
CHSi = the average number of children under 5 in
zone i
Ai = the average number of cars per household
in zone i
Ii = the average number of workers per
household in zone i
and a household model for home-based trips:
Th = -1.42 + 1.46HHSh - 1.65CH5h + 1.69Ah + 0.75lh R2=.38
(2)
where: Th = the number of person trips for household h
HHSh = the size of household h
CHSh = the number of children under 5 in household h
Ah = the number of cars available to household h
Ih = the number of workers in household h
Even though equation 1 appears to provide a better fit, there is no
doubt from a theoretical viewpoint that the coefficients of
equation 2 are more likely to be "correct", in the sense that
equation 2 is based on the full scope of individual behavior while
equation 1 reflects an averaging process that is highly dependent
on a specific zone structure. If one employs a regression
equation, then, the estimation methodology always should use
disaggregated data even though the forecasting procedure likely
will use zonal averages.
The cross-classification trip generation examples (Tables 1 though
4 shown previously) reveal clear non-linearities for each variable
(as values of the other variable are held constant). While some of
these non-linearities could be accounted for by additional
variables (as in eqs. 1 and 2 above), there is no real reason to
expect the effects of any of the variables to be linear throughout
the range of interest. Recognizing this, practitioners have
developed a strong preference for cross-classification methods in
basic applications. Again, however, it is not so easy to apply
cross-classification in advanced practice, because of the geometric
increase in the number of cells as variables are added.
P93168-14/SA 19
DKS Associates
Examples: Non-Residential Trip Generation models generally serve
two purposes: to estimate the number of attractions of home-based
trips (i.e., at the non-home end); and to estimate the number of
productions and attractions of trips which are non- home-based.
(Commercial trips are often estimated separately.) Non-residential
trip generation rates also are widely used to determine the traffic
consequences of development proposals for specific sites. In
general, this latter application requires the estimation of an
average trip generation rate, per unit area, for uses of different
types.
Non-residential person trips may be further categorized as employee
work trips, other employee trips, and visitor (or "other) trips.
Work trip rates are closely related to the number of employees per
unit area for a particular use; visitor trip rates (including
customers, clients, etc.) vary considerably with the land use.
Non- home-based trips from nonresidential land uses will largely be
trips made by workers traveling to other nonresidential locations
(especially work-related business, but also including lunch hour
trips to restaurants and shopping). Other non-home-based trips are
the result of trip chaining (e.g., stopping at a gas station and
then proceeding to work produces a home-based trip followed by a
non-home based trip; a shopping excursion to several stores may
produce a lengthy chain of non-home based trips.) Hence these trips
tend to be related in part to the number of employees, and in part
to the type of land use.
Non-residential trip generation has received considerably less
attention than has home-based trip generation, and the techniques
that have been used are generally less sophisticated. Typically,
simple cross-classification or regression schemes are used to
relate nonresidential trip making to various attributes of the land
uses from which those trips are produced, or to which they are
attracted.
In cross-classification applications, the most commonly used
classifications are for land use (offices of various types,
industrial, retail, medical, education, etc.). Trip generation
rates are expressed per unit area or size (acres, square footage,
employees, etc.). The rates are typically derived from data
aggregated over the entire region, although in some cases separate
rates are calculated for a typology of areas, e.g., CBD, inner
suburb, outer suburb, rural. A further breakdown into peak and
off-peak periods is commonly used for site studies.
Regression equations also have been developed and usually are
extremely simple - one or two variables are common. These
equations are not necessarily more rigorous than the simple cross-
classification schemes, in part because the variables used as
correlates of the number of trips made are highly correlated with
each other.
Freight trips are sometimes handled as part of the non-residential
trip generation analysis, but few transportation modeling efforts
have included much analysis of freight trips (except as an
afterthought). A typical approach is to express truck trips as a
percentage of person trips or vehicle trips based on counts. A few
areas have occasionally developed freight O-D matrices which they
update using VMT, employment, or population growth factors. While
it is recognized that such methods impose strong assumptions to the
effect that truck travel will maintain a
P93168-14/SA 20
DKS Associates
constant relationship to overall travel (or to population or
employment growth) over the forecast period, this simplification
has been accepted on the grounds that truck traffic is only 5-10
percent of overall trip-making. Obviously, local variations e)dst
and may be significant
Advanced Practice
Advanced practice in trip generation begins with the inclusion of a
wider range of socio-economic variables in cross-classification
models and extends to the estimation of separate models for a wider
range of personal trip purposes, e.g, home-work, home-school, home-
shop, home-other, non-home-based work-related, non-home-based
other. In addition, greater attention may be given to time of
travel, e.g., peak, off-peak, midday.
While virtually all trip generation models in basic practice deal
with vehicle trips only (either as vehicle trips per se, or as
person trips by vehicle), some advanced approaches estimate person
trips by all modes (including walking and biking.) However, a
focus on 'true' person trip generation implies greater complexity
in trip distribution and mode choice (in order to distinguish
vehicle trips from person trips in a spatial context). For this
reason, many modelers continue to work with vehicle trip generation
models, even though vehicle trip generation may be harder to
estimate effectively (because they are sensitive to a far wider
range of factors such as land use densities).
A critical issue in advanced practice is whether trip generation
models should include land use and transportation variables.
Accessibility, defined as the proximity of desirable activities, is
a function of both the distribution of relevant land uses and ease
of access to each location. Whether improved accessibility leads
to more trips, or primarily to longer trips and to modal shifts, is
the crux of the issue. if the influence is primarily on trip length
and mode, then accessibility effects are logically treated in trip
distribution and mode choice rather than in trip generation. If
there also is an effect on the number of trips, then accessibility
may belong in all three model stages. It is clear, however, that
the practice of purging non-motorized trips prior to trip
generation has the effect of complicating trip generation modeling
by adding elements of trip distribution and mode choice to the
phenomena subsumed in the trip generation model.
Recent work with high quality survey data has somewhat clarified
the role of accessibility in travel choices (Harvey [1994]). When
the survey captures most of the substantive walk trips, and walk
trips are fully included in the analysis, one finds a significant
role for accessibility in explaining the choice of the walk mode
and the choice of nearby destinations, but only a minor influence
of accessibility on the actual number of trips. Whether the muted
effect on trip generation should be taken at face value, or
reflects a more complex process of greater activity participation
offset by an ability to carry out more than one activity on a
single trip, is not yet known.
The practical lesson is clear, however. "True" person trip
production is a relatively simple process dependent on a few
demographic factors, with land use and transportation playing only
a secondary role. The production of person trips by vehicle is an
entirely different matter. People are more likely to walk when
walking is convenient and desirable land uses are close by. People
P93168-14/SA 21
DKS Associates
are more likely to drive or take transit when highway and transit
accessibility is improved and desirable land uses are not close by.
Thus, a model of person trips by vehicle must account for proximate
land uses and for accessibility by motorized and non-motorized
modes in order to provide a reasonable explanation of trip rate
variations among zones.
Given the strength of this statement, it is puzzling that so few
agencies have adopted a "true person trip approach to trip
generation. A simplistic explanation might refer to the
traditional focus on vehicle trips in transportation investment
analyses and the inertia accompanying any well-codified, long-
established procedure. One might even argue that the above
assertions are not quite true, since the frequent ploy of using
different trip generation models for broad categories of land use
(e.g., urban, suburban, rural) is at least a crude reflection of
accessibility effects. But the fact is that there are a number of
reasons for analysts to move cautiously in converting from a
vehicle trip approach to a 'true' person trip approach. Among
these are:
- it may be easier to build a complex trip generation model than
to redo the full model system to accommodate non-motorized
trips.
- many surveys do not include a full representation of non-
motorized trips, particularly for non-work purposes. This is
sometimes explicit, e.g., when surveyors were instructed to
focus on vehicle trips, but is always at least something of a
problem because walk and bike trips are easiest for the
respondent to forget.
- network-based level-of-service measurements are valid only
over distances longer than, say, twice the typical zone
dimension. Thus, as a practical matter, it is difficult to
obtain accurate estimates of walk and bike times without
extensive analysis of point-by-point accessibility in a GIS
framework. In other words, most agencies do not have on-hand
the information needed to accurately model non- motorized
trips in trip distribution and mode choice.
Since the development of a complex trip generation model is no easy
task either (as demonstrated by the Bay Area example to follow),
most agencies now rind themselves in the uncomfortable position of
knowing that traditional trip generation models fall short in terms
of policy relevance and perhaps, theoretical validity), yet lacking
the resources especially data) for timely development of a better
approach.
The following examples illustrate how two MPOs at the forefront of
advanced practice have addressed this problem.
Example: Portland (OR) Trip Generation. Portland METRO has made a
decision to seek the best possible treatment of non- motorized
trips, both because land use issues are prominent on their policy
agenda and because in-house research has shown that a more
theoretically correct treatment of non-motorized travel can improve
the overall performance of their model system. Portland's trip
generation models predict "true" person trip origins and
destinations for each of
P93168-14/SA 22
DKS Associates
six trip purposes:
- Home-based work
- Home-based school
- Home-based college
- Home-based other
- Non-home-based work-related
- Non-home-based non-work-related
In general, two models exist for each trip purpose: one to predict
trip ends at home (for home-based trips) or at work (for non-home-
based work-related trips), and one to predict the other end of
these trips. The models depend on a set of demographic
characteristics that are projected for each zone:
- Household size (4 categories)
- Household income (4 categories)
- Age of the 'head" of household (4 categories)
- Workers per household (4 categories)
- Vehicles per household (4 categories)
- Children per household (4 categories)
and on a set of attraction factors characterizing non-home
locations:
- Total employment
- Retail employment
- Total households
- Students and employees at colleges
The METRO trip generation procedures make extensive use of
household characteristics, but they do not reflect land use
indicators such as housing type (single family vs. multiple
family), residential density, mix of uses, or density of trip
attractions. METRO's evidence suggests that these factors strongly
influence the number of person trips by vehicle, but have an
important effect on total person trips only in the extreme.
Similarly, METRO's trip generation procedures ignore the
accessibility of a zone to various desired activities.
Conceptually, one might expect an individual to participate in more
out-of-home activities as a larger number of suitable opportunities
come within easy travel range of the home (or work) zone. In
practice, this effect appears to be secondary.
METRO's work does point to one possible effect of accessibility on
trip generation. An earlier model step represents auto ownership
as a function of transit accessibility at the home zone (among
other factors). Since auto ownership is a factor in some elements
of person trip generation, it follows that changes in transit
accessibility can have an effect on person trip generation. Still,
the effect would be much more important in a vehicle trip
generation model,
P93168-14/SA 23
DKS Associates
where auto ownership is by some accounts the single most important
variable.
Metro's treatment of person trips extends into trip distribution
and mode choice. Trip distribution uses a traditional gravity form
and a single highway level-of-service variable (but with logit
estimation of the friction factors). While more explicit treatment
of modal accessibility would have been preferred, these simplified
distribution models appear to perform adequately as placeholders
until a better modeling database can be developed.
The non-motorized component of mode choice is treated as a binary
"pre-mode" choice (i.e., motorized vs. non-motorized), based only
on trip distance and household auto ownership. Again, these
simplified models appear to perform adequately as placeholders
until a better modeling database can be developed.
Basically, METRO has elected to cope with the difficulties of a
'true" person trip model rather than attempt to deal with non-
motorized trips through a more complex trip generation model The
problem of survey under-representation was addressed through a
reweighting of observed walk and bike trips. Similarly, the
problem of LOS inaccuracies for short trips was addressed through a
review of time and distance assumptions for intrazonal travel. The
decision to proceed in this way was based as much on political as
on technical considerations - it is easier to explain to non-
technical users of model information that incremental improvement
is possible within an existing structure than that a different
structure would be more appropriate.
Example: Bay Area Home-Based Shopping Trip Frequency. The
Metropolitan Transportation Commission (MTC) of the San Francisco
Bay Area followed the opposite path in dealing with non-motorized
trips. For non-work trip purposes in its innovative 1978 model
system, MTC adopted elaborate non-linear regression models with
both land use and accessibility components.
For example, the MTC shopping trip frequency model is a non-linear
regression yielding an inverse function of household
characteristics, home zone characteristics, and aggregate
destination attractiveness (as embodied in the expected utility for
shopping destination/mode choice). The exact model specification
is:
Click HERE for graphic.
(3)
where:
hbshopi is the number of daily home-based shopping trips by
household i (person trips by vehicle);
hhsizei is the number of persons in household i; inci is the
income of household i ($);
E[Uidm] is the expected utility from the shopping destination
/mode choice model for household i, defined as the
natural log of the denominator of that model's logit
equation;
P93168-14/SA 24
DKS Associates
edeni is the service and retail employment density in
household i's home zone, expressed in workers per gross
acre.
The inverse exponential form of this function makes it somewhat
difficult to interpret. Basically, household shopping trips
increase with household size, income, and accessibility to shopping
opportunities, and decrease with rising local density of retail
opportunities. The latter relation emerges because this is a
vehicle trip generation model, and residents are less likely to
make a shopping trip by vehicle if there are plentiful shopping
opportunities within walking distance. In a fully person trip
model (including walk), the density variable would have a
negligible or slightly positive effect on shopping trip generation.
Recent Developments
Although the emphasis of this paper is on ways of improving the
analysis of trip generation in a conventional modeling framework,
there is some value in looking farther ahead at the direction of
future developments. The emerging paradigm, as suggested by a wide
range of academic research over the past 15 years, focuses on the
program of activities undertaken by each member of each household
in a region. While specific analytical approaches have varied, the
consistent goal of this work has been to predict the sequence of
activity (and related transportation) choices over the course of a
typical day or week.
Imagine, for simplicity, a household consisting of a single person.
The conditions of weekday travel for such a person will depend on:
1) whether ' the person is currently employed; 2) where the person
is employed; 3) the person's typical work hours; and 4) the
likelihood a person will go to the workplace on a given day. Each
of these can be modeled as a function of personal, land use, and
transportation system characteristics. On days when a work trip is
made, the overall travel pattern will be organized around the work
trip. Modeling might consist of a representation of the likelihood
that the worker will engage in additional activities before leaving
for work, on the way to work, during the work day, on the way home
from work, and after returning from work. Each of these also could
be modeled as a function of personal, land use, and transportation
system characteristics.
A set of models would be developed for each characteristic
household type, where number of workers, presence of children,
marital status, and other life-cycle factors might be used to
distinguish among household types. Another set of models would be
used to represent the dynamics of household formation and change
over time.
Forecasts would be made for a sample of households in the analysis
region. Such a sample might be based on an existing home interview
survey, or it could derive from a source such as the Census Public
Use Microdata Sample (with additional statistical procedures to
'synthesize' spatial detail). It has been suggested that a Census-
based procedure could aim to reproduce the full regional
population, in which case the travel "model" would provide a
complete representation of the regional travel pattern.
P93168-14/SA 25
DKS Associates
Forecasting for a specific future scenario would involve aging each
household in the sample from the present to the forecast year(s),
and then calculating the likely household travel pattern in the
forecast year(s).
While these procedures represent a large departure from current
practice, they are neither as innovative nor as demanding of
resources as they may seem. Household-level simulation (sample
enumeration) has been used by economists since the early 1960s for
a variety of policy studies. Similarly, one version of the MTC
travel model developed in the late 1970s is a household simulation
("STEP"). The enumeration approach is in fact the only practical
way to use all of the feedback features of the MTC model, and was
the method used to satisfy court requirements in MTCs lawsuit over
TIP conformity.
Moving to a population rather than a sample-based approach would
involve a computer-intensive style of analysis, but hardware and
software experts who have looked at the issue believe that
presently available workstations are sufficient for such analyses
in a typical metropolitan area. Furthermore, advances in hardware
and software (faster microprocessors, multiple processor
configurations, parallel multi-threaded execution, etc.) are
expected to make it possible not only to simulate the entire
population's travel behavior but to maintain a simultaneous dynamic
microsimulation of highway and transit operations.
While there may be something of a consensus about the long term
direction of regional travel modeling, there has been no clear
enunciation of the path by which agencies will be able to move from
the current generation of models to a new modeling paradigm. One
possibility, suggested by Kitamura and others, is that a household-
based simulation initially could be "grafted onto" the front end of
a traditional travel mode to provide forecasts of demographic
(e.g., "life-cycle") changes. As experience and knowledge were
gained, such a model could be expanded to account for household
(and perhaps worker) location behavior, then to provide predictions
of gross household activity participation (i.e., to substitute for
traditional trip generation analysis), and finally to provide
predictions of daily household activity patterns (i.e., to
substitute for the entire traditional model system).
It probably is not realistic to expect a single MPO to fund the
development and testing of such a model. However, money from
Federal sources, and possibly from a number of private foundations,
will be available to MPOs willing to venture into this territory.
P93168-14/SA 26
3. General Strategy for the Regional Council's
Trip Generation Model
DKS Associates
3. General Strategy for the Regional Council's Trip Generation
Model
Implementation Issues
The literature review in the previous section does not offer us an
unambiguously "best" way to proceed with a trip generation model
development for Regional Council. The most important points from
the literature are:
- Cross-classification is the currently-preferred method of trip
production analysis in a traditional modeling framework.
However, under typical sample size constraints (e.g., 5000
households or less), cross-classification is most effective
when only two or possibly three dimensions are sufficient to
capture the major differences in household behavior.
- Representing all person trips in the trip production analysis,
rather than the more common approach of representing only
trips in motorized modes, greatly reduces the number of
variables necessary for a good model. This is because
congestion, land use densities, and other determinants of
accessibility play an important role in the decision to use a
motorized mode but appear to have only a modest effect on
overall activity participation. Cross-classification nearly
always is a feasible strategy for a model of all person trips,
but a model of motorized mode person trip production may
require a more complex and less tractable form such as a non-
linear regression.
- Representing all person trips in trip production analysis
necessitates changes in trip distribution and mode choice
modeling, to reflect the decision on whether to use a
motorized vs. a non-motorized mode. To some degree, this
increases the complexity of mode choice analysis and requires
a more extensive set of changes to the model system, but it
also places the modeling of non-motorized choice at a more
logical point in the model hierarchy and makes the model
system easier to explain to a lay audience.
- One difficulty in carrying non-motorized trips through trip
distribution and mode choice is that both models depend
heavily on travel time and distance estimates, yet time and
distance values are notoriously inaccurate for short trips in
a traditional network analysis. This issue does not arise in a
conventional model system because most short trips (i.e., walk
trips) are excluded at the outset. While there is no easy fix
without a change in the basic paradigm, reasonable results can
be achieved through detailed review of intrazonal and
centroid-to-network access times and careful model
development.
- Finally, even for a trip production model covering all person
trips, household activity simulation holds the potential for
greatly improved conceptual validity and policy sensitivity.
However, while household simulation is now a practical method
for applying traditional models and perhaps for generating
socioeconomic forecasts, full activity
P93168-14/SA 27
DKS Associates
simulations will require additional research. Individual MPOs may
well choose to underwrite such research, but only with the
recognition that a fully practical application may take some time
to develop.
In thinking about how all of this applies to the Regional Council,
it is important to be mindful of the context of this trip
generation model development project For instance:
- Regional Council is operating under a tight time constraint.
An upgraded trip generation model is needed as soon as
possible for conformity and other analyses.
- The existing regional model system, within which any upgraded
trip generation model must fit, represents person trips by
motorized modes. Initially, an improved trip generation model
would have to produce trip estimates compatible with the
existing model system. However, updates to the trip
distribution and mode choice models are planned in the next
year or two, so that the trip generation step, and indeed the
whole model system, could be modified within that time frame
to represent all person trips.
- Regional Council has good data resources for model
development, with a range of recent cross-sectional and panel
surveys. Multi-day trip diaries are available in both cases.
It should be possible to build a large trip generation data
set from these surveys, which (among other things) might
support a relatively complex cross-classification scheme.
Trip Generation Model Development
Any model improvement strategy should focus both on making
immediate adjustments to enhance model validity and credibility and
on ensuring a smooth path to more extensive but longer-range and
more fundamental changes. While it may be an obvious statement, it
is important to note that virtually any strategy which uses recent
data to produce new models will do some good. Beyond this, there
are several ways to proceed with trip production model development.
In our view any strategy should have at least the following
characteristics:
- It would deal explicitly with non-motorized trips.
- It would use cross-classification as its basic methodology.
- Analysis of variance (ANOVA) methods for unbalanced data
should be applied rigorously in the trip production analysis.
- It would aim for a partitioning of trip types that preserves
as much information about the activity pattern as possible.
For the majority of home-based trips, this would involve a
distinction among work, college, school, shopping, and other
personal travel. For non-home-based (and some home-based)
trips, this would involve a closer relationship with the home-
based trips that anchor the overall trip chain, e.g., work-
work, work-other,
P93168-14/SA 28
DKS Associates
and other-home trips modeled in terms of the basic work trip.
- Despite an innovative partitioning of trip types, it would
yield predictions for a set of trip purposes that could be
recombined to match the purposes in the remainder of the model
system.
- It would use a basic core of household life cycle and economic
variables to build the person trip cross-classification
models. Household size, number of workers, and income would
always be considered, with the addition of other variables
suited to each purpose. Experience suggests that "first-
order' effects, as identified through ANOVA methods, would be
limited to a small number of variables in each case.
- It would incorporate an additional post-generation, pre-
distribution model to split person trips among motorized and
non-motorized modes. This could be achieved through an
additional cross-classification dimension based on
accessibility or density, but a simple set of percentages by
classification that separates motorized from non-motorized
trips is an acceptable interim solution since the Regional
Council intends to update its trip distribution and mode
choice models in the near future.
- It would begin a longer-range process of developing a more
explicitly activity- based treatment of travel behavior.
The structure of the trip production model was determined by
developing and testing a number of alternative cross-classification
schemes, as described in Section 5. Before the trip production
analysis could be conducted however, the travel survey data needed
to be carefully evaluated and systematically screened to eliminate
trips and households associated with missing, invalid or
inconsistent household and/or travel data. The evaluation of the
survey data is described in the following Section 4.
P93168-14/SA 29
4. Adoption of Local Survey Data for
Trip Generation Analysis
DKS Associates
4. Adaptation of Local Survey Data for Trip Generation Analysis
Overview
This project's key resources are two travel diary surveys, 1) a
"cross-sectional" survey of about 4500 households in the Puget
Sound region, and 2) multiple-year observations of the Puget Sound
Transportation Panel, consisting of about 1600 households.
The cross-sectional survey is actually six surveys of different
areas in the Puget Sound region, each observed in different times
from 1985 to 1988. One day of travel per household was recorded.
Household Travel Surveys, 1985-1988, Puget Sound Region, available
from PSRC, documents and summarizes results from this survey.
The Puget Sound Transportation Panel was begun in 1989 to study
demographic and activity change by individuals through time "waves"
of demographic and travel diary observation have been performed in
1989, 1990, and 1992, each including two days of travel recording.
Information and documentation of this survey are also available
from PSRC.
The "cross-sectional" and the "panel" surveys each provide three
main data tables described as follows:
- a Household table, containing information pertaining to the
whole household, such as the total number of persons
(including those not providing diaries), the number of autos
owned, household income, and location (census tract).
- a Person table, containing information particular to each
diary participant in the household. Depending on the survey,
this includes some or all of the following
information: employment status, occupation, sex, possession of
a driver's license, and age.
- a Travel table, a diary of travel by all household members age
5 and up for the crosssectional surveys, and age 15 and up for
the panel surveys. This includes the purpose of each trip,
the time of beginning and arrival, the locations of origin and
destination (census tracts), the mode of travel, and number of
occupants.
There are differences in the details of the information contained
in various areas of the crosssectional surveys as well as
differences between the information for the panel survey versus the
cross-sectional survey. Each table has identifiers that link
travel records to the respective persons, and persons to the
respective households.
Practical Issues
As noted above, a key definitional difference between the two
surveys is that the cross-sectional survey collected travel diaries
of persons age 5 and older, while the panel survey collected
diaries from persons age 15 and older.
P93168-14/SA 30
DKS Associates
Due to the need for a timely analysis, and because of definitional
and analytical issues that make it difficult to simply combine the
two surveys, it was decided to perform broad exploration of
alternative models using the larger cross-sectional survey only.
Data from the Puget Sound Transportation Panel were only examined
in terms of the selected models, to identify differences between
the cross-sectional and panel surveys.
Validation of the Cross-Sectional Survey Database
Travel surveys are large, concentrated data-collection efforts
involving a large number of people, including respondents,
interviewers, and data processors. It is difficult to ensure that
all persons involved will follow instructions perfectly, and record
or enter all information accurately, or cooperate fully, so
compiled tables of travel surveys generally contain some
incomplete, unclear, conflicting, or highly unlikely observations.
For trip generation analysis, observations with doubt of the
household's composition (by which it is classified), or on its
members' trips, including trip purpose, or whether any trips are
missing or wrongly inserted, may contribute biases to trip
generation models.
Given this situation, there are many courses of action that may be
taken, including the following:
1) Scan for observations meeting certain criteria that can be
applied systematically by computer programs, drop the "bad"
observations thus found, and accept the remainder of the
observations despite the possibility that observations with
more subtle problems remain.
2) Systematically scan for "bad" observations, and attempt to
correct as many as possible, using the original data sheets
and judgment.
3) Have trained analysts examine the composition and travel of
each individual household, making judgments of their
likelihood, and attempting to correct as many as possible.
Option (1) was chosen because of the need for a timely yet accurate
analysis. The demands of timeliness of this study preclude any
effort to correct the rejected observations, as with options (2)
and (3). Such an effort requires case-by-case review and the
judgment of skilled analysts, and cannot be automated in the same
way that the problematic observations were detected. As discussed
below, the resulting "refined" database is still very adequate for
the purposes of developing a trip generation model.
Table 5 presents the specific validation checks developed and
chosen for this analysis. Also shown are the numbers of households
found in each category of problem. Note that some problems may
occur more than once in a household, and that some households are
associated with more than one problem. The sum of the numbers of
problem households is 1477, but the actual number of unique
rejected households is 1109. The analysis that followed used 3507
out of the original 4616 households, or 76 percent. This is a very
good validation rate compared to many other surveys in the U.S.
that have been similarly checked.
P93168-14/SA 31
DKS Associates
Table 5
Cross-Sectional Travel Survey Validations
Category/Specific Criteria Number of
Households
Number of Missing or inconsistent household data 286
Number employed more than household size (King
County)
Number employed is missing (King County)
Household size is zero or missing
Income is missing or not valid (i.e. not in the dictionary)
Number of vehicles available is missing
Number of drivers is missing
Number of drivers more than household size
Persons to household problems 91
More persons counted than household size
Number employed more than count of persons age 15 or older
Number of drivers more than count of persons age 15 or older
Households with no members age 18 or older 4
Missing or inconsistent person data 85
Age is missing
Age < 5 (usually = 0)
Age > 16 and employment status is blank or not in dictionary
(for non-King Co.)
Age < 16 and employed full- or part-time
Employment status "Refused"
Household with no or invalid persons 2
No person records found for household
Person number is not a valid digit
Trip against household problems 144
None employed but a work trip reported (King Co.)
No drivers but drove
No vehicles but drove
P93168-14/SA 32
DKS Associates
Table 5 (continued)
Cross-Sectional Travel Survey Validations
Category/Specific Criteria Number of
Households
Trip against person problems 189
Age < 15 (age known) and drove
Age < 15 (age known) and made work trip (It is acknowledged
that the above condition might not always be an error;
fortunately only 25 households meet this condition.)
Age < 16 (age known) and made college trip
Not employed or student, but made work trip (non-King Co.)
Trips with invalid date or time 154
Person made a trip, but any of these occur:
Begin time = 0 and end time > 500, i.e. begin time is missing
Begin time < 2000 and end time = 0, i.e. end time is missing
Day code missing or invalid
Invalid begin or end time, like 1480
Trip end time before begin time 38
Time at stop problems 334
Rectified end date and time before begin date and time
Over 8 hours shopping
Over 12 hours at school
Over 16 hours at college
Arrived at either of these activities before midnight at end
of day, and not departed at end of day.
Trips duplicated 3
Trip number used for a person more than once
Trip mode problems 147
Trip mode is Car, Vanpool, Carpool, or Motorcycle (01, 02, 03,
09) AND either:
Drive-Ride indicator (D_R) not valid, or
NUM of occupants invalid (except for Kitsap County Carpool and
Motorcycle trips, for which this field is systematically zero
for many cases)
P93168-14/SA 33
DKS Associates
A carefully screened database with 3507 households provides an
adequate survey size for development of the trip generation model.
Estimation of a sophisticated mode choice model, however, would
benefit from a larger sample size. Therefore, the Regional Council
is rightly considering option (2) or (3) above prior to development
of a new mode choice model. While quite rigorous, by no means have
all problems been detected. Most additional checks that could be
proposed require particularly difficult computer programming, a
high dependence on judgment, and/or additional data.
Trip Purposes
Trip generation in standard travel models deals with trip ends,
which are either productions (the end at the home or the demand
side of the trip) or attractions (at the location of the out-of-
home activity). The production end of a trip can be either the
origin or the destination of the trip; similarly, the attraction
end can be the origin or destination of the trip. Most models deal
with trip purposes related to the activities of both the production
and the attraction ends of the trip. Simple travel models
partition their trips between home-based work, home-based non-work,
and non-home-based trips. More sophisticated models further
partition these purposes. Some general goals for a trip purpose
system are:
- The trip purpose is identifiable from the travel survey.
- The trip purposes can be associated with the types of
employment and/or land use available as zonal data.
- Each trip purpose should contain similar activities, where a
given need or desire can be satisfied in one trip purpose.
- Productions and attractions are logically paired. A trip
cannot go from a production to another production, or an
attraction to another attraction. (But some or all non-home-
based trip ends cannot be clearly labeled as productions or
attractions.)
- The trip purposes should relate logically to the model system.
That is, they should be appropriate for trip distribution,
mode choice, and time-of-day choice model estimation and
application. For this project, additional trip purposes will
be added to those currently used in the Regional Council
model, however, the purposes will be combinable into the
current model system's trip purposes.
The travel surveys encode the activity at the destination of each
trip as the "trip purpose." It was necessary to derive each trip's
"modeling trip purpose from its coded activities of both the origin
and destination ends. This is deduced from the preceding trip's
destination activity as well as each trip's own destination
activity. Table 6 shows the activity purposes coded in the survey,
and Table 7 shows how the activity purposes at the ends of a trip
are paired to identify its production attraction orientation and
its model trip purpose.
P93168-14/SA 34
DKS Associates
Some notes and issues on the proposed trip purposes:
- The first trip's activity of origin is not specified, so it
was presumed to be home. If a person's first trip is to home,
then its activity of origin is unknown; these trips are
labelled "Home-based Unknown". Their actual purpose could
include returns from late or graveyard work shifts, social
visits, etc.
- Home-based College trip production analysis should consider
proximity to a college as well as the internal characteristics
of the household. The data needed to test locational
hypotheses were not available for this paper's analysis,
however, One way to apply a model with this effect is ton
convert college trips into some combination of Home-based Work
and Home-based Other. Most other treatments require special
formulations of trip distribution.
- The production-attraction orientation of Work-Other trips was
arbitrarily chosen. This distinction preserves more
information than most models of non-home-based travel, for
which no orientation is identified. This purpose includes
much work-related travel, personal trips to and from work, and
part of the chain of trips that stop along the way to or from
work.
- Other-Other troops have no particulAr production-attraction
orientation.
General Procedure for Preparing Survey Data for Analysis
Preparation of the survey database for home-based trip production
analysis involved these treatments:
1) If any validation problem was identified for a household, then
the whole household was marked for exclusion from analysis.
(If one trip is in doubt, then the total number of trips by
the household is in doubt.)
2) Using the purpose of each trip and of its predecessor, the
trip purpose And trip orientation (production to attraction or
attraction to production) were identified.
3) For each "valid" household, the trips in each trip purpose
are counted. Also, to permit cross-classification analysis,
the numbers of persons employed and in certain age ranges was
counted for each household.
Attractions and non-home-based trip generation analysis use a
different treatment of the survey data, along with existing zonal
land use and employment data. This treatment is discussed later.
P93168-14/SA 35
Click HERE for graphic.
Click HERE for graphic.
36
5. Trip Production Analysis
DKS Associates
5. Trip Production Analysis
Overview
The objective of this analysis is to identify factors most
responsible for trip making, yielding conclusions that will lead to
selection of a new trip generation model for the Regional Council's
travel model system. This analysis explores the relationships
between household demographic and economic characteristics, and the
daily numbers of trips generated in each purpose, as observed in a
substantial sample of households in the Puget Sound region. For
this analysis, these relationships are expressed as cross-
classification models, in which each household is categorized
according to some function of its demographic/economic makeup, and
the mean observed trip productions for a given trip purpose are
determined for all households in each category.
The "core" demographic/economic variables available to categorize
the households are:
- number of persons in household,
- number of employed persons in household,
- household income,
- number of automobiles available to household members,
- numbers of persons in various age ranges.
The home-based trip purposes are appropriate for analysis by this
method. It might be feasible to analyze non-home-based travel in
the same manner, but it would not be possible to apply such a model
within the trip distribution, mode choice, and traffic assignment
model system.
An important statistical technique for analyzing and comparing
alternative classification models is Analysis of Variance, often
called ANOVA. An ANOVA experiment measures how much of the
variance in trips generated in each household in the sample dataset
is "explained' by a proposed classification system. It is desired
to find a classification system that maximizes "explained"
variance. Variance is defined as the sum of the squares of the
differences between each household's observed trip productions, and
the mean trip productions of all households in some category.
Of course, not all variance will be explained by any model. Trips
in some trip purposes are made almost daily, such as work and
school; an effective model can explain a high portion of their
variance, but it cannot explain which workers went to work and
which ones did not on the survey day. On the other hand, trips
made less regularly, such as shopping, have a high degree of day-
today variance, so even the best model can explain only a small
portion of variance.
More-complex classification systems will, in general, explain more
variance than simpler ones, but the apparent explanatory power of a
complex model can be misleading:
P93169-14/SA 37
DKS Associates
- Partitioning a given classification system into more
classifications will always "explain" more variance, even if
the additional classifications are made on an irrelevant or
random variable. The F-statistic test is used to determine
whether an added variable (or relaxation of a restriction)
significantly added to the explained variance, or that it had
a significant probability of being due to chance alone.
- More dimensions in a cross-classification system decrease the
sample sizes of households within each cell. The sample sizes
in each cell translate directly into the levels of certainty
(technically, the standard deviation of the mean, and the
confidence intervals) of the respective trip production rate.
- It is considerably more difficult to provide zonal counts or
forecasts of households in multi-dimensional cross-
classification systems than in one-way and simple two-way
systems. Most U.S. Census data, when provided at a
geographically detailed enough level for model input, only has
one-dimensional distributions by most variables, so cross-
classified zonal distributions must usually be synthesized.
The procedure used in the analysis of trip production rates for
each trip purpose involved the following steps:
- Identify the appropriate "core' household variables that
appear to be most reasonable for trip making for that trip
purpose.
- Estimate the trip production rates for each of the 'core"
variables taken individually.
- Compare the trip production rate estimates for the single
"core variables to each other using ANOVA techniques.
- For the most appropriate "core" variables, test whether the
addition of a second variable, either as a 'main effects"
model or a full cross-classification model, would improve trip
production estimation by explaining more of the variances.
These tests also require rigorous use of ANOVA techniques.
Definitions of terms used in the ANOVA analyses that follow are:
- Sample size is the number of units of the denominator of the
trip rate. For each trip purpose, the analysis begins with
trips per household, so the total sample size is the number of
households, or 3507. The home-based work analysis continues
with a study of trips per worker; the total "sample size of
workers is 4429 workers. Some of the other trip purposes
study trips per person age 5 and up; the total "sample size"
is 7951. It should be noted that where the trips are per a
unit other than households, "sample size" is labeled in quotes
because each person or worker's trips are actually household
trips spread upon each person or worker. (Such "clustering
reduced the variance from
P93168-14/SA 38
DKS Associates
what would be measured of trips by the respective persons, but
did not alter the mean trip rates.)
- Total sum of squares is a measure of household-to-household
variation in trip production, in the respective trip purpose,
among all sample households.
- Explained sum of squares is the portion of total variance that
exists between cells.
- Residual variance is the remaining portion of total variance,
not explained by the model. It is also known as within-cell
variance.
- The F-statistic is used to assess the significance of the gain
of explained variance by adding a variable or relaxing a
restriction of a simpler 'comparison" model. It equals
Difference in explained sums of squares / Added degrees of freedom
___________________________________________________________________
Comparison model's residual / Comparison model's degrees of freedom
- The significance of F is the probability that the difference
in variances is due to chance alone. It is a function of the
F-statistic and of the two models' degrees of freedom. In
many statistical contexts, a probability of more than 1 or 5
percent is considered a failure to demonstrate the
significance of the proposed relationship or variable. But
this test can not necessarily mandate inclusion of the added
variable. Some other variable, perhaps correlated to the test
variable, may be even more significant.
Analysis of Home-Based Work Productions
The number of workers in a household is an obvious determinant of
household work travel; some alternative classifications are also
examined. The estimated mean productions by classification are
shown in Table 8.
A comparison of the core variables for home-based work production
is shown in Table 9 including the F-tests of each of these
classifications, relative to no classification at all. An example
illustration of the F-test is as follows:
1) Identify the hypothesis. The first hypothesis tested is
whether the observed differences in the mean trips per
household, between the groups by the number of workers, can be
attributed to the natural variability among means possible
from an arbitrary sampling of the households, or whether
households with different numbers of workers really have
different work-trip production rates. In other words, is the
variance between samples significantly greater than the
variance within samples?
2) Compute the variance ratio F, that is, the ratio of the mean
sum of squares between groups to the mean residual sum of
squares. Here, it is
(3533/4) / ((8229-3533)/(3506-4)), or 658.7.
P93168-14/SA 39
DKS Associates
3) Determine the significance of F, the probability of obtaining
group means as different as observed when, in fact, mean work
trips per household are independent of the number of workers.
This is a function of the F-statistic and the degrees of
freedom in the numerator and the denominator. Because F Is
high, the significance of F is nil, that is, the probability
that there is no relationship is extremely small. By
comparison, the value of F at the 5% significance level for
the degrees of freedom (4 and 3502) is close to 2.37. (Many
statistical textbooks and references provide values of F,
typically at 5% and/or 1% significance levels, for various
combinations of degrees of freedom.)
P93168-14/SA 40
DKS Associates
Click HERE for graphic.
P93168-14/SA 41
DKS Associates
Click HERE for graphic.
All of these classifications show a significant positive
relationship to work travel. As might be expected, the number of
workers shows the most distinction among rates and the most
explained variance. The number of persons in an age range
associated with most employment (18-64) is a somewhat strong
indicator of work travel. Income, number of vehicles, and total
persons also show positive relationships, apparently because of
their colinearities to the number of employed persons.
Tables 10 through 15 explore additional effects upon work travel,
in combination with the number of workers. These further
explorations are normalized to the number of workers in the
household, so the production rates are trips per worker.
Normalizing in this manner makes the cell means more equal, and
reduces the tendency for households with larger numbers of workers
from dominating the measurement of additional effects. Note that
"sample size" is actually numbers of workers.
Some additional definitions must be given for two-way cross-
classified analyses in Tables 11 through 14. Each two-way cross-
classification actually provides two different models:
1) A "Main Effects" model, in which a household's estimated
production rate equals the overall mean rate plus an
adjustment due to its first variable's classification, plus an
adjustment due to its second variable's classification. These
are denoted with a plus sign, for example, "Workers +
Income."
2) A full cross-classification model, using the mean productions
in each cell. These are denoted with "x" as in "Workers x
Income". It captures "interactions" that are smoothed over in
the main effects model.
P93168-14/SA 42
DKS Associates
Table 10
Estimated Home-Based Work Productions
By Number of Workers
Mean Productions per
Classification "Sample Size" (Workers) Worker (all modes)
All Households 4429 1.30
Workers in Household
1 1463 1.44
2 2434 1.23
3 447 1.25
4+ 85 1.28
Table 11
Estimated Trip Productions for Home-Based
Work Workers + Household Income
(Main Effects, adjusted for interaction)
"Sample Size" Productions per
Variable Classification (Workers) Worker(all modes)
Mean - 4429 1.30
+ Main Effect
of Workers 1 1463 +0.16
2 2434 -0.08
3 447 -0.06
4+ 85 -0.02
+ Main Effect
of Household
Income $ 0 - 15,000 284 -0.14
$ 15 - 35,000 1771 -0.03
$ 35 - 55,000 1630 +0.06
$ 55,000 + 744 +0.00
P93168-14/SA 43
DKS Associates
Table 12 shows sparse sample sizes in some cells (remember to
divide each by the number of workers to obtain actual households).
This accounts for some of the 'bumpiness' of the mean productions
across the rows and columns.
The preceding main-effect model in Table 11 is somewhat smoother,
with the main effect of household workers being fewer trips per
worker with more workers. A curious main effect of income is seen,
however: increasing income accompanies more trips per worker except
for the highest income group.
Table 14 shows empty cells, sparse sample sizes in some other
cells, and "bumpiness" of the mean productions across the rows and
columns.
F-tests for these additional models are shown in Table 15. Each
analysis seeks to determine whether a proposed model explains
significantly more variance than a directly simpler model, called a
"comparison model." In the prior "core variables" ANOVA, Table 9,
the comparison model was the population mean, that is, the mean of
all households. For these additional models, the comparison model
is indicated in Table 15. For example, the probability is low
(0.1%) that the Workers+lncome model did not explain more variance
than the Workers model The probability is higher (2.9%) that the
interactions of workers and income (in the Workers x Income model)
did not explain more variance than the main effects of workers and
income (the Workers+Income model).
Click HERE for graphic.
P93168-14/SA 44
Click HERE for graphic.
Click HERE for graphic.
45
DKS Associates
Click HERE for graphic.
Using the number of workers as a classification dimension when the
trips are already on a per worker basis may seem redundant, but the
one-way classification test shows that non-linearity of trips per
household worker is significant. That is, the first worker makes
more work trips than any others. This is consistent with the
expectation that secondary workers are more likely to work part-
time; the survey coding even includes as workers some children who
help at a family business.
Thee two "main effect" models show positive effects of income and
vehicle availability in relation to work trips. A part-time worker
would tend to earn less income than a full-time worker, and is less
likely to need to own another car. (Auto ownership models have
demonstrated a strong relationship between number of workers and
number of autos owned, as may be expected.)
Not shown are main-effect tests of Income and Vehicles alone, on
per-worker trips, but without the Workers classification. These
models were significantly less effective than the main-effect
models shown above.
The fully cross-classified versions are weakly significant, and
suffer from "bumpiness" and the lack of general trends (except for
the trend revealed by the main effect models) and small sample
sizes in some cells. The main effect models serve as effective
'smoothing of the cross-classified models.' (Addition of the Puget
Sound Transportation Panel Survey's roughly 1600 households to this
analysis 3507 households should increase the sample sizes of cells
with small sample sizes, but not by much.)
Analysis of Home-Based Shop Productions
Unlike in work travel, there is not one overwhelmingly obvious
causal variable. A number of variables were tested to identify
primary variables. The estimated mean productions by
classification are shown in Table 16.
P93168-14/SA 46
DKS Associates
Click HERE for graphic.
A comparison of the core variables for home-based shop productions
is shown in Table 17. This includes the F-tests of each of these
classifications, relative to no classification at all. All of
these classifications show a significant positive relationship to
shop travel, though the significance of household income is weak.
The total number of persons is a strong indicator of shop travel,
but the number of persons age 5 and up shows even more distinction
among rates and more explained variance. This observation is
understandable given that person trips are counted only for persons
age 5 and up, assuming that shop travel is roughly proportional to
the number of persons. Income and number of vehicles show positive
but weak relationships, apparently from their colinearities with
the number of persons.
P93168-14/SA 47
DKS Associates
Click HERE for graphic.
These statistics support the expectation that the number of persons
in the household is the primary determinant of shop travel. If not
available, then other variables can provide a reasonable, though
indirect, indication.
To explore additional effects upon shop travel, further
explorations in Tables 18 through 25 arc normalized to the number
of persons age 5 and up in the household, so the production rates
are trips per person (age 5 and up). Normalizing in this manner
makes the cell means more equal, permitting the testing for subtler
effects, and it reduces the tendency for larger households from
dominating the measurement of those effects. Note that "sample
size" is actually numbers of persons age 5 and up, rather than
households. (In the ANOVA analysis, the trips by all household
members are still "bundled" together; greater variance is expected
of trips by each respective person.)
Special definitions for two-way cross-classified analyses are given
above in the Home-based Work analysis. In Tables 18 through 25,
"persons" and "per person" are understood to mean persons age 5 and
up, unless stated otherwise.
Click HERE for graphic.
P93168-14/SA 48
DKS Associates
Click HERE for graphic.
Click HERE for graphic.
49
DKS Associates
The cross-classification in Table 20 reveals a complex interaction
between persons and workers, the most apparent relationships being-
1) households composed only of one or two non-workers make
substantially more shop trips than average (probably because they
have more time to shop), and 2) households composed of only four or
more non-workers or four or more workers are sampled thinly, but
they appear to make fewer shop trips per person.
Click HERE for graphic.
Click HERE for graphic.
P93168-14/SA 50
DKS Associates
The "bumpiness" of the mean productions across the rows and columns
in Table 22 makes the full cross-classification model ineffective,
but a slight general trend is apparent in Table 21 but may be
unexpected. That is, with more income, shop trip production
declines. But the F-test below shows even this main effect to lack
statistical significance.
Click HERE for graphic.
Click HERE for graphic.
51
DKS Associates
The person x vehicles model in Table 24 has an empty cell, sparse
sample sizes in cells representing a large shortage or surplus of
cars to people, and has 'bumpiness" across the rows and columns.
F-tests for the per-person Home-based Shop models in Tables 18
through 24 appear below in Table 25. @ analysis cannot be compared
to the one on Table 17.
Click HERE for graphic.
Some conclusions from this more-detailed part of the analysis
(Tables 18 through 25) are:
- Persons (age 5 and up) should be continued as a classification
variable despite their use as a normalizing variable because
of the significant non-linearity of shop trips per person.
- The full interaction of persons and workers yields a
significant, complex, but somewhat understandable relationship
to shop travel.
- Income and vehicles were not effective in combination with
number of persons, either as main effects or in full cross-
classification. Unlike preceding analyses that found
significance in income or vehicles alone, these analyses
eliminated most effects of colinearities between persons and
those variables, and failed to establish effects of those
variables keeping other variables constant.
Analysis of Home-Based School Productions
This analysis considers home-based school travel by persons age 5
to 18, with the intent of capturing K-12 school travel only. The
travel survey has other trips coded to school, including adults
dropping kids off at school or day-care, and (apparently) some
college and trade school trips by adults. A mode choice/auto
occupancy model for school travel can add adult drivers of children
implicitly, since they are consequential to the mode choice
decision made by or for the child.
P93168-14/SA 52
DKS Associates
The most obvious causal variable of school trip production is the
number of children enrolled in school. In the survey, the closest
representation of this is the number of school-age persons.
School-age persons are taken as those age 5-17. (The slight
disagreement with the age limit for the trip will be reconciled in
a final analysis.) A number of other, less direct, variables are
examined as well. The estimated mean productions by classification
are shown in Table 26.
Click HERE for graphic.
Table 27 introduces a cross-classification of workers by persons
because of this classification's identification of non-workers in
relation to the whole household. Because the number of nonworkers
is constant along diagonals, the full cross-classification is more
appropriate than the main effects.
P93168-14/SA 53
DKS Associates
Table 27
Estimated Home-Based School Trip Productions
Persons x Workers
(Full Cross-Classification)
Workers in Household
Persons in
Household 0 1 2 3 4+
Sample Size (Households)
1 245 491 0 0 0
2 360 400 605 0 0
3 36 228 272 78 0
4+ 16 344 340 71 21
Mean Productions per Household (all modes)
1 0.00 0.00 N/A N/A N/A
2 0.04 0.18 0.00 N/A N/A
3 0.72 0.82 0.65 0.19 N/A
4+ 1.56 1.85 2.28 2.00 0.81
A comparison of the alternative cross-classification schemes for
home-based school trips is shown in Table 28 including the F-tests
of each of these classifications. All of these classifications
show a significant relationship to school travel, some much more
than others. As expected, the number of school-age persons shows
the most distinction among rates and explained variance. Persons
and Persons x Workers exhibit strong relationships. Income and
number of vehicles show moderately significant relationships,
apparently from their colinearities with the number of persons.
These statistics support the expectation that the number of school-
age persons in the household is the primary determinant of school
travel. If not available, then other variables can provide
reasonable, though indirect, indications.
Additional tests, normalized to the number of school age persons,
were made, but are not discussed in detail here. One result of
interest is that a non-linear classification model by schoolage
persons is not proven significant compared to a fixed rate of 1.45
trips per school-age person. (Significance of F = 52%.) This means
that for model application, a one-dimensional distribution (or even
zonal totals) of school-age persons may be used that is independent
of the classification systems chosen for the other trip purposes.
P93168-14/SA 54
DKS Associates
Click HERE for graphic.
Analysis of Home-Based Other Productions
In this analysis, home-based other travel covers all home-based
travel that is not for the purpose of work, shopping, school, or
college. Because home-based other travel is a catch-all for
numerous actual purposes of travel, there is not one overwhelmingly
obvious causal variable. A number of variables are here tested to
identify primary variables. The estimated mean productions by
classification are shown in Table 29.
A comparison of the core variables for home-based-other trip
productions is shown in Table 30 including the F-tests of each of
these classifications, relative to no classification at all. All
of these classifications show a significant positive relationship
to home-based other travel. The total number of persons is a
strong indicator of home-based other travel, but the number of
persons age 5 and up shows even more distinction among rates and
more explained variance. This is understandable given that person
trips are counted only for persons age 5 and up. Income and number
of vehicles show positive but less-strong relationships, apparently
from their colinearities with the number of persons.
These statistics support the expectation that the number of persons
in the household is the primary determinant of home-based other
travel, preferably where the persons that qualify as trip makers
are also used as the classification variable. If not available,
then other variables can provide a reasonable, though indirect,
indication.
To explore additional effects upon shop travel, further
explorations in Tables 31 through 40 are normalized to the number
of persons age 5 and up in the household, so the production rates
are trips per person (age 5 and up). Normalizing in this manner
makes the cell means more equal, permitting the testing for subtler
effects, and it reduces the tendency for larger households from
dominating the measurement of those effects. Note that 'sample
size' is actually numbers of persons age 5 and up, rather than
households. (In the ANOVA analysis, the trips by all household
members are still "bundled" together; greater variance is expected
of trips by each respective person.)
P93168-14/SA 55
DKS Associates
Special definitions for two-way cross-classified analyses are given
above in the Home-based Work analysis. "Persons" and "per person"
are understood to mean persons age 5 and up, unless stated
otherwise.
Click HERE for graphic.
Click HERE for graphic.
P93168-14/SA 56
DKS Associates
Click HERE for graphic.
Click HERE for graphic.
P93168-14/SA 57
DKS Associates
Click HERE for graphic.
The cross-classification above in Table 33 reveals an interaction
between persons and workers. The most apparent implication is that
workers tend to engage in less non-work out-of-home activities than
non-workers.
Click HERE for graphic.
P93168-14/SA 58
DKS Associates
Click HERE for graphic.
The sparse sample sizes in some cells in Table 35 account for some
of the "bumpiness" of the mean productions across the rows and
columns. (Recall the sample sizes must be divided by the number of
persons to obtain actual households).
Click HERE for graphic.
P93168-14/SA 59
DKS Associates
Table 37
Estimated Home-Based Other Trip Productions
Persons x Vehicles
(Full Cross-Classification)
Vehicles in Household
Persons in 0 1 2 3 4+
Household ("Sample Size" (Persons)
1 79 537 119 34 12
2 48 650 1764 640 254
3 15 219 690 489 279
4+ 0 161 845 596 530
Mean Productions per Person (all modes)
1 0.96 1.23 0.96 0.79 2.00
2 0.85 1.21 1.23 1.33 1.22
3 0.53 1.31 1.47 1.21 1.06
4+ N/A 1.34 1.33 1.16 1.24
The persons x vehicles model of Table 37 has an empty cell and the
sparse sample sizes in cells representing a large shortage or
surplus of cars to people, and has 'bumpiness" across the rows and
columns. Zero-car households appear to produce less home-based
other trips than others; for the others, little trend is evident.
A more detailed test of household composition upon home-other trip
production follows in Tables 38 and 39. It is a two-way
classification of workers and the presence of children under 5. To
simplify the analysis, persons are not used as a classification
dimension, but the trips are still per person (age 5 and up).
The test in Tables 38 and 39 also implies that workers make fewer
home-based other person trips than non-workers. It also found that
children under 5 (the household members not included as person-trip
generators) are related to the travel of the other household
members.
F-tests of the normalized home-other classification models appear
below in Table 40.
P93168-14/SA 60
DKS Associates
Table 38
Estimated Home-Based Other Trip Productions
Workers + Presence of Children Under 5
(Main Effects, adjusted for interaction)
"Sample Size" Production per
Variable Classification (Persons) Person (all modes)
Mean - 7951 1.25
+Main Effect of Worker 0 1089 +0.34
1 3018 -0.05
2 3211 -0.03
3 540 -0.15
4+ 93 -0.23
+Main Effect of Children None 6346 -0.03
Under 5 1 or more 1605 +0.14
Table 39
Estimated Home-Based Other Trip Productions
Workers x Presence of Children Under 5
(Full Cross-Classification)
Children Under 5 in Household
None 1 or
