Beruflich Dokumente
Kultur Dokumente
534
Travel Behavior
and
Values
~0~3
TRANSPORTATION
RESEARCH BOARD
NATIONAL RESEARCH
COUNCIL
Washington, D. C., 1975
Transportation Research Record 534 These papers report research work of the authors
Price $3.20 that was done at institutions named by the authors. The
Edited for TRB by Joan B. Silberman papers were offered to the Transportation Research
Board of the National Research Council for publication
and are published here in the interest of the dissemina-
Subject areas tion of information from research, one of the major
55 traffic measurements functions of the Transportation Research Board.
84 urban transportation systems Before publication, each paper was reviewed by
members of the TRB committee named as its sponsor
and accepted as objective, useful, and suitable for publi-
cation by the National Research Council. The members
of the review committee were chosen for recognized
scholarly competence and with due consideration for
the balance of disciplines appropriate to the subject con-
cerned.
Responsibility for the publication of these reports
rests with the sponsoring committee. However, the
opinions and conclusions expressed in the reports are
Transportation Research Board publications are those of the individual au th ors and not necessarily
available by ordering directly from the Board. They those of the sponsoring committee, the Transportation
are also obtainable on a regular basis through organiza- Research Board, or the National Research Council.
tional or individual supporting membership in the Each report is reviewed and processed according to
Board; members or library subscribers are eligible for the procedures established and monitored by the Re-
substantial discounts. For further information, write port Review Committee of the National Academy of
to the Transportation Research Board, National Academy Sciences. Distribution of the report is approved by the
of Sciences, 2101 Constitution Avenue, N .W., Washing- President of the Academy upon satisfactory completion
ton, D.C. 20418 . of the review process.
BEHAVIORAL MODELING OF
EXPRESS BUS-FRINGE PARKING DECISIONS
F. C. Kavak and M. J. Demetsky. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
v
all cases, the analysis is conducted by examining the implied values of the attributes
based on a comparison of relevant coefficients with the cost coefficient and with the
ratio of the travel-time and travel-cost coefficients. Waiting time is analzyed directly
by the input of data on actual waiting times incurred. Transfers and seat availability
are analyzed by constructing models using various dummy variable constructs. The
authors conclude that these variables are all important in the modal-choice decision
and indicate that they have a pronowiced effect on the individual's value of travel time.
This exercise was carried out in the context of the development of planning models for
Stockholm.
Liou, Cohen, and Hartgen compare aggregate and disaggregate modal-choice models
for the same data set. The work, which was done as part of the Niagara Frontier
Transportation Study, compares disaggregate models developed from a 12 percent
subsample of the original household-interview sample and from a conventional aggre-
gate model. The models are developed for four categories of trips and use only two
independent variables, travel-time ratio and trip length (used by the conventional aggre-
gate model). After correction factors for aggregation were applied, the aggregated re-
sults of the disaggregate models were found to be comparable to those of the aggregate
models. Thus, this research documented that disaggregate modal-choice models can
perform as accurately as conventional aggregate models but require a much smaller
data set.
Much work, particularly in Australia and Europe, has been concerned with the de-
velopment of estimates of the value of time. Guttman points out that there are a num-
ber of sources of serious error in estimating values of time, all of which must be
avoided carefully or compensated for if satisfactory values of time are to be obtained.
Several problems are investigated in the paper: accounting for traveler uncer-
tainty in regard to perceptions of alternative times and costs; substitution of times of
day, routes, and modes; effects of comfort differences; and measurement procedures
for important variables in modal- or route-split estimation. Guttman finds that models
can be successfully reformulated to avoid the bias caused by traveler uncertainty, and
other errors can largely be avoided by correct selection of the sample for value-of-
time estimation. Guttman also concludes that measured data are less suitable for
value-of-time estimation than reported data and that the supposed reporting bias ap-
pears less important than has been previously expected. Taking into account these
various error sources, Guttman points out that the value of work-travel time based
on data from Stanford Research Institute should be about $5.17 rather than $2.84,
which was obtained by ignoring these errors.
Much yet remains to be done in all of the areas encompassed by this RECORD. How-
ever, it is clear from these papers that significant steps are being made in many of the
relevant research areas. The papers in this RECORD clearly move disaggregate
models substantially closer to full operation, particularly for the modal-choiee de-
cision, and cast some extremely important highlights on the value of travel times and
other attributes.
- Peter R. Stopher
vi
COMPARISON OF EFFECTIVENESS OF VARIOUS
MEASURES OF SOCIOECONOMIC STATUS IN
MODELS OF TRANSPORTATION BEHAVIOR
Timothy J. Tardiff,* University of California, Davis
•THE major purpose of this paper is to compare the strengths of the relationships be-
tween several measures of socioeconomic status and variables measuring transporta-
tion behavior. In particular, the relationships involving income, a variable often in-
cluded in transportation models, will be compared to those involving other measures
of socioeconomic status, e.g., education and occupation. Such comparisons will in-
dicate whether these other measures should be included in future transportation studies
in place of, or in addition to, the income variable.
Models of individual (or household) travel behavior include, as the dependent variable,
a measure of transportation behavior such as frequency of travel or chOice of mode.
In addition, the independent variables can be divided into the following classes (not all
classes are necessarily present in a given model): (a) variables that describe the
available transportation systems, (b) variables that describe the individuals (or house-
holds), and (c) variables that are the results of interactions among variables in the
first two classes. Travel time is an example of variables in the first class, age is
an example from the second, and attitudes toward modes are examples of variables
from the third.
*Mr. Tardiff was a graduate student at the University of California, Irvine, when this research was performed.
1
2
The variables in the second class, which includes measures of socioeconomic status,
play a role somewhat different from that of the other variables. Whereas variables in
the other classes often have a definite behavioral interpretation, (e.g., such variables
might represent the decision process involved in the choice of modes), the variables in
the second class have no such interpretation. Consequently, when large samples are
available, such variables are sometimes used to stratify the sample into homogeneous
groups before the models are caUbrated. This strategy is used in place of calib1·ati.ng
a single unstratified model that contains the variables in the second class as additional
linear terms. That is, the strategy of stratification is based on the assumption that
the behavioral determinants of travel are somewhat different for various strata.
Stopher and Lavender (13) have found that stratification did, indeed, offer a closer
fitting model in the caseof modal-choice models. However, whether stratification
is used or not, a variable in the second class must be rather strongly associated with
the dependent variable if it is to be a useful addition to the mod.e l in question.
A review of the literature indicates that income has been the measure of socioeconomic
status most often used in transportation studies (8,~ 10,Q,.!_g.. g 15). In a few cases,
education has been used in addition to income (4, 'T).
The income variable is often included under-sfrictly economic assumptions. For
example, Lave (8) included income in his miodel of mode choice under the assumption
that, as income increases, a person's value of time also increases. A second example
is the use of income in models of household trip generation (10, 15). In these studies,
the variable is an indicator of the household resources availablefor travel.
However, sociologists have long recognized that there are other status indicators
on which people in society, or a subset thereof, diffe1·. Among these are education,
occupation, prestige, and interpersonal interactions (3, 5). Often these alternative
measures are used in empirical sociological applications Wlder the assumption that
the noneconomic indicators are more appropl'iate than income for describing individ-
uals or families. Relying on similar sociological arguments, Aldana et al. (1) chose
an occupational indicator rather than income as their measure of socioeconoriiic status.
They asserted that socioeconomic status in a sociological, rather than strictly econom-
ical, sense was the more appropriate describer of households with respect to travel
demand.
Rather than relying on a priori theoretical arguments, one may select the indicators
of socioeconomic status that are most strongly associated with the dependent V'ariable
in question. The empirical section of this paper is concerned with this method of se-
lection; that is, the same data base will be used to test relationships between the alter-
native measures of socioeconomic status and measures of h·anspo1·tation behavio1·.
RESEARCH DESIGN
The data necessary to test the relationships between socioeconomic status and trans-
portation behavior were gathered in November and December 1973 as part of a larger
study on transportation attitudes and behavior (14). A stratified probability sample of
223 households i:n the Santa Monica-west Los Angeles, California, ai·ea was selected.
One member from each household was interviewed. The size of the sample was based
on the fact that, on any variable in a study using a random sample of this size, the prob-
ability of the mean response being within 6.5 percent of the true mean for the entire
population is 95 percent.
The1·e are four possible measures of socioeconomic status available from the ques-
tionnaire used. Household income was measured on a seven-point scale. The highest
grade level completed by the respondent was recorded as the education variable, and
both the occupation of the household head and the respondent's occupation (if different)
were recorded. These were measu1·ed on a seven-point occupational scale developed
3
by Hollingshead (2). The precise categories used for each variable are given in
Table 1. -
The variables measuring transportation behavior (dependent variables) can be di-
vided into two classes: (a) those measuring monthly frequencies of travel (trip gener-
ation) and {b) those indicating modal selection for particular trips. The variables in
the first class are stratified by the purposes of shopping, visiting, and entertainment-
recreation. In addition, frequencies of travel for any of the preceding three purposes
were used. For each strata of trip purpose, the total frequency of trips as well as the
vehicular frequency was available. Vehicullar frequency is generally the measure of
trip generation used in studies based on origin-destination data. It is a measure that
excludes trips by bicycle and walking.
The second class includes variables indicating the usual modes selected for the
work trip and the most frequent nonwork trip. Although other modes were used, in
the results that follow only the car and bus modes are considered. In addition, whether
an iudividual used the bus at all for any reason in the month (actually, four weeks) pre-
ceding the interview was recorded.
The entire sample was not available for the relationships to be tested here. First,
the seven respondents who failed to answer the income question were excluded from
further analyses. Second, based on the nature of the dependent variable, a different
number of respondents were used to test the relationships involving the various depen-
dent variables. For example, many respondents either did not take a work trip or used
neither the car nor the bus. Similarly, some respondents had to be excluded from tests
of relationships involving the measw·es of travel frequencies because of inconsistencies
in their answers to the relevant questions.
(1)
Income Occupation
1 <$4,000 1 Higher executives of larger concerns, proprietors, and
2 $4,000 to $7,999 major profe:sslo1mls
3 $8,000 to $11,999 Business m:tnRgera, proprietors of medium-sized busi-
4 $12,000 to $14,999 nesses, and lessor profes.slonnls
5 $15,000 to $24,999 Admtntstrative pentonnel, O\'inere o( emall businesses,
6 $25,000 to $50,000 and rntnor pro[cssionals
7 >$50,000 Cle rici:d and Bnl~ workers, te chnicians, and owners of
little businesses
5 Skilled manual workers
6 Machine operators and semiskilled manual workers
7 Unskilled manual workers
Note : The highest grade level completed by the respondent is the edu cation variable.
Shopping 11.23 -0.27 0.43 +0.49 l.52 +2.44 l.79 -0.14 0.45 0.11 0.01 -0.01
(N = 205) B.23 -0.18 0.45 +0.50 l.52 +2.22 1.78 ~a.11 0.22 0. 12 0.01 0.08
10.37 -0.32 0.44 +0.46 l.52 +2.41 l.78 +0.18 0.41 0.11 0.01 0.01
10.02 -0.41 0.44 +0.02 l.57 +2.40 1.77 .49 0.41 0.14 0.02 0.06
Visiting 20.46 -1.30" 0.41 -2.44 1.41 -0.09 1.77 -0.83b 0.41 0.29 0.08 -0.13
(N = 211) 23.17 -1.54" 0.44 -2.37 1.42 -0.50 I. 76 -0.32 0.21 0.27 0.07 -0.02
17.51 -1.40" 0.42 -2.43 1.43 -0.68 1.76 +0.28 0.52 0.26 0.07 -0.01
17.40 -1.46" 0.42 -2.84 l.47 -0.71 1.75 +0.50 0.38 0.27 0.07 -0.00
Entertainment- 13.73 -1.27" 0.29 -2.40.. l.02 -0.23 1.23 +0.53 0.59 0.36 0.13 0. 12
recreation 8.92 -1.00· 0.30 -2.47b l.00 -0.26 1.21 +0,4:.. J , 15 0.39 0.15 0.27°
(N = 213) 17. 18 -1.10' 0. 30 -2.32b 1.01 -0.09 1.21 -0.69~ 0.27 0.38 0.14 0.24°
16.19 -1.13" 0.30 -2.01' 1.04 +0.05 1.22 -0.50 0.27 0.36 0.13 -0.22°
Shopping1 44. 82 -2.53" 0.74 -4.69 2.63 +0.04 3.20 +0.07 0.77 0.29 0.08 0.01
visiting, or 37.21 -2.19" 0.78 -4. 72 2.61 -0.30 3.16 +0,51 0.38 0.30 0.09 0 . 17d
entertainment- 46.31 -2.39" 0.76 -4.49 ?. . fi~ -0.01 3.15 -0.56 0.71 0.29 a.OB -0.13
recreation 44.51 -2 .62" 0.76 -5.09 2.73 +0.10 3.16 +0.3ti 0.'10 0.29 0.08 -0.07
(N = 197)
Not&: RC is the regression coefficient, A is tht multiple correlation coafficient, A2 is the squared multiple com1lation coefficient, rses is the simple correlation between the dependent variable and the indi·
cator of socioeconomic status, and SE. is the standard error.
'Significant at p <0.01. bSignificant at p <0.05 csignificant at p < O 01 (two-tail test) , dSignificant at p <0.05 !two-tail te~t!
Shopping 3.09 -0.21 0.35 +l.85 1.24 +4.41" 1.46 +0.20 0.36 0.23 0.05 0.06
(N = 205) 2.76 -0.16 0.37 +l.81 l.24 +4.48" 1.45 +0.06 0.18 0.23 0.05 0.06
3.48 -0.21 0.3'3 ... 1.79 1 2.4 +4.4!t 1.45 +0.08 0.34 0.23 0.05 -0.00
2.94 -0.31 0.36 +1.37 1.28 +4.56"° 1.44 +U.4ti u.JJ U.25 0.06 0.08
Vioiting 13.86 -0.85' 0.36 -2.13 1.24 +2.01 1.56 -0. 56 0.36 0.26 0.07 -0.08
(N = 211) 15.69 -1.00· 0.38 -2.0B 1.25 +l.74 1.54 -0.21 0.19 0.25 0.06 -0.00
11.93 -0.91' 0.37 -2.12 l.25 +1.61 1.54 +0.17 0.34 0.24 0.06 -0.02
12.07 -0.92' 0.37 -2.25 1.29 +l.58 1.54 +0.17 0.34 0.24 0.06 -0.05
Entertainment • 8.87 -0.96" 0.23 -1.12 0.80 +1.~:i~ U.':i7 +U.1'9 0.2S 0 :.tR 0 . 1~ 0,07
recreation 5.76 -a.so· 0.24 -1.14 0.80 +1.85 0.96 +0.24b 0.12 0.38 0.14 0.23°
(N = 213) 10.58 -0.86' 0.23 -1.05 0.80 +l.94~ 0.96 -0.42' 0.22 0.37 0.14 -0.20°
10.03 -0.87" 0.24 -0.83 0.82 +2.02b 0.96 -0.34 0.22 0.37 0.13 -0.20'
Shopping, 25.49 -1.87" 0.68 -1.40 2.41 +7.0Bb 2.94 +0.10 0.71 0.29 0.09 0.04
visiting, or 22.64 -1.73' 0.72 -1.43 2.41 +6.99b 2.91 +0.20 0,35 0.29 0.09 0. 13
entertainment- 26.76 -1.76b 0.70 -1.25 2.42 +7 .07b 2.90 -0.43 0.65 0.29 0.09 -0.11
recreation 25.50 -1.92" 0.70 -1.62 2.51 +7.J5b 2.90 +0 , 19 0. 65 0.29 0.09 -0.07
(N = 197)
stratified by trip purposes, AGE is the age variable, SEX is the sex vru:iable, C/DL is
the automobile availability ratio, and SES is one of the four indicators of socioeconomic
status (INC is the income variable, EDU is the education variable, OCl indicates the
occupation of the household head, and OC2 indicates the occupation of the respondent).
As mentioned earlier, there are four strata of trip purposes (including the stratum
of the variables measuring travel frequencies for any of the three specific purposes)
and two measures of travel frequency, total and vehicular frequency, for each stratum.
For each combination of trip purpose and measure of travel frequency, four versions
of the basic model are tested, i.e., one version for each of the measures of socioeco-
nomic status. The results of the multiple regression analyses are given in Tables 2
and 3.
Before the relative performances of the various indicators of socioeconomic status
are discussed in detail, some general observations are in order. It is apparent from
the results in Tables 2 and 3 that none of the indicators of socioeconomic status are
especially good predictors of travel frequencies for the purposes of shopping and visit-
ing. On the other hand, almost all of the indicators are significantly associated with
the trip·frequencies in the entertainment-recreation stratum. A possible reason for
these variations in explanatory power is that the first two purposes are relatively more
basic, and hence, there may be less systematic variance to be explained. Similarly,
entertainment-recreation travel is relatively more discretionary, and, therefore, there
may be more systematic variance in the dependent variables.
The three independent variables other than the measures of socioeconomic status
follow general patterns throughout the regression analysis. The age variable is in all
cases negatively associated with the dependent variables. In all relationships, exclud-
ing those involving shopping travel, the sex variable is such that women travel less
frequently. Finally, the pattem involving the automobile availability ratio is inter-
esting. When measures of total travel frequency (Table 2) are the dependent variables,
the coefficients of the automobile availability ratio are insignificant and often are in a
direction contrary to intuition. On the other hand, in Table 3, the coefficients of the
automobile availability ratio are much larger and often statistically significant. This
finding appears to indicate that differences in vehicular travel frequencies resulting
from differential automobile availability may be offset by more nonvehicular travel by
those with fewer automobile resources available.
The relative performances of the various indicators of socioeconomic status will
now be observed, and special attention will be focused on the income variable. Only
for the visiting stratwn is the income variable superior to the other indicators. Fur-
ther, for the entertainment-recreation stratum and the stratum containing travel fre-
quencies for any of the three purposes, the income variable appears to be clearly infe-
rior to the other three indicators. This finding is especially important because overall
fit of the models in these last two strata, especially the entertainment-recreation stra-
tum, is better than the fit in the shopping and visiting strata.
An interesting finding is that, in three of the four possible combinations of trip pur-
pose and measure of travel frequency for the last two trip purpose strata, the educa-
tion variable performed the best in terms of yielding the version of the basic model
that explained the most variance. In the case of vehicular travel for any of the three
purposes, although the variable indicating the occupation of the head of the household
yielded a model explaining slightly more variance than that involving education (the dif-
ference in multiple correlation coefficients was indiscernible when the coefficients
were rOWlded to two decimal places), the simple, two-variable correlation between
the dependent variable and the measures of socioeconomic status was highest for the
relationship containing the education variable. Inasmuch as the education variable is
probably easier to measure in terms of definition, operation, and interviewing than
the other three indicators, this particular finding assumes added importance.
6
The assumption common to many of the existing models of mode choice is that individ-
uals trade off time savings for increased costs in their selection of modes (!!_, .!!_, 11, 12,
13) .. Such models are calibrated by using statistical techniques appropriate for situa-
tions in which the dependent vai:iable is categoric rather than continuous. The partic -
ular technique that has been used most often is multivariate logit analysis. Based on
this technique, a modal-choice model in which there ai·e two alternative modes might
be
(2)
where
n
L(X) =CONSTANT+ a1ATIME + a2ACOST + L ai+ 2 11 (3)
i=l
where ATIME is the variable comparing the times of the competing modes, ACOST is
the analogous variable involving the costs, 11 are variables describing individuals, and
P(Mode 1) is the probability of selecting the first mode.
The data needed to calib1·ate such models must meet at least two requirements.
First, data on the times and costs for all competing modes must be available. The
present data base includes this information. Second, there must be a method of sep-
arating those who choose among modes from those who are captive to one of the modes,
and therefore are not included in the data used to calibrate the model. In this paper,
the criterion used to separate choosers from captives was whether or not the respon-
dents perceived the existence of a mode that is an alternative to their usual mode.
Unfortunately, when this criterion was applied, only 38 respondents were choosers
between the car and bus for the work trip (34 car users and 4 bus users). Similarly,
there were 41 choosers for the most frequent nonwork trip (33 car users and 8 bus
users). The small number of bus users as well as the small overall number of choosers
prevents the calibration of modal-choice models for these trips.
There are alternative tests that may shed some light on the issue of which indicator
of socioeconomic status is most strongly associated with modal selection. First, for
both the work trip and the most frequent nonwork trip, it is possible to observe the dif-
ferences in means between the car and bus user groups based on the various indicators
of socioeconomic status. In this case, the modal-use groups include both choosers and
captives.
Second, inasmuch as the data indicate whether each respondent used the bus at all in
the month preceding the interview, it is possible to calibrate a modified modal-choice
model by using this information as the dependent variable. Because travel time and
trip costs are trip specific, such information is not available for the model. However,
there are other variables describing the qualities of the competing transportation sys-
tems. These are the distance of the respondent's home from the bus line (in blocks)
and the automobile availability ratio. In addition, age, sex, and one of the variables
indicating socioeconomic status are included in the model, which is represented as
where
7
Work trip
Car users ( N = 107)
Mean 3.60 14. 56 3.10 3.27
Standard deviation 1.57 3.36 1.71 1.58
Bus users (N = 7)
Mean 2. 71 12.29 5.29 5. 71
Standard deviation 0.95 3.40 1.70 1.25
z-statistic 2.28' 1.71 -3.30' -4.9lb
Nonwork trip
Car users (N = 173)
Mean 3.52 14.06 3.47 3.94
Standard deviation 2.11 3.55 1.82 1.92
Bus users (N = 14)
Mean 2.36 12. 29 4.86 4.79
Standard deviation 1.55 2.92 1.61 1.48
z-atatistic 2. 61' 2. 14' -3.08' -2.02•
Table 5. Linear function values from modified modal-choice model in which dependent variable is whether or
not bus was used in month (N = 210).
-0.77 0.92 +0.20 0, 14 -1.57' 0.48 +0.29 0.38 -0.042 0.11 -0.006 0 .11 0.33
-0.95 1.20 +0. 20 0.14 -1.59' 0.48 +0. 29 0.38 -0.036 0.12 +0.011 0.057 0.33
-1.44 0.92 +0.17 0.14 -1.46' 0.47 +0.24 0.39 -0.083 0.11 +0.22~ 0.10 0.35
-1.04 0. 88 +0.19 0,14 -1.54' 0.47 +0.17 0.40 -0.070 0.11 +0.13 0.10 0.34
Note: p 2 is the logarithm of the likelihood ratio divided by the logarithm function when all coefficients of the independent variables (and the constant) are zero, LC is 1he logarithmic
coefficient, and S,E. is the standard error.
1
Significant at p <O 01. bSignificant at p <O 06.
where P(Bus Use) is the probability of using the bus for any reason in the month, BDIS
is the distance from the respondent's home to the closest bus line, and the other vari-
ables are as defined previously. In addition, the simple differences in means between
user and nonuser groups based on the various indicato1·s or s ocioeconomie status are
given in Tables 4 and 6. The logit values given in Table 5 were obtained by use of
the CLOGIT program written by Daniel McFadden of the University of California, Berke-
ley.
ID two of the three. situations (most frequent nonwork trip and whether or not the
bus was used at all}, the variable indicating the occupation of the household head was
superior to the others. For the work trip, the respondent 's occupation was the best
variable, and the occupation of the household head was next best. Although the income
indicator was superior to the education variable in all three situations, the difference
in means involving income was insignificant in the situation of whether the respondent
used the bus at all. Finally, it should be noted that the variable indicating the occupa-
tion of the household head was superior to the income indicator in all these situations,
and the variable indicating the respondent's occupation was superior in two of the three
situations.
The structure of the modified modal-choice model is interesting. Although the signs
of the coefficients of the independent variables are almost always in the expected direc-
tion (the coefficient of the education variable is an exception), only the coefficient of
the automobile availability ratio is significant in three of the four versions of the model.
In the fourth version, which is the best one, the coefficient of the variable indicating
the occupation of the household head is also significant.
The main conclusion from these exercises is that some of the other indicators of
socioeconomic status were superior to the income indicator in the tests performed.
This finding is consistent with the conclusion that the income indicator was generally
inferior to at least one of the other indicators.
CONCLUSIONS
Before the implications of the results are discussed, it is necessary to mention some
of the features of this study that might affect the generality of the results. First, spe-
cific indicators of socioeconomic status were used in the cases of the income and oc-
cupational variables. It is possible that different scales for these variables might yield
different results. Second, inasmuch as the sample was relatively small, the indicators
of socioeconomic status were used as additional linear terms in the models rather than
as stratification criteria. It is possible that, if a stratification strategy were followed,
the conclusions for the relative effectiveness of the various indicators might be differ-
ent. For example, if the relationships between the indicators of socioeconomic status
and the dependent variables are nonlinear, the stratification strategy might yield re-
sults different from those that are obtained by using the indicators as additional linear
terms. Finally, it was not possible to test the conventional modal-choice modal.
Therefore, one cannot generalize these results for this type of model with much cer-
tainty, and the results of this study must be regarded as suggestive rather than defin-
itive.
On the other hand, the results do have implications for transportation modeling and
surveys. There is some fairly strong evidence that the income variable, which has
been the most frequently used indicator of socioeconomic status in previous transpor-
tation studies, is not as effective as other indicators of socioeconomic status for ex-
plaining transportation behavio1·.
This is even more important in view of the success in obtaining usable responses
in this and previous surveys. Specifically, in the present survey, all respondents
supplied the info1·mation on education and occupation variables; however, about 3 per-
cent of the respondents refused to answer the income question. The refusal rates for
9
income questions have been even higher in other transportation surveys. For example,
it is possible to calculate the following refusal rates from three other studies: Lave
(8), 7.57 percent; Lisco (9), 17.61 percent; and Stopher (12), 3.87 percent. Therefore,
other indicators might be-more efficient because the chances of obtaining relevant in-
formation about them might be greater than for income.
In light of the greater effectiveness of some of the other indicators of socioeconomic
status in this study, it is recommended that future transportation surveys should con-
tain such indicators in addition to or in place of the income variable to explain travel
behavior. Of course, future research with different scales for these indicators [or
based on a large sample so that (a) the stratification strategy can be tested and (b) con-
ventional modal-choice models can be calibrated] would either strengthen or weaken
the basis for the preceding recommendation.
ACKNOWLEDGMENTS
The data used in this paper were collected in a research project supported by funds
from a National Science Foundation Grant for Improving Doctoral Dissertation Re-
search in the Social Sciences. In addition, the author would like to thank Gordon J.
Fielding of the University of California, Irvine, for helpful comments on the first
draft of this paper.
REFERENCES
1. E. Aldana, R. de Neufville, and J. H. Stafford. Microanalysis of Urban Travel
Demand. Highway Research Record 446, 1973, pp. 1-11.
2. C. M. Bonjean, R. J. Hill, and D. S. McLemore. Sociological Measurement.
Chandler Publishing Co., San Francisco, 1967, pp. 441-448.
3. R. Brown. Social Psychology. The Free Press, New York, 1965.
4. D. P. Constantine, R. Dobsen, and E. T. Canty. An Investigation of Modal
Choice for Dual Model Transit, People Mover, and Personal Rapid Transit Sys-
tems. Paper presented at International Conference on Dual Mode Transportation,
Washington, D.C., May 1974.
5. J. A. Kahl. The American Class Structure. Holt, Reinhart, and Winston, New
York, 1953.
6. H. Kassoff and H. P. Deutschman. Trip Generation: A Critical Appraisal.
Highway Research Record 297, 1969, pp. 15-30.
7. J. Lansing. The Effects of Migration and Personal Effectiveness on Long-
Distance Travel. Transportation Research, Vol. 2, No. 4, Dec. 1968, pp. 329-338.
8. C. A. Lave. Modal Choice in Urban Transportation: A Behavioral Approach.
Department of Economics, Stanford Univ., PhD dissertation, 1968.
9. T. E. Lisco. The Value of Commuters' Travel Time: A Study in Urban Trans-
portation. Department of Economics, Univ. of Chicago, PhD dissertation, 1967.
10. W. Y. Oi and P. W. Shuldiner. An Analysis of Urban Travel Demands. North~
western Univ. Press, Evanston, 1962.
11. D. A. Quarmby. Choice of Travel Mode for the Journey to Work: Some Findings.
Journal of Transport Economics and Policy, Vol. 1, No. 3, 1967, pp. 1-42.
12. P. R. Stopher. Factors Affecting Choice of Mode of Transport. University
College, London, PhD thesis, 1967.
13. P. R. Stopher and J. 0. Lavender. Disaggregate Behavioral Travel Demand
Models: Empirical Tests of Three Hypotheses. Transportation Research Forum,
Vol. 13, 1972, pp. 321-335.
14. T. J. Tardiff. The Effects of Socioeconomic Status on Transportation Attitudes
and Behavior. School of Social Sciences, Univ. of California, Irvine, PhD dis-
sertation, 1974.
15. H. J. Wootton and G. W. Pick. A Model for Trips Generated by Household.
Journal of Transport Economics and Policy, Vol. 1, No. 2, May 1967, pp. 137-153.
BEHAVIORAL MODELING OF
EXPRESS BUS-FRINGE PARKING DECISIONS
F. C. Kavak, Virginia Highway and Transportation Research Cowicil; and
M. J. Demetsky,
University of Virginia and Virginia Highway and Transportation Research Cowicil
A disaggregate data base was developed to analyze the behavior of com-
muters based on their choice of using fringe parking and riding the express
bus or driving directly tothe central business district. Statisticru summa-
ries of the information obtained were analyzed and interpreted to provide a
basis for selecting explanatory variables and for stratifying mathematical
choice models using the logistic fwiction. It was fowid that separate
models should be calibrated for groups of trip makers according to the ac-
cessibility of the fringe lot to their zone of residence. The models were
then calibrated by using alternative methods for specifying the time and cost
differentials between the options of driving directly to the central business
district and using the bus service. In this respect, models were calculated
to predict the probability of choosing the express bus as a function of the
age and sex of the trip makers, the ratio of automobiles to licensed drivers
for their respective households, and the relative times and costs of the
competing modes. Evaluation criteria, including statistical tests, were
used to select the model strategy that best described the choice situation.
Consequently, a model that related time and cost as the difference between
modes divided by the average of the two proved best. Finally, the models
were used to obtain elasticities of choice with respect to the independent
variables and the value of time, which agreed with the theory and the find-
ings of other studies.
•THE express bus-fringe parking concept is an example of a situation in which area-
wide travel models that represent an arbitrary automobile-transit choice cannot be
directly applied because of the aggregate data base. In such circumstances, the se-
lected transit operation serves only a segment of the travel market along a corridor
of the urban area and provides varying levels of service to different residential zones.
This problem, therefore, provides an opportunity to use disaggregate behavioral mod-
eling concepts to represent travel behavior for a defined subarea. In this respect,
firsthand data can be obtained that clearly relate the available choices.
Accordingly, this paper describes how survey data from a recent fringe parking
project, the Parham Express in Richmond, Virginia, were collected and used to for-
mulate a set of hypotheses for constructing a mathematical model of the choice between
driving to the CBD OJ,' parking at a fringe lot and using the express bus for CBD-destined
work trips. Alternative behavioral models based on the logistic function are then for-
mulated, tested, and evaluated.
Avenue). Although using the Interstate requires a toll of $0.25, it is generally pre-
ferred to using the arterials so that delays due to traffic congestion can be avoided.
The Parham Express bus, which originates at the fringe lot, provides a nonstop
trip to the CBD, and has 10 bus departures from 7 a.m. to 9 a.m. The one-way fare
is $0.50 compared with $0.35 (on the day of survey) for the regular area routes.
DATA
The system and trip-maker {behavioral) data used in this analysis of travel behavior
include
The questionnaire used for the bus survey is shown in Figure 1. It was distributed
on a weekday at the beginning of the bus trip and collected as riders left the bus, or it
was returned by mail. The license numbers of automobile travelers entering the ex-
pressway during the same time period (7 a.m. to 9 a.m.) on the two ramps in the
vicinity of the fringe lot were recorded. Their questionnaires were mailed. The auto-
mobile survey form is shown in Figure 2.
In the bus survey, 302 questionnaires were handed out, and 285 were returned.
After incomplete forms were eliminated, 229 usable responses remained.
Because license numbers of all vehicles entering the freeway could not be recorded,
traffic counters provided a control total for the automobile users that could be associ-
ated with the 302 bus users. Of the 4,030 automobile trips, 1, 165 valid questionnaire
responses, a 28.9 percent sample, were obtained. Of this sample, 381 responses
concerned work trips to the CBD. It was further assumed that the sample was rep-
resentative of the trip end distribution of the total population and, accordingly, the
381 responses represented a 28.9 percent sample of CBD-destined work trips by auto-
mobile.
Because the usable automobile responses made up a 28.9 percent sample, a similar
proportion of bus-user responses was developed so that the data set would be propor-
tionately representative of the automobile and bus populations, as would·be the case
for a home-interview survey. Inasmuch as such a requirement may result in a statis-
tical bias, a program was used to randomly select a 28.9 percent bus sample.
Furthermore, because the model is designed to reflect only real choices, captive
riders were removed from the data set. Individuals were classified as transit captives
if their alternative mode choice was another bus and if they indicated that they could not
have used a household automobile to make the trip. Trip makers were considered to be
captives of the automobile if they needed cars for their jobs or worked at locations in
the CBD that were remote from a transit stop. The development of the data set is
given in Table 1.
The transportation system data were obtained from the transit operator and the
Virginia Department of Highways and Transportation and were supplemented with traf-
fic engineering measurements.
MODEL FOUNDATIONS
The travel survey data were examined by summary statistics and correlation analyses
so that some basic hypotheses could be developed for stratifying the travel market into
homogeneous choice groups and so that explanatory variables for structuring a model
could be specified. The results, combined with information from previous efforts,
were then implemented to derive a modal-choice model.
The potential explanatory variables examined in terms of sample means and
Figure 1. Bus survey questionnaire.
THE FOLL<MING QUESTIONS CCllCERN THE BUS TRIP YOU ARE N<lol MAXING
5.
Time you began your trip : A.M.
How did you get to the Parham Road Lot to board this bu1?
__ drove and parked
IIIII
another bus
car passenger-car parked
::::::: dropped off-car not parked
::::::: walked, how many minutes
other (specify) - - -- - - I I II
What time did this bus leave the Parham Road Lot ? _ _ _ _ A.M.
IIIII
6.
11 . If this bus service were not available, how would you make this trip?
drive a car another bus
IIIII
12 .
ride H a car paasenger
::::::: participate in a carpool
other (specify) - - - - -
£3. Do you have any recommendations as to how this bus service could
be improved? - - - - - - - - - - -- - - -- - - -- - - - - - - -
16. Could you have u1ed one of the care to make this trip? __yea __no
18. What i• the combined annual income of all members of your household?
_$0-$4000 _$4000-$8000 _$8000-$12000 _over $12000
Figure 2. Automobile survey questionnaire.
roa OFFICIAL
USE <»!LY
D
Errors in recording license plates do occur. If this form
was sent to you by error, please check here and return__ Other-
wise, please continue.
3.
home
work other (specify)
5.
Time you began your trip:
8. After you parked the automobile, how did you get to your final
destination?
walk, how many minutes __ taxi
- - bus other (specify) _ __
10.
=
Do you usually make this trip: (check one}
alone
carrying passengers, how many? ---:---
within a carpool, how many members (count yourself)? - - - - III!
11. Could you have used the express bus from the Parham Road Lot for
this trip?
yes, but I chose not to b e c a u s e - - - - - - - - - - - - - - -
no, b e c a u s e . - - : - : - - - - : - - - - - - - - - - - - - - - - - - - -
not aware of this service D
12. Are there any improvements possible regarding the Parham Express
bus service which would make it acceptable enough to influence
you to use it? __ yes no If yes, what might they be? __
13. How many licensed drivers reside in your household (count your-
self}?-------
14. How many autos are owned bY. members of your household?
16. What is the combined annual income of all members of your house-
hold?
_$0-$4000 _$4000-$8000 _$8000-$12000 _over $12000
Table 1. Summary of data set.
380
) 374
373
-
~·
'··'
,. 367 I
J
15
correlation analyses were automobile ownership (number of automobiles per number
of licensed drivers), household income, sex, age, travel cost, residence zone, and
travel time. Initially, all variables except automobile ownership and household income
were determined to be significant.
Automobile Ownership
The distribution of automobile ownership for those who drive to the CBD (1.8 automo-
biles per household) does not differ significantly from that of express bus riders (1.9
automobiles per household). However, automobile ownership per se is not sensitive
to the number of household members who have access to cars, i.e., the licensed
drivers. In this respect, the presence of more licensed drivers than automobiles
places a constraint on the availability of cars to household members, and this may
encourage household members to consider alternative modes. Therefore, it washy-
pothesized that, for a given household, the ratio of automobiles owned to the number
of licensed drivers may be a more significant variable than automobile ownership.
This hypothesis was tested by using a chi-square test that showed that the distribution
of the aforementioned ratio for the bus sample was significantly different from that of
the automobile sample at the 0.01 level. Therefore, the ratio of automobiles owned to
the number of licensed drivers for a given household was selected for testing in the
model.
Sex
Female commuters showed a higher propensity to use the express service than did
males. It was assumed that the differing percentages of females in the bus and auto-
mobile populations (41. 7 versus 35.5 percent) warranted that the sex variable be in-
cluded in the model testing.
Household Income
The .data showed that age should be considered for inclusion in the model. Variations
in behavior were observed among the following age groups: 16 to 24 and 45 or over
versus 25 to 44.
Travel Cost
Cost figures for automobile travel were estimated by summing the freeway toll ($0.25),
half of the parking cost, and an assumed out-of-pocket operating cost of $0.04/mile
($0.025/km). (In the case where there were passengers as well, it was assumed that
riders shared the travel cost with the driver.) Transit travel costs consisted of the
$0. 50 fixed fa.re and the cost associated with getting to the lot. A statistical analysis
was carried out to determine whether the distribution of travel cost for the bus popula-
tion was different from that of the automobile population. The travel cost figures for
the automobile sample were greatly dispersed with a standard deviation of $0.296,
whereas the travel cost values for the bus sample varied only slightly from the average
16
(standard deviation = $0.043). The t-tests and F-tests further indicated that both the
mean and the variance of the bus population were significantly different from those of
the automobile population at a 0.01 significance level. Therefore, travel cost was se-
lected as a variable for the modal-choice model.
The set of zones that made up the market area for the service is shown in Figure 3.
Only 16 zones exhibited substantial trip productions, and four of those 16 had very
low express bus ridership. These observations were assumed to arise from the fol-
lowing factors:
The effect of accessibility to the fringe lot from the zone of residence was also ex-
amined as an influence on travel behavior. For each zone, minimum time paths to the
CBD and to the lot were plotted. The zones were then grouped into accessibility groups
1, 2, and 3 according to the following criteria:
1. The zone is located adjacent to the zone where the lot is;
2. The zone's minimum time route to the CBD falls close to the fringe lot; or
3. The lot is out of the way of the best route to the CBD from the residential zone.
Analysis of the data showed that accessibility of the fringe lot was highly related to
the transit use in a zone because relatively high split numbers for transit were indi-
cated for zones convenient to the fringe lot. Accordingly, it was concluded that some
measure of the accessibility of the lot from the residence zone should be reflected in
the model.
Travel Time
The survey provided information on the perceived time for each trip, and, because the
engineering data were incomplete, the former was interpreted for use in the models.
A chi-square test was carried out to compare the automobile population with the bus
population in relation to perceived travel time. The test indicated a substantial dif-
ference in distribution at the 0.005 significance level. The results of at-test at the
0.005 significance level further showed that the automobile and bus populations also
differ with regard to mean values of travel time.
Total travel time has been broken into running time and excess time in some cur-
rent research (1, 2). Excess time includes walking, waiting, and transfer time and is
usually weightea by a factor greater than 1. [The Washington model (3) used 2.5.J A
similar approach with varying forms of the time variables was considered in the model
development.
These observations from the survey data provided a set of variables for consider-
ation in the model development.
MODEL APPLICATION
Models were estimated by using logit analysis where the basic form is
17
e Gfxl
p - - -- {l)
b - 1 + e GCxl
where Pb is the probability of choosing the express bus, and G{x) is a linear function of
explanatory variables for which the parameters can be estimated by the maximum like-
lihood method.
The s-shaped graph of the logit model agrees with the common assumption that the
marginal utility of increments of a commodity, say time difference, is highest in re-
gions where the difference between the two modes is close to zero and, conversely,
approaches zero in regions where the difference is substantial. The logit methodology
can also be conveniently used to derive some important microeconomic properties of
the behavior group, i.e., elasticities and values of time.
The models presented here were calibrated with a computer program devel1Jped by
J . G. Cragg, University of British Columbia, and adapted to the CDC 6400 by Peter R.
Stopher, Northwestern University. Three models were calibrated, all of which in-
cluded the socioeconomic variables cited previously, but they differed from each other
by the nature of the specification of the time and cost measures. A surrogate measure
of the accessibility of the residential zone to the lot was initially entered as a dummy
variable to reflect the three accessibility groups described earlier. The estimated co-
efficient of the accessibility variable was found to be highly significant, and, for prac-
tical reasons, separate models were calibrated for each accessibility level as well as
one model with all the data.
A listing and definition of all of the variables used in developing the models are given
in Table 3. The three types of models presented are referred to by terms indicative
of the manner in which the time and cost measures are treated; they are the difference
model, log-of-ratios models, and relative values models.
The criteria used to evaluate the credibility of the alternative models included the
sign of the coefficients, the significance of the parameters, the probability of choice
at zero difference, and the predictive ability of the models.
The first step is a check on the rational performance of the model. The second
evaluation, which relates the significance of the parameters, uses the t-test and the
chi-square test. The t-test shows the significance of each parameter individually (4),
whereas the chi-square method tests the hypothesis that all coefficients except the con-
stant are equal to zero against the hypothesis that they all do not have zero values (5, 6).
The probability of choice at zero difference given by the model is considered refii"--
tive to the basic premise of the logit model that states that, when the values of the sys-
tem characteristics are equal for the competing modes, the trip maker is indifferent to
either mode. In application, however, this figure is difficult to attain because all fac-
tors influencing travel behavior are not reflected in the model. Accordingly, a slight
bias in favor of the bus mode was envisioned for accessibility group 1, and a high prob-
ability of automobile choice was anticipated for accessibility group 3. The group 2
models were expected to give values close to 0.50 because there appeared to be no
real advantage to either mode for those trip makers.
Finally the ability of the models to predict observed travel behavior was considered.
One way to accomplish this was to compare the expected value of the estimated number
of users with the number of observations from the data set. Here the percentage of
predictive error, e 1 , is defined as
A further test is suggested by Watson (7) and involves using one-half of the data to
calibrate the models and the other half to study predictions. This prediction error, e2,
is computed in a fashion similar to the way in which e 1 was.
18
Independent variable
Sex-0 = female; 1 = male X1
Age-0 = (25-44); 1 = otherwise X2
No. household automobiles
No. licensed drivers
Running time difference
"" =(RT, - RT,)
X•
Excess time difference Xs = (ET, - ET,)
Total cost difference x. ~ (C. - c,)
Natural log of ratios of total times X1 c 0o (T,/ T,)
Natural log of ratios of total costs xe c-e. (c,/c,)
Total time difference divided by X9 - T. - Tb
- (T, + T,)/2
average total time c. - c,
Total cost difference divided by x,. = (c. + c,J/2
average total cost
Dependent variable
Automobile transit choice
Calibration
For automobile trips p, = 0
For bus trips p, = 1
Application, probability of bus
choice P,
Note: a= automobile measure; b =bus measure. Accessibility groups 1, 2,
and 3 were also used in model development.
Independent variable
X1 -1.1080 -1.3083 0. 70169' -0.4901'
X2 1.1720' 0.3846' 1. 5550 1.0520
-2.17 63 -4.3375 -3 .4558 -3.2798
""
X4
Xs
0.1083
0.1530
0.2383
0.3752
0.2326
0.5803
0.1891
0.2262
x. 0.0378 0.0490 0.0591 0.03R9
Constant 2.1544' 4.5191 0.99814' 2.6365
Evaluative measures
x' 23.51 28.77 35. 84 80. 04
e1, percent 0.06 0. 55 0.19 0.28
p, at zero dHference 0.573 0.500 0.251 0.48
e2, percent 8.02
'Indicates that the variable or constant was nonsignif icant at the 0.05 level.
19
Difference Model
The system characteristics used in the difference model given in Table 4 include the
running time, excess time, and cost differences. The values for the reported mode
(choice) were obtained directly from the survey responses concerning perceived times
and costs, and the measures for the alternative mode were estimated by averaging the
perceived data for each zone. The use of averaged data for the alternate mode pre-
sents potential problems; but, inasmuch as the survey provided no information on the
alternative choice and the available engineering data were incomplete, this was the
only option available.
Estimated coefficients exhibited the expected signs. All of the system variables
were significant at the 0:05 level with the exception of the excess time, which was sig-
nificant at the 0.10 level for group 3. Chi-square tests proved the hypothesis that the
coefficients were significantly different from zero at the 0.005 level. Probabilities of
choice at zero difference were in agreement with the prior assumption (i.e., the bias
of the accessibility groups). The model predicted the original data almost perfectly;
however, the value of ea was the highest among the three models.
Log-of-Ratios Model
The difference model makes the assumption that the modal-choice decision is based
strictly on the absolute values of the differences in times and costs. In this respect,
the model implies that the choice between travel times of 15 and 20 min is equivalent
to a choice between 35 and 40 min when all other variables are kept constant. The
log-of-ratios model given in Table 5 was introduced to correct this fault, and it pre-
dicted two significantly differing transit choice probabilities for the hypothetical travel
times indicated previously (0.214 and 0.430 respectively). The time and cost variables
were calculated by taking the natural logarithm of the ratio of automobile time and cost
to bus time and cost. Also, inasmuch as the automobile data included some observa-
tions in which excess time was equal to zero (i.e., 07!0 = -co), total time figures rather
than a breakdown of running time and excess time were used in this model.
In the log-of-ratios model, all coefficients exhibited the hypothesized signs. All
the system variables were significant at the 0.05 level. Age and sex failed the signif-
icance test for the group 2 and group 3 models respectively. The chi-square test was
satisfactory at the 0 .0 5 significance level. Probabilities of choice at zero difference
were as hypothesized. The model predicted the original data almost perfectly, and the
value of ea was the second best among the three models.
A further model form, given in Table 6, was designed to use time and cost differences
relative to total cost and time. However, there was no evidence available to indicate
whether the actual mode taken or the alternate mode was being considered as a base
by the trip makers. Therefore, it was assumed that the average of the total times
from the alternative choices would be suitable.
The general characteristics of this model, although slightly better, were similar
to those of the log-of-ratios model. The values of the estimated coefficients were
within 8 percent of those of the latter model, and similar variables were insignificant.
The relative values model satisfied the expected sign test and had the lowest ea value.
Probabilities of choice at zero difference complied with the hypothesized trend.
Independent variable
X1 -1.3291 -1.2840 0.8103" -0.5259"
"2 1.1447 0.3235" 1.4357 1.0838
-2.3681 -3.8899 -4.6697 -3.5466
""x, 4.1488
3.3688
10.525
4.4247
8.3734
4.5293
6.4931
3.3654
Xe
Constant 2.3994 4.3030 2.0240" 2. 7782
Evaluative measures
x' 29.60 32.41 35.96 93.7
e1, percent 0.13 0.15 0.18 0.61
Po at zero difference 0.559 0.543 0.245 0.455
e2, percent 3.2
a Indicates that the variable or constant was nonsignificant at the 0.05 level.
Independent variable
X1 -1.3416 -1.3092 0.8207' -0.5294'
"2 1.1430 0.3443' 1.4384 1.0883
-2.3536 -3.9319 -4. 7517 -3.5738
""""
X10
4.2932
3.3990
10.899
4.7533
8.5377
4. 7783
6.6795
3.5717
Constant 2.3732 4.3230 2.0465' 2. 7839
Evaluative measures
x' 30.05 33.03 36.20 94.8
ei, percent 0.15 0.16 0.17 0.63
Po at zero difference 0.554 0.532 0.236 0.451
e2, percent 1. 59
a Indicates that the variable or constant was nonsignificant at the 0 .05 level .
Significance Probability
of Individual Chi-Square Expected Prediction at Zero Final
Model Variables Test Sign (e1, e,) Difference Ranking
Relative values 1 OK 1 OK
Log of ratios 2 OK 2 OK
Difference 3 OK 3 OK
discussed earlier, the models were ranked and the relative values model, as given
in Table 7, was selected as the best. Thus, this model primarily is used in the fol-
lowing microeconomic analyses.
The aforementioned models derive basic descriptive measures of the travel behavior
of the sample group, namely, elasticities of choice and the trip maker's value of time.
The elasticity (Em k) is a dimensionless number defined as the relative percentage
change in the probability of the specified choice that results from a 1 percent change
in any explanatory variable. Mathematically, the elasticity of the probability of using
mode m(pm) with respect to a given variable Xk is stated as
ap/pm
Em,k = oX k /X k (2)
With the relative values model, the probability of using the express bus relative to
using all of the system variables and the ratio of the number of automobiles to number
of drivers were elastic. However, automobile choice was elastic only with respect to
time (i.e., Em,k > 1). The elasticities computed for each mode are given in Table 8.
These values can then be applied to estimate the effect of various policy and social
changes on modal choice. For example, the analysis shows that a 1 percent increase
in transit cost will lessen the former probability of using that mode by 1.51 percent.
The value of time is found by calculating the change in cost due to a unit change in
the time variable and by keeping all other variables, including the choice probability,
constant. Accordingly G(x) must remain unchanged if the choice probability remains
constant. Given this fact, it has been shown that(~
(3)
where i =a, b; j =a, b; a= automobile; b =bus; and dC /dTJ =the rate of commodity sub-
stitution and, accordingly, the change in the value of cost of mode i (dC 1 ) for a unit
change in the travel time of mode j(dT J) is the value of travel time.
The value of travel time for the study area was estimated by substituting the appro-
priate terms for each mode in the relationship discussed, which results in four expres-
sions of the value of travel time for average sample values. These four values of time
relate to changes in the cost associated with each of the two competing modes, which
are due to a unit time change in either mode. As Watson (6) argues, it can be assumed
that the trip makers will perceive a change in bus and automobile costs due to a travel
time change for either mode. Therefore, it was assumed that the average of the two
values for each mode would be a reasonable estimate of the value of time for that par-
ticular mode. Under this assumption the values of bus and automobile time were cal-
culated to be $2.10/hour and $2.69/hour respectively. This indicated that bus users
value time differently than automobile users do. Furthermore, the values of travel
time were found to be 39.6 percent and 50.8 percent of the wage rates for the bus and
automobile users respectively.
Inasmuch as travel time was not broken into running and excess time in the relative
values model, the difference model was used to estimate the values of these times
separately. The values obtained were $2.33/hour for the rwming time and $4.55/hour
for the excess time. This amounted to 44 percent and 85.8 percent of the wage rates
respectively. However, the accuracy of these figures may be questionable because the
difference model was comparatively less reliable, as was discussed previously.
22
Disaggregate behavioral models of the choice between driving to the CBD and using
an express bus-fringe parking service proved highly successful. This conclusion is
based on the statistical tests and other evaluation methodology that showed that models
calibrated with all the survey data closely predicted the observed behavior of the entire
sample and that mOdels calibrated with only half of the data predicted the behavior of
the remainder of the sample. The findings showed that one method of specifying the
system variables of time and cost (the relative values model) performed slightly better
than the other two. This indicates that the better the model represents the way indi-
viduals perceive the differences among the attributes of competing modes, the more
reliable the model becomes. The variations in behavior among trip-maker groups
based on the accessibility of their residence zone to the lot are significant and can be
used to determine the service area for proposed lots. They should also be of use in
the design of feeder lines where they are being considered.
The survey design, by which data were collected directly from travelers on the com-
peting modes, proved adequate for the purposes of developing the model. It is envi-
sioned that the model developed here can be applied in planning new express bus-fringe
parking services with a minimal amount of socioeconomic and system data (Table 9).
Recent origin-destination study data should also satisfy these needs.
The significant variables in the models included sex, age, time, cost, and the ratio
of household automobiles to household drivers, all of which were recommended in the
preliminary analysis of the data. The income variable, which was initially nonsignifi-
cant, was similarly dismissed by t-tests for models in which it was included to test
the prior analysis. This finding is appropriate because this was not an areawide study,
in which income usually becomes appropriate, but it was a study of a relatively homo-
geneous suburban area.
Finally, the models gave the measures for elasticities of choice in relation to the
various independent variables and for the value of time. They were consistent with
theory and the findings in other studies.
The models developed here should be tested to see how well they predict the behav-
ior of trip makers for similar services in other areas. They should also be used in
the planning stages of a proposed project so that their effectiveness as a planning tool
can be evaluated.
A logical extension of binary-choice modeling techniques is to consider the submodal
split of access modes to the express bus such as walking, bicycling, feeder buses,
and kiss and ride. Subsequently, appropriate multimodal techniques should be applied
and tested for their validity relative to that of the binary models.
Improved problem-oriented surveys, such as the one described in this paper, or
origin-destination surveys for a selected subarea should be carried out for similar ser-
vices and should be expanded to include psychological data on trip-maker perceptions
Data Source
Socioeconomic
Numb r.r o r zonnl work trips terminating at destination z<>ne
or iiorvice (e .g., model output, m1mbel' of CBD work trips)' Census or gravity
Zonal distribution of workers by sex Census
Zonal distribution of workers by age Census
Average zonal automobile ratio (number of household auto-
mobiles to number of licensed drivers) Census
Transportation system
Average cost per trip via each alternative Trip assignment model
Average total travel time per trip via each alternativ e Trip assignment model
Zonal classification relative to lot acc essibility (3 groups
in Richmond model) Trip assignment model
• u only the expected proportion of CBD-destined drivers who will change to the bus mode is desired, then this
information is not required.
23
of comfort, convenience, and other subjective system measures. In addition, the sur-
vey respondents should be asked questions dealing with the competing modes as well
as with their chosen mode. These additional data are necessary to add to the behav-
ioral accuracy of the models.
REFERENCES
24
25
1. Given a correctly calibrated disaggregated behavioral model, how can aggregate
interzonal flows be forecast?
2. What biases are introduced in the calibration of aggregate models when zones
are not homogeneous? How do these biases affect demand forecasts? What correc-
tions can be applied?
3. What biases are introduced in the calibration of disaggregated behavioral models
when some independent variables are approximated by zonal averages? What correc-
tions can be applied?
The axioms of disaggregated behavioral demand modeling are that individuals represent
the basic decision-making unit and that each individual will choose one alternative
among those available that he or she finds most desirable or useful. This depends on
the attributes of the alternative and the socioeconomic characteristics of the individual.
For simplicity, we will develop this model and the results only for the classical prob-
lem of modal split for a work trip. The substance of our conclusions continues to hold
for more complex aspects of travel behavior. The full power and generality of the dis-
aggregated approach become apparent primarily in forecasting complex demand sys-
tems.
Suppose an urban area is partitioned into zones, indexed 1, ... , J. Let k =
1, ... , K index the individuals in the population, and let N1J denote the set of indi-
viduals commuting from zone i to zone j . Each individual is assumed to face a binary
modal choice, and a and bare the modes. Now, consider a specific individual k having
a vector of socioeconomic characteristics sk and facing vectors of attributes x•k and xbk
for the two modes. According to the behavioral model, the individual will have a utility
function of u = Uk(x, sk), which summarizes the desirability of a mode with attributes x,
and will choose the alternative that gives the higher utility; i.e., mode a will be chosen
if Uk(x•k, sk) > Uk(xbk, sk). Not all attributes of alternatives and socioeconomic charac-
teristics determining tastes can be measured. Consequently, the utility function of an
individual drawn randomly from the population can be thought of as containing a random
component reflecting his or her unmeasured idiosyncrasies in tastes. We therefore
write the utility function in the form
where V(x, s) is common to all members of the population and can be interpreted as the
representative utility, and Ek(x, s) is the random component. The condition for mode a
to be chosen by individual k can now be written
(2)
The unmeasured term on the left side of Eq. 2 has a statistical distribution in the
population and a cumulative distribution function G. Consider an individual drawn ran-
domly from the subpopulation who has socioeconomic characteristics sk and who faces
alternative trip attribute vectors x•k and xbk. The probability that this individual will
choose mode a is given by
(3)
26
The methodology of disaggregate behavioral demand analysis is to specify paramet-
ric functional forms for the representative utility function V and the distribution func-
tion G. [In general, the parameters of the distribution function G will be functions of
the measured trip attributes and socioeconomic characteristics of the subpopulation.
We suppress these arguments and later assume G to be independent of these variables.
This is a strong restriction which is inconsistent with some descriptions of the struc-
ture of taste variation in the population. For example, the disaggregated behavioral
model of Quandt (8) is incompatible with this restriction. A more complete discussion
of this specification is given in Domencich and McFadden (4).] Observed choices from
a sample are treated as drawings from binomial distributions, and the probabilities are
given in Eq. 3. Statistical methods such as the maximum likelihood procedure are used
to calibrate the unknown parameters. Following this procedure, we first specify that
the representative utility function be linear in unknown parameters, and the form is
L
V(x, s) = I: f3QZQ(x, s) (4)
l=l
where the ZQ(x, s) are numerical functions. [Numerical functions ZQ can be simple or
complex functions of the trip attributes and socioeconomic characteristics. For ex-
ample, Z 2 may be a trip attribute such as trip cost or on-vehicle time or a transfor-
mation such as the logarithm or square of one of these variables. It may be a dummy
variable that is one for mode a and zero for mode b, which corresponds to a pure mode
preference effect. It may involve interactions of trip attributes and socioeconomic
variables such as trip cost divided by trip time, trip cost divided by wage, or trip
walking time times an index of physical health. It may involve interactions of trip at-
tributes or socioeconomic variables with mode dummy variables, for example, a vari-
able that is income for mode a and zero for mode b or on-vehicle time for mode band
zero for mode a. Because the utility comparison involves only differences of the ZQ
variables, choice will be influenced only by factors that vary between modes. This
means a pure socioeconomic characteristic such as income, unless interacted with a
pure mode preference dummy, will not influence tastes and should not be included as
a Z 2 variable. However, it should be noted that an arbitrary smooth utility function
V(x, s) can be approximated as closely as desired by the form in Eq. 3. An important
class of models will be those in which no pure mode effects appear either singly or in
interactions with other variables. These generic models can be used to forecast new
mode demand.]
Defining z~ = ZQ(xak, sk) - Z2 (xbk, sk) and zk = z~, ... , z~' and letting f3 denote the
column vector of unknown parameters, we rewrite Eq. 3 as
(5)
Equation 5, derived from the behavioral model, has a simple conventional interpreta-
tion. Vector zk measures the differential attributes of the two modes, which are
weighted to account for the effect of measured socioeconomic differences. For ex-
ample, z~ might be the cost of mode a less the cost of mode b, and z~ the on-vehicle
time of mode a less the on-vehicle time of mode b, which is multiplied by the individ-
ual's after-tax wage. The coefficient vector f3 weights the components of zk into a
single measure {3'zk of the differential impedance of the two modes, and G represents
the response curve that gives the proportion of individuals choosing mode a at each
level of relative impedance. To simplify further analysis, we will assume that the
distribution function G is standard normal. This condition allows us to obtain simple
closed formulas but is not critical for our general conclusions. If ~ denotes the cu-
mulative standard normal distribution, Eq. 5 has a final form
27
(6)
that determines the probability ·of the choice of mode a by an individual drawn randomly
from the subpopulation of individuals facing a vector zk of differential attributes of the
two modes. This form in statistics and transportation demand analysis is known as the
probit model.
(7)
Equation 7 is just the expectation of the response probability of the empirical distri-
bution of the vector of the independent variables z. When n 1 J is large, this formula is
closely approximated by the expectation of the response probability of the underlying
distribution of the independent variables. We asswne that the distribution of z for in-
dividuals in N1 J is normal; the mean is z1l and the covariance matrix is A1 J. (This is a
plausible large sample approximation even when the individual observations are clearly
nonnormal, e.g., discrete. It should be noted that some of the conclusions, e.g., the
result concerning the consistency of disaggregated models calibrated from zonal aver-
ages, depend critically on the symmetry of this distribution and would change substan-
tially if the distribution were skewed.} Then, the differential impedance y = {3'z is dis-
tributed normally with mean {3'z 1J and variance a~J = f3' NJ{3. Therefore, the expecta-
tion of the response probability is
(8)
(9)
28
Setting w = µ 1 = O, a1 = 1, aa = aq, and µ2 = fj'ziJ, Eq. 9 implies a final form for the
aggregate modal split between zones i and j:
(10)
Equation 10 can also be obtained directly by returning to the condition given in Eq. 2
for an individual to choose mode a. Given tllat V(xa\ sk) - V(xbk, sk) = {3 1 zk, this con-
dition becomes
(11)
The first two terms in Eq. 11 taken together are assumed to have a standard normal
distribution, and fj'zk is assumed to be normal with mean {3'z1 J and variance a=.
There-
fore, the left side of Eq. 11 is normal with mean -fj'-ziJ and variance 1 + a~ 3 • This im-
plies that the probability of the event in Eq. 11 is given by Eq. 10.
By comparing Eqs. 10 and 6 for the individual response probability, one sees that
the components of differential impedance have the same relative weights, given by the
component of (3, but that the effect of the mean differential impedance for the zones is
attenuated by the factor ..j 1 + a;J, which reflects the degree of heterogeneity of the
variables facing the individuals in the zone. If the z ones are homogeneous, so that
each individual has the same socioeconomic characteristics and faces the same trip
attributes yielding a~ 3 = f3' N 3{3 = 0, then the aggregate and disaggregate models co-
incide. This conclusion provides a condition under which the disaggregated model can
be calibrated directly from interzonal data.
Equation 10 provides a method of forecasting interzonal modal splits from a knowl-
edge of the zonal average z1 J of the variables entering differential impedance and of the
intrazonal covariances A1 J of these variables. It is important to note that one does not
require a sample of individuals going from i to j or observations on individual data
points, although an alternative approach to computing the aggregate flows is to use
Eq. 7 directly for a random sample from the population. The zonal means z1J are
often available from transportation surveys; the covariances NJ are usually not re-
ported but could be constructed from the underlying data. The effect of a policy change
can be forecast from Eq. 10, provided the effect of the policy on z13 and A1 J can be de-
termined. The most straightforward case is a policy change that has a homogeneous
impact on the zone, as for example a $0.05 increase in the basic transit fare. This
changes the cost component of zu and leaves N; unchanged. A more complex example
wbuld be an increase in parking charges given that zones are heterogeneous with re-
spect to the availability of free parking to different individuals. This increase would
change the corresponding component of z1J but would also spread the distribution of
parking charges more widely, which would increase the variance of this component in
A13 and possibly also change the covariance of this component and other variables.
Specifying the precise effect of a policy change on NJ may be challenging; various ap-
proximations may become necessary, including the extreme approximation implicit in
conventional aggregate models that NJ is always zero and is therefore unchanged by
policy. The error introduced by this last approximation may be substantial.
An empirical example shows the order of magnitude of the factor ..J 1 +~Jin Eq. 10.
We consider an automobile-bus modal-split model calibrated on 160 individual workers
in the San Francisco Bay area (7, model 1). This model (7) has one pure socioeconomic
variable (income), one pure transportation variable (cost difference), and four mixed
variables defined by the after-tax wage times the time difference for on-vehicle, walk,
initial wait, and transfer wait times. Table 1 gives means and standard errors of the
29
Normalized
Variable Standard Error
Standard Probit of Intrazonal
Number Description Mean Error Coefficient Variances
1
For automobile and bus trips.
Variable
variable 2 4 5 6
variable
variable 2 4 5 6
variables and their probit coefficients (3. Table 2 gives the overall correlation matrix
of these variables. Using the MTC 440 zone network for the San Francisco Bay area,
we grouped the sample into zones. Assuming a common intrazonal covariance matrix
A= A1 J, we obtained the estimate of A (expressed as a proportion of the overall co-
variance matrix), which is gi-ven in Table 3. We note that 91 percent of .iilcome vari-
ation is intrazonal. The percentage is only 13.6 for cost difference variation but is as
high as 65 for one of the variables for mixed wage times cost. For this A matrix,
a~J = {3' A{3 = 0.485 and the factor '11
+ a~J = 1.22. This implies that the elasticity of
any independent variable of aggregate demand for the first mode between zones i and j
will be 82 percent of the average of the corresponding individual demand elasticities
for trips from i to j. [This is a conservative estimate of the bias in typical modal-
split models because (a) the calibration on which our calculation is based underweighs
transit walk time because of a failure to distinguish walk access and automobile access
transit trips and (b) a variety of additional automobile availability and socioeconomic
variables are excluded from the model. Jn each of these cases, the contribution of
intrazonal variance can be expected to be a high proportion of the total.] We conclude
that direct application to aggregate modal splits of elasticities calculated from disag-
gregate models will tend to overestimate the magnitude of demand response, and
30
elasticities calculated from aggregate data will underestimate, on the average, indiv-
idual demand elasticities.
To determine whether the assumption of a common intrazonal matrix A= A1 J was a
good one, we computed the standard errors over the origin-destination pairs in our
sample of the diagonal variance elements in the AiJ· These standard errors divided
by their means (i.e., the diagonal elements of A) are given in Table 1. We conclude
that there is substantial variation in the AiJ over origin-destination pairs and that it
would be a poor practice in applications to assume a common A matrix. Finally, we
note that over half the contribution to the variance a~J = {3A' f3 in our example is con-
tributed by off-diagonal (covariance) terms. Thus, it is not sufficient to look only at
the variance elements of A. Because of the small data base, these conclusions are
necessarily tentative and are intended only to suggest orders of magnitude.
With any sigmoid-shaped choice function, attempts to represent aggregate modal
splits by the disaggregated function with homogeneous zone assumptions will similarly
bias the results. The bias will be toward the extreme possible values of the modal
split-too low for values of Eq. 6 below 0.5, too high above 0.5. Only in the special
cases where zones are homogeneous or where interzonal splits are equal to 0.5 and
their differential impedances are distributed symmetrically about their mean would
there be no bias. The magnitude of the bias found in the example is sufficient to re-
quire that the zone variances be accounted for in the aggregation process.
We next examine the effects of calibrating an aggregate model directly from interzonal
flows when, in fact, individual responses conform to the disaggregate behavioral model
of Eq. 6. The typical aggregate modal-split model assumes that the frequency of choice
of mode a for trips between i and j is given by a function of the differential impedance
of the two modes, and differential impedance is measured in turn by a weighted com-
bination of the zonal average differential attributes of the two modes. Given that the
response curve is normal, this model is
(12)
where 'Y is the vector of unknown parameters weighting the components of differential
trip attributes into a measure of differential impedance. Although Eq. 12 is identical
to Eq. 6 for individual response probability, it would be justified in traditional aggre-
gate demand analysis on the basis of its success as an empirical law rather than on
behavioral grounds. Calibration of the model of Eq. 12 from data on interzonal flows
would typically be done by applying least squares for origin-destination pairs to the
equation
(13)
(14)
31
Now, suppose that travel demand behavior is in fact determined by the disaggregated
model in Eq. 6. Aside from sampling error, which is inversely proportional to the
number of trips between zones, and which can be ignored when the total number of
trips is large, the observed interzonal modal-split frequencies Pu from Eq. 10 will
satisfy the following:
(15)
or
(16)
If one substitutes Eq. 16 in Eq. 14, the value of .Y converges, with a probability of one
as sample size increases, to
(17)
1, J 1, J
From Eq. 17 we can draw inferences on the biases introduced in the estimation of 'Y by
the presence of heterogeneity within the zones. First, we note that, in homogeneous
zones where a~ 3 = O, .Y is a consistent estimator of {3. When zones are heterogeneous,
the effect of the term ~ will gene1·ally be to bias the estimates .Y downward in
magnitude. If large magnitu~es of a component of z1 J tend to be associated with large
values of a~ J' the bias in the corresponding coefficient will tend to be larger than for
a coefficient for which the component of ziJ and a~J tends to be uncorrelated. Given that
a~J is a constant a for all zones, Eq. 17 reduces to a si mple expression for the bias
2
y = 1 f3 (18)
~
(19)
Equati on 19 gives a consis tent estimate of the parameter vector fj. Then, Eq. 10 can
be used with the estimates y ancJ a~J to correctly forecast the effects of policy changes
that are either homogeneous or heterogeneous in their impact on zones.
In practice, it may be feasible to obtain a consistent estimate of the intrazonal co-
variance matrix A1J from external s ources, but it may not be feasible to obtain the
initial consistent estimators of fj necessary to construct estimators of a~J = {3 ' A1 Jfj.
An alternative is to consider the estimator in Eq. 19 as an implicit fwlction of y:
(20)
(21)
where
M Z.2. = "z
L#I
1
J z1 J 1
i, j
33
and
M.y = :E ziJ 1
~- cP1J)
i, j
Given data for a sample of individuals, which include accurate measurements of the
attributes of the transportation alternatives faced by each person, the disaggregated
model of Eq. 6 can be calibrated by straightforward application of a variety of statis-
tical techniques. These techniques include maximum likelihood procedures and, under
suitable circumstances, minimum chi-square procedures (3, 6). A second approach to
calibration is to use the conclusions of the preceding discussfOn to estimate the behav-
ioral model from zonal data when the intrazonal covariance matrix of the independent
variables is known or can be estimated consistently from external sources.
In practice, a third approach is important. Data on choices, socioeconomic char-
acteristics, and some attributes of alternative trips are collected for a sample of in-
dividuals, and the remaining attributes of trips are measured only by zonal averages.
For example, individual data may be collected on income, age, and travel costs, and
zonal averages obtained from transportation grids may be used for access and on-
vehicle travel times. This introduces a measurement error in the independent vari-
ables, which may bias the estimates obtained by applying the statistical methods or-
dinarily used in calibrating disaggregated models . (The reason calibration of a mixed
model with individual response frequencies and interzonal average explanatory variables
gives consistent estimates and calibration of an aggregate model with zonal average re-
sponse frequencies does not give such estimates is that an arithmetic average of inverse
cumulative normal transformations of frequencies does not equal the inverse cumulative
normal transformation of the arithmetic average of the frequencies. It is this nonlin-
earity of the response curve that makes it nece ssary to distinguish disaggregate and
aggregate models.)
We will now determine the conditions under which biases will occur and derive cor-
rection formulas. The first useful conclusion is that replacing all independent vari-
ables by zonal averages yields consistent estimates of the parameters of the behavioral
model when the usual statistical methods are applied. (This conclusion depends criti-
cally on the assumption that the intrazonal distributions of the independent variables are
not skewed. If, to the contrary, these distributions are skewed, and modes lie between
the zonal mean vectors and the total population mean vector, then the use of zonal aver-
ages introduces a regression-to-the-mean effect, Which generally biases estimates
downward in magnitude.) (Note that individual observations are used as the dependent
variables and, thus, differ from the zonal average frequencies used in the previous
section.) The loss of efficiency in estimation resulting from use of this procedure may
be substantial because not all information is being used. However, in very large
34
samples, this loss may be offset by the saving in cost of providing accurate measures
of the travel attributes for each individual. The second result provides a straightfor-
ward linear transformation of the estimates obtained from the usual statistical proce-
dures, which makes the resulting estimators consistent. These correction formulas
require external estimates of the intrazonal variance of the independent variables.
The behavioral model of Eq. 6 can be written as
(22)
Partition zk = (zl, zU, where z~ is the (possi bly empty) subvector of components that
can be measured for each individual and z~ is the subvector of components that will be
approximated in the estimation process by the zonal averages. Then,
(23)
(24)
where (k are error terms with a zero mean, which converge to zero with a probability
of one as the numbers of members of the groups increase. Suppose now that z~ is re-
placed by the zonal average z1J and apply ordinary least squares to the equation.
(25)
This general estimation procedure is known as Berkson's method, which under a cor-
rect specification of the independent variables provides estimates of the parameters,
which are consistent and equivalent in asymptotically large samples to the maximum
likelihood estimator.
We rewrite Eq. 25 in vector notation as
(26)
where y and ( are column vectors with components yk and (k respectively, and z 1 and
"2"2 are matrices with rows z~' and z~J' respectively.
The ordinary least squares es-
timates from Eq. 26 then satisfy
(27)
But Eq. 24, written in the vector notation used above, is y = z 1 {31 + z 2 {32 + (, and this
implies
35
(28)
The last term in this expression converges to zero with a probability of one as the num-
ber of individuals of each type becomes large. Then &1 , &.i converge (with a probability
of one) in this limit to values a1, a:i and satisfy
(29)
We will now consider further the structure of the matrices in this expression. De-
fine E to be a square block-diagonal matrix with a block for each zone pair i, j of the
form (1/n 1 J) e 1 Je 1 J ', where e 1 J is a column vector of ones of length n!J. Then, E satis-
fies
Z2 = Ez2 (30)
and
2
E =E
(31)
Partition
_(An A12)
A - ~1 A22 (32)
(33)
36
(34)
(35)
(36)
1
°'1)- _! (M11 M12 + A12)- (0 nA12) (a1)
( a2 (37)
n M21 M22 0 0 a2
Define
(38}
Then, writing out the expression for the inverse of a partitioned matrix, we obtain
1
f31) (°'1) (M.lf(M12 + A12)C- M21 + 111) _1
( R = ,.. - c-1 M21 M11 A12°'2 (39)
1'2 ""2 -
The second term in Eq. 39 reflects the bias in the estimators of the disaggregated
model introduced by using zonal average values for the variables z2. First, we note
that, if the intrazonal correlations of the variables measured by zonal averages are all
zero, i.e., A12 = O, then the usual estimators are consistent, and no correction is nec-
essary. This is true, in particular, if all variables are measured as zonal averages
so that z 1 is empty. Second, we note that the magnitude of the bias is determined by
the degree of intrazonal variability relative to total variability. If, for example,
A12 = 8M12 for some e with O ~ e < 1, then
(40)
where C = M22 - (1 + 8)M21M;fM12 . Third, Eq. 39 provides a correction for the usual
disaggregate estimators, which makes them consistent for this problem. Application
of this correction requires that consistent estimates of A12 be obtained from external
sources. Estimates of A might be obtained from previous transportation surveys or
from more limited data sets when it is possible to limit, a priori, the structure of A.
An example of the latter construction would be an estimate of access-time covariance
with income in each zone based on the geometry of the zone, the location of transit
routes, and census block statistics on income.
37
SUMMARY
This paper has established links between aggregate travel demand models and disaggre-
gate behavioral models. Equation 10 provides a formula for computing interzonal flows,
given a calibrated disaggregate model, and zonal averages and intrazonal variances for
the independent variables. Biases resulting from direct calibration of an aggregate
model when individual behavior conforms to the disaggregate model are derived. The
implications of these biases for forecasting are discussed. Equation 20 provides a
method for obtaining consistent estimates of the parameters of the disaggregate model
from interzonal flow data. Biases introduced in calibration of disaggregate models
when some independent variables are approximated by zonal averages are also dis-
cussed. It is shown that the use of zonal averages for all independent variables results
in consistent estimates of the parameters of the behavioral model. Equation 39 gives
a formula for correcting the bias in the estimates of the behavioral parameters when
a mix of individual and zonal average variables is used, and an estimate of the intra-
zonal covariance matrix can be obtained from external sources.
ACKNOWLEDGMENT
This research has been supported by a grant from the National Science Foundation.
This study has not been carried out as part of the research program of the Metropolitan
Transportation Commission or the U.S. Department of Transportation, and the views
expressed herein do not necessarily reflect the opinions of these agencies.
REFERENCES
•BY MEANS of a questionnaire, the greater Stockholm Labor Force survey of 1968
(AKU-68) presented an excellent opportunity t o analyze the greater Stockholm com-
muting pattern at the disaggregate l evel. T he Stockholm ar ea trans portation study of
1971 (TU-71) gave us an opportunity to further investigate the unexplored field of
evaluation of comfort and convenience at a disaggregate behavioral level. Based on
these surveys, a number of special studies were carried out at the Joint Planning
Board of Stockholm and the Stockholm County Counci.l in 1970 t o 1974 @ ).
This paper is concerned with the application of a disaggregate binary-choice model
for the commuters' modal choices reported in AKU-68 and TU-71 and places special
emphasis on the impacts of changes in travel comfort, convenience, and transit waiting
times.
Our reasons for this choice of emphasis are simple: Research on travel choice
has for the past 10 years been concerned with the value of time savings and estimates
of time and cost elasticities. The role of comfort and convenience was always referred
to as important but rarely was incorporated explicitly as a policy-oriented variable in
econometric models.
The first major contributions to empirically clarify travelers' rankings (and, im-
plicitly, their relative evaluation) of various transportation system improvements were
carried out by sociopsychological research techniques such as paired comparison and
semantic scaling. Studies along these lines strongly suggested that comfort and con-
venience play a major role when travelers evaluate alternative means of travel.
38
39
Gustafson, Curd, and Golob (7) found that the following characteristics seem to be
more important than lower fares; in order of importance these are
All these factors are concerned with time savings or improvements in comfort and con-
venience. This indicates what variables to concentrate on in trying to divert automobile
users to public transportation, although it should be emphasized that the above ranking
depends on the levels of service experienced by those who were interviewed.
In addition to these empirical findings, theoretical developments took place in dis-
aggregate choice analyses. DeDonnea (5) modeled travel demand so that a value of
time savings would result from an algebraic model of the "economic allocation of time"
as contrasted to preceding theoretical models in this field, in which the time values
derived contained the composite value of the time savings and the corresponding change
in comfort and convenience.
He therefore concluded that an ideal, econometric modal-choice model should ex-
press modal choice as a function of three types of explicatory trans por tation policy
variables (~): (a) the cost of the trip by each mode, (b) the time of the t rip by each
mode, and (c) the objective comfort characteristics of each mode.
Independently, DeSerpa (~) derived similar results. McGillivray (.!E.) used a s imilar
theoretical approach to show the relationship between the modern microeconomics of
choice behavior and the econometric binary-choice models (logit, probit, and linear
probability function). This has been considered the theoretical justification for deriving
money values on nonmarket factors such as time, comfort, and convenience from such
statistical choice models.
Anot her approach or iginally introduced by Luce (12) was fu rther developed by
Marschack (ll), McFadden (ill, and Charles River Associates (4). This showed that
the logit model of choice analysis can be derived from probabilistic utility theory.
One may therefore conclude that applying the logit model to analyze individual travel
choices is compatible with the theory of demand, and trade-off values .may consequently
be derived.
Both the algebraic and the probabilistic point of departure rest on some strong as-
sumptions regarding separability or independence of irrelevant alternatives (g_, i, fil,
but the lack of better alternative means for analyzing our sample data made our choice
of the econometric approach quite easy.
1. The socioeconomic characteristics that vary between individuals are included ex-
plicitly as variables in the model.
2. The sample is subdivided into categories so that socioeconomic variables are not
needed as explanatory variables of the modal choices.
Both approaches were used in this study, although it is difficult to succeed with
category analyses because the subsamples usually turn out to be too small. Which
approach to use depended on whether the variable in question influenced the traveler's
evaluation of the other variables or whether it influenced the a priori probability to
use a car.
40
The basic issue, however, had to do with the transportation policy variables inas-
much as one aim of the study was to provide public planning officers, politicians, and
transit operators with a firmer knowledge of what public transportation system char-
acteristics to focus on when they suggest urban transportation options in the future.
Travel time, travel costs, and travel comfort and convenience are the major groups
of transportation policy variables. As in many other similar studies, each segment of
travel time and travel cost is identified. But unlike most previous studies in this field
(exceptions include~ lb .!!!., ~. our models also include travel comfort and con-
venience variables, the most important of which seem to be "number of transfers" and
"chance of getting a seat" (7).
The use of each trip segment as a basi.s for separate explanatory variables is well-
founded in modern microeconomic consumer theory (~ .?., ~ and is compatible with the
intuitive policy-oriented segmentation that is relevant to the supply of transportation
services. The travel segments used as explanatory variables not only reflect the trip
segmentation perceived by the travelers; they also reflect the action parameters of
transportation planning authorities.
This raised the question of whether travelers perceive the same variable values as
those measured by traffic engineers. We assumed that no bias of this kind was present;
this assumption was supported by the findings of Lansing and Hendricks (.!Q_) in the
United States, Quarmby (17) in the United Kingdom and by recent findings in Sweden
(1) (Figs. 1 and 2).
- This is an important assumption because both perceived and engineering data had to
be used in the same models as a consequence of the limitations of perceived data for
the travel mode not chosen for the journey to work.
It was also assumed that the travelers in our samples have a good knowledge of the
values of the attribute dimensions of the alternatives available. This was a reasonable
assumption because we were dealing with work trips. This assumption seemed to be
supported by a recent study (1) and TU-71, in which 68 percent of the automobile com-
muters reported that they were conscious of public transit costs and 84 percent were
conscious of the time it took to travel by public transit. Among transit users, similar
percentages were 40 and 76 respectively. Similar results were also derived by Lansing
and Hendricks (10).
Socioeconomic variables are important for making valid predictions for future
situations with significantly different socioeconomic structures within and between the
zones of the study area. In addition to sex, age, and income level, a variable express-
ing the use of the automobile while the respondent was at work was included because
we believe that frequent car use during working hours makes commuting by public transit
less attractive. From AKU-68, we found that about 50 percent of the automobile com-
muters also used their cars during working hours. The frequency of car use during
work time turned out as follows according to TU-71:
Daily 14
At least once a week 20
At least once a month 11
Only occasionally 20
Never 35
Total 100
A brief definition of the variables and the nature of the data (in parentheses) used in
AKU-68 follow. Both reported and estimated values refer to round trips. Cost and
income are measured in Swedish crowns (Kr).
41
Figure 1. Perceived versus estimated travel time for automobile and public transit users
in Stockholm .
30 30
~ 25
·;::
25
Qi
> 20 20
g
al 15 0 I- + 15
>
'iii
e., 10 + 10
Q.
5 5
0 0
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30
estimated travel time
Figure 2. Perceived versus estimated travel time for automobile and public transit users in Uppsala.
100
90
Auto travel time
80
70 70
GO
~
·;:::
Qi
50 50
g> 40 40
.,
'O
>
'iii 30 30
eQ)
Q.
20 20
10 10
10 20 30 40 50 60 70 80 90 10 20 30 40 50 60 70 80
Stockholm, the automobile running costs were estimated at 2.46 Swedish Kr/ 10 km in
peak periods and 2.25 Swedish Kr in off-peak periods. When the automobile was re-
ported to be used for work purposes, the cost for the individual was defined as zero
because of tax deduction allowances.
c: data were basically provided by the survey. For those respondents who answered
that no parking was available, 20 Swedish Kr were coded because this would be the cost
of using a parking lot. Ct was estimated by using the lowest monthly cost alternative.
NOT was derived through the reported work trip composition for transit users and
from the alternative transit arrangement reported by automobile users. Inasmuch as
transfer times were not explicitly recorded, the coefficients of the NOT variable mea-
sured the value of transfer inconvenience.
The 1968 waiting times are estimated by the British formula (16). Because we are
particularly interested in the sensitivity of our conclusions with respect to waiting-
time behavior, the underlying distribution of time intervals between transit arrivals
should be known. A brief definition of the variables in TU-71 follows. The data for
all the variables except UOC were reported by automobile and transit users; UOC was
reported only by car users. Costs are measured in Swedish Kr.
The symbol T~ illustrates so-called excess time (not including transfer time) and con-
tains the sum of T~ait(o) and T~• 1 k.
The most striking feature of Table 2 is the drastic change in the waiting-time coef-
ficient when T~ait is replaced by T~• 1 t(O). All the other coefficients remain unchanged
44
Table 1. Distributions of transit frequencies
(AKU-68).
0 to 2.5 23 2.2
2.5 to 7.5 247 23.5
7.5 to 12.5 459 43. 7
12.5 to 17.5 109 10.4
17 .5 to27 .5 51 4.9
> 17.5 161 15. 3
1~
12
RANDOM WAITING TIME
11
10
5
_ _ --
-- .... ....
_.... ..... - EXPECTED
TIME
WAITING
_ - .... ....
j
~
'7~ -- ----
~
~
0 TRIP HEADWAY
4 b 8 10 11 1~ 1& 18 20 22 24 2&
Model
L(x, • )
0
p =
l+eL(x,a)
where
and where
The other variables are as defined previously for AKU-68. If Z=Y=X=Q=O, this refers
to transfers between other combinations of transit modes, such as bus plus rail modes
with transfers or underground plus bus transfers and others. The mean number of
transfers is thus highest for the other combinations, e.g., bus to rail and rail to rail
(Table 4).
By use of the illustrated dummy variable technique, the parameters for the different
types of transfer can be derived as follows. The estimated, traditional policy param-
eters are significant. The transfer dummies are not at all significant, but the approach
indicates one possible way to use data to understand how different transfers are re-
garded by the travelers.
The derived values of travel time and waiting time are equal to those found in model
1 in Table 3. The (highly speculative) transfer-comfort values derived from the esti-
mates in Table 5 are given in Table 6.
Not much can be relied on in Table 6, but it is interesting to note that the two most
significant transfer types, bus to bus and other transit mode combinations, are valued
somewhat higher than the average transfer value (1.85 Swedish Kr) derived above from
the other models. This may indicate that transfer discomfort is regarded as more
onerous for bus modes and mixed bus-rail mode combinations than for pure rail modes.
There are several intuitive explanations that support this. First, transfers between
different bus lines take place at bus stops that have fewer conveniences compared with
commuter train and underground stations that are well-equipped with climate shelter
and other convenience facilities. Second, our values of transfer discomfort reflect
time spent walking to and waiting for transit. As buses in the greater Stockholm area
generally depart less frequently than the underground, the time spent waiting for a
transfer-on the average-is higher for the bus than for the underground. Because of
the mix of buses and cars in the street, buses also tend to give rise to a lower regu-
larity than rail modes on exclusive rights-of-way. This also tends to prolong time
spent waiting for transfers. Third, neither the different bus route schedules nor
Table 3. Trade-off values. Table 4. Structure of transfers (AKU-68).
Table 5. Empirical results from model based on type of Table 6. Transfer comfort values.
transit transfer.
Variable
Total
sample 0.0084 2.46 0.1136 11.02 0.212 3 3.37 0.6559 3. 54 1.1270 7.00 4.44 1.98
Low
income 0 .0068 1.30 0.1319 8.38 0.2421 2.32 0.1780 3.34 1.4335 6.20 3.10 1.83
High
income 0 .0089 1.93 0.0983 6.83 0 .2359 2. 85 0.6572 2.71 0.7153 3.03 5.45 2.40
Note: LC • logit coefficient. t·values indicate level of statistical significance,
•swedish Kr/hour, hSwedish Kr/transfer.
Table B. Modal choice and seating comfort (TU-71 ). Variable Model 5 Model 6
the locations of bus stops are usually synchronized. Distance and time spent waiting
for transfers tend to increase more for buses.
The transfer value with respect to other mode combinations is our most reliable
transfer value. As this type of transfer mainly refers to the combinations of bus plus
several rail modes or underground plus several bus routes, this comfort value should
be interpreted as reflecting either feeder bus or rail transfers in the outskirts of the
metropolitan region or city bus or rail transfers in the inner city of Stockholm. The
convenience level is generally rather low, especially at the suburban commuter train
and underground stations. The distance and the difference in the physical convenience
level between train station platforms and feeder bus stops are usually considerable.
In 1968, none of the stations was equipped with escalators or elevators.
However, the convenience level of transfers between different underground lines and
between the commuter train and the underground (located in the inner city of Stockholm)
is exceedingly high compared with the other transfers discussed. Transfers between
rail modes always take place across a platform or by elevators or escalators. These
enable commuters to be unaffected by weather, and there are other conveniences such as
kiosks and ~inishops.
Our point estimates given in Tables 5-and 6 are certainly compatible with these
observations, but, because of the extremely low significance levels associated with
these dummy coefficients, further research is needed to provide conclusive evidence
to transit planners.
The AKU-68 data are insufficient for deriving more information through this kind of
transfer differentiation. Consequently, there is an urgent need for research to develop
similar modal-choice models that emphasize differentiating comfort values into dif-
ferent types of transit use.
Effects of income level on trade-off values were also studied. Because of sample
limitations, only two income categories were used: (a) low income (< Swedish Kr 30 000/
year) and (b) high income (>Swedish Kr 30 000/ year). The logit model used for this
purpose was somewhat less detailed than the ones used previously in that total door-
to-door cost and time differences were used. We first estimated the coefficients for
the whole sample (1050 observations) and then estimated for each of the two income
groups. The results are given in Table 7.
Table 7 shows that the statistical level of significance is highly sensitive to sub-
divisions of the sample. The time difference variable of the low-income category is
significant at 80 percent only, which cannot be considered satisfactory. The other
coefficients are significant at 95 percent. Attempts to subdivide further by income led
to completely unsatisfactory results in regard to level of significance.
The conclusion to be drawn from the findings presented in Table 7 is that both value
of time and transfer seem to increase with income. This is quite in line with similar
studies in other cities, although little evidence so far exists in regard to the evaluation
of transfers as a function of income.
The final resuits in this paper are based on the TU-71 data, 352 observations in all.
These data made it possible to test the hypothesis that people who do not have a seat
during the trip value their travel time higher than people who have seats.
Two logit models were used for this purpose. The first (20) contained cost and time
differences (door-to-door) in addition to NOT and UOC. Theother one is somewhat
more general in that door-to-door times and costs enter separately for each mode,
except for the slight modification that public transit costs (Ct) is not included. The
reason for this is that the uniform 50 Swedish Kr/month season ticket applies to the
whole commuting population in the study area and can be considered a constant rather
than a variable.
The seating question is dealt with similarly in both models. A dummy variable, S,
takes the value one if the individual has a seat and zero if otherwise. It is incorporated
in the model by attaching it to the time variable in the following way:
(a1 + a2 • S)T t
49
or
If a sufficiently large sample were obtained it would have been useful, as an alterna-
tive, to stratify by S to see if there were any variations in values of time when the
stratification procedure was used in contrast to our dummy method. However, limited
sample size made such a stratification unsuccessful in terms of significant coefficients.
The estimates obtained via our two models are given in Table 8.
Our immediate reaction is that model 6 gives more reasonable results than model 5.
The level of significance in model 6 is much better and "the value of having a seat",
which is the difference between the two time values for public transit, seems more
reasonable in model 6. We found the seating value of model 5 somewhat high (20). In
model 6, this value is 65 percent of the same estimate in model 5. -
The absolute level of the time values is extremely high compared to those derived
in AKU-68 and other studies. A 30 percent increase compared with AKU-68 data for
inflationary and real income increases should be expected, but those values given in
Table 8 are just too high to be readily accepted.
The opposite conclusion is to be drawn for the transfer value. Instead of an ex-
pected 30 percent increase, the transfer value is slightly less than those of AKU-68.
The difference is, however, so small that it is tempting to conclude that the result has
added further support to our beliefs about the commuters' willingness to pay to have
one transfer less. This is not incompatible with the findings of Watson (22).
The analytic approach used in this study has yielded important empirical results.
Most models used are satisfactorily significant. In those models with no category
subdivision, almost every transportation policy parameter is highly significant and
supports the following conclusions.
The travel-time value differs substantially with the type of trip segment. The
models estimated reveal a significantly higher value of the time spent waiting for
transit than of the corresponding line-haul time on transit.
The commuters tend to be conscious of waiting time. When a more realistic as-
sumption about actual waiting time is used instead of half the headway, the waiting-
time value becomes between 7 to 12 times the value of the in-vehicle travel time. Our
findings indicate that the waiting-time value depends on the level of service of the
transit system in terms of transit headway.
The level of the travel-time value also depends to a large extent on the in-vehicle
comfort in terms of seating opportunities. The overall travel-time value decreases
substantially when commuters can enjoy a seat as compared to when they cannot.
A transfer-comfort value is deduced and amounts to about 1.90 Swedish Kr/transfer.
In terms of travel time, the commuter is equally well-off with 1 less transfer and 24
min longer for door-to-door time. This transfer-comfort value tends to be valued as
more onerous for such transfer combinations as bus to bus and bus plus several rail
modes than for an average transfer or transfers between rail or underground modes.
·When the sample is stratified into different categories such as income groups,
travel-time groups, and types of transfers, the level of statistical significance drops
markedly in the subgroups compared to the overall models formulated.
For income groups, the results support the hypothesis that the travel-time or
transfer-comfort value increases with income.
These results may be useful in various transportation planning situations:
The results of this analysis indicate an urgent need for further research and de-
velopment. Existing and future traffic surveys should be better adapted to handle
planning problems that are relevant for policy-making authorities, e.g., by permitting
relevant modal- and traffic-choice analyses in which subdivisions into trip segments,
multiple-choice situations, and different socioeconomic groups are practicably possible.
Inasmuch as one of the comfort and convenience variables estimated in these models
(number of transfers) seems to reflect a rather composite factor of discomfort, it is
necessary to focus more investigations on model components. Among these components,
number of transfers may reflect the time spent walking to a transfer point and waiting-
time discomfort as well as the uncertainty of getting a seat in the vehicle entered at
the transfer point.
ACKNOWLEDGMENTS
We wish to acknowledge the following persons and agencies. Kristina Lennartsson and
Lars Wikstrom of the Department of Planning and Building Control in Stockholm; Ole
Havard Hansen of M~re og Romsdal Regional College; Nils Bruzelius of the Stockholm
County Council; AB Stor-Stockholms Lokaltrafik; the National Central Bureau of Statis-
tics; Ulf Nyberg of the Joint Planning Board of Stockholm; J. G. Cragg of the University
of British Colombia; and David A. Hensher, Commonwealth Bureau of Roads, Melbourne,
Australia. Finally, we wish to thank the Economic Research Fund of the Bank of Nor-
way and the National Swedish Council for Building Research for financial support, which
made our international cooperation much more efficient. The entire responsibility for
the contents including all conclusions derived from the analytic results presented
rests solely on the authors.
REFERENCES
1. The Role of Parking Impedances for the Journey to Work. Allmanna Ingenjt.Jrsbyran
AB (AIB), Stockholm, Sept. 1974.
2. D. Brand. Travel Demand Forecasting: Some Foundations and a Review. HRB
Special Rept. 143, 1973, pp. 239-283.
3. N. Bruzelius. Economic Theories of the Allocation of Time, a Survey. Faculty of
Commerce and Social Science, Univ. of Birmingham, Master's dissertation, 1972•
unpublished.
4. A Disaggregated Behavioral Model of Urban Travel Demand. Charles River
Associates, CRA Rept. 156-2, Cambridge, Mass., 1972.
5. F. X. DeDonnea. The Determinants of Transport Mode Choice in Dutch Cities,
Some Disaggregate Stochastic Models. Rotterdam Univ. Press, 1971.
6. A. C. DeSerpa. A Theory of the Economics of Time. The Economic Journal, 1971.
7. R. L. Gustafson, H. N. Curd, and T. F. Golob. User Preferences for a Demand-
Responsive Transportation System: A Case Study Report. Highway Research
Record 367, 1971, pp. 31-45.
8. S. Hansen. The Structure of Disaggregate Behavioral Choice Models. Resource
paper for EF Conference on Issues in Behavioral Travel Demand and the Valuation
of Travel Time, South Berwick, Maine, 1973.
9. D. A. Hensher and S. J. McLeod. Comfort and Convenience: Theoretical, Con-
ceptual, and Empirical Extensions of Disaggregate Mode Choice Models. Common-
wealth Bureau of Roads, Melbourne, 1974, unpublished.
10. J. B. Lansing and G. Hendricks. How People Perceive the Cost of the Journey to
51
•THE simulation package for conventional urban transportation planning is used pri-
marily for forecasting demand for transportation facilities and systems in urban areas.
This package has been applied by New York State Department of Transportation in the
preparation of its long-range systems plans in upstate New York cities. The calibra-
tion process for several of the simulation package models is quite extensive and in-
volves massive amounts of data. Accordingly, each of the urban studies in upstate
New York conducted an extensive survey program as part of its initial activity. Be-
tween 13,000 and 20,000 households were interviewed in each area, and the resulting
travel files contained more than 100,000 trip records.
There is a general feeling on the part of the profession, however, that the use of
such data in these models is inefficient at best and that, as the transportation planning
studies move into the continuing phase, the emphasis of systems planning has shifted
from concentration on urban areawide system plans to more detailed analysis of cor-
ridor locations and questions of modal alternatives within corridors. Furthermore,
a general tightening of the budget has made it infeasible to collect the same amount
of information as was done in the first stages of these studies. These constraints re-
quire, therefore, that new modeling techniques be developed that are capable of
operating at the corridor and systems level and that use much smaller data sets.
The problem is how this may be done within the contexts of the simulation process
described without unduly sacrificing the reliability of travel forecasts.
In the last 2 years, considerable research on disaggregate models has been under-
taken. The most important characteristics of these procedures for our use are that
(a) they operate on considerably fewer data than do conventional modeling procedures
52
53
and (b) they use specific mathematical fwictional forms from which estimates of travel
demand elasticity may be obtained external to the model. These advantages are ob-
tained by using calibration methods that operate on individual trip records and thereby
use all of the information in travel data sets for model development. Conventional
model calibrations, however, generally aggregate trip information before models are
constructed and thereby lose much of the information.
Use of disaggregate models as part of the repertoire of transportation planning tech-
niques, however, has not followed easily from their theoretical development. Most of
the research on disaggregate models has been confined to the calibration of mathemat-
ical forms from small data sets and to the extension of mathematical model theory.
Much less emphasis has been placed on the use of these tools in transportation system
planning. This is because a typical disaggregate model ideally would require specially
collected or augmented data information, and most state and regional transportation
planning agencies have available to them only conventional home-interview (HI) survey
information. Another problem is that aggregation is generally necessary if these tools
are to be applied to aggregate zone-to-zone flows for computation of modal shares.
MODELS COMPARED
The conventional modal-split procedure used at New York State DOT determines the
zonal percentage of transit use based on the transit-to-automobile service ratio and
the air distance of each zone pair. Four trip categories (by purpose and automo-
bile availability) are used in which the splits are performed within each category (4).
Percentages of zone-to-zone transit use are multiplied by the volume of travel so that
transit and automobile trip volumes can be obtained. Various post-split aggregations,
such as the district-to-district summaries and other relevant information, are also
provided as part of the model output (!, .§_).
Currently, to obtain the zonal transit use percentage, New York State DOT works with
a table "look-up" (e.g., a response surface) that arrays the transit trip percentage by
the air distance and the transit-to-automobile travel-time ratio. Figure 1 shows
the response surfaces for each trip category: long trip, automobile available (LAA);
long trip, no automobile availabl.e (LNAA); short trip, automobile available (SAA); and
short trip, no automobile available (SNAA). Long trips are basically work trips, and
short trips are basically shopping, social, and recreational trips. A more detailed
definition of these trip categories is documented in(~).
Figure 1. Response surfaces.
work; no auto available non ~ worki no atitlo available wortl; auto available non-worki auto ovailablt
11rvtc1 quallt1 (rat IOI of timt via transit to lime via auto)
N
@
~ I
I
I
@ @ !
r ·j
I
i
!
@ I
!
I
i
@ l
I
I
j
I
!
55
The main shortcoming of the response surface method is that a large data set is
required to assure sufficiently stable estimates of the transit proportion in each cell,
especially those near the edges of the surface. The set of response surfaces for the
Niagara Frontier, for instance, were constructed from over 100,000 individual trip
records.
The disaggregate modeling technique has two advantages. First, it requires a much
smaller sample size than aggregated methods. Because the data are treated at the
individual or household level, all of the information contained in each trip record is
fully used when the models are calibrated. Consequently, the disaggregated models
can be estimated on the basis of a small sample and are not restricted to zoning.
The second advantage is that the method deals with the most basic travel unit: the
individual or the household. In other words, although the aggregate method searches
for the associative relationships between travel demand and the various explanatory
variables, the disaggregate method attempts to extract the traveler behavioral pattern
associated with making a trip. Therefore, the resulting models promise transferability,
temporally as well as spatially.
This particular study aims to capitalize mainly on the first advantage of the disag-
gregate modeling approach, the savings in data requirements. (Because of the peculiar
data information used in this paper for model estimation, the estimated models are not
believed to be truly behavioral, and therefore their transferability is questionable.) It
is intended that, instead of relying on the response surfaces that require a large amount
of sample data, we can compute the percentage of zonal transit use by applying a set of
disaggregate modal-choice models, which can be estimated on a much smaller sample.
The disaggregate model form used in this study is the binary logit model; i.e.,
1 (1)
p k : 1 G( X )
+ e
where Pk is the probability that a certain travel mode is chosen by an individual and
G(X) is usually a linear combination of transportation level of service and individual
socioeconomic characteristics.
DATA
The 1962 HI survey for the Buffalo region (Fig. 2) contains 103,328 individual trip
records. Inasmuch as one of the objectives of this study is to determine if one can
feasibly apply disaggregate techniques to the relatively small 1973 update travel sur-
vey collected in the Niagara Frontier, this study uses, as a data base, a subsample of
the 1962 HI survey that has a sample size comparable to that anticipated in the 1973
survey. Accordingly, trip records in the quasirandom 12 percent subsample, con-
taining 12,916 records (Table 1), were drawn from the original file. [A detailed de-
scription of the original sample and the subsample is discussed elsewhere(~).]
APPROACH
For each category of trips, binary logit models (Eq. 1) were developed. So that an
exact comparison between the response surface method and the disaggregate approach
could be made, only two variables, air distance and service ratio, were used in con-
structing the modal-choice models.
As a typical, conventional travel survey, the 1962 HI file contains zone information
56
(2)
where Var [G(X)J is the within-zone (or within-interchange) variance of G(X), and its
value depends on the variances and covariances of modal attributes. [Another method
of binary-choice model aggregation is found elsewhere (10).J
The need to aggregate the estimated choice models has been questioned because
zonal values will be used for the explanatory variables and, therefore, the estimated
models will not be disaggregate at the individual level. Instead they will be zonal
models and will not require aggregation for zonal prediction. The other interpretation
is that, even though the individuals are assigned to zonal service levels, the models will
still be estimated on individual records by individually observed modal choices. There-
fore, the models will still be disaggregate at the individual level and, thus, will require
aggregation for zonal prediction. Recent research results indicate that disaggregate
models may .be obtained from a combination of disaggregate and aggregated data if the
variables are not correlated (7). This seems to support the latter interpretation. Jn
this study, both of these interpretations are considered empirically. Prediction results
based on both the nonaggregated and the aggregated estimates are compared.
EVALUATION CRITERIA
The choice models in this study are evaluated on reasonableness, statistical indicators,
and performance prediction for transit systems. This will be done by inserting· the
models (i.e., calibrated equations) and other related information into the modal-model
program, by replacing the response surfaces, and by applying the program to zone-to-
zone flows in the Niagara Frontier. Model performance will be tested by the following
methods:
L (Tpredicted - Tabserved )
2 1la
all
districts
RDS = T2
(-:l)
L
all
districts
This also reflects the smoothness of the deviation fluctuation from district to district
over the entire study area.
4. Comparison of predicted and observed transit trips in districts.
RESULTS
The disaggregate method for modal-split prediction is developed over a number of
modeling and trial stages. In each stage, the prediction results were evaluated ac-
cording to the set of criteria established earlier. Logical improvements in the mod-
eling approach were made after each stage, and the revised method was investigated
in the next stage. This process was continued until a method was derived that was
able to predict as accurately as the existing response surface technique and that, at
the same time, was based on sound theory. The approaches taken at the various mod-
eling stages and their prediction performances are described in the following.
12 Percent Method
An initial approach was taken to estimate logit models based on 12,916 trip records.
The coefficients of the four models are given in Table 2.
The coefficients of the service-ratio variable in these models have positive signs,
which are consistent with intuitive reasoning. The model constant is largest for SAA
and smallest for LNAA. This indicates that the travelers are most inclined to use an
automobile for SAA trips and least inclined for LNAA trips. The model constant and
the service-ratio variable are statistically significant at 0.95 level of confidence in all
four models; however, the air-distance variable is not significant in the LAA and
LNAA models.
These models were inserted into the on-line modal model, which was then applied
to the Niagara Frontier Transportation Study (NFTS) regional data. [It is assumed for
this test that the probability of transit use for a zone or zone pair is also the proportion
of transit use for the zone pair, (e.g., no aggregation bias).] Whereas the response
surface method overpredicts the number of transit trips by 0 .17 percent of the total
person trips, the 12 percent method underpredicts the number of transit trips by 1.35
percent of the total person trips (Table 3).
This is a considerable discrepancy, considering that the total number of person
trips in this region is in excess of 2 million. The sum of deviations and the RDS also
reveal the relatively poorer prediction performance of the 12 percent method.
The cause of such undesirable performance is that there are interzonal and intra-
zonal groups of trips in this region, and over 8 percent (or""' 1,800) of the trips of the
12 percent sample are intrazonal. For these trips, the values of the air-distance
and service-ratio variables are both assigned values of zero.
Because the response surface method only deals with the average values within each
cell, intrazonal trips do not materially affect its predictive performance. However,
inasmuch as the disaggregate method takes into account each trip record, these trips
must be eliminated before reasonable models can be calibrated.
As an extension of these models, the interzonal and intrazonal trips were separated
from the set of trip records of 12 percent of the sample. Furthermore, among the
interzonal trips, those that had a service-ratio index greater than 6.0 were also
sorted out. This is because a relevant choice situation between the two alternative
travel modes is not expected to exist when the transit travel time is more than six
times as much as the automobile travel time.
The remaining 10,004 trips were categorized according to their trip classifications,
and a logit model was estimated for each trip type. The air-distance variable
turned out to be insignificant in all the estimated models. Therefore, a set of models
were reestimated by using only the service-ratio variable. The coefficients of these
models and the statistical indicators are given in Table 4.
The coefficients of the service-ratio variable in these models also have positive
signs. The model constants indicate the same rank order toward choosing the transit
mode as that indicated by the previous models. However, the values of the coefficients
are significantly different in the two sets of models. In particular, the values Qf the
coefficients of the service-ratio variable are 1.2 to 3.3 times greater than the corre-
sponding coefficients in the previous models.
For the intrazonal trips, two constants were used to represent the transit proportion.
For zones in the inner urban area contained within the first four traffic rings, the tran-
sit pr oportion is 3.5 percent. For zones in the outer urban area where there is little
transit service, the transit proportion is 0. 73 percent. These proportions are empir-
ically derived from the 1962 data.
The interzonal models and the constants for intrazonal transit use were inserted into
the modal model, which was again applied to the NFTS regional data. The prediction
is given in Table 3.
The predictive performance
of the disaggregated approach
Table 3. Model performance for response surface, 12 percent, is improved considerably over
disaggregated, and aggregated methods. the 12 percent approach. For
the area-wide transit use, this
Absolute method underestimates the tran-
Transit Error• Sum of
Prediction Method (percent) (percent) Deviation RDS sit usage by 0.84 percent, as
compared to an underestimation
Response surface 7.59 +0.17 24.1 x 10 3 0 .131
12 perc ent of sample 6.07 3
-1.35 63 .3 x 10 0.344 of 1.35 percent by the previous
Disaggregated 6. 58 -0.84
3
31.8 x 10 3 0 .166 approach. Nevertheless, its
Aggregated 7. 53 +0.11 26. 2 x 10 0 .128
predictive ability is still not
8
0bserved area-wide transit use is 7.42 percent. satisfactory when compared with
Table 4. Coefficients for Constant Service Ratio
interzonal logit models.
Trip Type LC t-Ratio LC t-Ratio Pseudo R'
152
76
@) 723
355
125 5 60
0
40
@ 53
13
103 37
lnlrozonol Tr ips
1,800
that of the response surface method. Further attempts to stratify trips by character-
istics of the trip ends, e.g., urban interchange (trips with one end in Buffalo or Niagara
Falls) and suburban interchanges (the rest of the interzonal trips), yielded, at best,
predictive accuracy similar to that of this approach.
For the various approaches investigated up to this point, it has been assumed that the
choice models are applicable directly to the zones. Therefore, proportion of transit
use for a certain zone is the same as the disaggregate transit probability at its zonal
means, Pk(X). Results of these attempts show an underestimation of transit use over
the entire region.
As mentioned earlier, another interpretation of these choice models is that they are
disaggregate at the individual level and therefore require aggregation for zonal predic-
tions. In the following tests, an aggregation term is added to the disaggregate proba-
bilities as shown in F,q. 2. The a priori lmowledge of this aggregation term is that, for
noncorrelated explanatory variables, it has a positive value if P.(X) is less than 0.5.
The magnitude of this term depends on Var [G(X)].
The procedures taken in this stage were almost identical to those taken in the last
stage, except that information about Var [ G(X)] and the aggregation term was also
inserted into the modal model. Because the set of estimated choice models contain
only the service-ratio variable, the value of Var [G(X)] is
2
Var [G(X)] = {J, r · Var (X.r) (5)
where Var (X.,) is the within-interchange variance of the service-ratio variable, and
{3,, is its model coefficient.
Two assumptions (or approximations) are made in computing the values of Var (X,r).
The first is that the within-interchange variance of the service-ratio variable is similar
to the between-interchange variance of this variable. Although previous studies (3) in-
dicate that the within-zone variance is greater than the between-zone variance, tills
assumption is necessary because the Niagara Frontier data in zone-to-zone form do
not contain any within-interchange information. The second assumption is that many
zonal interchanges have approximately the same Var (X.r). This assumption simplifies
the prediction process because it would be infeasible, in practice, to compute the vari-
ance for each zone pair. Intuitively, the validity of these two assumptions may be jus-
tified if a single Var [G(X)] is computed for trips of similar service characteristics.
On examination, the urban interchanges (e.g., trips with either origin or destination
in Buffalo or Niagara Fails) and the suburban interchanges (e.g., the rest of the inter-
zonal trips) appeared to be the two trip groups within which trip characteristics are
similar. The means and variances of the service-ratio variable are respectively 1.870
and 1.076 for urban trips and 3.523 and 1.196 for suburban trips.
Consequently, two values of Var [G(X)] are computed, one for urban interchanges
and the other for suburban interchanges. This information and the choice models were
inserted into the modal model. Results of this prediction approach are given in Table 3.
CONCLUSION
The predictive ability of the aggregated approach compares closely with that of the
response surface method. The response surface method overpredicts areawide tran-
sit use by 0 .17 percent; this method overpredicts it by 0 .11 percent. The sum of devi-
ation and the RDS indicators also show the predictive accuracy of this method to be as
good as that of the response surface method.
A comparison of transit trips by origin district is shown in Figure 3. These results
61
show that the disaggregate method of approach can indeed produce, by using the same
type of information, as plausible predictive results as the response surface method,
but it uses only 12 percent of the original sample. Furthermore, the empirical results
of this study suggest that such modal-choice models require aggregation corrections
for application to the zonal level. ·
USE IN PRACTICE
A number of modeling approaches have been investigated in developing an accurate and
workable disaggregate method for modal-split prediction. The best approach involved
the following procedures:
1. Separate the regional trip records into interzonal trips and intrazonal trips;
2. Use constants to represent proportions of intrazonal transit use for intrazonal
trips;
3. Eliminate those interzonal trips with service ratios greater than 6.0 and develop
logit models by trip classification;
4. Separate the interzonal flows into urban interchanges and suburban interchanges
and compute the variance of the logit model index, G(X), for each type of interchange;
and
5. Insert the constants for intrazonal transit use, the interzonal logit models, and
the values of urban or suburban interchange into the on-line modal model and apply it
to regional flows.
explored and studied, the current application appears to offer a valid and efficient way
to approach travel demand forecasting for large urban areas.
ACKNOWLEDGMENTS
We wish to thank Antti Talvitie of the University of Oklahoma, Richard Westin of North-
western University, and Patricia L. Milic for their valuable comments. Special thanks
are extended to James McDermott and his staff whose continuous assistance in data
processing is greatly appreciated. The authors alone remain responsible for errors
of fact, concept, or opinion. This research was supported by the Highway Planning
and Research Program of the Federal Highway Administration.
REFERENCES
This paper suggests economic criteria for choosing among estimates of the
value of time. Four assumptions that are implicit in most models of the
value of time are identified and discussed. The accuracy of an estimate of
the value of time depends on whether the data and model underlying the
estimate take into account the unrealistic nature of these assumptions,
either by excluding observations for which the assumptions are especially
incorrect or by modifying the model so that the assumptions may be relaxed.
The first assumption of modal- or route-split models is that travelers believe they
have perfect information about the characteristics of all alternatives open to them.
As will soon become clear, this assumption can be avoided through a rather unorthodox
specification of the modal- or route-split equation. Nevertheless, it is usually im-
plicit in studies attempting to estimate the value of time. Such studies [with one ex-
ception known to me (~)J implicitly assume that the responsiveness of travelers to
63
64
changes in time and money costs of the alternate mode is independent of the frequency
with which tnvelers make their trips. This assumption would be valid if information
about the alternate mode were free, and required no experience or time-consuming
inquiries. In this case, travelers on their first trip would acquire as much information
as t r avelers who makethe trip everyday. If, however, such information is costly, the
information possessed by each traveler would depend on the benefit of acquiring such
information; this in turn is larger the more often the commuter makes the trip. One
would then expect more frequent travelers to be better informed and, thus, more re-
sponsive to changes in the time and money costs of the alternate mode.
What is the effect of this assumption, if it is not entirely correct? The less frequent
commuters will be less responsive to changes in the differences in time and money
costs so that a sample of a relatively high proportion of less frequent commuters will
yield coefficients that are smaller in absolute value for the two crucial variables, the
differences in time and money costs. If the coefficient on the difference in time cost
(AT) declines in absolute value at a faster rate than the coefficient on the difference in
money cost (AM), the less frequent commuters will appear to have smaller values of
time. An earlier study (2) found that, on the contrary, the coefficient on AM went to
zero at a faster rate so that the less frequent commuters appeared to have higher
values of time. Regardless of which is generally the case, there is no reason to ex-
pect travelers' values of time to differ depending on how often they make a J;Rrticular
trip. Effects such as these are spurious and result from the assumption that ther e is
perfect information when, in fact, travelers' information is imperfect.
This assumption will be roughly correct if all travelers in a sample are daily
travelers. Such commuters would have a large incentive to acquire information about
the alternatives open to them because the benefit of such information would be quite
significant. This suggests one criterion for choosing between estimates of value of
time: If the study asks the individuals in the sample how often they made the trip in
question and then examines only daily commuters, the value of time estimate will be
less biased. Not even among work trips are all travelers daily commuters; therefore,
it is important not t o simply assume the problem is solved in all s tudies concentrating
on work trips . In an earlie r s tudy (2 ) I found t he bias resulting from t his assumption
to be about 50 percent. -
For nonwork trips, it is usually difficult to find individuals who make the trip daily.
As long as a few such individuals can be found, however, it is not necessary to restrict
the sample size so that the bias discussed can be avoided. One can build into the modal-
or route-split equation interaction terms that control the effect of the frequency of the
trip on the coefficients of AT and AM. A simple specification is
L(x) (1)
where F is the frequency with which the individual makes the trip, and L(x) is the logit,
which, in this model, determines the mode choice. The ellipsis represents whatever
other variables, such as driver's income, which seem to be relevant to the mode choice.
This equation would actually be estimated as
L(x) (2)
where F~M and FAT are the interaction terms. The value of time for commuters of
frequency F would then be -(as+ 34F)/(a1 + a2F). By inserting the value of Finto Eq. 2,
which represents daily commuters, one can obtain a more accurate estimate of the
value of time than if the interaction terms were left out and nondaily commuters were
included in the sample. This method has been applied (2) with good results. The
coefficients of the interaction terms and the other variables were statistically significant,
65
and those of the predicted signs were interesting; plausible estimates of the value of
time resulted. For off-peak work trips, the value of travel time was estimated as ap-
proximately $1.91 per adult per hour, compared with $2.95, which was the value when
the effect of imperfect information was not taken into account.
A second assumption of modal-split models is that the two modes have identical levels
of comfort and other intangible characteristics. This assumption is necessary so that
the single value of time derived from such models will make economic sense. To see
this clearly, one must recognize that the value of time savings on any particular mode
is composed of two parts: the opportwiity cost of travel time, which is independent of
the mode chosen, and the utility or disutility of traveling, which is specific to the mode
chosen. The more positive the utility (or less negative the disutility) of traveling is,
the lower the amount a traveler will pay to save travel time will be. Thus, if two modes
have different levels of comfort, they will have different values of time savings. In
this case, to derive one number and call it the value of time between the two modes is
meaningless from an economic point of view. One would be estimating neither the value
of time on one mode nor the value of time on the other mode, but rather a compromise
would be reached between the two. The estimate would be closer to the value of time
on the mode that the majority of the individuals in the sample use. Thus, it would be
inappropriate to use the estimate of the value of time when time savings on the mode
are valued. This is less commonly used in the sample.
There are three ways in which tWs problem can be avoided. The first is to examine
not only choices within a particular mode but also those between two routes or lines of
travel. In this case, the assumption of identical, intangible characteristics will be
roughly correct. The difficulty is that it is often impossible to find a route-split situa-
tion for the mode desired; therefore, one is forced to resort to modal-split situations.
A second alternative is to reformulate the model of the value of time so that the as-
sumption is unnecessary (!). This can be done by considering separately the travel
times of the two modes by estimating different coefficients for the two travel times.
The difficulty here is that the two travel times will be highly correlated with each other,
and this makes estimation extremely difficult. It is not sufficient for one to simply
insert a new variable into the modal-split equation and purport to measure the comfort
level of the mode, for, by doing so, one implicitly assumes that the coefficient on aT
(and thus ~he value of time} is independent of this comfort level. If, however, the
comfort level could be accurately measured (which is doubtful), it would be appropriate
to multiply the difference in this variable by aT and thus capture the effect of differing
comfort levels on the value of time.
The third and most practical alternative is to alter the overall ratio of observations,
e.g., between the two modes, so that most of the observations are on the mode whose
value of time one wants to measure. In this way the resulting estimate will approxi-
mate the desired value of time. This suggests a second question the user of estimates
of the value of time might ask before choosing which estimate to use: On which mode
did the majority of observations appear? The answer will, in general, be crucial to
an intelligent interpretation of the estimate of the value of time.
Note that if the assumption of identical, intangible characteristics is approximately
correct, altering the overall ratio of observations, e.g., betwe.en the two alternative
modes or routes, will leave the estimate of the value of time approximately unchanged.
This fact (2) is convenient for analysts attempting to estimate the value of time because
it means that they need not sample the population being studied s0 that the overall
modal- or route-split ratio exactly equals the ratio in their samples. For large
samples, it is sufficient that each subpopulation is randomly sampled (the two sub-
populations are the groups using each mode); an unrepresentative overall modal-split
ratio will affect only the constant term in the modal-split equation.
66
COMPLETE ACCOUNTING OF ALTERNATIVES
OPEN TO TRAVELERS
The third implicit assumption of modal- and route-split models is that they account
for all alternatives open to travelers. This assumption is especially s uspect during
peak hours when traveling at alternative times rather than on alternative modes be-
comes an important option to the typ 'cal commuter. During peak hours, travel times
change so rapidly that leaving a few minutes earlier or later can significantly affect
the time cost of the trip. When this is true, a change in the difference in travel time
between modes will have a second effect besides the usual one of a substitution between
modes: Substitution will occur between times of travel. If this latter substitution is
not taken into account (and available data generally preclude properly accounting for
it), the coefficient on .:iT will be misleadingly small. It will fail to .reflect the signif-
icant way in which commuters reveal their values of time. The result is tbat the value
of time during peak hours will, 'i n general, be underestimated.
The relevance of this problem is essentially restricted to peak hours. It is also
much less serious for rapid transit than for automobile travel, because variation of
travel times on rapid transit (e.g., subways, not buses) will usually be smaller than
for travel times on the highway. This suggests that samples in which subway or rail-
road users dominate will be less subject to this source of bias. This, of course, is of
little help to those who desire accurate estimates of t he value of time for peak-hour
automobile travel. Unless studies appear that use data broken down by time of day of
trip as well as by the mode or the route and that use sophisticated multinomial choice
models, one must sim,ply keep in mind a possible source for a negative bias in peak-
hour estimates. [Even so, available evidence suggests that the value of time during
peak hours is more than twice the value of time during off-peak hours for the same
type of trip (2). This difference is evidently due to the greater disutility of traveling
dur~ng peak !'lours.]
The fourth implicit assumption of models attempting to estimate the value of time is
that the differences in (or ratios of) time and money costs in the analyst's data set are
accurate measures of the differences in (or ratios of) time and money costs perceived
by travelers. This is a troublesome assumption because, as was argued earlier,
travelers' information about .:iT and 6.M is imperfect. If either variable changes ob-
jectively and an equal change does not occur in the perceptions of the people in the
sample, these individuals will respond diffeTently from the way they would have
answered had the change been fully perceived. In this case, the response that reflects
the individuals' values of time will be confounded with the response (or lack of it) that
reflects their misinformation. This type of problem can arise with either measured
data, in which .:iT and !SM are measured with engineering estimates, or with reported
data, in which the individuals report .:iT and .:iM. In particular, it is misleading to call
reported data perceived data because the individuals may be misr eporting their per -
ceptions.
This assumption is distinct from the assumption that individuals in the sample think
they have perfect information of .:iT and 6.M (i.e., that they are certain of their percep-
tions). Uncertainty may be present whether or not misperception occurs. Similarly,
misperception may occur whether or not travelers are uncertain.
The problem here is that the travelers' perceptions are not directly observed by
analysts of the value of time. This should not prevent discussing thei.r role in travelers '
behavior- the analysis, after all, rests on the existence of another variable that is not
directly observed: the traveler's value of time. But the fact that one cannot observe
travelers' perceptions makes it difficult to do more than speculate on which data, mea-
sured or reported, are more appropriate for estimating the value of time. Such
speculations are discussed elsewhere (~ ~ i). Here I wish to discuss ways in which
67
SUMMARY
This paper identified four assumptions implicit in most studies attempting to estimate
the value of time. These assumptions were shown to have important implications for
which data and which model specifications are more appropriate for estimation of the
value of time. In particular, biases tend to result when
1. Nondaily travelers are included in the sample, unless the model is specified to
take account of their uncertainty of mode characteristics;
2. The two modes examined differ significantly in intangible characteristics, unless
this also is properly accounted for in the model;
3. Cross-time substitutions take place, particularly during peak hours; or
4. Significant misperception by travelers occurs when measured data are used.
One can often avoid these biases, and estimates that do so can be distinguished from
those that do not if the sources of these errors are kept in mind.
ACKNOWLEDGMENT
This research was supported by the National Science Foundation. I wish to thank Gary
Becker, Arnold Harberger, Thomas Lisco, Herbert Mohring, Margaret Reid, and the
participants in an Urban Economics Workshop at the University of Chicago for helpful
comments on an earlier version of this paper.
REFERENCES
The organizational units and the chairmen and members are as of December 31, 1974.
69