Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

T~JPR Ru.. Vol. 10. w. 299410. Pergmnoa Press.

Rinted h orutBritain

AN ATTITUDINAL MODAL CHOICE MODEL

WILPREDW. REcrcr& and THOMASF. GOLOB


Transportation
and Urban AnalysisDepartment,ResearchLaboratories,
GeneralMotors Corporation,Warren,
Ml 48090,U.S.A.

(Receiued 25 August 1975;in reuisedform15March 1976)

Abstract-Explanatoryvariables in this modelof urban residents’modechoice behavior for the home-to-worktrip


are attitudinalratingsby survey respondentsof descriptiveattributesof modesperceived to be availableto them.The
modelmethodologyis presented in threeparts:individualsare clusteredwith respect to homogeneityof theii choice
constraints;the covarinnce structureof attitudinalvariablesis characterizedthrough development of latent factors
describingrespondents’perceptions of choice alternatives;and logit probabilisticchoice models are estimated in
termsof subsets of descriptive attributeratingschosen to representlatentperceptionfactors. Empiricalresultsfrom
applicationof the model in the Ottawa,Ontario/Hull,Quebec CanadianNational Capital Region are extremely
encouraging.Goodness-of-fit summan‘es indicate model performancebetterthan that obtained using conventional
logltmodelsbased upon perceived timeandcost data. Moreover,it is proposedthat the attitudinalmodels provide
valuableinsights concerning the importancesof factors underlyingthe choice decision structureof the sample
population.

lNlRODUCl’lON fully employing attitudinal variables can be attributed to


Signiticant research effort has been spent in developing two omissions in previous research. Firstly, the past
travel demand forecasting models which arc “behavioral’* travel experiences and current circumstances of individu-
in the sense of describing the process by which als can affect individuals’ attitudes much to the same
individuals choose when, where, and how to travel. Many extent that attitudes are in turn determinant in current
of these recently developed travel demand models employ choice of travel alternative. This potential circular
methods developed in the social and biological sciences to causality is addressed herein through a segmentation
model decision problems involving acts or outcomes scheme in which travelers facing ditIerent choice con-
which are not continuous, as in traditional consumer straints are allowed to possess different attitude-choice
economic analyses, but which are indexed in countable relationships. Further refinements in which individuals’
sets. McFadden (1974) provides a state-of-the-art review travel histories or general preconceptions are taken into
of this general class of quanta1 choice models. account are the subjects of future research.
Sign&ant effort has also been spent in developing Secondly, the interrelationships between the explanat-
methods to assess individuals’ attitudes and to determine ory variables in most travel demand models are seldom
underlying structures of these attitudes for use in urban explicitly structured. Expected high multicolliiearities
transportation planning. As pointed out elsewhere (Golob, between attitudes and objectively measured or perceived
1973; Golob and Dobson, 1974; D. G. McFagden, 1975) systems variables can in these cases lead to rejections of
much of this research has its foundation in psychology the abilities of attitudinal variables to account for
and often follows guidelines established through exten- variances which are residual to the explanation of the
sive applications in the market research field. systems variables. Moreover, interrelationships among
However, with a few exceptions (e.g. Costantino, attitudes themselves can be expected to cloud attitude-
Golob and Stopher, 1975; Nicolaidis, 1975), studies in behavior relationships if such interrelationships are left
behavioral demand modeling and those in attitudinal unstructured. This is especially true when attitudes are
analyses have remained separate. In fact, in a recent measured UBhtg a barrage of often redundant survey
glossary accompanying a compendium of urban travel questions. This problem is addressed herein through the
demand forecasting state-of-the-art reports the terms structuring of explanatory variable interrelationships in
“behavioral” and “attitudinal” were assigned meanings such a manner as to define not a single model but an entire
which are unmistakably mutually exclusive (Brand and family of models.
Manheim, 1!?73,page 307). The scope of the present application is limited to a
This distinction between attitudinal sad behavioral binary mode choice situation in which the two alternatives
approaches is contrary to well-accepted theories in social are private automobile and conventional urban bus
psychology (e.g. Fishbein, 1%3) and to applications of a service. Nevertheless the models have been designed to
range of attitude-behavior models in marketing research accommodate any number of choice alternatives. The
(e.g. Sheth,.,,l!974). The present research attempts a application is further liited to the home-to-work trip;
reconciliation of these two approaches. Spec&ally, applications to other travel purposes could dictate use of a
attitudes are shown to be useful explanatory variables in Merent set of attitudinal and choice-constraint variables.
quantal choice models.
Possibly the dearth of travel demand models success- DATA

The data used in this study were collected through an

ton leave from: Department of Civil Engineering, State attitudinal home-interview survey administered in the
University of New York at Butfalo. NY 14214,U.S.A. Regional Municipality of Ottawa-Carleton, Ontario,

299
300 W. W. RECKERand T. F. GOLOB

Canada and the Outaouais Regional Community, Quebec, Table 1. Twenty-five descriptive
work trip attributes and their
Canada. These two areas effectively comprise the abbreviated forms
National Capital Region of Canada, with a total popula-
tion in 1971 of approximately 610,000 persons. The
attitudinal survey was au integral part of the Regional
Municipality of Ottawa-Carleton Transportation Planning
Study initiated in 1971. The survey was conducted in 1973
and 1974.
The primary sources of data for the present study are
ratings by survey respondents of their alternative
automobile and bus modes. The respondents rated their
perceived satisfactions on a set of twenty-five descriptive llIlltW EXPOSUllE TO WLULV CKWGIIY. 1 BM)tLV
CROWtNG
attributes with respect to their home-to-work trip. There LCUDUT-OF-WCKEV COST 1 OUT-OF-FGCKET COST

were 543 usable survey questionnaires after removal of LOU RlDIKG TIlL RI0I.GTIM
LOHllALKlM
jTII(E WALKING TIME i
incomplete responses. The page of the questionnaire LW u4111)16
WE WAtTIll TtltF
eliciting information on evaluation of the automobile for OFfUFfU~p:IIl llE6lAUDTALKm
OPFORTUNITV
the work trip is reproduced in Fig. 1; there was a similar
page for the bus mode. The twenty-five descriptive
attributes are listed again in Table 1, together with
abbreviations used to identify them in the remaining
tables and figures. )IoTl!AVIR TO CIIAME VEHtClES VEHICLE TAANSFER
vEA&FaJKD1EmxA1uAEcOfFOm
VEHtCLE TEIPERITUAE
111VEHICLE
CHOICE CONSTRAINT SEGMENTATION ASSLMHCEOFK4V1ffiA SEAT SEAT ASSLMNCE
Previous analyses have shown that attitudes can be SEtR;~~ROWUKOESIfU6LEAClS
PERSONAL SECUAITV
used as effective descriptors of consumer preferences FaulVlow PER PEKW
I
muurmn PER PERSON
I
(Golob, 1973;Golob and Dobson, 1974).Considerably less

INTERVIEI : CIRCLE WV i SELEC


y r
DlI,atll iatll-
Fmtura: fld - 'id
bnfortabl* saalng ..~..........,...........26-l 3
13 4 5 6
DaPmdabllltv of on-time lrrtval ............ 29-l 3 4 5 6
Avbllebilltv IDI-.0, I.,$ when vw w.nt It .. 30-l 3 4 5 6
AttrKtlVWIl of wklcl* ...................31-I 3 4 5 6
Low nois. 1.~1 In whlsle ..................32-I 3 4 5 K
Vehicle .af.tv ..............................
33-l T4 3- 6
Sloothmrs of rl& ..........................34-I ~3 4 5 6
Prlvacv frm other pmpla ...................35-l 3 5 6
Avoldlns .xpo.ur. to traffls con~,artlon..... 36-l ~: : 5 6
lllniw .Jmo.".. to bad,," cradlna ......... 37-I 3 4 5 6
Lax out-of-p&d SO,t ...................... 38-l i3 4 z- 7 0
Lar riding (h-0............................. 35-1 I3 4 5 6 0
Lar wlklns tlma ............................40-l 5 6 0
Lo* waltlng tim ............................41.1 ~: :: 5 6 0
+f,ortunltv to .m.t md talk to other ~opls. 42-l 3 4 5 6 0
wortunltv to r*IU( ........ ... .. ........... 43-l 3 4 r d 0
opportun1tq to read ......................... 44-I 3 4 5 6 0
bntinuou. ride, fu *top. .................. 45-l 3 4 5 6 0
protutlon fromrather m entire trip ...... 46-l 3 4 5 6 0
Flmxlbla dertlnatlon,CM go lny*l.re ....... 47-l 3 4 5 6 0
Not having to changevehlcl,r ............... 46-l 3 4 b 0
vow-round t-r,tw. cafort In v&ICI. ... 4%1 3 4 5 cl
A~wr8 m. of h.rlng. ,e.t .................. 50-I 3 4 5 0
Suurlty fra und.,,r&l..st 5 of oth.r, .... 51-l 3 4 5 0
tllnlmu poll"tlon per permn carried ........ 52-l 3 4 5 0

-b) I& overall, taking wwvthing Into accamt, ha, mtl8fl.d .I. you, p.rw~lly, rlth
. S.T . ..mn. of tr.valllng fra ham, to work?

Very VW*
olr~~tl*fled s*tI9fiad a

53-l 2 3 4 5 6 0

Fig. 1. Sample work trip questionnaire page.


Anattitudinalmodalchoice model 301

successhas been achieved in using attitudes as descrip- including time to walk to and from the bus and waiting
tors and predictors of actual modal choice (Hartgen, time; and (2) number of transfers requiredto reach the
1974).It is proposed that this latter result is due in part to work trip destination.
the exclusion of Constraints which may intervene between A clustering procedure WCSused to identifysubgroups
travelers’ preferences and their actual choices. Such of the sample population which are relativelyhomogene-
constraints are typically associated with supply-side ous with respect to the three accessibility variables. The
characteristics of the alternatives. variables were standardized to zero mean and unit
Whife attitudes toward these supply-side characteristics variance to eliminate clustering bias due to scale
can be measured, it is in many cases inappropriate to differences. The clustering technique used was the
consider that these attitudes enter the choice decision ISODATA algorithm of Ball and Hall (1967).
process in a compensatory manner (e.g. Wallace and The 543 attitudinal Survey respondents were deter-
Sherret, 1973).As the most general case, individuals faced mined to be best clustered into five groups or segments,
with different sets of supply-side constraints will possess The positions of the five group centers in the standardized
different relationships between their behavior and their three-dimensional mode accessibility space was Used to
attitudes toward the choice alternatives. Since most label the segments in terms of the two bus accessibility
models of choice inherently assume that the Sample variablesand the auto accessibility variable(Fig. 2). The
population is homogeneous with respect to the structure segments were IabeUecIas the “mobile”, the “inapprop
of the decision process, variations in constraints on riate bus routing”, the “poor bus accessibility”, the
choice among the population can lead to erroneous “carless”, and the “busless” segment. The sample sizes
estimations of parameters in a supposedly common were 211, 99,94, 91 and 48, respectively. The aggregate
structure. modal splitsforthesefive segmentsareshowninFii. 3.
To ekninate estimation problems associated with The sample size for the “busless” segment(48 persons)
heterogeneity of choice constraints, a market segmenta- was too small to permit estimation of model coefficients.
tion technique was used to obtain po&+tion groups that This segment was removed from further consideration.
are relatively homogeneous with respect to such con-
straints. The accessibitity to auto for each survey
respondent in the Ottawa-Carleton/0 Regions 0 Car
was measured by constructing a swe variable based on
the respondent’s answers to questions regarding percep-
tions of auto availability. The scale points (assumed to be
at equal intervals) from lowest to highest accessibility are
as follows:
1. No access even as passenger
2. Occasional access as passenger
3. Occasional access as driver
4. Access anytime as passenger
5. Access anytime as driver
6. Ownership of personal automobile
Buabsa Carlear F%orbur lroprxoprote Moms
Survey respondents’ accessibility to bus was measured accMalb(lltYboaruutha
by two variables: (1) total perceived access time, Fig.3. Modechoice by marketsegmcats.

i%?ss
time
t -5

sAh$ E
SEGMENT IDENTITY
I MOBILE I 211
2 INAPPROPRIATE BUS ROUTING] 99
3 POOR BUS ACCESSIBILITY 1 94
4 CARLESS I 91
5 BUSLESS 1 48

Fig.2. Choicecoaseaint cl&W’ing.


302 W. W. RECKERand T. F. GOLOB

STlWClWRING PERCEPTION by deleting from consideration factors associated with the


Methodology smaller latent roots of the correlation matrix of the x
There is no reason to expect that the survey respon- variables.
dents’ ratings of a mode on each of the twenty-five The factor scores of eqn (1) can be readily estimated
attributes are independent from one another. The (Van de Geer, 1971):
attributes were chosen for inclusion in the survey on the
basis of their potential importance in planning policy
determination; they were not chosen in an attempt to
correspond with the latent psychological dimensions by
which consumers evaluate alternatives when making their where Aqk= the eigenvalue corresponding to factor q for
modal choice decisions. Moreover, psychologists con- perception of alternative k.
cerned with human decision-making have presented The varimax rotation applied to the fjqcoordinates in
convincing evidence that there are significantly fewer the pdimensional attribute space (specified in the
than twenty-five of these latent dimensions (e.g. Newell, principal components solution only up to an arbitrary
Shaw and Simon, 1958). orthogonal rotation) forces many of the fiq loadings
Factor analyses were employed to identify the latent toward zero and a few toward 1.0 in absolute value. Thus
perception dimensions for each of the four market the yk scores can be approximated by including only
segments with enough observations to yield statistically attributes in the SFk set of eqn (2) which have absolute
reliable results. Principal components analyses with values of fiqsignificantly different from zero:
varimax rotation (Harman, 1967) were applied to the
correlation matrices of each segment’s satisfaction ratings
for the two modes. The latent factors are linear
composites of the original attributes, each factor account-
ing for the maximum variance of the original attribute set, where gFk = the set of all attributes j for which I&l > D (a
subject to the condition that it is orthogonal to each of the is an appropirate confidence limit on the order of 0.30);
other factors for that group and mode. and qqk = the error introduced by ignoring the attributes
Not all of the twenty-five attributes were factored. included in set S,” but not included in set spk.
Attribute inclusion was determined by an iterative A comparison of factors across alternatives was
process in which attriiutes having low correlations (factor performed for each segment to determine the similarity of
1oadings)‘with each of the factors, or those not having a factor structure of perceptions toward each of the two
single dominant factor loading, were deleted from the choice alternatives. Factors which were found to be
correlation matrices. The adjusted matrices were then similar across alternatives were classified as “generic”
refactored until only attributes having significantly high factors. Factors which were unique to perception of a
loadings on a single factor remained. Deleted attributes single choice alternative were classified as “altemative-
were termed “non-factorable” attributes. Attributes specific” factors. In addition, a non-factorable attribute
which were included in the final factors were termed was classified as “generic” if it was factorable with
“factorable” attributes. The set of all factorable attributes respect to perceptions toward only one of the alternatives.
for perception of alternative k is denoted by S,“, and the These distinctions are important in developing a better
set of all non-factorable attributes is denoted by S,“. understanding of the interrelationships among the exp-
The number of factors retained for each modal choice lanatory variables of the choice models proposed in the
alternative and segment was determined by a comparison remaining stages of the methodology.
of the set of eigenvalues obtained from analyses of
random data matrices of the same order as the actual data “Mobile” segment perception structure
matrices and by consideration of the “Kaiser rule” in Factor analyses of the attribute satisfaction ratings for
which eigenvalues z 1 are retained (Horn, 1%5). Where the bus and auto modes by the 211 respondents classified
these two criteria did not result in selection of the same into the “mobile” market segment resulted in the selection
number of factors, selection between the two criteria was of four factors to represent perceptions of the bus mode
made through subjective judgment based on ease of and five factors to represent perceptions of the auto
interpretation. The set of factors selected to represent mode. This is illustrated in Fig. 4, where the first ten
perceptions of any choice alternative k is denoted by Q k. eigenvalues or latent roots of the two correlation matrices
The factor analytic expression for individual i’s rating of the attribute ratings are plotted together with the means
of the jth original attribute (j = 1,2,. . . ,25) for the kth of the eigenvalues found through repeated decomposi-
modal alternative can then be written tions of correlation matrices of random data of the same
orders as the correlation matrices. Table 2 lists the
xi= qE Lyfqfbfiq+E: for all j E S,“, (1) non-factorable attributes for the “mobile” segment.
3 Table 3 shows the bus perception factors comprising
where x$ = manifest rating by individual i of alternative k the set QBus. The “factor description” column gives a
on attribute j; yt= latent (i.e. unobserved) scores for subjective label for each factor, together with the
alternative k on factor q for individual i; ff4= percentage of the variance in the original attribute set
correlations (i.e. factor loadings) between attribute j and which is accounted for by this factor. The “attribute”
factor q for alternative k; and e$= the error introduced column lists the attributes with loadings on each factor
An attihdid modalchoicemodel 303

Table3. “Mobile” segment: bus worknip factors

‘Mobile”sepmentxworkhip taton
eigenbulwplot
(Samplesire-211)

4
x ___-- L\m
f - Bus
....... . .. .. .... Ron&,,&,&,)
P
g ?
P
W

respectively the strength and uniqueness of the attribute-


factor relationship.
The auto perception factors comprising the set QAW
, , , I I a I I#,
2 4 6 8 are presented in Table 4. Here the most dominant factor,
Factor “convenience”, does not explain as great a percentage of
Fii. 4. attribute variance as does the dominant bus “service”
factor. Moreover, none of the bus and auto factors were
Table 2. determined to be directly compatible, and consequently
‘m!j,LE. sEBENT : NORK TRIP
the set of generic factors is a nufl set.
IV,“-FACTORABLE ATTRIBUTES

ws 18) Aum (11) Perception structures of other segments


SEATII CawUAT SEAwiG CnFlFoAT
VEHICLE AT,AACTlVEMESS VEHYCLE ATTAACTIVEXEXS The sample sizes (99, 94 and 91) for each of the
VE”I‘,E YFETV VEHICLE SAFETY segments other than tbe “mobile” segment were too small
PEMDlUL SEWRITV PERYH(IC SECURITV
PULUTIOX PER PmfOA Pu1UT1m PER PE6cu
to permit factor analyses. The decision rule used in
__~~~_____~__________~_._~__~~~~~~~_~~~~~~~~___.
evaluating the sample sizes was that for a statistically
C”JT-UF-POCK1 EOSF DEPENDABLE IRRIV& TIM
WEATHER PXOTECTIMI AVAILABlLm
reliable analysis the difference between the number of
SEAT ASSUPJJKE mrvAcv individuals in the sample and the number of factor
YlIllM TIHE
FLEXIBLE UEsTI”ATb3,,
loadings to be estimated (the maximum expected number
VEHICLE TOPEAAME of factors times the expected number of factorable
attributes) must be greater than 20. Consequently, the
latent perception factor structure determined for the total
which are signiticantly di#erent from zero and the work trip sample minus the “mobile” segment was
“per cent variance” columns listed the per cent variance of employed as a proxy for these segments. This group of
each attribute which is accounted for by the factor in respondents was titled the “composite immobile” sample.
question and the per cent which is accounted for by all The eigenvalue plot from factor analyses of the
other factors combined. These two pieces of iuformation perceptions of the bus and auto modes by the “composite
(which sum to the total percent of the attriiute variance immobile” sample is given in Fig. 5. Comparison of this
accounted for by all factors, or the communality) indicate figure to the “mobile” segment plot of Fig. 4, indicates

Table4. “Mobile” segment:autowork trip factors

PERFOMAMCE RIOIN TM .?2


TRAFFIC CONGEST1ON .71 :z
OGOF-POCKET COST .a 40
(14%) RIDE CONTINlJITY .62 38 :i

PERSONAL OFWRTWTY TO READ .79 62


OPPOKNJNITY TO &ET OTHERS ::
ENVIRONWEM
(12%) OWORTUNITY TO RELAX .72
.70 :i 13
1 I I ,
VEHICLE RIOL VEHICLEWISE -85 07
RIDE WOTHNESS .77 :; 13
@%:::

CRlkllING
m%) I BODILY CRDHOING

‘IX Vol. 10.No. 5-B


304 W. W. RECKER and T. F. GOLOB

Table 7. “Composite immobile” sample: auto work trip factors

*Composite Immobile”sayk:work trip factors


eigenvalue plot
(Sample size=3321
, -----Auto
I
I YzFx $;dom(both)
I

L. I 51 * I 81, I
2 4 6 6 IO
Factor
Fig. 5. while two factors, crowding and sociability, account for
the remaining interdependencies for the “mobile” seg-
that one less factor is needed to describe the latent ment bus perceptions, only a single factor, personal
structure of both the bus and auto perceptions for the environment and autonomy, accounts for the perceptions
“composite immobile” segment. Table 5 lists the non- not described by the service and vehicle quality factors
factorable attributes for the “composite immobile” for the “composite immobile” segment. The inability to
sample, and Tables 6 and 7 describe the bus and auto separate perceptions of the five attributes comprising the
perception structures, respectively. personal environment and autonomy factor may be a
For the bus perceptions, comparison of Tables 3 and 6 result of confounding different patterns of perception by
reveals that the service factor accounting for the greatest including the four relatively inhomogeneous market
proportion of variance iu the attribute ratings is essen- segments into the single segment.
tially the same. The vehicle ride quality factor of the For the auto perceptions (Tables4 and 7), the personal
“mobile” segment is closely related to the vehicle quality environment, convenience, and performance latent per-
factor for the “composite immobile” sample, but the latter ception factors are approximately the same for the
factor includes “vehicle attractiveness”, a unique or “mobile” and “composite immobile” segments. The
unfactorable attribute for the “mobile” segment. Finally, “mobile” segment vehicle ride quality factor is expanded
to include “vehicle attractiveness” and “seating comfort”
Table 5. in the case of the “composite immobile” segment. The
“COMWSITEIMKWLE” SA)PLE : WORK TRIP
attribute “bodily crowding” is important enough in its
WON-FACTORABLE ATTRIBUTES description of the “mobile” segment’s perceptions to be
considered a latent factor itself, while its low relative
importance in describing perceptions for the “composite
immobile” segment relegates it to unfactorable status.

SEATING CONFORT AVAILABILITY


TRAFFIC CON6ESTMN VEHICLE SAFETY
OPPORTUNITV TO MEET OTHERS I PERSONAL SECURITY
RANL.rOM
WrUtlYCAOICEMODEL
Methodology
It is hypothesized that an individual decision maker’s
Table 6. “Composite immobile” sample: bus work trip factors overall preference ranking of a choice alternative is a
function of the utility which that alternative holds for the
individual. Utility is specified in terms of individual i’s
attitudes toward alternative k. Consistent with social
psychology theories of attitude and behavior (e.g.
Fishbein, 1%3) and with extensive market research
applications in consumer buying behavior and informa-
DEPENDABLE ARRIVAL TIME tion processing (e.g. Howard and Sheth, 1%9), a utility
FLEXIBLE DESTINATION
form which is linear and additive in terms of attitudes
toward the attributes of the alternatives is assumed. A
review of the use of this genre of attitude models in
consumer research is provided by Wilkie and Pessemier
(1973).
In light of the division of perception toward each
alternative k into a set of non-factorable attributes SNk
and a set of latent factors QL,the deterministic utility
PERSONAL SECURITY
component is given by
Anatlitudid modalchoicemodel 30s

choice criteria for establishing sFk’ must then be related to


specific planning objectives. This is judged to be a definite
where LJ/ = utility of alternative k to individual i; advantage of the present methodology: it can be used
a,k = utility weight reflecting the importance of the jth potentially to test a variety of policy issues without
attribute in contributing to the overall utility of alternative change in analytical structure. It is the antithesis of
k to any individual in a particular market segment; “single model” methodologies represented in an extreme
a,’ = utility weight reflecting the similar importance of the case by the use of step-wise or screening linear
qth latent factor; and 5: = random utility component, regressions to lind “optimal” subsets of independent
assumed to be independent and identically distributed variables (e.g. Constantine, Dobson and Canty, 1974).
across all individuals. By dividing the set of factor-representative attributes
Latent factors more properly take up the contribution dFk*into a generic set 9% and an alternative-specific set
of the attributes in the factorable set S,” because they &;, and by observing the previous differentiation
remove the strong multicollinearities in this set and are between generic and alternative-specific nonfactorable
purported to better represent the latent dimensions of attributes, utility can be written
perception independent of the survey designer’s selection
of attributes. However, specification of a choice model in
terms of latent perception factors as explanatory vari-
ables leads to major problems in interpretation of results where the utility weights a and b are now divided into
and prediction of consequences. Planners necessarily alternative-specific and generic weights: a: = utility
think in terms of attributes such as travel time, waiting weight of non-factorable attribute j when evaluating
time, or protection from the weather while waiting for a choice alternative k; a, = utility weight of non-factorable
vehicle, and not in terms of linear composites of these alternative-specific and generic weights: br = utility
attributes. The more exact form of eqn (4) must weight of factorable attribute j (chosen to represent an
consequently be approximated by a more readily inter- alternative-specific latent perceptual factor) when
pretable form. evaluating alternative k ; and b, = utility weight of
Substituting the factor scores approximation of eqn (3) factorable attribute j (chosen to represent a generic latent
into expression (4) yields perceptual factor) when evaluating any alternative.
More simply written, expression (7) states that the
utility of alternative k to individual i can be written as the
sum of a deterministic component (the first four terms on
the right-hand side of the expression) and a random
component:
where S: = an error term = X aqkgat+ .$ assumed to be
CiEQ' vi” = v: + e:. (8
independent and identically distributed across all indi-
viduals i. At this point it is convenient to assume a WeibuU
Since the attributes in the set sFk with high loadings on distribution for the random components and to derive the
each of the factors in the set Qk are in general mutually multinomial logit specification for the probability that
individual i will prefer alternative k from a set of
highly correlated, the term ,& x t jp for each q E Q k can available alternatives A:
be approximated by 8x&f;.,, where s is a proportionality
P,(k:A)= l/(1 +,&exp(V,’ -VI)). (9)
constant determined by the number of attributes in the set
I+k
and the respective scales of the one remaining attribute j*
and the deleted attributes. The use of attributes to Such a derivation is provided by McFadden (1973).
approximate latent perception factors is also found in the The present methodology differs from previous applica-
development of demand models advanced by Hartgen and tions of the logit model in the specification of the
Tanner (1971)and Constantino, Golob and Stopher (1975). deterministic utility component. For apparently the first
Thus, time, attitudinal variables are used exclusively, these
variables are explicitly selected in light of their structural
interrelationships, and they are distinguished as either
generic or alternative-specific variables. These features
where b: = a,k8f: for all j E s”‘, a modiied utility are best highlighted by rewriting (9) in terms of definition
weight; $,“‘= the set of attributes chosen each to (7) for Vi”:
represent one and only one of the factors in set Q”
describing perception of alternative k(8,’ c i?:); and
e: = an error term representing the cik error term of eqn
(5) and errors introduced by approximating latent factors
by attributes.
Operationally, the set $*’ can be specified as being
comprised of the attributes j* such that Ifi.,1is maximum
over all j E sPk,for each q E Qk. However, this choice
is arbitrary if for some q there are two or more attributes
ies’;,
I, (10)
with factor loadings with equally high absolute value. The
306 W. W. RECKER and T. F. GOLOB

Interpreting this expression, it is stated that individuals “Mobile” segment modal choice
compare pairs of choice alternatives on the basis of Table 8 lists results of the maximum likelihood
absolute levels of their perceived satisfactions with estimation of one choice model for the “mobile” segment.
alternative-specific attributes together with perceived Presented are the crk, a,, b; and bj coefficients for all
differences in satisfactions with generic or altemative- attributes j for which the coefficients are signiftcantly
independent attributes. Decision makers will make different from zero at the 95% confidence level. This
relative comparisons of alternatives on all attributes which significance test was performed by comparing the ratio of
have consistent meaning for each alternative, but will the coefficient values to their estimated standard errors,
make absolute evaluations of the alternatives on all which are asymptotically distributed as t-statistics in a
attributes which have unique meanings. Such a concep- linear model, to a critical value of approximately 1.65.
tualization considers attitudinal preference model specifi- This represents a one-tailed test of the null hypothesis for
cations in which all evaluations between alternatives are each variable. All attributes with coefficients insignifi-
treated as differences between attribute scores (e.g. cantly different from zero were not included in the tinal
Wallace and Sherret, 1973; Constantino, Golob and estimations. The variables are listed in order of their
Stopher, 1975; and Costantino, Dobson and Canty, 1974) t-statistics.
as special cases. Moreover, it simpliies the selection of The model goodness-of-fit indices shown in Table 8 are
“choice criterion” presented by Hartgen (1974) and is one judged to be very good for such probabilistic choice
approach to solving the “perceptual consistency” prob- models. Over 80% of the 211 choices can be predicted
lem outlined by Gensch and Golob (1975). correctly through utility comparisons using the estimated
Equation (10) represents the attitudinal modal choice parameters, and even for the less frequently chosen bus
model in the present methodology. The dependent mode, the prior probability of 20.8% expected correct
variable is the observed choice, where P,(k : A) takes the classifications can be approximately doubled to 38.5%
value 1 when k is chosen, 0 when k is not chosen. using the model results. (Insuthcient sample sizes
Parameters to be estimated are the mode-independent prevented use of a “hold-out” sample in these classifica-
utility weights a, for all attributes in set SN and bi for all tion tests.)
attributes in set ,!$*,,and the mode-dependent utility Due to the nonlinearity of the choice specification the
weights a,’ for all attributes in set Sk for all alternatives 1 elasticities of choice probabilities with respect to each of
and b,’ for all attributes in set 3:; for all alternatives 1. the explanatory variables are not directly proportional to
These were estimated using a maximum likelihood the coefficients of the variables as they are in linear
technique (McFadden, 1968). demand models. An elasticity is a dimensionless quantity
Because of the lack of goodness-of-fit measures with defined as the percentage change in the dependent
well-defined statistical properties (such as the linear variable which would result from a one percent change in
regression coefficient of determination, Rz) for such an independent variable:
probabilistic choice models emphasis was placed on
coefficient significance tests and on predictive perfor- -- @,(k :A) IdI
Ey _
(13)
mance criteria for evaluating the models. One such ax$ P,(k:A)
indicator is the ratio of choices predicted correctly by the
models; this is determined as the ratio of the number of where, Et = elasticity of the probability that alternative k
times the predicted probability of the chosen alternative is is chosen from set A by individual i with respect to the
greater than that of a non-chosen alternative. This ratio rating by individual i of alternative 1 on attribute j.
was also disaggregated by alternative chosen (bus or auto). For the logit specification, the individual elasticities can
In addition, two different measures (both termed “pseudo be written as
R*” or p*) which are nonlinear analogies of the linear R*
measure were used to evaluate overall model performance. (14)
The 8rst measure was apparently tirst used by McFadden
(1%8), and the second measure was used by Cragg (1968)
Tatde 8. “Mobile” segment: work trip logit choice model (sample
and is attributed to Theil. These measures are size=211)

McFadden p* = I- log I.( &log L (0), (11)


Theil p* = 1 - exp {2(log L(i)- log L(O))/N}, (12)

where L.(i) = likelihood function value for the vector of


estimated coefficients e^; L(0) = likelihood function value
with all coefficients set equal to zero; and N = sample size.
Unfortunately neither p* measure has well defined
distributional properties so that significance levels can be
defined. Moreover, they have maximum values which can PERCENT OF CHOICES PREDICTED CORRECTLY

be less than 1.0. They are used here mainly as indices for PERCENT OF CHOICES PREDICTED CORRECTLY

comparing choice model results for different market


segments with approximately equal sample sizes and PERCENT 0~ CHOICES PREDICTEO c0nREcnY
WHEN BUS HAS THE CHOSEN CaDE (20.8%
number of explanatory variables.
Auattidudinal modal choice model 307

Table 9. “Mobile” segment: worktrip aggregate logit choice elasticities (sample size = 211)

The elasticities for I = k are referred to as direct Table 10. “Mobile” segment: work trip correlations among
elasticities, or simply elasticities, since they represent the variables included in a logit choicemodel
changes in probability of choosing an alternative with
respect to perceived satisfactions toward the attributes of
the alternative itself; the elasticities for 1# k are referred
to as cross-elasticities.
Aggregate elasticities can be developed from the
individual elasticities to estimate the overall sensitivities AUTO-TRAFFIC , o.
of choice probabilities to uniform per cent changes in CGNGESTIGN

explanatory variables for all individuals. These aggregate K-wK*NG 0.10 1.w
elasticities are given by
$g$""'" 0.12 0.06 1.w

GUS-
AVAILAGILITY 0.05 -0.16 0.05 1.00

(1% They are, as shown in Table 10. Indeed one of the


objectives in the development of the proposed methodol-
ogy was to insure that problems of multicollinearity among
where E,” = aggregate elasticity of the probability of the independent variables do not adversely affect the
choosing alternative k with respect to uniform changes in predictive capabilities of the model.
the perceived evaluation of alternative I on attribute j for Another important objective was improved flexibility
all i=l,..., N individuals. over single-equation models in addressing planning policy
Aggregate elasticity estimates are listed in Table 9 for issues. As previously discussed, such flexibility is
the “mobile” segment. The probability of choosing the accomplished herein through calibration of the model
bus alternative was found to be most sensitive to with alternative attributes selected to represent the latent
respondents’ evaluations of the convenience of the auto perception factors (i.e. with alternative attribute sets 3::
alternative; a ten percent decline in satisfactions with auto and 9%). As an example, the choice models for the
walking time would be expected to result in over a twenty “mobile” segment were reestimated with the bus attri-
per cent increase in the probability of choosing bus. Other butes “wait time” and “vehicle transfer” replacing
explanatory variables exhibiting relatively high elasticities “availability” as representative of the bus service factor.
for choice of bus include the bus “availability” attribute The goodness-of-fit statistics for the three estimations
representing the bus service factor and the auto “traffic were virtually identical, and, as shown in Tables 11 and
congestion” attribute representing the auto performance 12, the elasticities and cross-elasticities for the unchanged
factor. The only equally high elasticity on choice of auto variables remained approximately the same.
is that associated with the bus service factor. A logit model employing only the respondents’ pcr-
These interpretations of the relative importances of the ceived time and cost for the two modes as explanatory
attributes of the bus and auto modes in explaining the variables was also estimated for the “mobile” sample.
modal choice decisions of residents of the Ottawa- While the “goodness of fit” measures for this model were
Carleton/Outaouais Regions imply that the explanatory approximately the same as those obtained with the
variables in the choice models are relatively independent. attitudinal models, much of the predictive capabilities of
Tat de 11. “Mobile” segmeat: work trip aggregate logit choice elasticities; model version 2
308 W. W. RECKERand T. F. GOLOB

Table 12. “Mobile”segment: work trip aggregate logit choice elasticities; model version 3

I VAR
LIABLE FACTOR -
AUTO ALTERNATIVE

ELASTICITY
I mncc_
"..""__
EWTICITY
I
BUS ALTERNATIVE
I rancc_
-..--"
ELASTICITY 1 ELASTICITY
BUS-Y
':!j!CLE BUS-SERVICE -1.485 1.883
TRANS..., I I I 1 I
AUTO-WALKIWG AUDI
0.377 ( -2.136 1
TIME UINVENIENCE
AUTO-TRAFFIC AUTO
CONGESTIOK PERFMUlANCE I 0.191
I 1 -1.123 1

AUTO- AUTO-
;;PO$NITY PERSONAL 0.098 -0.631
FNVIRGB)IENT
I I I

the time and cost model must be discounted because of an Table 14. “Poor bus accessibility”segment: work trip logit choice
incorrectly signed coefficient estimated for the cost model sample size N = 94
variable. One explanation for the incorrectly signed
coefficient may be found in relatively constant transit
fares, compared to automobile costs which are highly
dependent on trip length; there was greater dependency
on auto for trips of greater length.

McFadden p2- 0.735 The11 p2- 0.639


“Znappropriate bus routing” segment modal choice
Results of one choice model estimation for the
“inappropriate bus routing” segment are listed in Table PERCEM OF CHOICES PREDICTED CORRECTLY

13. Goodness-of-fit measures indicate that the explanat-


ory power of this model is significantly greater than that PERCENT OF CHOICES PREDICTED CORRECTLV
WHEN BUS WAS THE CHOSEN WOE (8.6% - 37.5 I
of the “mobile” segment models (whichwere themselves OF TOTAL CASES)
judged quite acceptable for planning purposes). The
model is also slightly more parsimonious than the
previous models, with only the auto convenience and bus case of the “inappropriate bus routing” model, goodness-
service factors and the “out-of-pocket cost” attribute of-fit indices are very satisfactory. Also consistent with
displaying coefficients significantly ditTerentfrom zero at the previous model, an attribute representing the bus
the 95% confidence level. There again is no problem with service factor and the attribute “out-of-pocket-cost” were
multicollinearity among the three explanatory variables to found to have significant explanatory power. This consis-
adversely atfeet use of the model for predictive purposes. tency between two segments faced with different aspects
The aggregate elasticity calculations for this model of bus supply-side problems is intuitively satisfying.
indicate that choice of the bus mode is most sensitive to Moreover, the inclusion of the auto convenience factor
the auto convenience factor, followed by the bus service as a significant explanatory variable in the “inappropriate
factor and “out-of-pocket cost” on both auto and bus. For bus routing” model and exclusion of this factor in
choice of the auto mode, a one percent increase in the “poor bus accessibility” model emphasizes dif-
perceived satisfactions toward the “dependable arrival ferences in bus supply-side problems faced by these
time” attribute of bus service would be expected to lead two market segments. Individuals in the “inappropriate
to a greater than 2% decrease in the probability of bus routing” segment can get to a bus, but the service
choosing auto. provided, as related to auto performance, might be
unacceptable to them. Individuals in the “poor bus
“Poor bus accessibility” segment modal choice accessibility” segment, on the other hand, experience
Results of the logit estimation for the “poor bus trouble in actually getting to and from the bus and thus
accessibility” segment are displayed in Table 14. As in the have no reason to compare the bus with performance of
the automobile. Omitted factors related to problems of
Table 13. “Inappropriate bus routing” segment: work trip logit
choice model (sample size= 99) getting to and from the bus might account for the
significance of the constant in the “poor bus accessibility”
model; this was the only model in which a significant
~~ constant was found.

“Carless” segment modal choice


Estimation results for the “carless” segment are
kFadden p2- 0.660 Theil A'= 0.606 provided in Table 15. Goodness-of-fit is poorer than for
PERCENT OF CHOICES PREDICTED CORRECTLY any of the other work trip modal choice models but is still
PERCENT OF CHOICES PREDICTED CORRECTLY judged to be very good. Three of the four factors
WREN AUTB WAS THE CHOSEN MODE (87.81
exhibiting sign&ant explanatory power in this model are
PERCENT OF CHOICES PREDXCTED CGRRECTLY related to social and psychological factors involving
J
relationships between an individual and those inanimate
Anattitudinal modal choice model 309

Table IS. “Carless” segment: work trip logit choice model urban areas or data collected both before and after
(samplesize = 91) substantial changes in travel mode characteristics.

Acknowledgements-The authors Bratefulfy acknowledge the


cooperation of the Regionaf Municipality of Ottawa-Carleton,
Ontario, Canada in providing data and motivationfor the study. fn
particular, Messrs. Richard Clark and Rena Safavian Bave
8enerously of their time. Messrs. Lawrence D. Burns,Abraham D.
Horowitz, K. S. fi~hmiu, and Gregory C. Nicolaidisof General
Motors ResearchLaboratories provided u&l critiquesand MS.
I I !
MagdalineHubbard and Mr. Raymond Voglerprovidedtyping and
I
graphics. Only the authors are implicated in any errors or
kFadden P'- 0.342 Theil P*- 0.374
misconceptions.
PERCENT OF WObXS PREOICTED CORRECTLY

Ball G. H. and Hall D. J. (1967) A clustering te&niqUe for


wnmarizing multivariatedata. Beheuior~Science12,153-M.
Bock R. and Jones L. (1968) The Measuremeat and Prediction of
Mgement and Choice. Holden Day, New York.
Brand D. and Manheim hf. (Editors) (1973)~.J&UITravelDemand
objects and other persons immediately around him or her. Forecasting. Special Report 143, Highway Research Board,
The abiity to identify the causality of such factors in National Academy of Sciences.
travel demand decisions is taken as direct evidence of the Constantino D. P., D&son R. and Canty E. T. (1974) An
usefulness of attitudinal models. Conventional demand investigation of modal choice for dual modetransit, People
mover and personal rapid transit systems. Proceedings, First
models based upon times and costs simply cannot deal Internationalhtal Mode Transportation Confcnnce.Transpor-
with such decision influences. tation Research Board, Washington. D.C.
There was no prior expectation of the negative sign on Costantino D. P., Golob T. F. and Stopher P. R. (1975).Consumer
the “opportunity to meet others” coefficient. There is no preferences for automated public transportation systems.
reason why opportunities for personal contact cannot be TransportationResearch Record No. 527,81-93.
Cragg J. G. (1968)Some statistical models for limited dependent
perceived to have a strong negative connotation possibly variables with application to the demand for durable goods.
related to threatening behavior or cultural alienation. In Discussion Paper No. 8, Department of Economics,University
light of this result, wording of the particular attribute of British Columbia.
evaluation question (Fig. 1) should probably be changed, Fishbeii hf. (1%3). An investigation of the relationships between
beliefs about an object and the attitude toward that object.
and further research is called for prior to execution of
Human Relations 16, 233-240.
future surveys with attribute rating tasks encompassing Gensch D. and Golob T. F. (1965) Testing the consistency of
similar modal attributes. attribute meaning in empirical concepttestiag. .I. Marketing
Research. To be published.
CONCLUSIONS Golob T. F. (193) Resource paper on attitudinal models. (In
Urban Travel Demand Forecasting (Edited by Brand D. and
This research addresses the question: Can attitudes Manbeim hf.). Special Report 143,Hiiway Research Board,
expressed by survey respondents rating their choice National Academy of Sciences.
alternatives on the basis of a comprehensive set of Golob T. F. and Dobson R. (1974) Assessmentof preferences and
descriptive attributes be used to explain travel modal perceptions toward attributes of transportationalternatives. In
Behavioral Demand Modeiing and Vafaerionof Travel Time.
choice decisions? The answer is yes, provided that certain
Special Report 149, Transportation Research Board, National
other information involving perceived accessrbilities to Academy of Sciences.
the choice alternatives is also employed. Harman If. H. (1%7) Modem Factor Analysis. University of
The attitudinal models were judged to provide useful Chicago Press, Chicago.
diagnostic information on which of the many attributes Hartgen D. T. (1924) Attitudinal and situational variables
intluencing urban mode choice: Some empirical findings.
describing modal alternatives are determinaat in choice. Transportation 3,377-392.
Such information can be of great value to planners and is Hartgen D. T. and Tanner G. H. (1971).fnvestipationsof the effect
unobtainable with models using only time and cost of traveler attitudes in a model of mode-choice behavior.
variables. In contrast to many previous mode choice HighwayResearch Record No. 369,pp. l-14.
Horn J. L. (1%5) A rational and test for the number of factors in
studies, this information is developed by explicitly taking
factor analysis. Psychometrika 30.
into account interrelationships between independent Howard J. A. and Sheth J. N. (1969)The Theory of Buyer
variables, as well as the usual relationships between each Behavior. Wiley, New York.
independent variable and the dependent choicevariable. McFadden D. (1%8) The revealed preferences of a government
Lie other models purported to explain the decision bureaucracy. Unpublished paper, Departmentof Economics,
University of California, Berkeley.
processes of individuals or groups of individuals whose McFadden D. (1973) Conditional logit analysis of qualitative
descriptive power has been demonstrated using a single choice behavior. fn Frontiers in Econometrics (Edited by
data set, these attitudinal models must be further Zarembka P.). Academic Press, New York.
validated. One step in the direction of validation would McFadden D. (1974)Quantal choice analysis:A survey. Presented
at NSF-NBER Conference on DecisionBales and Uncertainty,
involve goodness-of-fit testing of the presently calibrated
22-23 March, University of Caliiornia,Berkeley. Unpublished
models on an independent (i.e. hold-out) data set. More paper, Department of Economics University of California,
satisfactory steps would involve using data from different Berkeley.
310 W. W. RECKER and T. F. GOLOB

McFadden D. G. (1975) Transportation mode-choice research: Sheth J. N. (1974) Models of Buyer Behouior. Harper and Row,
Recent contribution from the social sciences. Working Paper New York.
No. 7502, Travel Demand Forecasting Project, Institute of Van De Geer J. P. (1971)Introduction to Multivariate Analysis for
Transportation and TratTic Engineering, University of Caliior- the Social Sciences. W. H. Freeman, San Francisco.
nia, Berkeley. Wallace J. P. aud Sherret A. (1973) Estimation of product
Newell A., Shaw J. C. and Simon H. A. (1958)Elements of a attributes and their importances. Lecture Notes in Economics
theory of human problem solving. Psychological Review 65, and MathematicalSystems, No. 89. Springer-Verlag, New York.
151-166. Wiie W. L. and Pessemier E. A. (1973)Issues iu marketing’s use
Nicolaidis G. C. (1975) Quautitication of the comfort variable. of multi-attribute attitude models. Journal of Marketing
Tmnsp. Res. 9, 51-66. Research 10, 428-41.

You might also like