See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/230663774

Estimation of Logit Choice Models Using Mixed Stated-Preference and

Revealed-Preference Information

Chapter · January 1996

CITATIONS READS
113 1,711

4 authors, including:

Mark Bradley Andrew Daly

Resource Systems Group University of Leeds
38 PUBLICATIONS   1,628 CITATIONS    93 PUBLICATIONS   2,513 CITATIONS   

SEE PROFILE SEE PROFILE

Peter Stopher
The University of Sydney
154 PUBLICATIONS   2,528 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Travel Surveys View project

Retirement View project

All content following this page was uploaded by Andrew Daly on 23 April 2015.

The user has requested enhancement of the downloaded file.

Submitted to the 6th International

Conference on Travel Behaviour

Quebec, May 1991

______________________________________________________________

ESTIMATION OF LOGIT CHOICE MODELS

USING MIXED STATED PREFERENCE AND

REVEALED PREFERENCE INFORMATION

______________________________________________________________

M.A. Bradley

&

A.J. Daly

Hague Consulting Group

ESTIMATION OF LOGIT CHOICE MODELS
USING MIXED STATED PREFERENCE AND
REVEALED PREFERENCE INFORMATION

Bradley, M.A. & Daly , A.J.

Hague Consulting Group

Abstract

The need for cost effective research techniques in transport has led to increasing use of Stated
Preference data, as well as the development of mixed models, based on multiple data sources.
Different types of data, as used in such models, may have different accuracy and sources of error.
Such differences exist between Revealed Preference and Stated Preference data, but also
additional differences may exist within each of these two categories: network journey times and
costs based on a very detailed zoning system are typically much more accurate than data based on
a coarser zoning system, and self-reported journey times and costs have a different accuracy
again. Within Stated Preference data, responses from "within mode" experiments usually reflect
less error than those from "across mode" experiments, and the type of the response variable
(ranking, rating, choice) may also affect accuracy.

The magnitude and source of errors in different types of data will be refle cted in differences in
both the measured components and in the unmeasured components (variance) captured in the
corresponding models. When choice models are estimated on multiple data sources, such as
various types of Revealed and Stated Preference data, these differences must be explicitly taken
into account in the specification of the model structure and the utility functions. Similarly, when
such models are used to provide forecasts for future scenarios, the accuracy and error sources of
the input data describing those scenarios must be taken into account when specifying the model in
estimation.

The paper discusses the main issues involved in the estimation and application of such models
estimated on mixed data, covering:

- the theoretical framework for combining data sources;

- the specification of the model and the likelihood function;
- the suitability and accessibility of different estimation techniques;
- a case study which compares probit and logit estimation methods;
- a case study which incorporates a new mode in a forecasting model;
- issues involved in interpreting and applying mixed model results.

In particular, the paper describes a new approach for estimating models on mixed data sources
using the "tree logit" estimation technique.

Key Words: Choice theory; Discrete choice modelling; Logit analysis; Probit analysis; Random
utility model; Stated preference
1. INTRODUCTION

Methods for using "revealed preference" (RP) disaggregate data to model travel demand have
now become well established [1]. The usual form of such data is that of choices reported from
traveller surveys, which are then related to the characteristics of the travel alternatives available to
them, usually measured in the form of highway and public transport network data, zonal-level
demographic and socioeconomic data, and background data collected as part of the survey. To
apply such models in forecasting, more aggregate data (e.g. a coarser zoning system) must often
be used, and a number of techniques have been de veloped for doing so [2].

Increasingly, "stated preference" (SP) data, based on responses to hypothetical travel situations in
a survey context, have been used in cases where observed choice behaviour is not adequate to
model the context of interest. SP data may be used, for example, because one is interested in
types of travel alternatives or characteristics which do not yet exist, or in qualitative attributes
which can be very difficult to measure in actual circumstances. In the earliest transport
applications using SP data, such as Kocur et.al. [3] and Sheldon and Steer [4], the methods used
remained quite close to those which had been widely adopted in the field of market research [5].

In recent years, the emphasis has shifted somewhat from "stated preference" data to "stated
choice" data which resembles revealed preference data as closely as possible and can be analysed
with similar techniques, such as discrete choice logit or probit estimation [6]. Concerns remain,
however, that actual choices and stated choice may actually be measuring different aspects of
behaviour, with the data subject to different levels of accuracy and different types of error and
bias. Such differences may be particularly important in forecasting, where the amount of
variation in actual behaviour which cannot be explained by the SP model becomes critical [7, 8,
9].

For forecasting purposes, one would ideally like a methodology which can derive the essential
information from SP data (that information which cannot be obtained from RP data) and to
incorporate it into a model which is suitable for application to RP-type data reflecting actual
choice conditions. That is, it is desirable to combine the stronger features of RP and SP data. The
issue addressed by the present paper is how this combination can be achieved.

Recently, Morikawa, Ben Akiva and McFadden [10, 11] have developed an approach for using
RP and SP data together in an integrated model estimation framework. They tested both a
sequential estimation approach and a more efficient approach which used all information from all
data sources simultaneously with a joint probit/ordered probit likelihood function. As described
later in this paper, we have applied a similar integrated approach to use logit estimation
techniques which are generally more accessible.
Structure of the Paper

This paper discusses and illustrates the main issues involved in the estimation and application of
mixed RP-SP models:

Section 2 provides background on the theoretical framework of the integrated estimation

approach, the structure of the model and the specification of the joint utility function. This
section includes a statistical derivation of the "tree logit" estimator of the joint likelihood
function.

Section 3 describes a case study of the tree logit estimation method. This case study was
performed on the same data set used previously by Morikawa (in [10]), meaning that a direct
comparison can be made of the logit and probit estimation techniques.

Section 4 presents a second case study, a fairly typical case where stated preference data is used
to supplement a revealed preference model in order to add into the model a mode choice
alternative which does not yet exist.

Section 5 provides a brief summary and conclusions regarding the methods and results described
in the paper.

2. THE THEORETICAL FRAMEWORK

The theoretical framework in which analysis is set is that of individual utility maximisation by
travellers. While other behavioural paradigms can be used for analysis of this type, utility
maximisation has been studied in detail by many other researchers and there is a consistent and
coherent body of theory based on utility maximisation to which reference can be made.

2.1 Formulation of the Analytical Problem

For further consistency, notation similar to that used by Morikawa et al. [11] is used. In this
notation, the (latent) utility maximised by travellers in their revealed preferences, uRP , for a given
traveller, for a given alternative, is given by:

uRP = β.xRP + α.w + ε (1)

where:

xRP , w are vectors of the measured variables influencing the RP decision;

β, α are vectors of unknown parameters (to be estimated);

ε represents the sum of the unmeasured utility components of utility influencing the
RP decision.
Similarly, the SP decision can be represented as being decided on the basis of the `utility function
uSP :

uSP = β.xSP + γ.z + ν (2)

where:

xSP , z are vectors of the measured variables influencing the SP decision;

β, γ are vectors of unknown parameters (to be estimated);

ν represents the sum of the unmeasured utility components of utility influencing the
SP decision.

In this notation, it is possible to allow for the existence of measured variables w that occur only in
the RP context or z that occur only in the SP context. These variables would also incorporate
average biases or modal preferences, which may of course be different in the two contexts. This
incorporation of mean bias in the measured variables allows the assumption that the mean value
of the unmeasured components ε and ν is zero for each alternative.

The key advantage of the approach is that variables x are incorporated that appear in both utility
functions. Their coefficients β can therefore be estimated using the information from both
surveys. By this formulation, therefore, the objective is achieved of exploiting data from the two
contexts to make joint estimates; this can be done of course only when there is an overlap of some
of the variables as indicated by equations (1) and (2) above. It is also necessary to assume that
the effect of the variables x in the RP and SP contexts is the same, i.e. that the marginal
contribution to utility indicated by βx is identical. (This assumption should be tested in
estimation.)

Among the variables that are present only in one of the contexts must be included all likely
context-specific biases. Among these it is necessary to allow for the "inertia" effect, whereby SP
respondents tend to be reluctant to change their preferences. This effect has been found to be of
substantial importance in a number of empirical studies. Generally, "alternative-specific
constants" are included in the utility functions for the alternatives. These constants represent the
mean net effect of all the unmeasured components of utility; they make it reasonable to assume
that the mean value of the unknown components ε and ν is zero. The existence of inertia and
other context -specific biases make it necessary in many cases to include these alternative-specific
constants within the context-dependent variables w and z rather than in the joint variables x.

The theory of utility-maximising models (see e.g. Ben-Akiva and Lerman [1]) derives choice
probabilities directly from the assumptions made about the distributions of the unobserved
components ε and ν. Thus to derive the probability models necessary to analyse the data it is
necessary to specify distributions for these variables.
2.2 Distributional Assumptions

Apart from the assumption that the marginal utility of each common variable is equal in each of
the contexts, which is necessary to make any joint use of the data, it is also necessary to make an
assumption about the distribution of the unmeasured components within the population. Because
of the presence of alternative -specific constants, as explained above, it is acceptable to assume
that the mean value of ε and ν is zero. However, further assumptions are required about the
interdependence (both between observations and between alternatives), variance and
distributional form to be taken by these variables. As is usual in analysis based on utility
maximisation, ε and ν are treated as random variables.

The issue of interdependence between observations is difficult. In most practical cases, a single
RP choice observation is taken from each traveller and there is little question that it is reasonable
to assume independence of these observations. However, it is one of the chief advantages of SP
surveys that repeat observations can be taken from individuals. As a result, however, the
assumption of independence must at best be a poor approximation. Nevertheless, this assumption
is routinely made and the success that has been achieved in previous studies suggests that the
problems caused by this assumption are limited in their practical effect. Until a more satisfactory
theoretical framework is established the assumption will be maintained. Interdependence
between RP and SP observations from the same individual can similarly not be treated fully at
present.

However, the existence of interdependence between SP observations from the same individual
makes it impossible to assume that the variance of the unmeasured components ε and ν is equal.
The existence of the correlation in the SP data suggests that the variance of ν will be less than that
of ε. However, there are many other effects at work here, such as the omission of many relevant
variables from most SP contexts and the existence of substantial measurement error in most RP
contexts. In general it cannot be stated a priori which of the two unmeasured components will
have the greater variance; it is however clear that it is not possible to assume that their variances
are equal.

The issue of interdependence of the unmeasured utility components between alternatives is of

relevance in models with more than two alternatives. The assumption that is adopted determines
the structure of the model of choice probability. This issue is treated fully in the literature [1].

The functional form assumed for these distributions of ε and ν determines the functional form of
the probability model that is used to estimate the coefficients α, β and γ. In the work of
Morikawa et al. the assumption is made that ε and ν are distributed normally (with zero mean
and unequal variance, as suggested above), leading to 'probit' models of probability. This
assumption of normal distribution facilitates analytical study of the models, but causes severe
difficulties in empirical work. Special computer programs are required (e.g. the use of 'GAUSS'
subroutines), which themselves can require long run times and present other difficulties in finding
optimum values of the parameters.
An alternative dis tributional assumption is to take the limiting-value distribution, often described
as 'Gumbell' or 'Weibull' to describe the variation of ε and ν. This assumption, suggested for the
present context by Daly [12], leads to the use of logit models for the probability calculations.
These models are characterised by their much greater simplicity in empirical work; this feature is
illustrated in the subsequent chapters of the present paper.

From a theoretical viewpoint, the use of the logit model is equally as appropriate as the use of the
probit. Intuitively, the logit model is appropriate when an alternative is viewed as being an
aggregation of a number of 'sub-alternatives'. The probit model, on the other hand, is appropriate
when the unmeasured part of the utility function is viewed as the sum of a large number of
components. Since both of these viewpoints appear to be partial descriptions of the reality being
modelled, there appears little theoretical basis for the choice of one model or the other.

From an empirical point of view, for binary models, apart from the computational difficulty of the
probit model, the results achieved with the two forms are very similar, as is illustrated in Chapter
3 below. The reasons for this are explained by Daly and Zachary [13]. Briefly, the major
difference between the two model forms is the greater sensitivity of the probit model to 'outliers'
(unusual observations) in model estimation. For less unusual observations estimation is almost
identical; in forecasting there is also almost no difference between the two model forms.

For more complicated models with more than two alternatives, the differences between logit and
probit forms are more significant. Logit models extend easily to a large number of alternatives,
provided that an assumption of symmetry among the alternatives is made. Some forms of
asymmetry can be handled with 'tree' logit forms. Very general model forms can be included in
the probit family, although the computational difficulties become very severe indeed. It seems
reasonable to suggest that different circumstances will require the use of different models.

It may be concluded that there is scope for the use of logit models in the integration of RP and SP
data. While some researchers may prefer to use probit models for specific research needs, the
logit model is theoretically attractive and empirically advantageous.

2.3 Estimation Procedure

The primary advantage offered by the use of logit models in this context is that of simplicity in its
empirical applications. To preserve and exploit this simplicity, it is desirable to be able to use or
adapt existing estimation procedures. This can be achieved as follows.

It is assumed, as sketched out in the previous discussion that the non-measured components of the
utility ε and ν are distributed independently (across both individuals and alternatives and
independently of each other) with the limiting-value distribution ('Weibull', 'Gumbell') but with
unequal variance. Define θ2 to be the ratio of the variances

θ2 = var(ε) / var(ν). (3)

Then the SP utility can be scaled by θ:

θ.uSP = (θ.β).xSP + (θ.γ).z + (θ.ν) (4)

so that the random variable (θ.ν) now has a variance equal to that in the RP utility (ε). It is now
possible to use both RP and SP observations in a logit estimation procedure that requires equal
variance across the observations.

However, while for the RP estimation the parameters to be estimated are the simple values of α
and β, for the SP the coefficients are (θ.β) and (θ.γ). This scaling has no other effect on the
distributional assumptions or on the conversion of the utility functions to choice probabilities. In
the case of (θ.γ), since the scale of γ is arbitrary, the presence of θ is not necessary, but in the case
of (θ.β) the scaling by θ is the essential link between the RP and SP models. The SP model,
however, is now non-linear - an estimation problem which can be solved by existing 'tree logit'
estimation software, by setting up an artificial tree structure as follows.

The artificial tree is constructed to have twice as many alternatives as there are in reality. Half of
these are labelled as RP alternatives, the other half as SP alternatives. The utility functions in
each case are as given in equations (1) and (2) (i.e. without θ). As indicated in the diagram, the
RP alternatives are placed just 'below' the 'root' of the tree; however, the SP alternatives are each
placed in a single-alternative 'nest'. For an RP observation, the SP alternatives are set unavailable
and the choice is modelled as in a standard logit model. For an SP observation, the RP
alternatives are set unavailable and choice is modelled by the tree logit structure.

Diagram: Artificial Tree Structure for Joint Estimation

'Root'

RP Alternatives (β,α) Dummy Alternatives (θ)

SP Alternatives (β,γ)

For SP observations, the mean utility of each of the dummy composite alternatives is composed
as usual in a tree logit model (see Daly, [14]):

vCOMP = θ . log Σ exp (vSP ) (5)

in which the sum is taken over all of the alternatives in the nest corresponding to the composite
alternative and

vSP = uSP - ν = β.xSP + γ.z (6)

is simply the measured part of the SP utility. Then, because each 'nest' contains only one
alternative in this specification,

vCOMP = θ.vSP = (θ.β).xSP + (θ.γ).z (7)

which is exactly the form required by equation 4, as long as the value of θ is constrained to be the
same for each of the dummy alternatives. This approach can also be generalised to two or more
different types of SP data by using a different set of dummy alternatives for each type, each with
its own θ parameter.

Since the dummy composite alternatives are placed just below the root of the tree, as are the RP
alternatives, a standard estimation procedure, such as Hague Consulting Group's ALOGIT
program, will ensure that θ is estimated to obtain uniform variance at this level.

It is important to note that this artificial construction requires no assumption about consistency of
preference between RP and SP choices: individuals are not modelled as choosing from the whole
set. Thus the usual requirement for tree logit models that θ should not exceed 1 does not apply.
The scale for SP relative to RP may be either greater than or less than 1.

Finally, the structure indicated above in which the models of choice over the RP and over the SP
alternatives are each simple logit in form can be generalised. Tree logit structures among the
actual alternatives may be introduced. To exploit this method for joint estimation it is sufficient
to introduce one additional level in the way outlined above.

3. CASE STUDY 1: COMPARISON OF LOGIT AND PROBIT MODELS

This Section describes the application of logit estimation techniques to the data set studied by
Morikawa [10]. Comparisons are made between Morikawa's probit estimations and estimations
using logit methods.

3.1 The Study Context

Hague Consulting Group, along with Steer Davies Gleave, was contracted by the Netherlands
Railways (NS) to study the potential for substitution between car and train for intercity travel in
the Netherlands as a function of rail service levels and fare. A sample of 235 individuals was
selected who had recently travelled by car or rail from the Dutch city of Nijmegen, near the
German border, to Amsterdam, Rotterdam or Den Haag, all located about 125 kilometres to the
West.

A computer-based home interview was designed to first elicit a detailed account of the journey
which the respondent had made, including all journey costs, times and interchanges. Respondents
were then asked to give their perception of making the same journey by the alternative mode.
This data was used as a basis for the attribute levels to be varied in stated preference exercises,
but also provided data suitable for estimating revealed preference (RP) models, albeit with self-
reported values of the car and train attributes.
The first SP experiment (SP1) was designed to measure the relative importance of four train
attributes: fare, journey time, number of interchanges, and comfort level. This was thus a
"within-mode" game, where all respondents compared different train options against each other.
Pairwise choice questions were used, with a five-point response scale: (1) Definitely A, (2)
Probably A, (3) Not Sure, (4) Probably B and (5) Definitely B.

A second SP experiment (SP2) was a mode choice game, always comparing car against train.
The same four attributes for train were varied as in the first game. For car, although both travel
cost and time were shown as attributes, only the cost was varied during the experiment. In order
to create situations in which people may be induced to switch from the actual mode, the train
attributes were varied to become worse than the reported levels for train users and better than the
reported levels for car users. The same five-point response scale was used as in SP1. Further
information on the survey and results is reported in Bradley et.al. [15].

3.2 Results

Morikawa [10] used ordered probit estimation to ana lyse the interval information in the five-point
scale responses in the SP1 and SP2 data, and binary probit estimation for the RP train vs. car
choice. For the mixed RP/SP models, he programmed the joint likelihood function in the GAUSS
microcomputer package. His results, reproduced in Table 1, contain a number of results which
are typical for the types of data used:

_ the RP model does not yield a significant estimate for in-vehicle time, presumably due to
correlations with other variables such as cost and number of transfers;

_ the SP1 within-mode model has the most significant estimates, but coefficients can be
estimated only for the four variables which were varied in the experiment;

_ the SP2 across-mode model contains fairly significant estimates for all level-of-service
variables, but has values of the rail mode constants which are different than for RP, and
shows a strong "inertia" or "self-selection" effect, as discussed in Section 2;

_ the scale parameters show that the within-mode SP1 responses have significantly less
unexplained variance than the RP data (θ>1), and the across-mode SP2 responses have
significantly more unexplained variance when compared to the RP data (θ<1).

Table 2 contains results for models estimated on exactly the same data, with the same variable
specifications as in Table 1. The models in Table 2, however, were estimated using binary logit
models and tree logit models for the mixed RP/SP models, as explained in Section 2. Estimation
was done using Hague Consulting Group's ALOGIT microcomputer package.
Table 1: Probit/Ordered Probit Parameter Estimates
(from Morikawa [10], except RP2 from Morikawa, et.al., [11])

Model : RP SP1 SP2 RP+SP1 RP+SP2 RP+SP1+SP2 RP2

N.Obs. : 235 2965 1628 3200 1863 4828 228

ρ2 : .295 .315 .372 .314 .366 .334 .299

(coefficients, with absolute t-values in parentheses)

(β)
Cost/pers -.026 -.081 -.0072 -.024 -.024 -.025 -.027
(gld.) (4.4) (25.3) (5.3) (4.9) (4.9) (5.4) (4.4)

In-V.Time .0005 -.016 -.0032 -.0047 -.0046 -.0050 -.0050

(min.) (0.2) (11.7) (2.7) (4.6) (1.9) (5.1) (1.3)

OutV.Time -.026 -.0066 -.026 -.025 -.025 -.026

(min.) (4.4) (3.2) (4.6) (4.7) (4.9) (4.7)

Transfers -.344 -.156 -.053 -.049 -.245 -.052 -.144

(number) (2.6) (4.9) (1.6) (3.6) (2.7) (3.9) (1.0)
(α)
Non-work -.900 -.870 -.900 -.864 -.691
(0/1 -RP) (4.0) (4.1) (4.2) (4.1) (2.2)

Rail ASC 1.74 1.44 1.61 1.43 0.35

(0/1 -RP) (5.5) (5.9) (5.5) (6.0) (1.2)
(γ)
Comfort -.500 -.150 -.156
(0/1/2-SP1) (14.8) (4.7) (5.2)

Non-work -.025 -.129 -.075

(0/1 -SP2) (0.3) (0.5) (0.3)

Rail ASC -.966 -2.96 -3.18

(0/1 -SP2) (7.9) (3.4) (3.8)

Inertia 1.49 4.84 4.90

(0/1 -SP2) (17.5) (4.2) (4.4)
(θ)
SP1 scaled to RP 3.33 3.22
(4.8) (5.3)

SP2 scaled to RP .294 .298

(4.3) (4.5)

_________________________________________________________________

Table 2: Logit/Tree logit Parameter Estimates

(divided by factor of 1.814 to give probit scale [13])

Model : RP SP1 SP2 RP+SP1 RP+SP2 RP+SP1+SP2

N.Obs. : 235 2965 1628 3200 1863 4828

ρ2 : .303 .148 .338 .158 .332 .218

(coefficients, with absolute t-values in parentheses)

(β)
Cost/pers -.025 -.081 -.0061 -.026 -.022 -.026
(gld.) (4.3) (19.8) (3.5) (4.7) (4.6) (5.0)

In-V.Time -.0040 -.016 -.0045 -.0050 -.0083 -.0051

(min.) (1.0) (10.7) (3.2) (4.4) (2.8) (4.7)

OutV.Time -.024 -.0043 -.024 -.023 -.025

(min.) (3.9) (1.7) (4.0) (4.1) (4.5)

Transfers -.286 -.176 -.065 -.058 -.269 -.058

(number) (2.3) (5.4) (1.3) (3.6) (2.7) (3.8)
(α)
Non-work -.690 -.660 -.684 -.664
(0/1 -RP) (3.3) (3.2) (3.3) (3.2)

Rail ASC 1.44 1.27 1.33 1.28

(0/1 -RP) (5.1) (5.0) (4.9) (5.1)
(γ)
Comfort -.517 -.166 -.165
(0/1/2-SP1) (14.6) (4.5) (4.8)

Non-work .092 .278 .319

(0/1 -SP2) (1.1) (1.0) (1.0)

Rail ASC -.880 -2.67 -3.14

(0/1 -SP2) (6.9) (3.2) (3.2)

Inertia 1.42 4.50 4.86

(0/1 -SP2) (15.3) (4.2) (3.9)
(θ)
SP1 scaled to RP 3.12 3.16
(4.6) (4.8)

SP2 scaled to RP .299 .276

(4.2) (3.9)
_______________________________________________________________
* Two additional dummy variables were included in the RP2 model,
mainly affecting the value of the rail constant.
In order to estimate binary logit models on the five-point scale SP responses, we interpreted
"Definitely A" and "Probably A" as simple choices for A, "Definitely B" and "Probably B" as
simple choices for B, and "Not Sure" as half of a choice for A and half for B. (One could also
omit the "Not Sure" responses to give a true discrete choice variable.)

There is some information lost in the transformation from the scale response to binary choice, as
indicated by the lower ρ 2 measures of fit for the SP models in Table 2 as compared to Table 1.
This extra explanation in the ordered probit models is provided by two extra parameters (omitted
from Table 1) which indicate the threshold utility differences that define the five symmetric scale
intervals. In terms of likelihood measures, this loss of information appears more important for
SP1 within-mode choices, where individuals are comparing very similar alternatives, than for the
SP2 mode choices. In terms of the coefficients and their significance, however, the difference is
quite small, particularly for the SP1 model.

The clearest way of comparing the probit and logit results is to compare the parameter estimates
and their statistical significance in Tables 1 and 2. The logit coefficients have been divided by a
factor of π〉+3 in order to normalise the Weibull error distribution to the standard normal
distribution used in probit [13]. None of the results are significantly different from each other in
the two tables. The two estimates which change sign - in-vehicle time for RP and the rail
alternative-specific constant (ASC) for SP2 - are not significant in either case. The estimates
appear closest and most robust in the SP1 within-mode data and least robust in the RP data. The
t-statistics are slightly lower for the logit models than for the probit models, but are very similar
in the case of the RP+SP models. Perhaps most importantly, the scale factors for the RP+SP
models are nearly identical in Tables 1 and 2, indicating that the tree logit method for estimating
these factors gives comparable results.

An interesting additional point is that some of the results mentioned above suggest that the
difference between the probit and logit results is caused by the presence of "outliers", i.e.
responses which are very poorly predicted by the model, often due to errors in the data. Because
the error distribution assumed in probit has smaller tails than that assumed in logit, probit
estimates tend to be more distorted by outliers [13]. Further evidence is provided by the RP2
model in Table 1, reported in a subsequent paper by Morikawa, et.al. [11] after 7 respondents
had been deleted from the RP sample. These results now appear somewhat more similar to those
for RP in Table 2, particularly the change of sign for in-vehicle time. The sensitivity to outliers is
not a strong point for or against the use of either technique, since the best modelling practice in
both cases is to identify such outliers, examine them, and eliminate them from the sample if
appropriate.

In summary, this case study has shown that the tree logit approach, which is quite easily
performed with currently available logit estimation software, gives results consistent with those
obtained from a more complex probit estimation procedure. This difference is especially
important for models with more than two alternatives, such as the one described in the following
case study. Existing logit estimation packages, such as ALOGIT can estimate nested,
multinomial logit models, while such estimation is not currently possible with probit ana lysis.

In terms of application, the use of RP in this case study was as a small supplementary data source
obtained during the SP interviews, in order to scale the sensitivity of the SP models to that
reflected in current choices. A more challenging case for the use of SP data, perhaps, is when it
itself is the supplementary data source, used to add additional parameters and confidence to a
model which is based on an independent RP data, and is estimated and applied using objective
network-based data. Such a case study is described in the next section.

4. CASE STUDY 2: MIXING OF INDEPENDENT RP AND SP DATA

A situation in which the logit approach offers particular advantages relative to probit modelling is
when it is required to model choice among seve ral alternatives. A study applying a model of that
type is described in this Chapter.

4.1 The Study Context

Hague Consulting Group, along with Macquarie University and Cambridge Systematics, was
contracted by CSIRO in Australia to study the market potential for a high-speed rail system
connecting the cities of Sydney, Canberra and Melbourne. The objective was to identify the
sensitivity of demand to key variables such as fare and travel time, and also to predict the absolute
level of demand, requiring estimation of a choice model including a mode which did not yet exist.

The first stage of the study included intercept surveys among car, air, bus and train travellers
along all routes for which the proposed Very Fast Train (VFT) would be a relevant alternative.
Over 30,000 such questionnaires were completed. This data was expanded to produce origin-
destination matrices of existing travel in the corridor, and was also analysed to produce RP
models of mode choice among the existing alternatives.

The second stage involved over 2500 home interviews, with roughly half of the respondents
recruited out of the intercept samples, and the other half contacted at random to include non-
travellers. The SP experiment began by collecting a number of details of the most recent actual
journey and of certain attributes of the alternative modes for that journey (e.g. access/egress
times from airports and train/coach stations). The VFT service was then introduced, using a
realistic and informative colour brochure. The stated preference experiment used an orthogonal
design, with the travel times and costs of the existing modes and the VFT varied independently,
with each respondent considering three relevant sets of hypothetical mode alternatives,
supplemented with the information which they provided on access/ egress times and costs.

Although only data from the mode choice SP experiment is discussed here, the survey also
included a second, within -mode SP experiment which asked for choices among different
"versions" of the VFT, with attributes including fare, type of seating, and entertainment facilities.
Further details of the study are given in Gunn, et. al. [16].
4.2 The Results

First, we should note that, because of continuing development, the models described in this
section are not those which are currently used to forecast demand for the VFT system.

Table 3 shows model results for Business trips. Two model specifications were estimated: the
first uses the travel cost per person in the travelling party divided by the average wage per worker
in the party, giving the cost in units of minutes of "work-equivalent" time. Other variables are
main line-haul time for the mode, access/egress time to the mode, and a variety of mode
constants. Model 2 is identical except that the logarithm of cost per person divided by the
logarithm of the average wage is used as the cost variable. The second specification is less
conventional and more difficult to interpret than the first, but gives a substantially better model fit
to the RP data.

A number of points can be made about these results. First, the change in the specification of the
cost variable between Model 1 and Model 2 has a large influence on the RP -only results -
indicated by the change in ρ 2 and the change in the cost and time coefficients. The change has
much less effect on the SP-only results. This might be expected in that RP data generally contain
more unexplained relationships and background effects, which are less common in SP data due to
the highly controlled nature of the independent variables and the choice context. This difference
is also evident in the extra mode-specific effects related to day trips, travelling alone, age of the
traveller, etc., which appear much more significant in the RP data than in the SP data.

One result is very similar to the results of Case Study 1: with Model 1 estimated on the RP data,
the coefficient for access time has the wrong sign, as a result of correlation with the other time
and cost variables. Model 1 on the SP data, on the other hand, shows a much more reasonable
estimate for access time: two to three times as high as that for main mode time. In contrast to
Case Study 1, however, combining the RP and SP data for Model 1 did not produce a significant
access time coefficient. The difference here is the fact that the RP sample is quite large compared
to the SP sample and has a similar scale (θ+1). There is no within-mode SP data with high
enough measurement precision to "override" the RP data, as was the case in Tables 1 and 2. In
practice, if one believed that the SP results were more "correct", it would be best to simply
constrain the main and access time coefficients in the RP or RP+SP model to have the same ratios
as in the SP-only model - the "sequential" estimation approach.

With Model 2, the ratio of access time to main mode time is very similar for both the RP and SP
data, so the issues discussed above do not arise. Another issue does arise, however: in Model 1
the ratio of the main mode time coefficient to the "work time" cost coefficient is about 3.0 for all
data types. In Model 2, however, the SP time coefficient is much higher than the RP coefficient,
while the SP cost coefficient is lower than that for RP. For the mixed RP+SP model, the results
are somewhat in-between, closer to the RP-only results due to sample size.
Table 3: Logit Model Parameter Estimates for Business Trips

----- Model 1*----- ----- Model 2*-----

Data : RP SP RP+SP RP SP RP+SP

N.Obs. : 12586 1930 14516 12586 1930 14516

ρ2 : .364 .258 .340 .442 .254 .400

(coefficients, with absolute t-values in parentheses)

RP/SP Generic Variables (β )

Cost/Wage -.0013 -.0021 -.0013 -7.08 -3.19 -6.72
(min.) (28.6) (7.0) (29.6) (39.0) (6.2) (36.3)

Main Time -.0045 -.0057 -.0047 -.0012 -.0045 -.0017

(min.) (31.9) (13.3) (33.5) (10.1) (12.5) (13.4)

Access Time .0026 -.0090 .0001 -.0026 -.0090 -.0035

(min.) (4.7) (7.5) (0.2) (4.2) (7.4) (6.4)

RP-Specific Variables (α)

Air -1.27 -.947 3.97 3.65
(RP) (11.1) (8.3) (21.5) (20.1)

Coach -1.80 -1.58 -.851 -.740

(RP) (22.6) (20.8) (9.9) (8.9)

Train -1.98 -1.80 -.266 -.218

(RP) (25.6) (24.0) (2.9) (2.4)

Air-Travel 1.94 1.93 1.21 1.27

Alone (RP) (25.4) (25.3) (14.1) (14.8)

Air-Age -.698 -.690 -.728 -.749

< 35 (RP) (11.1) (11.0) (10.9) (11.2)

Air-Day .337 .321 .611 .626

Trip (RP) (4.6) (4.4) (7.3) (7.5)

SP-Specific Variables (γ)

Air .084 -.906 .941 2.68
(SP) (0.4) (7.6) (3.5) (14.0)

Coach -1.95 -2.37 -1.75 -2.74

(SP) (3.2) (4.1) (2.9) (3.5)

Train .406 -.300 .577 .063

(SP) (1.3) (1.1) (1.8) (0.2)

VFT 2.68 1.74 3.46 5.59

(SP) (12.1) (6.4) (13.3) (13.0)

VFT-Travel .175 .198 .076 -.077

Alone (SP) (1.6) (2.1) (0.7) (0.6)

VFT-Day -.137 -.110 -.177 -.275

Trip (SP) (1.3) (1.2) (1.7) (2.2)

Inertia 2.65 2.35 2.60 3.37

(SP-intercept) (15.1) (8.7) (14.9) (8.9)

Inertia 2.37 2.02 2.24 2.85

(SP-random) (11.4) (7.5) (10.8) (7.6)

Scale Parameter (θ)

SP scaled to RP 1.12 .826
(11.3) (11.0)

______________________________________________________________

* Model 1 uses travel cost per person in the travelling party

divided by the average wage per worker in the party (in $A/min).
Model 2 uses the logarithm of cost per person divided by the
logarithm of the average wage per worker.
This difference shows another way in which SP information can be useful. Namely, Model 2 has
a better fit to the RP data, but the SP results indicate that Model 1 may be a more "correct" model.
Such a contrast might suggest that the better fit of Model 2 to the RP data may be due to other
factors correlated with time or cost or to extreme values in the input data. In such a case, further
testing of the RP model specification could be useful. In further tests of these models, for
example, we found that a nested model structure with all modes other than car in a separate
"branch" of a tree was justified for both the RP and the SP data. The tree logit scaling approach
could also be applied to combine these models - which themselves had a tree structure. This level
of complexity - a doubly nested structure - is not likely to be feasible using other estimation
methods.

The SP models in Table 3 contain an inertia effect to reflect the higher probability of selecting the
actual mode in the SP choice context. The size of the inertia effects is similar to that found in
case study 1. Those recruited from the intercept survey show a somewhat larger inertia effect,
perhaps because they are likely to be more frequent travellers than those in the random sample.
Although the inertia variables may already reflect a number of the unobserved effects which are
captured in the RP mode -specific variables, it is also useful to estimate additional mode-specific
effects from the SP data. The most important such variables are those related to the new VFT
mode, since these are necessary for forecasting and cannot be derived from the RP data. Mode-
specific constants for the existing modes are also estimated from the SP data. These are not
constrained to be the same as those from the RP data, as different types of unobserved effects may
influence the two types of data. The two data sources give the same general trends for these
constants, however, with air most positive and rail usually most negative.

The scale parameters for the SP data show that the unexplained variance is similar to that of the
RP data. As the RP model improves from model 1 to model 2, the scale on the SP data becomes
slightly less than one, although the difference from 1.0 is not highly significant in either case.
This result can be viewed positively for the SP mode choice data: it has provided significant
estimates for all of the important variables, while the variance is not significantly higher than that
of the RP data, as it was in Case Study 1.

Given that the RP data is that which the models are applied to in forecasting, it was used as the
"primary" data source, and the SP utilities were scaled relative to it. As a result, any coefficients
in the RP+SP model which apply to the RP data can be applied as they are. Any coefficients
which are specific to the SP data, however, must be multiplied by the scale factor before use in
forecasting. The VFT constant for day trips in Model 2 in Table 3 would thus be -0.275 * 0.826 =
-0.227, a value closer to that in the SP -only model. The SP constants for the existing modes and
the inertia constants would usually not be applied in forecasting, as they are already reflected in
the RP mode constants.

Finally, note that simultaneous estimation is not necessary when SP experiments only have a
single variable in common with RP models. An example is a within-mode experiment trading off
a number of VFT qualitative service variables against fare level. Models based on such data can
be applied by translating the coefficients into monetary units and then using them directly in
combination with the cost coefficient in the forecasting model.
5. SUMMARY AND CONCLUSIONS

Experience in transportation planning has confirmed the usefulness of both Revealed Preference
and Stated Preference data. However, these two types of data each offer specific and different
strengths and weaknesses. Analysis exploiting both types of data simultaneously can use the
strengths of each type of data to help overcome the weaknesses of the other.

The key points in combining the two types of data are the differing biases and errors that occur in
the two differing contexts. Coping with biases introduces no new methodological issues, but the
fact that different levels of error are present requires special treatment in modelling.

The proposal presented in this paper is that the combination of RP and SP data can be achieved
using models of the logit form. Accommodation of the differing error variances then requires the
use of an artificial 'tree' logit construction. However, the resulting model can be estimated
conveniently using existing commercial software.

The method developed has been tested on two separate case studies.

In the first case study, it was shown that the results obtained from the logit models proposed here
correspond closely with those obtained from the probit models investigated by Morikawa et al.
[11]. Remaining differences appear to be largely associated with the higher sensitivity of probit
models to unusual observations ("outliers").

The second case study confirms the results of the first in showing that the combined RP+SP data
can improve and expand upon the results of RP-only data, while retaining the reliability of
association with actual behaviour.

The methods presented are not difficult to apply and can be used with readily available software.
The advantages of combining data of these two types are substantial and the results thus far
appear reliable. Further examples of research and applications of this new approach will be
important in making it a part of standard travel demand modelling practice.
 REFERENCES
[1] Ben Akiva, M. and Lerman, S. Discrete Choice Analysis: Theory and Application to Travel Demand. MIT
Press. Cambridge, Mass. 1985.

[2] Daly, A.J. and Ortúzar, J.de D. "Forecasting and Data Aggregation: theory and practice". Traffic Engineering
and Control. Surrey. December 1990.

[3] Kocur, G., Adler, T. and Hyman, W. "Guide to Forecasting Travel Demand with Direct Utility Assessment".
U.S. Dept. of Transportation. Wash. D.C. 1982.

[4] Sheldon, R.J. and Steer, J.K. "The Use of Conjoint Analysis in Transport Research". Proceedings of PTRC 10th
Annual Summer Meeting. Warwick. July 1982.

[5] Green, P.E. and Srinivasin, V. "Conjoint Analysis in Consumer Research: Issues and Outlook". Journal of
Consumer Research. v.5. pp. 103-123. 1978.

[6] McFadden, D. "The Choice Theory Approach to Market Research". Marketing Science. v.5. no.4. 1986.

[7] Ben Akiva, M. and Boccarra, B. "Integrated Framework for Travel Behavior Analysis". Presented at the
International Conference for Travel Behavior. Aix-en-Provence. October 1987.

[8] Bradley, M.A. and Kroes, E.P. "Forecasting Issues in Stated Preference Survey Research". Presented to 3rd
International Conference on Survey Methods in Transportation. Wash. D.C. January 1990.

[9] Bradley, M.A. and Kroes, E.P. "Simultaneous Analysis of Stated Preference and Revealed Preference
Information". PTRC 18th Annual Summer Meeting- Proceedings of Seminar H. University of Sussex. September
1990.

[10] Morikawa, T. "Incorporating Stated Preference Data in Travel Demand Analysis". Doctoral dissertation,
Massachusetts Institute of Technology, Cambridge. June 1989.

[11] Morikawa, T., Ben Akiva, M. and McFadden, D. "Incorporating Psychometric Data in Econometric Demand
Models". Prepared for Banff Invitational Symposium on Consumer Decision Making and Choice Behavior. Canada.
May 1990.

[12] Daly, A.J. "Integration of Revealed Preference and Stated Preference Data in Model Estimation". Discussion
Paper for Banff Invitational Symposium on Consumer Decision Making and Choice Behaviour. Canada. May 1990.

[13] Daly, A.J. and Zachary, S. "Commuters' Values of Time". Local Government Operational Research Unit.
Report T55. Reading. January 1975. pp. 55-60.

[14] Daly, A. J. "Estimation of 'Tree' Logit Models", Transportation Research, 21B, pp 251-267, 1987.

[15] Bradley, M.A., Grosvenor, T. and Bouma, A. "An Application of Computer-Based Stated Preference to Study
Mode Switching in the Netherlands". PTRC 16th Annual Summer Meeting- Proceedings of Seminar D. University
of Bath. September 1988.

[16] Gunn, H.F., Bradley, M.A. and Hensher, D.A. "High Speed Rail Market Projection: Survey Design and
Analysis". Presented at the 3rd International Conference on Survey Methods in Transportation. Washington, DC.
January 1990.

View publication stats