Forecasting and Data Aggregation: Theory and Practice

You might also like

Download as doc, pdf, or txt
Download as doc, pdf, or txt
You are on page 1of 20

Draft 12 August 1990

FORECASTING AND DATA AGGREGATION: THEORY AND PRACTICE

A. J. Daly and J. de D. Ortúzar

An important issue in the design of a transportation planning study is the level of


aggregation that is selected for the measurement of data. In general, a higher level of
disaggregation, giving a greater level of accuracy, will improve the quality of forecasting models
but increase the cost of data collection and of other aspects of the modelling exercise. Clearly,
some data items give a better return on the investment of effort in increasing their accuracy: the
problem is to optimise the return on the investment of a fixed budget over all the data to be
collected. The resulting pattern of investment may well be different for different sub-models
in the transportation planning system.

This paper addresses the problem of determining appropriate levels of data aggregation for two
key travel demand forecasting models: those predicting destination choice (or trip distribution)
and mode choice. These choice processes may be treated separately or, as is now becoming
more common, in a joint modelling procedure. They have been chosen for study in this paper
because of their intrinsic importance and because of the availability of results from a number
of studies that serve to illustrate key points.

The objective of the paper is to provide indications for the practitioner for appropriate
aggregation levels to be used in empirical studies. The recommendations are supported by
theoretical discussions and practical findings.

1. Introduction: Data Aggregation in Travel Demand Modelling

The modelling of travel demand is a crucial part of most transportation planning studies.
Forecasting of the likely usage of proposed new facilities is an essential precursor to
rational decision-making about the value of those facilities, whether that value is assessed in
revenue to the operator, in benefits to the users or in cost-benefit surplus to the community as a
whole. The sensitivities of demand to important variables under the control of the operator,
such as the price charged, also need to be known. Thus the transport planner is frequently
confronted with the need to construct a forecasting mechanism, i.e. a model (of one sort or
another), to provide these forecasts and sensitivities. Clearly these forecasts need to be
"aggregate", i.e. to represent the behaviour of an entire population.

In very many practical studies, the models used are often still those of the classical "four-
stage" process, despite the many justified criticisms that have been made of their inflexibility,
implausibility, inaccuracy and cost. Among other reasons for the persistence of these old
models, such as their familiarity, it is important that they offer a 'recipe' for the complete
modelling process, from data collection through to the provision of forecasts at link level.
Many modern approaches to the analysis of travel demand adopt a sophisticated treatment of
the choice processes and choice situations (e.g. constraints) of the individual travellers, but are
unable to take the process through to the production of forecasts, perhaps because data is
required that cannot reasonably be forecast. In this paper attention is given throughout to the
Draft 12 August 1990
problems of forecasting as well as to model development.

Rather than considering the whole modelling process, this paper focusses on the forecasting of
mode and destination choice, the latter usually termed 'trip distribution' by classical modellers.
This selection has been made partly because consideration of the whole field would go beyond
the scope of a single paper, and in any case the boundaries of the modelling process have
themselves become more variable, in response to the criticism of inflexibility of the classical
models. The two choices that are considered are themselves very closely related in ways that
will be discussed in more detail below and are particularly interesting from the point of view
of aggregation. Moreover the two choices have been studied intensively, yielding many relevant
results from the point of view of data aggregation.

This paper is concerned with the aggregation of exogenous data, that is the extent to which
data that is used for items other than the travellers' behaviour (which is 'endogenous' in the sense
that the model aims to predict it) represents an average over a group of travellers, rather than
being calculated or collected specifically for one individual. In the case where the model at
base represents the behaviour of more than one individual, as is always the case with the four-
stage models (which usually represent the behaviour of a population segment, e.g. non-car-
owners, living in a zone), some degree of aggregation of the exogenous data is inevitable;
in these models the issue of data aggregation is the extent to which greater accuracy (i.e.
smaller zones or more segments) is desirable. When the model at base represents the
behaviour of single individuals, i.e. a disaggregate model, it is conceivable that exogenous data
can be obtained and used separately for each traveller; the issue is then whether, as often, it is
preferable on cost or other grounds to use less accurate data.

The possibility of using disaggregate models for the development of aggregate forecasts has
been known for some time.(1) These methods turn out to be particularly advantageous for the treatment of mode and destination choice.
While current disaggregate methods do not pretend to deal with all of the issues raised by some critics of the four-stage models, they do offer substantial advances
over the traditional models while remaining practical in applications studies. Forecasting systems, including mode and destination choice components and based
on disaggregate models, have been reported from a large number of countries
(2), including The United States, The Netherlands, Chile, France, Brazil, Israel,
Australia, Sweden, Norway, Japan and even Britain, while further systems known to the authors are in operation or development in most of those countries and in
Denmark and New Zealand. It is clear that there is a substantial and growing body of practical experience with these methods.

The differences in practice between purely aggregate and disaggregate-based forecasting systems have often been overstated. In many cases there is complete
equivalence between the forms of the forecasting models (Daly, 1982b). The essential difference is the treatment of the description of behaviour in the two
approaches, particularly during the model development process. It will be argued below that, for the development of models, the disaggregate approach is clearly
superior to the aggregation of behaviour zonally (e.g. into trip matrices) or over pre-defined segments. Because of this, and because of the wide application of
disaggregate modelling in practical studies already noted, the paper will deal exclusively with disaggregate modelling, i.e. with models that represent, at base, the
behaviour of individuals. The issue of data aggregation remains the central concern.

In summary, the topic of the paper is the determination of the appropriate level of aggregation of the exogenous data in disaggregate models of mode and
destination choice. This subject has substantial interest in its own right, as shown by the numbers of practical studies completed and in progress. The methods
presented are also applicable to disaggregate modelling of other choices, and, to some extent, to aggregate modelling as well.

The following section attempts to define the objectives that the process of model design must try to satisfy: essentially the trade-off of errors of various kinds, and
the losses they may cause, against the costs of reducing them. Section 3 sets up a theoretical framework for the analysis of aggregation, and shows how different
Draft 12 August 1990
kinds of measurement or aggregation error in the base data will lead to different kinds of error or bias in the forecasts made. The fourth section presents some
results from practical studies, indicating the effects of data aggregation on model estimation in varying contexts. Finally, the concluding section tries to draw
together the findings of the paper in a specification of "good practice" with regard to data aggregation.

2. Errors in Modelling and Forecasting

In setting out criteria for the design of a forecasting model, whether for travel demand or in any other field, it is necessary to take account of the errors that are
inevitably present in the information to be used, both in setting up the model and in applying it to make the actual forecasts. The correct form of the model is
also not known and approximations or mistakes in this respect will also give rise to inaccuracy in the forecasts. It is worth noting that many of the statistical
procedures used in model estimation and application assume, implicitly or explicitly, that both data and model form are known exactly.

A model to derive forecasts of demand for transportation systems - i.e., in the disaggregate framework, forecasting the numbers choosing each of the alternatives -
must balance the issues of model complexity and data accuracy, within a defined study budget, to give the most accurate forecasts possible. Model complexity and
data accuracy interact to influence different types of error in different ways, and it is important to try to be clear about these types of error. Broadly, two groups can
be distinguished:

- errors that would cause even correct models to yield incorrect forecasts; e.g. errors in the prediction of the explanatory variables, in model transfer and
in aggregation;

- errors that cause incorrect models to be estimated; e.g. measurement, sampling, specification and estimation errors.
In this section, consideration is given first to the types of error that may arise, with the broad effects they may cause. Then, the trade-off between model
complexity and data accuracy is examined before turning finally to the "design variables" that can be altered in the specification to improve the model within the
budget available. Throughout, interest is focussed on the issue of data accuracy as a function of aggregation.
2.1 Classification of Error Types
Several authors have offered classifications of the types of error that may enter forecasting models (Alonso, 1968; Wilson, 1974; Tardiff, 1979, Horowitz, 1981).
The following list is based on their work.
2.1.1 Measurement Error
The term "measurement error" is used here for the errors that arise in the actual measurement process, i.e. in the base year. Many of these are connected with
survey procedures, such as misinterpretation of the questions by respondents, erroneous recording of answers by the interviewer, coding and computation errors.
Others arise in the preparation of other data, such as in digitising or other network preparation steps; where there is also plenty of scope for computational error.
Measurement error can obviously be reduced both by improving the techniques used (e.g. by appropriate use of computerised interview support) and simply by
expending more effort; however, both of these cost money. For this reason, in particular, it may be expected that measurement errors will be greater in
developing countries.
Measurement error as defined here should be distinguished from the difficulty of defining the variables that ought to be measured. The complexity that may arise
in this area is indicated in Figure1. Ideally, modelling should be based on the information perceived by individual travellers. While reported data may give some
insight into perception, the use of this type of data raises the problem of how it is to be forecast. However, the use of the alternative " engineering" estimates, which
can readily be used in forecasting, introduces the error that perceptions may vary across alternatives, being particularly inaccurate for non-selected alternatives.
[insert Fig 1 here]
2.1.2 Sampling Error
Sampling error arises from the fact that models must be estimated using finite data sets. These errors are approximately inversely proportional to the square root of
the number of observations (i.e. to halve the error it is necessary to quadruple the sample size), so that reducing them is very costly indeed. Daganzo (1980) has
examined the problem of defining sampling strategies optimal in the sense of optimising estimation accuracy in this respect.
In this context it is worth noting that many statistical tests commonly used in practice, for example the approximate "T" ratio tests in logit models, are based on
asymptotic properties of the models and do not necessarily apply when the sample is "small". Very little seems to be known about the behaviour of typical
models with small samples or even how small a sample has to be to cause problems.
Draft 12 August 1990
2.1.3 Computational Errors
Computational errors arise because modelling is based on iterative procedures for which the exact solution, if it exists, has not been found for reasons of
computational cost. Typically it seems to be believed that these errors are very small in comparison to other errors, although recent work (Horowitz, 1989) suggests
that, for data extracted from assignments to congested networks, this belief may be misplaced. Problems are particularly severe when working with model systems
that do not ensure consistency between supply and demand
(3).
2.1.4 Specification Errors
Specification errors arise because the analyst does not perfectly understand the choice and other processes being modelled or because it is necessary to simplify
the process for whatever reason. The following are important subclasses of this type of error.

- omission of a relevant variable, perhaps the most common specification error; interestingly, models incorporating an explicit random error term, as in the
utility functions of the disaggregate models considered in section 3, are designed to accommodate this error through the random term, i.e. the extent of
the error is known; Horowitz (1981), however, shows that problems can arise when the excluded variable is correlated with included variables or when
its distribution in the population to be forecast is different from its distribution in the sample used for model estimation;

- inclusion of an irrelevant variable (i.e. a variable that does not influence travellers' choices); Tardiff (1979) shows that, in models where the variables
appear in linear form, the inclusion of such a variable will not bias the model or the forecasts, although sampling error will be increased; in a non-
linear model, however, biases may be caused;

- exclusion of taste variation, which has not been incorporated in practical models, with a few exceptions
(4); Horowitz (1981) shows that this error will
always produce biased models;

- simplification of the form of variables, for example using a variable in linear form to represent a non-linear effect;

- other specification errors, in particular the use of inappropriate model forms, such as the use of compensatory models to represent behaviour that
might be non-compensatory, the omission of behavioural effects such as habit or inertia, or the use of logit model forms when a more complex
correlated model might be appropriate; a full discussion of these forms of error is given by Williams and Ortúzar (1982a).
Clearly, errors of this type occur in most if not all models applied in practice. Moreover, most statistical analyses ignore the possibility of specification error
(except for the simple case of omitted uncorrelated variables). Removal of specification errors would in many cases require extensive behavioural research; it must
simply be conceded by the modeller that such errors may be present in all feasible models.
2.1.5 Transfer Errors
Transfer errors occur when a model developed in one context (time and/or place) is applied in a different context. Adjustments may be made to compensate for the
transfer, but ultimately the fact must be faced that behaviour may be different in different contexts (Gunn and Bradley, 198x).
In the case of spatial transfers, the errors can be reduced or eliminated by partial or complete re-estimation of the model to the new context, although to do so is
to discard some of the substantial cost savings obtainable from model transfer. However, in the case of temporal transfer (e.g. forecasting), this re-estimation is
obviously not possible and the potential errors have to be accepted.
2.1.6 Data Aggregation
In most practical model systems, the data that is used to define the choice situation of each traveller is to a greater or lesser extent aggregated. Even when the
traveller is asked to report the characteristics of the alternatives, he or she can only have based the choice on the expected values of these characteristics. When
network models are used, there is inevitably an aggregation over routes, departure times and locations (i.e. zones) which means that the values obtained for the
explanatory variables are, if correct, averages for a group of travellers rather than exact values for any one traveller.
Clearly the quality of any model can in principle be improved by reducing aggregation error of this type. Equally obviously, an important limitation of the extent to
which such improvements can be implemented is the cost. Generally, reducing aggregation error implies making measurements under many more sets of
circumstances: more zones, more departure times, more routes, more socio-economic categories; this costs money and time and increases complexity, which brings
with it further costs and difficulties in all aspects of the modelling.
The issue of the appropriate level of data aggregation is the main topic of this paper. The above discussion indicates that error introduced by data aggregation is
only one of many sources of model error. The determination of the appropriate allocation of modelling resources to the reduction of aggregation error has
Draft 12 August 1990
therefore to be done with reference to these other errors, whose severity will of course be different in different modelling contexts.
In the following sections attention is given to the impact of data aggregation error on model accuracy and how the model designer can manipulate the design to change
the accuracy so as to best meet the conflicting needs of accuracy and simplicity.
2.2 The Model Complexity - Data Accuracy Trade-Off
Data accuracy, as indicated above, is affected by measurement error (and possibly computational error) as well as by aggregation. It is therefore not reasonable to
expect that variables can ever be measured or forecast with perfect accuracy. Nevertheless, in the light of the previous discussions, it is interesting to consider how
changes in measurement error could change the overall error in the model.
Forecasting and model estimation are considered separately.
2.2.1 Forecasting
Suppose that some statistic z of interest, such as the flow on a particular transport link, or total revenue, is forecast by a function f of measured variables x
r:

z = f ( x1, x2, ..., xn ). [1]

Of course, in most realistic cases f is a very complicated function that cannot easily be
stated.

The variables xr are each measured with an error, part of which is due to the process of measurement (section 2.2.1 above) and part of which is due to the
fact that the variable measured may be an aggregated measure of the true value for a given decision maker (section 2.2.6 above). If for each variable x r the
sum of these two components gives a standard error rr, then the total error rz in the forecast z due to measurement error can be
approximated by
rz2 = R (Zf/Zxi)2.rx2 + R R (Zf/Zxi).(Zf/Zxj).rx .rx .qij,
i j
i i j]i
[2]

where qij is the correlation in the measurement error of x i and xj, which can reasonably be taken to be zero in most contexts. Equation [2] is exact when
the function f is linear and a good approximation in other cases (Alonso, 1968).
Using the approximation appropriate to a linear function, and making the assumption of zero correlation, the rate of improvement in the total error with respect to an improvement in measurement
of a variable is given by
2
z i
Zr /Zr = (Zf/Zx ) i . rx / rz. [3]
i

This equation quantifies the obvious point that the variables of highest priority for error reduction are those:

(a) with the largest relative error;

(b) with the greatest impact on forecast demand.

In principle, based on equation [3] and the relative costs of reducing the measurement error in each variable,
an overall "optimum" investment of effort in measuring each variable could be determined. Equation [3] shows
clearly that the return from improving the accuracy of a particular variable will diminish as more and more effort
is put in; this adds to the fact that it is easier to make the first step in reducing the error in a variable than to obtain
very high levels of accuracy. The conclusion is very strong that a good experimental design will balance the
investment in measurement of the important variables.

In practice, of course, equation [3] omits all the sources of error, other than measurement error, that were
discussed in sections 2.1.2 to 2.1.5 above. Of these error sources, sampling error, computational error and
transfer error can reasonably be taken to be independent of model complexity. However, specification error can
be reduced by increasing the complexity of the model: including more behavioural effects. Unfortunately, not
only does this increase the cost of measurement of the increased number of variables, it is clear from equation [2]
that including more variables (giving more terms in the equation) can also increase the total contribution of
measurement error to the forecasting error. Again it can be concluded that an effective model design will
involve finding a balance: specification error against measurement error.
Draft 12 August 1990
This last point is illustrated in Figure 2, which shows the increase of measurement error rz with complexity, set
against the reduction of specification error, labelled rs. An interesting aspect of this trade-off is that the
"optimum" balance may depend on the technology available. In a richer country, measurements can more easily
be made accurately than in a poor country: the consequence, shown in the figure, confirms the intuitive
expectation that, ceteris paribus, more complex models should be more appropriate for richer countries and
simpler models for poorer countries.

[insert Fig 2 here]

The ultimate trade-off with which the modeller needs to be concerned is that of cost against accuracy. Thus the
full costs of complexity need to be considered: not only the measurement in base and forecast situations of all the
variables, but also computer resources and, not least, the need to make forecasts credible by explaining them
clearly to decision-makers. Equally, the broader aspects of accuracy need to be weighed carefully; these include
the credibility obtained by giving a good base-year explanation of behaviour. The full trade-off is rather subtle.

This discussion of measurement accuracy and specification complexity must of course be seen in the context that
there are several other forms of error on the list, and these severely restrict the extent to which the accuracy of
the model can be improved. Nevertheless, measurement accuracy remains an important issue in determining the
forecasting accuracy of a model and thus a key influence on the best design to minimise forecasting error.
Further attention will be given to this point in Section 3 below.

2.2.2 Model Estimation

During model estimation, the analyst obtains estimates of the unknown parameters incorporated in the model and
also estimates of the error with which those parameters are estimated. These error estimates, on the basis of which
decisions must be taken about the model specification, may be considered correct to the extent that the errors listed
above are made explicit in the basic specification of the model. Other types of error that are not made explicit
in this way are not properly assessed.

Among the errors that may be considered to be assessed correctly in common model types (e.g. multinomial,
linear-in-parameters logit models) are sampling error and the specification error due to an uncorrelated missing
variable. Measurement error, computational error, other forms of specification error and transfer error are not
properly assessed, although they will to some extent be reflected in the reported standard error of estimation given
by an estimation procedure.

In this context, the analyst must exercise careful judgement in determining the appropriate model specification.
For example, a variable that seems a priori relevant to the choice being modelled may not appear a posteriori to
add greatly to explanation; the decision whether or not to incorporate the variable in the model must weigh the
issues of complexity and measurement accuracy on the one hand with the risk of increasing specification error on
the other hand. Once again, a careful balance is indicated.

In summary, therefore, it is clear that although the errors reported by estimation procedures for the estimates of
key parameters are an essential input to the assessment and improvement of the model, the analyst must also
Draft 12 August 1990
consider the reduction of specification error as an important objective. In meeting this objective, there is no
substitute for judgement based on a priori notions, enlightened by findings based on the data.

2.3 Balancing Competing Requirements

From the discussion in Section 2.2 above, it can be concluded, both for model estimation and for model
application (forecasting), that measurement error is an important contributor to the total error in the model. In
particular, measurement error influences the analyst's choice of the appropriate level of complexity of the model:
more measurement error implies that the model should be less complex (as always, ceteris paribus).

The objectives of the analyst in designing the model should be to obtain the best forecasting accuracy possible
within the resources available to the study. It is clear from the above discussion that this is a very difficult
problem; this paper will not attempt a complete solution, nor is it likely that trying to find a complete solution
is worthwhile. The above discussions indicate that the best plan is to aim at a reasonable balance of effort in
attempting to reduce the various sources of error, taking into account their relative error contributions,
importance in influencing forecasts and the relative costs of reducing error.

For measurement error itself, however, it is possible to take the discussion further. This is begun in a discussion of
theoretical aspects in Section 3, indicating that different designs may be appropriate in different circumstances.
This indication is confirmed for some practical circumstances in the empirical results reported in Section 4.

3. Theory of Aggregation

The preceding section listed the various types of error that can occur in the data and in the specification of a
travel demand forecasting model. This section describes in more detail how these errors in input and
specification can work through into the forecasts output from the model.

The discussion is greatly simplified by the explicit assumption of a specific behavioural paradigm: we shall assume
that travellers choices are motivated by the wish to maximise their 'utility'. As noted by McFadden, (5) this can be very
general; indeed, it is difficult to imagine the circumstances in which the hypothesis could be disproved (Daly, 1982a). In any case, most of the following discussion could probably be based
on an alternative paradigm, if that were wished, but the utility-maximising assumption is by far the best known and most studied in the transportation context.
Following a brief outline of model specification in the utility-maximising framework, attention is switched to the consequences of the types of error noted in section 2 for the crucial assumptions

about error variance and then to the implications of these findings about error variance for forecasting.
3.1 Utility-Maximising Choice Models

Models based on utility maximisation attribute to each alternative c in the choice set C for each individual (the choice set may vary between individuals) a utility U
c , for which the
analyst is able to obtain a measured approximation Vc:

Uc = Vc + ec, for c + C . [4]

The error difference e between the true value of the utility and the analyst's approximation is not known and
cannot be observed; assumptions can however be made about its distribution in the population. The nature of
these assumptions are the key to determining the form of the model, since the probability p c of choosing alternative c from the
set C is just
pc = Pr{ Uc = Uk , ¦ k + C } [5]

(the utility-maximising rule), which can be expressed by

= Pr{ ec - ek = Vk - Vc , ¦ k + C } , [6]

which obviously depends on the measured values V and the assumptions made about the distribution of e only.
The simplest such assumption, that the e's are independent and identically distributed (i.i.d.), leads to the simplest
models, such as the familiar multinomial logit model.(6)
The form of the approximate-utility function V is also of interest. Most often a linear form is chosen:
Draft 12 August 1990
V
1 c1 + h2.xc2 + ... + hr.xcr ,
c = h .x [7]

in which the x's represent directly-measured variables, such as travel time, and the h's represent unknown
parameters that are to be estimated.

It is in this context that we can begin to quantify the effects of the errors listed in Section 2. The error term e
summarises all of the error in the model, since it gives the difference between the analyst's approximation V and
the traveller's true utility U for each alternative. Each of the various types of error described in Section 2 can
be expressed in terms of a failure of V to model U accurately. In many cases the magnitudes of these errors can
be quantified, giving insight into the best directions to look for improvements in the model.

However, if we look at the nature of the types of error described in the preceding section, it is apparent that the
simplest assumptions such as linear V's and i.i.d. e's can be true only as approximations. For example, 'taste
variation' implies changes in the values of at least one of the h's (say ht) between individuals; it is then clear from
the form [7] of V that the contribution of taste variation to the total inter-personal variation e will depend on the
value of the associated xct for each alternative: i.e. the i.i.d. assumption can no longer hold exactly. The consequences of the failure of these assumptions for the accuracy of
model estimation require to be investigated.
Similarly, if a model of this form is to be used in forecasting, it is necessary that the probabilities [6] can be calculated in the same way as when the model was estimated. Essentially this requires

that the errors e have the same distribution under forecast conditions as they had when the data was collected. For reasons similar to those mentioned above for model estimation, several
components of error cause this assumption to be questionable.

Thus the various types of error have different consequences for the assumptions about the variance of the error term e; some of these are acceptable within the simplifications commonly
incorporated into the model framework, others are not. The nature of these different error types and their influence within the utility-maximising framework will now be investigated.
3.2 Consequences of Error for Variance Assumptions
In Section 2, we noted that several different types of error contribute to the total error in a forecasting model. These different types of error can now be considered from the point of view of

utility maximisation to illustrate and begin to quantify their impact on model estimation. The consequences for forecasting will be considered in the following section.
Sampling error is in principle incorporated explicitly in the statistics underlying model estimation. That is, the magnitude of the error caused by the smallness of the sample is reflected in the
reported errors in the estimated parameters, provided the sample is not too small. This proviso is required to ensure the validity of the standard statistics reported by estimation programs,
which usually assume that the data set is large. Not much is known about the small-sample behaviour of standard model estimators.

Computational error in this context refers to an error in the calculation of variables used in the utility functions of the alternatives. This type of error is by its nature extremely difficult to
handle, because any error of this type is liable to be systematic, because the direction and magnitude of the error is difficult to predict and because the error may well be closely associated with one

or two variables only. Treatment of computational error is beyond the scope of this paper.
Terms omitted from the model (i.e. omitted from the V function) are in principle simply a contribution to an overall error, not related to any other term. In terms of equation [4] above, an omitted
term can be represented as
ec = Uc - Vc - Oc [8]

where Oc is the value of the omitted term for alternative c. Here it seems that the variance of ec is simply increased by the variance
of Oc over the observations, i.e. there is a simple addition to the error without further consequences.

However, equivalently to the point quoted from Horowitz (1981) in section 2.1.4 above, problems arise when the
omitted term is correlated with other terms or when its variance changes between model estimation and model
forecasting. In this case the model will be in error by an unpredictable amount, so that specification error of this
kind should be avoided or reduced wherever possible by including all possible variables believed to be relevant to
the choice process.

The inclusion of an irrelevant term will present a problem when its coefficient is large, since the error will be
increased by the product of the variable and its coefficient. Again in terms of equation [4], we can write

ec = Uc - Vc + Ic [9]

where Ic is the value of the irrelevant term for alternative c


c, already included in V . The magnitude of the problem is obviously determined by the size of the term, i.e. the product of the value
of the variable and its coefficient. Care should therefore be exercised to ensure that variables included in the model can reasonably be expected to influence behaviour, and are not simply the

result of statistical fluctuation.


Taste variation implies that the value of h varies between different travellers. Since the usual assumption is that h is constant, an error will be introduced that is proportional to the value of the

corresponding x variable. This can easily be seen from the equations


e c = Uc - Vc = Uc - (h1.xc1 + .. + hr.xcr) [10]
and therefore
Draft 12 August 1990
2 2
c c1.Var(h1) + r(1)
Var(e ) = x [11]
2
where r(1) is the variance due to terms other than 1 in the approximation V (assuming this to be independent of the variation in term 1).
Clearly, the omission of this type of effect will (if the effect in fact exists) make both the estimated model and forecasts incorrect. Unfortunately, the incorporation of taste variation would make
the model very much more complicated; however, if the effect is substantial, it must be included to avoid these inaccuracies. Models incorporating an analytical description of taste variation have

been proposed (see note 4), but they have not been carried through to use in practical forecasting. A simpler solution is to use segmentation of the population to allow for the most important
variations in taste.

Other specification errors are naturally difficult to discuss, because the effect they have depends on the nature of the error. However, models should be carefully checked for the presence of non-
.
(7)
linearities with respect to the important variables
Measurement error is almost always present. In terms of the equations above

2 2
Var(e c) = h1.M(xc1) + .. + hr.M(xcr) + V*(ec) [12]

where M(x) is the variance of the measurement error in x (assuming independence among the measurement errors)
and V*(ec) is the error in ec arising from sources other than measurement error. Equation [12] gives essentially the
same information as equation [3], that the contribution to the total forecasting error variance of a measurement
error in a variable is proportional to the square of the variable's impact on demand.

It is through measurement error that the main influence of aggregation is felt, larger zone sizes (for example)
leading to larger measurement error. Depending on the measurement procedure used, there may well be a
dependence of the measurement error on the value of the variable in question: high values would usually be
measured with a larger (absolute) error. In this case the i.i.d. assumption would no longer be applicable. In some
cases it is possible that increased aggregation could lead to an increase in the correlation of the error variance with
the values of the variables being measured.

In conclusion it is clear that the types of error listed in Section 2 will often lead to errors in the model that are
not consistent with the assumptions (a) of i.i.d. and (b) that the error variance remains constant between
estimation and forecasting. This is apart from the problems of specification error (e.g. the existence of
unrecognised non-linearities), for which some improvements can be made in the model in some circumstances,
but which will always remain as a hidden and poorly understood source of inaccuracy.

In particular, there seems to be good reason to suppose that there is a dependence of the error variance on the
values of the variables, larger values being measured (absolutely) less accurately (i.e. M(x) increases with x).
This failure of the i.i.d. assumption is a special reason for concern, because it could easily lead to distortions
in the value of h as model estimation tries to meet as closely as possible the input i.i.d. assumption.

This connection is clear in equation [12]. To maintain constancy in the error variance, small values of h are
required for large values of M(x) (i.e., probably, large values of x); this can work though in model estimation as
an apparent need for non-linear transformations to the variables, but these cannot be identified separately from
changes in h itself as a function of x (Ben-Akiva et al., 1987).

3.3 Aggregate Forecasting with Disaggregate Models

The problem of aggregate forecasting with a disaggregate model can be stated as follows.

Given a model predicting the probability of behaviour of any individual with


known characteristics, facing alternatives with known characteristics, how do we
make a forecast over a whole population?

Clearly, the essential information needed is a description of the distribution of the characteristics
of the individuals and the alternatives available to them across the whole population for which
forecasts are to be made.
Draft 12 August 1990
Given such a description of the base variables, the demand model has to be evaluated explicitly
or implicitly for every combination of variable values. Mathematically, an integration must
be performed (Daly, 1982c). There are two classical methods of performing the integration:
averaging over subpopulations (e.g. zones and population segments) or random sampling
from the population. These two methods have different advantages and disadvantages so that
the choice between them would be different in different contexts.

Random sampling from the population, usually called sample enumeration (Ben-Akiva and
Atherton, 1977; Daly and Zachary, 1977) is a simple form of Monte Carlo method
(Hammersley and Handscomb, 1964). This method is generally to be preferred when it is
feasible, because of its lower cost and ability to accept more sophisticated models: in
particular, the exact model that was estimated. There can be problems in long-term
applications, although the method of 'prototypical sampling' solves many of these (see Gunn et
al., 1987). However, for large-scale forecasting (e.g. mode-destination choice over a large area)
the computer time required for sample enumeration can be excessive.

Calculating the average demand for a part of the population (e.g. for a specific population
segment and geographical area) is the classical method of transportation planning. This method
can be quicker in computer time but can also reduce the accuracy with which the model is
evaluated. A serious problem arises when this reduction of accuracy causes a change in the
error variance between the estimation and forecasting models.

3.4 Changes in Error Variance in Forecasting

A serious problem that can arise in forecasting with a demand model is "aggregation error".
This error is caused by a difference in the error variance between the data on which the model
was estimated and the data used for forecasting. Usually, a change of this type arises when the
data used for model estimation is more accurate than the data used for forecasting.

This problem is serious because, ceteris paribus, the predicted change in demand for a given
change in the value of a variable included in the model is inversely proportional to the
standard error of the attractiveness functions. This follows directly from the basic equation
giving the probability for choosing an alternative [6]. In fact, if the e's are i.i.d. with variance re2,
then the probability of that alternative c is preferred to alternative k is
p
c|k = D( ( V c - Vk ) / r e ) [13]
where D is a fixed cumulative distribution function (zero mean and unit variance) whose form depends on the specific assumption made concerning the distribution of e.
Thus, if a model is applied with coefficients estimated on data containing less error than is present in the forecasting data, r e will be too small, and changes in

demand will be overestimated. This problem applies to model applications made using both the sample enumeration and averaging procedures.
A change in error variance can arise in two ways: from a change in the actual circumstances or a change in the measurement method.

3.4.1 Change Between Base and Forecast Circumstances


A real change between base and forecast circumstances, may change the error variance in several ways. One example is given by Gunn (1985); there, measurement

error is supposed to be proportional to the value of the variable to be measured; a doubling of the value (as a result of policy, perhaps) thus leads to a doubling of the
measurement error and to a bias in the forecast. The assumption of exact proportionality is not essential to Gunn's analysis, any change in measurement error
resulting from a change in the value of a variable will give the same effect, although perhaps to a reduced extent.
In the notation of equation [12], the possibility studied by Gunn is that M(x) varies with x. Given the nature of many variables used in modelling, this dependence is

very likely. For example, the cost of buying petrol for a specific trip (a common variable in modelling) can be calculated by multiplying the trip length by fuel
consumption by the price of petrol. The main source of measurement error is likely to be in measuring the trip length. Thus, if fuel consumption or petrol prices

change (a scenario commonly investigated), both the value of x and the measurement error will change.
However, other causes can also have the same effect. For example, if there is any 'taste variation' present in the data (see equation [11]), then changes in the

value of the corresponding variable will lead to changes in the total error and hence to bias in forecasting. A similar result can be obtained from various forms of
specification error, for example a perception error that depends on the value of a variable (so that 5 minutes is perceived more accurately than 60 minutes), so that a

change in total error could result from a change in the value of a model variable.
Thus with respect to the introduction of forecasting error as a result of changes in the values of variables, aggregation and measurement error play a rôle, but not a
unique rôle. Every model application outside the range of values used in estimation must be viewed with increased suspicion; the introduction of forecasting bias is
Draft 12 August 1990
simply another reason for this well-known conclusion.
3.4.2 Change between Base and Forecast Measurement Method

A change in variance can also arise because of a change from the measurement method used in estimation to a new method for forecasting. Clearly, this can
change the whole nature of the function M in equation [12]. Most often such a change arises when accurate data is available for model estimation, but forecasting has
to rely on zone and/or population segment averages. When there is such a change in measurement method, forecasting bias can be expected unless explicit steps are
taken to correct the model.

The error caused when models estimated on accurate data are applied to less accurate data has usually been called aggregation error. This error has long been
(8) A crucial difference between this error and those
recognised, and techniques have been developed to reduce or eliminate it.
caused by changes in the variables is that aggregation error is much more easily controlled by
the analyst.

3.4.3 Avoidance of Forecast Bias

Two mechanisms have been described that can lead to changes in the total error variance in the
model between estimation and forecasting. These changes are particularly unwelcome because
they lead to biases in forecasting.

In general, changes of the first type, often caused by a link between the values of the variables
and the accuracy with which they are measured, are difficult to deal with. In contrast,
changes in error variance caused by changes in the measurement method can usually be
eliminated completely.

Dependence of total error on the values of variables that are likely to change in the forecast
situation can of course be reduced by changing the specification of the model or by improving
the methods used for calculating ("measuring") the variables in the forecast situation. While
in principle the changes in variance caused by this dependence can operate in both directions,
in practice it would be reasonable to expect that variance would increase (there may be errors
caused by unforeseen features of the forecast situation) and therefore that the forecast changes
in behaviour may be overestimates of the change that will in fact occur.

Dependence of total error on the measurement procedure can usually be eliminated by using
the same measurement method in both estimation and forecasting. When more accurate data is
available for estimation than can be provided for forecasting, a common situation, the model
can usually be corrected by calculating the increase in variance explicitly, i.e. by applying the
forecast measurement procedure to the estimation data. Often a convenient way of using this
procedure is use it to prepare data for re-estimation of the model (Cambridge Systematics,
1982). Either the model may be completely re-estimated or, preferably, a parameter may be
estimated to scale the coefficients estimated on the accurate data to the variance level of the
less accurate data.

3.5 Conclusions from Theoretical Discussion

Each of the error sources listed in Section 2 can be represented within the utility-maximising
framework. This method has been used to give an analytical framework within which the
contribution to total error of each of the various sources can be analysed and the contribution of
measurement error and aggregation can be put in context.

As in the less formal discussion of Section 2, it is apparent that measurement error is only a part
of the total error and that the variables for which it is most productive in reducing the
measurement error are those that are most important in influencing demand and those with the
largest error.

A specific problem arises when the model that has been estimated is to be used for forecasting.
Draft 12 August 1990
A change in the total error variance in the model between estimation and forecasting will lead
to biases in the forecasting. Such changes can arise because of dependence of measurement
error on the value of variables or because of changes in the measurement method. The former
cause of change can be reduced by improving the model specification but seems impossible to
eliminate completely; the latter cause can be eliminated by using the same measurement
procedure for estimation as for forecasting, perhaps in a two-stage estimation.

However, it is clear that a theoretical analysis of this problem cannot give a complete solution.
Many of the decisions that the analyst needs to take depend on the relative values of the errors
and the coefficients of the variables. Naturally, this dependence implies that the conclusions
reached may be different in different contexts. For this reason, the following section reviews
the empirical findings obtained in previous studies and gives new results from a particular
data set.

4. Empirical Results

In this section we refer to previous results and present new analyses from a series of
practical studies which have been conducted precisely with the aim of answering some of
the questions posed in the previous sections; although they use data and models developed for
Santiago (Chile) their generality is supported by similar information available for European
experience.

An important consideration in the discussion is the relative merits of modelling mode and
destination choice separately or together. Separate modelling often allows more accurate data
to be used; this must be set against the improved model specification and increased amount of
relevant information available when modelling the two choices together. As is indicated in
the theoretical discussions above, different circumstances can lead to different decisions as to the
better choice of these two.

4.1 Previous Experience: Mode Choice Modelling

Interest on the problem of disaggregate mode choice model estimation using attribute data
measured at different levels of accuracy has prompted empirical studies on:

- the effect of using more precise measures for the variable 'waiting time' in public
transport options; the results reported by Ortúzar and Ivelic (1986), which clearly
support the need for more detailed values, entailed replacing crude measures based
on the average frequency at different distances to the CBD by more accurate values
obtained with the aid of state-of-the-art public transport assignment models (e.g. see
De Cea et al., 1988);

- the impact of using temporally disaggregated information (i.e. replacing peak averages
by more precise values for each individual depending on the exact timing of their
journeys) for the in-vehicle travel time variable on the specification and stability of
mode choice models (Ortúzar and Ivelic, 1987a); it was concluded in this case that
although the more accurate models dominated their less detailed counterparts, the
differences were not striking suggesting that in this case cost could be given priority
over accuracy; it is interesting to mention, however, that a similar study by Reid and
Small (1976) concluded that the use of peak averages resulted in models with biased
coefficients and that the bias could become important for relatively complex functions;

- the effects of using spatially aggregated (i.e. zonal) data for level-of-service and
individual attributes in the estimation of multinomial and nested logit mode choice
functions; in this case the results were quite definitive, as the coefficients of the
Draft 12 August 1990
detailed models were significantly different and better than those of the models
estimated with zonal data; also, with the latter the preferred tree structure (see Daly,
1987) for the detailed data was rejected (Ortúzar and Ivelic, 1987b).

Thus, it appears that previous empirical experience in mode choice modelling supports the
conclusion that the relative importance of the various considerations depends on the context, as
suggested in the preceding sections of this paper.

4.2 Spatial Aggregation and Destination Choice

It is clear from the discussions of Section 3 above that it is important to use the most accurate
and detailed data available in modelling traveller behaviour. However, these arguments can
lead to different conclusions in different contexts: in particular in the contexts of mode and
destination choice modelling.

For modelling mode choice alone, the task of collecting fully-detailed data describing the
alternatives available to each traveller is feasible. Since the results of spatially aggregated
models have often been substantially inferior to those using more detailed data (see 4.1 above),
detailed data should be used whenever possible.

However, for modelling mode and destination choice together, as is now becoming more common
because of the improved model structure obtained, even the specification of the alternatives is a
daunting task, let alone the determination of their characteristics. The spatial aggregation of
elementary alternatives into zones and the measurement of the characteristics of the alternatives
at this level of aggregation becomes a more attractive proposition. Clearly, measurement
at zonal level can reduce data collection costs substantially, as well as solving the problems of
defining alternatives. If, in addition, the quality of the models thus obtained can be shown to
be competitive with those obtained from fully-detailed (mode choice) analysis, the procedure of
working at zonal level can be recommended as standard practice.

A study in Grenoble (Daly, 1985) had shown that the inclusion of destination choice in the
choice framework could improve the quality of models relative to those estimated solely on
mode choice. However, the data measurement methods of that study were somewhat unusual
and it could not be claimed that an improvement would generally follow from the extension of the
choice context.

It was therefore desirable to set up an experiment in which this issue of the relative importance
of data aggregation and model structure could be assessed in circumstances more general
than those of the Grenoble study. The design and results of this work are now described.

4.2.1 Design of the Experiment

The experiment was designed precisely to test the relative merits of data disaggregation and
mode-destination choice integration; journey-to-work data collected in morning peak-hour
conditions in 1983 and 1985 for two corridors of Santiago was augmented by similar but
city-wide data collected in 1986 for the Great Santiago Strategic Transport Study
(ESTRAUS) yielding a total sample of 1,591 observations. As the data was collected at the
detailed level information was available from the start on each attribute at this level; as part of
the experiment the data was later recoded (see Ivelic, 1988 for details) at a spatially
aggregated (zonal) level following the recommendations of the ESTRAUS consultants (see
Daly, 1986). This was the data used by Ortúzar and Ivelic (1987b) in the study mentioned
above, further analysed by the same authors (1988).
Draft 12 August 1990
Table 1: Choice and Availability in the Complete Sample

Alternative Choice (%) Availability (%)


------------------------------------------------------------
1. Car Driver 18.7 53.8
2. Car Passenger 7.3 48.2
3. Shared Taxi 4.5 71.8
4. Metro only 24.2 31.8
5. Bus only 27.9 96.8
6. Park & Ride (metro) 6.2 26.9
7. Kiss & Ride (metro) 3.4 27.3
8. Shared Taxi & metro 2.6 23.5
9. Bus & metro 5.2 32.1
------------------------------------------------------------

Table 1 shows the percentages choosing and having available each alternative for a total
sample size of 1,512 individuals (79 observations were eliminated, mainly bus users,
because of inconsistencies in the definition of alternative availability at the two levels of
aggregation). Table 2 gives the variables used for model estimation.

Table 2: Definition of Explanatory Variables

Variable Definition.......
----------------------------------------------------------------
1. TIME : in-vehicle travel time (min)
2. WAIT : waiting time including transfers (min)
3. WALK : walking time (min)
4. COST : travel cost divided by per capita family income (min)
5. COMP : min (N° of cars / N° driving licences, 1)
6. SEX : dummy; 1 for males, 0 for females
7. LSIZE : logarithm of the destination attraction variable
(number of places of employment in the zone)
----------------------------------------------------------------

To estimate the destination/mode choice models, it was also necessary to generate a set of
alternative destinations for each individual in the ESTRAUS sample (the corridor data is a
choice based sample for destination choice so that only mode choice could be modelled without
difficulty). Following usual practice (based on McFadden, 1978; see Ben-Akiva et al., 1987)
six alternative destinations were sampled randomly (9), applying uniform sampling probabilities
to each of the zones of the study area (i.e. satisfying McFadden's (1978) "uniform
conditioning property"), and zonal attribute values were assigned for each of these. The
number of jobs in each zone was taken as the attraction variable; see Daly (1982) for a
description of the function and form of this type of variable.

In this case mode and destination choice were modelled using a multinomial logit model with
approximate alternative utility given by:
6
Vi = di + R hk.xki + u.LSIZE [14]
k=1

where d is a modal constant (estimated values are not reported) and the x variables correspond
Draft 12 August 1990
to the attributes 1 to 6 in Table 1; it is important to remark that u must be greater than 0 and
less than or equal to 1 for a consistent model; also, a value of u significantly different from 1
would indicate a significant correlation between the zoning pattern and choice behaviour (Daly,
1982).

Before presenting the results it is necessary to note that the income data used by Ortúzar and
Ivelic (1987b) in their detailed models has varying levels of accuracy (it is exact, i.e. the
number of Chilean $ reported, for the corridor data of 1983 and 1985, but corresponds to the
mean of the income band reported in the case of the ESTRAUS data; furthermore in this latter
case there is no information on the number of hours worked per month by each individual so an
average figure is used to transform monthly per capita family income to Chilean $/min). To
isolate the influence of the effect of the differing accuracy of income measurement it was
decided to estimate separate models incorporating the exact figures for the corridor data; the
subsample for the corridors alone was also used separately.

4.2.2 Results of the Experiment

The results of the three model specifications are shown in Table 3. The first column of the table
shows the model results (Z1) that are obtained from modelling mode and destination choice on
data (variables 1-4) specified exclusively at zonal level. The second and third columns (D1
and D2) illustrate the results of modelling mode choice when data coded to the more detailed
level of accuracy is used. Comparing these last two models, it can be seen that the improvement
obtained from using detailed income data is small but not negligible.

Table 3: MNL Models for the Complete Sample

Coefficients (standard errors)


Variables (options) MNL-Z1 MNL-D1 (1) MNL-D2 (2)
-----------------------------------------------------------------
TIME (all) -0.036 (0.006) -0.084 (0.013) -0.084 (0.013)
WAIT (all but 1,2) -0.303 (0.047) -0.346 (0.052) -0.351 (0.052)
WALK (all) -0.087 (0.011) -0.120 (0.011) -0.121 (0.011)
COST (all) -0.020 (0.004) -0.020 (0.004) -0.022 (0.004)
COMP (1,6) 1.614 (0.310) 1.588 (0.318) 1.591 (0.318)
SEX (2,3,7) -0.804 (0.164) -0.767 (0.166) -0.762 (0.166)
LSIZE (all) 0.326 (0.029) - -
-----------------------------------------------------------------
l(C) (3) -2049.68 -1607.66 -1607.66
l(h) -1913.23 -1452.24 -1449.93
q² (4) 0.067 0.097 0.098
Sample size 1512 1512 1512
-----------------------------------------------------------------

(1) Using mean family income and number of hours worked per month.
(2) Using detailed values for these variables for the corridors.
(3) Likelihood obtained with modal constants only.
(4) q²=1-l(h)/l(C), the "likelihood ratio", measures the overall
goodness of fit of the model.

In terms of the overall fit, measured by q², it may appear that the more-disaggregate mode
choice models are preferable. However, it must be remembered that the choice situation being
"explained" by Z1 is more complicated and, in a real sense, more difficult to explain than the
Draft 12 August 1990
simple mode choice treated by the D models. Thus q² seems to be an inadequate criterion.

Moreover, in terms of the accuracy of estimation of the key para-meters, i.e. the forecasting
accuracy of the model as internally estimated, the model Z1 actually seems to perform better than
the more detailed models. Partly this is due to a general reduction in scale in Z1 relative to D1
and D2, presumably associated with the increase in total error arising from the inclusion of
destination as well as mode components in the utility function. In proportional terms (i.e. as
measured by 't' ratios) the D models perform better.

4.2.3 Discussion and Further Analysis

To investigate further the relative merits of the Z and D models described above, statistics are
presented in Table 4 for the mode choice component of model Z1. That is, a model is specified
that predicts the choice among modes for the chosen destination only using the coefficients
estimated in Z1. These results show clearly that the detailed data in model D2 gives a much
better representation of mode choice in terms of likelihood, or equivalently, the q² measure
(directly applicable here, since the choice situations being modelled are the same). It is worth
mentioning, however, that both models satisfy the v² test of reasonable structural fit to the
data used by Gunn and Bates (1982); that is, it cannot be said on the basis of this test that the
zonal-data model totally fails to represent the choice situation.

Table 4: Statistics for Mode Choice Models

Statistic MNL-Z1 (mode choice) MNL-D2


--------------------------------------------------------------
l(h) -1517.12 -1449.93
q² 0.058 0.098
--------------------------------------------------------------

Another comparison of the Z and D models is given by comparing the relative values of the
coefficients within each model. A convenient means of achieving this is to compare subjective
(implied) values of time, calculated by

SVOT = ht/hc

where ht is the coefficient of a time variable (e.g. TIME, WAIT or WALK) and hc the coefficient of COST. The values obtained are set out in Table 5.
Table 5: Subjective Values of Time as Percentage of Income

Variable MNL-Z1 MNL-D1 MNL-D2


-------------------------------------------------------------

TIME 183.08 414.85 387.56


WAIT 1552.31 1714.85 1618.43
WALK 447.18 596.04 556.68
-------------------------------------------------------------

It can be seen that in general the values arising from the zonal model are more reasonable (i.e. smaller) than those of the detailed models (see Gaudry et al., 1989
for a discussion of this issue). It is important to note that although much higher than conventional wisdom similar time values have been reported for cases as

different as the USA (McFadden et al., 1977) and Brazil (Swait and Ben-Akiva, 1987).
An important difference between the Z and D models is the presence in the former of the LSIZE variable. The results of Table 3 show that the coefficient of this variable

is significantly different from 1, a finding confirmed by a likelihood ratio test (with one degree of freedom) comparing this model with another model in which the
value of the coefficient was constrained to 1. This coefficient value indicates that the destinations within each zone have some common characteristics not measured in

the model - a finding intuitively acceptable to those with knowledge of the zoning system and local conditions. Thus the Z model is able to incorporate a real aspect
of travel behaviour excluded from the D models.
It appears that for the Santiago data described here, which is in general a typical example of transportation study data, that either detailed data can be used to
Draft 12 August 1990
analyse mode choice or zonal data can be used to analyse mode and destination choice. The use of zonal data to analyse mode choice alone gives clearly inferior and
misleading results, although that is not apparent from the models themselves.

Comparing the acceptable models, the balance lies with the zonal mode-destination model, because of its intuitively preferable and more accurately estimated
coefficient values. Nevertheless, because of the important differences in cost in setting up the data between these two models, it cannot be said conclusively that one
is preferable to the other as a general procedure. The best model would clearly be obtained by modelling mode and destination choice together, using fully-detailed
data. However, the theoretical basis of this procedure has not yet been worked out.

The difficulties of application of models of this type must not be forgotten. Forecasting detailed-level data over a wide area is also not yet possible, so that models
estimated at detailed level must pass through an aggregation procedure to avoid bias before they can be used for forecasting. This naturally adds to the cost and

difficulty of the procedure, so that it may be preferable to start with zonal-level data if wide-area forecasting is required.
5. Conclusions and Recommendations

The preceding sections have, in turn, attempted to set the problem of model design in an appropriate context, to describe the theoretical position of measurement
error relative to other sources of model error and to describe an empirical investigation of the magnitude of the errors arising in mode and destination choice

models. In this final section, these three areas are drawn together to obtain conclusions and then to derive recommendations for practice in future studies.
5.1 Conclusions
A substantial number of different sources may lead to error in the models obtained and in forecasting based on those models. Thus although error caused by excessive
aggregation or other measurement problems may be important and should wherever possible be reduced, other sources will ensure that the models remain far from

perfect.
To improve forecasting accuracy, the variables that are most important to measure accurately are those with a large impact on demand and/or those with a large

relative error (quantified in equation [3] above). In model estimation, while some sources of error can be taken into account and their impact reported by
statistical estimation procedures, other error sources (such as model specification error) cannot be treated in this way and the analyst must seek to minimise them

based only on his or her professional judgement. In every case, it seems that modelling resources should be deployed so as to balance the competing
requirements of reducing error from the various sources.
The conclusion of balancing resources also applies to the issue of model complexity. Increasing complexity can reduce specification error, but it inevitably increases
both cost and measurement error. The best point of balance will depend not only on the policy needs but also on measurement technology and cost, i.e. model

design is a function of the economic development of the country being studied.


The more detailed analysis of Section 3 shows that certain common sources of error, most notably "taste variation" and measurement error that depends on the value
of the variable being measured, are particularly unwelcome, since they cause the basic assumptions of most models to be invalidated. These errors can be reduced by
segmentation where that is possible, or by more sophisticated model specifications.

In forecasting, care must be taken that there is no substantial change in the error variance in the model from the circumstances of model estimation: failure to
maintain this constancy will result in "aggregation bias". Avoidance of aggregation bias usually requires either that the model is estimated using the same

measurement procedure that will be used in forecasting or that an explicit aggregation step is undertaken. Unfortunately, the need for one or other of these steps can
often reduce the usefulness of detailed measurement procedures.
The theoretical discussion leads to decisions about model form and structure that depend on the value of certain error variances. These vary between choice contexts,
measurement methods and data sets. It is not possible to derive universally valid conclusions.

To put these discussions in a practical context, results were presented in Section 4 from recent practical studies in Santiago (Chile) which are believed to be
representative of a wide range of transportation studies world-wide. Here it was found that data aggregation was a key determinant in the quality of mode choice
models. However, when the model specification was improved to include the choice of destination as well as that of mode, the use of aggregate data became much
more justifiable.

In considering the issue of model system design, it is also necessary to take into account the relative cost of data collection the current impossibility of developing
detailed measures of variables describing destination attractiveness and the needs of application, which would in any case require an aggregation procedure to be

carried out before using the model in wide-area forecasting. In many cases, this would lead to the modelling of mode and destination choice jointly with zonally
measured variables.

5.2 Recommendations
The issue of model design should be approached from the viewpoint of the most efficient allocation of modelling resources. This means deriving the most accurate
possible results, taking into account the costs of data collection, model estimation, forecasting and presentation. While experience in other contexts can be used to
indicate likely solutions, it must be expected that time and money budgets, policy issues, staff capabilities, existing data availability and other technical resources will

vary between studies, thus leading to different model specifications being appropriate.
Model error arises from many sources and cannot be totally eliminated. Nevertheless error should be minimised, particularly when the type of error in question

invalidates basic assumptions in the model. Errors of this type may be "taste variation", some types of measurement error and various kinds of specification error.
Measurement error is an important component of the overall error in the model and should therefore be reduced wherever possible, given other demands on the

budget. In particular, it seems to be dangerous to model mode choice on zonally aggregated data. The combination of relatively high measurement error and
relatively limited model specification can lead, it is clear, to inaccurate and misleading models.

It appears to be acceptable to model mode choice alone when detailed data is available for each individual. Given that sample sizes for this type of work are
generally fairly small, it may be competitive to undertake this detailed coding. Further, if wide-area forecasts are not needed, so that sample enumeration procedures
can be used for forecasting, problems of aggregation need not arise.
Draft 12 August 1990
When modelling destination choice, it is not currently known how detailed data should be specified for non-chosen alternatives. In this case it seems that zonal-level
variable measurement is inevitable. Fortunately, it appears that mode-destination choice models specified at this level give competitive results (both in Chile and in

European studies).
Further advantages of modelling at zonal level are those of cost (once networks have been created) and that the same calculation procedures can be used both for
generating variables for model estimation and for forecasting. The expensive and complicated step of model aggregation is thus avoided without introducing
"aggregation bias" into the model.

Notes
(1) Ben-Akiva et al. (1978) describe a number of applications studies from the mid-1970's. See also Williams and Ortúzar (1982b) for a review of travel demand

forecasting methods up to 1980.


(2) The first model system of this type seems to have been that reported by Ruiter and Ben-Akiva (1977); these models have recently been updated, as reported by

Kollo and Purvis (1989). Model systems in The Netherlands are reported by Projectbureau IVVS et al. (1976-78), by Cambridge Systematics (1978-82) and by Daly
and Gunn (1986). For Chile, see Fernández et al. (1983), Hall et al. (1987) and ESTRAUS (1989). For France, a model of Paris is described by Moïsi et al.

(1981), of Grenoble by Daly (1985) and of Nantes by CSE and CETE Ouest (1985). International comparisons of France, Brazil and Israel are given by Ben-Akiva et
al. (1984). An applications study in Australia is reported by Gunn, Bradley and Hensher (1990). A comparison of Swedish models is given by Algers et al. (1985), a
Norwegian model is presented by Tretvik (1989) and a survey of Japanese applications is given by Yai (1989). A comparatively simple application in London is
reported by Copley et al. (1988). It is clear that these methods enjoy widespread application.

(3) A example of this type of problem was encountered in the recently-completed Greater Santiago Strategic Transport Model (ESTRAUS, 1989), which is the first
modelling system known to the authors to ensure supply-demand equilibrium in deriving variables for estimation. In this study, it was found that there would have been

severe distortions in the values of travel times and distances obtained from congested networks had full equilibration not been required.
(4) There are a few value-of-time models that incorporate this effect, such as Daly & Zachary (1975); see also Bates et al. (1987) and Daganzo (1979).

(5) "An object can have no value unless it has utility. No one will give anything for an article unless it yield him satisfaction. Doubtless people are sometimes
foolish, and buy things, as children do, to please a moment's fancy; but at least they think at the moment that there is a wish to be satisfied." quoted from F.M.Taussig,
Principles of Economics, 1912, by McFadden (1981).
(6) This assertion obviously depends also on the specific form chosen for the distribution of e. But reasonable 'bell-shaped' distributions will lead to models little

different from each other. In fact, given that the "alternatives" represented in the model generally imply maximisation over a number of sub-alternatives, the most
plausible functional form is the extreme-value distribution often described as "Weibull" which leads to models of exactly (generalised) logit form.
(7) A key specification issue is the form in which explanatory variables should enter the utility function, even if this is linear in the unknown parameters. Three
approaches are commonly used to approach this problem: stated preference experimentation (Lerman and Louviere, 1978); statistical transformations, such as Box-

Cox transformations (Gaudry et al., 1989) or other empirical transformations (Ben-Akiva et al., 1987); and the constructive use of economic theory (Train and
McFadden, 1978; Jara-Díaz and Farah, 1989). The last approach has the advantage of being possible to connect to evaluation measures (Jara-Díaz and Ortúzar,

1989).
(8) See Ben-Akiva and Lerman (1985), Chapter 6, for a summary of the various methods, or Cambridge Systematics (1982) for a practical application.
(9) Actually, as in all other transportation studies (and, indeed, almost all practical statistical work), "pseudo-random" sampling was used. The philosophical
differences of this form of sampling from "true" random sampling and how the latter may be achieved, if at all, are discussed in statistical texts, e.g. Hammersley and

Handscomb (1964).
References
Algers, S., Colliander, J. and Widlert, S. (1985) Logitmodellens användbarhet och generaliserbarhet, report prepared for Statens råd för byggnadsforskning, (in
Swedish).

Alonso, W. (1968) Predicting Best with Imperfect Data. Journal of the American Institute of Planners 34(3), pp. 248-255.
Bates, J. et al. (1987) The Value of Travel Time Savings. Policy Journals, Newbury, Berkshire.

Ben-Akiva, M.E. and Atherton, T.J. (1977) Methodology for Short- Range Travel Demand Predictions: Analysis of Car-Pooling Incentives, Journal ofTransport
Economics and Policy 11(3), pp. 224-261.

Ben-Akiva, M.E., Daly, A.J. and Gunn, H.F. (1987) Destination Choice Models: Design and Appraisal. Proceedings 15th PTRC Summer Annual Meeting,
University of Bath, September 1987, England.
Ben-Akiva, M.E., Gunn, H.F. and Silman, L.A. (1984), Disaggregate Trip Distribution Models, Proceedings Japan Society of Civil Engineers, 347/IV-1.
Ben-Akiva, M.E., Kullman, B.C., Sherman, L. and Daly, A.J. (1978) Aggregate Forecasting with a System of Disaggregate Travel Demand Models. Proceedings 6th

PTRC Summer Annual Meeting, University of Warwick, July 1978,England.


Ben-Akiva, M.E. and Lerman, S.R. Discrete Choice Analysis. MIT Press, Cambridge, Massachusetts, U.S.A..

Cambridge Systematics (1982), Report of the Zuidvleugel Study: Volume 7: Travel Demand Models. Ministerie van Verkeer en Water-staat, The Netherlands.
Cambridge Systematics Europe and Centre d'Etudes Techniques et de l'Equipement de l'Ouest (1984) Etude de la possibilité de transfert de modèles désagrégés de

choix modal, Ministère des Transports et de l'Equipement, France (in French).


Copley, G.,Hyman, G., Meehan, A.H. and Swain, D.A. (1988) Distribution and Mode Choice Modelling in London. Proceedings 16th PTRC Summer Annual

Meeting, University of Bath, September 1988, England.


Daganzo, C.F. (1979) Multinomial Probit: The Theory and its Applications to Demand Forecasting. Academic Press, New York.
Daganzo, C.F. (1980) Optimal Sampling Strategies for Statistical Models with Discrete Dependent Variables. Transportation Science 14(4), pp. 324-345.
Draft 12 August 1990
Daly, A.J. (1982a) Applicability of Disaggregate Behavioural Modelling, Transportation Research 16A, pp. 363-370.
Daly, A.J. (1982b) Estimating Choice Models containing Attraction Variables. Transportation Research 16B(1), pp. 5-15.

Daly, A.J. (1982c) Policy Analysis using Sample Enumeration - an Application to Car Ownership Forecasting. Proceedings 10th PTRC Summer Annual Meeting,
University of Warwick, July 1982, England.
Daly, A.J. (1985) Les Déplacements Domicile-Travail et Domicile-Ecole: Modèles Désagrégés de Choix Modal (in Grenoble). CETUR, French Ministry of Transport
(in French).

Daly, A.J. (1986) Proposal for Zonal Mode and Destination Choice Modelling. Technical Discussion Note, Great Santiago Strategic Transport Study (unpublished).
Daly, A.J. (1987) Estimating "Tree" Logit Models. Transportation Research 21B(4), pp. 251-268.

Daly, A.J. and H.F. Gunn (1986) Een nieuw prognosesysteem voor het personenvervoer in Nederland, Verkeerskunde 37, pp. 512-514 (in Dutch).
Daly, A.J. and Zachary, S. (1975) Commuters' Values of Time, Report T55, Local Government O.R. Unit, Reading, England.

Daly, A.J. and Zachary, S. (1977) The Effects of Free Public Transport on the Journey to Work . Supplementary Report SR 338, Transport and Road Research
Laboratory, Crowthorne, England.

De Cea, J. Bunster, J., Zubieta, L. and Florian, M. (1988) Optimal Strategies and Optimal Routes in Public Transport Assignment Models: an Empirical
Comparison. Traffic Engineering and Control 29(10), pp. 520-526.
ESTRAUS (1989) Estudio Estratégico del Gran Santiago. Informe Final a la Secretaría Ejecutiva de la Comisión de Transporte Urbano, Santiago (in Spanish).
Fernández, J.E., Coeymans, J.E. and Ortúzar, J. de D. (1983) Evaluating extensions to the Santiago underground system. Proceedings 11th PTRC Summer Annual

Meeting, University of Sussex, July 1983, England.


Gaudry, M.J.I., Jara-Díaz, S.R. and Ortúzar, J. de D. (1989) Value of time Sensitivity to Model Specification. Transportation Research 23B.

Gunn, H.F. (1985) Artificial Sample Applications for Spatial Interaction Models. Colloquium Vervoersplanologisch Speurwerk, The Hague, November 1985,
Holland (see particularly the Appendix).

Gunn, H.F. and Bates, J.J. (1982) Statistical Aspects of Travel Demand Modelling. Transportation Research 16A(5/6), 371-382.
Gunn, H.F., Bradley, M.A. and Hensher, D.A. (1990) High Speed Rail Market Projection: Survey Design and Analysis. Presented to 3rd International Conference on
Survey Methods in Transportation, Washington D.C., January, U.S.A..
Gunn, H.F., Bradley, M.A. and Ben-Akiva, M. (1987) Tests of the Scaling Approach to Transferring Disaggregate Travel Demand Models, presented to Annual

Meeting of Transportation Research Board, Washington, D.C., January, U.S.A..


Gunn, H.F., VanderHoorn, A.I.J.M. and Daly, A.J. (1987) Long-Range County-Wide Travel Demand Forecasts from Models of Individual Choice. Proceedings
International Conference on Travel Behaviour, Aix-en-Provence, October, France.
Hall, M.D., Davies, R.F., Daly, A.J. and Russell, C.H. (1987) Modelling for an Expanding City. Proceedings 15th PTRC Summer Annual Meeting, University of

Bath, September 1987, England.


Hammersley, J.M. and Handscomb, D.C. (1964) Monte Carlo Methods, Chapman and Hall, London.

Horowitz, A.J. (1989) Convergence Properties of Some Iterative Traffic Assignment Algorithms, presented to Annual Meeting of Transportation Research Board,
Washington D.C., January, U.S.A..
Horowitz, J. (1981) Sampling, specification and data erors in probabilistic discrete choice models. Appendix C of D.A. Hensher and L.A. Johnson, Applied Discrete
Choice Modelling. Croom Helm, London.

Ivelic, A.M. (1988) Impacto de la Agregación de Atributos en la Estimación de y Estabilidad de Modelos Logit de Partición Modal. Magister in Engineering Thesis,
Departamento de Ingeniería de Transporte, Pontificia Universidad Católica de Chile, Santiago (in Spanish).
Jara-Díaz, S.R. and Farah, M. (1987) Transport Demand and User's Benefits with Fixed Income: the Goods-Leisure Trade-Off Revisited. Transportation Research
21B(2), pp. 165-170.

Jara-Díaz, S.R. and Ortúzar, J. de D. (1989) Introducing the Expenditure Rate in the Estimation of Mode Choice Models. Journal of Transport Economics and Policy
XXIII(3), pp. 293-308.

Kollo, H.P.H. and Purvis, C.L. (1989) Regional Travel Forecasting Model System for the San Francisco Bay Area. Presented to T.R.B. Annual Meeting,
Washington, D.C., January 1989, U.S.A..

Lerman, S.R. and Louviere, J.J. (1978) The USe of Functional Measurement to Identify the Form of Utility Functions in Travel Demand MOdels. Transportation
Research Record 673, pp. 78-85.
McFadden, D.L. (1978) Modelling the Choice of Residential Location. Transportation Research Record 673, pp. 72-77.
McFadden, D.L. (1981) Econometric Models of Probabilistic Choice. In C.F.Manski and D.L.McFadden (eds.), Structural Analysis of Discrete Data with

Econometric Applications, M.I.T. Press, Cambridge, Mass..


McFadden, D., Talvitie, A.P. and Associates (1977) Demand Model Estimation and Validation. Special Report UCB-ITS-SR-77-9, Institute of Transportation

Studies, University of California at Berkeley.


Moïsi, F.,Raison, M. and Silman, L. (1981) A Modal Split Model with Multiple Transit Modes. Proceedings 9th PTRC Summer Annual Meeting, University of

Warwick, July 1981, England.


Ortúzar, J. de D. and Ivelic, A.M. (1986) Efectos de la Desagregación Temporal de Variables de Servicio en la Especificación y Estabilidad de Funciones de

Demanda. Actas del IV Congreso Panamericano de Ingeniería de Tránsito y Transporte, Santiago, December 1986, Chile (in Spanish).
Ortúzar, J. de D. and Ivelic, A.M. (1987a) Effects of using More Accurately Measured Level of Service Variables on the Specification and Stability of Mode Choice
Models. Proceedings 15th PTRC Summer Annual Meeting, University of Bath, September 1987, England.
Draft 12 August 1990
Ortúzar, J. de D. and Ivelic, A.M. (1987b) Impacto de la Agregación de Atributos en la Estimación de Modelos Logit de Partición Modal. Actas del III Congreso
Chileno de Ingeniería de Transporte, Universidad de Concepción, November 1987, Chile (in Spanish).

Ortúzar, J. de D. and Ivelic, A.M. (1988) Influencia del Nivel de Agregación de los Datos en la Estimación de Modelos Logit de Elección Discreta. Actas del V
Congreso Panamericano de Ingeniería de Tránsito y Transporte, Universidad de Puerto Rico en Mayagüez, July 1988, Puerto Rico (in Spanish).
Projectbureau IVVS, Cambridge Systematics Inc., Nederlands Econ-omisch Instituut and Nederlands Vervoerswetenschappelijk Instituut (1976-78). Reports of the
SIGMO Study, Volumes 1-7, Ministerie van Verkeer en Waterstaat, The Netherlands (in English and Dutch).

Reid, F.A. and Small, K.A. (1976) Effects of Temporal Disaggregation of Trip Data on Traveller Behaviour Models. Transportation Research Forum Annual Meeting,
MIT, October 1976, U.S.A.

Ruiter, E.R., and M.Ben-Akiva, A System of Disaggregate Travel Demand Models: Structure, Component Models and Application Procedures , prepared for
Metopolitan Transportation Commission, 1977.

Swait, J. and Ben Akiva, M.E. (1987) Empirical Test of a Constrained Choice Discrete Model: Mode Choice in Sao Paulo, Brazil. Transportation Research 21B(2),
pp. 103-116.

Tardiff, T.J. (1979) Specification Analysis for Quantal Choice Models. Transportation Science 13(3), pp. 179-190.
Train, K.E. and McFadden, D. (1978) The Goods-Leisure Trade-Off and Disaggregate Work Trip Mode Choice Models. Transportation Research 12(5), pp. 349-
353.
Tretvik, T. (1989) Logitmodeller for Transportplanlegging. Thesis for Doctorate in Engineering, Norges Tekniske Ho/gskole, Trondheim, Norway (in Norwegian with

English summary).
Williams, H.C.W.L. and Ortúzar, J. de D. (1982a) Behavioural Travel Theories of Dispersion and the Mis-specification of Travel Demand Models. Transportation

Research 16B(3), pp. 167-219.


Williams, H.C.W.L. and Ortúzar, J. de D. (1982b) Travel Demand and Response Analysis - Some Integrating Themes. Transportation Research 16A(5/6), pp. 345-

362.
Yai, T. (1989) Disaggregate Behavioural Models and their Applications in Japan. Transportation Research 23A(1), pp. 45-51.

You might also like