Some lessons in stated choice survey design

Stephane Hess and John M. Rose

Stephane Hess
University of Leeds
[email protected]

John M. Rose*
The University of Sydney
[email protected]

Abstract
A growing majority of discrete choice studies are now based on data collected through
stated preference (SP) surveys, primarily in the form of stated choice (SC) questionnaires.
The state-of-the-art in this area has evolved dramatically over recent years, as witnessed in
a burgeoning literature. At the same time however, the state-of-practice has stagnated,
especially in some countries. Additionally, the growing emphasis on theoretical
developments, primarily to do with efficiency, has meant that a number of fundamental
issues are often no longer talked about. In the present paper, we look in detail at the entire
process going from initial survey planning to actual data collection, discuss, often with
examples, a number of common but avoidable mistakes, and provide some guidance for
good practice.

1. Introduction
Over recent years, there has been a hype of activity in the field of experimental design
for stated preference (SP) surveys, leading to a move away from orthogonal design
techniques to efficient design techniques. The advantage of these design techniques in a
practical context is that more robust results can be obtained with smaller sample sizes,
potentially leading to significant financial savings, especially with surveys involving face to
face interviews.
Whilst these developments represent theoretical advancements that are gradually
making their way into applied research, the literature as a whole appears to have largely
focused on these advances to the neglect of more fundamental issues. The design and
implementation of surveys for the collection of choice data is an in-depth process that adds
to the already existing complexities of more traditional questionnaire construction and data
collection. It is to these issues that the present paper returns, specifically dealing with the
basic principles of good practice in the field of survey design.
The topics covered in this paper are survey technique, survey context, choice set design,
experimental design, survey testing, and survey administration. For each topic, we discuss

1
© Association for European Transport and contributors 2009
 the basic issues, highlight possible mistakes, often with the help of examples, and provide
some guidance for good practice.

2. Survey technique
The majority of surveys looking at hypothetical scenarios are of the stated choice (SC) type, in
which a respondent is faced with a choice between a finite number of mutually exclusive
alternatives. Figure 1 shows an example of this type of response format for an unlabelled route
choice experiment. It is this type of experiment that the majority of this paper focuses on.

Figure 1: Example of a typical choice response format

Nevertheless, it should be acknowledged that there exist a number of alternative (or in some cases
complimentary approaches) that we will now touch on briefly. Indeed, the choice of SC should not
be an automatic one, once the analyst has settled on SP methods rather than RP methods.

2.1 Rating and ranking

Rather than have respondents choose their single most preferred alternative, some researchers
prefer to have respondents rate each alternative on some form of scale. Typically used rating scales
involve respondents having to rate each alternative from 1 to 10 or 1 to 100 where higher values
represent higher degrees of preference for that alternative. An example of this type of response
method is shown in Figure 2. Some researchers prefer respondents to rank (with or without allowing
for ties) each alternative in order of preference. An example of this is shown in Figure 3.

Figure 2: Example of a typical conjoint ratings response format

Together, ranking and rating type response data combine to form a single SP methodology referred
to in some literature as traditional conjoint analysis or simply conjoint analysis. This differs to choice
type responses (which for historical reasons is referred to as choice based conjoint in the same

2
© Association for European Transport and contributors 2009
 literature) in a number of important ways. Firstly, the analysis used for such data typically relies on
linear regression models as opposed to non-linear logit or probit type models often employed for
choice data. Although ordered discrete choice models may also be used on such data, the literature
dealing with traditional conjoint methods typically ignore such models in favour of linear regression
models. This is because such literature usually seek to derive individual specific models as opposed
to a model estimating the population ‘average’ parameters. This has proven somewhat controversial
given that linear regression models assume interval or ratio scaled data for the dependent variable,
with debate raging as to whether rating or rankings data meet this criterion. Secondly, the response
metric (and in particular ratings scales) has also proven somewhat controversial from a psychological
perspective with many researchers questioning whether different respondents assign the same
psychological value to the values of the scale (i.e., does a rating of 4 on a 1 to 10 point scale have the
same meaning to 2 different respondents). Discrete choices do not suffer from this issue.
Nevertheless, ratings and rankings tasks offer two significant advantages over discrete choice tasks;
firstly they provide full information on the relative preferences of all alternatives unlike choice which
informs the analyst only what is the most preference option, and secondly, ranking of responses may
allow the analyst to rank explode the data which may provide more observations per respondent
from which to model with.

Figure 3: Example of a typical conjoint rankings response format

2.2 Best worst and best worst scaling

Over time, two different forms of best worst response formats have come into existence. The
most recent format, as shown in Figures 4a and b, require that respondents indicate what they
consider to be the best and worst alternatives from amongst those on offer. Whilst providing similar
data to pure rankings data, those advocating this approach suggest that respondents are better
equipped psychologically to distinguish between their most preferred and least preferred
alternatives than they are to rank all alternatives, particularly when presented with large number of
alternatives to consider. Where more than three alternatives are shown in any single choice task,
requiring respondents to suggest only the best and worst alternatives will result in partial preference
rankings of the alternatives (see Figure 4a). If the analyst wishes to obtain a full ranking of the
alternatives, subsequent response tasks involving the remaining alternatives can be used (see Figure
4b). See Louviere et al. 2008 for a more detailed review of this response format.

3
© Association for European Transport and contributors 2009
 Figure 4a: Partial ranking best worst response format

Figure 4b: Full ranking best worst response format

Originally proposed by Finn and Louviere (1992), best worst scaling offers yet another alternative
response mechanism for collecting SP type data. Best worst scaling methods differ significantly to
the other response methods in that the response mechanism is not active at the level of the
alternatives, but rather at the level of the attributes (see Figure 5). Rather than present respondents
with a number of alternatives to choose from amongst, the best worst scaling approach presents
respondents with a single alternative and asks them to select the best and worst attribute for that
alternative based on the attribute levels shown. The (log of the) frequency of times a particular pair
of attributes is selected as the best and worst combination is then used as the dependent variable in
a linear regression model to determine the desirability of each attribute and attribute level for
different respondents.

Figure 5: Best worst scaling response format

Proponents of best worst scaling point to two significant benefits in its use over more traditional SP
response formats. Firstly, they argue that the traditional pick one choice responses, which represent
the predominant data collection method used to date, are largely inefficient in terms of the amount
of data obtained from the respondent. This criticism, whilst warranted, has been partially addressed

4
© Association for European Transport and contributors 2009
 via the other response mechanisms outlined here. The second criticism relates mainly to the ability
of discrete choice type models to untangle the base levels of categorical variables that are dummy
coded from any estimated alternative specific constants (ASCs). Whilst effects or orthogonal coding
overcomes this, there still remains a problem in interpreting the wiliness to pay (WTP) values
obtained for such categorical attributes. Such coding structures allow for a determination of the
WTP values for the non-base levels however these values are estimated relative to the base level,
the WTP value of which is not calculable. For example, consider a categorical attribute ‘comfort’ with
levels, low, medium and high. Assuming this attribute is dummy coded with low as the base level,
then further assuming the experiment contains a cost attribute, the WTP for the for the medium and
high attribute levels can be determined. Unfortunately, these WTP values are relative to the base
low level, the WTP for which is not known. Effects and orthogonal coding unconfound the base
levels from the ASCs; however, interpretation of the WTP outputs remains equally problematic. The
best worst scaling response format allow for the calculation of the WTP for all attribute levels, and
hence is argued as being preferred if an experiment has non numeric attributes (see Marley and
Louviere, 2005).

2.3 Frequency data

The final ‘choice’ based response mechanism that has been applied in the past involves
respondents being assigned some value that they may then parcel out to the various alternatives. In
transport studies, this typically involves respondents being asked to assign a number of trips to
alternative routes as shown in Figure 6. Once collected, frequency data may then be converted to
proportions which can then be estimated using the same statistical methods used for ‘pick one’ type
choice data.

Figure 6: Frequency data

2.4 Transfer price / Contingent Valuation / WTP question

Another tool used especially in the context of applied work consists of so called transfer price,
contingent valuation or willingness to pay questions. In the more blunt willingness to pay approach,
a respondent is directly asked how much he/she would be willing to pay for a saving in travel time by
a certain amount, while, in the slightly more refined transfer pricing approach, the respondent is
asked whether he/she would be willing to pay a certain amount x, which is then gradually increased
or reduced until the boundary willingness to pay (WTP) for that respondent is reached. Transfer
price approaches are potentially affected by significant levels of strategic bias, and have come under
considerable legal scrutiny in some countries, such as Australia1. In fact, it is the risk of strategic bias

1
In a case dealing with the valuation of music, the Copyright Tribunal of Australia in 2007 ruled that transfer
price methods were inappropriate and that discrete choice modelling was the preferred method for valuation
studies (PPCA (the Nightclubs Matter) CT2/2004 [2007] ACopy1).

5
© Association for European Transport and contributors 2009
 in such direct questioning approaches that is one of the motivations behind using multi-attribute
hypothetical choice scenarios, masking at least up to a degree the true aim of the research. Even
though, in some work, the results from the transfer price exercise seem to be largely consistent with
those from the traditional SC work, the authors of the present paper are not convinced as to the
motivation for retaining transfer price as a tool for future studies.

3. Survey context
In some cases, the topic of a study directly determines the context of the SC surveys, such as for
example in the case of mode choice experiments. In some cases however, most notably in the
context of work looking at valuation of travel time (VTT) measures, the analysis is more results than
context driven, and various possible approaches arise, as discussed in the following subsection. Later
on in this section, we also discuss issues with inappropriate and unrealistic contexts.

3.1 Context for valuation of travel time studies

The VTT is the core WTP measure in a transport context, and a large share of SC surveys are
commissioned precisely with the purpose of eliciting such VTT measures. Clearly, VTT measures can
be obtained from surveys with very different contexts, and analysts have variously made use of
route choice experiments, abstract choice experiments, and mode choice experiments.
Independently of the context of a survey, the main emphasis is on using the hypothetical choice
scenarios to study the relative sensitivity of respondents to time and money components. The latter
especially deserves some special attention. While in a public transport context, the cost component
of the journey is relatively easily understood (i.e., journey fare), complications arise for car journeys.
Indeed, the main cost component of a car journey is the running cost, which, although some weight
should also be given to maintenance costs and depreciation, is essentially the fuel cost for the
journey. Many studies do rely extensively on running costs when looking at the cost sensitivity of car
drivers, but it is important to recognise that this is a difficult concept for respondents to
comprehend, not least because fuel bills are not generally paid on a journey by journey basis.
Given the difficulties of relying solely on running costs, there is considerable interest in exploring
the use of other cost components. With the absence of parking costs, which have little or no
relationship to travel time, the main emphasis falls on road tolls. Route choice experiments are
increasingly being framed as toll road experiments. In many ways, toll road studies offer one of the
most realistic settings for studying valuations of travel time reductions, with a higher toll (or indeed
a non-zero toll) applying to more rapid routes. This could be in the form of comparing several tolled
routes, with negative correlation between the level of toll and the travel time (i.e., a more highly
tolled route is faster), or a comparison between slower untolled routes and faster tolled routes.
Two main complications however also arise in toll road studies. The first of these is that of
experience. Indeed, in some countries, toll roads are still relatively rare, and a large share of the
driving population will have little or no experience of toll roads. This not only causes problems with
sampling strategies, which we will return to below, but also potentially means that many
respondents of the survey will have difficulties relating to the toll attribute.
The lack of experience with toll roads arguably also accentuates the second problem with such
studies, namely that of a high aversion by respondents to choose tolled options. Road tolls are a
contentious issue, and surveys including a toll attribute are often affected by a high level of political
voting or lexicographic behaviour, with respondents refusing to choose a tolled option (if untolled
options are available) or always choosing the option with the lowest toll. This is indeed especially the
case in surveys making use of respondents with limited or no exposure to the benefits of toll roads,
and respondents where the current route is untolled. Not surprisingly, results in such studies are
often affected by misunderstanding as well as strategic bias, as recently observed by Chintakayala et
al. (2009a).

6
© Association for European Transport and contributors 2009
 Significant effort has also gone into using mode choice experiments in the study of VTT
measures, notably in Switzerland (see e.g., Axhausen et al., 2008). Such experiments can be useful in
producing VTT measures jointly for different modes, while they arguably also have an advantage in
masking the aim of the work. However, mode choice studies often face major issues with mode
allegiance, with many respondents being unwilling to switch mode even in return for large time
savings. This is then however not necessarily a reflection of a high or low VTTS, but simply of high
modal allegiance.
In an attempt to avoid problems with toll road and mode choice studies, VTT studies in many
countries have made use of abstract choice scenarios, presenting respondents with explicit time
money trade-offs. Not only are there potential issues with unrealistic time cost trade-offs, as
addressed in the next section, but such abstract scenarios bear little resemblance to real world
scenarios. This in itself can pose significant problems.
David Hensher, one of the leading advocates for realism in SP design, has put forward the notion
of “experientially meaningful configurations”, i.e., ensuring that respondents are presented with
choices that would be reflective, at least up to a degree, of real life scenarios so as to ensure an
acceptable degree of realism and response quality. This is arguably not the case in such abstract
choice scenarios, and it is not immediately clear whether getting respondents to make such a leap of
faith in completing a SP scenario can be guaranteed to have no influence on results. Crucially, there
is very little evidence on this issue to date, but abstract scenarios continue to be used quite widely in
an applied context.

3.2 Unrealistic contexts

As already stressed above, it is of crucial importance to present respondents with realistic choice
experiments. In this section, we look in particular at the realism of presented attribute level
combinations before turning our attention to two examples of SC surveys affected by problems with
realism.

3.2.1 Unrealistic attribute level combinations: the time/cost example

SC surveys present respondents with scenarios where alternatives are described by a range of
attributes, and where respondents are expected to make trade-offs between different attributes. To
this extent, surveys will often make use of designs in which attributes that we expect respondents to
trade on being negative correlated with one another. This is clearly not the case with orthogonal
designs, but problems with dominated choices may arise as a result, as discussed later.
The two attributes that we focus on in this discussion are time and cost. With the aim of
encouraging trading between time and money, independently of the survey context, there is a
natural temptation to allow for negative correlation between travel times and costs in the
hypothetical choice sets. As such, a faster journey is more expensive than a slower journey. This
generally makes good sense in a public transport context, but can create problems in a car context
where running costs are included. Indeed, by definition, in real life scenarios, travel time and running
costs are strongly correlated, while, with the above rationale, a SC choice set would tend to give
respondents a choice between a cheaper but slower option and a faster but more expensive option.
This however arguably fails the realism test in that the faster route now leads to greater running
costs, which would not generally be the case in reality. Indeed, the faster route would arguably have
to be significantly longer than the slower route for this to apply, leading to unrealistic speed
assumption, coupled with the fact that fuel consumption per miles on faster roads is generally lower
than on slower roads. If the cost attribute used in the surveys is a toll component, such negative
correlation would be acceptable, with a faster route incurring a higher toll. However, toll road
studies often make use of both toll costs and running costs, and once again the issue of correlations
between travel time and running costs need to be addressed.

7
© Association for European Transport and contributors 2009
 3.2.2 Example 1: Swissmetro
Swissmetro is a hypothetical underground railway system, using maglev technology and
travelling at speeds of over 400 km/h under the whole of Switzerland, with extensions to other
European cities (see Figure 7). The highly ambitious project is arguably not likely to ever be
completed. Nevertheless, a SC survey was conducted, giving respondents a choice between car, rail
and the Swissmetro (cf. Bierlaire et al., 2001). With a headline figure of Zurich to Berne in 12 minutes
(where a conventional train takes 57 minutes), the advantages of the Swissmetro option are
however so big that it should come as no surprise that Swissmetro was chosen in 58 percent of
choice sets.

Figure 7: Swissmetro (figures copyright of Pro Swissmetro)

3.2.3 Example 2: The Sydney bush fire example

A large number of discrete choice projects, both consulting and research oriented, are typically
constrained not only by budget, but by time considerations. This often means that such projects are
rushed, with many aspects of the research design done in such a way as to cut corners. The easiest

8
© Association for European Transport and contributors 2009
 and one of the most frequently left out part of many studies, is the use of qualitative research to
refine the questionnaire design.
Consider a bush fire evacuation choice study conducted in Sydney, Australia in 2003. The project
was designed to examine what factors would result in respondents evacuating their residence given
an approaching bush fire. A preliminary examination of the literature resulted in a traditional grid
like choice task design, where respondents were to be asked to select which bushfire they would be
most likely evacuate from given two possible fires. An example of the proposed choice task is shown
in Figure 8.

Figure 8: Initial Choice task based on literature review

Fortunately, qualitative research was conducted whereby focus group participants were shown the
above choice task and asked whether it made sense to them and whether they could answer the
question accurately. The focus group participants were unable to understand the task, arguing that
in reality, individuals were unlikely to be faced with having to choose between which of two
different bushfires they would evacuate from. Furthermore, asking questions as to the likely timing
of evacuation, as was proposed, was not realistic given that such decisions are based on quickly
changing circumstances. Finally, focus group participants indicated that the decision to evacuate was
not as simple as choosing to evacuate or not, with many indicating that they may only evacuate
some of the household members, whilst others would remain behind. Given the above as well as
discussions related to the specific attributes and the levels that they could assume, the final version
of the survey used a completely different choice context, as well as relying on graphics and videos
for presentation. An example of the final survey task is shown in Figure 9.

9
© Association for European Transport and contributors 2009
 Figure 9: Final bushfire choice task based on qualitative research

4. Choice set design

The next step after choosing a survey context is to make decisions on the number of alternatives,
attributes, and attribute levels and values. We will look at these different topics in turn, before also
looking at the topic of referencing.

4.1
Alternatives and attributes
A choice situation in a SP survey presents a respondent with a fixed number of mutually
exclusive alternatives, each described by a number of attributes. In generating a design for a survey,
the analyst first needs to decide on the number of alternatives and attributes.
Many designs used in applied work still rely on binary choice experiments, i.e., involving only two
alternatives in each choice situation2. Here, there is a major gap between theory and practice, with a
large share of applied work relying on simplistic binary choice sets, while work of a more academic
nature regularly presents respondents with choices involving three or more alternatives, sometimes
up to five or six. The main argument in favour of using binary designs has been that of a reduction in
respondent burden. However, work has not only shown that respondents can adequately deal with a
larger number of alternatives, but that unnecessarily restricting the number of alternatives may in
fact make the surveys too simplistic and transparent, while also bearing little resemblance to real life
scenarios (see e.g., Caussade et al., 2005, who recommend four as the optimal number of
alternatives). Additionally, a case can be made for increasing the number of alternatives on the
grounds that this allows for greater variability in each choice set, increasing data richness while also
reducing the overall sample size requirements. Finally, as discussed in the vehicle choice example

2
In some applied work, the reliance on paper based surveys plays at least a partial role in the use of binary
experiments.

10
© Association for European Transport and contributors 2009
 below, it is not just the number of alternatives but also the type of alternatives that is of great
importance.
Many surveys make use of only the most relevant attributes, typically time and cost. Such
simplistic choice scenarios clearly avoid any risk of overburdening respondents, and this, in
conjunction with the use of paper based surveys, was the main motivation for such an approach.
However, simplistic scenarios can also be criticised, primarily on the grounds that they bear little
resemblance to the more complex real life choices undertaken by travellers.
The incorporation of other attributes into the choice situations, such as departure time,
reliability, or different travel time components, may be advantageous for three reasons. Firstly, they
lead to a higher degree of realism, potentially improving response quality. Secondly, they mask the
aim of the study, arguably reducing the risk of political voting. Finally, they obviously allow for the
study of valuations in a broader context. As mentioned above, the main argument against increasing
the complexity of stated choice scenarios is that of respondent burden. However, it has now been
shown conclusively that not only are respondents able to cope with relatively complex scenarios (see
e.g., Caussade et al., 2005; Chintakayala et al., 2009a), but that making choice sets relevant by
including all important information may in fact improve response quality (see e.g., Hensher, 2006).

4.1.1. Example 1: A vehicle choice example

A study currently being conducted in Sydney Australia dealing with Automobile choice involves
respondents first having to provide information as to their most recently purchased new vehicle, the
levels of which were then used as an alternative in the subsequent choice tasks they were asked to
undertake. Aside from the most recent purchase, respondents were also shown three other
hypothetical vehicles to choice from. The experimental design strategy involved selecting randomly
selected vehicles of different sizes to make up each choice scenario. An example choice screen is
shown in Figure 10.

Figure 10: Automobile choice example

11
© Association for European Transport and contributors 2009
 In selecting randomly sized vehicles to construct each choice task, inevitably, many respondents
who took part in a pilot of the survey instrument were confronted with situations of having to select
vehicles that were priced twice as much as their most recent purchase. The result of this was that
the current vehicle was selected almost 90% of the time as the most preferred vehicle. For the main
field phase, the experimental design was changed so that at least one hypothetical vehicle would be
the same size as the current vehicle, one would be either smaller, the same size or one size larger,
and the other would final third alternative would be randomly selected from any size model. This
resolved the issue, with respondents then trading off between the available alternatives.
Around the same time as the Australian study, a study on vehicle type and fuel type choice was
also carried out in California, and similar problems were noted where respondents were being
presented with the choice between say a small current vehicle, and a very large alternative vehicle
(an example choice task is shown in Figure 11). Similarly, respondents were initially presented with
choices between often very different fuel types. As a result, a weighting approach was used,
ensuring that the more relevant options had a higher probability of being included while however
still guaranteeing that all possible combinations had a non-zero probability.

Figure 11: 2nd automobile choice example

12
© Association for European Transport and contributors 2009
 4.2. Presentation of difficult attributes: the case of variability
While attributes such as travel time or cost are generally easily understood by most respondents,
complications may arise with more abstract attributes, such as for example comfort. In this section,
we focus on one such attribute, namely travel time variability. While some studies are purely
dedicated to the study of the WTP for improvements in reliability, and can hence make use of e.g.
graphical representation, the majority of studies simply want to include travel time variability as one
additional attribute. This however raises the important question of how to present it. An example of
where the attribute might possibly have been better chosen or represented is shown in Figure 13.
Figure 13 presents an example choice task from a toll road study conducted in Sydney Australia in
2004. In the study, travel time reliability was considered to be an important attribute influencing
route choice. The attribute was shown as a ± value around the current travel time, a representation
which proved problematic for two reasons. Firstly, the experiment dealt with the travel times and
costs for a specific trip, whereas travel time variability presented in this way represents an
accumulation over many trips. Secondly, the presentation of the attribute as both a plus and a minus
can be somewhat confusing to respondents as well as the analyst, as it is not certain whether they
are reacting to the possibility of arriving earlier or later to the intended arrival time. In this way the
attribute might be considered somewhat ambiguous, which may certainly explain why it often
produces random parameter estimates with zero mean, but significant standard deviations (see e.g.,
Hess and Rose 2009a or Hensher et al. 2006).

Figure 13: Example of possible poor attribute representation (travel time reliability)

Figure 14 represents a more recent stylisation of travel time reliability from an experiment
conducted in Brisbane Australia in 2008. In it, the travel time reliability attribute is presented as
probabilities or more accurately percentages (as respondents understood the concept of
percentages much better than probabilities) of arriving earlier, on-time or later than expected.
Qualitative research and pilot studies showed that this representation of the attribute was much
more realistic for respondents and far less ambiguous as to its meaning (Li et al. 2009).

13
© Association for European Transport and contributors 2009
 Figure 14: Example of a potentially better attribute representation (travel time reliability)

4.3 Realism and availability of alternatives

Realism in SC experiments arises from the fact that respondents are asked to undertake similar
actions as they would in real markets (i.e., respondents are asked to make ‘choices’ just as they do in
real markets, yet another reason why ‘pick one’ choice response formats have become the
predominant data collection mechanism in SP studies). However, for any individual respondent,
realism may be lost if the alternatives, attributes and/or attribute levels used to describe the
alternatives do not realistically portray that respondent’s experiences or, in terms of ‘new’ or
‘innovative’ alternatives, are deemed not to be credible. As discussed above, concerns related to the
attributes and attribute levels used within a SC experiment may be alleviated with significant prior
preparation on behalf of the analyst (Hensher et al., 2005). Additionally, for quantitative variables,
pivoting the attribute levels of the SC task from a respondent’s current or recent experience is likely
to produce attribute levels within the experiment that are consistent with those experiences, and
hence, produce a more credible or realistic survey task for the respondent (see for example, Rose et
al., 2008).
Nevertheless, a significant proportion of products and services offered in real markets exhibit
uneven degrees of distribution coverage (Lazari and Anderson, 1994). Such unevenness in availability
may be either geographical, temporal or both. For example, train services may not be available to
certain suburbs due to a lack of existing infrastructure or a train strike on a specific day might
temporarily remove the train alternative from an individual’s choice set. Such constraints on
availability are likely to be population wide (or at least impact upon a large proportion of the
population), and as such, have an even impact over the entire study population. In SC experiments,
such impacts may easily be handled through the removal of the affected alternatives from choice
sets shown to respondents (removal may be from all choice sets within the experiment or subsets of
choice sets to test availability effects in the presence or absence of an alternative). Other factors,
however, may result in the non-availability of an alternative at the individual level. For example, for
a specific journey, an individual may not have access to a car because a partner is using the vehicle,
or alternatively, the household may not be able to afford a car in the first place.

14
© Association for European Transport and contributors 2009
 Rose and Hensher (2006) discuss the generation of experiments that adapt to the reported
availability to respondents of different alternatives. The purpose behind such experiments is to
construct SC experiments in which the alternatives present within the choice tasks are respondent
specific, and hence, reflect individual differences in the choice contexts that are likely to exist within
real markets. In this way, alternatives that would never be available to specific individual
respondents are not shown to them, and hence the preference structure that they reveal in
undertaking the survey is much more likely to mirror that which they would exhibit in real markets.
Figures 12a and b show one such adaptive survey where the alternatives shown to respondents was
determined by whether the respondent had access to a car or not for a recent surveyed trip, and the
origin and destination of that trip.

4.4 Referencing and pivoting: pros and cons

With a strong interest in making surveys more relevant to respondents, choices are often framed
around specific current levels. One approach in this context is to include a reference trip as one of
the alternatives in the survey, typically alongside two further hypothetical alternatives. This is for
example the standard approach in many Australian studies. While this has the advantage of putting
a reference trip in front of the respondent, it also potentially leads to large levels of inertia (in the
form of non-trading) and special care is required during the design and the analysis. This approach
has also been shown to be of great use in analysing the differences between gains and losses (see
e.g., Hess et al., 2008), where the question however arises as to how the SP presentation may in fact
influence the results in terms of gains and losses, and whether this is an area where the
incorporation of RP work may be desirable.
Independently of the nature of the design, another issue in this context is the actual definition of
the reference point. Here, the question needs to be asked if the current trip is actually the most
natural reference point for an individual when it could equally well be the ideal trip. This issue needs
to be kept in mind at the design stage (to enable adequate pivoting) but is also of crucial importance
at the modelling end, as discussed later on in this document.

Figure 12a: Example mode choice experiment with 6 alternatives

15
© Association for European Transport and contributors 2009
 Figure 12b: Example mode choice experiment with 4 alternatives

A further issue with reference point designs which is only now starting to become apparent is
that such designs appear to induce a significant proportion of respondents to exhibit inertia or non
trading behaviour. Whilst such effects may indeed be the norm in many real markets (and hence
suggest that the SP task is somewhat realistic or at the very least, produces realistic behaviour),
inertia or non-trading does cause model estimation problems. If respondents always choose an
alternative irrespective of the attribute levels of that and other alternatives, then unrealistic
parameter estimates may result (e.g., the reference alternative may have a higher travel time, which
if always chosen may produce a positive travel time parameter when modelled). Furthermore,
respondent non-trading between alternatives may not provide any information as to the trade-offs
that respondents are willing to make between the various attributes. If that is the case, then there
might exist a possible trade-off for researchers between making choice experiments more realistic
and making choice experiments that force trade-offs which may be useful in modelling respondents
preferences. This is one of the arguments used by proponents of abstract experiments.

4.5 Attribute levels

Another important decision relates to the number of levels used for each attribute in the design,
and the actual values for these levels. Here, the main emphasis is generally on using a set of levels
that is broad enough to allow for a diverse set of possible combinations and trade-off values while
also not being so wide as to lead to unrealistic combinations.
Relatively little effort is however generally invested in the number of levels and the number of
times each level is used in the survey. Raising the number of levels increases the number of possible
combinations, improving the richness of the data, up to a point where the effects become
detrimental (see e.g., Chintakayala et al., 2009aAdditionally, the more levels shown in an
experiment, the more able the analyst is to detect non-linear marginal utility functions. However,
from a statistical efficiency perspective, the more levels used, the less efficient the design will
probably be. This is because the statistical efficiency of the design (which relates to the t-ratios one

16
© Association for European Transport and contributors 2009
 will likely obtain from using that design) is a function of the choice probabilities. Contrary to what
many might believe, the more attribute levels that are used, the more constrained the design will be
in terms of the possible choice probabilities that it can achieve. For this reason, end point designs
(designs with two levels at the extremes of the attribute levels) have often been found to produce
the most statistically efficient results. However, this is also related to the range used for the
attribute levels, as too wide a range may result in completely dominated alternatives. As such, there
are several trade-offs that need to be considered in selecting what and how many attribute levels to
use in a SC study.

4.6 Counteracting effects of inertia: moving away from single choices

As alluded to in Section 4.1.2, a significant issue with many experimental designs is inertia or
non-trading between the alternatives. Whilst recent evidence is being gathered that suggests that
this effect is a particular problem for experiments based on individual specific reference alternatives,
similar effects may also be found in many experiments that involve status quo or no choice
alternatives (for a discussion of this, see Hess and Rose, 2009b and Rose and Hess, 2009). A number
of potential solutions to this issue exist, many of which we have already alluded to. The use of best
worst, or rankings type response mechanisms, rather than the typical ‘pick one’ choice responses
that are typically used, allow the analyst to capture information on additional preferences.
Even when ‘pick one’ type choices are used, the analyst may allow for partial rankings within the
‘pick one’ alternative approach. This solution involves the use of dual responses in SC experiments
where respondents are first asked to select from amongst all non status quo alternatives (a forced
choice) after which they are asked to make a second choice in which the status quo alternative is
added (a non-forced choice). The use of dual responses in SC experiments, whilst potentially
improving the statistical efficiency of estimated models as well as providing further information that
can be used to refine the parameter estimates, may however lead to other potential modelling
problems, in particular violations of the IID assumption if the data from the two choices are pooled
into a single data set. IID violations may occur if the error variances between the two choice tasks
are different. Brazell et al. (2006) acknowledge this potential problem, and found that in simulated
data as well as in two empirical data sets, no such violations occur. Nevertheless, Dhar and Simpson
(2003) who also explore issues related to the use of dual responses within SC choice tasks, did find
limited evidence of such violations occurring. An alternative, though corresponding approach, is to
first give respondents a choice from the full set of alternatives, and to force a choice between the
purely hypothetical alternatives if the reference alternative is chosen in the first task.
This dual choice approach proved particularly useful in a recent toll road study conducted in the
UK, where respondents have limited exposure and hence experience with such roads. In that data
set, the reference alternative was selected in the majority of cases (75%), with resulting models
being affected by major problems with retrieving significant estimates. Only once the second tier of
choices was analysed where the reference alternative was not available could reliable models be
estimated (see Chintakayala et al., 2009b). Of course, the discussions in this paper would point
towards a need for adequate pilot work in such a context.

5. Experimental design issues

One of the most critical components to any SP experiment is the underlying experimental design.
Unfortunately, experimental design theory remains one of the least understood aspects of SP
studies. To highlight this point, Bliemer et al. (2009) undertook a literature review of four top tier
transport journals over the past decade and found that out of 61 SP studies in which the
experimental design type and dimensions could be determined, 40 (66 percent) utilized an
orthogonal design, 12 a D-efficient designs (20 percent), seven (11 percent) randomly assigned
attribute levels shown to respondents and three (3 percent) used an adaptive design approach,
alternating the levels shown to respondents based on the respondent’s previous answers. Whilst

17
© Association for European Transport and contributors 2009
 such a range of different design types would be understandable if the objectives of these papers
were to explore the impact of different design methodologies, not a single one of these studies was
specifically addressing experimental design issues. As such, despite decades of experience with SP
studies, the disparity of design types employed suggests that the practical implications of using one
design type over another is yet to be recognised within the literature.

5. 1 What do we want from a good design?

As we have seen, one of the key aspects of SP design is the construction of realistic choice tasks.
However, as we have also argued, realism may result in some undesirable choice behaviour, such as
inertia or non-trading. Independent of how realistic the choice experiment is designed to be, there
does exist within the literature a number of properties that are considered to be desirable in terms
of the underlying experimental design used. Unfortunately, many of these properties may be
considered to be myths that have developed over the years or be based on limited or untested
experiences. For example, for many years, it was accepted wisdom that respondents could only
complete three or four choice tasks at most with three or four attribute levels, thinking that many
researchers today still adhere to. In their review of the literature, Bliemer et al. (2009) found choice
experiments capturing between 1 and 25 choice tasks per respondents, with no evidence that
respondents could not complete larger numbers of choices (although research does suggest more
error variance occurs with larger numbers of choices captured per respondent; see Caussade et al.,
2005). We now discuss properties considered desirable for experimental designs specifically for
capturing discrete choice type responses.

5.1.1 Is attribute level balance a desirable property of designs?

Attribute level balance occurs when each level of an attribute occurs an equal number of times
over the course of the experiment. The argument for such a constraint is that it will minimise
behavioural bias in so far as respondents will not be exposed to situations with more or less ‘better’
or ‘worse’ attribute levels. For example, if low cost attribute levels are shown more than higher cost
attribute levels, respondents may react much more strongly to the higher levels when viewed than
would otherwise be the case if they saw high and low prices equally over the experiment. Evidence
for this was found by Wittink et al. (1982, 1989, 1992) where experimental manipulations of the
number of attribute levels and degree of balance resulted in systematic differences in estimated
attribute sensitivity rankings. As such, evidence exists as to the behavioural impact that attribute
level balance has upon derived model outputs.
Unfortunately, from a purely experimental design perspective, attribute level balance potentially
produces some undesirable outcomes. Firstly, attribute level balance may result in larger than
necessary experimental designs. This is particularly the case where odd and even attribute levels are
combined in an experiment. For example, it might be possible to generate a design with 8 choice
tasks if each attribute has either 2 or 4 attribute levels, but if one or more of the attributes has 3
levels, then the smallest design would require 12 choice tasks. Secondly, attribute level balance
represents a constraint in generating a design with any constraint limiting the statistical efficiency of
the design (see Kanninen 2002). This is not to suggest that statistical efficiency should be considered
to be the most important aspect in generating the design, but simply to note that attribute level
balance will likely produce less efficient designs. Finally, in the case of orthogonal designs, it might
not be possible to find a design with the desired number of choice tasks that exhibits zero
correlations between each of the attributes. This too may result in larger than necessary designs
being generated.

5.1.2 Are larger designs better?

A common assumption in SP studies is that more choice tasks (in the design, not necessarily for
each respondent) are required to produce greater levels of data variability which will aid in model
estimation. Simply put, the assumption is that smaller designs will not provide enough coverage of

18
© Association for European Transport and contributors 2009
 preference or utility space and hence generating designs with more choice tasks will produce better
model outcomes. Bliemer et al. (2009) compared the results for the same choice problem using
designs created with either 18 or 108 choice tasks. In that study, it was found that the an efficient
design with specifically chosen alternatives outperformed an orthogonal design with 108 choice
tasks in terms of producing much smaller standard errors. This finding suggests that using more
choice tasks is not necessarily better. What is important is how much information each choice task
provides in terms of the trade-offs respondents are required to make. This also means that analysts
should strive to produce designs that do not contain choice tasks that provide no additional
information (e.g. dominated choices).

5.1.3 Are simpler designs better?

As we have seen, many researchers ascribe to the theory that simpler and smaller (per
respondent) designs are better. Whilst this may or may not be the case in terms of the behavioural
impact that more complex designs may produce, it does often result in one undesirable outcome.
Where experimental designs considerations are allowed to dominate the SP study, the experiments
may become less realistic. Consider the experiments shown in Figure 12a and 12b. In that particular
study, 453 respondents completed the choice experiment without being offered any form of
incentive. This is partly explained by the fact that the sampled area is bereft of public transport
options, and respondents when asked about completing the survey stated that they saw completing
the survey as a means of expressing their frustrations in this regard to the government despite the
survey taking between 45 minutes and an hour to complete. This suggests that the involvement of
respondents may be just as important as the complexity of the task they are asked to complete
when completing such surveys. As further evidence that using less complex experiments does not
necessarily mean better modelling outcomes, contrast the experiments shown in Figures 13 and 14.
The experiment shown in Figure 14 is far more complex in terms of the number of attributes shown
to each respondent; however, the ambiguous nature of the travel time variability attribute used in
the study represented in Figure 13 produces worse model outcomes. Indeed, the results produced
from the second study, not only resulted in more intuitive WTP outputs, but also new insights into
the VTT itself and how it might be confounded with the concept of reliability (see Li et al. 2009).

5.1.4 Is orthogonality important in choice experiments?

As suggested by Bliemer et al. (2009), the vast majority of experimental designs used within the
literature are orthogonal in nature. This suggest that either researchers view the correlation
structure of the design as being important, or that they are simply following processes and
procedures used in the past blindly. This is not to suggest that reducing correlation is not important.
Indeed, for several reasons, minimising correlations in the data may represent an ideal.
Nevertheless, research into the design of experiments specifically for the use of SC studies has
shown that the econometric models usually associated with such data do not require that the data
be orthogonal. Indeed, the non-linear nature and the fact that these types of models are estimated
as differences in utility, and hence differences in data, suggests that the correlation structure of the
data is not what is important. This has resulted in the construction of so called efficient designs
which generally trades orthogonality with the aim of reducing the standard errors of the estimated
parameters. As such, efficient designs look not at the correlation structure, but rather at the
expected asymptotic covariance matrix that will result from use of the design.
If one takes the view that the correlation structure of the data is not what is of primary
importance for the econometric models that are estimated on SC data, provided that said data is not
perfectly correlated, then questions arise as to whether the typical nomenclature commonly
associated with experimental designs generated specifically with linear models, such as main effects
only or main effects plus interaction effects, have any real meaning for SC experiments. Indeed, Rose
and Bliemer (in press) demonstrate that despite the persistent use of such language in the SC
literature, such designations do not actually match the reality of the models estimated. This is

19
© Association for European Transport and contributors 2009
 because so called main effect and/or interaction effect designs are generated to produce the
smallest possible standard errors for the parameter estimates as well as reduce to zero the
parameter covariances (i.e., they are designed to produce independent parameter estimates of each
effect). As Rose and Bliemer (in press) show, unlike linear models, once the parameters of a discrete
choice model are no longer zero, then the parameter variances and covariances are no longer zero,
with the values of the covariance matrix becoming larger for orthogonal designs as the parameters
move further away from zero. In effect, this suggests that if an experiment has the desired outcome,
that being to estimate non-zero parameter estimates (very few researchers upon suspecting an
attribute will not play an important role in terms of observed choice behaviour will include that
attribute in the experiment), then the more orthogonal the design, the worse the standard errors
will be. As such, generating a design that have zero correlations for the main effects and/or selected
interaction effects does not necessarily mean that such a design will deliver independent parameter
estimates when discrete choice models are estimated upon data collected using the design.
Notwithstanding recent developments, which we return to below, the majority of SP
questionnaires are still based on orthogonal designs. In an orthogonal design, the different columns
in the design are uncorrelated. However, the use of orthogonal designs also poses a number of
complications, primarily to do with dominance. Especially in simplistic designs, a potentially large
number of choice sets will include dominated alternatives. Many studies largely ignore this issue,
and retain such choice situations in the design, not realising that presenting respondents with such
no brainer choices not only adds nothing to our understanding of the choice processes but
potentially also has detrimental effects on response quality. Other studies take a more aggressive
approach, simply removing these problematic choice situations. A problem with this approach is that
it often leads to a loss of orthogonality, and almost invariably also leads to a loss of attribute level
balance. Whilst not always perfect, efficient design techniques will mostly be able to be adapted to
incorporate constraints and be set up to avoiding dominance.
Finally, it should be noted that manual designs, such as the so called Bradley design (see e.g.
AHCG, 1996), and the design proposed by Hess & Adler (2009), also move away from orthogonality
with a view to avoiding dominated choice scenarios and encouraging trading between relevant
attributes.

5.1.5 What are the effects of blocking?

Often the total number of choice tasks generated for a design will exceed the number of choice
tasks the analyst is willing to givn any one respondent. When this occurs, the analyst needs to decide
how to allocate subsets of the design to different respondents. In their review of the literature,
Bliemer et al. (2009) also examined how such allocations were apportioned within SC surveys. In
their review, they found that the majority of studies reported using a blocking column to allocate
choice tasks to respondents (39, or 64 percent of designs), however, a number of studies where
found to randomly assign choice tasks to respondents (8, or 13 percent). Three (5 percent) studies
were found to use provide the full factorial to each respondent whilst it was not possible to
determine how the choice tasks were allocated to respondents in 11 (18 percent) of the studies
reviewed. The authors are also aware of studies in which the rows in the design are allocated
sequentially. The aim behind including the block as an additional design column is that it avoids
correlation between the blocks and the remaining attributes. In other words, we avoid a situation in
which one block (and hence one group of respondents) is allocated say all the high cost scenarios.
Hess et al. (2008) compared the results of three different experimental designs including an
orthogonal design with randomized choice set assignment, an orthogonal design with an orthogonal
blocking column and an efficient design. In that study, they found that the efficient design
performed only marginally better than the orthogonal design with blocking, but that the design with
random assignment of choice tasks to respondents performed significantly worse than both the
efficient design and the orthogonal design with blocking. As such, they concluded that the blocking

20
© Association for European Transport and contributors 2009
 of the experiment was far more important than the underlying experimental design. In a similar vain,
Bliemer et al. (2009) also empirically examined the impact of blocking, examining the effects of
maintaining equal representation of blocks within a data set versus allowing an uneven sampling
across each block. In that study, they found that statistical differences occur in terms of the standard
errors of the parameter estimates that were found depending on the sampling over blocks that
occur. As such, it is recommended that formal blocking columns be used in SC surveys and that
random assignment of choice tasks be avoided where possible. Sequential blocking is of course
worse still.
Finally, it should be noted that the use of inappropriate blocking approaches will jeopardise the
characteristics of the data. Indeed, analysts sometimes forget that what matters are the qualities of
the data, not of the base design. Even if the underlying design is perfectly orthogonal, using random
blocking will mean that the final data is not, especially with small sample sizes.

6. Survey testing
6.1 Inclusion of consistency checks
It is becoming more popular of late to include so called no brainer choices in surveys, generally in
the form of dominated choices, and to eliminate any respondents failing these tests. While we
recognise the importance of such tests, especially in the case of a departure from current methods,
we feel it is important to mention that such tests have to date often been performed in a potentially
inappropriate manner. As an example, the recent Danish VTT study included a dominated choice as
the sixth (out of nine) choice task. The problem with this approach is that, very much in the same
way as retaining dominated choices in standard orthogonal designs, the presence of this choice
scenario may lead to respondents not taking the remainder of the survey seriously. Work by Hess et
al. (2009) shows some evidence of different behaviour before and after this choice scenario. For this
reason, it is our recommendation that if such tests are to be included, this should be done at the end
of the survey, that way avoiding any biasing influence on the remainder of the data.

6.2 Using simulation to test designs

Aside from what type of experimental design to use (i.e., orthogonal, efficient, etc.), many
researchers fail to test how the design might perform in practice. One such example where such
testing would have assisted in practice occurred as part of a research project conducted in South
America in 2002 (reported in Efron et al., 2003). In that project, an orthogonal design was
constructed, with orthogonal blocking columns. Each block was replicated an equal number of times
over the sample so that orthogonality could be maintained through to the data set.
Once data was collected, the data was reformatted so that qualitative variables could be
estimated using dummy codes. Unfortunately, despite each block being represented equally in the
data, the correlation structure of the dummy codes was such that for some variables they were near
perfectly correlated, meaning that these effects for the dummy codes could not be estimated. This
was despite the design being orthogonal. Had the researchers tested (possibly via simulation) the
design’s performance in terms of how it was to be used in practice, this issue may have been
identified earlier and the problem rectified prior to data being collected.
Nevertheless, simulations may not necessarily detect all possible issues that may arise once the
experiment has gone to field. Indeed, a point that is often not recognised is that while simulation
can show whether a design is able to capture specific effects, the impact of the design on actual
behaviour cannot be tested in this manner. Indeed, most simulations assume that respondents act in
a logit like fashion, when in reality, respondents often implement various behavioural rules of
differing complexity (such as employing lexicographic choice rules). Unless such behavioural rules
are incorporated in the simulation process, then the simulation is unlikely to mimic the real

21
© Association for European Transport and contributors 2009
 outcomes associated with discrete choice data. For example, Hess et al., 2009 produced results that
suggest that respondents lacked the ability to adequately distinguish between the alternatives in
presented in the Danish VTT survey, where simulations had not revealed any problems.

6.3 Pilot and skirmish surveys

Independently of the decisions made in relation to survey design, significant pre-testing of the
survey instrument is highly recommended, involving comparisons of various approaches prior to the
main survey. Given the above discussion in relation to the shortcomings of simulation, such pre-
testing and piloting should include actual data collection. Unfortunately, in academic settings
involving transportation contexts, little evidence exists that such testing takes place (alternatively,
one could equally argue that little evidence also exists that such testing does not take place). Unlike
other discipline areas such as the environmental economics literature where pretesting and survey
design can take 12 months or more, little credence is placed on discussing or even commenting upon
whether and to what extent any pretesting and piloting has occurred within the transportation
arena. Indeed, whilst it is hard to judge the extent of any pre-testing that occurs given what little
comment is devoted to such an issue in published journal articles, it can only be assumed that in
academia, as with consulting, very little if any pretesting and piloting occurs due to a lack of funding
or available time. Nevertheless, when conducted, piloting and pretesting can provide invaluable
insights (see Section 3.2.3).
In addition to the previously discussed Sydney bushfire survey, consider the process used to
derive the choice survey shown in Figure 13. Prior to going to field, the survey instrument was first
created in Microsoft Excel. In setting up the Excel version of the survey, the survey mock up was
made to look as close to possible to the final survey form. Further, using the functionality available in
Microsoft Excel, the survey was set up so that it could actually be used to collect data (shown in
Figure 15; see Black et al. (2005) for a description of Microsoft Excel and its functionality in terms of
being able to collect data). Once complete, the Excel version was sent to all stakeholders for
comments allowing for changes to the survey instrument prior to it being formally programmed.
Next, after the survey questions and layout were agreed upon, the survey was pre-piloted using a
small number of respondents. This allowed not only for the survey instrument to be tested in terms
of respondents answering each question, but also for a small scale test of the logistical requirements
associated with administering the survey in field. This pre-test was followed by further discussions
with additional changes to the survey then made (e.g., the inclusion of the travel time variability
attribute shown in Figure 13). A more traditional pilot survey using the fully programmed computer
aided programmed interview (CAPI) instrument was then held, further testing the logistical
requirements of the survey, before the project was finally used in the main field phase.

22
© Association for European Transport and contributors 2009
 Figure 15: Example of Pre-test CAPI screen in Excel

7. Survey administration
7.1 What other information should be collected?
SC surveys invariably collect some form of additional information on top of the data relating to
the respondents’ preferences. This includes socio-demographic and attitudinal data, as well as data
relating to the questions put to respondents to get them to explain the process that led to their
choices, e.g. information processing strategies. Analysts should always attempt to collect whatever
additional data they feel may be useful at the modelling stage, and should use as a warning the large
number of studies that fail, for one reason or another, to collect a vital piece of information and
then need to rely on arbitrary processes such as imputation. However, analysts should also be
mindful of the fact that some of these additional questions can be of a personal nature, such as for
example to do with attitudes and attributes such as income. For this reason, it is crucial that such
data be collected after the actual choice experiments, so as to avoid influencing the actual
behaviour. Analysts also need to make careful trade-offs between asking for enough information
and not to overdo it and risk non-response, for example when using too high a level of
disaggregation for income.

7.2 Sampling size requirements

Once the survey instrument is ready to go to field, important decisions need to be taken in
relation to the overall sample size as well as the number of segment quotas, if any, that will be used.
In terms of the overall sample size, many studies employ rather arbitrary approaches to calculate the
sample size, e.g., using some arbitrarily chosen round number of respondents (such as 300) for each
quota group. Typically, budget considerations underlie such strategies. Ad hoc formulae, constructed
with little scientific or theoretical backing, also exist to determine the minimum number of
respondents required for estimating model parameters (see e.g., Orme, 1988). More recently
however, significant progress has been made in developing theoretically valid formulae that take
additional prior information into account to provide more accurate guidance on necessary sample
sizes (see e.g., Bliemer and Rose, 2009). In terms of placing quotas on various sub segments of the

23
© Association for European Transport and contributors 2009
 population, Rose and Bliemer (2006) show how optimal sample size requirements can be
determined for each segment. Whilst Bliemer and Rose (2009) suggest that sample sizes derived
from such equations should represent a minimum bound in terms of the actual sample sizes
required for data collection, Bliemer et al. (2009) found that the real sample sizes required were
remarkably close to those suggested by the equations.
Nevertheless, in studies aimed at drawing wide and general conclusions, a certain degree of
representativeness needs to be achieved in the sample. If only a small sample is collected, there
does exist the potential that sampling bias will mean that the sample selected will not be
representative of the overall population from which it is drawn. As such, even if smaller sample sizes
are required than are generally collected, as suggested by Bliemer and Rose (2009), other external
requirements may necessitate larger sample sizes be collected in practice.

7.3 Inappropriate contexts

The only way for data collected from SC surveys to be of use in the estimation of choice models
is for respondents to make informed choices between the alternatives they are faced with. In some
ways, this relates to the realism of the choice tasks presented to each respondent, where realism
need not match reality, but rather match the sampled respondent’s perceptions of reality.
Unfortunately, in most contexts, different respondents will have different levels of experiences with
the alternatives offered and given that experience is likely to be one of the key inputs into the
formation of these perceptions, it is likely that different respondents will also have different
perceptions of the offered alternatives. Furthermore, based on their previous experiences, some
sampled respondents may be totally inappropriate for the study. Unfortunately, there exist many
such examples where respondents have been faced with choice contexts in which they are unable to
make informed decisions. For example, one mode choice survey that the authors are familiar with
involved sampling respondents without a drivers license in a choice between car and rail.. Other
such examples found within the literature include for example presenting a departure time choice to
respondents with completely inflexible working times.

7.4 Data collection methods

There are still analysts who suggest that paper based surveys are the only acceptable survey
delivery mechanism, given the greater flexibility for example for on-train surveys. Whilst paper
based surveys offer many advantages, in particular they tend to allow a much wider reach in terms
of sampling than computer based methods, as well as much lower costs, they also suffer from a
number of disadvantages. As well as the obvious potential for coding mistakes (during the
translation from paper to data files), paper based surveys may also be somewhat limited in terms of
the degree of customisation possible. Typically, where customisation is required, separate surveys
need to be prepared for a small number of segments. Furthermore, paper based surveys are
generally faced with having to relate values of alternatives not in absolute terms relative to some
reference option, but rather in such a way that respondents are required to carry out calculations to
determine the precise value for themselves. As an example, an alternative may be described as:
Current travel time + 10 minutes
Current travel cost – 50 pence
Here, two main issues arise. Firstly, there is obvious scope for numerical mistakes. Secondly, any
digression between what the respondent uses for the current travel time and current travel cost and
the values used during modelling will potentially lead to significant problems. Here, even asking
respondents for their current values at the start of the survey may not be sufficient. With both
problems, the end result would be that what is modelled is not necessarily what was used by the
respondent in evaluating the alternatives. Thirdly, given the generally accepted knowledge that
respondents struggle with the notion of percentage changes, absolute time changes have to be

24
© Association for European Transport and contributors 2009
 used, which may create problems if the number of separate surveys is low, i.e., the same savings are
used for potentially quite different journey times.
All the above listed problems can to a large degree be avoided by making use of a computer
based survey, either in the form of an interviewer assisted survey or an internet based survey. Such
surveys allow for a high degree of customisation and also carry out calculations automatically,
avoiding numerical issues while also guaranteeing a correspondence between those values used in
the survey and those values used in the models. Furthermore, such surveys can rely on percentage
variations, which may produce more realistic attribute levels, while also increasing data richness.
Finally, a point that is rarely discussed is that in paper based surveys, respondents see all choice
situations at the same time, potentially leading to cross-scenario comparisons of alternatives, an
issue that does not arise in computer based surveys where one screen is used per scenario.
The cost of interviewer assisted surveys may be prohibitively high, paving the way for internet
based surveys. Here, issues of sample representativeness may be avoided by still sampling
respondents in the same way (e.g., roadside) and by handing out login details for the internet based
survey, an approach that is again becoming more common place, for example in many river crossing
and toll road surveys in the United States.

8. Using SP data inappropriately

Stated choice surveys have proven useful in examining many transportation related issues. For
example, SC data has been used to examine the demand for cycle-way networks (e.g., Ortúzar et al.,
2000), to examine the benefits derived from various calming measures on traffic (e.g., Garrod et al.,
2002)., to study the influences on parking choice (e.g., Shiflan and Bard-Eden, 2001; Hensher and
King, 2001; van der Waerden et al., 2002) and to establish the VTT of commuters and non-
commuters (e.g., Hensher, 2001a,b). Despite the flexibility of the method which allows it to be
applied to almost any topic that one might wish to study, there do exist some limitations in terms of
what outputs might be considered relevant to generate and report. In particular, very rarely should
SP studies be used to generate measures of elasticities.

8.1 Elasticities and SP data

To understand why this is the case, consider the direct and cross elasticity equations for the MNL
model. Both equations, shown as Equations (1) and (2), require the calculation of the choice
probabilities. These choice probabilities are a function of the data as well as the parameter
estimates, including any estimated ASCs. In discrete choice models, these ASCs represent the means
of the random error terms, which reflect, after controlling for the modelled components of utility,
the choice shares of the data. Given that SC are constructed as hypothetical markets, these choice
shares are therefore reflective of the choice shares based on these hypothetical markets.
Unfortunately, even if the choice survey is made as realistic as possible, the choice shares are
unlikely to match those observed from equivalent real markets given that the experimental design
imposed will not likely reflect the true market situation. As such, the ASCs from SC data, whilst
required for estimation purposes, have limited meaning in terms of their interpretability.
Unfortunately, unless the ASC values are calibrated so that the predicted market shares match those
from real markets, the choice probabilities will also not reflect the true real market shares. Taking
this argument to its logical conclusion, it follows that the elasticity values derived from such data will
also not be correct.

dPi X ik dVi (1)

E XPiik    X ik (1  Pi )
dX ik Pi dX ik

25
© Association for European Transport and contributors 2009
 dPi X jk dV j (2)
E XPi jk    X jk Pj
dX jk Pi dX jk
This does not mean that the calculation of values such as elasticities (and likewise marginal effects)
should never be contemplated. Indeed, the opposite is true. Where one wishes to compare the
results across different models or data sets, then the generation of elasticities should be considered.
However, unless the model constants have been calibrated, the actual values of the elasticities
should be interpreted with care.
Additionally, there is a clear risk that the scale in SP data is different from that in RP data, i.e.
that respondents’ response to changes in attribute values is higher or lower than it would be in a
real life scenario. Thus, it is not only the ASCs that need recalibrating, but also the scale of the
marginal utility coefficients.

8.2 WTP and SP data

Whilst several studies have shown that SP experiments are able to reproduce the behavioural
outputs, such as WTP measures, obtained from RP choice experiments (e.g., Carlsson and
Martinsson, 2001; Lusk and Schroeder, 2004), contradictory evidence also exists that calls into
question whether results obtained from SP experiments do in fact mirror those obtained from real
markets. For example, Wardman (2001) and Brownstowne and Small (2005) found significant
differences between WTP values derived from RP and SP choice studies. In both these studies, VTT
from SP experiments were found to be undervalued in comparison to the results from RP studies.
Interestingly however, the opposite is typically observed in traditional contingent valuation studies
where WTP values have been found to over value those observed in real markets (e.g., Harrison and
Rutstrom, 2006; List and Gallet, 2001; Murphy et al., 2004; see Hensher (2008) for a detailed
overview of differences obtained between WTP values from different survey methodologies).

9. Conclusions
In this paper, we have presented examples of previous SP studies to demonstrate some practical
aspects of conducting such studies. In doing so, we have sought to provide practitioners and
academic researchers alike with recommendations that might prove useful in order to avoid
problems and mistakes that have been made in the past. We have also sought to provide discussion
on other aspects of using SP surveys that hopefully will improve the reporting and use of SP survey
results. Our discussions herein have led us to a number of conclusions, primary of which is that
qualitative research is a must and that piloting and pretesting of SP surveys is a necessity.
We wish to conclude however by stating that such survey problems do not only afflict SP
surveys. Many of the issues identified here may equally impact upon RP data collection as well. To
this end, we offer the following example, taken from Hensher et al. (2005). “Some years ago a
student undertook research into a household’s choice of type of car. The student chose to seek
information on the alternatives in the choice set by asking the household. Taking one household
who owned one vehicle, their chosen vehicle was a Mazda 323. When asked for up to three
alternatives that would have been purchased had they not bought the Mazda 323, the stated
vehicles were Honda Civic, Toyota Corolla and Ford Escort. After collecting the data and undertaking
the model estimation it was found that the vehicle price attribute had a positive sign (and was
marginally significant). After much thought it became clear what the problem was. By limiting the
choice set to vehicles listed by the respondent, we were limiting the analysis to the choice amongst
similarly priced vehicles. Consequently more expensive vehicles (and much less expensive ones)
were not being assessed in the data although some process of rejection had clearly taken place by
the household. The price attribute at best was a proxy for quality differences amongst the vehicles
(subject to whatever other vehicle attributes were included in the observed part of the utility

26
© Association for European Transport and contributors 2009
 expression). Price would be better placed in explaining what alternatives were in or out of the choice
set. If the student had simply listed all vehicles on the market (by make, model, vintage) and
considered all eligible, then regardless of which grouping strategy was used (as discussed above) one
would expect price to have a negative parameter; and indeed the model if well specified should have
assigned a very low likelihood of the particular household purchasing a vehicle in a higher and a
lower price range. An important lesson was learnt.”

Acknowledgements
The first author acknowledges the financial support of the Leverhulme Trust in the form of a
Leverhulme Early Career Fellowship.

References
AHCG (1996), Value of Travel Time on UK Roads, report by Hague Consulting Group and Accent
Marketing & Research for the UK Department for Environment, Transport and the Regions.

Axhausen, K.W., Hess, S., König, A., Abay, G., Bates, J.J. & Bierlaire, M. (2008), Income and distance
elasticities of values of travel time savings: New Swiss results, Transport Policy, 15(3), pp. 173-185.

Batley, R., Grant-Muller, S., Nellthorp, J., de Jong, G., Watling, D., Bates, J., Hess, S. & Polak, J.W.
(2008), Multimodal travel time variability, final report for the UK Department for Transport.

Bierlaire, M., Axhausen, K. and Abay, G. (2001). Acceptance of modal innovation: the case of the
Swissmetro, Proceedings of the 1st Swiss Transportation Research Conference, Ascona, Switzerland

Bliemer, M.C. and Rose, J.M. (2009) Efficiency And Sample Size Requirements for Stated Choice
Experiments, Transportation Research Board Annual Meeting, Washington DC January.

Bliemer, M.C., Rose, J.M. and Beelaerts van Blokland, R. Experimental Design Influences on Stated
Choice Outputs, European Transport Conference, Leeuwenhorst, October 5-7.

Black, I., Efron, A., Anthony, C.I. and Rose, J.M. (2005) Designing and implementing internet
questionnaires using Microsoft Excel, Australasian Marketing Journal, 13(2), 62-73.

Brazell, J.D., Diener, C.G., Karniouchina, E., Moore, W.L., Severin, V. and Uldry, P.F. (2006) The no-
choice option and dual response choice designs, Marketing Letters, 17, 255-268.
Brownstone, D. and Small. K. (2005) Valuing time and reliability: assessing the evidence from road
pricing demonstrations. Transportation Research Part A, 39, 279-293.
Carlsson, F and Martinsson, P. (2001) Do hypothetical and actual marginal willingness to pay differ in
choice experiments? Journal of Environmental Economics and Management, 41, 179-192.
Chintakayala, P.K., Hess, S., Rose, J.M. & Wardman, M.R. (2009a), Effects of stated choice design
dimensions on estimates, paper presented at the inaugural International Choice Modelling
Conference, Harrogate.
Chintakayala, P.K., Hess, S., & Rose, J.M. (2009b), Using second preference choices in pivot surveys
as a means of dealing with inertia, paper presented at the European Transport Conference,
Noordwijkerhout, The Netherlands.

27
© Association for European Transport and contributors 2009
 Dhar, R. and Simonson, I. (2003) The effect of forced choice on choice, Journal of Marketing
Research, 40(2), 146-160.
Efron A., Rose, J.M., and Roquero D. (2003) Truck or Train? A Stated Choice Study on Intermodalism
in Argentina, presented at XVII Congresso de Pesquisa e Ensino em Transportes, Rio de Janeiro,
Brazil, November 10th -14th.
Garrod, G.D., Scarpa, R. and Willis, K.G. (2002). Estimating the Benefits of Traffic Calming on Through
Routes: A Choice Experiment Approach, Journal of Transport Economics and Policy, 36(2), 211-232.
Harrison, G.W. and Rutström, E.E. (2006) Experimental evidence on the existence of hypothetical
bias in value elicitation methods, In: Handbook of Experimental Economics Results, C.R. Plott and
V.L.Smith, Eds., Amsterdam: North-Holland.
Hensher, D.A. (2008) Hypothetical bias and stated choice studies, submitted to Transportation
Research Part B.
Hensher, D.A. (2001a) The valuation of commuter travel time savings for car drivers: evaluating
alternative model specifications, Transportation, 28(2), 101-118.
Hensher, D.A. (2001b) Measurement of the Valuation of Travel Time Savings, Journal of Transport
Economics and Policy, 35(1), 71-98.
Hensher, D.A., Greene, W.H. and Rose, J.M. (2006) Deriving willingness to pay estimates of travel
time savings from individual-based parameters, Environment and Planning A, 38, 2365-2376.
Hensher, D.A. and King, J. (2001) Parking demand and responsiveness to supply, pricing and location
in the Sydney central business district, Transport Research Part A, 35(3), 177-196.
Hensher, D.A., Rose, J.M. and Greene, W.H. (2005) Applied Choice Analysis: A Primer, Cambridge
University Press, Cambridge.
Hess, S. & Adler, T. (2009), Experimental designs for the real world, ITS working paper, Institute for
Transport Studies, University of Leeds.
Hess, S. and Rose, J.M. (2009a) Allowing for intra-respondent variations in coefficients estimated on
stated preference data, Transportation Research Part B, 43(6), 708-719.
Hess, S. and Rose, J.M. (2009b) Should reference alternatives in pivot design SC surveys be treated
differently?, Environment and Planning A, 42(3), 297-317.
Hess, S., Smith, C., Falzarano, S. & Stubits, J. (2008) Measuring the effects of different experimental
designs and survey administration methods using an Atlanta Managed Lanes Stated Preference
survey, Transportation Research Record, 2049, 144-152.
Hess, S., Rose, J.M. & Polak, J.W. (2009), Non-trading, lexicographic and inconsistent behaviour in
stated choice data, Transportation Research Part D, accepted for publication, January 2009.
Kanninen, B.J. (2002) Optimal Design for Multinomial Choice Experiments, Journal of Marketing
Research, 39, 214-217.
Li, Z., Hensher, D.A. and Rose, J.M. (2009) Willingness to Pay for Travel Time Reliability for Passenger
Transport: A Review and some New Empirical Evidence, submitted to Transportation Research Part
E.
List, J. and Gallet, G. (2001) What experimental protocol influence disparities between actual and
hypothetical stated values? Environmental and Resource Economics, 20, 241-254.
Louviere, J.J., Street, D., Burgess, L., Wasi, N., Islam, T. and Marley A.A.J. (2008) Modeling the
choices of individual decision-makers by combining efficient choice experiment designs with extra
preference information, Journal of Choice Modelling, 1(1), 128-163.

28
© Association for European Transport and contributors 2009
 Lusk, J. and Schroeder, T. (2004) Are choice experiments incentive compatible? A test with quality
differentiated beef steaks, American Journal of Agricultural Economics, 86(2), 467-482.
Marley, A.A.J. and Louviere, J.J. (2005) Some probabilistic models of best, worst, and best–worst
choices, Journal of Mathematical Psychology, 49, 464–480.
Murphy, J., Allen, P., Stevens, T. And Weatherhead, D. (2004) A meta-analysis of hypothetical bias in
stated preference valuation, Department of Resource Economics, University of Massachusetts,
Amherst, January.
Orme, B. (1998) Sample Size Issues for Conjoint Analysis Studies, Sawtooth Software Technical
Paper, http://www.sawtoothsoftware.com/technicaldownloads.shtml#ssize.
Ortúzar, J. de D,, Iacobelli, A. and Valeze, C. (2000) Estimating demand for a cycle-way Network,
Transport Research Part A, 34(5), 353-373.
Rose, J.M. and Bliemer, M.C.J. (in press) Constructing Efficient Stated Choice Experimental Designs,
Transport Reviews.
Rose, J.M. and Bliemer, M.C. (2006) Designing Efficient Data for Stated Choice Experiments,
presented at 11th International Conference on Travel Behaviour Research, Kyoto, August 16-20, 2006,
Japan.
Rose, J.M., Bliemer, M.C., Hensher and Collins, A. T. (2008) Designing efficient stated choice
experiments in the presence of reference alternatives, Transportation Research Part B, 42(4), 395-
406.
Rose, J.M. and Hensher, D.A. (2006) Handling individual specific non-availability of alternatives in
respondent's choice sets in the construction of stated choice experiments, Stopher, P.R. and Stecher
C. (eds.) Survey Methods, Elsevier Science, Oxford, pp347-371.
Rose, J.M. and Hess, S. (2009) Dual Response Choices In Reference Alternative Related Stated Choice
Experiments, Transportation Research Board Annual Meeting, Washington DC January.
Shiflan, Y. and Bard-Eden, R. (2001) Modeling Response to Parking Policy, Transport Research
Record, 1765, 27-34.
van der Waerden, P., Timmermans, H. and Borgers, A. (2002) PAMELA: Parking Analysis Model for
Predicting Effects in Local Areas, Transport Research Record, 1781, 10-18.
Wardman, M. (2001) A review of British evidence on time and service quality Valuations,
Transportation Research Part E, 37, 91-106.
Wittink, D.R., Huber, J., Zandan, P. and Johnson, R.M. (1992) The Number of Levels Effect in
Conjoint: Where Does It Come From and Can It Be Eliminated?, Sawtooth Software Conference
Proceedings.
Wittink, D.R., Krishnamurthi, L. and Nutter, J.B. (1982) Comparing Derived Importance Weights
Across Attributes, Journal of Consumer Research, 8, 471 -4.
Wittink, D.R., Krishnamurthi, L. and Reibstein, D.J. (1989) The Effects of Differences in the Number of
Attribute Levels on Conjoint Results, Marketing Letters, (2), 113-23.

29
© Association for European Transport and contributors 2009