Stated Preference survey design in air travel choice behaviour modelling Andrew Collins Institute for Transport and Logistics Studies - University of Sydney Stephane Hess Centre for Transport Studies - Imperial College London Institute for Transport and Logistics Studies - University of Sydney John Rose Institute for Transport and Logistics Studies - University of Sydney

Abstract Like in most other areas of choice analysis, the vast majority of studies of air travel choice behaviour now make use of Stated Preference (SP) rather than Revealed Preference (RP) data. However, experience as to the optimal design of such surveys in the context of aviation is still very limited. This is a crucial issue given the complexity of the choice processes undertaken by air travellers, involving choices among a multitude of dimensions. In this paper, we present results from a pilot study making use of an innovative SP survey that combines traditional SP style questionnaires with choice experiments modelled on typical online booking engines. Even with the very limited sample size used in this pilot study, significant and meaningful effects of all major attributes are retrieved. Furthermore, there are interesting insights into the different ways in which respondents process the information from the two types of survey questionnaires.

1. Introduction As set out in Section 2, there is increasing interest in modelling the choice processes undertaken by air travellers. Typically, these studies make use of discrete choice models that allow analysts to relate observed or stated choices to the attributes of the various alternatives that the respondents are faced with. A major problem in the area of air travel choice behaviour research is the poor quality of Revealed Preference (RP) data, leading to an inability to estimate significant effects for a number of crucial attributes such as air fares. Similarly however, the Stated Preference (SP) surveys used in the field of air transport have generally not been able to do justice to the complexity of the choice processes faced by air travellers. In this paper, we present initial results from a pilot study making use of an innovative new SP survey for air travel choice behaviour. Here, in addition to a standard SP survey, respondents are faced with a questionnaire modelled on

online booking engines, giving respondents full flexibility in searching for flights, sorting flight options by attribute levels, and eliminating subsets of alternatives that do not meet certain criteria. This approach not only allows us to face respondents with a more realistic choice situation (large choice set with many attributes per alternative), but also provides us with important insights into respondents’ information processing strategies. The remainder of this paper is organised as follows. The next section presents a brief overview of relevant existing work in the area of air travel behaviour research. Section 3 discusses the design of our SP survey, while Section 4 presents some initial insights into how respondents process the information presented to them in the survey. This is followed in Section 5 by a description of a choice modelling analysis carried out on this dataset. Finally, Section 6 presents the conclusions of the research.

2. Literature review A wide range of studies have investigated air travel choice behaviour using both SP and RP methods. Kanafani & Sadoulet (1977) modelled the choice among fare types for long haul journeys. Proussaloglou & Koppleman (1995) examined the choice of airline for recent trips using mail-in RP data. In recent years, a majority of studies have used the SP methodology. Bradley (1998) used SP data to examine the choice of departure airport and route from Schiphol, Brussels and Eindhoven airports. Hensher et al. (2001) use SP data for airline choice between New Zealand and Australia. Hess et al. (2007) and Hess (2007) also make use of SP data collected via the internet and retrieve effects of a number of attributes which often cause problems in RP data (fares, frequent flier benefits). Despite research evidence suggesting that SP experiments are capable of replicating decisions made in real markets (Burke et al., 1992; Carson et al., 1994), contradictory evidence exists to suggest that they do not. For example, Wardman (2001) found that the values of travel time savings (VTTS) from SP experiments are typically undervalued in comparison to the results from revealed preference (RP) studies. In the aviation context, Algers and Beser acknowledge the limitations of both SP and RP data in the choice of flight and booking class. RP data does not typically allow insights into how people respond when their desired ticket class is not available. SP data can introduce biases. They propose using both SP and RP data for the analysis. Rose and Hensher (2006) argue that one factor in the ability of SP experiments to replicate real market decisions is the degree of realism imposed in SP surveys, where realism in SP experiments arises not simply from the fact that respondents are asked to undertake similar actions as they would in real markets, but also in how much the experiment is made to look and feel like choices made in real market transactions. In this study, we test whether a choice environment that mimics online travel agent websites leads to different conclusions than those formed from a traditional SP experiment. Proussaloglou & Koppleman (1999) performed a novel air travel SP survey that incorporated the way that travellers search for information when talking to a travel

agent on the phone. After an initial collection of data by mail, they collected the SP data in a telephone survey. Presented with a travel scenario, the respondents had to elicit from the interviewer the available flights as described by schedule and fare. Flights could be revealed in any order the respondent wished – according to schedule or fare, and a choice could be made at any stage. The interviewer had a record of what flights had been revealed when the choice was made. Key methodological aspects of this study include the ability of the respondent to drive the search process prior to making a decision, and the use of flights that were close to real world options. The online search component of our study takes a similar approach to that of Proussaloglou & Koppleman, and examines search and choice decisions in the context of the online purchase of airline tickets. The traditional SP component allows for a useful comparison of the two approaches.

3. Survey design At the present time, we have only collected pilot data from a limited number of respondents. Specifically, the internet based survey was sent out to two university transport groups in the UK (Imperial College London and Leeds University), one transport group in Switzerland (ETH Zurich), and a set of other transport professionals from the UK. For the pilot study, data from 36 respondents were available. Respondents were asked to imagine they were taking a three week holiday to Australia or New Zealand, paid out of their own expenses, with Sydney and Auckland available as destinations. Long haul flights were selected as long haul travellers are more likely than short haul travellers to be discerning about comfort attributes such as in flight entertainment and seat pitch. Willingness to pay measures are likely to only be valid on flights of equivalent duration, so no mix with short haul and long haul destinations was used. With the small sample size used for this pilot study, only economy flights for leisure travellers were considered. No choice of departure airport was presented for respondents departing from London, where this is arguably less important given the distance travelled. Both destination areas are only served by a single airport. The survey presented to respondents contained two choice components: a traditional Stated Preference component with 7 choice tasks, and a ticket search component modelled on online search engines, with 4 search tasks and a practice search. The order of the two components was randomised, as was the order of the tasks within each component. Both types of task consisted of the same attributes, listed below in Table 1. For each attribute, the ranges of attribute levels for the two sets of choice tasks (SP/search) are shown. While an experimental design was used for the SP tasks, the search tasks made use of information from real world flights (where available). Two price components were shown: a carbon tax, and the ticket price excluding the carbon tax. While all taxes are typically bundled together, or included in the ticket price, we separated a possible carbon tax to test for differences in perceptions of carbon tax costs. Real airline names were

displayed, always with their logo visible. Some of the comfort related attributes are not typically shown on ticket booking websites, including seat pitch and the onboard entertainment system. In part these attributes were included simply to test for the impact of these attributes on airline ticket choice. However, they also represent types of information that could easily be added to ticket booking websites for ready comparison between flights, but remain difficult to search for. Airline specific information may not be available or buried in a website. Sites such as SeatGuru (www.seatguru.com) are paving the way, providing seat pitch and entertainment system information for many airline-aircraft combinations. The cost of an itinerary change is usually shown by online travel agents however this is typically in the fine print once a flight is chosen. Attribute

SP levels

Search levels or range

Search: From real flight?

Typical online travel agent attribute?

Price

AUD1600, AUD1900, AUD2200, AUD2500

AUD1895 – AUD19402

Yes

Yes

Carbon tax

AUD0, AUD120, AUD240, AUD360

AUD0 - AUD581.32

No

Yes

Airline

9 possible

18 possible

Yes

Yes

Departure time

6am, 10am, 5pm, 10pm

Continuous

Yes

Yes

Arrival time

Based on departure time and flight duration

Continuous

Yes

Yes

Total duration

20hrs, 22hrs, 24hrs, 26hrs

20hrs – 36hrs 45mins

Yes

Yes

Flight duration

Based on total stopover duration

19hrs 25mins – 25hrs 50mins

Yes

Yes

Stopover duration

1hr, 2hrs, 3hrs, 4hrs

55mins – 14hrs 5mins

Yes

Yes

Number of stops

1, 2

1, 2, 3

Yes

Yes

and

Plane type

747, 777, A330, A340

No

Yes

Seat pitch

31”, 32”, 34”

No

No

Yes/No

No

Yes

No

No

No

Often hidden

Seat allocation available? Entertainment system

Itinerary change cost

Overhead televisions (shared), Personal screens with limited movie Personal screens with video on demand

selection,

AUD0, AUD100, AUD200, AUD300 Table 1: Attributes in SP and search tasks

There are other factors that influence air travel choice in the literature that we have not included. Frequent flyer membership is widely recognised to have a significant influence on choice (Proussaloglou & Koppelman, 1995; Chin, 2002; Proussaloglou & Koppelman, 1999; Hensher et al., 2001). We have not included any frequent flyer attributes such as point multipliers. However, we asked what frequent flier programs the respondent belonged to, and this can be used in the analysis. Furthermore, airport and access mode choice were ignored, where the effect of this is possibly mitigated by the long haul nature of the flights presented.

3.1. Traditional SP tasks The Stated Preference component consisted of 7 choice tasks, each with 3 alternatives described by all of the attributes listed in Table 1. Respondents were asked to list their preferred flight, but were also given a no choice option. Furthermore, respondents were directed to indicate if any attributes were ignored, and were asked if some of the alternatives would never be chosen. An example of the choice screen is shown in Figure 1. A D-optimal efficient design was used, with 12 blocks of 7 choice tasks each..

Figure 1: Stated Preference task

3.2. Search tasks Whereas traditional Stated Preference experiments typically have a small number of alternatives, online travel agents (OTA) usually present a large number of ticket options for any given day. To help the customers deal with the large amount of information, most OTA sites allow the flights to be sorted on at least some of the attributes. Furthermore, after an initial search for specific dates, the search can be refined in various ways, so that flights that do not meet a specified criteria are not shown. In a departure from traditional SP methodology, we have tried to mimic the decision making process of a visitor to an OTA site. This approach includes a large number of alternatives that are based largely on real flight information, and the provision of extensive search and sort tools.

The flights for the search tasks were based on real flights that were obtained from a popular OTA site. To prevent extensive correlations within airlines, the plane type, seat pitch, seat allocation, entertainment system and cost of itinerary change attribute levels were not drawn from the real flights. Instead, for each attribute, each level was allocated an equal number of times. The levels were then swapped between flights such that the correlations between attributes were minimised. Four search tasks were presented to the respondents, in addition to a practice search task which contained four flights only. Real flight prices vary over time for the same flight due to yield management systems. Also, travel at certain times of the year will always be more expensive due to high demand. Consequently, it was decided to populate each of the four search tasks with flights that were selected at different times in the future. This allows for a good coverage of flight prices in the sample. Flights were selected for departure in two weeks’ time, in a month’s time, in five months’ time, and over Christmas. These timeframes were explicitly mentioned to the respondents to help them understand why the average prices varied from task to task. In real decision environments, a decision about both the departing and return flights must be made. In the interests of simplicity, we only required a choice for the departing flight, and asked the respondent to assume that the return flight would have similar service levels. Figure 2 shows how the tickets appear in the search tasks, with all attributes shown in this example.

Figure 2: Two flights displayed in the search tasks

The top of the search task screens contained a set of tools that included sort, search, a description of the attributes, and a means to hide some attributes. Figure 3 shows this set of tools. All attributes could be sorted on, with the flights instantly being resorted on the screen. Sorting was performed in the most obvious way for each attribute: with the first flights on the screens having the lowest prices, shortest durations, earliest flights, largest seat pitch, best entertainment system and so on. By default, the flights were sorted on price. Sort selections were preserved from one task to the next.

Figure 3: Sort, search and hide attribute tool

All attributes except for departure and arrival time were able to be searched on. All costs and most durations could be searched on with a respondent specified maximum. Other attributes could be searched on by choosing a category. Searches on stopover duration were limited to distinct categories that did not overlap. This was done both for simplicity and to test whether some respondents wanted a minimum stopover time while others wanted a maximum. Unlike sorting, the search tool had the potential to reduce the size of a respondent’s consideration set. It was possible to enter overly restrictive search criteria and find no flights. This resulted in a message being shown suggesting less stringent search criteria. Any number of searches could be performed. By default, no search criteria were applied, although search criteria were preserved from one task to the next. Price, carbon tax, airline name, departure time and arrival time are always shown. All other attributes can be hidden and shown at whim via the set of tools. We provided this option as a way to remove unnecessary clutter on the screen and help facilitate easier, faster decision making. Further, for the practice task, we did not show the attributes that could be hidden, and forced the respondent to show the attributes that mattered to them.

3.3. Collection of other information In addition to the choice observations, information was collected on how many times the respondent had travelled domestically, internationally, and to Australia and New Zealand over the last three years, broken down by whether the ticket had been paid for by themselves or others. The number of unique airlines flown with over the previous three years was obtained, as was frequent flyer membership and the usual class of ticket purchased for international flights. Sociodemographic information was collected on age, work status, typical work hours per week, income and gender, where in each case respondents were given the possibility not to disclose this information.

4. Initial insights into respondent behaviour In the search tasks, the flights can be sorted on any attribute, with the initial default being a sort by price. Table 2 indicates both which attributes were sorted on at the time a choice was made, and how many times an actual sort was explicitly performed. Since sort information is preserved between tasks, for any given attribute there may be fewer sort actions than tasks that were sorted on that attribute at the time of choice. Furthermore, since many sorts can be performed before a choice is made, there may be more sort actions than tasks that were sorted on that attribute at the time of choice. Note that in Table 2, only the 4 real tasks were considered, and the practice task data was excluded. Some respondents may have established their sort preferences in the practice search task, and maintained these preferences for the remaining tasks. Clearly price is the dominant sort attribute. However this is not for lack of engagement with the sort tool: most respondents performed some sorts, even if they reverted to the default of sorting by price for their choices. Only 3 of 36 respondents never performed a sort. Total duration and seat pitch had some share of the sorts, but the remaining non-price sorts were shared fairly evenly between the remaining attributes. Tasks with this sort at Sort actions performed time of choice Price 76 53% 39 40% Price (by default) 13 9% Carbon tax 4 3% 11 11% Airline 5 3% 4 4% Departure time 4 3% 4 4% Arrival time 3 2% 5 5% Total duration 15 10% 19 19% Flying duration 3 2% 4 4% Stopover duration 1 1% 1 1% Number of stops 7 5% 2 2% Plane type 0 0% 0 0% Seat pitch 8 6% 5 5% Seat reservation 0 0% 0 0% Entertainment 1 1% 2 2% system Ticket change 4 3% 2 2% charge Total 144 100% 98 100% Table 2: Sorting strategies

While some information is gleaned from the above aggregate results, a breakdown of sort behaviour at the individual level is revealing. Of the 36 respondents, 19 made all their choices while sorted on price. 7 more respondents

made choices while sorted on price for some tasks, and other attributes for the remainder. The remaining 10 respondents never chose while sorted on price. Two of these respondents sorted on a mixture of attributes, but eight respondents chose while sorted on the same attribute for all their tasks. All of these eight bar a pair sorted consistently on different attributes: departure time, airline, number of stops, seat pitch, ticket change charge, carbon tax, and total duration, with the later being used by two respondents. What this shows is that there is heterogeneity of sort preference. Beyond price, which is clearly the most popular sort attribute, there exists a variety of sort preferences. Yet these preferences are usually consistent within individuals, with 75% of respondents sorting on the same attribute at the time of choice for all 4 tasks. Number of sort actions per Number of tasks task 0 106 74% 1 16 11% 2 7 5% 3 8 6% 5 5 3% 8 1 1% 11 1 1% 144 100% Table 3: Number of sort actions per task

Table 3 indicates the level of sort activity per task. 74% of tasks presented resulted in no sort actions by the respondent. This is not to say no sort was applied: the sort criteria may have been carried over from previous tasks. It is worth noting then that 11% of tasks have 1 sort action and 15% have 2 or more. It is likely that these respondents are finding use for each of the attribute sorts in their decision making process. Evidence of this at the individual level is that of the 36 respondents, 6 performed multiple sorts on at least 3 out of the 4 tasks. A significant minority of respondents are making use of multiple searches. As with sorting, only some of the respondents used the search tool. Out of 36 respondents, 17 performed no search in any of the non-practice tasks, while 19 respondents performed at least one search and of these 14 performed searches in at least 3 out of 4 tasks. Table 4 shows, at the attribute level, the number of tasks for which a search criterion was applied at the time of choice. As with sorting, price is the dominant attribute, and total duration and seat pitch were notable. Unlike sorting, number of stops and the entertainment system were also selected with some frequency. The video on demand option was commonly sought after in the search, and this is reflected in generally high significance levels in the various models estimated (cf. Section 5). Interestingly though, the number of stops attribute was never significant in the models and often of the wrong sign. This is likely to be a consequence of sample size; however the search criteria anecdotally suggest that the number of stops is important for at least some of the respondents.

Price Carbon tax Airline Total duration Flying duration Stopover duration Number of stops Seat pitch Seat reservation Entertainment system Ticket change charge

Number of tasks with search criteria applied for this attribute at time of choice 28 11 5 20 8 13 17 23 2

19% 8% 3% 14% 6% 9% 12% 16% 1%

29

20%

0

0%

Table 4: Number of tasks with search criteria applied for each attribute at time of choice

The manner in which each attribute was sorted is interesting. Some attributes have a clear preference sign, including price and entertainment system. Price limits were typically close to the bottom end of the price range, and all entertainment system searches specified video on demand as essential. Other attributes are likely to be considered in different ways across the population. The stopover duration levels were mutually exclusive, and searches on this attribute were split between a desire to minimise stop time (up to 2 hours), and a desire to have a lengthy stop time (2-4 hours and even 4+ hours). Either strategy is plausible. The former would minimise total travel time, while the later would provide a lengthy break from a confined environment, or perhaps provide an opportunity for shopping. Taste heterogeneity can be captured in advanced models such as mixed logit. However, the search tool can help reveal individual level taste signs with some certainty (while noting that an absence of a search does not imply an absence of equivalent tastes). An area for future research is to consider how the search information may be incorporated in advanced model structures. Unlike sorting, search criteria can be applied across multiple attributes concurrently. Table 5 indicates for how many tasks multiple attribute searches were performed. 24% of all tasks were completed with multiple search criteria applied. 8 respondents had multiple search criteria for all 4 tasks. Respondents tend to search on the same attribute(s) from task to task, with some extra attributes potentially being included in some tasks. So a significant minority of respondents perform complex searches to assist in their decision making process. Indeed it is the respondents who perform complex searches that are most likely to obtain benefits from the search tool. For example, trying to

compare 7 attributes across 24 alternatives is a cognitively complex process, but also, for some respondents, a realistic decision making environment. Respondents who make decisions on one or perhaps two attributes are unlikely to waste their time on a search. For them, a sort is likely to prove more useful. Number of concurrent search Number of tasks criteria 0 79 55% 1 31 22% 2 12 8% 3 7 5% 4 9 6% 5 1 1% 6 1 1% 7 4 3% 144 Table 5: Number of concurrent search criteria

Whereas sort actions only reorder the flights on screen, search actions actually add or remove flights from view. In a sense, this makes a search a stronger type of filter, as any flight that fails to meet the search criteria cannot be chosen. Figure 4 shows a distribution of the number of flights shown at the time of choice as a percentage of the potential number of flights shown if no search criteria were applied. Of the tasks where a search was performed, the median band is a showing of just 30-40% of potential flights. These reductions are quite large in absolute terms when some search tasks contain 24 potential flights. The distribution shows that the search tool is providing an effective way to reduce the size of the consideration set of alternatives. 90

83

80

60 50 40 30

10

13 6

18

6

6

8 3

1

0 80-90%

20

70-80%

Number of tasks

70

90-100%

60-70%

50-60%

40-50%

30-40%

20-30%

10-20%

0-10%

0

Figure 4: Number of flights shown at choice as percentage of number of potential flights shown

5. Choice modelling analysis This section summarises the estimation results of a modelling analysis making use of discrete choice structures. We first look at model specification before discussing the results in detail. A mixture of NLogit (Greene, 2007) and Biogeme (2003) were used for estimation.

5.1. Model specification For the present study, we had information from 36 respondents, each completing a set of standard SC tasks along with the search tasks described in detail in Section 3. The final sample contained information on 417 choices. All models used a linear in attributes specification of the observed utility function, and the analysis was restricted to the use of Multinomial Logit (MNL) models. Before proceeding with the description of the actual modelling results, a matter that deserves some special attention is the specification of the scale of the utility functions. As described in detail later on, there are various possible scenarios that could lead to differences in scale across different subsamples of the data. By differences in scale, we refer to a larger or smaller relative impact of the observed part of utility. In the present study, the following specification was used to allow for differences in scale. Here, in the presence of K groups, with utility for alternative i in each group given by: Ui,1 = Vi,1 + εi,1 …

Ui,k = Vi,k + εi,k … Ui,k = Vi,k + εi,k

[1]

For normalisation, we set the scale in group 1 to value of 1. We then estimate K1 separate scale parameters α and multiply the utilities of all (but the first) alternatives by an appropriate scale parameter. From this, we get: Ui,1 = Vi,1 + εi,1 …

αkUi,k = αkVi,k + αkεi,k … αKUi,k = αKVi,k + αKεi,k

[2]

This process is repeated for all alternatives, with scale parameters generic across available alternatives. We now have that the variance of εi,k is π2/(6µ2), and, to ensure homoscedasticity in our model with rescaling, we get that:

var(εi,1) = αk2var(εi,k) for all k

[3]

such that: αk=µk/µ1.

[4]

Hence, with αk>1, we have that µk>µ1, and hence var(εi,k)< var(εi,1). As such, in a group where the scale parameter is larger than 1, the relative weight of the observed part of utility (compared to unobserved) is larger than in the base group.

5.2. Model results Twelve separate models were estimated. Table 6 contains a summary of these models, and the key differences between them. Table 7 to Table 10 present detailed results for each of the models. Before estimation of the models, all monetary values were transformed to Australian dollars (AUD). Model

Description

1 Base model 2 Based on 1, but with slot constants 3 Based on 2, but with airline constants All remaining models based on model 3: 4 With timing effects 5 Separate scales for search and SP tasks 6 Only for search tasks 7 Only for SP tasks Separate scales for UK and Switzerland 8 departures Separate scales depending on whether 9 search or SP tasks were presented first Only respondents who answered SP tasks 10 first Only respondents who answered search 11 tasks first 12 Separate scales for 4 sampling locations

Number of adj. observations ρ2 417 0.246 417 0.287 417 0.286 417 417 151 266

0.294 0.285 0.288 0.239

417

0.285

417

0.285

194

0.2

223 417

0.254 0.299

Table 6: Summary of model results

•

Model 1 was the initial model estimated. As expected, airfare and flight time are highly significant. The comfort attributes of video on demand and seat pitch are also significant. However, carbon tax is only significant at

•

•

•

the 41% level and the number of stops at the 30% level. Seat reservation is of the wrong sign, but is in any case only significant at the 42% level. The value of travel time savings (VTTS), at AUD35.78 per hour of flight time, seems acceptable. The fact the sensitivity to stop over time is slightly lower is also consistent with intuition, with such a long flight. Willingness to pay (WTP) for reducing the number of stops is low at AUD19.36 per stop, but must be treated with caution in any case given the low significance level for the number of stops. WTP measures for the comfort attributes seem quite plausible for a long haul flight. Respondents are willing to pay AUD122.63 to have personal TV screens instead of shared TV screens, where the WTP for on demand TV is even higher, at AUD218.04. WTP for seat pitch might seem a little extreme at AUD70.64 per inch, but given that the seat pitches shown range only from 31” to 34”, this represents a cost of AUD211.92 for the largest seat pitch over the smallest. Model 2 considers the effect of the order of the alternative on the screen. Particularly in search tasks, where there may be up to 24 alternatives, the position of an alternative on the screen, or its slot, will have an effect on the utility of the alternative. Slot 1, the first alternative on the screen, is used as a base. A single constant is estimated for all slots beyond slot 10, as these slots are often not shown, and rarely chosen. Model 2 provides a significant improvement in model fit over model 1, with an adjusted pseudo ρ2 of 0.287. The significance of all slot constants shows that later alternatives are less likely to be chosen. There is likely to be some correlation between slot position and the sorting of flights, which are often sorted on price. However, the visibility of alternatives is also likely to have some effect. Looking at the parameter estimates, the coefficients for the number of stops and seat reservation now have the correct sign, but are still not significant. Interestingly, all WTP values except for seat pitch decrease slightly. Model 3 adds dummy terms for the airlines, with airline 1 used as the base. It was not possible to estimate effects for airlines 7, 11, 17 and 18, where this is most likely due to the fact that these airlines were rarely presented and even more rarely chosen. After taking into account the increase in the number of parameters, the improvement in model fit obtained by this model is only just significant. Most WTP values are higher than in model 2. This model serves as the base for all following models. Model 4 examines departure and arrival time effects by adding dummy terms for early and late departure and arrival, and for flights that arrive two days after departure. The results are not very intuitive, with late arrival (after 8pm) and 2 days flights having significant positive parameters. The VTTS are also very high. A larger sample might improve these results, but at present departure and arrival times will be omitted from our models. A potential reason for these results is the non-random distribution of these attribute values in the sample. Most flights arrive two days after departure, and this is clearly reflected in the associated constant.

•

•

• •

•

•

Model 5 is based on model 3, but tests for differences in scale between search and SP tasks. Here, the difference between the scale parameters is only significant at the 29% level, such that no evidence of differences exists. Different results might be obtained with a larger sample. Model 6 and model 7 look separately at the results for the search tasks and standard SP tasks. Of note are the much higher WTP and VTTS values for the search sample. However, This has to be treated with caution, as flight time is significant for the SP sample but not for the search sample. The combined model fit of -395.16 is an improvement over the log likelihood for model 3 (-401.98), but the difference is not significant. The WTP differences look promising, but a larger sample is required to quantify differences and significance levels of differences. Model 8 tests for differences in scale between departures from the UK and Switzerland. The difference in scale is only significant at the 10% level. Model 9 has separate scales depending on whether the search or SP tasks were presented first. The number of observations were fairly balanced, with 194 observations with SP first and 223 observations with the search first. As with model 8, there were not significant differences in scale. Model 10 is estimated only on the sample that answered the SP tasks first, while model 11 only uses the sample that saw the search tasks first. The combined log-likelihood of -408.41 is worse than that of model 3. Nevertheless, it is interesting to note that the WTP in the model for respondents who were given the search tasks first are significantly higher. This is consistent with the results for models 6 and 7 and could suggest that the process of learning resulting from completing the search tasks first ahs an impact on behaviour in the traditional SP tasks. Model 12 estimates separate scale parameters for respondents in the four survey locations, but not significant differences are observed.

Model

1

Observations Final LL

2

3

417

417

417

-448.29

-414.65

-401.98

adj. ρ2

0.246

0.287

0.286

estimate

rob. p-val.

estimate

rob. p-val.

estimate

rob. p-val.

δ(airline 1)

0

-

0

-

0

-

δ(airline 2)

0

-

0

-

0.2020

0.39

δ(airline 3)

0

-

0

-

-0.6770

0.06

δ(airline 4)

0

-

0

-

0.1010

0.76

δ(airline 5)

0

-

0

-

0.6350

0.03

δ(airline 6)

0

-

0

-

0.2840

0.42

δ(airline 7)

0

-

0

-

-0.0534

0.89

δ(airline 8)

0

-

0

-

0.1210

0.78

δ(airline 9)

0

-

0

-

0

-

δ(airline 10)

0

-

0

-

0.1230

0.76

δ(airline 11)

0

-

0

-

0

-

δ(airline 12)

0

-

0

-

-0.9290

0.01

δ(airline 13)

0

-

0

-

-0.4240

0.70

δ(airline 14)

0

-

0

-

0.8270

0.43

δ(airline 15)

0

-

0

-

0.2670

0.70

δ(airline 16)

0

-

0

-

-0.9350

0.45

δ(airline 17)

0

-

0

-

0

-

δ(airline 18)

0

-

0

-

0

-

δ(slot 1)

0

-

0

-

0

-

δ(slot 2)

0

-

-0.3930

0.01

-0.3780

0.01

δ(slot 3)

0

-

-0.5830

0.00

-0.5710

0.00

δ(slot 4)

0

-

-0.9370

0.01

-0.9370

0.01

δ(slot 5)

0

-

-1.3500

0.00

-1.3100

0.00

δ(slot 6)

0

-

-1.6300

0.01

-1.6600

0.00

δ(slot 7)

0

-

-2.9700

0.01

-2.9600

0.01

δ(slot 8)

0

-

-2.6300

0.01

-2.6300

0.01

δ(slot 9)

0

-

-2.5000

0.01

-2.4700

0.01

δ(slot 10+)

0

-

-2.8600

0.00

-2.8200

0.00

β(air fare)

-0.0033

0.00

-0.0028

0.00

-0.0027

0.00

β(carbon tax)

-0.0003

0.59

0.0002

0.81

0.0005

0.45

β(fee for changes)

-0.0015

0.01

-0.0008

0.21

-0.0007

0.25

β(flight time)

-0.0020

0.00

-0.0015

0.01

-0.0018

0.00

β(stopover time)

-0.0013

0.11

-0.0011

0.21

-0.0011

0.21

β(stops)

0.32

-0.0633

0.70

0.0416

0.80

0.1710

δ(overhead TV screens)

0

-

0

-

0

-

δ(personal TV screens)

0.4010

0.03

0.2250

0.23

0.3300

0.10

δ(video on demand)

0.7130

0.00

0.5070

0.00

0.5570

0.00

β(pitch)

0.2310

0.00

0.2550

0.00

0.2090

0.00

δ(seat reservation)

0.23

-0.0728

0.58

0.1480

0.26

0.1710

δ(747)

0

-

0

-

0

-

δ(777)

0.2830

0.12

0.0871

0.64

0.1420

0.47

δ(A330)

0.4840

0.01

0.3430

0.06

0.4520

0.02

δ(A340)

0.2900

0.13

0.0705

0.71

0.2720

0.25

WTP flight time (AUD/hr)

35.78

33.60

38.90

stopover time (AUD/hr)

24.59

23.78

25.05

stops (AUD/stop)

19.36

-15.13

-62.64

personal TV screens (AUD)

122.63

81.82

120.88

video on demand (AUD)

218.04

184.36

204.03

pitch (AUD/inch)

70.64

92.73

76.56

seat reservation (AUD)

-22.26

53.82

62.64

Table 7: Results for models 1-3

Model

4

5

6

417

417

151

-392.31

-401.56

-195.81

0.294

0.285

Observations Final LL adj. ρ2 δ(airline 1)

0.288

estimate

rob. p-val.

estimate

rob. p-val.

estimate

rob. p-val.

0

-

0

-

0

-

δ(airline 2)

0.2630

0.27

0.1900

0.47

0.2120

0.59

δ(airline 3)

-0.6880

0.06

-0.8330

0.20

0.0806

0.95

δ(airline 4)

0.1320

0.70

0.1570

0.72

2.0100

0.00

δ(airline 5)

0.7390

0.01

0.6850

0.05

0.7440

0.11

δ(airline 6)

0.2810

0.44

0.3010

0.47

0

-

δ(airline 7)

0.0930

0.80

-0.1050

0.82

0.3950

0.77

δ(airline 8)

0.3050

0.50

0.0458

0.93

0.3170

0.66

δ(airline 10)

0.2440

0.55

0.1010

0.83

0

-

δ(airline 12)

-0.7960

0.05

-1.1700

0.18

0

-

δ(airline 13)

-0.1050

0.93

-0.3010

0.79

0.0500

0.97

δ(airline 14)

1.0300

0.33

0.8890

0.42

0.8880

0.45

δ(airline 15)

0.5790

0.43

0.1650

0.83

1.3200

0.09

δ(airline 16)

-0.5850

0.63

-1.0200

0.44

-0.8840

0.47

δ(slot 1)

0

-

0

-

0

-

δ(slot 2)

-0.3470

0.02

-0.4060

0.02

-0.3900

0.12

δ(slot 3)

-0.4990

0.00

-0.6210

0.01

-0.5540

0.07

δ(slot 4)

-0.8520

0.02

-0.9170

0.01

-0.9740

0.02

δ(slot 5)

-1.3300

0.00

-1.3000

0.01

-1.3100

0.01

δ(slot 6)

-1.6500

0.01

-1.6500

0.01

-1.7800

0.00

δ(slot 7)

-3.1000

0.00

-2.9700

0.01

-3.1900

0.00

δ(slot 8)

-2.5600

0.01

-2.5800

0.01

-2.6700

0.01

δ(slot 9)

-2.5200

0.02

-2.4000

0.02

-2.4100

0.02

δ(slot 10+)

-2.7900

0.00

-2.7600

0.00

-3.1600

0.00

β(air fare)

-0.0029

0.00

-0.0032

0.04

-0.0032

0.00

β(carbon tax)

-0.0001

0.85

0.0007

0.45

-0.0006

0.68

β(fee for changes)

-0.0012

0.08

-0.0008

0.28

-0.0009

0.40 0.19

β(flight time)

-0.0038

0.00

-0.0022

0.17

-0.0063

β(stopover time)

-0.0028

0.01

-0.0015

0.33

-0.0016

0.22

β(stops)

0.1400

0.44

0.1650

0.42

-0.3050

0.51

δ(personal TV screens)

0.4230

0.03

0.3750

0.15

0.6940

0.04

δ(video on demand)

0.6520

0.00

0.6060

0.01

0.6880

0.01

β(pitch)

0.2350

0.00

0.2430

0.04

0.2410

0.01

δ(seat reservation)

0.10

0.2340

0.11

0.2370

0.37

0.5670

δ(747)

0

-

0

-

0

-

δ(777)

0.1120

0.59

0.1030

0.68

0.2130

0.51

δ(A330)

0.4390

0.03

0.5100

0.07

0.2650

0.69

δ(A340)

0.3060

0.21

0.2650

0.34

-1.0800

0.15

δ(departure prior to 8AM)

0.22

0.53

1.00

-

1.00

-

δ(departure after 8PM)

-1.29

0.00

1.00

-

1.00

-

δ(arrival prior to 8AM)

-0.63

0.13

1.00

-

1.00

-

δ(arrival after 8PM)

0.95

0.00

1.00

-

1.00

-

δ(arrival 2 days later)

2.42

0.00

1.00

-

1.00

-

α(search tasks)

1.00

-

1.00

-

1.00

-

α(SP tasks)

1.00

-

0.82

0.71

1.00

-

WTP flight time (AUD/hr)

78.40

41.19

119.05

stopover time (AUD/hr)

57.49

28.03

30.86

stops (AUD/stop)

-48.78

-51.72

96.83

personal TV screens (AUD)

147.39

117.55

220.32

video on demand (AUD)

227.18

189.97

218.41

pitch (AUD/inch)

81.88

76.18

76.51

seat reservation (AUD)

81.53

74.29

180.00

Table 8: Results for models 4-6

Model

7

Observations Final LL

8

9

266

417

417

-199.36

-401.93

-401.53

adj. ρ2

0.239 estimate

0.285

rob. p-val.

estimate

0.285

rob. p-val.

estimate

rob. p-val.

δ(airline 1)

0

-

0

-

0

-

δ(airline 2)

0.0798

0.83

0.1920

0.46

0.2020

0.43

δ(airline 3)

-0.7660

0.10

-0.6970

0.09

-0.7560

0.07

δ(airline 4)

-0.3150

0.48

0.0870

0.81

0.0800

0.82

δ(airline 5)

0.5620

0.19

0.6300

0.03

0.6690

0.03

δ(airline 6)

0.2580

0.55

0.2740

0.46

0.3040

0.42

δ(airline 7)

-0.3430

0.45

-0.0633

0.87

-0.0626

0.88

δ(airline 8)

0

-

0.1100

0.80

0.1480

0.75

δ(airline 10)

0.0199

0.96

0.1030

0.82

0.0624

0.89

δ(airline 11)

0

-

0

-

0

-

δ(airline 12)

-0.7930

0.06

-0.9360

0.01

-1.0500

0.02

δ(airline 13)

0

-

-0.4230

0.70

-0.3560

0.76

δ(airline 14)

0

-

0.8230

0.43

0.8590

0.44

δ(airline 15)

0

-

0.2670

0.71

0.1730

0.82

δ(airline 16)

0

-

-1.0200

0.48

-0.9520

0.47

δ(slot 1)

0

-

0

-

0

-

δ(slot 2)

-0.3500

0.07

-0.3880

0.02

-0.4360

0.02

δ(slot 3)

-0.6090

0.00

-0.5810

0.00

-0.6430

0.00

δ(slot 4)

0

-

-0.9470

0.01

-1.0200

0.01

δ(slot 5)

0

-

-1.3300

0.01

-1.3700

0.01

δ(slot 6)

0

-

-1.6900

0.01

-1.8700

0.01

δ(slot 7)

0

-

-2.9900

0.01

-3.1400

0.01

δ(slot 8)

0

-

-2.6500

0.01

-2.7900

0.01

δ(slot 9)

0

-

-2.4800

0.01

-2.7800

0.02

δ(slot 10+)

0

-

-2.8500

0.00

-3.1700

0.00

β(air fare)

-0.0027

0.00

-0.0028

0.00

-0.0029

0.00 0.46

β(carbon tax)

0.0009

0.32

0.0005

0.45

0.0005

β(fee for changes)

-0.0004

0.63

-0.0007

0.27

-0.0008

0.23

β(flight time)

-0.0015

0.02

-0.0018

0.00

-0.0019

0.00

β(stopover time)

-0.0005

0.79

-0.0011

0.22

-0.0011

0.27

β(stops)

0.3050

0.14

0.1720

0.32

0.1820

0.33

0

-

0

-

0

-

δ(overhead TV screens) δ(personal TV screens)

0.2040

0.49

0.3280

0.10

0.3530

0.10

δ(video on demand)

0.5030

0.09

0.5590

0.00

0.5960

0.00

β(pitch)

0.2340

0.00

0.2110

0.00

0.2290

0.00

δ(seat reservation)

0.1760

0.34

0.1740

0.24

0.1850

0.23

δ(747)

0

-

0

-

0

-

δ(777)

0.0223

0.94

0.1400

0.48

0.1590

0.45

δ(A330)

0.4520

0.06

0.4580

0.03

0.4830

0.03

δ(A340)

0.2820

0.36

0.2860

0.29

0.2960

0.25

α(UK departure)

1.00

-

1.00

-

1.00

-

α(CH departure)

1.00

-

0.94

0.91

1.00

-

α(search tasks first)

1.00

-

1.00

-

1.00

-

α(SP tasks first)

1.00

-

1.00

-

0.87

0.55

WTP flight time (AUD/hr)

33.78

39.05

37.88

stopover time (AUD/hr)

10.07

24.87

22.32

-112.96

-62.55

-62.12

stops (AUD/stop) personal TV screens (AUD)

75.56

119.27

120.48

video on demand (AUD)

186.30

203.27

203.41

pitch (AUD/inch)

86.67

76.73

78.16

seat reservation (AUD)

65.19

63.27

63.14

Table 9: Results for models 7-9

Model

10

Observations Final LL

11

12

194

223

417

-186.50

-221.91

-390.98

adj. ρ2

0.200 estimate

0.254

rob. p-val.

estimate

0.299

rob. p-val.

estimate

rob. p-val.

δ(airline 1)

0

-

0

-

0

-

δ(airline 2)

0.1190

0.75

0.5340

0.09

0.0255

0.90

δ(airline 3)

-1.2200

0.03

-0.1640

0.74

-0.6670

0.29

δ(airline 4)

-0.4130

0.40

0.4920

0.31

-0.1250

0.75

δ(airline 5)

0.6210

0.12

1.0300

0.01

0.3130

0.48

δ(airline 6)

0.4050

0.46

0.3960

0.48

0.0726

0.83

δ(airline 7)

0.0466

0.94

0.3700

0.44

-0.1730

0.61

δ(airline 8)

0.0601

0.92

0.3270

0.62

-0.0575

0.89

δ(airline 10)

-0.8210

0.24

1.1500

0.04

-0.1990

0.54

δ(airline 12)

-1.2100

0.03

-0.3030

0.58

-0.7390

0.32

δ(airline 13)

-0.5510

0.62

0

-

-0.8380

0.44

δ(airline 14)

0.3750

0.80

1.2400

0.37

0.8350

0.39

δ(airline 15)

-0.7490

0.36

1.4400

0.25

0.1260

0.84

δ(airline 16)

-0.7870

0.52

0

-

-1.2500

0.40

δ(slot 1)

0

-

0

-

0

-

δ(slot 2)

-0.5460

0.02

0.1480

0.51

-0.2170

0.45

δ(slot 3)

-0.7550

0.00

-0.0077

0.98

-0.4080

0.36

δ(slot 4)

-0.3220

0.48

-0.4360

0.39

-0.6340

0.38

δ(slot 5)

0.1730

0.74

-1.9000

0.07

-0.7440

0.39

δ(slot 6)

0

-

-0.2400

0.66

-1.0900

0.42

δ(slot 7)

-1.5000

0.19

0

-

-2.4100

0.25

δ(slot 8)

-0.9780

0.27

0

-

-2.1600

0.10

δ(slot 9)

0

-

-1.0500

0.25

-1.9800

0.11

δ(slot 10+)

0

-

-1.3800

0.02

-2.3500

0.03

β(air fare)

-0.0033

0.00

-0.0034

0.00

-0.0021

0.17

β(carbon tax)

-0.0005

0.62

0.0004

0.70

0.0009

0.10

β(fee for changes)

-0.0025

0.02

-0.0005

0.61

-0.0001

0.89

β(flight time)

-0.0016

0.04

-0.0029

0.00

-0.0013

0.30

β(stopover time)

-0.0007

0.53

-0.0023

0.07

-0.0008

0.53

β(stops)

0.0384

0.88

0.1470

0.53

0.1120

0.53

δ(overhead TV screens)

0

-

0

-

0

-

δ(personal TV screens)

0.4000

0.18

0.5480

0.04

0.2410

0.24

δ(video on demand)

0.5120

0.05

0.9580

0.00

0.4360

0.20

β(pitch)

0.2570

0.00

0.1660

0.06

0.1580

0.13

δ(seat reservation)

0.0989

0.64

-0.0526

0.81

0.1320

0.43

δ(747)

0

-

0

-

0

-

δ(777)

0.3940

0.15

0.1680

0.56

0.0043

0.98

δ(A330)

0.4050

0.17

0.6450

0.03

0.2910

0.47

δ(A340)

0.4720

0.19

0.4860

0.16

0.1960

0.57

α(CTS)

1.00

-

1.00

-

1.00

-

α(IVT)

1.00

-

1.00

-

1.00

1.00

α(ITS)

1.00

-

1.00

-

1.87

0.69

α(other)

1.00

-

1.00

-

3.27

0.67

WTP flight time (AUD/hr)

29.63

51.25

37.14

stopover time (AUD/hr)

12.75

40.06

22.00

stops (AUD/stop)

-11.78

-42.86

-53.33

personal TV screens (AUD)

122.70

159.77

114.76

video on demand (AUD)

157.06

279.30

207.62

pitch (AUD/inch)

78.83

48.40

75.24

seat reservation (AUD)

30.34

-15.34

62.86

Table 10: Results for models 10-12

6. Conclusions In this paper, we have presented the results from a pilot study making use of an innovative SP design that is modelled on typical online air travel booking engines. Despite the very small sample size, our analysis has shown that respondents are able to cope with the amount of information presented to them in this survey. Furthermore, there is evidence that respondents make extensive use of the searching and sorting tools available to them when completing the individual choice tasks. Initial estimation results suggest some differences in how respondents react to the choice situations in the two types of choice situations they are presented with (traditional SP/search & sort). More work is required to quantify these differences. Along with the collection of a larger sample, several other avenues for further research exist: • Use of models allowing for taste heterogeneity • Use of choice set consideration models for data from search & sort games • Making use of information on attribute ignoring

Acknowledgments The authors would like to thank the volunteers at Imperial College London, ETH Zurich and the University of Leeds for taking time to complete the survey. We would also like to thank Jun Zhang for all his work in coding the internet survey.

References Algers, S. & Beser, M. (2001), ‘Modelling choice of flight and booking class - a study using stated preference and revealed preference data’, International Journal of Services Technology and Management, 2(1-2), 28–45. Bates, J., Ashley, D. & Hyman, G. (1987), ‘The Nested Incremental Logit Model: Theory and Application to Mode Choice’, Proceedings of the PTRC summer annual meeting, Volume 290, PTRC, London. Bierlaire, M. (2003) BIOGEME: a free package for the estimation of discrete choice models, Proceedings of the 3rd Swiss Transport Research Conference, Monte Veritá, Ascona. Bradley, M. A. (1998), ‘Behavioural models of airport choice and air route choice’, in J. de D. Ort´uzar, D. Hensher & S. R. Jara-Diaz, eds, ‘Travel behaviour research: updating the state of play (IATBR 94)’, Elsevier, Oxford, pp. 141–159. Burke, R.R., Harlam, B.A., Kahn, B.E. and Lodish, L.M. (1992) ‘Comparing Dynamic Consumer Choice in Real and Computer-Simulated Environments’ Journal of Consumer Research, 19 (June), 71–82. Carson, R., Louviere, J.J., Anderson, D., Arabie, P., Bunch, D., Hensher, D.A, Johnson, R., Kuhfeld, W., Steinberg, D., Swait, J., Timmermans, H., and Wiley, J. (1994) ‘Experimental Analysis of Choice’, Marketing Letters, 5 (October), 351-367

Chin, A. T. H. (2002), ‘Impact of frequent flyer programs on the demand for air travel’, Journal of Air Transportation, 7(2), 53–86. Greene, W. H. (2007) Nlogit 4.0, Econometric Software, New York and Sydney. Hensher, D., Stopher, P. R. & Louviere, J. (2001), ‘An exploratory analysis of the effect of numbers of choice sets in designed choice experiments: an airline choice application’, Journal of Air Transport Management, 7(6), 373–379. Hess, S. (2007) Posterior analysis of random taste coefficients in air travel choice behaviour modelling, Journal of Air Transport Management, paper accepted for publication (February 2007). Hess, S., T. Adler and J. W. Polak (2007) Modelling airport and airline choice behaviour with stated-preference survey data, Transportation Research Part E: Logistics and Transportation Review, 43, 221–233. Kanafani, A. & Sadoulet, E. (1977), ‘The partitioning of long haul air traffic – a study in multinomial choice’, Transportation Research, 11(1), 1–8. Proussaloglou, K. & Koppelman, F. S. (1995), ‘Air carrier demand: an analysis of market share determinants’, Transportation, 22(4), 371–388. Proussaloglou, K. & Koppelman, F. S. (1999), ‘The choice of air carrier, flight, and fare class’, Journal of Air Transport Management, 5(4), 193–201. Rose, J.M. and Hensher, D.A. (2006), ‘Accounting for individual specific nonavailability 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. Wardman, M. (2001), ‘A review of British evidence on time and service quality Valuations’, Transportation Research E, 37, 91-106.