Menu
Home
Log in / Register
 
Home arrow Economics arrow Bounded rational choice behaviour
< Prev   CONTENTS   Next >

4.3. Data and Model Specifications

The empirical setting for this chapter is a stated choice study of commuters, in which several generic route alternatives are described by travel time uncertainty, and one or more time and cost attributes. Crucially, the study systematically varied the information load across respondents, along the following dimensions:

• Number of choice tasks – 6, 9, 12 or 15 tasks;

• Number of alternatives – 3, 4 or 5, inclusive of the recent trip;

• Number of attribute levels per attribute – 2, 3 or 4;

• The range of the attribute levels – a base range, narrower than this base, and wider than the base; and

• Number of attributes – 3, 4, 5 or 6.

Sixteen sub-designs were generated, each with a unique combination of the above dimensions. Table 4.3, presented later in this chapter along with some of the model results, details these combinations. Since the alternatives were presented as being generic, the numbers of alternatives could be varied without inducing obvious alternative specific effects.[1] The number of attributes was varied by providing varying degrees of decomposition of the travel time and cost attributes. Table 4.1 summarises this decomposition, as well as which attributes were presented for any given number of attributes shown. A full decomposition of time yielded time in three conditions: free flow traffic, slowed down traffic and stop/start traffic. In some sub-designs, slowed down and stop/start time were aggregated to form congestion time, and in other sub-designs, all time components were aggregated to form total time. Cost was either presented as a total cost, or decomposed into toll and running cost attributes. An uncertainty in travel time attribute was framed as the possible variation in minutes around the total travel time.

Attribute levels were pivoted around a recent trip, with the specific shifts from the recent trip levels being a function of the sub-design, and its associated number of and range of attribute levels. These shifts are documented in Table ΑΙI' of Hensher (2006).[2] A d-efficient experimental design (see Rose, Bliemer, Hensher, & Collins, 2008) was employed in the study, to maximise the statistical efficiency of the data collection effort. A typical stated choice screen from the study is presented in Figure 4.1, wherein the choice task has four alternatives and six attributes.

The survey was administered as a computer-aided personal interview (CAPI) in the Sydney metropolitan area in 2002. A stratified random sample was applied, based on the residential location of the household. Since this chapter uses the same dataset as Hensher (2006), and has some similar objectives (albeit with different methodology),[3] it is useful to highlight some of the key differences in data usage and model specification. One such difference is the number of observations retained. Whereas Hensher eliminated sub-designs with only three attributes, all sub-designs are retained in the present analysis, resulting in 4593 observations over 419 respondents.

To account for preference heterogeneity across respondents, all models presented will specify random parameters for travel time uncertainty, and all of the time attributes. Use of the RPANA model with continuous attributes necessitates the use of a distribution that is constrained in sign (Collins et al., 2013). The lognormal distribution was chosen over the constrained triangular distribution, as it led to better model fit, and with two structural parameters instead of just one, allows a more flexible distribution to be estimated. In particular, this chapter will test for the influence of ANA on the mean, median, mode and standard deviation of the distribution. The sign of the time attributes was reversed to ensure identification with the lognormal distributions. In all models, 400 Halton draws were employed, which was found to be sufficient to ensure stability of the parameter estimates. To simplify the

Table 4.1: Composition of attributes, and distribution across sub-designs.

Attribute

Attribute composed of

Present in sub-designs with these number of attributes

Free flow time

Slowed down time

Stop/start

time

Running

cost

Toll

cost

3

4

5

6

Total time

X

X

X

X

Free flow time

X

X

X

X

Slowed down time

X

X

X

Stop/start time

X

X

X

Congestion time

X

X

X

Uncertainty in

X

X

X

X

travel time

Total cost

X

X

X

X

X

Running cost

X

X

Toll cost

X

X

A typical state choice screen from the study.

Figure 4.1: A typical state choice screen from the study.

calculation of willingness to pay (WTP) measures, fixed coefficients were used for all cost parameters.

The key motive of this chapter is to investigate the impact of information load on inferred ANA. Whilst any of the presented attributes could potentially be left unattended to by respondents, only non-attendance to uncertainty and free flow time will be modelled. This is because free flow time is present in 12 of the 16 subdesigns, and uncertainty is present in all of them. This provides more potential to find statistically significant influences of information load on non-attendance to these attributes. The RPANA model is specified such that it relies on the assumption of independence of non-attendance across attributes, as this avoids problems for those sub-designs in which uncertainty is present but free flow time is not.

In addition to the base RPL model (Model 1), three models that accommodate ANA will be estimated. Model 2 is an RPL model that accommodates stated non- attendance. Where respondents indicate that they ignored an attribute, the coefficient for that attribute is constrained to zero. Model 3 infers non-attendance with the RPANA model, although the influence of information load is not modelled. Against these baselines, Model 4 infers non-attendance whilst allowing measures of information load to moderate the inferred non-attendance rate, by entering them as zn in the ANA component of the RPANA model, and estimating the associated parameters Θ. The next section will detail the results from these four models, with a particular focus on the inferred non-attendance rates, their link with the systematically varied information load, and the implications of this link. The impact on WTP measures is also investigated.

  • [1] A left-right bias is possible, but no such bias was identified.
  • [2] The interested reader is also referred to other papers by Hensher and colleagues who have used this dataset, including Hensher (2004) and Hensher, Rose, and Greene (2005).
  • [3] The same dataset has also been used in other papers such as Hensher et al. (2005). A cross cultural study of the same survey instrument and experimental design has been investigated in Rose, Hensher, Caussade, Ortuzar, and Jou (2009).
 
Found a mistake? Please highlight the word and press Shift + Enter  
< Prev   CONTENTS   Next >
 
Subjects
Accounting
Business & Finance
Communication
Computer Science
Economics
Education
Engineering
Environment
Geography
Health
History
Language & Literature
Law
Management
Marketing
Philosophy
Political science
Psychology
Religion
Sociology
Travel