6.2. Results and Discussion
Several model specifications were initially developed and compared against the proposed LCCM framework using the pooled dataset.
Table 5 shows a comparison of the goodness-of-fit statistics for the alternative model specifications. The models 1 to 7 used in
Table 5 are described as follows:
Model 1: MNL estimating and
Model 2: MNL estimating , and where is only multiplied with *
Model 3: MNL estimating , and where is multiplied with
Model 4: Extending Model 1 by considering lognormal distribution for in-vehicle travel time
Model 5: Extending Model 2 by considering lognormal distribution for in-vehicle travel time
Model 6: Extending Model 3 by considering lognormal distribution for in-vehicle travel time
Models 1–3 apply an MNL that assumes invariant parameters across the population. Models 4–6 model taste-variation through in-vehicle travel time by assuming a log-normal distribution. Finally, model 7 is the LCCM. The idea behind developing these seven models is to compare the model performance of LCCM against the other two, which further supports the selection of LCCM for this study. The models are developed and estimated using the Pythonbiogeme software package [
51]. Model 1 shows the MNL results on the pooled dataset, which assumes equal unobserved variance (i.e., fixing
) for auto and transit sub-samples. The model gives poor model fi statistics (e.g., adjusted rho squared, AIC and BIC values). Models 2 and 3, which estimate the scale parameter (
) show: (1) a significant effect size for
, and (2) improved model fit statistics when compared to model 1. This indicates that the auto sub-sample (comprising 130 participants or 910 observations) has a statistically different unobserved variance when compared to the transit sub-sample (46 participants or 322 observations), which eventually shows up in different effect sizes for
and
across the two sub-samples [
50]. Similar observation is made when comparing models 4–6 where the goodness of fit improves when
is estimated. Furthermore, models 4–6 perform better than models 1–3 indicating that randomising the parameter (using 1000 standard Halton draws) for in-vehicle travel time further improves the model fit. Model 7, which also estimates class specific
along with
and
, has the highest adjusted rho-squared and the least AIC and BIC values indicating the best model fit among the other models reported in
Table 5. Hence, the results from the LCCM (Model 7) are reported and discussed in this section.
For the class membership model, socio-demographics are included as dichotomised variables, which represent: (1) Females (1: females; 0: males), (2) low-income individuals (1: Income less than
$25 K; 0: Otherwise) [
40], (3) work trips (1: Trip purpose is work; 0: Otherwise), and (4) short-distance trips (1:Trip distance is less than 30 km; 0: Otherwise). The value 30 km is selected as it represents the average trip length revealed by the participants (refer to
Table 4). The choice model comprises trip-related attributes (in-vehicle travel time, travel cost, access time, etc.) along with the error component term (σ), which is simulated using 1000 standard Halton draws during the maximum likelihood estimation (Equation (8)).
Table 6 presents the parameter estimates for the two-segment LCCM using the pooled dataset. Several other LCCMs are also tested by trying different combinations of variables (socio-demographics and trip attributes) and the model has been chosen based on the following criteria: (1) Higher goodness of fit statistics (shown in
Table 6), and (2) meaningful parameter interpretation of the trip specific attributes. Furthermore, additional LCCMs are developed by increasing the number of latent segments to three. However, the chosen model is preferred over three-segment LCCMs as it has: (1) A lower BIC value (2-LCCM: 1170.0 vs. 3-LCCM: 1193.5), and (2) model parsimony (2-LCCM: 16 parameters against 3-LCCM: 25 parameters).
Table 6 shows that the participants belonging to class 2 have a negative (and highly significant) effect towards in-vehicle travel time (−2.59), travel cost (−3.79), and access time (−5.19) attributes. In other words, participants belonging to this segment experience increased disutility (hence, lesser probability of selecting the two alternatives) with every unit increase in the trip-related attributes. While the parameter for egress time bears a negative sign (−0.166), which again corresponds to disutility, it is found to be statistically insignificant. On the other hand, the participants in class 1 have insignificant parameters for in-vehicle travel time, travel cost, and access time, indicating that they are indifferent towards them, and are sensitive towards egress time, which can be seen through a negative and significant parameter value (−0.27). Thus,
Table 6 shows contrasting tastes differences towards trip-related attributes between the two identified segments. The ASC parameter for class 1 is negative (−0.557) and significant indicating that the participants in this segment have a lower brand effect towards the SQ alternative (auto or transit) when compared to the new service. On the other hand, the ASC parameter for class 2 (2.8) is statistically insignificant, which means that the participants in this segment do not have any affinity towards either mode.
The scale parameter for segment 2 is highly significant (0.414). As discussed earlier in
Section 6.1.1,
. This indicates a greater unobserved variance in the auto sub-sample when compared to the transit sub-sample. Additionally, this scale parameter leads to two sets of parameters, one each for auto and transit as the SQ alternative. For example, the parameter for in-vehicle travel time for transit and auto is −2.59 and 0.414*(−2.59) = −1.07, respectively. Similarly, the parameters for cost (transit: −3.79; auto: −1.57) and access time (transit: −5.19; auto: −2.15) can be obtained. It can be observed that the participants having auto as the SQ alternative have lower effect sizes when compared to the participants with transit SQ alternative. This finding can be explained as follows: Car travel is generally more flexible when compared to transit. For example, the time to access car park is usually much lower when compared to bus stops and train stations. Additionally, (1) travel is typically not quite onerous in a car in contrast to transit (which often experiences standing and over-crowded conditions), and (2) car users generally have more disposable income and are thus not very sensitive towards price. Due to these reasons, car users are usually not greatly perturbed with a unit increase in the attributes. The scale parameter for segment 1 (0.203) is found to be insignificant, which indicates that there is no statistical difference between the unobserved variances. Similarly, the error variance parameter (σ = −6.46) is significant for segment 2, which indicates the presence of the unobserved correlation across multiple-choice scenarios in the SP experiment. However, this parameter turns out to be insignificant in segment 1.
Table 6 also shows the average market uptake for the new service in class 1 and 2 as 96 and 44 percent, respectively. In other words, the participants belonging to class 1 are more likely to use the new service over cars when compared to class 2.
The results from the class membership model show that individuals commuting to/from work are more likely to belong to class 1 in comparison to class 2 (fixed as the base category). Thus, it can be said that the people who make work-related trips are more likely to be in class 1, which represents the user segment characterised by a high uptake (96 percent) for the new service. This finding can be explained as follows: As discussed earlier in
Section 5, around half of the participants have a trip length of at least 30 km, which is the distance to the areas around Sydney CBD. An express bus service, called the B-line service, operates at frequent intervals connecting Northern Beaches to the Sydney CBD. The proposed DRT service aims to supplement the B-Line service by providing easy and direct access to the latter while maintaining similar level of comfort and convenience as offered by private car. Thus, the new service provides a conducive mode allowing commuters to do other activities such as reading, responding to emails, etc., while on DRT or B-line bus. This leads to reduced (or no) disutility towards in-vehicle travel time, cost, and access time (since DRT facilitates close to home pickup), which justifies the insignificant parameters for these three attributes. As the DRT service is not currently available around the Sydney CBD, reaching the destination (workplace) from the bus stop is rather challenging which could explain the negative and significant parameter for egress time. On the other hand, class 2 is more likely to be composed of people who travel for non-work purpose (social, shopping, etc.) and are less inclined to use the new service. This observation can be interpreted as follows:
Table 3 shows that more than 60 percent of the trips are made due to non-work purposes, and 70 percent of the non-work trips are made using cars. The household travel survey report of Sydney also brings out a similar finding where cars account for around 65 percent of the social/recreational trips with transit contributing only 6 percent [
52]. In other words, as cars offer greater flexibility, these are the preferred mode of travel for the participants belonging to segment 2 with only 44 percent uptake for the new service.
The value of travel time savings (VTTS) for the individuals under this segment is calculated as AU
$13.7/hr, which is similar to the one (AU
$15.5/hr) found by Saxena et al. (2018) [
47]. The obtained VTTS value implies that the participants in this segment are willing to pay additional AU
$13.7 to reduce the travel time by an hour. This indicates travel disutility for individuals in general, which makes sense and is consistent with previous literature. A similar value, AU
$17.3/hr, is also obtained from Model 3 (where in-vehicle travel time, cost, and scale are significant), the results of which are presented in the
Appendix A. Using the income distribution shown earlier in
Table 3, the weighted average salary is calculated around AU
$75 K (assuming the median of income range). Assuming work duration of 35 h/week and 52 working weeks, in general, the hourly wage rate of the participants comes to be AU
$41 on average. This value is much larger than the obtained VTTS indicating meaningful interpretation of the results.