Identifying commuter preferences for existing modes and a proposed Metro in Sydney, Australia with special reference to crowding

David A. Hensher · John M. Rose ·Andrew T. Collins

Published online: 1 February 2011

8  The modelling results

In this section we present the final error component logit model, arrived at after extensive investigation of alternative specifications.9

An important point to be aware of when evaluating the findings is that the catchment area was defined to ensure that the metro was able to be evaluated as a physically available alternative, with a time-cost profile that was deemed sensible in satisfying in-scope eligibility. Hence the sample has selectivity bias in that the choices are conditional on metro being deemed a considered alternative in the choice set. In contrast, all other means of transport (apart from the mode actually chosen in the recently sampled trip for an in-scope interview), may or may not be in the choice set, depending of the responses of respondents. Consequently it will come as no surprise that the current modal share is not a reproduction of the wider modal share of the Sydney metropolitan area or of those who live and travel from a location within the catchment area.

The pre-defined catchment was assumed to be the locality where eligible trips would commence; however many trips in the catchment area would not have a destination that would allow the Metro to be a candidate alternative; hence the use of population data such as that provided by the NSW Transport Data Centre for exogenous weighting turned out to be of little value since the data assumes that everyone who is a ‘Residents of occupied private dwellings in the CATCHMENT’ (defined by a list of postcodes in Fig. 4) was a candidate potential users of Metro.10 Given that it is unlikely that one would be able to extract the relevant sub-sample from the HTS survey (given the physical catchment area), we have chosen to focus on the unweighted models, which provide parameter estimates that are a true reflection of the preferences of the sampled respondents for whom the Metro is a genuinely relevant alternative.

A descriptive profile of the data used in model estimation is given in Table 10.

The final model is presented in Table 11.

This was selected from a large number of models. The overall goodness-of-fit on accepted criteria is impressive (rho2 of 0.6075). Some parameter estimates are alternative-specific (e.g., car costs and times), and some are generic to a subset of modes (e.g., train and metro or all public transport modes). There are no fully generic parameter estimates (i.e., across all modes), suggesting that there are mode-specific sensitivities on many influencing attributes. For example, the mean parameter estimates (or marginal disutilities) for public transport fares are −0.4345 for bus, and −0.3994 for train and metro, and all are statistically significant (lowest t-value is −5.21).

Taking a closer look at each explanatory variable within class, we see that for access, park or kiss and ride time is the most statistically significant influence on main mode choice for train and metro, followed by the use of a public transport mode (in terms of both time and cost). Note that for the main public mode, it is feasible to access it by either bus, train or the proposed metro. Walk time to a public transport mode was marginally significant. The access public mode cost is also statistically significant, so overall the access trip does have a strong role to play in establishing the preference for public transport. There is clear evidence of a high degree of sensitivity to travel time to drive to, and park near or be dropped off (kiss n ride), a train or proposed metro station, as well as using public transport to access the train or metro.

Thus station spacing can have a significant influence on the role of this attribute, given the marginal disutility weight.

The crowding attributes associated with public transport are statistically significant.

The number standing is defined as a quadratic interacted with in-vehicle travel time for the main mode, and the proportion seated is defined by the natural logarithm transformation interacted with the main mode in-vehicle travel time. The nonlinearity is intuitively plausible and is very useful in establishing the willingness to pay to get a seat and to stand under various loading scenarios by time of day (or time slice during a particular period such as the peak period) and trip length. The willingness to pay to obtain a seat or to avoid standing is set out in the next section for a range of scenarios. The (dis)utility function for the number standing and the proportion seated are respectively:

Of particular note is that the mode-specific constants for all public transport modes are negative (relative to the car-specific constant set to 0.0) and statistically significant.  What this suggests is that, after accounting for the differential and correlated variances associated with the unobserved influences associated with each alternative, as well as the rich specification of factors that really do matter to travellers who are in-scope, that there remain significant unobserved influencing effects on average. The metro has the least negative mode-specific constant (with bus the most negative) indicating its preference over train and bus after accounting for service and cost levels.

In application one can construct a modified metro-specific constant from all those attributes in the model that are statistically significant that the analysts may choose not to incorporate in their application model. One would have to define a specific (fixed) level for each of these attributes, and use that to obtain the additive metro-specific constant.  We have introduced dummy variables to represent the mode currently being used.

These variables are measures of inertia and suggest, all other factors remaining unchanged, that an individual is more likely to choose the alternative they currently use.  Since many trips are essentially habitual (in contrast to variety seeking), this is an important conditioning effect. The dummy variables for the revealed preference (RP) modes car and bus are statistically significant, but not for train. This might suggest that the proposed metro will appeal mostly to existing train users. The strongest RP inertia effect is for car, with a parameter estimate of 4.7780, in comparison to the bus estimate of 1.1163.

We see a very statistically significant public transport frequency attribute (an impressive t -ratio of −6.47). Frequency is defined as a quadratic of minutes between services, allowing for the marginal disutility to vary by headway. The negative parameter estimate indicates all other influences remaining constant, that an increased frequency associated with reduced time between services, will reduce the marginal disutility of public transport, and increase the probability of choosing public transport.

The willingness to pay for increased frequency is statistically significant and non-marginal.  Finally, the presence of a transfer within public transport is negative as expected, and statistically significant for bus, and marginally significant for train and metro.  Thus relative to metro, the transfer penalty is lower for train and higher for bus.

10  Conclusions

The study has focussed on obtaining new empirical evidence on the factors influencing the choice of mode of transport for commuting trips in Sydney, and especially the contribution to the relatively small literature on valuation of crowding.

The primary focus is on establishing the role of a number of trip attributes in the definition of the modal preference expression for the proposed metro in the presence of existing modes of transport for access, main and egress stages of door-to-door travel activity.

The evidence on the relative marginal disutility (or utility) associated with the influencing attributes, essentially components of service level (time, frequency, transfers), costs and crowding as well as the metro-specific constant that captures served influences associated with the metro option, enable us to derive estimates of willingness to pay for each component.

The application modelling system can import the various Willingness to Pay estimates and use them to build up the empirical expressions of relative utility associated with the competing modes of transport. This can be achieved in a generalised time or cost specification.

A number of comments are appropriate as a way of reinforcing some important assumptions that underlie (or condition) the behavioural research herein:

1. The estimated model has been developed in a context where the metro is deemed to be a ‘feasible’ alternative. Essentially, the catchment origin-destination set has been premised on the metro being included in each respondent’s choice set as if they were asked, in contrast to imposing it on them.

2. The application system within which the model can be applied must recognise that the parameter estimates and associated Willingness to Pay (WTP) estimates are applicable to a specific context in Sydney, and as such, their incorporation in a Sydney-wide modelling system must recognise this, and be prepared to assume that the parameters are transferable to the wider context.

3. On the reasonable assumption that the key outputs are portable to the application model system, the only constant that can be ported is the metro-specific constant. The numerical estimates for all segments are intuitively plausible relative to the car, train and bus and so we can be somewhat confident that we have identified the relative magnitude of the mean unobserved influences associated with the metro.

4. All modelling was undertaken as if the entire metro network will be in place. Consequently any applications that focus on a staged introduction must be mindful of the context in which all parameter estimates are derived.