Explaining the surge: Using segmentation and choice modelling to make sense of the huge growth in public transport patronage

Introduction

For the Victorian Government’s Department of Transport (herein DOT), accurate forecasts of future public transport patronage are critical in forward planning, particularly because investments in rolling stock and infrastructure can be of considerable cost and have long lead times in project delivery. As a result, DOT has typically developed long-term patronage forecasts using strategic demand models.

In 2004/05, DOT’s official patronage forecasts based on the strategic models for public transport (PT) were 3.6% (metropolitan train), 2.7% (tram) and 4.0% (metropolitan bus). This gave an overall annual growth forecast of approximately 3.3% for metropolitan Melbourne.

However, the subsequent four years saw an unprecedented and unpredicted surge in public transport patronage, with year-on-year overall metropolitan public transport growth figures of 4.9%, 7.9%, 7.7% and 9.0% respectively. By 2008/09 Melbourne’s total patronage was at 491.5m trips per year – a figure that the strategic models had not predicted to be reached until 2012/13.

The growth was particularly evident on Melbourne’s metropolitan train network, which had patronage growth of 43% between 2005 and 2009. However, annual growth rates on both metropolitan trams and buses reached 12.5% (trams) and 9.0% (buses) in 2008/09 – a point by which time extreme growth on trains had began to abate.

The surge in public transport usage – seen both in peak and off-peak periods – placed considerable pressure on train services, particularly during peak periods. The tram network also experienced increasing crowding problems during peak periods at this time. (Department of Transport, 2009).

Figure 1. Melbourne’s Public Transport Annual Growth Rates: 1990/91 to 2008/09

Figure 1

By 2008, Melbourne’s already stretched public transport system was balanced on a knife’s edge; while the DOT’s imperative was to increase public transport patronage, any improvements to the system, and/or factors that drove travellers away from private transport could at any moment cause a massive inflow of patronage which in turn could create excess demand. On the flipside, desired demand for PT could easily fall away if service improvements could not keep pace with this demand, which would ultimate lead to declines in operational performance.

DOT put forward a number of hypotheses about why its models had failed to predict the surge in PT usage. Initially, these included:

  • Inaccuracy in terms of model inputs. Strategic models relied on multiple inputs including population growth, land use and petrol price forecasts that proved not to be accurate.
  • Outdated travel data used to calibrate the models. Strategic models rely on travel survey data to calibrate the models. In 2004/05, these models relied on data from the Victorian Activity and Travel Survey (VATS) which was largely undertaken in the 1990s. The most recent home interview travel survey (VISTA, 2007) can now be used to calibrate new models.

However, there was also a hypothesis within DOT that other, less tangible and potentially less readily quantifiable, changes in the transport market could also have been at play during this period.

  • For example, had increasing congestion on the road network reached a ‘tipping point’ for users?
  • Had service improvements on public transport (or the contrast between service improvements and road conditions) made public transport relatively more attractive than other modes?

As well as factors that are intangible yet relatively within the realm of control of DOT, it also hypothesised that broader attitudinal changes within the community could also have been playing a part in driving the lack of predictive accuracy in its models. For example, had an increased awareness of environmental issues and increased societal focus on personal health, wellbeing and fitness played a role in PT patronage surge and a possible migration toward modes such as walking or cycling? Further, and independent of these factors, to what degree were increased PT usage rates a function of possible increases in the degree to which consumers were deriving utility from PT usage – for example to have ‘down time’, the opportunity to read, or communicate with others by virtue of mobile handheld devices?

In simple terms, DOT formed a hypothesis that societal attitudinal and preferential shifts may have been playing an important part in influencing system-level PT usage patterns, and as such, explained why their forecasting models could not account for the massive surge in PT described earlier. In an ideal world, DOT would have been able to test this and related hypotheses using its own demand models, but the very nature of the factors hypothesised to be at play (that is, at times intangible factors, which are traits of people rather than the PT system) meant that they were unable to shed light on the matter in this fashion. This fact led to the need for research to form input on the matter of the degree to which attitudinal shifts played a part in explaining the surge in PT usage. As well making sense of past patterns, the research was also viewed as playing an important part in helping DOT more accurately forecast future shifts in PT patronage. In 2009, Metlink commissioned Nature to embark on a major research study with these goals in mind.

The Research Approach

In developing the approach to tackling this research problem, what became immediately apparent was that along with a distinction between hard-edged tangible and softer-edged intangible factors, some of the variables that were hypothesised to influence choice of transport mode were traits of the transport mode in question, while others were traits of the people involved. This had a direct and fundamental bearing on how the research design was approached. Examples of traits ‘owned’ by modes include:

  • Ticket price
  • Door-to-door transport time
  • Petrol, parking, and tolls cost
  • Probability of service disruption
  • Traffic congestion
  • Passenger Density

Examples of variables which were in fact traits of the people involved including:

  • Perceived utility of public transport
  • The degree to which one was ‘rusted on’ to a particular mode of transport
  • Attitude towards environmental issues / environmental consciousness
  • Sensitivity to value for money
  • Attitudes to health and wellbeing

This split between these mode-specific variables and traits of individuals posed a particular problem in terms of research design. While a well crafted piece of quantitative research could in principal determine how an individual’s behaviour would change in the face of an increase in ticket price, or the availability of a seat on their preferred train service, the matter of accurately predicting the impact of attitudinal change on behaviour is much more complex. Determining this matter through a classic ‘direct question’ approach would lack validity, and inferring it through choice modelling in which individually-owned traits (e.g. attitudes) are included as model attributes would be nonsensical given that such features are both intangible, and outside the direct control of any organisation. It would make no sense to include an attitudinal attribute in a choice model given attitudes are the traits of individuals. There are also several reasons why one would not wish to test the impact of an attribute such as ‘environmental impact of train’ on mode choice behaviour. A key challenge for this project was therefore to design a methodology that would enable DOT to understand without design-based bias the degree of influence on PT patronage of shifts in tangible mode-specific attributes, and changes in attitudes. To address the challenge, Nature designed a 3-stage research project, detailed herein.

Stage 1: Depth Interviews with Melbourne Commuters

DOT had a number of hypotheses about the sorts of factors that were important to Melbournians with regard to their travel behaviour. Nevertheless, it was still prudent to engage in a phase of qualitative research to both confirm that the hypotheses were founded, and to identify any other key factors. As such, Nature conducted a series of in-depth interviews with Melbournians who traveled for work, study, or leisure using either private or public transport. This set of interviews enabled Nature to understand that there were a core set of dimensions which underpinned consumer transport mode choice for their main weekday trip from place of home to place of work / study (the main focus of the study):

  • Convenience
  • Value for Money
  • Environmental Concerns
  • Status
  • Utility in the form of activities that could be engaged in during journey
  • Health and Fitness

Stage 2: Segmentation of the Melbourne Metropolitan Population

In August 2009, a representative sample of n=1,500 respondents from Melbourne’s metropolitan population (over 16 years of age) was administered a CATI survey which incorporated a lengthy battery of attitudinal items based on the dimensions revealed during the qualitative phase, as well as a set of questions designed to elicit detail of travel behaviour. Based on responses to the attitudinal items, Nature developed a segmentation solution to understand core group differences in attitudes that related to public transport, using multivariate techniques. The segmentation was developed specifically in relation to the degree to which an individual’s transport mode choice was driven by factors such as:

  • the degree to which they are in-principle wedded to one mode over another for no other reason than simple “preference”
  • the extent to which the individual was able to use, and derived value from using, the in transit for activities such as reading, communicating, working, etc.
  • their degree of environmental focus in making transport mode decisions
  • their degree of health and wellbeing focus in making transport mode decisions.

The primary design reason for conducting the segmentation was to overcome the research challenge discussed above, that is, that the effect of intangible factors on behaviour is not readily measureable. A well crafted segmentation would, we expected, offer a powerful ‘window’ through which consumer decision making and trade-off behaviour could be interpreted and understood. In very simple terms, in order to test DOT’s hypothesis that attitudinal shifts had played a part in driving behavioural patterns, a segmentation defined in relation to the key attitudinal dimensions of interest would provide a means of explicitly testing this hypothesis. In other words, a segmentation of this nature would allow us to examine the degree to which mode choice behaviour, and in particular trade-off behaviour for other more tangible factors, is influenced by attitudes of key interest.

The final segmentation solution revealed 6 segments. Broadly speaking, the segmentation solution revealed two “public transport aligned” segments:

  • “PT lifestylers” are attracted to public transport because it is a cleaner, greener way to travel.
  • “PT works for me” are attracted to public transport because they enjoy the utility it offers – faster and easier travel, and the opportunity to undertake activities in transit (reading, sleeping, listening to music, answering emails on Blackberry, etc.). There were also two “car aligned” segments.
  • “PT rejectors” prefer private vehicle travel so much that many of them would not even consider using public transport if it were free. They reject public transport as a slow, unsafe and unclean way of travel.
  • “Car works for me” segment enjoys the flexibility and speed associated with private vehicle travel based on the trips they have to make.

Outside of this PT vs. car alignment, there were two other segments.

  • “Agnostics” were not particularly engaged in the transport debate and do not hold strong views one way or the other.
  • “Convertibles” were the ‘swinging voters’. This was a segment that was open to public transport but which was being held back by perceptions of lack of the ease of access, service quality, and convenience of PT.

The key attitudes expressed by each segment have been summarised in a quote, as shown in Table 1.

Table 1. Essence of the Attitudinal Segments

Table 1

The prevalence of transport mode segments in metropolitan Melbourne

Choice Model

Interestingly, the segments had almost no relationship with any demographic variable. That is, each segment was not significantly skewed by age, gender, residential distance from the centre of Melbourne, education or income. Segments did, however, report very different patterns of behaviour. As one might predict, the public transport aligned segments reported taking far more trips by public transport than did the car aligned segments. That said, there were still a proportion of the PT rejector segment (21%) who took public transport almost every day, and a proportion of the PT lifestylers (28%) who rarely took public transport. Presumably, due to life circumstances and the availability (and quality) of transport services, not everyone is able to convert attitudes to behaviour to the same extent.

Stage 3: Choice Model

The segmentation covered off the means by which we could understand shifts in the prevalence of particular community attitudes as they pertain to public transport, such that market-level trade-off and mode choice behaviour could be examined at a segment level in order to explicitly test the core hypothesis of the research. The second stage of quantitative research involved a discrete choice model which was designed to determine how transport mode choice varies as a function of trade-offs between more ‘hard-edged’ factors.

There are two main ways in which one can go about understanding this issue. Taking the example of ticket price, we could ask a respondent which transport option they would choose if ticket price was perhaps 10%, or 20% more than current. What we might expect in this circumstance is that successive increases in ticket price would drive an increasing proportion of respondents away from their preferred option. The danger in this approach is that it over-simplifies a complex decision making process. The issue of ticket price is not the only determinant of mode choice, because while one mode may be cheaper than another, there might be a trade-off in terms of convenience, or reliability. Likewise, one might imagine that an increase in petrol cost for private car transport might be offset to some extent by a decrease in traffic congestion.

To address the complex nature of the decision to choose one mode or another, Nature developed a choice model that included a set of mode options for travel in metropolitan Melbourne capturing the vast majority of the means by which the population usually travel:

  • Public Transport: Train, Tram, Bus
  • Private Transport: Car, Motorbike
  • Non-motorised: Bicycle, walking
  • Multi-mode: Train and Car, Train and Bus, Train and Tram

Model attributes that were hypothesised to influence transport mode choice were developed and included. The table below outlines these variables and the boundary of manipulation for each of them. What is immediately apparent about these variables are that some are applicable across transport modes (total travel time, for example, could apply equally to car, train, or tram), some were specific to PT (such as passenger density, ticket cost), and some were specific to private transport (such as petrol or parking cost).

Variables tested in phase 2 Market Segmentation project
T2

The simplest means in which to deliver choice scenarios to respondents would of course be to show all respondents all possible transport modes and ask them to choose which they would take for a particular trip given the levels of the attributes associated with each. However, for almost every person we had hoped to interview, the reasons for travel, and the transport options available to them in real life would be a subset of all possible options. Thus, offering a person the option of Car when they don’t have one, or Train when they live nowhere near a station would be nonsensical, and would ultimately lead to a poor quality model. Likewise, making the context of the trip study, or work commute, or for a sporting event would make sense for some, but not others.

To address this issue, the choice model was tailored for each individual respondent based on what they told us at the start of the survey about each of the travel options available to them for their trip to work or study. In Melbourne, more than 80% of trips on public transport are made from home to place of work or study, so it made sense to constrain the context of the choice model to these situations, and therefore also constrain the sample frame such that each respondent needed to be a person who regularly commuted from home to their place of work or study. Direct questions in the survey elicited from each respondent:

  • Which transport mode options were available to them
  • Which they had ever used
  • And for each they had ever used:
  • How long their usual door-to-door commute was by that mode
  • The chance of a significant delay
  • Their usual ticket type (for PT)
  • Whether, and if so how much, they pay for parking and tolls (for Car)
  • Their perception of travel conditions (passenger density for PT, road congestion for Car and road-based PT)

The choice model was then tailored for each respondent both in terms of the modes available to them, and how changes to variables such as travel time were expressed. For example, a respondent who told us that they travelled by train, and it usually took them 30 minutes to travel door to door by train might see travel durations of 23 minutes (-20%), or 42 minutes (+40%) for that mode. Each respondent saw 12 choice cards which represented different variations in travel conditions. Below is a schematic table of a choice card shown to a respondent who told us that they had Tram, Train, Car, and Bicycle available to them.

Sample Choice Card

Transport mode

The final piece in the puzzle was to administer a predictive subset of the segmentation items to this sample of respondents in order to allocate each respondent to one of the segments defined in stage 1. The survey itself was delivered online to a sample of n=3000 Melbourne metro commuters where the prevalence of each segment was constrained with the use of quotas to ensure an approximately equal distribution.

Results

As discussed earlier, DOT hypothesised that broad societal attitudinal trends and shifts in preference were playing an important role in influencing Melbourne metropolitan transport usage patterns, and were the key reason that its forecasting models were unable to fully account for the massive surge in PT use between 2005 and 2009.

The research design described in this paper has enabled us to directly test DOT’s hypotheses at two levels. First the choice model has allowed macro level measurement of the degree to which consumer mode preference changed as a function of changes in journey quality. From this measurement, likely future patronage patterns can be estimated using the model, and compared with the current outputs of the strategic models. At a macro level, the following key insights on patronage trends have been generated from the research:

  • Future rises in petrol prices may not have the same positive impact on public transport patronage as they have done in the past five years.
  • Future patronage growth will be heavily reliant on improvements in the quality of public transport services (reliability, comfort, travel times) – with projected growth as a result of petrol price rises and road congestion alone are only modest.
  • The main “competitor” mode to car travel is train, the main “competitor” mode to train is tram, and the main “competitor” modes to tram are walking and cycling. Thus, there is a risk that by improving tram services, there will be a more significant shift to tram from walking and cycling than there would be from car travel.
  • Trade-off behaviour exhibited between ticket cost and non-monetary aspects of journey quality such as passenger density and chance of delay means that a dollar value can be placed on changes in non-monetary aspects. From this, we have determined that greater value is placed by commuters on service improvements for tram versus train and bus.

Second, the segmentation in conjunction with choice modelling analyses have delivered an understanding of the degree to which changes to journey attributes impacted on transport mode preference as a function of attitudinal segment – directly addressing DOT’s key hypothesis For example:

  • PT Lifestylers are not surprisingly quite rusted on to PT as a transport option. It follows from their attitudes that they are extremely sensitive to deterioration in car journey quality, in particular shifts in petrol cost or worsening traffic conditions that have almost no impact on other segments.
  • PT Works for Me’s mode preference is similar to that we see for PT Lifestylers except when it comes to Bus, where they exhibit a high level of sensitivity to passenger density and traffic conditions on that mode.
  • Car Works for Me is aligned with car because of the utility it offers. The choice model results show that improvements to journey utility (in particular improvements in traffic conditions and passenger density on tram and bus) cause a shift towards those modes for this segment.
  • PT Rejectors are almost completely immune to service improvements on all PT modes. However, when extreme improvements are made to PT, such as 40% decreases in travel time, this segment does exhibit a shift in preference towards Train in particular. That is, while rejecting of PT, they can be ‘bought’.
  • Convertibles exhibited mode preference behaviour consistent with their perception that PT is inconvenient and unreliable. Further deterioration in train conditions in particular would drive Convertibles away from this mode, while reductions in usual transport time, and cost, would be attractors. The research clearly highlighted this segment – which account for 15% of Melbourne’s metropolitan population – is a key battleground in that they can readily be attracted either toward or away from PT.
  • Agnostics occupy the middle ground in terms of their sensitivity to changes in the quality of the travel experience. As would be hypothesized for this segment, most changes to journey quality went un-noticed with the exception of substantial improvements to journey time on Tram in particular.

Overall, this research has clearly identified a direct relationship between segment membership and the degree to which particular journey attributes drive mode preference. Thus beyond the limits of the research discussed here, tracking of segment sizes will become a worthy ongoing exercise, since it will allow DOT and associated bodies to better determine the impact of changes to journey quality under circumstances in which the Melbourne metropolitan population has a different segment distribution. Because of this, a predictive subset of the segmentation items has been recently added to the Public Transport Tracker operated by Metlink to ensure that an ongoing understanding of segment size is retained.

Conclusion

The current study has successfully provided DOT with a lens through which to view both the push and pull factors that drive patronage to and from public transport modes. In addition, it has also provided the Department with a powerful tool to simulate the impact of possible, realistic, scenarios on patronage for public transport.

The overlay of the attitudinal segmentation, and the ability to manipulate the size of the segments provided DOT with the hitherto missing piece of demand forecasting. While the outcome of the project is not without limitations, (e.g. the survey used to develop the model focused on work and study trips only and does not cover in any way other trip types) it does provide DOT with a solid foundation on which to develop a more effective means of patronage forecasting. Nature and DOT are currently looking at ways in which the quantification of attributes by segment that it is possible to extract from this segmentation model can be used as inputs into its existing patronage models.