Allowing for Heterogeneity in the Consideration of Airport Access Modes: The Case of Bari Airport

Mode choice models traditionally assume that all objectively available alternatives are considered. This might not always be a reasonable assumption, even when the number of alternatives is limited. Consideration of alternatives, like many other aspects of the decision-making process, cannot be observed by the analyst, and can only be imperfectly measured. As part of a stated choice survey aimed at unveiling air passengers’ preferences for access modes to Bari International Airport in Italy, we collected a wide set of indicators that either directly or indirectly measure respondents’ consideration of the public transport alternatives. In our access mode choice model, consideration of public transport services was treated as a latent variable, and entered the utility function for this mode through a “discounting” factor. The proposed integrated choice and latent variable approach allows the analyst not only to overcome potential endogeneity and measurement error issues associated with the indicators, but also makes the model suitable for forecasting. As a result of accounting for consideration effects, we observed an improvement in fit that also held in a validation sample; moreover, the effects of policy changes aimed at improving the modal share of public transport were considerably reduced. The number of air travelers in the European Union has significantly increased in recent years (1). This growth has been largely driven by low-cost carriers, which made air transport economically affordable to a larger share of the population. However, this expansion continuously imposes a challenge for airport managers and regional mobility planners, who not only have to deal with the increasing number of (infrequent) travelers, but also the additional staff and accompanying persons needing to access the airport. There is no generic solution to this challenge that is valid everywhere; in addition to this, each user segment (e.g., resident vs. non-resident, business vs. non-business, or airport employees) has its own needs and preferences for airport access services (2, 3). Most studies investigating the drivers of airport access mode decisions have relied on revealed preference (RP) and (or) stated preference (SP) data in combination with discrete choice models. These studies were aimed at understanding the choice between existing access modes (4, 5), or focused on the implications of introducing a new access mode (6, 7). In some cases, access mode decisions have been modeled jointly with airport, airline decisions, or both (8–10). The underlying assumption in all these studies is that all objectively available airport access modes are effectively considered by each airport user. However, this assumption might be questioned as some access modes might be discarded a priori, that is, regardless of their characteristics. For example, in the case of air travelers, trips to the airport are only the first leg of a longer trip and are associated with a hard constraint: the departure time of the flight. Hence, the possible consequences of a delay in arriving at the airport may be severe. Even though unexpected delays might occur with all modes, air travelers might consider as feasible only those alternatives that they perceive to have a sufficiently low risk of getting to the airport late. Other factors that might influence which alternatives are considered include concerns for personal safety, and the need to access a train station/bus stop that is inconveniently located with respect to their departure point. Comfort also matters, particularly because passengers perceive the requirement to transfer and wait (e.g., with public transport) as a significant “discomfort” (11). The assumption that individuals might consider only a subset of the available alternatives has been tested in several transport contexts, particularly route and mode choice (12, 13). However, to the best of our knowledge, this assumption has never been tested in the specific context of airport accessibility, which is the focus of this paper. The biggest challenge with the consideration of alternatives is that this aspect of the decision-making process is not observable to the analyst. Some researchers have tried to incorporate consideration effects into probabilistic models only on the basis of the observed choices (14–16). Others have explored the possibility of using supplementary information as direct (but imperfect) measures of consideration, including for example the perceived availability (17) or acceptability (18) of the alternatives and thresholds for attributes (19), elicited using ad hoc questions in travel surveys. These indicators, however, might not correspond to actual levels of consideration, that is, there is potential for measurement error, and they may be correlated with other unobserved factors, that is to say, there is scope for endogeneity bias (20). Given this, rather than using them as “error-free” measures of consideration, it might be preferable to recognize that these are a function of latent consideration, and treat them as dependent rather than independent variables using an integrated choice and latent variable (ICLV) model (21). The ICLV approach has been used extensively in many fields to incorporate either psychological factors such as attitudes and perceptions (22) or respondents’ processing strategies (20) into models based on random utility maximization (23). In addition to allowing the analyst to overcome potential endogeneity and measurement error issues with the indicators, the ICLV approach also allows the analyst to make the indicators suitable for forecasting. In this paper, the ICLV framework is adopted to measure consideration of airport access modes using three distinct sets of indicators collected as part of a stated choice (SC) survey. The first set consists of the level of agreement with various perception statements and of a preference-based ranking of the alternatives; the second refers to thresholds for attributes inferred from respondents’ previous choices; the third set comprises direct reports of consideration of the alternatives. These indicators were chosen because they represent additional sources of information that are generally collected during travel surveys (the first two), or because they have been used in previous studies to measure consideration of the alternatives (the third). In our proposed formulation, latent consideration explains the indicators and enters the utility of an alternative through a discounting factor. The discounting factor accounts for consideration lowering the utility, and therefore choice probability, of a supposed unconsidered alternative. Data for this study came from an SC experiment on a sample of air travelers of Bari International Airport, in Italy. This airport has recently experienced a substantial increase in travelers (1) as a result of the increase in the number of low-cost connections available. A direct train connects the airport with the city center in 15 min; however, more peripheral areas within the Metropolitan City of Bari and the Apulian region are not as easily accessible, as the railway link to the airport is not interconnected with the main regional railway networks. Other public transport means are available (e.g., local buses), but these involve at least one interchange, are even less frequent, and their timetables are not coordinated. As a result, travelers from these areas mainly access the airport by car. Given these premises, in this paper we estimate mode choice models in which we allow for the possibility that some air travelers might not consider public transport as a feasible alternative. Both RP and SP data are used in the estimation, and the proposed ICLV models are compared with two reference models: the first is a traditional mixed multinomial logit (MMNL) model in which all alternatives are assumed to be considered. The second is a reduced-form MMNL model of the proposed ICLV models, which only infers the latent consideration of public transport through the observed choice data. The remainder of the paper is structured as follows. We describe the data in Section 2. Section 3 explains the proposed model. In Section 4, we report and discuss the estimation results, and in Section 5 we present the validation exercise. Finally, in Section 6 we draw conclusions from our study.