A survival analysis-based choice set formation approach for single-destination choice using GPS travel data

1 This research investigates how land use and road network structure influence home-based single2 destination choice in the context of trip chains, using the in-vehicle GPS travel data in the Minneapolis3 St. Paul Metropolitan area. We propose a new choice set formation approach which combines 4 survival analysis and random selection. Our empirical findings reveal that: (1) Accessibility and 5 diversity of services at the destination influence destination choice. (2) Route-specific network 6 measures such as turn index and speed discontinuity also display statistically significant effects 7 on destination choice. Our approach contributes to methodologies in modeling destination choice. 8 The results improve our understanding on travel behavior and have implications on transportation 9 and land use planning. 10 Arthur Huang and David Levinson 2 INTRODUCTION 1 Non-work trips are an important part of daily trips. According to the 2009 National House2 hold Travel Survey, non-work trips make up 81.3% of individuals’ personal trips, where so3 cial/recreational activities make up roughly 27.4 %, shopping trips constitute 21%, family/personal 4 errands share 21.6 %, and school/church trips account for 9.6%. Because of its significant share 5 in daily travel, it is of interest to investigate individuals’ non-work travel behavior and individ6 uals’ decision-making processes using empirical data. The goal of this research is to study the 7 impacts of land use and transportation network structure on individuals’ home-based non-work 8 single-destination choice. The contributions of this research include: 9 • Proposing a new method of forming choice sets for the non-work destination choice 10 problem. 11 • Proposing a procedure to justify the choice set size for destination choice. 12 • Examining the impacts of land use and the characteristics of route-specific network struc13 ture on destination choice based on the in-vehicle GPS data. 14 This paper focuses on home-based non-work single-destination trip chains (i.e., home15 destination-home) for the simplicity of illustrating our approach. 16 The rest of this paper is organized as follows. The following section reviews existing stud17 ies on destination choice. The modeling procedure is proposed in the next session, after which the 18 survival analysis-based choice set formation method is illustrated. The model evaluation criteria, 19 results and analysis are further described. This paper concludes with a summary of key findings. 20 LITERATURE REVIEW 21 With the invention of the discrete choice model, destination choice problems have been a topic of 22 extensive study. This section first describes different approaches of constructing choice sets, and 23 then reviews models used in modeling shopping destination choice models. 24 Choice set formation 25 In a choice problem, a choice set is comprised of all alternatives considered by a traveler. Given 26 a large number of potential destinations in an area for an individual, it is probably unrealistic that 27 one individual considers all of them. Furthermore, it is computationally burdensome to incorporate 28 all of them into a model. The key challenges lie in the methodologies for constructing a choice set 29 in a model and for deciding the choice set size. 30 In terms of selecting choice alternatives, Ben-Akiva and Lerman (1) introduced a series 31 of choice sampling approaches: simple random sampling (a sample is drawn at random from the 32 whole population), general stratified sampling (partitioning the population into a certain number 33 of exclusive strata, selecting sampling fractions, and randomly drawing a designated number of 34 alternatives from each strata), exogenous sampling (defining strata by segments only on attributes 35 and not on actual choices), choice-based samples (choices are defined and controlled for subjects), 36 enriched sampling (the pooling of exogenously stratified samples with one more choice-based 37 samples), and multistage sampling (multiple surveys on preferences of choices). Among these 38 approaches, the most widely used approaches are simple random sampling and general stratified 39 sampling. Both approaches assume that very subject has perfect knowledge of all locations in their 40 Arthur Huang and David Levinson 3 decision-making process. Spiggle and Sewall (2) and Shocker et al. (3) suggested that the decision1 making process involves nested sets of alternatives from higher to lower levels: total set, awareness 2 set, evoked set, and choice (consideration) set; the lower level set is a subset of the higher level set. 3 Spiggle and Sewall (2)’s studies revealed that such proportions vary by individuals and by different 4 brands/products. 5 Hoogendoorn-Lanser and Van Nes (4) and Zhu (5) argued that there exists a hierarchical 6 structure in choice sets. All available choices comprise the universal set. An individual knows 7 a proportion of it (subject choice set) but only considers feasible ones (consideration set), a pro8 portion of which are defacto placed into the actual choice set for a particular decision-making 9 process. Empirically there are different variations for determining a choice set. One direction to 10 impose time constraints on the universal choice set. Travel time budgets are assumed to be the lim11 iting source that restricts the feasible alternatives to only a subset of the universal choice set (6). 12 For example, Thill and Horowitz (7) used travel time constraints to downsize destination choice 13 sets. Several other models used the space-time prism to reduce choice set size given their spatial14 temporal constraints in formulating choice sets for activity-based models (8, 9, 10, 11). (12) built 15 a simulation module to simulate activity patterns. In modeling destination choice, the module cal16 culates the time budgets for each activity and select subsets of destinations for each activity, based 17 on final destinations are selected from the subsets. However, the model is more applicable to fixed 18 activities and less applicable to flexible activities and it does not consider the spatial relationships 19 of destinations for destination choice. Another direction is to select alternative destinations based 20 on certain criteria (such as distance, priorities, lifestyle, and utilities) in combination with afore21 mentioned methods to reduce the universal choice set size (13). More complicated models include 22 stochastic constrain model (14) and the GenL model (15). Nonetheless, they tend to be computa23 tionally inefficient. This area needs to be further investigated with a balance of the complexity of 24 sampling techniques and computational efficiency. 25 Based on previous work, Willumsen and Ortuzar (16) summarized three methods to address 26 the choice set construction problem: (1) random/stratified sampling, (2) using the heuristic or de27 terministic approach, and (3) learning about preferences about locations through surveys. Each 28 approach has its limitations. The random sampling approach, though simple and good at incorpo29 rating all possible alternatives, treats all locations with the same weight yet does not fully consider 30 the spatial effect. Given one trip’s starting point, there are more locations that are farther away than 31 locations nearby. Therefore, a random selection will tend to incorporate more farther-away desti32 nations than nearby locations into the choice set. The deterministic choice-set generation approach 33 aims to limit the number of destinations by defining a radius from the starting point. Nevertheless, 34 the value of radius is hard to pinpoint and may vary by person thanks to inter-personal hetero35 geneity. The stratified importance sampling approach tries to improve simple random sampling by 36 grouping locations with the same or close weights. Nevertheless, it is challenging to justify the 37 number of groups that should be defined, and different numbers of groups may produce different 38 estimates of coefficients. The surveying approach, though can provide useful information regard39 ing priorities, preferences, and tastes, suffers from that fact it is often impossible to ask individuals 40 to precisely document all their alternatives in their decision-making process. Furthermore, there 41 may exist differences between stated preference and revealed preferences. 42 Choice set size is also worth studying. Adler and Ben-Akiva (17) proposed a theoretical 43 model to combine destination choice and mode choice and the size of actual choice set used for 44 modeling can be huge. The advantage of this model is that it considers trip chaining behavior 45 Arthur Huang and David Levinson 4 as an alternative and combines destination choice with mode choice. The disadvantages of this 1 model include: (1) It does not consider the spatial and land use connections of the destinations in 2 a trip tour. (2) The number of alternative destinations used is a hypothetical number which needs 3 to be justified in application. (3) If one more alternative destination is added, the choice set size 4 becomes 2584 (303 more choices than the base case). If one more mode of transportation is added, 5 the choice set will have 3881 choices which are 1100 more than the base case. 6 It should be noted that most of destination choice studies lack a systematic investigation on 7 the appropriate choice set size for modeling. Nerella and Bhat (18) performed numerical experi8 ments to examine the effects of choice set size on the performance of Multinomial (MNL) models 9 and mixed multinomial logit (MMNL) models. The research recommended a choice set size of a 10 fourth of the full choice set for MNL models and one-half of the full choice set for MMNL models. 11 Yet these numbers may still be too large for shopping destination choice models if there are a large 12 number of potential destinations in a metropolitan area. The impact of choice set size on shopping 13 destination models is a research niche. 14 GPS DATA 15 The in-vehicle GPS data used in this research are the same as the data used in Zhu (5) and Carrion 16 and Levinson (19). The Minneapolis-St. Paul Metropolitan Area is selected as the study area 17 because of the availability of the unique GPS travel behavior data. The area typifies many regions, 18 with high-density CBDs and low-density suburbs, allowing comparison of the difference between 19 suburban non-work trips in post-World War II non-work venues and urban non-work trips with a 20 streetcar-derived urban form built in the late 19th and early 20th centuries. 21 The collection process lasted from September to December of 2008, during which 141 22 surveyed subjects made over 20,000 trips. The in-vehicle GPS data collection process includes 23 three stages. The first stage is to recruit the subjects. The announcements on recruiting subjects 24 were posted on various media such as Craiglist.com and Citypages.com, and were sent out via other 25 forms such as postcards handed out in downtown parking ramps and emails to 7000 University of 26 Minnesota staff (excluding students and faculty). More than 900 people responded, from which 27 subjects were selected based on the following criteria: 28 • Age between 25 and 65 29