Systematic Review of Transportation Choice Modeling

Abstract

This research presents an overview of transportation mode choice, emphasizing key influencing factors and a range of methodological approaches from traditional Random Utility Theory (RUT) models to modern Machine Learning (ML) techniques. A comprehensive review covered 875 papers, which were screened for relevance. The search was conducted on ScienceDirect and Google Scholar between October and November 2024 using the keywords transport and choice model. Search results were reviewed until several consecutive entries no longer contained content relevant to the topic. After the screening and exclusion process, 106 papers remained for analysis. The review reveals that the Multinomial Logit (MNL) model remains the most widely used approach for modeling transportation mode choice, despite a growing interest in ML methods. Cars and buses dominate in passenger transport studies, while trucks, trains, and ships are most common in freight research. Data is typically collected through surveys (for passenger transport) and interviews (for freight), though some studies use secondary sources. Geographically, Asia and Europe are most represented, with regions like South America underrepresented. Travel time and cost are key variables, with increasing attention to the built environment in passenger studies and service reliability in freight studies. Overall, most studies aim to address real-world transport challenges. The review highlights the persistent gap between theoretical advancements and real-world applicability. To support this analysis, it examines the specific research objectives and findings of each study.

1. Introduction

The specific choices an individual makes are influenced by a variety of factors. Through modeling, it is possible to predict individual behavior by identifying the variables that influence decision-making. This approach is particularly relevant in the context of transportation mode selection. Modeling helps us understand individual travel behavior, including which mode of transport is chosen and the reasons behind that choice. The need to apply modeling techniques may arise from various sources, such as research objectives, regulatory requirements, geographical characteristics of certain areas, or demographic patterns. Regardless of the underlying motivation, the insights obtained through modeling support the development of effective measures. These measures may aim to improve the current state of the transport system, introduce new modes of transport, or enhance our understanding of how individuals use existing transportation infrastructure.

The factors influencing the choice of transportation mode have evolved over time. Initially, the primary goal was simply to reach the destination safely. Later, speed and cost became more important. Recently, criteria such as environmental sustainability, reliability, and predictability have come to the forefront. Decision-making regarding the choice of transport mode is accompanied by uncertainty [1]. Leong et al. [2] highlights the randomness of transportation mode selection throughout the day. For instance, if an individual chooses public transport for commuting to work, this decision impacts the rest of the day. Without a car, that individual may not be able to make large purchases after work. Such situations are often overlooked in transportation mode modeling. Asgarpour et al. [3] emphasize the unpredictability of human behavior, as each individual makes decisions differently [4]. To better understand an individual’s choice of transportation, modeling can be used to illustrate decision-making patterns. While Random Utility Models (RUM) were initially dominant, ML models, such as Neural Networks (NNs), have been gaining increasing attention in recent years.

As noted by Sekhar et al. [5], Discrete Choice Models (DCM), particularly Logit-based approaches, are widely adopted in transportation research due to their ability to model complex travel behavior using relatively simple mathematical structures. Despite their broad application in mode choice studies, such as MNL, these models face notable limitations regarding accuracy and flexibility. Consequently, there has been growing interest in Artificial Intelligence (AI) techniques as promising alternatives to traditional statistical models in transportation engineering. As highlighted by Sekhar [6], the fundamental purpose of mode choice modeling lies in effectively managing transportation demand and adapting the existing system to meet that demand.

The modeling of transportation flows has become more complex due to the introduction of new transportation modes. However, only through modeling can we assess the usability and attractiveness of these new transport options for users. It is crucial to determine how individuals will decide when multiple transport options are available at a given location [7]. Obregón-Biosca [8] aptly captured the essence of transportation modeling with the question: “Why do some people drive, others use transit, and others use active modes of transportation?”

2. Choice Modeling

The choice of transportation mode is a complex decision-making process that involves weighing multiple factors, such as travel time, cost, convenience, and personal preferences. Traditionally, this decision has been modeled using DCMs, based on RUT. This theory assumes that individuals choose the option that maximizes their perceived utility, where utility represents the overall satisfaction or benefit derived from a specific travel mode. The utility of a transportation mode depends on observable factors, such as travel time, cost, comfort, and reliability, as well as personal characteristics of the traveler, including income, age, travel habits, and accessibility needs. In addition, unobserved factors, such as mood, weather conditions, or unexpected disruptions, introduce randomness in the decision-making process, which is captured in models as a random error component [6,9,10].

One of the most widely used DCMs is the MNL model, which estimates the probability of choosing a specific transportation mode based on its utility relative to other available options. Due to its analytical tractability and relatively high predictive accuracy, it remains one of the most widely applied models in DCM [6]. However, MNL relies on a key assumption known as Independence from Irrelevant Alternatives (IIA), which means that introducing a new transportation option does not affect the relative probabilities of choosing between existing options. In reality, this assumption is often unrealistic, particularly when travel modes share similarities, such as bus and tram or ride-sharing and taxis. To address these limitations, researchers have developed more flexible models, such as the Nested Logit (NL) model, which groups similar alternatives into hierarchical structures to better reflect real-world substitution patterns. The Mixed Multinomial Logit (MXL) model introduces variation in preferences across individuals, allowing for more realistic travel behavior predictions. Other advanced models, such as Generalized Extreme Value (GEV) and Probit models, further relax the IIA assumption, but they require more complex computations and are less commonly applied in practice [6,9,10]. Interpretation of results is a critical component in the application of these models. Traditional models like MNL and NL provide coefficients that can be directly interpreted in terms of marginal effects on choice probability, making them especially valuable for policy analysis. However, the validity and stability of these interpretations heavily depend on the quality and granularity of the input data. Models built on incomplete, outdated, or overly aggregated data may yield biased or misleading outcomes. Thus, accurate data collection, whether through carefully designed surveys or reliable administrative sources, is essential to ensure robust and policy-relevant insights [6].

As transportation networks become more dynamic and complex, researchers have increasingly turned to AI and ML techniques. Unlike traditional models that rely on predefined equations and rigid assumptions, ML models learn directly from data, identifying patterns without requiring a specific functional form. This makes them particularly useful for handling large, diverse, and nonlinear datasets that traditional models struggle with. AI-based methods, such as Support Vector Machines (SVMs), Classification Trees (CTs), Random Forests (RFs), Extreme Gradient Boosting (XGB), and NNs, have been widely adopted in transportation research due to their ability to capture intricate relationships between variables. These methods offer better accuracy, greater adaptability, and the ability to process real-time data, making them highly effective for travel demand forecasting, policy evaluation, and urban transportation planning [9,10].

AI-based models offer significant advantages in terms of predictive accuracy, adaptability, and scalability. NNs, for example, can model highly nonlinear relationships and interactions between variables, often outperforming traditional models in terms of fit and forecast quality. Its performance depends significantly on the network’s architecture, training process, and activation functions [6]. However, these benefits come at a cost. The interpretability of ML models is often limited, especially with deeper architectures, which can hinder their use in policy-making where transparency and justification are critical. Furthermore, ML methods typically require large amounts of high-quality, granular data to train effectively and avoid overfitting, and their computational demands, in terms of processing power and training time, are substantially higher than those of traditional DCMs. In practical applications, the choice between traditional and AI-based models involves a trade-off between interpretability and flexibility, as well as between computational efficiency and predictive power. While DCMs remain dominant in theoretical research due to their transparency and analytical tractability, AI and ML models are gaining ground in applied contexts, particularly where decision-making is driven by real-time data and requires high adaptability. As mobility patterns continue to evolve, shaped by trends like shared mobility, electrification, and autonomous vehicles, AI-driven models are expected to play an increasingly central role in supporting dynamic, data-informed transportation planning. In practical applications, the choice between traditional and AI-based models involves a trade-off between interpretability and flexibility, as well as between computational efficiency and predictive power.

The primary advantage of AI and ML models over traditional DCMs lies in their ability to model complex, nonlinear relationships without imposing restrictive assumptions. Unlike MNL models, which assume homogeneity in individual responses, AI techniques can account for variations in traveler behavior across different demographics and geographic regions. Additionally, they can continuously improve as they incorporate new data from sources such as sensors and travel apps. While traditional DCMs remain valuable for theoretical insights, AI-driven models are increasingly being used in real-world applications, providing more precise and adaptable solutions to the challenges of modern transportation systems. As mobility patterns evolve with emerging technologies such as autonomous vehicles and smart transportation networks, AI and ML will play an increasingly central role in shaping the future of transportation modeling [9,10].

2.1. The Need for a Review

This review is a part of a broader study. The idea is to create a model that determines, for any location (as a point object) in, e.g., Europe, whether it is more convenient to travel within a predefined radius by plane, train, or another mode of transport. The radii would gradually increase, for example, from 100 to 1000 km, allowing us to observe how distance influences transportation choices. The final result would be a map indicating the preferred mode of transportation for each point for each given radius. Such a map would have a significant practical value, helping to identify areas where investments in specific transport infrastructure are needed, for example, in a high-speed rail network in Eastern Europe. The model’s usefulness could be further enhanced by incorporating additional variables beyond travel time, such as CO₂ emissions, travel comfort, and connectivity, elevating its practical value even further. For this purpose, a review of the field of transportation mode choice modeling is needed. Also, in researching decision-making regarding the use of different transportation modes, we identified shortcomings in the presentation of results. Most studies focus on the performance of models such as MNL, while clear and publicly accessible interpretations of findings are lacking. Although key variables are analyzed and models properly evaluated, greater emphasis should be placed on presenting final results in a format with immediate and strong communicative value, such as through maps.

We reviewed other review papers in the field of transportation mode choice modeling. The review by Hillel et al. [11] focuses on ML classification algorithms, but we aimed to cover a broader spectrum, including other mode choice models. The review by Sekhar [6] concentrates on mode choice models from a theoretical perspective and lacks practical examples. The review by Jing et al. [12] addresses the choice of sustainable transport modes, considering Random Regret Optimization, with a particular emphasis on the environmental aspect. Our review of decision models will not be focused solely on a specific viewpoint. We aim to highlight the results of research, with the environmental aspect being just one of many. Compared to the presented studies, our review will be more comprehensive and will not be limited to a specific type of method or a particular aspect of analysis. Special attention will also be given to the results, focusing on the specific cases in which transportation mode choice has been studied.

2.2. Overview of the Paper

The remainder of the paper is laid out as follows. Section 3 presents the methodology for the review, including the research questions, review protocol, and study selection. Section 4 presents the results of the review. Section 5 summarizes the findings, identifies potential limitations of the review, and presents the conclusions.

In the text, we will provide examples of studies. All studies that address a specific aspect of the research will be presented in tables. The general limitations of the research are related to the fact that our knowledge at the beginning of the review, when we were searching for relevant papers, was at a somewhat lower level than it is now, which may have influenced our selection process. There are still certain gaps in our theoretical knowledge of this field, which require future work.

3. Methodology

The procedure for this systematic review is adapted from that given by Hillel et al. [11], which was adapted from Kitchenham and Charters. Hillel et al. [11] excluded study quality assessment from their review, since they did not want to draw conclusions from the aggregate results or combined findings of the studies. So, no assessment of the quality of each study is made. We have adapted the structure from Hillel et al. [11] to the following format:

• Planning the review

  ○

  Identification of the need for the review;

  ○

  Specifying the research questions;

  ○

  Developing a review protocol.

• Conducting the review

  ○

  Identification of the research;

  ○

  Selection of primary studies;

  ○

  Data extraction;

  ○

  Data analysis.

• Reporting the review

  ○

  Specifying dissemination mechanisms;

  ○

  Formatting the main report.

This review is focused on summarizing the methodologies used in each study, and as such, no attempt is made to draw conclusions from the aggregate results or combined findings of the studies. No meta-analysis was conducted. Additionally, the review was not registered, and no protocol was prepared.

3.1. Research Questions

The focus of this review is on the methodologies used to model transportation mode choice. This review serves to investigate the following research questions:

• Which model was used to determine the optimal mode of transport?
• What types of transportation were included in the study?
• What data sources were used in the research?
• What geographical area was covered by the study?
• What specific aspects were examined in the research?
• What is the practical value of the research?

3.2. Review Protocol

3.2.1. Search Strategy

The search strategy is implemented to identify relevant papers for the review. In order to collect relevant papers, papers are retrieved from the database ScienceDirect and the website Google Scholar. Relevant papers were searched between October and November 2024. The keywords transport AND choice model were used for the search. Entering these keywords into ScienceDirect yielded approximately 500,000 results, while Google Scholar returned around 5,000,000 results. Based on the relevance of the results generated in the search, we reviewed the entries until there were no relevant contents related to our topic, transportation mode choice, in several consecutive results. Our aim was to capture a broad range of studies analyzing how transport mode choices were modeled. We have considered the time-consuming nature of such a search for relevant papers.

3.2.2. Selection Criteria

Papers from any period up until the search date are included in the review. The following eligibility criteria are determined for the papers found in the search to be included in our review:

• Studies should be published in peer-reviewed journals;
• Studies must be accessible to the academic public;
• Studies must be written in English;
• The study must use at least one decision-making model;
• The study must address a specific problem in a particular area.

Paper selection is carried out using the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) guidelines:

• In accordance with these guidelines, duplicates are removed from search records;
• Record titles, abstracts, introductions, and conclusions are screened against the eligibility criteria;
• Remaining full-text papers are reviewed for eligibility.

All stages of the selection process were conducted independently by the first author, as part of the doctoral study program, and under the supervision of the academic supervisor and the working mentor.

3.2.3. Data Extraction Strategy

Each study is reviewed in detail. Key elements for each study are determined and tabulated in various spreadsheets. Based on a review of abstracts, introductions, and conclusions, we selected a set of 114 papers. All 114 papers were then thoroughly reviewed. Eight papers were excluded from our final selection due to inconsistencies between their titles, abstracts, introductions, conclusions, and actual content, which did not align with the scope of our review. An example of such a case is the study by Guo [13], which examined path choices in public transit. Based on the title, the topic fell within the scope of our review, but upon examining the actual content, we found that no model was used in the study. Therefore, we excluded it from our review. Figure 1 shows the PRISMA flowchart of the paper selection process. Data extraction is carried out independently by the first author as part of the doctoral study program, and under the supervision of the academic supervisor and the working mentor.

4. Results

This section presents the results obtained from the systematic review process. Firstly, an overview of the 106 papers used for data extraction is given. Later sections then use evidence from 106 studies to explore each of the six research questions. The presented evidence is complemented by informative figures and tables.

4.1. Basic Information About Papers for Data Extraction

A total of 89 studies focused on passenger transport, while 17 studies addressed freight transport. Arunotayanun and Polak [14] noted that decision-making models in freight transport receive little attention, citing the complexity of the freight transport system as a major obstacle. Similarly, García-Menéndez et al. [15] highlighted that significantly fewer studies focus on freight transport compared to passenger transport. The numerical ratio of one to five in the review confirms this.

Our review includes papers from 46 different sources, including journals, proceedings, open access forums, and a digital library. The journal Transportation Research Part A contributed the highest number of papers, 15, followed by the journal Transportation Research Record with eight papers, and the journal Transportation Research Part E with six papers.

Table 1. Number of publications in scientific Journals.

Publication	Type	No. of Publications
Transportation Research Part A	Journal	15
Transportation Research Record	Journal	8
Transportation Research Part E	Journal	6
Transport Policy	Journal	5
Journal of Air Transport Management	Journal	5
Transportation Research Interdisciplinary Perspectives	Journal	4
Transportation	Journal	4
Transportmetrica A: Transport Science	Open access forum	3
Transportation Research Procedia	Proceedings	3
Sustainability	Journal	3
Journal of Transport Geography	Journal	3

presents publications with at least three papers included in the review.

The number of references by publication year is presented in Figure 2. If we examine the distribution of papers by publication date, we can see that most were published in 2024, 21 papers. The oldest paper was published in 1990 [16], and there are just a few papers published till 2003. As early as 1990, a computer program was used for parameter estimation.

4.2. Models Used in References

The number of papers with specific classification algorithm models is presented in Figure 3. Among decision-making models, the MNL model is the most frequently used, both as a primary model and as a benchmark for comparison, appearing in 62 papers: 54 in passenger transportation studies and 8 in freight transportation studies. Researchers also use the MNL model to evaluate the effectiveness of other decision-making models, either newly adapted or developed from the existing MNL framework [4].

Among other decision-making models, the NL and Mixed Logit (MLog) models stand out, together appearing in 21 and 14 studies, respectively. Multiple models can be used within a single study. For example, in the research by Al-Salih and Esztergár-Kiss [17], the MNL model was applied to determine the impact of attributes on transport mode choice, while the NL model was used to assess the influence of demographic factors. Additionally, we identified numerous individually applied models and model combinations developed to address specific case studies.

As shown in Figure 3, most of the models in the review appear only once. The example of the study with a specific classification algorithm is presented in Table 2.

Since 2017, there has been a notable acceleration in the number of studies utilizing ML, a trend that can be attributed to the emergence of advanced modeling approaches grounded in ML methodologies. Including the year 2017, ML methods were employed in 19 studies. Examples include Paredes et al. [18], who applied RF, SVM, DT, and XGB to investigate how ML models can outperform DCM in predicting car ownership in Singapore, and Díaz-Ramírez et al. [19], who utilized DT, RF, SVM, and XGBoost for predicting transportation mode choice preferences in a university district. In contrast, prior to 2017, ML methods had been used in only five studies. For instance, Omrani [20] applied NNs (Multi-Layer Perceptron and Radial Basis Function) as well as SVM to predict travel mode choice of individuals in the city of Luxembourg. The shift toward the use of ML models can also be linked to the increasing availability of travel data in recent years. These include datasets generated through integrated public transport smart cards [21], or data from shared mobility applications [22]. Such data sources were not available in the past, when researchers relied primarily on traditional methods of data collection, including government statistical databases collected through traditional methods, travel surveys, and questionnaires. Despite recent developments in ML methodologies, MNL still remains the predominant modeling approach. The popularity of the MNL model can be attributed to its widespread use, simplicity, and interpretability. These characteristics make it not only a common choice as a primary modeling approach but also a valuable benchmark against which newer and more complex methods, particularly ML models, can be compared. The established theoretical foundation and relatively straightforward implementation of the MNL model provide researchers with a reliable starting point, facilitating model development and validation. As ML techniques gain prominence, the MNL model continues to serve as a standard reference, helping to evaluate the added value and practical applicability of these emerging approaches. The dynamics of movement can be observed in Figure 4.

Many studies focus on comparing different decision-making models. Hassan et al. [22] argue that the larger the research sample, the more accurate the MNL model is compared to the NL model. Lee et al. [23] state that the MLog model automatically resolves the IIA problem and that incorporating additional variables improves curve fitting. Patil et al. [24] claim that the MLog model performs better than MNL. Jung and Yoo [25] and Zhou et al. [26] demonstrated that the NL model is superior to MNL. Mepparambath et al. [21] highlight that the NL model prevents overlap in travel sequences, ensuring a more structured representation of alternatives, as confirmed in Rich et al. [27]. Kamargianni and Polydoropoulou [28] emphasize the advantages of the HCM over MNL. Shahrier and Habib [29] found no significant difference between ICLV and MNL. Meanwhile, Siqueira et al. [30] concluded that the ML model outperforms NL.

Srivastava and Ravi Sekhar [9] assert that NN is superior to both MNL and NL due to its higher accuracy and improved explainability. Tan et al. [31] found that PSL models outperform the basic MNL model in terms of estimation accuracy. Cheng et al. [32] used the RF method, as it better captures variations in data. Hamadneh and Jaber [33] state that DTs offer an advantage due to their simple result interpretation and the ease of integrating new transport modes for comparison. Sekhar et al. [5] found that the RF achieved a higher prediction accuracy, 98.96%, compared to Logit models with 77.31%. Zhao et al. [34] demonstrated that RF significantly outperforms MNL and ML in predictive accuracy. However, they also noted that ML results are often difficult to interpret and clearly articulate. Additionally, they found that ML and Logit models generally agree on variable importance and the direction of their influence on transportation mode choice.

Table 2. Classification of used algorithms.

Classification Algorithm	Reference
Binary probit (BIP)	Zamparini et al. [35]
BL	Abdel-Aal [36]
Conditional Logit (CL)	García-Menéndez et al. [15]
Cross-nested logit (CNL)	Zhao et al. [34]
Error Component Logit (ECL)	Günay [37]
Hybrid Choice Model (HCM)	La Paix et al. [38]
Heteroscedastic Extreme Value (HEV)	Norojono and Young [39]
Hierarchical Nested Logit (HNL)	Iglesias and Raveau [40]
Integrated Choice and Latent Variables (ICLV)	Mohiuddin et al. [41]
Integrated Choice and Latent Variables based on Nested Logit Models (ICLV-NL)	Guo et al. [42]
Latent Class Choice Model (LCCM)	García-Melero et al. [43]
Latent Class Model (LCM)	Shahrier and Habib [29]
Latent Class Multinomial Model (LC-MNL)	Zhou et al. [44]
Latent Class Nested Logit Model (LC-NL)	Zhou et al. [45]
LOGIT	Hunt [16]
Multiple Discrete-Continuous Model (MDC)	Liao et al. [7]
Mixed PSC-Logit (Mixed PSC L)	Anderson et al. [46]
MLog	Hofer and Fellendorf [47]
Mixed Multinomial Logit (MMNL)	Birolini et al. [48]
Multi-Nested Generalized Extreme Value (MN-GEV)	Bovy and Hoogendoorn-Lanser [4]
MNL	Akar et al. [49]
Multinomial Probit Model (MNP)	Can [50]
Neurofuzzy Multinomial Logit (NFMNL)	Andrade et al. [1]
NL	Danaf et al. [51]
PSC-Logit (PSC L)	Anderson et al. [46]
Path Size Logit (PSL)	Nassir et al. [52]
Recursive Logit Model (RLM)	Leong et al. [2]
Rank-Ordered Logit Model (R-OLM)	Beuthe and Bouffioux [53]
Random Parameter Error Component Logit (RPECL)	Cordera et al. [54]
Random Parameters Logit (RPL)	Günay [37]

Table 2. Classification of used algorithms.

Classification Algorithm	Reference
Binary probit (BIP)	Zamparini et al. [35]
BL	Abdel-Aal [36]
Conditional Logit (CL)	García-Menéndez et al. [15]
Cross-nested logit (CNL)	Zhao et al. [34]
Error Component Logit (ECL)	Günay [37]
Hybrid Choice Model (HCM)	La Paix et al. [38]
Heteroscedastic Extreme Value (HEV)	Norojono and Young [39]
Hierarchical Nested Logit (HNL)	Iglesias and Raveau [40]
Integrated Choice and Latent Variables (ICLV)	Mohiuddin et al. [41]
Integrated Choice and Latent Variables based on Nested Logit Models (ICLV-NL)	Guo et al. [42]
Latent Class Choice Model (LCCM)	García-Melero et al. [43]
Latent Class Model (LCM)	Shahrier and Habib [29]
Latent Class Multinomial Model (LC-MNL)	Zhou et al. [44]
Latent Class Nested Logit Model (LC-NL)	Zhou et al. [45]
LOGIT	Hunt [16]
Multiple Discrete-Continuous Model (MDC)	Liao et al. [7]
Mixed PSC-Logit (Mixed PSC L)	Anderson et al. [46]
MLog	Hofer and Fellendorf [47]
Mixed Multinomial Logit (MMNL)	Birolini et al. [48]
Multi-Nested Generalized Extreme Value (MN-GEV)	Bovy and Hoogendoorn-Lanser [4]
MNL	Akar et al. [49]
Multinomial Probit Model (MNP)	Can [50]
Neurofuzzy Multinomial Logit (NFMNL)	Andrade et al. [1]
NL	Danaf et al. [51]
PSC-Logit (PSC L)	Anderson et al. [46]
Path Size Logit (PSL)	Nassir et al. [52]
Recursive Logit Model (RLM)	Leong et al. [2]
Rank-Ordered Logit Model (R-OLM)	Beuthe and Bouffioux [53]
Random Parameter Error Component Logit (RPECL)	Cordera et al. [54]
Random Parameters Logit (RPL)	Günay [37]

Díaz-Ramírez et al. [19] highlight the challenge of interpreting ML model results, a concern echoed by Jin et al. [55]. Despite this, they argue that ML models outperform statistical and econometric models. They also claim that ICLV is a superior model compared to MNL, as it allows for the integration of behavioral patterns into decision-making. Liao et al. [7] found that MDCEV and mixed MDCEV fail to capture the flexibility of individual transportation mode choices. Wang and Ross [56] state that the XGB model generally achieves higher predictive accuracy than MNL, particularly when the dataset is not highly imbalanced. The XGB model’s dataset is divided into training and test sets, and its lower error rate results in an overall prediction accuracy of 94.5%, compared to 92.7% for MNL. This suggests that the XGB model reduces overall prediction errors compared to MNL.

Wang et al. [57] found that MNP outperforms MNL, CNL, Heteroscedastic Independent MNP (HI-MNP), and Homoscedastic Non-independent MNP (HONI-MNP). Additionally, HI-MNP is better than both MNL and CNL, while HONI-MNP outperforms MNL. Zhang and Xie [58] concluded that SVM is superior to MNL. Zhang et al. [59] found that DNN outperforms MNL, NL, fully connected NNs, and RF. Paredes et al. [18] argue that ML models perform better when variables are specifically adjusted for them. If variables designed for DCM are used in ML models, the results tend to be inferior. Their study compared MNL with various ML models, including RF, SVM, DT, and XGB. They found that MNL yielded better results when ML models were not provided with optimized variables. Kamargianni and Polydoropoulou [28] stress the need to include unobserved variables in the decision-making process to create more realistic econometric models, enabling tailored and targeted policy-making. Standard variables alone do not sufficiently capture human behavior [60].

As support for the functioning of the models, authors also include specific software applications in their research, such as Geographic Information Systems [29,61] and various program packages, examples of which are provided in

Table 3. Used software applications.

Program Package	Reference
Biogeme	Leong et al. [2], Mepparambath et al. [21], Siqueira et al. [30], Zhao et al. [34], La Paix et al. [38], Iglesias and Raveau [40], Guo et al. [42], Anderson et al. [46], Hess et al. [62], Hess et al. [63], Ding and Zhang [64], Jin et al. [65], Yap et al. [66], Kurauchi et al. [67], Shah et al. [68], Ingvardson et al. [69], Kölker et al. [70], Tarkkala et al. [71]
NLOGIT	Obregón-Biosca [8], Srivastava and Sekhar [9], Zhou et al. [26], García-Melero et al. [43], Zhou et al. [45], Gokasar and Gunay [72], Román et al. [73], Spinney et al. [74]
Ngene	Zhou et al. [26], Zhou et al. [45], Román et al. [73], Arencibia et al. [75], Ilahi et al. [76], Hidayati et al. [77]

.

We also find the use of toy scenarios and real-life scenarios interesting. The toy scenario represents a hypothetical case, which is suitable for testing the model’s functionality, while the real-life scenario addresses a genuine problem. Liu et al. [78], after successfully testing the model’s functionality, tackled a real-world problem by studying the integration of Ride-Hailing services, Uber, with public transport in the city of Shanghai.

4.3. Modes of Transportation

In classifying transportation modes, we took certain assumptions into account. The jitney, a regional taxi vehicle, and the shuttle were considered under the Taxi mode of transportation [51,72]. On-demand transport services are considered based on the type of transportation used. On-demand car services are included under the Passenger in a Car mode [76]. The variants of motorcycles, such as autorickshaws, have been considered under the Motorcycle transportation mode [79]. Studies that addressed freight transportation have listed road, rail, and water as transport modes. We have considered road under the Truck mode, rail under the Train mode, and water under the Ship mode.

The frequency of transportation modes in papers is shown in Figure 5. Among the modes of transportation included in the papers, Bus and Car are the most frequently recorded for passenger transport, each appearing in 54 cases. It is important to note that some studies treat public transport as a single mode, while others disaggregate it into specific transportation modes available in the study area. Train, Bike, Walk, Metro, and Taxi have very similar representation, although Train is also commonly included in freight transport modeling along with Truck, which is the most frequently presented option in freight transport modeling, and Ship. This can be linked to the fact that these transportation modes are typically present in the same area, and thus, they represent mutual competition. Additionally, only a small number of studies consider multi-modal transport combinations. It is important to note that some authors overlook certain modes of transport in their studies due to their small share in the overall data [80].

Various variables were included in the studies. The time variable was considered in 88 papers, 73 of which focused on passenger transport and 15 on freight transport. Costs were analyzed in 67 papers, with 51 addressing passenger and 16 freight transport. Distance was included in 21 papers, 19 related to passenger and two to freight transport. Additionally, 53 papers incorporated other variables, which are presented in

Table 4. Variables in choice models.

Variable	No. of References	Variable	No. of References
Time	88	Comfort	12
Cost	67	Reliability of a service	12
Build Environment	21	Frequency	10
Distance	21	Number of transfers	6

. Among the remaining variables, comfort was commonly observed in passenger transport, while reliability of service was noted in freight transport. It should be noted that we did not distinguish between different types of time, cost, and distance. If any form of time, cost, or distance was used in the study, we categorized it as time, cost, and distance accordingly. We were also interested in which other variables were included in the studies. We found that most of the other variables can be classified into the following categories: built environment, which includes walkability, traffic index, and parking difficulty; comfort, which includes crowding and group size; reliability of a service, which includes punctuality; as well as frequency and number of transfers. A detailed overview of the variables is provided in Appendix A.1.

It is important to note that a given variable does not carry the same meaning across different contexts. Let us compare Europe and Asia. For example, from a cost perspective, car travel might appear to be a rarely used mode of transportation. However, personal car usage for mobility is widespread throughout Europe. This is partly due to the relatively affordable cost of car ownership and fuel, as well as the well-developed road infrastructure that supports private vehicle use. Similarly, in the context of freight transport, the dominance of road freight, particularly trucks, in Europe can be attributed to the extensive and efficient road network, high accessibility, and relatively short transport distances. Moreover, geographical constraints, such as mountainous terrain and water barriers, often make rail or maritime freight less practical, reinforcing the preference for trucking as an optimal solution. In contrast, many Asian metropolises face high population density and limited space, which has led to policies restricting private vehicle ownership through high taxes, congestion charges, or limited parking availability, making car travel more expensive and less accessible. Consequently, the cost factor plays a different role in travel mode choice in these regions, where even those who wish to purchase a car may find it practically impossible due to regulatory or spatial constraints. Asian cities have therefore been compelled to develop efficient and reliable public transportation systems capable of moving large numbers of people. When combined with such cost-related policies, these organized transportation systems become even more valuable. For instance, reliability and punctuality are critical factors in Asia, as they are essential for the smooth functioning of society and labor markets. Meanwhile, in Europe, public transportation delays are more common and often integrated into travel time considerations, or travelers simply opt to use private cars instead. The transportation infrastructure, or built environment, is thus tailored to accommodate large volumes of users in Asia, whereas in Europe it generally operates at a much lower capacity. When comparing regions in terms of the frequency and coverage of the transportation system, users in large Asian cities are accustomed to a frequent transportation service, which may require multiple transfers to reach their destination but offers short waiting intervals. Conversely, in cities where frequent transportation is not available, low service frequency can deter users from choosing this mode of transportation altogether. Similarly, in the freight sector, Asian economies increasingly rely on integrated multimodal logistics systems, including rail and port infrastructure, to handle large volumes of goods efficiently. In such contexts, cost and infrastructure access play a different role in shaping freight mode choice than in Europe.

4.4. Research Data

The frequency of data collection methods in papers is shown in Figure 6. The data sources used are author surveys with 56 entries, followed by statistical databases with 28 entries, interviews with 18 entries, and surveys by others with 14 entries. The interview method is mainly employed in freight transport modeling, where authors conduct interviews with representatives of transport companies, e.g., using the Delphi method [3]. Bergantino et al. [81] emphasize the importance of user opinions in the development of policies. The collection of opinions is therefore crucial.

Regarding statistical databases, it is important to highlight their accessibility. The same data can, and often is, used for various studies that explore different aspects based on the same databases [12]. The use of these shared databases is beneficial for comparison, but it raises concerns about the future development of the field due to the lack of sufficient data for modeling. Babic et al. [82] highlight data sharing issues among stakeholders as a problem for modeling transportation mode choice in freight transport. De Souza et al. [83] mention challenges in obtaining data for modeling. García-Menéndez et al. [15] searched for freight transport databases for the Valencia area for their research but were unsuccessful. Iglesias and Raveau [40] used an external contractor for survey implementation, namely Facebook (now Meta). In Jin et al.’s [65] study, participants received a financial incentive of 20 yuan upon completing the survey questionnaire. Jánošíková et al. [84] point out that smart card data is more suitable for modeling than traditional passenger surveys. We agree with Spinney et al. [74] that multiple data sources are necessary to accurately recreate the real-world environment in a virtual setting. Srivastava and Sekhar [9] argue that a good database leads to better decision models.

Stated Preference (SP) surveys were used in 34 cases, while Revealed Preference (RP) surveys were used in 24 cases. RP surveys are based on respondents’ actual behavior, while SP surveys allow researchers to study preferences in hypothetical scenarios tailored to research needs. SP methods are more flexible, enable analysis of choices under different conditions, and are cost-effective [32]. A detailed presentation of the data collection methods used in the papers included in the review is provided in Appendix A.2, which outlines the type of data collection and sample size.

We classified the collected data into four main categories: statistical databases, such as Baidu Maps [85]; smart card data, including public transit smart card records, taxi GPS trajectories, and taxi trip transaction data [20]; interviews, conducted either face-to-face [86] or using pen-and-paper methods [87]; and surveys. The type of data also influences the average sample size. For instance, the study by Asgarpour et al. [3], which employed the Delphi method, was based on four expert interviews, whereas the statistical database used in the study by [67] comprised as many as 250,000,000 entries. The average sample size for a certain type of data is provided in

Table 5. Data types in choice models.

Type	Average Sample Size
Statistical Database	15,048,422
Smart Data	5,177,676
Survey Questionnaire	3788
Interview	1229

. Statistical Databases and Smart Data typically encompass a substantially larger volume of collected data compared to Survey Questionnaires or Interviews, which is also reflected in the average sample size observed across individual studies, included in our review.

4.5. Geolocation

The distribution of the studied areas across regions is presented in Figure 7. The region most frequently covered in studies is Asia. High population density in small areas has necessitated adjustments to the transportation system, integrating various modes of transport, ranging from rickshaws to high-speed trains, which enable people to cover distances in these urban environments [25,42,55,59,65]. Europe comes second, where the promotion of public transportation over private vehicles makes it crucial to study the influences on transportation mode choice.

Figure 8 presents the distribution of applied models based on the geographical location of the studies. Data are shown for MNL, NL, and MLog models, while the remaining models were grouped under “Other” due to their low frequency. If we take a closer look at the geographical distribution of the studies and the corresponding models used, we can observe that the MNL model is the predominant choice across all regions. When comparing the values from Figure 8 with those from Figure 3, we can observe that the proportions of the models used are similar. The MNL model is the most frequently used, followed by the NL and MLog models. This trend is evident in Asia, Australia, and Europe. In North and South America, the proportions between NL and MLog differ slightly, but MNL still remains the most commonly applied model. Regarding the data sources employed, smart card data is more commonly used in studies conducted in Asia [21,22,34,85,88] compared to those in Europe [84]. In Asian transport systems, smart card technology is widely adopted and already integrated into the transport infrastructure, recording a large number of trips and serving as a current and valuable source of data for research.

4.6. Specific Aspects, Examined in the Research

The research addresses the issue of transportation mode choice in a given area, where the decision-making process occurs between established modes of transport, or the study includes a scenario of introducing a new transport mode to a particular area, focusing on how the new transport mode impacts the existing transport system, e.g., Hofer and Fellendorf [47] with the example of establishing a cable car in Graz. The areas represent both smaller, well-defined urban areas, such as student campuses [49,74,89], as well as larger urban regions, like the Asian megacity of Beijing [42,55,59,65]. Some studies focus on specific aspects of the transportation process, which are presented in Appendix A.3.

The included examples of studies do not treat travel exclusively as a single event. Travel can be divided into three parts: travel from the starting point to the point of entry into the mode of transportation [90], travel with the transportation mode on the central part of the journey [68], and travel from the point where the transportation mode can take us to the end of our journey [69]. Considering this division, more studies focus on the first part of the journey than on the third, final part of the journey.

The paper by de Souza [83] focused on freight transport in Brazil, specifically the possibility of using rail instead of trucks for transport due to the risk of cargo theft. The findings indicate that the key factor preventing the use of rail is the lack of existing infrastructure. This could be illustrated with a map to highlight the areas where railway infrastructure needs to be established. As an example of a map from which the preferred mode of transportation can be identified, we are attaching one that was created using the methodology presented in the paper by Fale et al. [91]. Figure 9 shows the preferred mode of transportation for each location within the study area, based on optimal travel from that point to all other destinations within a predefined radius. The mode that most frequently proves to be the most efficient—whether in terms of cost, time, or other relevant factors—is assigned to that point. For example, if air travel is most often the optimal choice from a particular location to others within the selected radius, that point is marked as best served by air travel.

The findings of the reviewed studies can be grouped into several categories: those that provide only recommendations for planning; those that focus solely on identifying the best-performing model; those that present detailed modeling results and explain the values of most variables included in the model; and finally, those that link the modeling outcomes to a real-world problem and offer concrete proposals for action. The majority of studies provide recommendations for solving real-world problems, while the least number present results where the relationships between the variables used in the model are explained in detail. Several studies offer concrete recommendations for practice and policy-making. These include increasing the frequency of public transportation services [76,79], improving intermodal connectivity [42], establishing larger transshipment hubs [81], and promoting sustainable forms of mobility through incentives [38,80]. Detailed information is available in Appendix A.4. The summary of key findings from all studies is provided in

Table 6. Findings of the reviewed studies.

Findings	No. of Research	Example of a Reference
Recommendations for planning	33	Kurri et al. [92]
The best model is presented	20	Zhang and Xie [58]
Detailed results	4	Jung and Yoo [25]
Other	49	La Paix et al. [38]

.

5. Conclusions

The focus of the paper could alternatively have been directed toward travel behavior, activity-time patterns, route choice, the review and classification of travel data, mobility patterns, or transportation safety. However, it specifically concentrates on decision-making models and presents a systematic review of transportation choice modeling. The review investigates six research questions covering the model used to determine the optimal mode of transport, types of transportation included in the study, data used in the research, geographical area covered by the study, specific aspects examined in the research, and practical value of the research. A comprehensive search methodology across the publication database and a website is designed and used to identify 875 unique records. Titles, abstracts, introductions, and conclusions are screened for relevance, leaving us with 114 papers. An additional screening process narrowed down the selection of papers to 106 studies, used for data extraction.

In relation to the first research question, the review shows that the MNL model is the most commonly used approach for modeling transportation mode choice. While recent years have seen a growing number of studies employing ML techniques, especially for large and complex datasets, Logit-based models remain dominant due to their interpretability and widespread acceptance in the transportation research community. With regard to the second research question, the majority of passenger transport studies include Car and Bus as primary transportation modes, while the majority of freight transport studies focus on Truck, Train, and Ship. Some passenger studies also consider additional options such as Train, Bike, Metro, or Taxi, but the inclusion of such alternatives varies depending on the study context and objectives.

In addressing the third research question, it is evident that most authors collect primary data themselves, typically through surveys for passenger transport studies, in order to capture detailed information on individual travel behavior and preferences. For freight transport studies, primary data is most often collected through interviews. However, a smaller subset of studies utilizes secondary data sources, such as national travel surveys or administrative transport databases. Concerning the fourth research question, the geographical focus of the reviewed studies is uneven. Asia and Europe are the most frequently represented regions, possibly due to stronger research infrastructure, better data availability, and greater academic output in the field of transportation modeling. Other regions, such as South America, are noticeably underrepresented. In relation to the fifth research question, travel time and cost are the most frequently included variables. In passenger-oriented models, there is increasing attention to the built environment as a factor influencing mode choice. For freight-oriented models, reliability of service is highly important. This indicates a shift towards more holistic and context-sensitive approaches in transportation modeling. Finally, in response to the sixth research question, the review finds that most studies aim to address practical, real-world challenges in transportation planning and policy. Model results are commonly used to generate recommendations for improving infrastructure, enhancing transport systems, promoting multimodal transport, and supporting sustainable mobility, thereby demonstrating the applied value of this body of research.

It is essential to emphasize the practical relevance of research. Investigations in the field of transportation choice should not be limited to validating the applicability of a model on a small sample. The findings should be extrapolated to the macro level. Research should not be conducted as an end in itself but should strive to contribute to real-world understanding and improvement. In this way, the practical applicability of empirical research gains additional value, reinforcing its role in supporting evidence-based decision-making in transportation systems. Presenting research results in a graphical form, such as maps like the one shown in Figure 9, can significantly enhance their communicative impact.

5.1. Recommendations and Further Work

It is meaningful to examine how the key elements of research have evolved over time, as this can reveal the impact of emerging modeling methods and the introduction of new transportation modes on transportation mode choice modeling. Furthermore, attention should also be given to methods and procedures aimed at reducing the computational complexity of the modeling process [15]. ML models, in particular, offer enhanced predictive performance, thereby opening new possibilities for modeling transportation mode choice decisions [93]. However, Paredes et al. [18] emphasize that this improved performance is often contingent upon careful feature engineering tailored specifically for ML methods. Without such adjustments, ML models may not consistently outperform traditional approaches like the MNL model. Moreover, researchers such as Zhao et al. [34] and Paredes et al. [18] caution that despite their predictive power, ML models can be challenging to interpret and explain, which limits their practical applicability in contexts where understanding model decisions is crucial.

We recommend using Logit models as a benchmark for comparison with ML models. Variables such as time and cost should be combined with other relevant factors, particularly those increasingly important today and in the future, such as emissions. It would also be worthwhile to explore hybrid models that allow flexible selection of variables based on specific needs. This approach would bring models closer to users by enabling choices that reflect their current priorities and preferences.

We highlighted the challenges associated with data collection for mode choice modeling in both passenger and freight transport. We believe that digital twin technology can significantly assist in overcoming these challenges. Digital twins enable real-time data collection and dynamic simulation of transportation systems, allowing for more accurate, adaptive, and detailed modeling of traveler and freight behavior. By incorporating these advanced technologies, future models could better capture complex system interactions and evolving preferences, thereby enhancing prediction accuracy and decision support.

5.2. Limitations of Systematic Review

A limitation of this review is the omission of a quality assessment of the included studies. This decision was made at the outset, as the primary objective of the review was not only to map the field of transportation mode choice modeling, but above all to generate new knowledge that will serve as a foundation for future research. The experience gained through this process will contribute to the development of a model that forms part of a broader study. The goal of this model is to determine, for any given location (represented as a point object) in a certain area, whether it is more efficient to travel within a predefined radius by plane, train, or another mode of transportation.

Another limitation of this study is that the search for articles was restricted to ScienceDirect and Google Scholar. As a result, there is a possibility that relevant studies published in specialized transportation databases, such as IEEE Xplore, were excluded from the review. We could improve our source search process by introducing additional keywords/phrases. However, since this research was our first in-depth exploration of this field, we consciously accepted the time-consuming nature of this method for collecting relevant studies for our review. Now, we have the foundational knowledge that enables us to focus on a specific aspect of transportation mode choice modeling in the future.

Whilst the procedure for the review was designed to be as objective as possible, the data extraction and interpretation were carried out by the first author under the supervision of the co-authors. The review was conducted as part of the first author’s obligations within a doctoral study program, under the guidance of both the academic supervisor and the working mentor. This is according to available resources. All results and decisions have been double checked, but there may be remaining errors, which are the responsibility of the authors.