A Neurobehavioral Study of Bus Crowding Valuations in Picture-Based and Immersive Choice Experiments

Abstract

9 This paper investigates how picture-based and virtual reality (VR) experiment formats influence 10 individuals’ valuation of bus crowding in stated preference settings and explores the cognitive 11 mechanisms underlying perception heterogeneity. Thirty-eight participants completed 40 choice 12 scenarios in each format. We estimate individual-level crowding multipliers, that is, the ratio of travel 13 time valuation in crowded versus uncrowded scenarios, across both formats and investigate their 14 correlations with neural biomarkers. Our analysis reveals no differences in crowding multipliers across 15 experiments. However, EEG analyses indicate distinct cognitive processes: the picture-based 16 experiment is associated with enhanced top-down control, reflecting greater reliance on internal beliefs 17 and prior experiences (elevated frontal alpha-band activity). This finding is further supported by the 18 association between crowding multipliers and self-reported willingness to avoid crowding in the 19 picture-based experiment, whereas no such association is observed in the VR-based experiment.

21 Keywords: Crowding valuation; Crowding multiplier; Virtual reality; Brain data.

22 1. INTRODUCTION

23 A growing body of evidence shows that crowding valuations vary across travelers, driven by travel 24 frequency (Aghabayk et al., 2021), trip purpose (Batley et al., 2019), personal attitudes (Márquez and 25 Poveda, 2019), and socio-demographics (Hong et al., 2024; Tirachini et al., 2017). To assess crowding 26 valuation, most studies estimate crowding multipliers, defined as the ratio of the value of travel time in 27 the crowded condition relative to the uncrowded one. Although previous studies have investigated 28 heterogeneity in crowding multipliers across travelers using both stated preference (SP) data (Márquez 29 and Poveda, 2019; Aghabayk et al., 2021; Rossetti and Daziano, 2024) and, more rarely, revealed 30 preference (RP) data (Tirachini et al., 2016; Yap et al., 2020; Bansal et al., 2022a; Kapatsila et al., 31 2025), the neurobehavioral mechanisms underlying this heterogeneity remain overlooked.

33 Prior research on crowding has mostly relied on SP experiments, which may exhibit hypothetical bias 34 and lead to deviations from real-world behavior (Mokas et al., 2021; Rid et al., 2018). Virtual reality 35 (VR) is emerging as a promising tool for enhancing the realism of choice situations, and systematic 36 comparisons between VR and other experimental formats have become a growing trend in the urban 37 and transportation planning literature (Patterson et al., 2017; Yin and Cherchi, 2024). In the crowding 38 valuation literature, only a few studies have conducted experiments in VR (Sadeghi et al., 2023a; 39 Sadeghi et al., 2023b), and they provide no conclusive evidence that VR is inherently superior to image40 based SP experiments.

42 Therefore, in this paper we address the following research questions. First, we explain the heterogeneity 43 in crowding multipliers across travelers by estimating the relationship between their crowding 44 multipliers and brain activity. Second, we test whether the type of experiment affects crowding 45 perception by comparing crowding multipliers and brain activity across similar experiments conducted 46 in VR and in a picture-based format. Finally, we test whether the relationship between crowding 47 multipliers and brain activity differs between VR and picture-based experiments.

48 2. METHODOLOGY

49 Participants and experimental design

1 Thirty-eight participants (19 males and 19 females; mean age = 23) were recruited from the National 2 University of Singapore. Sample size aligns with comparable EEG-based transportation studies 3 (Maksimenko et al., 2025). The study was approved by the institutional ethics committee (NUS-IRB4 2023-944).

6 The experiment consists of two parts. In the first part, we collected data on participants’ subjective 7 valuation of crowding using an average score derived from a five-question survey on a five-point Likert 8 scale, capturing five distinct behavioral dimensions related to crowding avoidance as stated below. In 9 the second part, participants completed a binary choice task in both picture-based format on a monitor 10 (Fig. 1A) and VR-based format (Fig. 1B), choosing between two buses based on travel time and 11 crowding level. An illustration of a choice situation is shown (see Fig. 1C).

13 Figure 1. Experimental setup for assessing crowding perception using EEG. (A) Picture14 based experiment. (B) Virtual reality-based experiment. (C) Example of a choice 15 situation (medium crowding on the left bus vs no crowding on the right bus).

17 Crowding level was shown through the density of people inside the bus and took three levels: no (0 18 pax/m²), medium (3 pax/m²), and high (6 pax/m²), consistent with Bansal et al. (2019). Travel time was 19 shown on the bus and it ranged from 20 to 40 minutes for short journeys and 40 to 60 for long journeys, 20 with five increase levels (10%, 20%, 30%, 40%, and 50%) relative to base. Each crowding-level 21 comparison (no vs medium, medium vs high) was paired with five travel-time saving scenarios for short 22 and long trips. In each choice situation, more crowded bus offered shorter travel time and less crowded 23 bus offered longer travel time. This yielded in 20 unique scenarios; each was repeated with 24 counterbalanced bus order, resulting in 40 situations per participant per format.

26 Choice scenarios were designed in Unity. The VR- and picture-based experiments utilized identical 27 static visual stimuli (screenshots from VR were shown in the picture-based experiment), ensuring that 28 any observed differences could be attributed to the level of immersiveness rather than dynamic 29 presentation. To control for order effects, half of the participants (N = 19) completed the VR block first; 30 the other half began with the picture-based block. The 40 choice situations were fully randomized within 31 each block. Between VR and picture-based blocks, participants were given a break for 5 minutes.

1 EEG data collection and analysis

2 Brain activity was recorded using multichannel EEG with 18 dry sensors arranged according to the 3 standard 10–10 system at a sampling frequency of 300 Hz, providing coverage of frontal, central, 4 parietal, temporal, and occipital regions.

6 Raw EEG signals were re-referenced to the common average, bandpass filtered between 4 and 30 Hz, 7 and cleaned using independent component analysis. EEG signals were segmented into 20-second trials 8 time-locked to scene onset.

10 EEG activity was quantified using wavelet power averaged across three frequency bands: theta (4–8 11 Hz), associated with cognitive control and mental workload (Kakkos et al., 2021); alpha (8–13 Hz), 12 shows whether participants rely more on external sensory information or on prior experience 13 (Bonnefond and Jensen, 2012; Jauk et al., 2012; Klimesch, 2012; Van Diepen et al., 2019). and beta 14 (13–30 Hz), linked to affective processing (Schubring and Schupp, 2019, 2021) and has recently been 15 used in transportation research to identify the effect of emotions on the evaluation of ethical decisions 16 involving autonomous vehicles (Maksimenko et al., 2025). Event-related spectral perturbation (ERSP) 17 was computed as the relative change in wavelet power during the choice process compared to the 418 second baseline.

19 Modelling

20 Given that each participant completed 40 choice tasks, the data exhibits a panel structure where 21 observations from the same individual are correlated. Therefore, our analysis relies on mixed 22 multinomial logit (MXL) model that explicitly accounts for intra-individual correlation and unobserved 23 heterogeneity in taste preferences (Train, 2009). We estimated MXL with 1,000 Halton draws using the 24 maximum simulated likelihood approach (Train, 2009). All models were estimated using the Apollo 25 package in R (Hess and Palma, 2019). Then, crowding multiplier is calculated by taking the ratio of the 26 marginal utility of travel time under crowded conditions to that under non-crowded conditions.

27 3. RESULTS AND DISCUSSION

28 Crowding Multipliers in VR and Picture-based Experiments

29 The distribution of crowding multipliers across VR and picture-based experiments is shown in Fig. 2, 30 showing the heterogeneity in crowding multipliers across individuals. The comparison of the ECDFs 31 of the estimated crowding multipliers under VR and picture-based experiments is shown in Fig. 3. Both 32 VR and picture-based formats exhibit increased crowding multipliers as crowding level rises, with 33 picture-based showing slightly higher crowding multiplier values. The shaded regions representing the 34 5th–95th percentile confidence intervals largely overlap between VR and picture-based at each 35 crowding level, which indicates that the observed differences in crowding multiplier distributions 36 are not statistically significant.

2 Figure 2. Histograms of crowding multipliers for each experiment type corresponding 3 to crowding level (pax/m2) using MXL model. The red dotted vertical line represents 4 MNL estimates.

2 Figure 3. Comparison of the empirical cumulative distribution function (ECDF) and 90 3 % confidence interval of crowding multipliers for each experiment type corresponding 4 to the crowding levels (pax/m2).

5 Reflection of Subjective Crowding Valuations

6 We then use Spearman’s rank correlation to assess the association between this subjective crowding 7 valuation score (see Methods) and the crowding multipliers estimated in the picture-based and VR8 based experiments (Fig. 4).

10 We observe a strong positive and statistically significant correlation between subjective valuation and 11 the crowding multipliers in the picture-based experiment (r = 0.48, p < 0.01). In contrast, the correlation 12 between these variables in the VR-based experiment is lower and not statistically significant at the 0.05 13 level (r = 0.28, p = 0.09). These results suggest crowding multiplier estimates in the picture-based 14 experiment more closely reflect the subjective valuation of crowding than the VR-based experiments.