From Spontaneity to Planning: Understanding Long-Distance Travelers' Mode Choice Preferences for BlaBla Carpooling, Bus, and Train in Central Europe

Abstract

Statistics of Variables

Level Category Variables Mean Standard Deviation Min Max Dependent Variables Demand Attributes Total number of trips 1,128 952 0 29,338 Total trips with 1-day advance booking 576 452 0 15,000 Total trips with 10-day advance booking 651 501 0 14,338 Independent Variables Trip Trip Characteristics Price 13 11 1 119 Distance 261 142 91 705 Travel duration 3 2 1 14 Speed 83 19 11 143 District Demographics Middle age 42,447 16,832 1,659 5,330 Old age 11,684 7,256 307 7,852 Young age 8,338 4,312 498 7,775 Means of Transport Bicycling 369 160 8 734 Walking 2,583 436 20 3,859 Personal Car 4,753 2,924 437 12,616 Municipal public transport 12,713 6,349 259 26,094 Education University 19,505 6,523 616 37,324 Without education 204 93 4 415 Economic Working students 17,181 8,168 669 8,059 Employees 19,505 6,523 616 7,324

variables, such as the number of working students and em- separately through MMLR. For 1-day advance bookings, the ployed individuals, were considered to assess how cost sen- average number of trips is 576, while for 10-day advance sitivity and employment status affect mode choice. Inte- bookings, the average is slightly higher at 651, reflecting a grating these district-level variables provides comprehensive tendency toward short-term planning. insights into how demographic and socio-economic factors The independent variables consist of two levels: tripshape the use of carpooling, especially under different ticket level and district-level, with the former showing mean values availability scenarios. per trip across the dataset. Summary statistics indicate that Table 2 presents descriptive summary statistics for the average trip prices stand at 13 euros, with distances averagkey variables associated with BlaBlaCar pooling demand ing 261 kilometers and trip durations around 3 hours. and district-level characteristics analyzed in this study. The At the district level, four categories significantly influtotal number of trips per district averages 1,128, with a ence BlaBlaCar users’ mode choice: demographics (middleconsiderable range extending up to 29,338 trips. These trips aged population dominates at 42,447 per district), transport were further classified into two categories based on booking modes (public transport is widely used, while cycling is lead time to establish the two dependent variables, calculated less common), educational attainment (university-educated

Sorath et al.: Preprint submitted to Elsevier Page 3 of 7

BlaBla Carpooling

residents average 19,505, with lower education linked to multilevel or clustered responses, which account for the exreduced carpooling adoption), and economic status (working isting degree of dependence within these clusters (Skrondal students show a preference for cost-effective, tech-driven and Rabe-Hesketh, 2004; Hedeker, 2003), which accounts options). These factors collectively shape carpooling de- for the existing degree of dependence within these clusters mand. Overall, these metrics are expected to capture the (Hedeker, 2003). A study by Hartzel, Agresti and Caffo demographic, economic, and mobility factors influencing (2001) emphasized that ignoring this dependence in the data carpooling demand. may lead to a loss of within-cluster information regarding intraclass correlations. Despite of the extension of these models practical applications remain relatively uncommon. In this paper, we explore the multilevel multinomial logit model—a mixed Generalized Linear Model (GLM) (McCullagh, 2019)—which incorporates both linear predictors and a multinomial logit link. We define the linear predictor for each response category 𝑚 = 1, 2, … , 𝑀, cluster 𝑗, and subject 𝑖 as:

′ 𝜂𝑖𝑗(𝑚) = 𝛼 (𝑚) + 𝛽 (𝑚) 𝑥𝑖𝑗 + 𝜉𝑗(𝑚) + 𝛿𝑖𝑗(𝑚) (1)

The corresponding multinomial logit link is:

exp{𝜂𝑖𝑗(𝑚) } 𝑃 (𝑌𝑖𝑗 = 𝑚 ∣ 𝑥𝑖𝑗 , 𝜉𝑗 , 𝛿𝑖𝑗 ) = (2) (𝑙) 1+ 𝑀 ∑ 𝑙=2 exp{𝜂𝑖𝑗 } Figure 1: Loaction of trips by Pragu’s Municipal Districts where 𝑌𝑖𝑗 is the response variable for subject 𝑖 in district 𝑗, taking values from a set of four differnt mode categories {1, 2, … , 𝑀}. Here, 𝑚 = 1 is the reference category that is train, and the probabilities are defined relative to it. This model assumes independent random effects at different levels. Specifically:

• 𝜉𝑗′ = (𝜉𝑗(2) , … , 𝜉𝑗(𝑀) )′ ∼  (0, Σ𝜉 )representing unobserved heterogeneity at the district level.

• 𝛿𝑖𝑗′ = (𝛿𝑖𝑗(2) , … , 𝛿𝑖𝑗(𝑀) )′ ∼  (0, Σ𝛿 )representing scenario-specific errors.

In practice, the parameters of the district-level covariance matrix Σ𝜉 are identifiable, while those of Σ𝛿 may face empirical challenges. In Section 4, we demonstrate that the subject-level covariance parameters are not empirically identified and are therefore omitted from our application. Nevertheless, we include them in this theoretical section to Figure 2: Distribution of Demand by Pragu’s Municipal Dis- showcase the full flexibility of the model. tricts One of the key features of this model is its treatment of the Independence from Irrelevant Alternatives (IIA) property. The odds ratio between any two chosen modes 𝑚 and 2.2. The Multilevel Multinomial Logit Model 𝑙 for a given scenario 𝑖 and cluster 𝑗 depends only on the 2.2.1. The GLM Formulation corresponding linear predictors: Multinomial choice models, commonly called "discrete choice models" in econometrics (McFadden, 1972; Train, 2009), are widely used across fields like economics, so- 𝑃 (𝑌𝑖𝑗 = 𝑚 ∣ 𝑥𝑖𝑗 , 𝜉𝑗 , 𝛿𝑖𝑗 ) = exp{𝜂𝑖𝑗(𝑚) − 𝜂𝑖𝑗(𝑙) } (3) ciology, and health sciences to analyze choices between 𝑃 (𝑌𝑖𝑗 = 𝑙 ∣ 𝑥𝑖𝑗 , 𝜉𝑗 , 𝛿𝑖𝑗 ) multiple distinct options. These models help reveal factors influencing individual decisions, such as transportation This conditionally satisfies the IIA assumption. Howmodes, product preferences, or medical treatments. Multi- ever, because IIA holds only conditionally on covariates nomial choice models are further extended to incorporate and random errors, the introduction of random terms in the

Sorath et al.: Preprint submitted to Elsevier Page 4 of 7

BlaBla Carpooling

Table 3 Results of MMLR models for 1-day and 10-day advance bookings.

Category Variables 1 Day Advance Booking 10 Days Advance Booking Bus Carpooling Bus Carpooling BlaBla Carpooling characteristics Price 0.0143*** -0.1126*** 0.0294*** -0.0714*** Distance -0.0085*** 0.0039*** -0.0093*** 0.0054*** Time duration 0.8851*** 0.1589* 0.8864*** -0.1070 Demographics Middle age -0.0035 0.0002 -0.0022 0.0019** Old age 0.0014* -0.0011† 0.0011* -0.0014** Mean of Transport Cycling -0.0131 -0.0087 -0.0127 -0.0014 Walking -0.0163** -0.0158*** -0.0106*** -0.0116*** Personal Car -0.0065*** -0.0054*** -0.0049** -0.0042** Municipal public transport -0.0035* -0.0048*** -0.0017† -0.0039*** Education University -0.0021 -0.0026† -0.0001 -0.0005 No education 0.0478* -0.0788* 0.0580 -0.0522** Economic Working students 0.0139* 0.0080† 0.0025 -0.0019 Employees 0.0066 0.0042 0.0038 0.0007 Model fit parameters Variance 0.030 - 0.201 Log likelihood - -5607.759 - -4252.477 ***p < 0.001, **p < 0.01, *p < 0.05, †p < 0.1

predictors helps relax the IIA assumption, increasing model A 1 fare increase raises bus selection odds by 1.4% but flexibility. reduces carpooling by 10.6%, indicating heightened price Finally, the likelihood of the model can be written as: sensitivity for spontaneous travel. Each additional kilometer decreases bus preference by 0.8% but increases carpooling 𝑛 by 0.39%. Longer trip durations significantly enhance the 𝐽 ∏𝑗 { } appeal of both buses (142%) and carpooling (17%). ∫ ∫ ∏ 𝐿(𝜃) = 𝑃 (𝑌𝑖𝑗 ∣ 𝑥𝑖𝑗 , 𝜉𝑗 , 𝛿𝑖𝑗 )𝑓 (𝛿𝑖𝑗 )𝑑𝛿𝑖𝑗 (4) Demographic factors further influence choices. Older 𝑗=1 𝑖=1 users show a slight preference for buses, while middle- ×𝑓 (𝜉𝑗 )𝑑𝜉𝑗 aged users exhibit no notable impact. Walking and public transport usage reduce the likelihood of selecting both buses Here, the likelihood involves integrals over the random and carpooling, favoring train travel. Education levels play effects, which do not have closed-form solutions. We esa role: university-educated individuals favor trains, while timate the parameters using adaptive Gaussian quadrature, those without formal education prefer buses over carpooling. as implemented by the gllamm command in Stata (RabeEconomically, working students lean toward cost-effective Hesketh, Skrondal and Pickles, 2004). This method effecoptions like buses (1.39%) and, to a lesser extent, carpooling tively approximates the required integrals, ensuring accurate (0.8%), while employed individuals show stable preferences. parameter estimation. These trends underscore the complex interplay of trip, demographic, and economic factors in mode selection. 3. Results 3.1. 10-day advance booking model The analysis of the Multilevel Multinomial Logistic ReThe 10-day advance booking model reveals greater price gression (MMLR) models for 1-day and 10-day advance sensitivity for buses, with a 1 increase raising selection odds booking scenarios is outlined in Table 3. This table reports by 2.94% and reducing carpooling by 7.14%. Distance dethe estimated coefficients, statistical significance, and key creases bus preference by 0.93% per kilometer but increases model fit metrics such as AIC and BIC. Both models display carpooling by 0.54%, showing its appeal for longer trips. strong predictive capability, as evidenced by lower AIC Unlike the 1-day model, trip duration does not significantly and BIC values when compared to simpler multinomial affect carpooling. logistic regression models. The random intercept standard Walking and personal car use reduce preferences for both deviations for the 1-day and 10-day models, at 0.174 and bus and carpooling, favoring trains. Older users maintain a 0.449 respectively, indicate that 17% of the variance in mode slight preference for buses, while middle-aged and younger choice for 1-day bookings is attributed to between-district demographics show no significant impact. Individuals withvariations, increasing to 45% in the 10-day scenario. This out formal education strongly favor buses (5.22%) over carshift underscores the greater influence of regional demopooling (-5.22%), reflecting accessibility factors. Working graphics in longer-term planning. students prefer buses (2.5%) but show no significant shift The 1-day advance booking model highlights significant toward carpooling. The model’s higher variance and better relationships between trip characteristics and mode choice.

Sorath et al.: Preprint submitted to Elsevier Page 5 of 7