Explaining shared micromobility usage, competition and mode choice by modelling empirical data from Zurich, Switzerland

13 14 Shared micromobility services (e-scooters, bikes, e-bikes) have rapidly gained popularity in the past 15 few years, yet little is known about their usage. While most previous studies have analysed datasets 16 from single providers, only few comparative studies of two modes exist and none so-far have analysed 17 competition or mode choice at a high spatiotemporal resolution for more than two modes. To this end, 18 we develop a generally applicable methodology to model and analyse shared micromobility competition 19 and mode choice using widely accessible vehicle location data. We apply this methodology to estimate 20 the first comprehensive mode choice models between four different micromobility modes using the 21 largest and densest empirical shared micromobility dataset to-date (~169M vehicle locations collected 22 in Zurich over two months). Our results suggest that mode choice is nested and dominated by distance 23 and time of day. Docked modes are preferred for commuting. Hence, docking infrastructure for 24 currently dockless modes could be vital for bolstering micromobility as an attractive alternative to 25 private cars to tackle urban congestion during rush hours. Furthermore, our results reveal a fundamental 26 relationship between fleet density and usage. A "plateau effect" is observed with decreasing marginal 27 utility gains for increasing fleet densities. City authorities and service providers can leverage this 28 quantitative relationship to develop evidence-based micromobility regulation and optimise their fleet 29 deployment, respectively.


Introduction 36 37
2. Literature review 84 85 The number and variety of shared micromobility services have rapidly increased in recent years and 86 now includes many different modes such as docked bikes / e-bikes, dockless bikes / e-bikes and 87 dockless e-scooters. Research on shared micromobility can be categorised mainly into supply-and 88 demand-side topics, of which the latter is more relevant to this paper. Demand-side research on shared 89 micromobility tends to focus on questions such as how and why specific services are used. Demand-90 side research can be further categorised by types of factors that influence demand such as internal (i.e., 91 user socio-demographics), external (e.g., built environment, geography, weather) and trip-related 92 (destinations, distance, time of day). The latter two are most relevant to the topic of this paper and thus 93 the focus of this literature review.

95
Research analysing external and trip-related factors that influence demand for shared micromobility 96 services began with studies on station-based bikesharing (which we refer to as "docked" in this paper 97 to contrast the "dockless" alternatives) (e.g., Shaheen et al., 2011). A number of factors have since been 98 identified that influence demand for shared bikes, such as population density, workplace density, social 99 and leisure centre density, public transport density, elevation difference and weather (Bachand-Marleau  of docked bikes and e-scooter users only showed a weekday evening peak. Docked bike trips were ~0.5 165 km longer than e-scooter trips and weather was less of a disutility for dockless e-scooter users than for 166 docked bike users. The authors explain these results with the egress walk often necessary from a 167 docking station. They further conducted an initial investigation into the interaction between the two 168 modes by measuring the impact of docked bike trips on dockless e-scooter trips using a negative 169 binomial regression model. As expected, the authors found that casual usage had a small negative and 170 significant coefficient. This implies potential competition. In contrast, regular usage had a small 171 positive and significant coefficient. This implies potential complementarity.

173
We identify two gaps in the reviewed literature. Lazarus et al., 2020) with varying temporal resolution (i.e., 1-5 min scraping intervals). We thus don't 186 know how the usage compares between more than two modes and in combinations that have not been 187 explored yet (e.g., dockless e-scooters and dockless e-bikes, docked e-bikes and dockless e-bikes). 188 Cross-inference from one place to another (even within the US) is difficult as city structures and travel 189 flows vary substantially (evidence of usage peaks for dockless e-scooters in some cities but not in others 190 supports this statement, see above). We also don't know how micromobility services are used anywhere 191 outside of the US as rigorous peer-reviewed studies have not appeared yet. Thus, a comprehensive 192 comparison of many different micromobility modes (e.g., docked bikes, docked e-bikes, dockless e-193 bikes, dockless e-scooters) at high spatiotemporal resolution could improve our current (limited) 194 understanding of the similarities and differences in usage.

196
This research aims to fill these gaps by developing a generally applicable methodology to model and 197 analyse shared micromobility usage, competition and mode choice at a high spatiotemporal resolution 198 using widely accessible vehicle location data. We estimate the first comprehensive mode choice models 199 between 4 different shared micromobility modes leveraging the largest and densest empirical shared 200 micromobility dataset to-date. We collect data in Zurich, Switzerland. Zurich is the largest Swiss city with 434K inhabitants (1.5M in 207 the metropolitan area). It is one of Switzerland's economic centres and has high-quality public transport 208 with a stop within 300m of each resident in the city. The overall modal split of public transport was 209 41% walk: 26%, car: 25%, (e-)bike: 8% in the latest Swiss mobility census (2015). Several 210 micromobility providers operate in Zurich. The most established is Publibike, which offers docked 211 bikes and e-bikes at ~160 stations. Bond (formerly Smide) offers dockless e-bikes that can travel up to 212 a speed of 45 km/h. Multiple dockless e-scooter providers started operating since 2019, among them 213 Lime, Bird, Tier, Voi and Circ.

215
Our raw dataset consists of vehicle location data from 5 shared micromobility providers in Zurich, 216 Switzerland. Between 1 January and 29 February 2020, we queried each micromobility providers' API 217 every ~60s for all available vehicles, thus collecting over 169M observations. Each observation contains 218 information on a vehicle's location (GPS lon/lat), its type and model, an ID, a timestamp, the provider 219 and, for dockless providers, the battery charge. Each vehicle appears as a sequence of observations over 220 time in our dataset only when it is available to be booked. Conversely, we define a disappearance of a 221 previously observed vehicle as a trip. It is, however, necessary to remove falsely identified trips due to 222 GPS inaccuracies. Thus, the following conditions have to be satisfied for a disappearance to be 223 considered a trip: (1) the time gap is at least 2 minutes and at most 1 hour, (2) the great-circle distance 224 between the origin and the destination is at least 200 meters and at most 15 kilometres, and (3) the 225 average speed is at most 45 km/h. Overall, we obtain 168'895 micromobility trips during the two months 226 of analysis (~2'800 trips per day). 227

Validation 228 229
We validate the calculated trips against actual trips (from booking data) which we obtained for 3 of the 230 5 providers (docked e-bikes, docked bikes, dockless e-bikes) with satisfactory results. Overall, we 231 correctly identified ~95% of all trips in terms of origin/destination, weekday, time of day and duration. 232 The only bias that we detected is fewer short rides for docked e-bikes and bikes (5-12 min) and slightly 233 more longer trips (17+ min), which may be due to "trip chaining" (i.e., if a bike is both returned and 234 rented out again between two queries, the successive rides are identified as one  relative to the total number of trips per provider. The plot by time of day shows that shared bikes in 251 general (i.e., dockless e-bike, docked e-bike, docked bike) are used most during the morning and 252 evening peaks. In contrast, e-scooters exhibit a much smaller morning peak, a pronounced evening peak 253 and much higher usage frequencies at night than shared bikes (i.e., between 8 p.m. and 4 a.m.).

262
The plot by battery charge reveals that very few e-scooters and dockless e-bikes show low battery 263 charges (i.e., below 20%) at trip start. This indicates that battery charge might be a relevant criterion 264 for mode choice. E-Scooter provider #2 exhibits further peaks at 60% and 80%, which we assume to be 265 due to programming of e-scooters' battery information or charging cycles. 266 267 4. Methodology 268 269 We identify choice sets from vehicle location data and vehicle trip data as follows. For each trip, we 270 identify all vehicles available within a 2 min walking distance (167 m at 5 km/h walking speed) from 271 the departure location and within 2 min to departure time ( Figure 2). Note that micromobility trips are 272 generally short, especially those made with e-scooters. It is therefore unlikely that users are willing to 273 walk a substantially longer distance to access a vehicle. Using this method, we were able to identify competing available providers for 139'559 trips (~82.6%). 280 For each of those trips, we can thus define a choice situation, where one provider was chosen while 281 others were available. Each choice set is composed of a number (1 to 5) of available providers and 282 attributes that vary by provider. This includes the number of available vehicles per provider ("vehicle 283 density") within 2 min walking distance from the departure location, the battery charge (only available 284 for three providers), prices and whether the provider was chosen to conduct the trip. Additionally, 285   In the following, we analyse the causes behind the different choice probabilities. We begin by exploring 306 bivariate relationships between our choice attributes (cf. Table 1) and the choice probabilities (cf. Table  307 2) for each provider and mode. Subsequently, we estimate a multinomial logit model (MNL) 308 (McFadden, 1974) to explore their joint effect on mode choice. Choice behaviour could also be nested 309 as some users might only be member of certain types of shared micromobility schemes (i.e., docked 310 bikes or shared e-scooters  Note the interesting difference between these plots and the descriptive plots (Figure 1) where e-scooters 355 show a slight morning peak and a pronounced evening peak. The difference in plots stems from the 356 difference in methods. Previously (Figure 1), we calculated the share of e-scooter trips observed during 357 a particular time bin relative to the total number of e-scooter trips over all time bins. Here (Figure 3), 358 we calculate the choice probability, i.e. the number of times an e-scooter was chosen over another 359 available mode during a particular time bin relative to the total number of times an e-scooter was 360 available during a particular time bin. While the descriptive plots (Figure 1) thus only reveal shares of 361 observed trips, the bivariate plots (Figure 3) reveal preferences in choice situations.

363
The plot by distance shows that as trips get longer, the probability of choosing an e-bike (docked / 364 dockless) sharply increases while simultaneously the probability of choosing an e-scooter drops. 365 Docked bikes show a bell curve with choice probability peaking at ~2'100m and then falling with further 366 distance. The e-scooter and docked e-bike curves cross at a distance of ~650m, which can be interpreted 367 as a competitive advantage of / general preference for docked e-bikes for distances greater than 650m 368 when compared to e-scooters (without considering further factors or interaction effects). Dockless e-369 bikes and e-scooters cross at a greater distance of ~1'500m.

371
The plot by elevation shows that the choice probability for e-bikes (docked and dockless) is greater 372 with increasing absolute elevation difference. In contrast, the choice probability for docked bikes peaks 373 at the highest possible negative elevation difference (i.e., down-hill) and gradually decreases as 374 elevation rises (up-hill). E-scooter choice probability is highest in flat terrain (i.e., 0 elevation 375 difference). 376 377 Next, we explore the impact of the battery charge on choice probability. As expected, a higher battery 378 charge at departure is related to a higher choice probability. Interestingly, there is a plateau for two 379 providers (dockless e-scooter provider #2 and dockless e-bike provider) at which users are (almost) 380 indifferent to a higher battery charge. From a consumer perspective, this represents decreasing marginal 381 utility gains from increasing battery charge. For dockless e-bikes, this plateau (or "saturation point") 382 appears to begin at ~30% battery charge, while for dockless e-scooter provider #2 it appears to begin at 383 ~50% battery charge. The difference can be explained with stronger batteries and propulsion of e-bikes 384 vs e-scooters, yielding a higher resistance to choose a low-charged e-scooter that might run out of power 385 during the journey. The variation in battery charges is much higher for dockless e-scooter provider #1 386 with several outliers. While there is no behavioural explanation for different effects between two 387 dockless e-scooter providers offering the same product, we speculate the effect to be due to rebalancing 388 in high-frequency areas after recharging or different recharging practices.

390
Vehicle density is measured as the number of available vehicles of each provider within 2 min walking 391 distance of the trip departure location. The plot shows an increasing choice probability with increasing 392 vehicle density for all providers as one would expect. Both the rate (i.e., marginal utility gain) and the 393 intercept differ by mode, however. Dockless providers in particular (both e-scooters and e-bikes) gain 394 from a higher vehicle density (steepest slope), while the gain is much less pronounced for docked e-395 bikes and almost non-existent for docked bikes. Inversely, the choice probability at low vehicle density 396 is much higher for docked e-bikes and bikes than for dockless modes. This indicates differences in the 397 choice process for docked and dockless micromobility variants. Potential users might decide to take a 398 dockless e-scooter / e-bike only as they see it (visually or in their app). In contrast, the decision to take 399 a docked bike / e-bike might be decoupled from visual stimuli as usage is more habitual due to 400 knowledge about the locations of the docking stations. It could also be evidence that user groups of 401 docked and dockless modes are distinct and that users typically only register with one type of shared 402 micromobility mode.

404
We observe a plateau-effect for dockless e-scooters in vehicle density. As vehicle densities of dockless 405 modes are generally much lower than those of docked modes, we plot vehicle density again (Figure 4) 406 with a focus on lower numbers (0-30) to better illustrate this effect. Here, we can observe lower 407 marginal utility gains for docked modes than for dockless modes, and decreasing marginal utility gains 408 for e-scooters. The plateau can be interpreted as a saturation point, where higher density does not 409 increase choice probability. For dockless e-scooters, this plateau appears to begin between 10 to 15 e-410 scooters within 2 min walking distance (i.e., a circle of 167 m radius at 5 km/h walking speed). The 411 difference between the two dockless e-scooter providers could stem from different repositioning 412 practices, for example, how many vehicles are placed and how closely they are placed to each other 413 after recharging.    Micromobility mode choice is most strongly and significantly influenced by distance (positively for 437 (e-)bikes and negatively for e-scooters). The morning peak strongly and positively influences mode 438 choice for docked micromobility (e-bikes and bikes) and strongly but negatively for dockless e-439 scooters. At night, this effect reverses itself (i.e., strong and positive effect on dockless e-scooters and 440 strong and negative effect on docked (e-)bikes). This suggests that docked (e-)bikes are preferred for 441 the commute while dockless e-scooters are preferred for other trips. Dockless e-scooter providers 442 exhibit the highest utility gains from increasing vehicle densities. Elevation has a negative effect for 443 docked bikes, which is intuitive as cycling up-hill takes time and energy; and has a positive effect for 444 dockless e-bikes. Finally, increasing the price has the expected negative effect on mode choice, while 445 the relative impact of battery charge is negligible.

447
The marginal probability effects (Hensher et al., 2015) for the NECLM model ( vehicles on the road do not increase choice probability any further (cf. Figure 4). We term this 492 fundamental relationship the "plateau effect" for micromobility fleet densities. While further studies are 493 needed to understand this effect in more detail, first evidence suggests that this effect also exists at the 494 city-level (Krauss et al., 2020). Vehicle operators can start using this knowledge to optimise their 495 relocating practices, for example by balancing marginal cost and utility for a better distribution of 496 vehicles in the network. Policymakers can also use this evidence to define maximum numbers of e-497 scooters that are simultaneously allowed in certain areas of the city to prevent unnecessary blockage of 498 public space. 499 500 7. Conclusions 501 502 This is the first study that comprehensively analyses usage, competition and mode choice for four 503 different micromobility modes (dockless e-scooters, dockless e-bikes, docked e-bikes and docked 504 bikes) at a high spatiotemporal resolution. We develop a generally applicable methodology to enable 505 these analyses using only widely accessible vehicle location data, and estimate the first comprehensive 506 mode choice models using the largest and densest empirical shared micromobility dataset to-date. 507 508 Our results suggest that mode choice is nested and dominated by distance and time of day. Docked 509 modes are preferred for commuting. Hence, docking infrastructure could be vital for bolstering 510 micromobility as an attractive alternative to private cars to tackle urban congestion during rush hours. 511 Furthermore, our results reveal a fundamental relationship between fleet density and usage. A "plateau 512 effect" is observed with decreasing marginal utility gains for increasing fleet densities. City authorities 513 and service providers can leverage this quantitative relationship to develop suitable micromobility 514 regulation and optimise their fleet deployment, respectively.

516
This study has some limitations that call for future work. First, the data used in this study is limited to 517 only one city. However, our methodology is generally applicable to any city worldwide as the data used 518 is widely accessible through a variety of data collection methods such as scraping openly accessible 519 provider APIs. Therefore, similar analyses could be conducted in any other city to verify the external 520 validity of our work. Second, our analysis focuses on the impact of provider-level and trip-level 521 attributes on micromobility mode choice. This could be extended by including more modes (e.g., public 522 transport and walking), user-specific attributes (e.g., sociodemographics, mobility tool ownership, 523 micromobility service membership), and destination-specific attributes (e.g., public transport 524 availability, type of destination). Third, transport network simulation is needed to fully understand the 525 impact of micromobility on urban mobility and its sustainability. Our results on mode choice and 526 underlying user preferences build the foundation for integrating micromobility in transport network 527 simulations.

529
As the variety, availability and use of micromobility modes grow rapidly worldwide, the questions 530 addressed in this study are likely to grow in relevance. This study provides first insights that help 531 evaluate the impact of micromobility at system-level and its potential to substitute private cars, alleviate 532 road congestion during rush hours, and reduce the footprint of urban transport. Using this evidence, city 533 authorities can develop suitable regulation on critical issues such as vehicle licensing and parking space 534 allocation, and plan transport infrastructure to support its sustainable use in conjunction with other 535 modes. Service providers can evaluate their competitive positions and further optimise their operations. 536 537