Modeling Dockless Shared E-scooter Demand by Time of Day: A Case Study of Austin

The goal of the current study is to identify and quantify the influence of various contributing factors on dockless e-scooter demand. Drawing on high-resolution e-scooter trip level data for 2019 from Austin, Texas, we develop Census Tract (CT) level demand data for four time periods of the day. The time-period specific data is partitioned for weekdays and weekends. Using the prepared datasets, we develop a joint panel linear regression (JPLR) model framework that accommodates for the influence of unobserved factors at multiple levels – CT, month, day, and time period levels. The analysis results indicate that the proposed JPLR models outperform the independent linear regression models for both weekdays and weekends. The results also manifest a significant association between e-scooter demand and several independent variables including sociodemographic attributes, transportation infrastructure variables, land use and built environment variables, meteorological attributes, and situational attributes. Further, several panel-specific correlation effects are found to be significant across four dimensions highlighting the importance of accommodating the influence of common unobserved factors on e-scooter demand across different time-of-day dimensions. Model validation exercise results revealed that the proposed models perform well compared to the independent models. Finally, the estimated models are employed to conduct a policy exercise illustrating the value of the estimated models for understanding CT level e-scooter demand on weekdays and weekends. The results indicate that land use mix, proportion of commuters, and season are some of the most influential factors for e-scooter demand.

also allow e-scooter agencies to develop a robust rebalancing plan (to move unused e-scooters to 1 locations with higher demand). 2 3 Literature Review 4 Prior research on e-scooters can broadly be classified along three directions: (a) survey-based 5 studies of e-scooter systems, (b) comparative analysis of e-scooter and other transportation modes, 6 and (c) e-scooter trip data analysis. In this section, we present a summary of the relevant studies 7 focusing on these three dimensions. 8 With regard to the first stream of studies, earlier e-scooter research efforts followed survey 9 based approaches to investigate and understand dockless e-scooter shared systems (Almannaa et 10 al., 2021;Campisi et al., 2021;Clewlow, 2019;Nikiforiadis et al., 2021;Sanders et al., 2020). 11 Most of these studies focused on understanding perceptions of e-scooter riders and non-riders 12 (Almannaa et al., 2021;James et al., 2019), differences in e-scooter renters and owners (Laa & 13 Leth, 2020), impact of age, gender and level of education on e-scooter usage (Huang & Lin, 2019;14 Laa & Leth, 2020), relation of e-scooter with transit (Nikiforiadis et al., 2021), differences in the 15 knowledge of rules and regulations among e-scooter riders and non-riders (James et al., 2019), and 16 behavior of long term users (Huang & Lin, 2019). An extensive survey was conducted across 17 eleven major US cities, and the study found that most of the people perceived e-scooters in a 18 positive way (Clewlow, 2019). In another study, surveying employed professionals at University 19 of Arizona, the authors identified safety concerns among women (Sanders et al., 2020). 20 Within the second stream of research, a number of studies compared docked bikes and 21 dockless e-bikes or e-scooters in several US cities including Washington, D.C., San Francisco, 22 Louisville, Chicago and Austin (Almannaa et al., 2020;Guo & Zhang, 2021;Hosseinzadeh, 23 Karimpour, et al., 2021;Lazarus et al., 2020;McKenzie, 2019;Wang et al., 2022;Yang et al., 24 2021;Younes et al., 2020;Ziedan et al., 2021). In terms of the interaction between e-scooter and 25 transit modes, previously published papers suggest that public transit and scooter complemented 26 each other (Baek et al., 2021;Nawaro, 2021;Yan et al., 2021). With regard to docked bikes and 27 dockless e-bikes or e-scooters, research studies found that the main difference between the two 28 modes is that the docked shared bikes are more likely to be used for commuting (Faghih-Imani et 29 al., 2017;Faghih-Imani & Eluru, 2015) while dockless e-scooters are less likely to be used for 30 commuting (McKenzie, 2019). Moreover, average dockless e-scooter trips were longer in terms 31 of travel distance by a third and approximately twice as long in terms of travel time than average 32 docked shared bike trips (Lazarus et al., 2020). Another study in Chicago found that the average 33 travel time of scooter trips is shorter than bike trips (Yang et al., 2021). Surprisingly, earlier work 34 found that dockless shared e-scooters are less sensitive to weather conditions than docked shared 35 bikes (Younes et al., 2020). The investigation in Washington, D.C. identified potential competition 36 between e-scooter and bikeshare use for non-members while complementarity was observed for 37 members. The result is interesting and indicates occasional users choose between the modes while 38 regular members combine the mode usage to improve their accessibility needs (Younes et al., 39 2020). Other studies also compared e-bike and e-scooter usage patterns and concluded that e-bikes 40 are relatively faster than e-scooters (Almannaa et al., 2020;Nawaro, 2021). Also, temporal 41 attributes were found to be crucial factors that influence e-scooter demand (Almannaa et al., 2020). 42 In terms of data analysis approaches, several methodologies were adopted for modelling these 43 systems including descriptive analysis (McKenzie, 2019), negative binomial count models 44 (Younes et al., 2020), and multi-objective clustering algorithms (Almannaa et al., 2020). 45 The current study falls within the third stream of research. This group of research efforts 1 focused on analyzing real-world dockless shared e-scooter trip data (Bai & Jiao, 2020;Caspi et 2 al., 2020;Hawa et al., 2021;Hosseinzadeh, Algomaiah, et al., 2021;Huo et al., 2021;Li et al., 3 2022;Mehzabin Tuli et al., 2021;Noland, 2019). Previous studies in this stream of research 4 investigated the primary purpose of using e-scooter and found that these emerging mobility 5 systems are mostly used for leisure rather than for commuting purposes (Caspi et al., 2020;Noland, 6 2019). In addition, several studies found that this mode is popular for short trips and for first-and 7 last-mile connectivity (Mathew et al., 2019;Milakis et al., 2020;Shaheen et al., 2020). Analyzing 8 data from Austin, contrary to expectations, authors found that e-scooters are not employed to 9 address first-and last-mile connections, but are shifting demand from transit to e-scooter mode 10 (Zuniga- Garcia & Machemehl, 2020). Previously published studies on shared dockless e-scooters 11 found that many factors increased e-scooter demand including commercial and industrial presence, 12 population density, land use mix, access to transit, bike score, central business district locations, 13 student populated regions and weather conditions (Bai & Jiao, 2020;Caspi et al., 2020;Cheng et 14 al., 2020;Hosseinzadeh, Algomaiah, et al., 2021;Jiao & Bai, 2020). The methodological 15 approaches employed to study e-scooter data include negative binomial count models, linear mixed 16 models and spatial regression models (and variants such as spatial error and autoregressive error 17 models) (Bai & Jiao, 2020;Caspi et al., 2020;Cheng et al., 2020;Hosseinzadeh, Algomaiah, et 18 al., 2021;Huo et al., 2021;Jiao & Bai, 2020 To that extent, the current study makes twofold contributions to shared micromobility literature 24 using 2019 e-scooter trip level data from Austin. The first contribution of the study stems from our 25 recognition that the impact of independent variables varies across the day. The recognition allows 26 us to incorporate the impact of independent variables accurately. For example, higher employment 27 density might contribute to higher demand for e-scooter in the morning peak period while not 28 having a significant influence during midday. In a model examining e-scooter demand as a daily 29 variable, the variation of the parameter impact across the day is lost. In addition to time of day, we 30 also recognize that e-scooter demand profiles are likely to be different for weekdays and weekends. 31 Thus, our study develops a time-of-day model with four time periods: Morning peak (6am-11am), 32 Midday (11am-4pm), Evening peak (4pm-9pm), and Nighttime (9pm-6pm). The daily trip level 33 data is aggregated to its census tract origin for each time period separately. The aggregate time 34 period data is partitioned for weekdays and weekends 1 . 35 The second contribution of our study arises from the flexible methodology employed for 36 our analysis in data samples with high number of repeated observations. The nature of the e-scooter 37 demand data offers multiple dimensions of unobserved impacts: CT level, Time of day, CT -Time 38 of day, day of the week, spatial factors, and observation resolution. In multiple studies modeling 39 such data, researchers have adopted spatial models such as spatial error and spatial lag models 40 (Faghih-Imani & Eluru, 2016;Rahman et al., 2021). While spatial factors are quite important, in 41 the presence of large number of repetitions such as is the case in our dataset, other dimensions of 42 unobserved effects are also important. For example, in our case, our data provides for repetitions 43 of demand at the CT level by four time periods for every day in the year. In the presence of such 44 large panels, the adoption of spatial models reduces the flexibility of the model system due to the 1 inherent complexity of developing spatial models. To elaborate, it is not readily possible to 2 estimate multi-level random effects while also accommodating for the spatial unobserved effects. 3 Further, as the size of the panel (repeated measure per CT) increases, estimating and interpreting 4 spatial models are not straightforward. Resorting to spatial model development will restrict the 5 model system to considering spatial unobserved factors while not considering for the presence of 6 multi-level unobserved dependencies identified. Towards addressing these challenges, in this 7 study, a viable middle ground is considered. Specifically, a multi-level mixed linear regression 8 framework that offers flexibility in accommodating for several types of unobserved dependencies 9 such as CT level, CT-Time of the day, day of the week and observation level is developed. The 10 mixed linear regression model framework is developed separately for weekdays and weekends 11 using an extensive set of independent variables including sociodemographic attributes, 12 transportation infrastructure variables, land use and built environment variables, meteorological 13 attributes, and situational attributes 2 . The performance of the estimated model is validated using a 14 holdout sample. A policy analysis is conducted to illustrate the applicability of the proposed model 15 system. 16 The rest of the paper is organized as follows: Section 3 presents data processing procedures 17 and summarizes the data employed for model estimation. Section 4 provides a discussion of the 18 econometric models employed in this study. The results from the models are discussed in Section 19 5. Section 6 presents model validation, and Section 7 presents policy analysis. Finally, the 20 conclusion section summarizes the findings and concludes the paper. The major focus of this study is to examine aggregate level e-scooter demand at a census tract 33 level across different times of the day for weekdays and weekends. Before aggregating the data at 34 a census tract level by time of day, the following steps were followed to process the trip level e-35 scooter data. First, e-scooter trip records with missing information were deleted (approximately 36 730 records). Second, to avoid including inaccurate or incorrect data in the analysis, we consider 37 the City of Austin official trips report criteria. Therefore, we delete any trips that do not meet the 38 following criteria: 39 ▪ Trip distance greater than or equal to .1 miles and less than 500 miles 40 ▪ Trip duration less than 24 hours 41 After applying the above-mentioned criteria around 600 thousand trips were deleted. Third, the 1 data was processed to eliminate CTs with very small number of records. Among the 265 CTs, 48 2 CTs account for 99.2% of total trips. For our analysis, we selected trips from these 48 CTs. Finally, 3 after cleaning the database based on the abovementioned criterion, the final e-scooter database had 4 approximately 4.98 million trips. The spatial distribution of the yearly e-scooter trips originating 5 in the selected 48 census tracts for the year 2019 is presented in Figure 1. From Figure 1, it is 6 evident that most of the e-scooter trips started near the city's center in close proximity to downtown 7 Austin and the University of Texas Campus. The time-of-day distribution of the yearly e-scooter 8 trip patterns are presented in Figure 2. From Figure 2, it can be observed that there are significant 9 differences in e-scooter demand across different times of the day. Furthermore, it is clear that e-10 scooter usage is considerably higher during midday and evening periods compared to morning and 11 nighttime periods. Therefore, in developing the e-scooter trip demand model, we consider four 12 time periods-Morning peak (6am-11am), Midday (11am-4pm), Evening peak (4pm-9pm), and 13 Nighttime (9pm-6pm). Further, to explore the trip patterns across different day-of-week, the day 14 specific trip distributions for the year 2019 are plotted in Figure 3. Figure 3 provides a 15 representation of e-scooter trips for weekdays and weekends. Figure 3 demonstrates that e-scooter 16 demand pattern remains stable across the weekdays (Monday-Friday) but varies on weekends 17 (Saturday-Sunday). Hence, we consider splitting the data into weekday and weekend samples for 18 each time period. Consequently, the e-scooter trips are aggregated by different times of day (4) 19 and days-of week (2) at the census tract level resulting in 8 dependent variables. 20 21 [ Figure 1 near here] 22 23 [ Figure 2 near here] 24 25 [ Figure 3 near here] 26 27 28 To obtain a reasonable sample for estimation purposes from the abovementioned samples, 29 we randomly select, for each census tract, 40 weekdays and 20 weekend days. Therefore, for 30 weekday samples we have 1920 records [48*40], while weekend samples resulted in 920 [48*20] 31 records. The descriptive stats of the dependent variables are presented in the first-row panel of 32 Table 1. The data compilation procedure including dependent and independent variables are 33 presented in Figure 4 for weekdays and weekends. 34 The independent variables considered in this study can broadly be categorized as: 1) 38 Sociodemographic attributes, 2) Land use and Built environment attributes, 3) Transportation 39 infrastructure attributes, 4) Meteorological variables, and 5) Situational attributes. The 40 sociodemographic, land use and built environment, transport infrastructure attributes are computed 41 at census tract level. The meteorological variables are generated specific to the time-of-day and 42 day-of week for which the e-scooter demand is computed. 43 The sociodemographic attributes include population density, employment density, the 1 proportion of students, the proportion of females, proportion of commuters, proportion of 2 commuters by mode (drive, carpool, public transport, walk and other modes) and median income. 3 Several land use and built environment variables are considered including the density of the single-4 family area, density of the multi-family area, density of commercial area (mixed-use houses, retail, 5 and wholesale), the density of office area, density of the industrial area, density of educational area 6 (colleges, universities, primary and secondary school), density of parking area (parking garage, 7 and parking lots), and density of parks and open space area, the density of other land-use areas 8 (cultural services, hospitals, utilities) and historic landmarks. Finally, land use mix is computed 9 as where is the category of land-use, is the proportion of 10 the developed land area devoted to a specific land-use, is the number of land-use categories in 11 a census tract. 12 The census tract level transportation infrastructure attributes include bus station density 13 (capturing the influence of availability of public transit on e-scooter usage), sidewalk density, bike 14 road density, major street density, and minor street density. The meteorological variables include 15 precipitation, humidity, and average temperature. Situational attributes include the day of the week 16 and seasons. A summary of the independent variables generated for our analysis are included in 17 Table 1. The reader would note that several functional forms such as logarithm and standardized 18 z-score were considered in our model estimation process. The functional form that offered the 19 most intuitive fit was retained in the model. This section presents the econometric framework for the JPLR model (see Rahman, 2018 for 26 similar approach). Let us assume that q (q = 1, 2, …, Q=48) be an index to represent census tracts, 27 t (t = 1, 2, 3, …, T=40 for weekdays and 20 for weekends) represents the different days, and r (r 28 =1, 2, …, R=4) represents different times of the day. Let, represents the observed log-linear 29 demand in census tract q, on day t and during time period r. Thus, the equation for modeling e-30 scooter demand can be written as: where, * is the predicted demand for census tract q, for day t and time period r. is 34 a matrix of attributes that influence e-scooter demand (including a scalar constant); is the vector 35 of coefficients corresponding to the attributes for the time of day and is a vector of 36 unobserved factors moderating the influence of corresponding element in in time of day 37 dimension, . Further, is an idiosyncratic random error term assumed to be independently 38 normally distributed with variance 2 . 39 represents the vector of coefficients representing the impact of common unobserved 40 factors that jointly influence e-scooter demand at different time periods across repetition level k. 41 As discussed earlier, in the current study context, we estimate for different levels (k) of 42 repetition measures including census tract, census tract-time of the day, day of the week and 43 observation level. In accommodating unobserved effects at different levels, random numbers are 1 assigned to the appropriate observations of the repetition measures. For example, we have a total 2 of 48 census tracts in the estimation set. Thus, in evaluating unobserved effect at the census tract 3 level, 48 sets of different random numbers are generated specific to each census tract and assigned 4 to the data records based on their census tract ID. Similarly, the census tract-time of the day level 5 repetition measure represents unobserved effects across different combination of census tracts and 6 time periods. Thus, the census tract-time of the day combination has a total of 192 (48 census 7 tracts*4 times of the day) records. For evaluating the unobserved effect at the census tract-time of 8 the day, 192 sets of different random numbers are generated and assigned to the data records based 9 on their census tract-TOD combinations. For other combinations considered, the random number 10 are generated and assigned following a similar process. 11 To complete the model structure of the equations (1), it is necessary to define the structure 12 for the unobserved vectors and . In this paper, we assume that these vectors are independent 13 realizations from normal distributions as follows: ~(0, 2 ) and ~(0, 2 ). 14 With these assumptions, the probability expressions for the observed demand may be 15 derived. Conditional on and the probability for census tract q to have e-scooter demand 16 in day t and time period r is given by: 17 18 where ϕ(.) is the standard normal probability distribution function. 19 The complete set of parameters to be estimated in the multivariate model system of 20 equations (2) are vector and the following standard error terms: and . Let Ω represent a 21 vector that includes all the standard error parameters to be estimated. Given these assumptions the 22 joint likelihood for e-scooter demand at four time periods for day-of-week (weekdays/weekends) 23 is provided as follows: 24 Finally, the unconditional likelihood function may be computed for census tract q as: 26 Now, we can express the log-likelihood function as follows: 28 29 The log-likelihood function in Equation (5)  presented in Figure 5 and Figure 6 for weekday morning peak and evening peak. In terms of the 24 sum of squared error (SSE), our model results indicate that adding variables gradually reduces SSE 25 of the updated models. 26 27 [ Figure 5 near here] 28 [ Figure 6 near here] 29 30

31
The results of the JPLR models for weekdays and weekends are presented in Table 2 and Table 3,  32 respectively. The final specification of the model development was based on removing the 33 statistically insignificant variables in a systematic process based on statistical confidence (90% 34 confidence level). The model estimation process followed scientific approach to model estimation. 35 We added the independent variables one at a time and estimated the model. After adding all the 36 variables, we examined the significance of all the variables in the model and dropped insignificant 37 variables one by one. For example, the variable with the lowest t statistic was dropped and the 38 model was re-estimated. The process was continued until no variables were insignificant. The 39 reader would note that potential correlation between the various independent variables were 40 carefully considered prior to model estimation. The variables that exhibited higher correlation 41 values were considered separately and the variable that offered the better fit was retained (while 42 excluding other correlated variables). The specification process was also guided by prior research 43 and parsimony considerations 3 . In estimating the models, several functional forms and variable 1 specifications are explored. The functional form that provided the best result is used for the final 2 model specification. In the estimated models, a positive (negative) coefficient corresponds to 3 increase (decrease) in e-scooter demand. Please note that only the results for weekdays are 4 described in detail for the sake of brevity. 5

Joint Panel Linear Regression Model for Weekdays 6
The estimation results of the joint model for weekdays are presented in Table 2. In the joint system, 7 the demand components for morning peak, midday, evening peak and nighttime are presented in 8 the second, third, fourth and fifth column panels of Table 2, respectively. The estimation results 9 of these components are discussed in the following sections by variable groups. 10

Sociodemographic Attributes 11
Several sociodemographic attributes at the census tract level are considered in the model. 12 Surprisingly, population density variable has a negative coefficient in morning peak, midday, and 13 evening peak for weekdays. The results imply that the e-scooter demand during weekdays is likely 14 to be less in the census tracts with higher population density. The variable also exhibits significant 15 variation across all time periods as indicated by the random parameter estimated for population 16 density. So, while the average impact might indicate lower demand with increasing population, 17 there is significant variability across census tracts. The reader would note that we retained the same 18 distribution variance across all time periods for maintaining a parsimonious specification. On the 19 other hand, employment density in a census tract is found to increase e-scooter demand at all times 20 (see (Caspi et al., 2020;Jiao & Bai, 2020) for similar findings). The results indicate that as the 21 proportion of females in the CT population increases, there is a reduction in e-scooter demand in 22 morning peak and nighttime. The result might reflect the lower exposure to e-scooters and/or safety 23 concerns among women. The proportion of students affects e-scooter demand positively across all 24 time periods. Thus, it is evident from the results that the e-scooter demand is likely to be higher in 25 census tract for specific cohorts of population rather than across all population categories in a 26 census tract. 27 The increase in proportion of commuters is likely to increase e-scooter demand across all 28 time periods. The proportion of commuters using public transit is found to affect e-scooter demand 29 negatively in all four time periods. Different trends by mode for commuters are perhaps alluding 30 to the competition between e-scooter and public transportation mode for commuting (see (Zuniga-31 Garcia & Machemehl, 2020) for a similar finding found to be significant influencers of e-scooter demand. The density of office has negative 39 association with e-scooter demand across the day. In the midday and evening peak demand 40 components, the e-scooter demand is found to be positively associated with higher density of 41 commercial area, while density of commercial area is not significant in the demand components 1 for morning peak and nighttime as most of the stores are closed in this time of the day. The density 2 of educational area is found to be negatively associated with e-scooter demand during morning 3 peak and nighttime periods. The result is to be viewed in conjunction with the proportion of 4 students' variable. When we consider the net values of proportion of student and density of 5 educational area in the census tract, the net result yields a positive value. The results reveal that 6 parks and open space, and other land use (cultural services, hospitals, utilities) areas in the census 7 tracts are likely to attract more e-scooter riders. 8 To test the relationship between land use diversity and e-scooter demand, we also consider 9 land-use mix as independent variable in the demand components. The results in Table 2 for 10 weekdays reveal that, land use mix is significant and positive across all time periods (see (Huo et  11 al., 2021) for a similar finding) . The results support the positive influence of diversified land use 12 that encourages an active and livable community. Given that the presence of historical landmarks 13 is a surrogate for recreational activity presence, it is not surprising that they are likely to encourage 14 e-scooter demand across all four time periods. 15

Transportation Infrastructure Attributes 16
Among different transportation infrastructure attributes considered, the effect of bus stop density, 17 rail and metro density, sidewalk density, and bike route density are found to be significant 18 indicators of e-scooter demand for weekdays. While proportion of commuters using public transit 19 affects scooter demand negatively, the bus stop density, rail and metro density are positively 20 associated with higher scooter demand. Hence, the results suggest that e-scooter may have a 21 complex relationship with public transit switching from competition to complementarity across 22 the region and by time of day (see (Yan et al., 2021) for a similar finding). Rail and metro density 23 is closely aligned with increasing e-scooter demand. E-scooter clearly serves as a fist-and last-24 mile connector for rail and metro alternatives. Higher level of sidewalk density and bike route 25 density reflect good infrastructure for riding e-scooter, possibly leading to higher demand. 26 27 Meteorological Attributes 28 Among meteorological attributes considered, precipitation, humidity, and temperature are found 29 to be significant determinants. Precipitation is found to contribute towards lower e-scooter demand 30 during midday and evening peak periods (see (Noland, 2021) for a similar finding). Humidity has 31 a negative coefficient across the time of the day (other than nighttime) indicating that with 32 increasing humidity, the likelihood of e-scooter ridership decreases, perhaps an indication of 33 discomfort resulting from higher humidity. E-scooter demand is found to be higher for the 34 weekdays with temperature higher than 15°C. The temperature>30°C does not have effect on the 35 morning peak and nighttime dimensions. The result may indicate the fact that e-scooter users are 36 likely to be more sensitive to cold weather (see (Noland, 2021) for similar finding). 37

Situational Attributes 38
With regard to seasons, spring is found to be associated with higher e-scooter demand for all time 39 periods. Fall is associated with increased e-scooter demand in morning peak and decreased e-40 scooter demand in the nighttime. With regard to different weekdays, the indicator for Tuesday and 41 Wednesday is found to have significant impact in midday, evening peak and nighttime demand 42 models. The indicator has a negative coefficient revealing that Tuesday and Wednesday are 43 associated with reduced e-scooter demand. Thursday is also associated with lower demand for 44 midday and nighttime periods. The results provide support to our hypothesis that variable impacts 1 vary by time period. 2

Panel Correlation Effects 3
In the joint panel model for weekdays, we consider several panel-specific (census tract, census 4 tract-time of the day, day of the week and observation level) correlation effects across four 5 dimensions. Among the different panel level parameters, two parameters were found to be 6 significant. These include (a) common unobserved factors at the CT panel level across all time 7 periods, and (b) CTnormalized population density (discussed earlier in Sociodemographic 8 attributes section). Overall, the results clearly highlight the importance of accommodating for the 9 common unobserved factors influencing e-scooter demand across different time-of-day 10 dimensions. 11 [ Table 2 near here] 12 [ performance of the models were compared using the Bayesian Information Criterion (BIC). The 21 results from the exercise are presented in Figure 7. Policy Analysis 30 The model specifications in Table 2 and Table 3 demonstrate how parameters affect e-scooter  31 demand. To further illustrate the applicability of the models developed, we perform an elasticity 32 analysis to identify the magnitude of the impacts of the independent variables. To evaluate the 33 impact of exogenous variables on e-scooter demand, we consider changes in aggregate scooter 34 demand in response to a 15 and 25 percent change in independent variables. In this research, we 35 perform elasticity analysis considering a selected set of significant factors. The results of elasticity 36 analysis for weekdays are illustrated in Figure 8 while the results of elasticity analysis for 37 weekends are shown in Figure 9. Regarding the weekday model components, we found proportion 38 of commuters, land use mix, proportion of other land use to be the significant factors that influence 39 the e-scooter demand positively for weekdays. Proportion of transit commuter and density of office 40 area are the most significant factors found to influence the demand negatively. In contrast, weather 41 factors are found to have the least influence on e-scooter demand. For weekend model components, 42 land use mix, density of other land use, medium temperature and proportion of commuters using 43 public transit are the most influential variables for e-scooter demand. 44 1 [ Figure 8 near here] 2 [ Figure 9 near here] 3 4

5
The current study contributes to our understanding of dockless e-scooter systems by identifying 6 and quantifying the influence of various contributing factors on dockless e-scooter demand. The 7 study recognizes the significant variation of e-scooter usage patterns across different time periods 8 and weekday/weekend. The study employs high-resolution spatiotemporal e-scooter trip level data 9 from Austin, Texas to generate census tract (CT) level e-scooter demand by time period (Morning,  10 Midday, Evening, Nighttime) separately for weekdays and weekends. 11 As data generated is available for multiple observations per CT (by day and time period), 12 the study develops a framework that accommodates for the influence of unobserved factors at 13 multiple resolutions including CT level unobserved factors, time period level unobserved factors, 14 and potential variation in the influence of various attributes (random parameters have affected e-scooter demand. 42

Declaration of Interest Statement 43
There is no competing interest to declare. 44 The authors would like to thank Dr. Bibhas Kumar Dey for initial discussions on the idea of the 2 paper. Also, the authors would like to acknowledge the City of Austin for providing access to their 3 datasets. 4 Author Contribution Statement 5 The authors confirm contribution to the paper as follows: study conceptualization and design: 6 Naveen Eluru, Nami Alsulami; data collection: Nami Alsulami, Sudipta Dey Tirtha, Naveen Eluru; 7 model estimation: Nami Alsulami, Sudipta Dey Tirtha, Shamsunnahar Yasmin, Naveen Eluru; 8 analysis and interpretation of results: Nami Alsulami, Sudipta Dey Tirtha, Shamsunnahar Yasmin, 9 Naveen Eluru; draft manuscript preparation: Nami Alsulami, Shamsunnahar Yasmin, Sudipta Dey 10 Tirtha, Naveen Eluru. All authors reviewed the results and approved the final version of the 11 manuscript. 12 Urban/University Environment.