Uncovering the spatio-temporal impact of the COVID-19 pandemic on shared e-scooter usage: A spatial panel model

Abstract

This study examines the spatio-temporal effects of the COVID-19 pandemic on shared e-scooter usage by leveraging two years (2019 and 2020) of daily shared micromobility data from Austin, Texas. We employed a series of random effects spatial-autoregressive model with a spatially autocorrelated error (SAC) to examine the differences and similarities in determinants of e-scooter usage during regular and pandemic periods and to identify factors contributing to the changes in e-scooter use during the Pandemic. Model results provided strong evidence of spatial autocorrelation in the e-scooter trip data and found a spatial negative spillover effect in the 2020 model. The key findings are: i) while the daily e-scooter trips reduced, the average trip distance and the average trip duration increased during the Pandemic; ii) the central part of Austin city experienced a major decrease in e-scooter usage during the Pandemic compared to other parts of Austin; iii) areas with low median income and higher number of available e-scooter devices experienced a smaller decrease in daily total e-scooter trips, trip distance, and trip duration during the Pandemic while the opposite result was found in areas with higher public transportation services. The results of this study provide policymakers with a timely understanding of the changes in shared e-scooter usage during the Pandemic, which can help redesign and revive the shared micromobility market in the post-pandemic era.

1. Introduction

The global COVID-19 pandemic has had a significant impact on transportation mobility worldwide, resulting in travel restrictions such as social distancing, bans on gatherings, stay-at-home orders, and work-from-home policies. These measures were put in place to limit viral transmission and consequently affected all transportation sectors (Menon et al., 2020, De Vos, 2020). As part of the transportation system, shared micromobility services, particularly shared e-scooters, were impacted similarly. Since their emergence, shared e-scooters have experienced rapid expansion throughout the US, with an average adoption rate of 3.6% and approximately 38.5 million trips generated in 2018 alone (NACTO, 2019). However, they were hit hard by the Pandemic, with half of the systems suspending operations and 12% facing permanent closure (US Department of Transportation, 2021). Only 37% remained in operation across different cities in the US from March to December 2020, and the impact of the Pandemic on shared e-scooters varied based on the size of the city and the nature of the spread of COVID-19 (Padmanabhan et al., 2021).

It is crucial to understand the impact of this unprecedented event on shared e-scooters, as this knowledge is necessary for capturing the demand of a new user market amidst the Pandemic. While there is a growing number of studies (e.g., Yan et al., 2021, Dean and Zuniga-Garcia, 2022) exploring the Pandemic's impact on shared e-scooter usage, there are still research gaps that need to be addressed. First, one of the methodological limitations of the existing studies is that they did not control for spatial and serial autocorrelation that could be present in the aggregate level (census tract) e-scooter trip data. As a result, the impacts were either underestimated or overestimated, as ignoring the presence of serial and spatial autocorrelation in the data could lead to inconsistent and biased estimates (Hausman et al.,1984; Lesage and Pace, 2009). Second, existing studies did not incorporate socio-demographic, built-environment, e-scooter availability, and spatio-temporal variables with COVID-19 policies in their models to capture the Pandemic’s impact on the usage of shared e-scooters.

Therefore, this study aims to fill these research gaps by answering three research questions. First, it seeks to understand how the e-scooter usage pattern changed spatio-temporally during the Pandemic. Second, it examines the differences and similarities in determinants of e-scooter usage during normal and pandemic periods. Third, it identifies the factors that contributed to the spatio-temporal changes in e-scooter usage during the Pandemic. To answer these questions, the study analyzed census tract-level daily e-scooter pick-ups and changes in Austin, Texas, using detailed descriptive analysis and random effect spatial panel models. Austin is one of the top 20 largest cities in the US, and after reporting the first COVID-19 case in Austin on March 13, 2020, the city authority took several actions and policies to curb the spread of the virus (Limón, 2020, Villalpando, 2021, Texas Department of State Health Services, n.d..).

This study will help future research by identifying the impact of COVID-19-related policy variables on the changes in trip demand during a pandemic. Overall, the outcome of this study will help policymakers, transport planners, and other concerned authorities to predict the e-scooter demand pattern in the future and make informed decisions to promote shared e-scooter as an alternative travel option in the post-pandemic era.

The remainder of the paper is organized as follows. Section 2 reviews recent literature related to changes in shared e-scooter usage during the Pandemic. An overview of the Austin Data portal, data from several other sources, and the random effects spatial autoregressive model approach is illustrated in Section 3. Section 4 presents the empirical results from the shared-e-scooter usage models. Section 5 discusses the results and its implications. Finally, the paper concludes (Section 6) with a summary of key findings, limitations, and directions for future research.

2. Literature review

There is a growing body of literature on the impact of the COVID-19 pandemic on shared e-scooter usage (Table 1 ). These studies are diverse in terms of the study area (Europe, Asia, and North America), modeling techniques, and mixed findings. For example, in the United States, Yan et al. (2021) reported that e-scooters in Washington DC mostly served trips with longer distances and duration during the Pandemic. Though the City of Austin experienced a decrease in the number of dockless e-scooters available during the first stage of the Pandemic as several companies stopped operating (Egan, 2020), Azimian and Jiao (2023) observed an increase in trip duration for dockless e-scooters in Austin. Dias et al. (2021) found that shared e-scooters mostly served short-distance trip demands in cities in Europe and US. Li et al. (2021) compared the usage variations of bike-share and shared e-scooters (both docked and dockless) in Zurich, Switzerland, before and during the Pandemic. Results showed that e-scooter services faced fewer changes than bike-share during the Pandemic. Table 2 .

Table 1.

Summary of Studies Related to the Impact of COVID-19 Pandemic on Shared E-scooters.

Authors (year published)	Data and Location [Year].	Methods/Models	Dependent Variable	Independent Variables	Key findings
Almanaa et al. (2021)	Survey on a sample of 439 respondents in Riyadh, Saudi Arabia [April 2020].	Binary Logistic regression model.	Willingness to use e-scooter in the future (Yes/No).	Socio-demographic variables: age, gender, income, car ownership. Trip characteristics variables: trip purpose, primary mode, purpose etc.	Gender, age, and experience of using ride-hailing services influenced peoples' willingness to use e-scooters.
Dias et al. (2021)	Cities in the US and Europe	Review paper.	–	–	Shared e-scooters mostly served short distance trip demands.
Glavi'c et al. (2021)	Survey conducted on a sample size of 1143 respondents in Belgrade, Europe, [July to September 2020].	Ordinal Logistic regression model.	Willingness to use e-scooter in the future (Yes/No/May be).	Socio-demographic variables: age, gender, income, car ownership. Trip characteristics variables: trip purpose, primary mode, purpose etc.	Available infrastructure, average traveled distance, environmental benefits, congestion, safety concerns had an impact on users' willingness to switch to e-scooters.
Yan et al. (2021)	Washington DC, 2020. [The week of July 15–21, 2019 and the week of June 15–21, 2020]	Spatiotemporal analysis (kernel density approach).	–	–	Service areas of e-scooters overlapped with the service areas of transit and bikeshare; shared e-scooters increased mobility in underserved neighborhoods.
Dean & Zuniga-Garcia (2022)	Austin, TX. [Feb 2019-June 2020]	Negative binomial models.	Daily e-Scooter Trip Counts.	Socio-demographic variables: household size, median income etc., built-environment variables: population density, employment density, residential density, etc.; road characteristics: number of lanes, shoulder width, etc.	Road segments with sidewalks, bus stops, lower vehicle miles traveled, and fewer lanes had more trips during the Pandemic than in the pre-pandemic period; the temporal usage patterns of shared e-scooter remained unchanged during the Pandemic.
Li et al. (2022)	Data from 30 European cities. [April-July 2021]	Construction of a trip frequency signature for each city.	–	–	Cities showed similarities and differences in usage of shared e-scooter; utilization efficiency had an association with the number of e-scooters per person and per unit area.

Table 2.

Model Variables and Summary Statistics.

Variables	Units	Source	Expected Relation	Time Variant?	Mean	Std. Dev.	Min	Max
Dependent Variables
Number of average daily pick-up trips in each census tract in 2019	Count	CAD1	N/A	Yes	65.8	510.8	0	27,330
Number of average daily pick-up trips in each census tract in 2020	Count	CAD	N/A	Yes	22.4	213.2	0.0	10,284
Δ Daily trips count	Count	CAD	N/A	Yes	−53.4	325.1	−5074.5	4267.0
Δ Daily trip miles	Count	CAD	N/A	Yes	−49.2	286.6	−4574.2	5152.2
Δ Daily trip duration	Count	CAD	N/A	Yes	−615.8	3871.5	−63464	58259.6
Explanatory Variables
Temporal Variables
Average temperature in 2019	oF	Wundergroun2	?	Yes	68.7	14.6	34.0	89.1
Average temperature in 2020	oF	Wunderground	?	Yes	69.3	13.0	39.7	90.4
Total precipitation in 2019	Inch	Wunderground	–	Yes	0.08	0.3	0.0	3.9
Total precipitation in 2020	Inch	Wunderground	–	Yes	0.1	0.4	0.0	3.5
Average wind speed in 2019	mph	Wunderground	–	Yes	8.2	3.6	1.6	18.9
Average wind speed in 2020	mph	Wunderground	–	Yes	7.8	3.4	1.1	17.9
Average Humidity in 2019	%	Wunderground	–	Yes	69.5	11.5	28.2	95.4
Average Humidity in 2020	%	Wunderground	–	Yes	71.0	11.9	31.0	94.4
Binary 1: if weekend in 2019	–	CAD	+	Yes	0.3	0.5	0.0	1.0
Binary 1: if weekend in 2020	–	CAD	+	Yes	0.3	0.5	0.0	1.0
COVID-19 Features
No. of daily new cases	Count	CAD	–	Yes	141.5	163.7	0.0	753.0
COVID-19 Policy		DSHS3	?
Pre-COVID: Binary 1, if the date corresponds to Pre-COVID phase	–			Yes	0.2	0.4	0.0	1.0
Pre-Lockdown: Binary 1, if the date corresponds to Pre-lockdown phase	–			Yes	0.0	0.1	0.0	1.0
Lockdown (baseline): Binary 1, if the date corresponds to state Lockdown	–			Yes	0.23	0.45	0.0	1.0
Reopen I: Binary 1, if the date corresponds to state Reopen I	–			Yes	0.2	0.4	0.0	1.0
Reopen II: Binary 1, if the date corresponds to state Reopen II	–			Yes	0.3	0.4	0.0	1.0
Socio-demographic and Commuting Characteristics
Median age	years	ACS4	?	No	34.4	5.6	0.0	49.4
Median income	$	ACS
Binary 1: Low median income	–		?	No	0.3	0.4	0.0	1.0
Built Environment Variables
Population density	Pl/sqml	SLD5	+	No	8.2	6.9	0.0	50.4
Land-use mix	–	SLD	+	No	0.6	0.2	0.0	0.9
Transit Supply Features
Binary 1: if a tract has 0 bus stops	–	CAD	?	No	0.3	0.4	0.0	1.0
Binary 1: if a tract has 1–5 bus stops	–	CAD	?	No	0.1	0.3	0.0	1.0
Binary 1: if a tract has 6–10 bus stops		CAD	?	No	0.1	0.9	0.0	1.0
Binary 1: if a tract has 11–20 bus stops	–	CAD	?	No	0.2	0.4	0.0	1.0
Binary 1: if a tract has more than 20 bus stops	–	CAD	?	No	0.1	0.3	0.0	1.0
E-Scooter Availability(in thousand)
# of devices in each tract in each month in 2019	Count	CAD	+	No	0.55	1.89	0.0	18.61
# of devices in each tract in each month in 2019	Count	CAD	+	No	0.2	0.8	0.0	9.3
Δ (# of devices in each tract in each month between 2020 and 2019)		CAD	?	No	−0.4	1.3	−13.6	0.043

Notes: 1. CAD = City of Austin Dashboard, 2. wunderground = https://www.weatherunderground.com; 3. DSHS: Texas Department of State Health Services Department; 4. ACS = American Community Survey, 5. SLD = Smart Location Database (https://www.epa.gov/smartgrowth/smart-location-mapping#SLD).

A handful of studies also explored the factors associated with shared e-scooter usage during the Pandemic. Rahimi et al. (2021) developed an ordered probit model to explore the usage of shared e-scooters during the Pandemic by analyzing data collected from shared e-scooter companies operating in Chicago. They found that households with zero- or one-vehicle ownership, income less than $50,000, and young adults (aged 18 to 35) increased their e-scooter trips during the Pandemic. More recently, Dean & Zuniga-Garcia (2022) employed two negative binomial models using route trajectory data from February 2019 to June 2020 in the pre-and during-pandemic periods in Austin, TX. They found that roadway choices between the pre-and during-pandemic periods. Results showed that road segments with sidewalks, bus stops, lower vehicle miles, and fewer lanes had more trips during the Pandemic than in the pre-pandemic period. They also reported that the temporal usage patterns of shared e-scooters remained unchanged during the Pandemic. Li et al. (2022) explored similarities and differences among shared e-scooter services by using trip frequency data from 30 European cities during the post-COVID-19 Pandemic. They grouped them into 7 clusters using a bottom-up hierarchical clustering method. Results showed that cities had similar temporal usage patterns of shared e-scooters within each cluster. Also, cities in the same country were more likely to have a similar temporal usage pattern of shared e-scooters.

The impact of the Pandemic on peoples' willingness to use shared e-scooter services in the post-pandemic era also received attention in some studies. Almannaa et al. (2021) conducted a study based on a sample of 439 responses in Riyadh, Saudi Arabia, to explore the willingness to use shared e-scooters if available after the COVID-19 pandemic. They employed logistic regression models and found that COVID-19 would not affect the majority of the respondents' willingness to ride e-scooters. They also found that gender, age, and user experiences of ride-share services are the key determinants of people's willingness to use shared e-scooters. Another study by Glavić et al. (2021) added other socio-demographic variables as influential factors, such as educational attainment, employment status, and factors concerned with individual travel characteristics, including car ownership, prior exposure to shared mobility services, and average distance traveled. They also found that environmental benefits and congestion avoidance positively affected peoples' preferences for using e-scooters. On the other hand, safety concerns, lack of infrastructure, and unavailability of shared e-scooter services, and unfavorable weather conditions were negatively associated with people's willingness to use e-scooters.

The review shows that existing studies focused on the usage changes and willingness to use the services, but little attention has been given to COVID-related policies’ impact. Additionally, no study has examined the full range of spatio-temporal factors affecting e-scooter usage during the Pandemic. Moreover, spatial and serial autocorrelatios in shared micormobility data have been ignored in previous studies. This study addresses these gaps using a random effects spatial autorgressive panel model that combines spatial and temporal variables including COVID related variables to analyze the impact of the Pandemic on e-scooter usage in Austin City, Texas, while controlling for spatial and serial autocorrelation.

3. Data and methodology

3.1. Shared micromobility vehicle trip data

The study used a shared micromobility vehicle trip data, which is publicly available in the City of Austin data portal (City of Austin, 2021). We downloaded the trip data for the entire years of 2019 and 2020.

The primary downloaded dataset contains a total of 11,921,020 shared micromobility trips. Among them, e-scooters cover 95% of trips. We cleaned the data by removing the missing values in the trip record. Moreover, by following Caspi et al. (2020), we also removed the trip data if the trip distance was more than 80 miles and the trip duration exceeded 12 h. Our final data set had 7,403,230 shared micromobility trip information, including e-scooter, bicycle, and moped trips. Among them, shared e-scooters had 6,943,814 trips (5,406,020 in 2019 and 1,537,794 in 2020),which were used for the data analysis.

3.2. Austin COVID-19 features and policy

On March 13th, 2020, the City of Austin announced the first and the second confirmed COVID cases, and by April 27th, 2020, Travis county reached 1000 cases. Austin saw two-wave patterns in reported cases of COVID-19 during 2020 (Fig. 1 ). Though spring passed by without any notable surge in cases, the tide began to turn in mid-June as the number of confirmed cases grew larger and larger, and the first wave (Wave-I) started, which continued till the end of August with its peak in early July. The second wave (Wave-II) started in late October and extended till the next year. The State of Texas and the City of Austin enacted different policy measures to contain the virus's spread, including lockdown measures, stay-at-home orders, mask mandates, etc.

Fig. 1.

Timeline of COVID-19 cases and COVID-related policies in Austin during 2020.

To examine how the usage of e-scooter services changed over the course of the COVID-19 pandemic, the entire year of 2020 was split into six phases, based on the timeline of COVID cases and non-pharmaceutical interventions enforced by the City of Austin:

• i)Pre-COVID (January 1, 2020-March 12, 2020): On March 13th, 2020, the City of Austin announced the first and the second confirmed cases.
• ii)Pre-Lockdown Phase (March 13, 2020-March 18, 2020): Time between the first COVID case reported date to the day first lockdown was issued.
• iii)Lockdown I (March 19-April 30, 2020): On March 19, the State governor signed his first coronavirus-related executive order (GA-08) (Mcconnell, 2021), which began Texas' lockdowns with a prohibition on social gatherings of more than ten people and mandated the closure of dine-in restaurants, nursing homes, and schools.
• iv)Reopening-I (May 1, 2020-June 25, 2020): Since there was no major spike or decline in April's COVID data, State (Friend, 2020) announced plans for “Phase I” of reopening to begin on May 1, with retail stores and dine-in restaurants permitted to resume operations but at a maximum of 25 percent capacity. Within a few days, the state also permitted office buildings and bars to reopen at 25 percent capacity and restaurants to expand to 50 percent capacity.
• v)Lockdown-II (June 26, 2020-September 20, 2020): On June 26, the state renewed the lockdown orders, limiting restaurant capacity to 50 percent and closing bars, prohibiting outdoor gatherings of more than 100 people. During this time, the State also issued a statewide mask mandate that made individuals liable for fines.
• vi)Reopening-II (September 21, 2020- December 31, 2020): Beginning September 21, a new order permitted many businesses to expand reopening capacity to 75 percent, though bars remained closed.

3.3. Dependent variables

We considered five dependent variables in this study: i) average daily e-scooter pick-up trips in each census tract from January 1, 2019, to December 31, 2019, ii) average daily e-scooter pick-up trips in each census tract from January 1, 2020 to December 31, 2020; iii) census tract level pointwise relative change (difference in each day relative to 2019) of e-scooter trips in each day between 2020 and 2019; iv) census tract level pointwise change of e-scooter trip distance in each day between 2020 and 2019; v) census tract level pointwise change of e-scooter trip duration in each day between 2020 and 2019. Our spatial panel models are linear models and require a normally distributed dependent variable. However, our first two dependent variables (number of e-scooter pick-up trips) are count data with zero values. We used log-transformed dependent variables for the average daily e-scooter pick-up trips to approximate a normal distribution, thus allowing us to estimate a linear model. To avoid the problem of log transformation of observations with zero trip, we added one to each observation following previous studies (e.g., Lee et al., 2021, Caspi et al., 2020; Lee and Noland, 2021).

3.4. Explanatory variables

Based on the literature review and study areas context, the study included five broader groups of independent variables: i) temporal variables, ii) socio-demographic variables, iii) built- environment variables, iv) public transportation supply features, and v) e-scooter availability. Table 1 provides the summary statistics and sources of these variables for the years 2019 and 2020.

3.4.1. Temporal variables

Previous studies (e.g., Mathew et al., 2019, Tuli and Mitra, 2021, Kimpton et al., 2022, Noland, 2021; Gabhardt et al., 2021) found weather variables as influencing factors in e-scooter usage. In order to compare the effects of weather variables on e-scooter usage during and before the COVID-19 pandemic, the study included the average temperature of the day (°F), total precipitation (inch), average wind speed (mph), and percentage of humidity. As there were no snow days in Austin during the study period, the model specification did not contain the snowfall variable. All weather variables were obtained from the portal of Wunderground (2021). The weather variables' spatial unit was Austin, which changed daily. The total precipitation (rain) and wind speed were expected to be adversely impacted by the number of e-scooter trips. On the other hand, warmer days were anticipated to create more e-scooter trips.

Jiao and bai (2020) observed that during the weekends, e-scooter rides were significantly longer in respect of distance and duration in Austin. Therefore, we also included weekends as a binary variable to compare the demands of e-scooter trips between weekdays and weekends.

The final temporal variables were the COVID-19 pandemic-related features which were only included in the 2020 model. The study collected the new daily COVID cases in Austin for the entire year of 2020 from the COVID-19 Dashboard of Austin city (COVID-19 Dashboards, City of Austin). As mentioned earlier (Section 3.2), based on the timeline of COVID cases and non-pharmaceutical interventions enforced by the City of Austin, we split the entire year of 2020 into six phases. To examine how the usage of e-scooter changed over the course of the COVID-19 pandemic, we considered binary variables to represent days of Pre-COVID, Lockdown-I, Reopening-I, Lockdown-II, and Reopening-II phases. We combined Lockdown-I and Lockdown-II periods to avoid the multicollinearity issues, and this variable entered as a baseline. We expected the Lockdown phase to experience less e-scooter usage and a greater decrease than other phases.

3.4.2. Socio-demographic variables

Several studies (e.g., Mitra and Hess, 2021, Lee et al., 2021, Caspi et al., 2020, Frias-Martinez et al., 2021) revealed a connection between the usage of e-scooters and various socioeconomic factors. To capture these effects, this study included age and income variables in the spatial autoregressive models. The median age of the population of each census tract was included to represent the age variable. The income variable was incorporated as a binary variable. Since the definition of low-income can vary widely across cities in the US, following (Frias-Martinez et al., 2021), we defined a census tract as low-income if that tract has a census-based average income within the lowest income quartile for Austin.

Built-environment variables

Jiao and Bai (2020) found that the densely populated and compacted land use areas of Austin were more likely to have higher e-scooter usage before the Pandemic. Therefore, the models included each census tract's population density and land use entropy index to observe the effects of these variables during the Pandemic. The American Census Survey provided the total population of each census tract of the observed area. The population density of each census tract was calculated by dividing the total population by the area of the tracts. Our measure of land use entropy came from the EPA’s Smart Location Database, where employment and occupied housing were included in the entropy calculation (EPA, 2021).

3.4.4. Public transportation supply features

E-scooters have gained popularity for being a first/last mile mode for transit users (Zuniga-Garcia et al., 2022; Hosseinzadeh, 2021). However, due to the higher risk of getting infected with the virus, several studies (e.g., Jenelius and Cebecauer, 2020, Eisenmann et al., 2021, Ali Sahraei et al., 2021) observed a dramatic decrease in public transportation ridership during the COVID-19 pandemic. Consequently, to capture the impact of transit facilities on e-scooter usage during the Pandemic, the study considered the number of bus stops in a census tract. These data were collected from the open data portal of the City of Austin, Texas (2021). The number of bus stops was included as five categorical variables: census tracts with no bus stop (baseline), 1 to 5 bus stops, 6 to 10 bus stops, 11 to 20 bus stops, and more than 20 bus stops.

3.4.5. E-scooter availability

The availability of e-scooters in a location could be an important determinant of e-scooter usage. During the initial stage of the Pandemic, several companies ceased their operations, resulting in a decrease in the number of both docked and dockless e-scooters available in many cities (Egan, 2020, Bureau of Transportation Statistics, 2021). For instance, Lime, Spin, and Jump, three of Austin’s five authorized e-scooter companies, paused their operations (Perez, 2020). Additionally, Lyft removed 2000 scooters from the streets of Austin by discontinuing its e-scooter services (Jankowski, 2021). This reduction in supply had an indirect impact on e-scooter usage. In a study analyzing trip and trajectory data of e-scooters in Austin, Dean and Zuniga-Garcia (2022) claimed that the operational decisions intrinsically affected their analysis of user demand for this mode and route choice (particularly rebalancing) since the availability of e-scooters affects users’ willingness to ride. To capture the effect of e-scooter availability, we included the number of unique e-scooter devices used in a month in a census tract. The Austin data portal provided data on the unique ID for an e-scooter that has been used for a trip. Therefore, counting those devices does not provide all the available e-scooters of that time. As a proxy for the e-scooter availability, we counted all the unique devices used in a month in a census tract by assuming that those devices were available at some point in that month in that census tract. The other option could be to consider the daily count, but it would increase the risk of excluding more devices available for use that day.

3.5. Modeling approach

3.5.1. Spatial dependence and model

To understand the differences and similarities in determinants of e-scooter usage during normal and pandemic periods and to examine what factors contribute to the changes in e-scooter usage due to the Pandemic, we employed a random effect spatial panel model. A spatial panel refers to the data containing time series observations (number of e-scooters trips each day) of a number of spatial units (census tracts). Since the e-scooter trip data is location-specific (census tract level), it is natural to expect spatial interactions between trips in nearby census tracts. In modeling term, this logic suggests that a combination of unobserved variables that may be spatially correlated plus the spatially distributed influence of variables that would influence the e-scooter demand of a census tract and would be included in any model of e-scooter demand.

On the other hand, the motivation for using a random effect model is that the census tract-specific effect is randomly distributed across locations. Also, depending on the deviation of the effects from the “average location” and across time, negative or positive serial correlation could occur. Failing to capture these spatial and serial effects will lead to biased and inconsistent estimates (Lesage and Pace, 2009). Thus, to analyze the e-scooter trips with the i number of locations (census tracts) and t periods (days), a random effect spatial panel model is appropriate which captures both the spatial autocorrelation and serial correlation over time. These spatial associations could operate via the number of e-scooter trips or other characteristics of nearby locations (census tracts), or they could be captured by error terms. To identify an appropriate spatial model, it is important to identify the sources of the spatial autocorrelations. Depending on the sources of the spatial autocorrelations, two types of model can be employed: Spatial Error Model (SEM) or Spatial Autoregressive Models (SAR) (Lesage and Pace, 2009). Spatial error occurs when errors are spatially correlated due to random features associated with locations and when both the dependent and independent variables have spatial autocorrelations. On the other hand, spatial autoregressive model is appropriate when the dependent variable (number of e-scooter trips) is spatially correlated.

Using Moran’s I test (Moran, 1950), we found that the residuals of the model are spatially correlated (P-value < 0.01), indicating the presence of spatial error in our data set. We also fitted a cross-sectional model with a spatial lag to check the presence of spatial autocorrelation among dependent variables, which yielded a significant statistic. Since our data set had both spatial error and spatial lag autocorrelation, we estimated a random effects spatial-autoregressive model with a spatially autocorrelated error (SAC) panel model. Our SAC model can be written:

$${\mathrm{l}\mathrm{o}\mathrm{g}(y}_{\mathit{nt}})=\lambda W{\mathrm{l}\mathrm{o}\mathrm{g}(y}_{\mathit{nt}})+X_{\mathit{nt}}\beta +C_n+u_{\mathit{nt}}$$

$$u_{\mathit{nt}}=\rho W{u}_{\mathit{nt}}+v_{\mathit{nt}}t=1,2,\cdots \cdots \cdots ,T$$ (1)

Where $y_{\mathit{nt}}={(y_{1t},y_{2t},\cdots \cdots .,y_{\mathit{nt}})}^{\prime }$ is an n × 1 vector of observations for the total number of e-scooter trips for time period t (days) with n number panels (census tracts). $X_{\mathit{nt}}$ is a matrix of time-variant and -invariant independent variables; W is an n × n spatial weight matrix; $C_n$ is random effects with mean 0 and variance ${{\sigma }_c}^2$. $u_{\mathit{nt}}$ is a spatillay lagged error. $v_{\mathit{nt}}$ is a vector of disturbances and is independently and identically distributed (i.i.d.) across panels and time with variance i and t with ${\sigma }^2$ In the first equation of (1), the term $\lambda W{\mathrm{l}\mathrm{o}\mathrm{g}(y}_{\mathit{nt}})$ reflects the impact of the number of e-scooter trips of neighboring census tracts and it accounts for locally constant omitted variables. The second equation of (1) captures residual spatial autocorrelation. When $\rho =0$, Eq (1) reduces to a SAR model and when $\lambda =0$, it becomes an SEM model; setting both $\lambda$ and $\rho$ to 0 yields a simple random effect regression panel model.

3.5.2. Weight matrix

Since our spatial unit of the trip data is census tract, we used a contiguity weight matrix which creates a symmetric weighting matrix, W, that has the same positive weight for contiguous spatial units and, a zero weight for all other units. The weight matrix is normalized by dividing the entries by the absolute value of the largest eigenvalue of the matrix. A normalized weight matrix is used to improve numerical accuracy and to make nonexplosive autoregressive parameters ($\widehat{\lambda }$, $\widehat{\rho }$) bounded by −1 and 1 (Kelejian and Prucha, 2010).

3.5.3. Model estimation

The SAC model is estimated using random-effect estimators where the random effects enter the Eq. (1) for ${\mathrm{l}\mathrm{o}\mathrm{g}(y}_{\mathit{nt}})$ linearly.

For (1), we can stack all the time periods and write the equations as an $n\mathrm{T}\times 1$ vector form

$$\mathrm{l}\mathrm{o}\mathrm{g}\left({\mathbf{y}}_{\mathit{nT}}\right)=\lambda ({\mathbf{I}}_T\otimes \mathsf{W}){\mathsf{y}}_{\mathit{nT}}+{\mathrm{Z}}_{\mathit{nT}}\beta +{\mathrm{\xi }}_{\mathit{nT}}$$ (2)

Where.

${\mathrm{l}\mathrm{o}\mathrm{g}(\mathbf{y}}_{\mathit{nT}})$ = ${\left({{\mathbf{y}}^{\mathbf{\prime }}}_{n1},{{\mathbf{y}}^{\mathbf{\prime }}}_{n2},\cdots ,{{\mathbf{y}}^{\mathbf{\prime }}}_{\mathit{nt}}\right)}^{\prime }$ is an $\mathit{nT}\times 1$ vector of observations of the dependent variable for $i=1,\cdots ,n$ and $=1,\cdots ,T$;

${\mathsf{v}}_{\mathit{nT}}$ = ${\left({{\mathsf{v}}^{\mathbf{\prime }}}_{n1},{{\mathsf{v}}^{\mathbf{\prime }}}_{n2},\cdots ,{{\mathsf{v}}^{\mathbf{\prime }}}_{\mathit{nt}}\right)}^{\prime }$ is an $\mathit{nT}\times 1$ vector of innovations;

${\mathrm{Z}}_{\mathit{nT}}$ = ${\left({{\mathrm{Z}}^{\mathbf{\prime }}}_{n1},{{\mathrm{Z}}^{\mathbf{\prime }}}_{n2},\cdots ,{{\mathrm{Z}}^{\mathbf{\prime }}}_{\ast }\left(\rho \right)\prime \right)}^{\prime }$ is an $\mathit{nT}\times k$ matrix of $k$ regressors for $i=1,\cdots ,n$ and $t=1,\cdots ,T$; and.

${\mathrm{\xi }}_{\mathit{nT}}$ is the overall disturbance vector $\mathit{nT}\times 1$ vector.

For (1), the overall disturbance vector ${\mathrm{\xi }}_{\mathit{nT}}$ is

$${\mathrm{\xi }}_{\mathit{nT}}={\mathbf{I}}_T\otimes {\mathbf{c}}_n+\left\{{\mathbf{I}}_T\otimes {\mathbf{R}}_n{\left(\rho \right)}^{-1}\right\}{\mathsf{v}}_{\mathit{nT}}$$

Where ${\mathbf{R}}_n\left(\rho \right)={\mathbf{I}}_n-\rho \mathbf{M}$. Its variance matrix is

$${\mathrm{\Omega }}_{\mathit{nT}}\left(\theta \right)={\sigma }_{\mathrm{c}}^2\left({\mathbf{I}}_T{{\mathbf{I}}^{\prime }}_T\otimes {\mathbf{I}}_T\right)+{\sigma }^2\left\{{\mathbf{I}}_T\otimes {\mathbf{R}}_n{\left(\rho \right)}^{-1}{\mathbf{R}\prime }_n{\left(\rho \right)}^{-1}\right\}$$

The log-likelihood function for (2) is

$$\mathrm{l}\mathrm{n}{L}_{\mathit{nT}}\left(\theta \right)=-\frac{\mathit{nT}}{2}\ln \left(2\pi \right)-\frac{1}{2}\ln \left|{\mathrm{\Omega }}_{\mathit{nT}}\left(\theta \right)\right|+T\mathrm{l}\mathrm{n}\left|{\mathbf{S}}_n\left(\lambda \right)\right|-\frac{1}{2}{{\mathrm{\xi }}^{\prime }}_{\mathit{nT}}\left(\theta \right){{\mathrm{\Omega }}_{\mathit{nT}}\left(\theta \right)}^{-1}{\mathrm{\xi }}_{\mathit{nT}}\left(\theta \right)$$

where ${\mathbf{S}}_n\left(\lambda \right)={\mathbf{I}}_T-\lambda \mathbf{W},$ and $\theta ={\left({\beta }^{\prime },\lambda ,\rho ,{\sigma }_{\mathrm{c}}^2,{\sigma }^2\right)}^{\prime }.$

All the SAC models were estimated using StataMP 17; in particular, our model parameters were estimated using “spxtregress, re” with a maximum likelihood estimator derived by Lee and Yu (2010).

3.5.5. Model interpretation

The interpretation of SAC model is not straightforward because of the spatial lag term ($\lambda W{\mathrm{l}\mathrm{o}\mathrm{g}(y}_{\mathit{nt}})$), which creates a feedback effect, meaning, the effect of an explanatory variable’s change for a specific unit will affect the unit itself and, potentially, all other units indirectly. This implies the existence of direct, indirect, and total marginal impact which can be derived from the reduced form mean of (1).

$${\mathrm{l}\mathrm{o}\mathrm{g}(\mathbf{y}}_{\mathit{nt}})={\left(\mathbf{I}-\lambda \mathsf{W}\right)}^{-1}\left({\mathbf{X}}_{\mathit{nt}}\mathit{\beta }+{\mathbf{c}}_n+{\mathit{∊}}_{\mathit{nt}}\right)$$ (3)

To simplify later notation, we define ${\stackrel{\sim }{\mathbf{y}}}_{\mathit{nt}}$ as $\mathrm{l}\mathrm{o}\mathrm{g}\left({\mathbf{y}}_{\mathit{nt}}\right)$ minus the spatial spillover of the individual effects ${\mathbf{c}}_n.$

$${\stackrel{\sim }{\mathbf{y}}}_{\mathit{nt}}={\mathrm{l}\mathrm{o}\mathrm{g}(\mathbf{y}}_{\mathit{nt}})-{\left(\mathbf{I}-\lambda \mathsf{W}\right)}^{-1}{\mathbf{c}}_n$$

$$={\left(\mathbf{I}-\lambda \mathsf{W}\right)}^{-1}({\mathbf{X}}_{\mathit{nt}}\mathit{\beta }+{\mathit{∊}}_{\mathit{nt}})$$ (4)

For the random-effects model, the individual effects ${\mathbf{c}}_n$ are treated as part of random errors. Thus, (3) implies that the mean of ${\mathbf{l}\mathbf{o}\mathbf{g}(\mathbf{y}}_{\mathit{nt}})$ conditional on the independent variables and their spatial lags is

$$E(\mathrm{l}\mathrm{o}\mathrm{g}\left({\mathbf{y}}_{\mathit{nt})}\right|{\mathbf{X}}_{\mathit{nt}},\mathbf{W})={\left(\mathbf{I}-\lambda \mathsf{W}\right)}^{-1}\left({\mathbf{X}}_{\mathit{nt}}\mathit{\beta }\right)$$ (5)

This is known as the reduced-form mean because the solution (3) is known as the reduced form of the model. The predicted reduced-form mean substitutes estimates of $\lambda$ and $\mathit{\beta }$ into (5).

The total impact of an independent variable $\mathbf{x}$ is the average of the marginal effects it has on the reduced-form mean of ${\mathbf{y}}_{\mathit{nt}}$,

$$\frac{1}{\mathit{nT}}\sum _{t=1}^T\sum _{i=1}^n\sum _{j=1}^n\frac{\partial E\left(\mathrm{l}\mathrm{o}\mathrm{g}\left({\mathbf{y}}_{\mathit{it}}\right)\right|{\mathbf{X}}_{\mathit{nt}},\mathbf{W})}{\partial x_{\mathit{jt}}}$$ (6)

Where $E\left(\mathrm{l}\mathrm{o}\mathrm{g}\left({\mathbf{y}}_{\mathit{it}}\right)\right|{\mathbf{X}}_{\mathit{nt}},\mathbf{W}$) is the $i$ th element of the vector $E\left(\mathrm{l}\mathrm{o}\mathrm{g}\right({\mathbf{y}}_{\mathit{nt})}|{\mathbf{X}}_{\mathit{nt}},\mathbf{W}$), whose formula is given in (4), and $x_{\mathit{jt}}$ is the $j$ th unit’s value for $\mathbf{x}$ at time $t$.

The direct impact of an independent variable $\mathbf{x}$ is the average of the direct, or own, marginal effects:

$$\frac{1}{\mathit{nT}}\sum _{t=1}^T\sum _{i=1}^n\frac{\partial E\left(\mathrm{l}\mathrm{o}\mathrm{g}\left({\mathbf{y}}_{\mathit{it}}\right)\right|{\mathbf{X}}_{\mathit{nt}},\mathbf{W})}{\partial x_{\mathit{it}}}$$ (7)

The indirect impact of an independent variable $\mathbf{x}$ is the average of the indirect, or spillover, marginal effects:

$$\frac{1}{\mathit{nT}}\sum _{t=1}^T\sum _{i=1}^n\sum _{j=1,j\ne i}^n\frac{\partial E\left({\mathrm{l}\mathrm{o}\mathrm{g}(\mathbf{y}}_{\mathit{it}}\right)|{\mathbf{X}}_{\mathit{nt}},\mathbf{W})}{\partial x_{\mathit{jt}}}$$ (8)

4. Results

4.1. Descriptive analysis

This section presents the descriptive analysis of spatial and temporal changes of the shared e-scooter usage during the Pandemic. Fig. 2, Fig. 3, Fig. 4 compare the temporal variation of shared e-scooter related attributes between the years 2019 and 2020 within the jurisdiction of Austin City (City of Austin Texas, 2021). For this comparison, February 29 was removed from the 2020 dataset to ensure corresponding comparisons. Fig. 5, Fig. 6, Fig. 7 show the census tract level spatial changes of e-scooter usage. This analysis was carried out for five different COVID-19 phases discussed in Section 3.2. We did not include the pre-lockdown phase (March 13, 2020-March 18, 2020) for this analysis because it had only six days.

Fig. 2.

Daily e-scooter trips in 2019 and 2020.

Fig. 3.

Average trip distance (miles) in 2019 and 2020.

Fig. 4.

Average trip duration (minutes) in 2019 and 2020.

Fig. 5.

Percentage change of e-scooter trips per day in 2020 compared to 2019 (periods are in clockwise order).

Fig. 6.

Percentage change of e-scooter trip distances per day in 2020 compared to 2019 (periods are in clockwise order).

Fig. 7.

Percentage change of e-scooter trip duration per day in 2020 compared to 2019 (periods are in clockwise order).

4.1.1. Temporal analysis

The number of e-scooter pick-up trips in 2020 was almost similar to the pattern of 2019 in the pre-COVID period (Fig. 2). However, a drastic reduction in daily e-scooter trips was observed after the reporting of the first COVID-19 case in Austin, and it almost dropped to zero when the first lockdown started. This was no surprise since the human movement was at its barest minimum during this time period. In Louisville, Kentucky, Hosseinzadeh and Kluger (2021) observed a similar steeper decline of e-scooter trips during the height of lockdown. The daily trip count in 2020 began to experience an increase as Wave-I started to slow down, and the Reopening Phase II started. This trend continued until the Wave-II started when the e-scooter pickups began to slow down again.

On the other hand, during the pre-COVID period, the average trip distance and duration were almost similar in both years (Fig. 3, Fig. 4); however, they increased significantly in 2020 as the Lockdown-I started and continued this trend until the end of the year. There are two possible explanations for this unusual trend. First, people might be willing to use these services only when necessary and when the travel distance might be longer, and other alternatives like public transportation were not felt as safe as e-scooters. Because these services allowed riders to travel while maintaining social distancing protocols. Second, people started to use this service for recreational purposes for which generally trip duration and distances are longer. A recent survey in a peer city e-scooter program in Chicago found that approximately one-third of e-scooter riders said they “sometimes” or “often” used e-scooters to make social visits, or to ride for enjoyment, to attend recreational activities during the Pandemic (Chicago Department of Transportation, 2020). It is also possible that this trend is related to an overall shift in trip patterns due to the COVID-19 pandemic.

4.1.2. Spatial analysis

Fig. 5, Fig. 6, Fig. 7 show the percentage change of daily e-scooter trips, trip miles, and trip duration of e-scooter usage for each period of 2020 compared to 2019. A negative change indicated a drop in 2020 compared to 2019 for all the metrics. Due to a very short period, the phase between the pre-COVID and Lockdown I are excluded from the maps in Fig. 5, Fig. 6, Fig. 7.

While some census tracts of Austin experienced an increase in percentage change of daily e-scooter use in terms of trip counts, trip miles, and trip duration, most of them died out as the Lockdown-I period started (Fig. 5, Fig. 6, Fig. 7). The majority of the census tracts experienced a larger drop (<−60%) during Lockdown-I which did not change even in the Re-opening-I period. However, some census tracts in the periphery of Austin experienced a positive percentage change of e-scooter usage during the Lockdown-II and Reopening-II periods. Though the central part of Austin city experienced an increase in the total distance of e-scooter trips during the first-lockdown period, it decreased during the re-opening phase-I (Fig. 6). However, the rest of Austin experienced an increase in total distance covered by e-scooter from the second lockdown period to the end of 2020. In general, the central part of Austin city experienced a major decrease in e-scooter usage during all the periods compared to other parts of Austin.

4.2. Model results

4.2.1. Model fits and diagnostic

We estimated five random effects SAC models to answer the second and third research questions. The final dataset of each year contains 82,125 longitudinal e-scooter trip data from 225 census tracts within the jurisdiction of Austin City. Multicollinearity was not an issue in any of the models, as the largest VIF was less than 4 in all four models. We first estimated two random effects SAC models to examine the differences and similarities in determinants of e-scooter usage during normal (2019, Model I) and pandemic periods (2020, Model II) (Table 3, Table 4 ). The model diagnostics presented in Table 3, Table 4 illustrated the necessity of using a random effect SAC model. We performed the Hausman test (Hausman, 1978) for the random-versus fixed-effects, which led to a rejection of the fixed-effect model (insignificant test statistics). The standard deviation of the panel effects (sigma_u) and the standard deviation of the errors (sigma_e) are statistically significant as per requirement. Moreover, the Wald Chi-squared for all the models are significant, indicating that the overall models are significant in all cases. The Pseudo R square values range between 0.34 and 0.77, indicating a good model fit.

Table 3.

Results of Trip Data for 2019 (Model I).

Variables	Co-efficient	95% Confidence Interval		Marginal Effects
				Direct Impact	Indirect Impact	Total Impact
Weather
Maximum temperature	4.0E-4***	2.6E-4	5.5E-4	4.6E-4***	3.3E-3***	3.7E-3***
Average Humidity	−1.6E-4*	−3.4E-4	2.0E-5	−1.8E-4*	−1.3E-3*	−1.5E-3*
Average wind speed	−5.2E-6	−5.7E-4	5.6E-4	−5.9E-6	−4.3E-5	−4.9E-5
Total precipitation	1.7E-3	−4.6E-3	8.0E-3	1.9E-3	0.014	0.016
Weekend vs Weekday
Binary: 1 if weekend	7.0E-3***	2.4E-3	0.012	7.9E-3***	0.057***	0.065***
Socio-Demographic Characteristics
Median age	−0.036***	−0.060	−0.013	−0.041***	−0.297***	−0.339***
Low median income (Binary 1, if the tract is a low- income tract, otherwise 0)	−0.677***	−0.961	−0.392	−0.767***	−5.526***	−6.293***
Built Environment Variables
Population density	0.056***	0.036	0.076	0.063***	0.457***	0.520***
Land use Mix	1.269***	0.592	1.946	1.438***	10.364***	11.803***
Public Transportation Supply Features Bus Stops (baseline: no bus stop)
Binary 1: if a tract has 1–5 bus stops	−0.318*	−0.645	0.010	−0.360*	−2.593*	−2.953*
Binary 1: if a tract has 6–10 bus stops	−0.221	−0.574	0.132	−0.250	−1.804	−2.054
Binary 1: if a tract has 11–20 bus stops	0.054	−0.285	0.393	0.061	0.442	0.504
Binary 1: if a tract has more than 20 bus stops	−0.318*	0.067	0.782	0.481**	3.467**	3.948**
E-Scooter Availability
Log (# of devices in each tract in each month)	0.012***	0.011	0.013	0.013***	0.094***	0.108***
Constant	−0.542	−1.584	0.499
Model Fit and Diagnostic
Hausman test for Fixed vs. Random Model	1.00
(Prob > chi2)1
Spatial autoregressive parameter (λ)	0.892***
Error Parameter (ρ)	−0.701***
Sigma_u (${\sigma }_c)$	0.795***
Sigma_e ($\sigma )$	0.511***
Wald Chi square (15)	80701.95***
Pseudo R Square	0.77
N (Number of observations)	82,125
n (Number of tracts)	225
T (Number of days)	365

Note: * Significance at 10%, **Significance at 5%, *** Significance at 1%. 1. Insignificant Hausman test statistic means a random effect is preferable to a fixed effect model.

Table 4.

Results of Trip Data for 2020 (Model II).

Variables	Co-efficient	95% Confidence Interval		Marginal Effects
				Direct Impact	Indirect Impact	Total Impact
Weather
Average temperature	5.5E-3	−1.5E-3	0.013	5.6E-3	−1.8E-3	3.8E-3
Average humidity	−1.7E-3	−7.0E-3	3.5E-3	−1.8E-3	5.5E-4	−1.2E-3
Average wind speed	−4.3E-4	−0.018	0.017	−4.4E-4	1.4E-4	−3.0E-4
Total precipitation	0.013	−0.154	0.181	0.014	−4.3E-3	9.3E-3
Weekend vs Weekday
Binary: 1 if weekend	0.018	−0.116	0.152	0.018	−5.7E-3	0.012
COVID-19 Features
No. of daily new cases	−6.0E-4***	−1.0E-3	−1.6E-4	−6.1E-4***	1.9E-4***	−4.2E-4***
COVID-19 Policy (baseline: Lockdown Phase) Pre-COVID: Binary 1, if the date corresponds before the day of the first COVID case in Austin, otherwise 0	0.395***	0.121	0.668	0.40***	−0.126***	0.274***
Pre-lockdown: Binary 1, if the date corresponds to Pre-lockdown phase, otherwise 0	0.497*	−0.040	1.034	0.503*	−0.158*	0.345*
Reopen I: Binary 1, if the date corresponds to state Reopen I, otherwise 0	−0.316***	−0.50	−0.132	−0.320***	0.10***	−0.219***
Reopen II: Binary 1, if the date corresponds to state Reopen II, otherwise 0	0.412***	0.203	0.621	0.418***	−0.131***	0.286***
Socio-Demographic Characteristics
Median age	−0.034	−0.075	6.9E-3	−0.035	0.011	−0.024
Low median income (Binary 1, if the tract is a low- income tract, otherwise 0)	−0.998***	−1.50	−0.496	−1.011***	0.318***	−0.694***
Built Environment Variables
Population density	0.105***	0.070	0.141	0.107***	−0.033***	0.073***
Land use mix	1.231**	0.038	2.425	1.248**	−0.392**	0.856**
Public Transportation Supply Features Bus Stops (baseline: no bus stop)
Binary 1: if a tract has 1–5 bus stops	−0.046	−0.623	0.532	−0.047	0.015	−0.032
Binary 1: if a tract has 6–10 bus stops	0.402	−0.220	1.024	0.407	−0.128	0.279
Binary 1: if a tract has 11–20 bus stops	0.630**	0.032	1.228	0.638**	−0.201**	0.438**
Binary 1: if a tract has more than 20 bus stops	1.595***	0.965	2.225	1.616***	−0.508***	1.108***
E-Scooter Availability
Log (# of devices in each tract in each month)	6.4E-3***	5.3E-3	7.6E-3	6.5E-3***	−2.0E-3***	4.5E-3***
Constant	−12.947***	−14.912	−10.982
Model Fit and Diagnostic
Hausman test for Fixed vs. Random Model	1.00
(Prob > chi2)1
Spatial autoregressive parameter (λ)	−0.44***
Error Parameter (ρ)	0.94***
Sigma_u (${\sigma }_c)$	1.40***
Sigma_e ($\sigma )$	0.487***
Wald Chi square (20)	211916.19***
Pseudo R Square	0.70
N (Number of observations)	82,125
n (Number of tracts)	225
T (Number of days)	365

Note: * Significance at 10%, **Significance at 5%, *** Significance at 1%. 1. insignificant Hausman test statistic means a random effect is preferable to a fixed effect model.

The spatial autoregressive parameter (λ) and spatial error parameter (ρ) were highly significant in all models, illustrating the necessity of using a SAC model. The values of these two parameters were between −1 and 1, which was required since we row-normalized our weight matrix (Kelejian & Prucha, 2010). However, the sign of the spatial autoregressive parameter (λ) and spatial error parameter (ρ) were opposite in the 2019 and 2020 models. While the sign of the spatial error parameter did not have any implication on the model interpretation, the spatial autoregressive parameter does have an implication on the model interpretation because of its feedback effect on the e-scooter usage of neighboring locations (Section 3.5.5 provides further details).

We also estimated three random effects SAC models to examine factors influencing the changes in e-scooter usage during the Pandemic in terms of the number of trips (Model III), trip distance (Model IV), and trip duration (Model V). All the model diagnostics illustrated the necessity of using a random effect SAC model (Table 5, Table 6, Table 7 ). The results are discussed in the following section based on the direct and total impacts. As mentioned in Section 3.5.5, the interpretation of the spatial model is not as straightforward as the nonspatial specification. The marginal effect of any variable on e-scooter demand may differ across census tracts because of spatial interactions. The key difference between the direct and total impacts is that the direct impact measures the impact of a unit change in variable X_k in census tract i on e-scooter demand in census tract i average over all states. In contrast, the total impact measures the impact of the same unit change in variable X_k in census tract i on the e-scooter demand of all census tracts, again averaged over all states.

Table 5.

Results of Change in Total Daily E-Scooter Trips (Model III).

Variables	Co-efficient	95% Confidence Interval		Marginal Effects
				Direct Impact	Indirect Impact	Total Impact
Weather
Δ Temperature	6.2E-4***	4.0E-4	8.4E-4	6.8E-4***	2.9E-3***	3.6E-3***
Δ Humidity	−1.2E-4*	−2.5E-4	9.3E-6	−1.3E-4*	−5.6E-4*	−6.8E-4*
Δ Wind speed	−3.8E-4*	−8.0E-4	4.5E-5	−4.1E-4*	−1.8E-3*	−2.2E-3*
Δ Precipitation	2.7E-3	−1.5E-3	6.9E-3	3.0E-3	0.013	0.016
Weekend vs Weekday
Binary: 1 if weekend in any of the year	8.1E-4	−3.4E-3	5.0E-3	8.9E-4	3.8E-3	4.7E-3
COVID-19 Features
No. of daily new cases	−1.6E-5**	−3.1E-5	−4.2E-7	−1.7E-5**	−7.4E-5**	−9.1E-5**
COVID-19 Policy (baseline: Lockdown Phase) Pre-COVID: Binary 1, if the date corresponds before the day of the first COVID case in Austin, otherwise 0	−8.9E-3**	−0.016	−1.8E-3	−9.8E-3**	−0.042**	−0.052**
Pre-lockdown: Binary 1, if the date corresponds to Pre-lockdown phase, otherwise 0	−1.9E-3	−0.020	0.017	−2.1E-3	−9.1E-3	−0.011
Reopen I: Binary 1, if the date corresponds to state Reopen I, otherwise 0	−2.0E-3	−8.2E-3	4.1E-3	−2.2E-3	−9.5E-3	−0.012
Reopen II: Binary 1, if the date corresponds to state Reopen II, otherwise 0	0.011***	5.1E-3	0.016	0.012***	−0.131***	0.063***
Socio-Demographic Characteristics
Median age	4.3E-3**	6.3E-4	7.9E-3	4.7E-3**	0.020**	0.025**
Low median income (Binary 1, if the tract is a low- income tract, otherwise 0)	0.046**	1.1E-3	0.090	0.050**	0.214**	0.264**
Built Environment Variables
Population density	−2.4E-4	−3.4E-3	2.9E-3	−2.7E-4	−1.1E-3	−1.4E-3
Land use mix	−0.063	−0.169	0.043	−0.069	−0.297	−0.366
Public Transportation Supply Features
Bus Stops (baseline: no bus stop)
Binary 1: if a tract has 1–5 bus stops	0.022	−0.029	0.073	0.024	0.103	0.127
Binary 1: if a tract has 6–10 bus stops	0.010	−0.045	0.065	0.011	0.048	0.059
Binary 1: if a tract has 11–20 bus stops	−0.090***	−0.142	−0.037	−0.098***	−0.422***	−0.520***
Binary 1: if a tract has more than 20 bus stops	−0.076***	−0.132	−0.021	−0.084***	−0.360***	−0.443***
E-Scooter Availability
Δ (# of devices in each tract in each month)	6.4E-5***	5.9E-5	6.9E-5	7.0E-5***	3.0E-4***	3.7E-4***
Constant	−0.105	−0.267	0.058
Model Fit and Diagnostic
Hausman test for Fixed vs. Random Model	1.00
(Prob > chi2)1
Spatial autoregressive parameter (λ)	0.83***
Error Parameter (ρ)	−0.69***
Sigma_u (${\sigma }_c)$	0.12***
Sigma_e ($\sigma )$	0.51***
Wald Chi square (20)	38394.54
Pseudo R Square	0.55
N (Number of observations)	82,125
n (Number of tracts)	225
T (Number of days)	365

Note: * Significance at 10%, **Significance at 5%, *** Significance at 1%. 1. insignificant Hausman test statistic means a random effect is preferable to a fixed effect model.

Table 6.

Results of Change in Total Daily E-Scooter Trip Mile (Model IV).

Variables	Co-efficient	95% Confidence Interval		Marginal Effects
				Direct Impact	Indirect Impact	Total Impact
Weather
Δ Temperature	1.8E-3***	1.2E-3	2.3E-3	1.9E-3***	4.4E-3***	6.3E-3***
Δ Humidity	−4.3E-4***	−7.3E-4	−1.2E-4	−4.5E-4***	−1.1E-3***	−1.5E-3***
Δ Wind speed	−1.5E-3***	−2.5E-3	−5.2E-4	−1.6E-3***	−3.8E-3***	−5.4E-3***
Δ Precipitation	5.5E-3	−4.4E-3	0.015	5.8E-3	0.014	0.020
Weekend vs Weekday
Binary: 1 if weekend in any of the year	1.4E-3	−8.5E-3	0.011	1.5E-3	3.5E-3	5.1E-3
COVID-19 Features
No. of daily new cases	−4.7E-5**	−8.4E-5	−1.1E-5	−5.0E-5**	−1.2E-4**	−1.7E-4**
COVID-19 Policy (baseline: Lockdown Phase) Pre-COVID: Binary 1, if the date corresponds before the day of the first COVID case in Austin, otherwise 0	0.030***	0.013	0.047	0.032***	0.075***	0.107***
Pre-lockdown: Binary 1, if the date corresponds to Pre-lockdown phase, otherwise 0	3.5E-3	−0.041	0.048	3.7E-3	8.7E-3	0.012
Reopen I: Binary 1, if the date corresponds to state Reopen I, otherwise 0	−1.9E-3	−0.017	0.013	−2.0E-3	−4.7E-3	−6.8E-3
Reopen II: Binary 1, if the date corresponds to state Reopen II, otherwise 0	0.017**	3.5E-3	0.030	0.018**	0.042**	0.060**
Socio-Demographic Characteristics
Median age	5.8E-3**	6.6E-4	0.011	6.1E-3**	0.014**	0.020**
Low median income (Binary 1, if the tract is a low- income tract, otherwise 0)	0.061*	−1.3E-3	0.124	0.065*	0.152*	0.216*
Built Environment Variables
Population density	−1.3E-3	−5.9E-3	3.2E-3	−1.4E-3	−3.3E-3	−4.7E-3
Land use mix	−0.093	−0.243	0.057	−0.098	−0.230	−0.328
Public Transportation Supply Features Bus Stops (baseline: no bus stop)
Binary 1: if a tract has 1–5 bus stops	0.017	−0.055	0.089	0.018	0.043	0.061
Binary 1: if a tract has 6–10 bus stops	0.029	−0.048	0.106	0.030	0.071	0.101
Binary 1: if a tract has 11–20 bus stops	−0.031	−0.105	0.043	−0.032	−0.076	−0.108
Binary 1: if a tract has more than 20 bus stops	−0.013	−0.091	0.065	−0.013	−0.031	−0.045
E-Scooter Availability
Δ (# of devices in each tract in each month)	8.0E-5***	6.9E-5	9.1E-5	8.4E-5***	2.0E-4***	2.8E-4***
Constant	−0.159	−0.386	0.069
Model Fit and Diagnostic
Hausman test for Fixed vs. Random Model	1.00
(Prob > chi2)1
Spatial autoregressive parameter (λ)	0.72***
Error Parameter (ρ)	−0.59***
Sigma_u (${\sigma }_c)$	0.16***
Sigma_e ($\sigma )$	1.13***
Wald Chi square (20)	11006.98***
Pseudo R Square	0.46
N (Number of observations)	82,125
n (Number of tracts)	225
T (Number of days)	365

Note: * Significance at 10%, **Significance at 5%, *** Significance at 1%. 1. insignificant Hausman test statistic means a random effect is preferable to a fixed effect model.

Table 7.

Results of Change in Total Daily E-Scooter Trip Duration (Model V).

Variables	Co-efficient	95% Confidence Interval		Marginal Effects
				Direct Impact	Indirect Impact	Total Impact
Weather
Δ Temperature	0.018***	0.013	0.024	0.019***	8.1E-3***	0.027***
Δ Humidity	−4.7E-3***	−7.5E-3	−1.8E-3	−4.7E-3***	−2.1E-3***	−6.8E-3***
Δ Wind speed	−0.010**	−0.020	−6.7E-4	−0.010**	−4.5E-3**	−0.015**
Δ Precipitation	0.045	−0.048	0.138	0.045	0.020	0.065
Weekend vs Weekday
Binary: 1 if weekend in any of the year	0.017	−0.076	0.111	0.018	7.6E-3	0.025
COVID-19 Features
No. of daily new cases	−5.1E-4***	−8.6E-4	−1.7E-4	−5.2E-4***	−2.2E-4***	−7.4E-4***
COVID-19 Policy (baseline: Lockdown Phase) Pre-COVID: Binary 1, if the date corresponds before the day of the first COVID case in Austin, otherwise 0	0.224***	0.065	0.384	0.226***	0.098***	0.324***
Pre-lockdown: Binary 1, if the date corresponds to Pre-lockdown phase, otherwise 0	0.352*	−0.062	0.767	0.355*	0.155*	0.510*
Reopen I: Binary 1, if the date corresponds to state Reopen I, otherwise 0	−0.226***	−0.365	−0.087	−0.227***	−0.099***	−0.326***
Reopen II: Binary 1, if the date corresponds to state Reopen II, otherwise 0	0.352***	0.224	0.480	0.355***	0.155***	0.509***
Socio-Demographic Characteristics
Median age	0.038***	9.9E-3	0.067	0.039***	0.017**	0.056***
Low median income (Binary 1, if the tract is a low- income tract, otherwise 0)	0.665***	0.313	1.017	0.670***	0.292***	0.962***
Built Environment Variables
Population density	−0.014	−0.040	0.012	−0.014	−6.2E-3	−0.020
Land use mix	−0.052	−0.891	0.787	−0.052	−0.023	−0.075
Public Transportation Supply Features Bus Stops (baseline: no bus stop)
Binary 1: if a tract has 1–5 bus stops	0.367*	−0.037	0.770	0.370*	0.161*	0.531*
Binary 1: if a tract has 6–10 bus stops	0.605***	0.173	1.037	0.610***	0.266***	0.876***
Binary 1: if a tract has 11–20 bus stops	0.681***	0.266	1.096	0.686***	0.299***	0.985***
Binary 1: if a tract has more than 20 bus stops	0.771***	0.334	1.209	0.777***	0.339***	1.116***
E-Scooter Availability
Δ (# of devices in each tract in each month)	1.6E-4***	8.1E-5	2.4E-4	1.6E-4***	7.1E-5***	2.3E-4***
Constant	−1.497**	−2.771	−0.222
Model Fit and Diagnostic
Hausman test for Fixed vs. Random Model	1.00
(Prob > chi2)1
Spatial autoregressive parameter (λ)	0.31***
Error Parameter (ρ)	−0.27***
Sigma_u	0.0.87***
Sigma_e	8.41***
Wald Chi square (20)	560.01***
Pseudo R Square	0.34
N (Number of observations)	82,125
n (Number of tracts)	225
T (Number of days)	365

Note: * Significance at 10%, **Significance at 5%, *** Significance at 1%. 1. insignificant Hausman test statistic means a random effect is preferable to a fixed effect model.

4.2.2. Parameter estimates

The results of Model I (Table 3) and Model II (Table 4) describe the differences and similarities between the determinants of e-scooter usage during the year of 2019 and 2020. Models III (change in the number of e-scooter trips; Table 5), IV (changes in the trip distance; Table 6), and V (changes in trip duration; Table 7) capture the factors influencing the changes in e-scooter usage during the pandemic year. Since we subtracted the 2020 observations from 2019 and more than 95% of data had a negative value, a positive coefficient indicates that an increase in independent variable leads to a smaller decrease in e-scooter usage (the higher the magnitude of the positive coefficient is, the lesser the reduction is; which could also lead to an increase for some observations) in 2020 compared to 2019. On the other hand, a negative coefficient indicates a greater decrease (the higher the negative coefficient is, the greater the drop is).

Temporal effect

The weather variables had mixed results in 2019 and 2020. While two weather variables (maximum temperature and average humidity) were statistically significant in Model I, none were significant in Model II. Direct effects of Model I show that one unit increase in maximum temperature increased the e-scooter usage by 0.045% of a census tract. In contrast, the total effects showed that one unit increase in maximum temperature increased the e-scooter use of all census tracts by 0.375%. The differences between direct and total effects were due to the positive spillover effect from nearby census tracts. Kimpton et al., 2022, Tuli and Mitra, 2021 also found a positive association between temperature and e-scooter usage by analyzing pre-COVID time data from different cities. Conversely, the average humidity had a negative association with e-scooter usage in the 2019 model, which was consistent with the findings from previous studies (e.g., Noland, 2021). However, none of the weather variables were statistically significant in the 2020 model. The possible reason for these different results in 2019 and 2020 was that the usage of e-scooters was down most of the time in 2020 due to the Pandemic, and there was less variability in e-scooter use throughout the year.

Results of the direct effects of Model III show that changes in maximum temperature between 2020 and 2019 were associated with a smaller decrease in e-scooter usage between 2020 and 2019. However, changes in humidity and windspeed were associated with a greater decrease in e-scooter use. Similar results were also observed for changes in the trip distance (Model IV) and duration (Model V).

The binary variable of weekends produced a statistically significant positive relationship with e-scooter trips in Model I. The values of the direct and total impacts of Model I suggested that during weekends there was 0.79% of higher e-scooter usage in each census tract and 6.7% of higher e-scooter usage in all census tracts in 2019. Espinoza et al., 2019, McKenzie, 2019 observed that e-scooters were mainly used for leisure purposes, which might lead to higher rides of e-scooters during weekends. However, this variable was not significant in 2020, and in general, leisure trips also went down during 2020 (Li et al.,2020). As a result, e-scooter usage did not vary much during weekdays and weekends.

The COVID-related variables were only included in Model II (2020). As expected, the number of daily new COVID-19 cases had a statistically significant negative association with e-scooter usage during the Pandemic. The direct and total effects results show that one unit increase in the overall number of new COVID cases in Austin resulted in a 0.06% decrease in e-scooter usage in each census tract and 0.04% in all census tracts. The reason for the smaller magnitude of the total effect was the positive spillover effects. However, the result is consistent with Tokey (2020), who also observed a significant drop in bike-share usage with the increase of daily new COVID cases in Boston and DC. We also estimated the impact of COVID-19-related policy variables on e-scooter usage by using six categorical variables based on the timeline of COVID cases and non-pharmaceutical interventions enforced by the City of Austin (see section 3.2 for further details). The baseline was Lockdown periods (combined Lowdowns I and II).

The total effect results show that the Pre-COVID and Pre-Lockdown periods had 31.55% and 41.21% higher e-scooter usage in all census tracts than the Lockdown period, respectively. Interestingly, the Reopen-I period had 19.69% lower e-scooter use than the Lockdown period. This is possible because the Reopen-I period was still the early stage of COVID-19 when people were frightened, and the case number was still on the rise; therefore, residents may not get out during that period. However, the usage rate recovered during the Reopen-II phase, as it experienced a 33.15% higher usage than the Lockdown period.

As expected, the increase of daily new COVID cases led to a greater decrease in total daily trips (Model III), trip mileage (Model IV), and trip duration (Model V) during the Pandemic. Results of direct effect showed that one unit increase in new COVID cases led to a 0.002% relative decrease in the number of e-scooter trips in a census tract. Different COVID period variables had mixed results. Interestingly, during the pre-COVID period, each census tract experienced a greater decrease (-0.97%) compared to the Lockdown phase. One plausible explanation is that during early March of 2019, there was a massive spike in e-scooter usage in Austin, which was the highest in that year (possibly because of different big festivals in Austin during that time, such as Rodeo, South by Southwest, etc.) However, there was no such spike in 2020 (See Fig. 2). As expected during Reopen-II, each census tract experienced a smaller decrease in e-scooter usage compared to the Lockdown period. On the other hand, for Model IV, all the phases experienced a smaller e-scooter trip distance decrease than the Lockdown period. The results of Model V were also similar except for the Reopen-I phase.

Socio-demographic characteristics

The socio-demographic variables were statistically significant in both models (Model I and Model II) and had similar signs. As expected, the median age was negatively associated with e-scooter usage in both models. Huo et al. (2021) also found a negative correlation between the e-scooter ridership and the median age of the population in a census block group. In line with previous studies (Dean and Zuniga-Garcia, 2022, Tuli and Mitra, 2021), median income was negatively associated with e-scooter usage in both models. Total effects results showed that all lower median income census tracts in Austin had 99% fewer e-scooter trips compared to the rest of the tracts in 2019; however, the differences reduced to 50% in 2020. According to Bonacini et al., (2021), the opportunity of working from home (WFH) mainly favors male, older, high-educated, and high-paid employees. Therefore, due to the opportunity of WFH, there could be a reduction in e-scooter usage in high-income areas, leading to a difference from 99% to 50% during the Pandemic.

Results of Models III-V showed that a census tract with higher median age experienced a smaller decrease in e-scooter trips. Likewise, a census tract with lower median income populations enjoyed a lesser reduction in total e-scooter trips (5.1%), total trip miles (6.7%), and total trip duration (95.4%) during the Pandemic.

Built environment variables

As expected, both built environment variables (population density and entropy index) were positively significant in both models (Model I and Model II). This result is consistent with the findings of previous studies carried out in Austin (Jia and Bai, 2020; Bai and Jiao, 2020) and other cities (Tuli and Mitra, 2021, Huo et al., 2021). Direct effects results of population density variable from Model I showed that one unit increase in population density in a census tract was associated with a 6.3% increase in e-scooter usage in that census tract and a 51.9% increase in all census tracts. While in Model II, the direct effect had a higher magnitude (10.7%), the total effects had a lower magnitude (7.3%) than in Model I. The negative spillover effect caused this magnitude reduction of Model II's total effects. On the other hand, the direct effects of the land use entropy index show that one unit increase in entropy index in a census tract would increase e-scooter usage by 143.8% of that census tract in 2019, whereas it would increase the usage by 124.7% in 2020. The reduction in magnitude for both variables in the 2020 model could be related to the effectiveness of the stay-at-home policies.

Public transportation supply features

The models included the number of bus stops as public transportation supply variables. Various specifications of the public transportation supply feature variable were considered in the model, such as continuous, a binary variable indicating the existence of one or more bus stops, and categorical variables with ordinal values. The final model included the number of bus stops as categorical variables with no bus stop as a baseline in order to capture the non-linear relationship between bus stops and e-scooter usage.

Direct effect results of the 2019 model showed that a census tract with 1 to 5 bus stops was associated with a decrease in e-scooter demand relative to a census tract with no bus stop, whereas a census tract with more than 20 bus stops was associated with an increase in e-scooter usage. In the 2020 model, while the variables had similar signs, the 1 to 5 bus stop variable was not significant. These results indicated that census tracts with very few bus stops (1 to 5 bus stops) generated fewer e-scooter trips than a census tract with no bus stop, whereas areas with higher public transportation services (20 or more bus stops) generated the most e-scooter trips. Moreover, the comparison of the 2019 and 2020 models showed that the census tracts with more than 20 bus stops had a higher magnitude in the pandemic year model (1.61***) compared to the pre-pandemic model (0.48***). This higher magnitude did not imply that e-scooter usage was higher in those areas compared to 2019, instead, it indicated the differences in e-scooter usage between those areas and census tracts with no bus stops increased.

Results of Model III on transit supply features show that census tracts with a higher number of bus stops (>10) experienced a greater decrease in e-scooter usage than census tracts with no bus stops. Since census tracts with a higher number of bus stops had a higher e-scooter usage than census tracts with no bus services (Model I and II), it was obvious that those census tracts with higher public transportation services would experience a greater decrease.

E-scooter availability

The e-scooter availability variable was included as monthly e-scooter availability in a census tract. We tried different model specifications, including the continuous and ordinal specifications and a log-transformed variable. Although there were not significant differences in the results of different specifications, the final model included the log-transformed variable because it gave the lower AIC and BIC values and VIF values. Since this variable was log-transformed, the direct effect result can be interpreted as average elasticity. The results of Model I show that a 1% increase in the availability of e-scooters in a census tract was associated with a 0.01% increase in e-scooter usage in that census tract and a 0.11% increase in all census tracts. However, the magnitudes were lower in Model II (direct effect: 0.006% and total effect: 0.004%). Note that the lower value of the total effects in Model II compared to its direct effects is consistent with result of the negative spatial effect coefficient of Model II. This implies that a census tract with a higher number of monthly e-scooter devices increases the e-scooter usage of that census tract but will have a negative spillover effect on the e-scooter use of neighboring tracts.

Finally, the differences in e-scooter availability between 2020 and 2019 were significant in all three models (Model III, Model IV, and Model V). Results showed that decreases in differences in e-scooter availability between 2020 and 2019 resulted in a smaller reduction in e-scooter trips, total travel distance, and travel duration, indicating the importance of the availability of e-scooters in a census tract.

5. Discussions

One of the most important and novel findings from the model results relates to the spatial dependence in the data. All the model results provided strong evidence of spatial autocorrelation and spatial spillover effects. In the 2019 model, the λ coefficient was positive, indicating a positive spillover effect on e-scooter usage for a positive coefficient. On the other hand, in the 2020 model, the λ coefficient was negative, indicating a situation in which the high use of e-scooter in one census tract had a negative effect on the usage of e-scooters in neighbor census tracts. This can also be explained by Kao and Bera (2016), who argued that negative spatial autocorrelation is likely to occur when competition between regions, in this case, the availability of e-scooter devices in neighboring census tracts, outweighs cooperative factors. During the Pandemic, e-scooter availability was limited across Austin (Perez, 2020) and the rebalancing of e-scooters among different stations was not executed as frequently as usual, which negatively impacted the usage of neighboring tracts. In contrast, this was not observed in 2019, likely because there were more e-scooter devices in Austin before the Pandemic, and rebalancing was done more frequently to ensure the adequate availability of devices at each station. This is supported by unique device count data from the City of Austin open data portal, which revealed that the number of unique devices in 2020 was less than half of that in 2019 (23,601 compared to 53,197). This decrease was largely due to three of the five licensed dockless e-scooter operators in Austin (Lime, Jump, and Spin) suspending their operations in March 2020, while Lyft and Bird removed thousands of dockless electric scooters from Austin's streets following the COVID outbreak (Egen, 2020).

This result indicated that the cause of the reduction of e-scooter trips during the Pandemic was not only the fall of demand itself; it might also be due to the unavailability of enough e-scooter devices during the pandemic year. Of course, COVID cases and stay-at-home policies had a negative impact on e-scooter usage, as evident in the model results, but the negative spillover effect indicated that there would be more e-scooter trips in Austin during the Pandemic if there were enough devices available, indicating the critical role of shared e-scooter operators’ strategy in e-scooter usage during the Pandemic. This result could be helpful for shared e-scooter operators. If the operators want a better usage rate during any future similar event, they need to ensure the availability of an adequate number of e-scooters in each station.

The results of this study show that e-scooter usage was negatively associated with low-income areas in both 2019 and 2020. The negative relationship between low-income areas and e-scooter usage may not be caused by lower demand but by other factors, such as the high price of e-scooter trips compared to other shared micromobility services (NACTO, 2019), or spatial disparities in e-scooter availability (Stickle and Felson, 2020, Mooney et al., 2019). Interestingly, in both the 2019 and 2020 models, census tracts with lower-income residents had fewer e-scooter trips than other areas, but they experienced a smaller decrease (or increase) in e-scooter usage during the Pandemic. Two possible explanations account for this result: first, higher-income people may have reduced their e-scooter usage at a higher rate than lower-income people, as they were disproportionately switching from other modes to cars during the Pandemic (Brough et al., 2021); and second, low-income residents, who rely more on public transportation (Pucher and Renne, 2003), may have used e-scooters more as an alternative when many public transportation services were closed or had reduced frequency (Sullivan, 2020, Qi et al., 2021).

This result indicates the possible role of e-scooters as an alternative transportation option for lower-income households during the Pandemic when many public transportation services were limited. These findings could be useful for policymakers to prepare for future pandemics as well as in the post-pandemic time when cities want to provide a viable alternative transportation option to public transportation services for lower-income residents. Planners and policymakers could incentivize e-scooter companies to provide services to low-income areas through regulations or pricing mechanisms, especially during the Pandemic. Also, Austin city authority can develop a policy to have a minimum rebalancing requirement for the areas with lower-income residents.

Another interesting result from both models (2019 and 2020) indicated a non-linear relationship between public transportation services and e-scooter usage. Results illustrated that: i) census tracts with limited public transportation services (no or fewer bus stops) were likely to generate very low e-scooter demand, and ii) census tracts with very high public transportation services (higher number of bus stops) had the highest e-scooter demand. However, results also showed that census tracts with a higher number of bus stop experienced a greater decrease than those without bus services during the Pandemic.

Interpreting these two results together could provide interesting insights into the relationship between shared e-scooters and public transportation services. First, shared e-scooters might serve as a complementary mode (first- and last-mile connections) to public transportation services in areas with very high bus services in 2019. However, in 2020 the shared e-scooters might serve a dual purpose (both as a complementary and substitute mode). Since bus services were reduced during the Pandemic and people stopped using public transportation due to the fear of infection (Qi et al., 2021), it might hit hard in areas with most bus stops. As a result, while those areas experienced a greater decrease in e-scooter usage, those areas still experienced a higher usage than areas with no bus stops. This was possible because many bus riders who used to ride e-scooters for first- and last-mile connections might switch to e-scooters for other trip purposes during the Pandemic. Further research is needed to understand this relationship using post-pandemic data as the question remains about how riders from those areas will use e-scooters in the post-pandemic era.

However, these findings could be useful for Austin transit authorities and shared e-scooter operators in Austin. They can integrate these two services to increase the ridership of both modes in the post-pandemic era by introducing different features such as a universal payment method, monthly transit pass with free e-scooter rides to bus stops, reduced transit fares for e-scooter members, etc.

Regardless of the effects of the pandemic, the usage of e-scooters showed a positive association with population density and land use mix index. The availability of e-scooters in each census tract was found to be a critical factor in increasing e-scooter usage. To increase the usage of e-scooters in the post-pandemic era, e-scooters operators in Austin should ensure higher availability of e-scooters in dense mixed land-use areas.

6. Conclusion

This study aimed to examine the impact of the COVID-19 pandemic on shared e-scooter usage. We analyzed shared e-scooter data of Austin from 2019 and 2020 to answer three research questions: i) how did the e-scooter usage pattern change spatio-temporarily during the Pandemic? ii) what were the differences and similarities in determinants of e-scooter usage during normal and pandemic periods? and iii) what factors contributed to the spatio-temporal changes in e-scooter usage during the Pandemic? The study relied on the exploratory analysis to answer the first question. To answer the second and third questions, this study estimated five random effects spatial-autoregressive models with spatially autocorrelated error (SAC) panel models that controlled for spatial autocorrelation and serial correlation. All the models included temporal variables (weather data, weekday/weekend, number of COVID cases, and COVID-related policies), socio-demographics, built environment, public transportation supply features, and e-scooter availability.

The descriptive temporal analysis revealed that while daily e-scooter trips reduced considerably during the Pandemic, the average trip duration and average trip distance increased during the pandemic period. The spatial analysis showed the majority of the census tracts experienced a larger drop (<-60%) during Lockdown-I, which did not change even in the Reopening-I period. However, the central part of Austin experienced the most decrease in e-scooter usage during all the periods compared to other parts of Austin. Perhaps this is because the central part had the highest usage before the Pandemic compared to other areas, as more e-scooters were deployed in downtown areas for operational reasons (Dean and Zuniga-Garcia, 2022). So, during the Lockdown-I, when everything was closed, the central part experienced the largest drop.

All the model results provided strong evidence of spatial autocorrelation and spatial spillover effects, indicating the necessity of using a spatial model for analyzing different shared micromobility services trip data. Models I and II revealed some similarities and differences in the determinants of e-scooter usage during the pre-pandemic and the pandemic period. For example, two weather variables (Temperature and average humidity) and weekend variables were statistically significant in the 2019 model but were not significant in the 2020 model. As expected, COVID cases and policy-related variables were significant in the 2020 model and had expected negative signs. The rest of the variables had similar signs and significance in both models; however, the magnitude of the direct and total effects varied between the two models. For example, the income variable had a lower magnitude in the 2020 model, whereas public transportation features had a higher magnitude.

Results of Model III, IV, and V showed that changes in weather variables (temperature, humidity, windspeed), COVID-related features, income, public transportation features, and changes in e-scooter availability were the main factors influencing the spatio-temporal changes in shared e-scooter usage in Austin during the Pandemic.

The study is not without limitations. One of the main limitations of the study was that the study relied on monthly device count data extracted from each trip as a proxy for daily e-scooter availability in each census tract since those data were not available. Moreover, the rebalancing information was also not included in the model. Future studies should include these variables. Another limitation of the study is that the study had to use city-level data for daily COVID cases as the census tract-level data were not available. The study could not include the changes in public transportation features in 2020, although transit supply was limited during 2020. However, we assumed that the changes in the frequency of public transportation services would be similar in each area irrespective of the number of bus stops in those locations. This study focused on the association between different pandemic-related policies with e-scooter usage without conducting any causal analysis. This is an avenue for future research. Finally, while the model results provided interesting findings regarding spatial dependence in the shared e-scooter data, they may be context-specific. Hence, future studies involving data from other cities, utilizing spatial models, are necessary to generalize the findings related to the impact of the COVID-19 pandemic on shared e-scooter usage.

Declaration of Competing Interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Suman Mitra reports financial support and article publishing charges were provided by the National Science Foundation.

Data availability

Data will be made available on request.