Introduction

The ongoing COVID-19 pandemic has affected numerous aspects of activity-travel behaviour, including activity participations/engagements, activity location choices, and travel mode choices. These changes can partially be attributed to the public health policies that have been implemented to curb the spread of COVID-19, including the introduction of stay-at-home orders, the temporary closure of schools and businesses, and the implementation of social distancing guidelines. However, changes in activity-travel behaviour can also be attributed to the impacts of the pandemic on the perceptions of risk associated with leaving one’s home and the use of different modes of travel. Consequently, travellers have demonstrated an increased preference for individual modes (such as private vehicles and active modes) and a decreased preference for shared modes (such as public transit, taxis, and ride-sourcing). As part of this shift in preferences, some have turned to cycling as a means for mobility needs, partly due to the ability to maintain social distancing while travelling (Beck and Hensher 2020a; Bucsky 2020; Marsden et al. 2021). In addition, cycling has the potential to help individuals supplement their mobility during the pandemic, particularly those without private vehicle access. In response to the growing prevalence of cycling during the pandemic, cities worldwide have installed temporary cycling infrastructure and implemented car-free hours on certain roadways (Budd and Ison 2020; Taylor 2020; City of Toronto 2021).

The increased prevalence of cycling during the pandemic has highlighted the potential to encourage individuals to shift to active modes, even after pandemic-related restrictions are lifted (Qu, Gao and Li 2020). Besides, active modes offer a more environmentally friendly alternative to private vehicles, particularly for short-distance trips. Additionally, encouraging people to use active modes can help build new habits that persist even after the pandemic. These behavioural changes can help to ensure that the increases in the popularity of private vehicles that have been observed during the pandemic (Beck and Hensher 2020b; de Haas, Faber and Hamersma 2020) will not persist post-pandemic. Amid the shortage of bicycles available for purchase during the pandemic (Goldbaum 2020; Mazerolle 2021), bicycle-sharing systems can help satisfy the increased demand for cycling. Consequently, understanding the factors influencing bicycle-sharing demand can help inform decisions about where to allocate additional capacity and where to locate new stations.

This study aims to investigate the factors influencing demand for bicycle-sharing trips during the COVID-19 pandemic in Toronto. Using data from Bike Share Toronto, this study utilizes econometric models to examine the relationships between built environment characteristics and monthly station-level bicycle-sharing demand. A multilevel model is estimated using data from 2019 and 2020 to examine the influence of the pandemic on the effects of weather, socio-demographic variables, and the built environment on the generation of bicycle-sharing trips. Additionally, a geographically weighted regression (GWR) model is estimated to capture spatial heterogeneity in the influence of these factors on the generation of bicycle-sharing trips during the COVID-19 pandemic. The results of this study can provide insights into the extent to which the COVID-19 pandemic has impacted the utilization of bicycle-sharing systems, which can help inform decisions regarding the management of the system. The remainder of the paper is organized as follows: first, the findings of pertinent studies are presented and discussed. Then, background information about the study area is provided, the key descriptive statistics are presented, and the formulations of the two econometric models used in this paper are described. Finally, the final specifications of the models are presented, and policy recommendations are discussed.

Literature review

Pre-pandemic bicycle-sharing studies

As bicycle-sharing systems become increasingly prevalent in urban regions, a growing body of research has explored the main determinants of their demand and usage patterns. The dependent variables in these studies are bike-sharing departures or arrivals aggregated at different spatial units (e.g., stations, traffic analysis zones, or census block groups) and separated by different temporal features (e.g., weekdays versus weekends, peak hours versus non-peak hours) or user groups (e.g., members versus casual users, low-income versus medium/high-income, etc.) (Faghih-Imani et al. 2014; El-Assi et al. 2017; Caspi and Noland 2019; Scott and Ciuro 2019). The independent variables can be categorized into three broad groups: trip-maker (e.g., age, gender, income, education, etc.), trip (e.g., purpose, time of day, weather, membership type, etc.), and built environment characteristics.

Trip maker characteristics

In terms of socio-demographic attributes, studies have found that bicycle-sharing users tend to be young in age (Nickkar et al. 2019; Wang and Lindsey 2019). Based on a study on bicycle sharing usage among Millennials, Gen Xers, and Baby Boomers, Wang et al. found that older Millennials make the highest number of trips, making up over 35% of all Citi Bike trips in New York City (Wang et al. 2018). In addition, males are disproportionately more likely to use bicycle-sharing services than females, indicating a potential gender gap in bicycle-sharing usage (Faghih-Imani and Eluru 2015; Nickkar et al. 2019; Wang and Lindsey 2019). The usage of bicycle-sharing services has been found to be positively associated with income and education level (Bao, Shi and Zhang 2018; Eren and Emre Uz 2020; Munira and Sener 2020), although some studies have reported a negative association between income and bicycle sharing usage as well (Hankey et al. 2017; Wang and Lindsey 2019).

Trip characteristics

Regarding trip purposes, the literature suggests that bicycle-sharing is used for both mandatory activities, such as work and school, and discretionary activities, such as recreational and social activities (Dung Tran et al. 2015; Shen et al. 2018; Wang et al. 2018). Studies have found that the presence of a university or college campus positively contributes to cycling activity (Scott and Ciuro 2019; Munira and Sener 2020), further supporting the utilitarian purposes of bike-share trips. The system may also provide first- and last-mile connections for public transit trips (Xing, Wang and Lu 2020). Studies also suggest that users’ predominant trip purposes may vary based on their membership type. For instance, non-members or “casual” users generally make recreational trips while members or “regular” users make more mandatory trips (Dung Tran et al. 2015; Nickkar et al. 2019). Moreover, bicycle-sharing systems have been found to have distinct temporal usage patterns, which may, in turn, depend on users’ trip purpose and membership type. For example, the bike-share systems of Montreal and New York City were found to be used more extensively in the PM period (Faghih-Imani et al. 2014; Faghih-Imani and Eluru 2016). Again, annual members use bike-share more frequently on weekdays, whereas casual members use these services more on weekends (Faghih-Imani and Eluru 2016; El-Assi et al. 2017).

Weather conditions play a significant role in bicycle sharing demand, although weather effects appear to depend on the geographical context (Faghih-Imani et al. 2014; Faghih-Imani and Eluru 2016; El-Assi et al. 2017; Bao, Shi and Zhang 2018; Shen et al. 2018; Ashqar et al. 2019; Scott and Ciuro 2019; Xing, Wang and Lu, 2020). For example, Ashqar et al. and Shen et al. found that temperature has a negative effect on ridership in San Francisco and Singapore, respectively (Shen et al. 2018; Ashqar et al. 2019). On the contrary, Faghih-Imani et al., El-Assi et al., and Bao et al. concluded that warmer weather increases ridership in relatively colder regions like Montreal, Toronto, and New York City (Faghih-Imani et al. 2014; El-Assi et al. 2017; Bao, Shi and Zhang 2018). Precipitation and snowfall are consistently found to have negative influences on ridership (El-Assi et al. 2017; Shen et al. 2018; Wang et al. 2018; Scott and Ciuro 2019).

Land use and built environment attributes

Bicycle-sharing ridership and usage patterns are significantly impacted by land use and built environment attributes. Distances of up to 800 m have been used to capture the influence of land use and built environment characteristics at the station level (Eren and Emre Uz 2020). However, the studies reviewed commonly use distances ranging between 200 and 500 m. In terms of density, previous studies suggest that station-level activity is higher in areas with higher population and/or employment densities (Faghih-Imani et al. 2014; Faghih-Imani and Eluru 2016; Wang et al. 2018; Scott and Ciuro 2019). More diverse land use has also been found to increase bicycle-sharing usage (Shen et al. 2018; Wang and Lindsey 2019). In terms of land use categories, studies typically use points-of-interest (POIs) to classify the land uses surrounding bicycle-sharing trip origins and destinations (Bao, Shi and Zhang, 2018; Shen et al. 2018; Xing, Wang and Lu, 2020). Using a K-means clustering algorithm to classify stations based on their predominant POIs and then using GWR models, Bao et al. identified spatial variability in the effects of land use on ridership (Bao, Shi and Zhang 2018). Additionally, parks and restaurants have been found to increase usage (Wang et al. 2016; Alcorn and Jiao 2019), particularly for casual users (Dung Tran et al. 2015).

Furthermore, bicycle-sharing demand appears to be significantly affected by the presence of and/or proximity to transportation infrastructures, such as subway stations, cycling lanes, and road intersections. In terms of the number of intersections surrounding a station, some studies have noted a negative influence (El-Assi et al. 2017; Wang and Lindsey 2019), while others have found a positive influence on ridership (Shen et al. 2018; Wang et al. 2018). Moreover, the length of major roads is a potential deterrent to the use of bicycle-sharing services (Faghih-Imani et al. 2014; Faghih-Imani and Eluru 2016), whereas the length of cycling infrastructure increases ridership (El-Assi et al. 2017; Bao, Shi and Zhang 2018; Shen et al. 2018; Scott and Ciuro 2019; Wang and Lindsey 2019). The availability of dedicated cycling lanes, especially off-road paths, is likely to increase users’ perception of cycling comfort and safety, which may encourage them to use bicycle-sharing. The number of bus stops surrounding a station has been found to increase ridership, although its effect is not consistently significant (Faghih-Imani and Eluru 2015; Shen et al. 2018). Meanwhile, the presence and proximity of subway stations are positively correlated with ridership (Faghih-Imani and Eluru 2016; El-Assi et al. 2017; Shen et al. 2018; Wang et al. 2018). This finding aligns with the use of bicycle-sharing systems for commuting purposes, especially in complement with rapid transit.

Bicycle-sharing studies focusing on the COVID-19 pandemic

Since the start of the COVID-19 pandemic, numerous studies have investigated its effect on urban mobility, especially how it relates to changed activity-travel behaviour and modal preferences. The studies found that attitudes towards shared modes, such as ride-hailing and transit, have been mainly negative compared to individual modes, such as driving and cycling, because of the higher perceived risk of infection on shared modes (de Haas, Faber and Hamersma, 2020; Shamshiripour et al. 2020). In particular, active (and individual) travel such as cycling and walking have experienced an increase relative to other modes (Beck and Hensher 2020b; Bucsky 2020; Marsden et al. 2021) due to a potential aversion towards public modes, and due to the health benefits, it offers (Budd and Ison 2020). As a hybrid mode, bicycle-sharing has been perceived to offer a greater level of safety from the virus than public transit, but is not as safe as entirely individual modes such as driving or walking (Nikiforiadis et al. 2020). This hybrid nature makes the impact on bicycle-sharing quite unique, which warrants in-depth investigation.

However, compared to the large number of studies that have analyzed COVID-19 and urban mobility, only a handful have focused on the impact of the pandemic on bicycle-sharing demand. The findings of these limited studies suggest that, as with other modes, bicycle-sharing systems have been significantly affected by the pandemic (Mehdizadeh Dastjerdi and Morency 2022). Specifically, ridership in the early days of the pandemic was found to be adversely impacted (Teixeira and Lopes 2020; Shang et al. 2021), with the number of bicycle-sharing trips being negatively correlated to the number of daily cases (Hu et al. 2021; Padmanabhan et al. 2021). The studies also consistently report an increase in trip duration (Teixeira and Lopes 2020; Hu et al. 2021; Padmanabhan et al., 2021; Shang et al., 2021). Given the negative attitudes towards shared modes, some of the studies infer that bicycle-sharing may be used in place of public transit (Pase et al. 2020; Teixeira and Lopes 2020), which partly explains why trip durations are longer during the pandemic. Furthermore, the decrease in bicycle sharing usage has been found to be more prominent among annual members than casual users (Hu et al. 2021; Padmanabhan et al. 2021), potentially because regular users tend to use the service for commuting purposes, which may no longer be required during the pandemic due to work-from-home policies. Studies also indicate that bicycle-sharing systems are more resilient than public transit since they have experienced a less significant drop in ridership (Teixeira and Lopes 2020; Hu et al. 2021).

The literature also suggests that the pandemic has had differential effects on bike-share users of different socio-economic statuses. For example, using generalized additive (mixed) models of Chicago’s Divvy system usage during the pandemic, Hu et al. found that areas with higher income levels experienced a more drastic decrease in ridership (Hu et al. 2021). This finding is in line with an analysis of the Citi Bike system in New York City, which indicates that higher-income areas can socially distance more than low-income areas, even though they have a higher percentage of essential workers (Pase et al. 2020). Survey responses from bike-share users in San Antonio (TX) suggest that people unemployed by the pandemic are most likely to increase their use of the system (Jobe and Griffin 2021). These findings suggest that the pandemic has a disproportionate effect across the population and highlights the potential for bicycle-sharing systems to help ensure that everyone’s mobility needs are met.

While existing studies provide useful insights on the general impacts of COVID-19 on bicycle-sharing systems, most of them have focused on the immediate short-term impact after the onset of the pandemic. Studies that have undertaken a comprehensive analysis of how socio-demographic, land use, built environment, and weather attributes affect long-term demand during the pandemic are rare. To contribute to this gap in the literature, our study utilizes ridership data from 2019 and 2020 to investigate the relationships between these factors and bicycle-sharing demand during the pandemic and how the relationships have evolved compared to the pre-pandemic period in Toronto. The study also analyzes the spatial heterogeneity of the relationships to better understand how the system is being used during the pandemic by different population groups and across different areas of the city.

Study area and background

Bike Share Toronto is the third-largest public bicycle-sharing system in North America (Bike Share Toronto 2021). Its service started in 2011 with 80 stations and 1,000 bicycles in the downtown core of Toronto. By 2019, it had expanded to include 465 stations and 5,000 bicycles distributed over 100 km2 of the city. In 2020, Bike Share Toronto further expanded its service by setting up 160 new stations in 30 new neighbourhoods. This resulted in a system that includes 625 stations and 6,850 bicycles distributed over 200 km2 of the city (see Fig. 1), with an average daily ridership of about 8,000 trips. Toronto is the most populous city in Canada and cycling is one of its fastest-growing transportation modes. According to the 2016 Census, 2.7 percent of Torontonians bike to work (city-wide average), up from 1.7 percent in 2006, with some downtown neighbourhoods in the range of 15 to 30 percent (City of Toronto 2018). Bike Share Toronto plays a key role in catering to the growing demand for bicycle trips in the city. Moreover, about 58% of all the bicycle-sharing stations are located within one kilometer of subway stations, providing a solution for the “last-mile problem” of public transit users. The city's well-connected cycling network and walkable urban form also offer a good opportunity for the success of the service.

Fig. 1
figure 1

Bike Share Toronto stations in the study area

Full size image

During the COVID-19 pandemic, the ridership of the service declined following the declaration of the state of emergency by the Government of Ontario on March 17, 2020 (see Fig. 2). However, bike-share usage rebounded quickly as soon as the city entered the first stage of the three-stage “reopening” plan on May 19, 2020 where restrictions on retail stores and elective surgeries were lifted. The ridership continued to increase steadily over the next few months as the city moved through stage 2 (where shopping malls, personal care services, restaurants, bars, and recreational facilities could operate as long as they complied with specific health and safety requirements) and stage 3 (where nearly all businesses and public spaces were allowed to reopen with health measures in place) of the reopening plan. As shown in Fig. 2, the ridership started declining as the weather turned cooler and Toronto entered a second wave of the pandemic. Nonetheless, Bike Share Toronto reported a record ridership of 2.9 million in 2020—the highest ridership in its 10-year program history (Bike Share Toronto 2021)—making it an interesting case study for this paper.

Fig. 2
figure 2

Seven-day moving average of relative volume of different modes of transport across the pandemic, from January 13th, 2020 to December 31st, 2020, compared to the baseline volume on January 13th, 2020

Full size image

On May 16, 2020, the City of Toronto launched the ActiveTO program (City of Toronto 2021) to support essential trips, front-line workers, and vulnerable road users. The program’s goal was to ensure that the residents have space to get around while adhering to the physical distancing protocols. The program included initiatives like closing major roads for active transportation on weekends, expanding the cycling network, and implementing quiet streets. Approximately 40 kms of on-street cycling lanes were approved for accelerated installation in 2020, focusing on routes that mirror major transit routes and/or connected trails and greenspaces.

Figure 2 presents the seven-day moving average of the relative volume of different travel modes throughout 2020 compared to the corresponding baseline volume on January 13, 2020. Specifically, the relative mobility ratio of a mode is calculated relative to its own baseline volume. The data for driving, transit, and walking modes are obtained from the Apple Mobility Trends Reports (Apple 2020), which approximates human mobility by the number of daily requests for directions from Apple Maps users. The bike-share demand profile is compiled from the Bike Share Toronto ridership data. From the figure, it is observed that the overall mobility trend is consistent across the four modes; however, the scales of change are not uniform. Transit suffered the most during the pandemic, losing the greatest percentage of trips and recovering at the slowest rate among all the modes. Bike-sharing, on the other hand, exhibited the fastest recovery and greatest rebound scale. It is important to note that temperature may also play an essential role in this rapid increase in bike-share ridership since cycling demand has an intrinsic seasonal trend. Moreover, the service expansion of the Bike Share Toronto system itself may also contribute to the high rebound rate. Nonetheless, it is evident that bicycle-sharing serves as a resilient travel mode during the pandemic potentially due to its high health benefit and low-risk perception, especially for those without access to a private vehicle.

Data description

The primary data used in this study was obtained from the City of Toronto’s Open Data Portal which provides historical Bike Share Toronto ridership information dating back to October 2014. The dataset includes information about trip start time and date, trip end time and date, trip duration, start station, the end station, bicycle id, and user type, i.e., whether the user was an annual member or a casual user. Additionally, the stations' capacity (i.e., the number of docks per station) and coordinates are also provided. In this study, we used ridership data for the years 2019 and 2020 to cover periods representative of the pre-pandemic “base” case as well as the pandemic case. After cleaning the data set by removing invalid/incomplete trip records (i.e., trips of durations less than 30 s and trips whose station information was not available in the database), the total number of trips completed using Bike Share Toronto in 2019 and 2020 were 2,423,305 and 2,905,158, respectively.

Table 1 shows the characteristics of the trips made over the two years. It is observed that the composition of trips in terms of user types is considerably different in 2020 compared to that in 2019. The proportion of trips generated by annual members decreased from 76.24% in 2019 to 64.31% in 2020. Considering that annual members use the service more for commuting purposes while casual members use it more for leisure/recreational purposes (Kaviti et al. 2019; Wang and Lindsey 2019), it is reasonable that the proportion of trips made by annual members decreased in 2020 due to the work-from-home policies implemented during the pandemic. Similarly, the system is found to be busier during the weekends in 2020 (presumably due to an increase in the proportion of recreation and leisure trips), whereas it used to be busier during the weekdays in 2019. Also, the average trip duration increased during the pandemic, which is in line with the findings of other studies (Hu et al. 2021).

Table 1 Description of bike share trips
Full size table

To investigate the factors affecting bike-share demand in the context of the pandemic, several categories of independent variables are considered in this study. These include weather information, socio-demographic attributes, land use, and built environment characteristics, and bicycle infrastructure variables. Monthly weather information collected at the Billy Bishop Toronto city airport (also known as the ‘Toronto Island Airport’) weather station was compiled from the historical weather data maintained by Environment Canada. The weather variables tested in this study include the mean temperature (°C), total precipitation (mm), relative humidity (%), and wind speed (kph). Socio-demographic characteristics such as population density, median household income, and age of the population were obtained at the Dissemination Area (DA) level of the bike share stations from the 2016 Canadian Census data (DA is the smallest standard geographic area for which all census data are disseminated in Canada). The number of jobs in each Traffic Analysis Zone (TAZ) was extracted from the 2016 Transportation Tomorrow Survey, which is a large-scale household travel survey conducted every five years in the Greater Golden Horseshoe Area (Data Management Group 2016).

The spatial determinants of bike-share demand at each station were examined using two classes of variables: (a) land use and built environment attributes, and (b) bicycle infrastructure variables. These determinants are considered at the station buffer level—a common approach adopted in bike-share literature (Faghih-Imani et al. 2014; Faghih-Imani and Eluru 2015, 2016; El-Assi et al. 2017; Wang et al. 2018; Scott and Ciuro 2019). Specifically, a 200 m buffer around each station was found to be appropriate, considering the distances between Bike Share Toronto stations and the dense urban form of the city (El-Assi et al. 2017; Scott and Ciuro, 2019). To examine the influence of land use, the areas of seven land use types within the station buffers were compiled from the DMTI Spatial Inc. database. These include: (1) open area, (2) parks and recreational, (3) residential, (4) waterbody, (5) commercial, (6) government and institutional, and (7) resource and industrial. We also tested the influence of various categories of Enhanced Points of Interest (EPOI) in the station buffers including food services EPOI, health care EPOI, retail trade EPOI, finance and insurance EPOI, professional, scientific, and technical services EPOI, arts, entertainment, and recreation EPOI, educational services EPOI (university/college) etc. To examine the influence of built environment, dummy variable indicating the presence of roadway intersections and the number transit stops (i.e., bus and streetcar stops and subway stations) within the station buffers were derived from the Toronto Transit Commission GTFS feed and the roadway intersection file obtained through the City of Toronto Open Data Portal, respectively. We also tested the effects of minor and major road lengths and bus and streetcar route lengths in the station buffers.

The cycling network of the City of Toronto includes many types of infrastructure, such as cycle tracks, bicycle lanes, shared roadway routes and multi-use trails. To test the effect of these infrastructure on bike-share demand, we compiled the total length of bicycle facilities in the station buffers (i.e., the summation of the lengths of available bike lanes, cycle tracks, multi-use pathways quiet street cycling routes, suggested on-street routes etc.). We also tested the effects of the number and capacity of bike-share stations in the buffer, the capacity of individual bike-share station, and the distance of each station from the Central Business District (CBD). A one-month lag of the dependent variable is also included in the analysis to capture the effect of short-term changes like relocation of a station from one zone to another or road construction in the vicinity of a station. Table 2 provides a descriptive summary of some of the important variables compiled for this study.

Table 2 Summary of variables
Full size table

Temporal usage pattern during the pandemic

Figure 3 compares the monthly trip generation patterns of Bike Share Toronto in 2019 and 2020. As expected, the 2020 ridership experienced a sharp decline towards the beginning of the stay-at-home order in Ontario (mid-March), reaching its lowest point around April 2020. From there, the ridership recovered steadily as the city went through the phased reopening plan and the temperature started to increase. Interestingly, the ridership remained consistently greater than that of 2019 for the rest of the year even as the winter season set in, and the city reverted to a lockdown situation in November 2020 in response to the second wave of the pandemic. This further indicates that some people have turned to bike-sharing to satisfy their mobility needs during the pandemic.

Fig. 3
figure 3

Monthly patterns of Bike Share Toronto trip generation in 2019 and 2020

Full size image

Figure 4 presents the daily patterns of trip generation over the two years. Both the daily profiles have a distinctive morning-and-afternoon-peak “commuting” pattern on weekdays and a midday-peak “leisure” pattern on weekends. However, the demand of the service during the weekday peak periods in 2020 is much lower than that of 2019, especially during the morning peak period. On the contrary, the off-peak period demand of the service is higher in 2020 compared to that in 2019. In fact, the overall weekly profiles indicate that the system was more popular during the weekends in 2020 while it was more popular during weekdays in 2019. These changes further support the argument that the proportion of commuting bike-share trips has declined during the pandemic while the proportion of exercise/leisure trips has increased. We also observe an increased mid-day usage during the weekdays in 2020 compared to pre-pandemic period, which might be indicative of more flexible working hours (due to more work-from-home policies) and thus more mid-day non-discretionary trips via bike-share.

Fig. 4
figure 4

Daily patterns of Bike Share Toronto trip generation in 2019 and 2020

Full size image

These changes also provide useful hint about the socio-demographic profile of the bike-share users in Toronto. Specifically, it indicates that a considerable segment of the user population is possibly white-collar workers (who had the opportunity of working from home during the pandemic) or younger individuals like post-secondary students (who have access to smartphones and credit cards to access the service and most likely had less part-time work available during the pandemic). Such a general impression of the user profile can be useful for future policy regulation and system expansion, especially given that user socio-demography is usually unavailable in the passively collected bike-share ridership data.

Spatial usage pattern during the pandemic

The average daily trip generation and attraction rates of each station in 2019 and 2020 are shown in the maps of Figs. 5 and 6. It is observed that stations located in the financial district (Planning District 1) of the City of Toronto suffered the highest decline in usage during the pandemic compared to 2019. At the same time, stations located outside the CBD and along the waterfront and near recreational areas became more popular. A disproportionately large increase in usage is observed along the southwest segment of the waterfront, which might be attributed to the increased use of bike-share for recreational purposes during the pandemic. Waterfront Toronto has a well-connected network of multi-use trails and parks and is one of Toronto’s most widely used biking routes. It became even more popular during the pandemic because it allowed people to enjoy the outdoors while maintaining physical distancing protocols. The southwest segment of the trail near Humber Bay Arch Bridge, Marilyn Bell Park, Coronation Park, and Queens Quay offers greener space than the eastern industrial segment, and as such became hot spot for cyclists during the pandemic. Overall, there exists spatial heterogeneity in the relative change of bike-sharing usage during the pandemic, which necessitates a detailed investigation of the factors related to such heterogeneity.

Fig. 5
figure 5

Spatial patterns of average daily trip generation in the Bike Share Toronto system in 2019 (left) and 2020 (right)

Full size image
Fig. 6
figure 6

Spatial patterns of average daily trip attraction in the Bike Share Toronto system in 2019 (left) and 2020 (right)

Full size image

Methodology

This study adopts a two-stage methodology to investigate the effects of weather, socio-demographic attributes, and land use and built environment on bike-share ridership in the context of the pandemic. First, multilevel modelling is used to analyze the factors affecting monthly bike-share trip generation before and during the pandemic. Then, to better understand how the factors vary by region during the pandemic, geographically weighted regression (GWR) analysis is performed at the station level. In each stage, separate models are estimated for weekdays and weekends to account for their different demand profiles. Since the number of trips is a non-negative variable (i.e., inherently skewed from the normal distribution), the natural logarithm of the trip counts is used as the dependent variable of the models (see Table 2). Thus, in this study, we estimate four models in total:

  1. 1.

    Station-level weekday trip generation multilevel model: The natural logarithm of the number of trips generated on weekdays from each station in each month between January 2019 and December 2020. The sample used to estimate the model includes 11,376 station-month observations across 609 stations.

  2. 2.

    Station-level weekend trip generation multilevel model: The natural logarithm of the number of trips generated on weekends from each station in each month between January 2019 and December 2020. The sample used to estimate the model includes 11,336 station-month observations across 608 stations.

  3. 3.

    Average weekday trip generation GWR model: The natural logarithm of the average number of trips generated on weekdays from each station during the pandemic (Mar 2020–Dec 2020). The sample used to estimate the model includes 609 station-specific observations.

  4. 4.

    Average weekend trip generation GWR model: The natural logarithm of the average number of trips generated on weekends from each station during the pandemic (Mar 2020–Dec 2020). The sample used to estimate the model includes 609 station-specific observations.

Multilevel/linear mixed-effects model

The bike-share ridership data has an inherent multilevel structure with repeated observations over time at each station. In this empirical analysis, we observe the monthly trip generation rates of each station between January 2019 and December 2020. To account for the repeated observations, we employ multilevel modelling to investigate the factors affecting Bike Share Toronto ridership during the study period. The modelling framework explicitly recognizes the dependencies associated with the trip generation rates of the same station. Specifically, we used a two-level random intercept model structure where levels 1 and 2 pertain, respectively, to the monthly counts of trips grouped within stations. Multiple fixed effects are captured (as shown in Table 2), and stations are modeled as a random effect to account for differences in ridership among them throughout the study period. The model takes the following form:

$$y_{ti} = \beta_{0i} + \beta x_{ti} + \varepsilon_{ti} , t = 1, \cdots ,T, \quad i = 1, \cdots ,I$$
(1)
$$\beta_{0i} = \gamma_{00} + u_{0i} , \quad i = 1, \cdots ,I$$
(2)

where \({y}_{ti}\) is a \(n\times 1\) vector of observed values (in this case, the natural logarithm of number of trips generated on each month \(t\) for station \(i\)), \({x}_{ti}\) is a \(n\times k\) matrix of observed independent variables for each observation, and \(\beta\) is a \(k \times 1\) vector of coefficients. \({\beta }_{0i}\) is a group-dependent intercept, consisting of an average intercept \({\gamma }_{00}\) and a group-dependent deviation (error) \({u}_{0i}\). \({\varepsilon }_{ti}\) and \({u}_{0i}\) are assumed to follow a normal distribution with mean 0 and variance \({\sigma }^{2}\).

This study estimated two random intercept multilevel models—one for weekday trip generation and the other for a weekend trip generation—to account for the distinct demand profiles across the two types of days. The models are estimated using the maximum likelihood method using the “lme4” package in the statistical software “R” (Bates et al. 2015). To explicitly capture the effect of the pandemic, the independent variables interact with dummy variables indicating whether the observation is from the pre-pandemic period (Jan 2019–Feb 2020) or the pandemic period (Mar 2020–Dec 2020).

Geographically weighted regression model

To explore the local variations of the impacts of socio-demographic, land use, and built environment attributes on bike-share demand during the pandemic, this study uses the Geographically Weighted Regression (GWR) model. The GWR is a useful technique to account for spatial non-stationarity, i.e., spatial heterogeneity of parameter estimates. It incorporates spatial variation into the modeling process by considering a local-specific relationship of the response variable to the independent variables. The GWR methodology extends the traditional global framework using a weighting scheme that puts more weight on geographically closer observations when estimating local parameters (Fotheringham, Brunsdon and Charlton 2002). The model can be represented as follows:

$$y_{i} = \beta_{0} \left( {u_{i} ,v_{i} } \right) + \mathop \sum \limits_{k = 1}^{K} \beta_{ik} \left( {u_{i} ,v_{i} } \right)x_{ik} + \varepsilon_{i} , \quad i = 1, \cdots ,n$$

Here \({y}_{i}\) is the \({i}^{\mathrm{th}}\) observation of the dependent variable (in this case, natural logarithm of the average number of weekday or weekend trips generated from each station \(i\) during the pandemic). \(n\) is the number of observations (i.e. total number of stations), and \(K\) is the number of independent variables. \({x}_{ik}\) represents the \({k}^{th}\) the independent variable for the \({i}^{th}\) station. The coefficients \({\beta }_{ik}\left({u}_{i},{v}_{i}\right)\) vary conditional on the location of the stations; \(\left({u}_{i},{v}_{i}\right)\) denotes the spatial coordinates of the \({i}^{th}\) station; and \({\varepsilon }_{i}\) is the random disturbance terms of independent and identical distribution.

The estimated coefficient, \({\beta }_{ik}\left({u}_{i},{v}_{i}\right)\) at station \(i\), could be obtained by the following equation:

$$\beta_{ik} \left( {u_{i} ,v_{i} } \right) = \left( {X^{T} W\left( {u_{i} ,v_{i} } \right)X} \right)^{ - 1} X^{T} W\left( {u_{i} ,v_{i} } \right)y$$

Here \(X=\left[\begin{array}{ccc}1& \begin{array}{cc}{x}_{11}& \dots \end{array}& {x}_{k1}\\ 1& \begin{array}{cc}{x}_{12}& \dots \end{array}& {x}_{k2}\\ \begin{array}{c}\vdots \\ 1\end{array}& \begin{array}{cc}\begin{array}{c}\vdots \\ {x}_{1n}\end{array}& \begin{array}{c}\ddots \\ \dots \end{array}\end{array}& \begin{array}{c}\vdots \\ {x}_{kn}\end{array}\end{array}\right]\), \(y=\left[\begin{array}{c}{y}_{1}\\ {y}_{2}\\ \begin{array}{c}\vdots \\ {y}_{n}\end{array}\end{array}\right]\), and \(W\left({u}_{i},{v}_{i}\right)\) is an \(n\times n\) spatial weight matrix that can be expressed as \({W}_{i}\):

$$W_{i} = \left[ {\begin{array}{*{20}c} {w_{i1} } & 0 & {\begin{array}{*{20}c} \cdots & 0 \\ \end{array} } \\ 0 & {w_{i2} } & {\begin{array}{*{20}c} \cdots & 0 \\ \end{array} } \\ {\begin{array}{*{20}c} \cdots \\ 0 \\ \end{array} } & {\begin{array}{*{20}c} \cdots \\ \cdots \\ \end{array} } & {\begin{array}{*{20}c} {\begin{array}{*{20}c} \cdots \\ \cdots \\ \end{array} } & {\begin{array}{*{20}c} \cdots \\ {w_{in} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \right]$$

where the diagonal element \({w}_{ij}\) \(\left(j=1,\cdots ,n\right)\) represents the spatial weight given to station \(j\) in the calibration of the model for station \(i\). Several common approaches have been used in the literature to define the weight matrix \({W}_{i}\), such as the Gaussian, tri-cube, and bi-square functions (Fotheringham, Brunsdon and Charlton 2002). The bandwidth of the function, i.e., cut off for the distance to assign non-zero weights, could either be fixed (based on distance) or adaptive (based on a specific number of neighbors). The GWR model outcomes are sensitive to bandwidth selection (Yu and Peng 2019). A small bandwidth covering a small number of observations may result in unstable fits, while too large a bandwidth may introduce bias (Munira and Sener 2020). In this study, we use an adaptive bandwidth coupled with a bi-square weighting scheme to ensure that each local regression encompasses enough regression points regardless of the surrounding station density. The bi-square function is written in the following form:

$$w_{{ij}} = \left\{ {\begin{array}{*{20}l} {\left[ {1 - \left( {d_{{ij}} /h_{i} } \right)^{2} } \right]_{i}^{2} } \hfill & {if\;d_{{ij}} < h} \hfill \\ 0 \hfill & {otherwise{\text{ }}} \hfill \\ \end{array} } \right.$$

\({d}_{ij}\) is the distance between two observations \(i\) and \(j\), and \({h}_{i}\) is the maximum distance to assign nonzero weights i.e., the bandwidth. Observations located beyond bandwidth \({h}_{i}\) are assigned a weight of zero. To find the optimal bandwidth (i.e., the minimum number of neighbors to be included in calibrating the local models), we used the cross-validation (CV) approach.

As previously mentioned, we estimated two GWR models corresponding to the pandemic period (Mar 2020–Dec 2020)—one for weekday trip generation and the other for weekend trip generation. The models are estimated using the “spgwr” package in the statistical software “R” (Bivand et al. 2020).

Empirical results

Multilevel/linear mixed-effects model

Table 3 presents the estimation results of the weekday and weekend random intercept multilevel models. The specifications of the final models are derived based on the accommodation of variables with proper signs and statistical significance. The critical value (1.96) of the t-stat with a 95% confidence limit is used as the threshold value of considering variables in the model. However, some parameters with t-stat values lower than 1.96 are retained in the models because the corresponding variables provide considerable insight into the behavioural process and policy relevance (Ortúzar and Willumsen 2011). Commonly used measures of model fit, such as Akaike’s Information Criterion (AIC), Bayesian Information Criterion (BIC), and − 2 log-likelihood values, are also reported. These models are compared to their respective null models (containing only random intercepts) to determine the degree to which the independent variables improved the model fit. The weekday null model has a − 2 log-likelihood of 33,171.3, an AIC of 33,177.3, and a BIC of 33,199.3. As shown in Table 3, adding the independent variables to the null model reduced the − 2 log-likelihood to 15,009.6, the AIC to 15,091.7, and the BIC to 15,392.6, signifying a better model fit. More importantly, adding these variables to the model decreased the proportion of total variance attributed to the group-specific intercepts, i.e., differences between the stations themselves. Nonetheless, the residual variance of the full model is still quite high, supporting the use of the random intercept multilevel model for the analysis of monthly bike-share ridership. A similar improvement of model fit is also observed when comparing the weekend full model (-2 log likelihood: 16,153.0, AIC: 16,235.1, and BIC: 16,535.8) to its corresponding null model (-2 log likelihood: 31,044.6, AIC: 31,050.6, and BIC: 31,072.6).

Table 3 Multilevel model estimation results
Full size table

In terms of fixed effects, most of the parameters are estimated with the expected signs and are statistically significant at the 95% confidence level. Many of the variables have similar effects on the monthly trip generation rates before and during the pandemic, reinforcing their effect on bike-share demand. However, the extent of their effects varies significantly between the two periods (as shown by the paired t-test results in Table 4), indicating that the nature of bike-share usage changed significantly during the pandemic.

Table 4 Paired t-test results comparing the pre-pandemic and pandemic parameters
Full size table

Among the socio-demographic attributes, the population density of the station DA is found to be positively correlated with weekday and weekend trip generation, both before and during the pandemic. However, the pandemic estimates are significantly higher than the pre-pandemic estimates, indicating that areas of higher population density generate more bike-share trips during the pandemic than before the pandemic. The number of jobs in the station TAZ is positively correlated with weekday trip generation and negatively correlated with the weekend trip generation before the pandemic. This indicates the likely use of the bicycle-sharing system for daily work commute trips, which is consistent with the findings of previous studies (Faghih-Imani et al. 2014; El-Assi et al. 2017; Scott and Ciuro 2019). However, during the pandemic, this variable is found to be insignificant in the weekday model, whereas it has more negative effects (than before the pandemic) in the weekend model. This finding further indicates that the proportion of commuting bike-share trips has declined significantly during the pandemic, potentially due to the widespread practice of work-from-home policies.

In terms of age distribution, areas with more people aged 20–34 years constitute the primary market for bike-share both before and during the pandemic. On the contrary, areas with more people younger than 20 years or older than 34 years are less likely to generate bike-share trips. The median household income of station DA is negatively associated with pre-pandemic and pandemic bike-share demand in both models. This indicates that, after controlling for other exogenous factors, higher-income neighbourhoods generate fewer bike-sharing trips, which is consistent with the findings of previous studies (Hankey et al. 2017; Wang and Lindsey 2019). In fact, these higher-income areas are found to generate even fewer weekday bike-sharing trips during the pandemic. People residing in the wealthier neighbourhoods tend to work at jobs better suited to social distancing, resulting in a sharper drop in commuting trips which consequently lowers the weekday bike-share demand in these areas. This suggests that the bike-share service’s income disparity may have become more polarized during the pandemic, which is consistent with the recent literature (Pase et al. 2020; Hu et al. 2021).

Variables characterizing the land use and built environment in the vicinity of the stations are important predictors of bike-share trip generation, both before and during the pandemic. Based on the model estimates, it is found that there is an increased demand for recreational trips during the pandemic. For example, in the weekday model, the proportion of the station buffer classified as a park presents a significant negative relationship in the pre-pandemic period which flips to a significant positive relationship in the pandemic period. Similarly, stations near water bodies and open areas are found to be more popular during the pandemic, especially on weekends. Both the models indicate that the number of food services EPOI in the vicinity of a station increases the usage of the station irrespective of the pandemic situation. However, this increase in usage is significantly greater during the pandemic, suggesting an increased demand for maintenance trips in this period. On the contrary, the presence of health care EPOI has a negative impact on trip generation, which worsens with the onset of the pandemic. This might be attributed to the reduced use by hospital workers during the pandemic or increased perception of risk associated with health care facilities, which discourages the use of nearby bike share stations.

The coefficient associated with the presence of a university/college campus within a station buffer is positive and significant irrespective of the pandemic. However, the effect is stronger during the pandemic. Given that educational institutions are closed for in-person classes during this period, these stations might be generating more non-education related trips. Moreover, weekday trip generation during the pandemic are found to be higher in stations that has metro/subway station within their buffers. This, coupled with the fact that the average bike-share trip duration increased considerably during the pandemic, indicates that bike-sharing may be substituting some weekday public transit trips. This is in line with the findings of previous research done on the topic (Pase et al. 2020; Teixeira and Lopes 2020; Hu et al. 2021; Wang and Noland 2021). As expected, the proximity of bike-share stations to Toronto’s CBD has a strong impact on ridership in both models. All effects are negative, meaning that regardless of the pandemic, the farther a station is from the CBD, the lower its monthly ridership. However, the effect is less negative during the pandemic indicating that stations located further from the CBD experienced a smaller decline in ridership.

Among the bicycle infrastructure variables, the length of bicycle facilities (e.g., bike lanes, cycle tracks, multi-use pathways, etc.) in the buffer presents significant positive signs in both models, which is consistent with prior studies (Faghih-Imani et al. 2014; Noland et al. 2016; Wang and Chen 2020). However, the effect is more positive during the pandemic, indicating that the pandemic trip generation is more sensitive to the availability of bicycle facilities within the station buffer. This might be attributed to the increased number of casual (recreational) users for whom the presence of bicycle facilities has been found to be an important motivator for cycling (City of Toronto 2019). In terms of station capacity, both models found that stations with a greater number of bicycle docks generate more trips irrespective of the pandemic situation. However, the effect of station size is greater on the pre-pandemic trip generation in the weekday model, whereas its impact is greater on the pandemic trip generation in the weekend model. The number of stations within a 200 m buffer area is significant for the pandemic period in both models and is negatively correlated with the number of trip generations. This suggests that multiple stations in close proximity to one another provide Bike Share users with more options to choose from, thereby distributing the trips amongst the nearby stations, which negatively impacts the station being analyzed. Thus, the optimal locations of new stations should be carefully selected by maximizing the total service area of the system and minimizing the overlap of the catchment areas of nearby stations, while ensuring that station-to-station distance is reasonable to provide improved and flexible service to the users.

Both the models indicate that there is a positive correlation between temperature and monthly bike-share ridership before and during the pandemic. However, the varying magnitudes of the parameter estimates indicate that temperature has the strongest effect on the decision to cycle on weekends during the pandemic. Finally, the one-month lag effect is positively correlated with a monthly trip generation throughout the study period. This implies that the consistent state of a station may cause an increase of trip activity over the months, although the effect is less positive during the pandemic. This might be attributed to the increased flexibility of users to change of station status or to the improved system performance during this period.

Geographically weighted regression model

In addition to the multilevel models, we estimated two GWR models to explore local variations of the impacts of socio-demographic, land use, and built environment attributes on bike-share demand during the pandemic. This analysis can potentially unveil inequalities in the impact of the pandemic across different areas of the city and on different population groups. Table 5 presents the estimation results of the weekday and weekend GWR models. The final models are derived based on the accommodation of variables with proper signs and statistical significance. Given that GWR is structurally different from multilevel modelling, the same variables might not be significant in both cases. Thus, the GWR models offer additional insights in terms of spatial heterogeneity of the influence of the explanatory factors. The models’ performances are compared with corresponding global models, which are equivalent to traditional ordinary least square regression models, in terms of R-square and AIC values. As shown in Table 5, the final GWR models have significantly smaller AIC values compared to the global models. This further emphasizes the spatial variation in the relationships between the predictors and bike-share activity during the pandemic. To finalize the specifications of the global models, we considered the signs and statistical significance of the estimates. Variance inflation factor analysis (VIF) was also conducted on the global models (see Table 7 in Appendix). The VIF measures the degree of multicollinearity between the ith independent variable and all other independent variables in a linear regression model. In our investigation, all variables had a VIF of less than 4, which is acceptable given the rule of thumb of excessive multicollinearity found in the literature (O’brien 2007).

Table 5 GWR model estimates
Full size table
Table 6 Elasticity estimates of GWR models
Full size table

We also tested the spatial non-stationarity of the study variables. Specifically, we investigated their spatial autocorrelation using the Global Moran’ I test (Moran 1950). The null hypothesis is that no spatial autocorrelation exists. The results showed that all the study variables are statistically significant at the 0.01 level and have positive z-score values, rejecting the null hypothesis and suggesting they are spatially clustered. This further highlights the need for a spatial model, like the GWR model, for this analysis.

Figure 7 presents the spatial variations of R-square obtained from the GWR models. It is found that the local R-squared estimates vary from 0.3 to 0.8 across the different neighbourhoods of the City of Toronto. The R-squared values are highest towards the peripheral suburban regions and lowest in the core downtown area. This indicates that the independent variables are not explaining as much of the variation of bike-share demand in downtown as the rest of the city. This emphasizes the importance of conducting an in-depth and more spatially focused analysis (with more relevant independent variables) for the downtown area of Toronto, which has a mix of different types of land uses and a higher bike share station density compared to the rest of the city. In general, the findings indicate that significant heterogeneity exists among the defining features when relating them to bike-share ridership during the pandemic, which the traditional global model cannot capture.

Fig. 7
figure 7

R-squared values derived from the weekday (left) and weekend (right) GWR models

Full size image

Figure 8 illustrates the spatial distribution, magnitude, and sign of the coefficient values describing the relation between the explanatory variables and bike-share ridership during the pandemic. To avoid clutter, only the local estimates that are significant at a 90% level of significance are shown in the figures. Darker colours are used to indicate stronger positive/negative effects in the figures. The significant estimates of most of the variables (except immigrants, residential land use, and parks) are found to have the same sign, which reinforces the effects of these variables on bike-share demand. However, their estimates showed notable spatial variation in terms of magnitude.

Fig. 8
figure 8figure 8figure 8figure 8

a Spatial distribution of the socio-economic coefficients from the GWR models (significant at p < 0.1) b Spatial distribution of the land use and built environment coefficients from the GWR models (significant at p < 0.1) c Spatial distribution of the bicycle infrastructure coefficients from the GWR models (significant at p < 0.1)

Full size image

Among the socio-economic variables, the effect of population density is similar in the weekday and weekend models, as observed in Fig. 8a. All the significant estimates in the study region are positive, indicating that higher population density is associated with greater bike-sharing demand during the pandemic. The effect is the strongest in the neighbourhoods located in the southwest just outside the financial district and the lowest in the core of the CBD. Similar to the multilevel models, the coefficient estimates of the age variable in the GWR models is found to be negative, confirming that areas with more people younger than 20 years generate fewer bike share trips during the pandemic. This might be attributed to the closure of educational institutions which likely reduced the overall travel demand of this age category. This could also be related to the need to have a credit card to use bike share and the relatively high cost of membership. However, the effect of the variable in the downtown core is less negative on weekdays than on weekends. This might indicate that some commute trips are made by part-time working students in these areas. In terms of median income, the models show that the estimates are less negative in lower-income areas (except the new stations located in the north), indicating that people living in these areas are more dependent on bike-sharing than people living in higher-income areas neighbourhoods during the pandemic. Interestingly, the number of immigrants in the station DA is found to have a differential effect in different parts of the city. While the effect is negative in the city core, it is positive in the downtown and the eastern part of the city. This suggests that neighbourhoods in these regions with more immigrants have experienced an increase in bike-share ridership during the pandemic, highlighting the increased popularity of bike-share among the marginalized communities during the pandemic.

In terms of land use (Fig. 8b), the residential neighbourhoods towards the periphery of the city are negatively associate with bike-share demand, whereas those in the city core are positively associated with demand. Stations located near parks and other recreational areas (especially near the Toronto waterfront) are positively associated with trip generation, which further indicates the use of bike-sharing for exercise/recreational trips during the pandemic. The estimates for the waterbody land use attribute are significant for almost 98% of the stations, suggesting the importance of this variable in increasing demand across the bike-share network. The number of food services EPOI in the station buffer also plays a very important role in increasing the trip generation rates across the entire network (all the local estimates for this variable are found to be positive and significant). As in the multilevel models, the presence of health care EPOI within the station buffer has a negative effect on trip generation during the pandemic, especially in the downtown area.

Station distance from the CBD is found to be negatively associated with bike-share trip generation in both models. The effect is stronger for stations located further from the CBD. Interestingly, the impact of this variable is less negative for the stations located in the eastern part of the city. This is intuitive, given that these stations are somewhat disconnected from the rest of the system, and with the 30 min return policy of Bike Share Toronto, it is more likely that these stations cater to the local demand only. Stations located near major intersections are associated with the greater bike-share trip generation, especially near the waterfront. This might suggest that bike-share facilitates the first- and last-mile connections of multimodal trips, or simply that most of the activity hubs accessed by bike-share in these areas are located near the intersections. The presence of bus and streetcar routes near a station is negatively associated with bike-share demand during the pandemic, likely because bus and streetcar lines tend to be located on major roadways, many of which do not have dedicated cycling infrastructure. This could dissuade less-experienced cyclists from using nearby stations because they feel less safe to share the right-of-way, or they consider transit stops as barriers on their path.

Among the bicycle infrastructure variables (Fig. 8c), the positive effect of the length of bicycle facilities is the most prominent towards the north and east parts of the city. The downtown stations show more positive effects on weekends than weekdays, which might be attributed to increased recreational trips being made by casual users of the system during the weekends. By simultaneously considering the spatial distributions of the positive effect of individual station capacity and the negative effect of total capacity of stations within 200 m buffer, it becomes evident that there is a trade-off between station distance and their capacities. Adding larger stations at reasonable distances (so as to increase the total service area of the system while minimizing the overlap of the catchment area of individual stations) is likely to increase Bike Share Toronto usage more than adding moderate sized stations in close proximity. New stations that started operation during the pandemic experienced lower trip generation demand, especially towards the north and east parts of the city. They also seemed to have a negative effect on nearby stations, likely due to the redistribution of existing demand. This indirectly suggests that new stations need time to add to the overall demand of the system. Their trip generation rates increase gradually with time, and the initial buffer period of low demand should be taken into consideration in the long-term expansion plan of the network.

Policy recommendations

Based on the findings of the study, it is evident that the bike-share system has been affected by the COVID-19 pandemic, and the effects may vary by location. As such, it is important to design targeted policies to promote this healthy and environmentally friendly travel mode in an equitable and efficient way, both during the pandemic and once the pandemic is over. To better inform policy recommendations, we compare pre-pandemic and pandemic elasticities of five important variables from the GWR models. For this, we first estimated two additional GWR models (weekday and weekend) for the pre-pandemic period (Jan 2019–Feb 2020). Then, we calculated the elasticity values of (1) length of bicycle facility in buffer, (2) station capacity, (3) capacity of all bike share stations in buffer, (4) station distance to CBD, and (5) length of bus and streetcar routes in buffer for each of the four models, as shown in Table 6. From the elasticity values, it becomes evident that although the rank order of effects among the variables did not change during the pandemic, their effects became stronger compared to the pre-pandemic period. For example, a 1% increase in bicycle facility length is associated with greater increase in trips during the pandemic than the pre-pandemic era. This holds true even after Bike Share Toronto went through the massive system expansion of 160 new stations in 2020 compared to 2019.

In terms of expanding the network, the study findings indicate that adding larger stations at reasonable distances (so as to increase the total service area of the system while minimizing the overlap of the catchment area of individual stations) is likely to increase Bike Share Toronto usage more than adding smaller stations in close proximity. This might also help reduce the chance of a person trying to return a bicycle at a station only to see that it is already full. It is also found that newly installed stations need time to add to the overall demand of the system. Their trip generation rates increase gradually with time and hence the initial buffer period of low demand should be taken into consideration in the long-term expansion plan of the network. Bike-sharing is found to provide a sustainable alternative for shorter trips, especially as users are concerned about the risk of infection from using public transit. In this regard, future extensions of the system can mirror the transit network to provide a more flexible and socially distanced means of satisfying the mobility needs of its users. However, many people might be concerned about getting on a shared bike not knowing who used it last. As such, the service operators need to ensure that bikes are cleaned/sanitised regularly—if feasible, provide users the means to do so at the bike stations. Finally, the modelling results show that it is the weekend use that offers the biggest potential for bike-share system operators, particularly encouraging increased physical activity among people who may have been sedentary during the week spent working from home.

Conclusion

This study investigated the effects of weather, socio-demographic attributes, and land use and built environment on bike-share ridership in the context of the pandemic. Using 2019 and 2020 ridership data from Bike Share Toronto, the study adopted a two-stage methodology where multilevel modelling was first used to analyze how various factors affect monthly bike-share trip generation during the pandemic compared to the pre-pandemic period. Then, geographically weighted regression (GWR) analysis was performed at the station level to better understand how the relationships vary by communities and regions. Based on the findings, several time-sensitive and neighbourhood-specific policy recommendations are made informing decisions regarding the locations of new stations and the expansion and management of existing stations.

The multilevel modelling results indicate that the proportion of commuting bike-share trips has declined significantly during the pandemic, potentially due to the widespread practice of work-from-home policies. After controlling for other exogenous factors, higher-income neighbourhoods were found to generate much fewer weekday trips during the pandemic, suggesting that the service’s income disparity may have become more polarized during this period. Given the closure of indoor recreational facilities and gymnasiums, bicycle-sharing served as an essential means of satisfying the need for recreational/exercise trips during the pandemic, as indicated by the positive estimates of parks, open areas, and waterbody land uses in the trip generation models. The system also experienced an increased demand for maintenance trips, including visits to restaurants to pick-up food. The findings of the models also suggest the possibility that bike-sharing may be substituting for some weekday public transit trips during the pandemic. This might be partly attributed to the fact that bicycle-sharing, being a natural physically distanced travel and movement option, is associated with a lower risk perception compared to riding transit during the pandemic (Pase et al. 2020; Teixeira and Lopes 2020; Wang and Noland 2021). With more work-from-home and less road traffic, bicycling face less barriers during the pandemic. However, proper policies are needed to ensure that the condition does not become infeasible for bike-sharing as we come out of the pandemic and commuting by car resumes. Moreover, the trip generation rates were found to be more sensitive to the availability of bicycle facilities within station buffers during the pandemic, which might be attributed to the increased number of casual (recreational) users of the service. Intuitively, warmer weather conditions were associated with increased bike-share ridership irrespective of the pandemic situation. Finally, the models indicated the presence of a one-month lag effect in the bicycle sharing demand throughout the study period.

The weekday and weekend GWR model results reveal the presence of significant spatial heterogeneity in terms of the level of influence of the explanatory factors on the bike-share demand during the pandemic. In particular, the estimates for median income were less negative in lower-income areas, indicating that people living in these areas are more dependent on bike sharing than people living in higher-income neighbourhoods during the pandemic. The models also suggest that neighbourhoods with more immigrants experienced an increase in bike-share ridership during the pandemic, which highlights the increased popularity of bike-share among the marginalized communities.

One of the key limitations of the study is the inability to disentangle the effect of the pandemic from that of induced demand due to system capacity expansion. Whether the rise in use during the pandemic means the system is getting closer to being used at its capacity or if the capacity is exhausted remains an open question, which should be investigated in future works as more ridership data becomes available. Moreover, the weekday and weekend trip generations are modelled separately. However, there might be common station-level unobserved factors that simultaneously influence the demand across the two types of days. Thus, future research should control for these station-level unobserved factors by modelling the weekday and weekend trip generations simultaneously via bivariate modelling techniques. Again, it is important to note that the analysis using the GWR models is mostly exploratory in nature, and the findings provide evidence of associations only. Future research should adopt advanced spatial panel modelling techniques to help improve the evidence for causality. Moreover, neighborhoods with low model fit can be investigated further, particularly to ascertain if additional variables that were not available in the current study (for example, number of students at a university or college, number of transit stops within the bike-share station buffer, and frequency of transit service) can explain the variability of bike share trip generation in those locations. Finally, the methodology proposed in this study can be extended to account for the difference between peak and off-peak trips and the difference between annual member and casual user trips. Further segmentation of trip types may improve our understanding of how these factors affect bike-share demand in the context of the pandemic.

Like many other North American cities, there has been a rush on bicycle shops in Toronto during the pandemic, suggesting that cycling is considered to be an attractive mode by many people (Goldbaum 2020; Lucs 2021; Mazerolle 2021). Amid the shortage of bicycles available for purchase, bicycle-sharing systems help satisfy the increased demand for cycling during the pandemic. However, based solely on the findings of this study, we cannot predict how sustainable this change may be once the pandemic is over. We need more detailed information like what proportion of trips were made by repeat users, what proportion of trips were once off, if there was an increase in the purchase of annual subscriptions during the pandemic etc. As more data becomes available, further research can complement our analysis to ascertain whether the rise in bike-sharing demand during the pandemic will be sustained in the future.