Substitution effect of bike-sharing and metro ridership in
Chicago: a natural experiment
Master’s Thesis
Abstract:
Bike-sharing has become a more prevalent form of transport in urban areas in the last decade. It is well established that bike-sharing comes with a lot of benefits to users and society as a whole. Therefore, it would be valuable for policymakers to know how travelers decide their mode of transport. This study aims to quantify the effect of metro ridership on bike-sharing ridership. Specifically, it investigates whether bike-sharing stations near a temporarily closed metro station experience more traffic than bike-sharing stations near opened metro stations. To test the hypothesis that bike-sharing stations close to temporarily out-of-service metro stations experience more traffic than other bike-sharing stations, regression models were set up with data on Chicago between 2014 and 2017. Through a natural experiment setting, the station’s daily bike-sharing rides were compared in periods of opened and closed nearby metro stations. Four models were created and analyzed. The results showed a positive effect between a temporarily closed metro station and bike-sharing ridership at a nearby station. These results suggest that there exists a substitution effect. However, the sample size of periods of closed metro stations was small, so this result cannot be seen as solid evidence of this effect. Future research should aim to increase the sample size of metro closings to increase the credibility of the outcome.
Jonathan Schermer
EBM877A20
Supervisors: dr. G.J. Romensen & A.L. Schippers MSc
University of Groningen
1
1. Introduction
The global focus on climate change of the last decade has inspired governments and corporations to think about sustainable transportation alternatives in and around urban areas. One of those alternatives is bike-sharing. Bike-sharing is a sustainable transportation system where people looking for short-distance travel can unlock a bicycle from a public bicycle docking station in the city, and return it at another bicycle station near their destination.
The impact of bike-sharing was noticeable after its introduction immediately; the cycling population increased, transit use increased, greenhouse gases decreased and public health improved. (DeMaio, 2009). It didn’t come at the cost of public transport much, as now some public transport areas became better accessible due to bike-sharing hubs. The environmental effects of bike-sharing are arguably the most profitable benefit for society. In a 2018 study on the effect of bike-sharing in Shanghai, one of the most populated urban areas in the world, Zhang and Mi found that bike-sharing saved 8358 tons of petrol and decreased CO2 and NOx emissions by 25240 and 64 tons respectively. (Zhang & Mi, 2018). In a study by Shaheen a positive effect on CO2 reductions in bike-sharing programs in the United States is found (Shaheen et al., 2013). Besides this direct effect on emissions, it can also trigger positive externalities in the form of long term effects like promoting cycling as a feasible alternative to other transport and getting more people aware of its benefits.
2
There are also positive health effects associated with the use of bike-sharing. Woodcock (2014) compares health benefits to the potential risk of injury of bike hire in London, and find that the benefits outweigh the harms in every age group. The older people get, the higher these benefits become, meaning that for seniors these benefits are stronger and more prevalent. Rojas and Rueda (2011) found these same health benefits, as well as a lower mortality rate caused by air pollution in specifically Barcelona when comparing car users with bike-sharing users. Bicycle users had a very small risk increase in traffic incidents1, but most of that increased risk is explained by inexperienced users cycling for the first time, implying that more experienced users do not encounter this risk increase.
The last prevalent types of positive externalities found in the use of bike-sharing are economic externalities. As previously mentioned, Yu et al. (2018) found benefits for working people who use bike-sharing to commute. It decreases travel times and increases worker productivity. Besides these positive effects, Schoner et al. (2012) find a positive relation between bike-sharing use and economic activity. For example, users will use their bikes during lunch breaks to create economic activity at food businesses nearby or create economic activity in stores nearby. Having the added transportation potential of bike-sharing can grow a local economy and benefit small business owners.
It is clear that having bike-sharing as a system of transportation comes with a lot of benefits to society. However, to actually harvest those benefits, availability is not the key; bike-sharing has to be used in order to produce benefits. Therefore, it is interesting to analyze to which extent bike-sharing is used as a substitute for traditional transport methods by users. If bike-sharing is viewed as a useful and convenient substitute to for example the underground metro, bus, car, or scooter, it has the potential to replace those modes of transport. This effect will strengthen the benefits to society even more, as bike-sharing can replace a transportation method which in itself creates negative externalities, like environmental and noise pollution.
Various literature on this topic exists. Bachand-Marleau et al. (2012) and Campbell & Brakewood (2017) find a negative relation between bus ridership and bike-sharing ridership. Fishman et al. (2015), Bullock et al. (2016), and Zhou & Wang (2019) find this effect for car ridership and bike-sharing ridership. For metro ridership, Goh et al. (2020) find a negative relation with bike-sharing ridership. However, Ma et al. (2018) find the exact opposite effect.
3
After examining the literature on traditional transport, there seems to not be a consensus on what the effect is that bike-sharing brings.
In this thesis, the potential substitution effect between bike-sharing and metro ridership will be examined. This will be done in a natural experiment setting, with datasets on metro ridership and bike-sharing ridership in the period 2014-2017 in Chicago. Through modeling and performing OLS regression, this thesis will try to find and describe the effect of a closed metro station on bike-sharing activity, and through that conclude if a substitution effect exists. In the model, a significant positive effect is found between the closing of a metro station and the local bike-sharing ridership. This result will be analyzed in detail.
This thesis proceeds as follows: In section 2, this literature will be discussed in more detail. The setting, design, and data collection will be described in section 3. Section 4 will contain the model design and methodology. In section 5, the results will be outlined. The discussion of this result can be found in section 6. In section 7, the areas of improvement will be discussed. Finally, in section 8 the conclusion will be drawn.
2. Literature review
This section will give a detailed overview of the existing literature in this field. In section 2.1, an introduction about the 1iterature on traditional modes of transport will be given. In section 2.2, the literature on bus ridership and bike-sharing will be analyzed. In section 2.3, the literature on private and public car use will be analyzed. After that, section 2.4 will review the relation between metro ridership and bike ridership. These sections will paint a picture of how bike-sharing and other forms of transport in urban areas might be related. In section 2.5 the purpose of bike-sharing will be analyzed, to give an idea of in which scenarios people tend to choose bike-sharing over other forms of transport, and why. In section 2.6, the objective of this thesis and the hypothesis will be formulated.
2.1 Traditional transport in urban areas
In urban areas, the most popular modes of transport are by car, taxi, or bus (road transport), and metro or train (rail transport)2 . According to the EPA3 the transportation sector
accounts for 30% of total emissions, having increased by 47% since 1990 in the United States. As established before by Schoner et al., Yu et al., Zhang and Mi, and DiMaio, bike-sharing is
2 Taken from “Figures in Transport” statistics (2019)
4
much cleaner than these industries, often faster on short distances, and hardly an increase in lethal traffic accidents. There has been extensive research as to which types of these traditional modes of transportation have a substitution effect with bike-sharing in one form or another.
Bachand-Marleau et al. (2012) conducted a large survey in bike-sharing. Their respondents were adults in Montreal, which has a large bike-sharing system called BIXI. This survey with 1,787 respondents was conducted to gather information on the factors influencing the likelihood of using shared bicycles. The survey found a lot of factors influencing this choice, for example gender, age, distance from downtown, number of bicycle theft, and experience with cycling (Bachand-Marleau et al., 2012). Additionally, they found significant results in factors correlating with what was defined as traditional transport. For instance, being a regular bus user decreased the probability of using BIXI bikes, and the number of bus stops in a 400m radius decreased the frequency of using BIXI bikes. This could be interpreted as proof that bike-sharing competes with public transit.
2.2 Bus ridership
This effect on specifically bus ridership was also found by Campbell and Brakewood (2017). Through a differences-in-differences analysis in New York City where bus routes with and without bike-sharing infrastructure get compared, they find a 2.42% fall in daily bus trips for routes with bike-sharing available. Even when controlling for the expansion of bike lanes the effect is still significant and negative (-1.69%). This finding again supports the hypothesis of a substitution effect between public transit and bike-sharing. Yet, Wang and Akar (2019) find that perhaps the substitution effect between bus ridership and bike-sharing is positive rather than negative, making it a complementary effect rather than a substitutionary. In their research, they find that the number of bus stops is positively associated with bike-sharing trips and that bikes get returned more often at docks located close to a bus stop. This could indicate that a synergy exists between bus ridership and bike-sharing; a combination of both could be the preferred way to travel instead of choosing one or the other.
2.3 Car ridership
5
of transportation, comparing bike-sharing with taxi trips. Through machine learning this model found a significant negative correlation between taxi trips and bike-sharing; they substitute each other. They are more similar to each other than bus ridership in terms of door-to-door service; often the bus requires more additional travel (mostly walking) than a taxi or bike-sharing. Thus, taxi and bike-sharing don’t have that potential synergy that bus ridership does.
Secondly, the research on private car use and bike-sharing will be examined. Owning a car might decrease your bike-share use, as you have another mode of transportation available that can take you from door-to-door. Nonetheless, a car can be inconvenient for short distances, requires a parking space, and can be costly. Therefore, it is not impossible to imagine a substitution effect between bike-sharing and private car use. Fishman, Washington, and Haworth (2014) examined this potential effect. For most cities studied, motor vehicle use decreased due to bike-sharing.4 Only in London, this effect was reversed; bike-sharing increased motor vehicle use. This finding is explained by a low car mode substitution rate and substantial truck use for rebalancing bicycles (Fishman et al., 2014). Barbour et al. (2019) researched this topic as well and found two significant effects with regards to car use. Owning a car makes respondents less likely to use bike-sharing. This makes sense, as the purpose of owning a car is transportation, and even an unused car comes with costs. The second finding is that a low average parking time during their most regular trip decreased the likelihood of using bike-sharing (Barbour et al., 2019). This reveals that the earlier mentioned downsides of using a car affect the substitutability of bike-sharing. In conclusion, it is in some cases seen as a relevant alternative to a private car.
2.4 Metro ridership
Another important traditional transportation method is the metro5. Many large urban areas have a metro network, often underground as to not congest traffic above ground, but still provide a quick, affordable and reliable transportation method to the city center, airports, and other important urban areas. Similar to bus ridership, metro ridership is often not door-to-door transportation and requires a small distance to a metro station to be covered in a trip. This opens the opportunity for a synergy as well as a substitution effect. Goh, Yan, and Jaillet (2020) find that in Boston there is a connection between bike-share usage and metro usage in certain areas; metro stations that are close to bike-share stations get used less than metro stations far away
6
from bike-share stations. This finding supports the idea of a substitution effect between these forms of transport.
In contrast, Ma et al.. (2018) find evidence for the opposite: they find that in Nanjing 89% of passengers integrate bike-share with using the metro. This effect is stronger right before and after business hours, during the most common commute times (07:00-09:00 and 17:00-19:00), also known as the rush hours. On weekend days, this effect is spread evenly throughout the day.
2.5 Bike-sharing trip purpose
It seems that the purpose of the trip plays a large role in how substitutable bike rides are for different types of transport. There is a lot of evidence in research for purposes of bike-sharing transportation. Mostly for commute purposes (Zhang et al., 2017; Fishman et al., 2015; Bachand-Marleau et al., 2012; Xing et al., 2020) by observing a significant increase in rush hour times. However, also recreational use is observed (mostly weekend and late evening trips) by Zhang et al., Fishman et al., and Xing et al.. One characteristic about bike-sharing trips that often comes up in research is the last mile problem that it seems to solve. This problem is described by Shaheen et al. (2010) as “the problem of covering the short distance between home and public transit or transit stations and the workplace, which may be too far to walk”. This last mile is often the mile that takes the most time, as there is no easy way of covering it if one does not live next to a transit station. With a large network of bike-sharing, this problem can be solved. Xing et al. (2020) also mention this problem as a reason for their prediction of trip purposes in bike-sharing. They observe that bike-sharing is regularly used to transfer between transit and other transit, or between transit and home. Martin and Shaheen (2014) support these findings by stating that bike-sharing “may serve prominently as a first-mile, last-mile facilitator in areas with less intensive transit networks”. DiMaio (2009) found the same pattern by observing an increase in transit trips on routes in areas with bike-sharing. In conclusion, for short trips, bike-sharing is seen as convenient and can be a direct competitor for traditional transit, but for longer trips, it more often serves as a way of connecting to transit rather than substituting it completely.
2.6 Objective and hypothesis
7
in Illinois in the United States. The objective of this thesis is to identify and quantify a potential substitution effect between metro trips and bike-sharing trips in the city of Chicago between 2014 and 2017. Through previous literature, it is reasonable to expect a substitution effect for bike-sharing and metro rides in a dense, urban area like Chicago.
3. Research Design
In this section, the research design will be laid out. In section 3.1, background information on bike-sharing and the setting in Chicago will be set out. In section 3.2, the natural experiment setting will be explained. In section 3.3, the data selection will be discussed. Finally, in section 3.4, the data will be described and summarized.
3.1 Background information
3.1.1 Field Setting
Bike-sharing in urban areas is an idea originating from the Netherlands, specifically Amsterdam, in 1965. (DeMaio, 2009). Original growth of this market was slow, as the technology for tracking and developing bikes was not on par with what an efficient market would demand. Over time, this 1st generation bike-sharing program was improved by the 2nd generation in Denmark and 3rd generation in England, equipped with technological improvements like electronic locks, smartcards, and on-board computers (DeMaio, 2009). In 2005, the first major city Lyon (and later Paris) got on board with the 3rd generation bike-sharing program and in 2008 cities all over the globe, from South and North America to Asia, followed suit. As of December 2020, there are 444 cities globally with a bike-sharing program.6
In the North American market, bike-sharing has grown rapidly since 2007. Along with previously mentioned New York City, urban areas like Chicago, Washington DC, and Montreal have large bike-sharing programs. Most of these are non-for-profit, and only 21% were privately owned programs. In terms of market share, not-for-profit programs accounted for 82% of total membership (Shaheen et al., 2013). These programs have not yet shown to be a profitable business model for private owners, despite gaining popularity. However, since the availability of bike-sharing has significantly positive externalities like Yu et al.. has found, most large cities and local governments don’t mind investing money in these projects.
8
New York City has the biggest bike-sharing program7 in North America as of 2020, with Chicago8 as a close second. In this thesis, the bike-sharing program of Chicago will be examined. Chicago is one of the largest urban areas in the world, with the 2nd largest business district in the United States9, making is one of the major world financial centers (Saracoglu, 2018). The combined factors of large bike-sharing infrastructure and a large urban area make Chicago a representative location for this field of research.
3.1.2 Metro network in Chicago
Chicago has an extensive metro network connecting important infrastructural spots like O’Hare airport and Midway airport with the business district, the city center, the university, and suburbs. This network is called the Chicago L10, and it is operated by the CTA.11 The L has 8 lines, most of them intersecting with each other, and 145 stations spread over the city.12 The
lines are all marked with a color and name. The network is 169km long, with 18km underground. This study uses information on this network as a form of traditional transport.
Fig 1. Location of L metro stations as of December
2020.
7 Citibike. 8 Divvy bikes. 9 Called “the Loop”. 10 Short for elevated.
11 Chicago Transit Authority.
9 3.1.3 Bike-sharing in Chicago
As mentioned in the field setting, Chicago has the 2nd largest bike-sharing system in the United States, called Divvy. Divvy is operated by Lyft for the CDT.13 As of December 2020, Divvy operates at 675 stations spread over the city of Chicago and some suburbs. The stations average at about 20 docks per station and serve around 22,000 travelers each day.14 The system is one of the last generation bike-sharing systems and includes real-time information on available docks at each station, and available stations in a selected radius (Saracoglu, 2018).
The system was introduced to the city in 2013, as an attempt to copy the bike-sharing model from Paris. The system gradually upgraded from 75 stations around the city center to the 675 stations now, slowly covering more downtown areas. Figure 2 contains the locations of all bike-sharing stations as of December 2020.
Customers can choose a single trip pass for $3, allowing them to take a 30-minute one-way bike trip from one station to another. Additionally, they can choose a day pass ($15 allowing unlimited 3-hour trips for 24 hours) or an annual membership for $99 per year (allowing for unlimited 45-minute trips). The bikes are accessible 24 hours a day, 365 days a year15.
Fig 2. Location of Divvy stations in Chicago as of December 2020
13 Chicago Departement of Transportation.
10
3.2 Natural experiment setting
To find a potential link between metro ridership and bike ridership, a way of setting up an experiment is to find a natural experiment. This is an empirical study that is not set up by the researcher, yet can still be analyzed as a normal experiment as the individuals are still exposed to experimental conditions. In the context of metro ridership, a natural experiment setting can be created by comparing days where a metro station is open and closed. Metro stations usually operate 365 days per year, but sometimes they require maintenance, experience technical issues, or get closed off for an event.16 By analyzing those temporary closings, and comparing them with regular days, the bike-sharing ridership of a close-by docking station can be reviewed. If there are significantly more rides near a temporarily closed metro station compared to a metro station that is actively in use, that could imply a substitution effect. If there are significantly fewer rides near a closed metro station compared to an open metro station, this could imply a synergy, or complementary effect.
3.3 Data Collection
In this thesis, a dataset is created using information from various, public data sets provided by the City of Chicago. The data is retrieved from the open database called “Chicago Data Portal”17. To create this dataset, information on the metro L ridership has been downloaded
in October 2020. Information on Divvy rides between 2014 and 2017, including weather information, has been downloaded from the same source in October 2020. The individual divvy rides have been grouped in daily ride totals. The daily bike-sharing rides and metro ridership datasets have been combined and complemented with additional data from this source; the distance between metro stations and bike-sharing stations, and the type of bike lane surrounding the bike stations. Data on temporary bike stations or permanently out-of-service bike stations was deleted. The dataset uses pure panel data on 77 stations that at some point in the time period 2014-2017 had a closed metro station nearby.
The metro ridership dataset was modified as follows: The dataset was filtered to only 2014-2017 ridership. Then, the rides were set to 0, to see which days the metro station was closed. In the period between 2014 and 2017, nine unique metro stations had periods in which
16 In this setting, we have to be aware of the anticipation effects that travelers might have. People might be (and are likely to be) informed beforehand about the closing of the metro station, and might pick out a different alternative beforehand. This problem makes this setting not a complete natural experiment. As there is little information on the awareness of travelers, we will note this effect as a potential flaw.
11
they were closed. For those nine metro stations, all bike-sharing stations within a 0.6 mile (1 kilometer) radius were identified. This radius is based on research by the US department of transportation, on how far an average pedestrian is willing to walk to access transit.18 The overview of all stations, including distance and bike lane information, can be found in the appendix. In Table 1 and Figure 3 examples of metro station Madison / Wabash and Montrose-O’Hare can be found.
In the cases where one bike-sharing station is within the radius of two metro stations, only the closest station is used in the analysis. Information on the dates that these metro stations were closed can also be found in the appendix.
MADISON / WABASH ID Distance to metro (in miles) Bike lane
Michigan Ave & Madison St 197 0 None Wabash Ave & Adams St 39 0,2 Protected
Millennium Park 90 0,3 None
Dearborn St & Monroe St 49 0,3 Protected Michigan Ave & Jackson Blvd 284 0,5 None Lake Shore Dr & Monroe St 76 0,6 None Michigan Ave & Washington St 43 0,3 Protected State St & Randolph St 44 0,3 Protected
Table 1. Information on metro station Madison / Wabash. This table contains the names of the bike-sharing stations within the radius, the ID
number of the stations, the distance to the metro station, and the nearby bicycle infrastructure.
Fig 3. The location of the 5 bike-sharing stations within 1-kilometer radius of metro station Montrose – O’Hare.
12
3.4 Data description
In the following section, the data used in the models will be described. Both models use time variables, meteorological variables, and station fixed effect variables. Section 3.4.1 describes the dependent variable and the structure of the dataset. Section 3.4.2 describes the meteorological variables and their practicality. Section 3.4.3 will describe the station fixed effect variables used in the first model. Section 3.4.4 will describe the remaining variables used in the regression models.
3.4.1 Dataset structure
Year 2014 2015 2016 2017 Total
Observations 20,938 25,550 27,084 28,105 101,677
Table 2. Observations per year in the dataset.
The dataset consists of 101,677 observations spread over 1,461 days and 77 unique stations. The dependent variable in the observations used in both models is Rides. The variable
Rides describes the amount of bike-sharing rides starting from the corresponding station
(StationID) on the corresponding day (Daynumber, Year). The day number describes the day of the year, 1 for the first of January all the way up to 365 or 366 for the 31st of December. For non-business days (Saturday, Sunday, and public holidays) the dummy Weekend is equal to 1 On any other day, it is equal to 0. The public holidays used are described in Table 8 in the appendix.
Year 2014 2015 2016 2017 Total
Rides 423,060 524,068 625,329 672,692 2,245,149
Avg 20.21 20.51 23.09 23.94 22.08
Table 3. Bike-sharing rides per year, total and average.
The dataset with 101,677 observations contains 2,245,149 rides in total, combined over 77 stations. The daily averages of the number of rides and the total number of rides each year can be found in table 3. For these years, dummies were generated, to account for the yearly growth that bike-sharing in Chicago underwent. The dummy firstyear is equal to 1 if the year of the ride corresponds with 2014, secondyear with 2015, thirdyear with 2016, and
fourthyear with 2017. Additionally, the models contain dummies for the week the rides were
13
Fig 4. Seasonal trends of weekly bike-sharing rides for years 2014, 2015, 2016, 2017, and the aggregated weekly rides of 2014-2017.
14
In figure 4, the seasonal trend of rides over 2014-2017 can be observed. The trend shows a clear increase in activity in the spring and summer, peaking at week 31 (End of July, beginning of August). In the winter activity decreases, with a low point in week 52 and week 53, the end of December. This trend shows that it is useful to control for weekly trends.
3.4.2 Meteorological variables
The dataset also includes five dummy variables describing the weather condition on the corresponding date. The weather conditions have been categorized into five states. These categories have been taken from weather data generated by a weather station called weatherunderground.19 This is a commercial weather station providing real-time weather information and historical data on weather conditions.
Tstorms Clear NotClear RainorSnow Cloudy Total
1,927 5,207 827 11,532 82,184 101,677 1,90% 5,12% 0,81% 11,34% 80,83% 100%
Table 4. Weather conditions in observations and in % of total observations (N = 101,667).
When the sky is clear, the dummy Clear is equal to 1. In all other cases, it is equal to 0. If the sky is cloudy, which is the most common weather condition, the dummy Cloudy is equal to 1. If the weather is very cloudy without rain or snow, the dummy NotClear is equal to 1. If it is raining or snowing, but not in a storm, the dummy RainorSnow is equal to 1. If there is a local thunderstorm, the dummy Tstorms is equal to 1.
Existing literature tells us to control for weather conditions, as it has a significant impact on bike ridership. Li et al. (2015) are one of the first researchers to mention the importance of meteorology in sharing. They mention that more people check outside options like bike-sharing on sunny days compared to cloudy days. Wang et al. (2013) and Rixey (2013) also find a strong correlation between favorable weather conditions and bike-sharing ridership. Gebhart and Noland (2014) conducted a research paper dedicated to the impact of weather conditions on bike-share trips in Washington DC. They find significant effects for multiple weather conditions, both in terms of temperature, humidity, rainfall, and snow.
15 3.4.3 Station fixed effects variables
To get an unbiased view on the effect of the closing of a metro station on bike-sharing ridership, the models need to control for station fixed effects. Station fixed effects are characteristics of stations that are time-invariant in the period between 2014 and 2017. The characteristics might have a significant effect on the prediction of rides on a certain day. The first important fixed effect that will be discussed is the bike infrastructure. The second fixed effect is the locational effect; we expect the bike stations close to the city center and business district to have more rides. The third fixed effect is the distance to the closest metro station, as more walking might discourage people from using the bike station.
The bike infrastructure around a station could influence the amount of bike ridership. With bike lanes nearby, people might feel safer to use a bike, even without bicycle experience. Chicago has four types of bike lanes in the city: regular bike lanes, shared bike lanes, protected bike lanes, and buffered bike lanes.20 In some parts of the city, there are no bike lanes at all and
bikers have to use the road to ride on. The dummies bikelane, sharedlane, protectedbikelane, and bufferedbikelane are equal to 1 if the corresponding bike-sharing station has a regular bike lane, shared bike lane, protected bike lane, or buffered bike lane nearly, respectively.
Fig 5. Type of bike lane in % of total observations (N = 101,677). Fig 6. Type of bike lane in amount of bike stations (N = 77).
20 Information retrieved from the Chicago Data Portal on 03/11/2020
35%
16% 13% 28%
8%
16
Fig 7. Bike lanes in Chicago as of February 2020.
The barrier protected bike lane (protectedbikelane) is a lane specifically and only for cyclists, protected by a barrier to separate cyclists and vehicles on the road. The buffered bike lane (bufferedbikelane) is a bike lane, buffered by an extra space to keep cyclists and vehicles more apart. A shared bike lane (sharedlane) is a lane that is shared between cars and cyclists and only contains a cyclists sign. A regular bike lane (bikelane) is a separate bike lane on the side of the road, separated by one large line. Figures with an example of each can be found in the appendix (figure 8-11).
The second fixed effects are locational fixed effects. As we expect more traffic in the city center (which is also the business district), the dummy citycentre was created. It is equal to 1 for every ride starting at a station within the Loop district in Chicago, and equal to 0 for every other station.
We also expect people to use bike-sharing for convenience and time-saving. Therefore, bike stations near metro stations might be used more than bike stations further away from metro stations. The dummy longwalk is equal to 1 if the bike station is not within 0.5 miles of a metro station, and equal to 0 if otherwise.
Longwalk = 1 Longwalk = 0 Citycentre = 1 Citycentre = 0
33,524 (32,97%) 68,153 (67,03%) 28,057 (27,59%) 73,620 (72,41%)
17 3.4.4 Other variables
Three more variables have been added to the models. Firstly, dummy variables for five weeks after re-opening of the closed metro station have been created. These variables test the effect on bike-sharing ridership after re-opening. The goal is to see if a positive or negative effect sticks even after re-opening, or slowly fades away over time. The variables are called
postclosingweek and range from 1 to 5, 1 is the first week after reopening, 5 is the fifth week
after reopening. Similarly, five dummies have been created to represent the five weeks before closing, called preclosingweek and ranging from 1 to 5. Preclosingweek5 represents the week that is 5 weeks prior to closing.
Secondly, the models control for the capacity of docks a bike-sharing station has. The more docks, the more potential for rides as there are more bikes available. The capacity often changed in the time period between 2014 and 2017, as the bike-sharing project keeps getting upscaled as demand rises. Therefore, the variable capacity is not a time-invariant variable.
The last variable added is the most significant one in the model. The variable
MetroClosed is a dummy, equal to 1 if the nearest metro station is closed and 0 otherwise. This
variable will test the hypothesis stated in section 2.6. Information on metro closings in the period 2014-2017 can be found in Table 9 of the appendix.
MetroClosed = 0 MetroClosed = 1
96,968 (95,38%) 4,709 (4,62%)
Table 6. Observations of MetroClosed variables (N=101,677).
4. Methodology
18
4.1 Model 1 & 2 research design
In the first regression, we attempt to predict the number of rides in a bike-sharing station. In the model, we use MetroClosed as a variable to indicate if the closest metro within the 1-kilometer radius is closed. The model also controls for weekend (as we expect fewer rides to happen on weekend days), the weather condition, the year, the week, and the capacity of the bike-sharing station. The results of this regression can be found in Table 7 and in the appendix. Potential fixed effects are still an issue in model 1, as the number of rides can also be predicted by certain characteristics of the station. The model takes the following form:
𝑅𝑖𝑑𝑒𝑠𝑗𝑡= 𝛼 + 𝛽1∗ 𝑀𝑒𝑡𝑟𝑜𝐶𝑙𝑜𝑠𝑒𝑑𝑗𝑡+ 𝛽2∗ 𝑊𝑒𝑒𝑘𝑒𝑛𝑑𝑡+ 𝛽3∗ 𝐶𝑎𝑝𝑎𝑐𝑖𝑡𝑦𝑗𝑡+ 𝜃 ∗ 𝑊𝑒𝑎𝑡ℎ𝑒𝑟′𝑗𝑡+ 𝛾 ∗
𝑌𝑒𝑎𝑟𝑠𝑡′+ 𝛿 ∗ 𝑊𝑒𝑒𝑘𝑠𝑡′+ 𝜀𝑗𝑡 (1)
Rides predicts the amount of bike rides on day t starting at Station j. MetroClosed and Weekend are both dummy variables, as described in section 3. 𝜃 is a vector of the coefficients corresponding with the vector of all weather conditions (Weather) described in section 3.4.2. The weather conditions in the vector are Clear, NotClear, Tstorms, and RainorSnow. The dummy for Cloudy weather is omitted. 𝛾 is a vector of coefficients corresponding with the vector of all years in the dataset 𝑌𝑒𝑎𝑟𝑠′. 𝛿 is a vector of coefficients corresponding with the vector of all weeks in the dataset 𝑊𝑒𝑒𝑘𝑠′. The vectors do not include week 1 and year 1, as they would be omitted otherwise. 𝛼 is the constant in this model, 𝜀𝑗𝑡 is the error term.
Model 2 is an attempt to fix issues of model 1 by adding fixed effect dummies. In this model, we attempt to predict rides using the control variables from model 1 and adding some more fixed effects. In the model, we add the location dummy citycentre, which is equal to 1 for all stations located within the Loop.
Secondly, the model controls for the distance between the bike station and the nearest metro station using the dummy variable longwalk is equal to 1 if the bike station is located more than ½ a mile away from a metro station. Lastly, the model controls for what type of bike lane is near the station, as a better cycling infrastructure could increase biking activity.
19
𝑅𝑖𝑑𝑒𝑠𝑗𝑡 = 𝛼 + 𝛽1∗ 𝑀𝑒𝑡𝑟𝑜𝐶𝑙𝑜𝑠𝑒𝑑𝑗𝑡+ 𝛽2∗ 𝑊𝑒𝑒𝑘𝑒𝑛𝑑𝑡+ 𝛽3∗ 𝐶𝑎𝑝𝑎𝑐𝑖𝑡𝑦𝑗𝑡+𝜃 ∗
𝑊𝑒𝑎𝑡ℎ𝑒𝑟′𝑗𝑡+ 𝛾 ∗ 𝑌𝑒𝑎𝑟𝑠𝑡′+ 𝛿 ∗ 𝑊𝑒𝑒𝑘𝑠𝑡′+ 𝛽4∗ 𝑐𝑖𝑡𝑦𝑐𝑒𝑛𝑡𝑟𝑒𝑗+ 𝛽5∗ 𝑙𝑜𝑛𝑔𝑤𝑎𝑙𝑘𝑗+ 𝛽6 ∗
𝑏𝑖𝑘𝑒𝑙𝑎𝑛𝑒𝑗+ 𝛽7∗ 𝑠ℎ𝑎𝑟𝑒𝑑𝑙𝑎𝑛𝑒𝑗 + 𝛽8 ∗ 𝑝𝑟𝑜𝑡𝑒𝑐𝑡𝑒𝑑𝑏𝑖𝑘𝑒𝑙𝑎𝑛𝑒𝑗+ 𝛽9 ∗ 𝑏𝑢𝑓𝑓𝑒𝑟𝑒𝑑𝑏𝑖𝑘𝑒𝑙𝑎𝑛𝑒𝑗+ 𝜀𝑗𝑡 (2)
4.2 Model 3 & 4 research design
In the 3rd and 4th regression models, FE regression was used. The model still predicts Rides as the number of bike rides on day t starting at Station j. Along with the first regression, this model includes weather dummies, time dummies, and the weekend and capacity dummy. The FE set up of this model will account for the station fixed effects, therefore the fixed effects dummies from the last regression are omitted in this model. Model 3 takes the following form:
𝑅𝑖𝑑𝑒𝑠𝑗𝑡 = 𝛼 + 𝛽1∗ 𝑀𝑒𝑡𝑟𝑜𝐶𝑙𝑜𝑠𝑒𝑑𝑗𝑡+ 𝛽2∗ 𝑊𝑒𝑒𝑘𝑒𝑛𝑑𝑡+ 𝛽3∗ 𝐶𝑎𝑝𝑎𝑐𝑖𝑡𝑦𝑗𝑡+ 𝜃 ∗ 𝑊𝑒𝑎𝑡ℎ𝑒𝑟′𝑗𝑡+ 𝛾 ∗ 𝑌𝑒𝑎𝑟𝑠𝑡′+ 𝛿 ∗ 𝑊𝑒𝑒𝑘𝑠
𝑡′+ 𝜖𝑗𝑡 (3)
Similar to models 1 and 2, 𝜃 is a vector of the coefficients corresponding with the vector of all weather conditions. 𝛾 is a vector of coefficients corresponding with the vector of all years in the dataset 𝑌𝑒𝑎𝑟𝑠′. 𝛿 is a vector of coefficients corresponding with the vector of all weeks in the dataset 𝑊𝑒𝑒𝑘𝑠′. The vectors do not include week 1. year 1 and Cloudy, as they would be omitted otherwise. 𝛼 is the constant in this model and ∈𝑗𝑡 is the error term.
In model 4, we add the ten time-dummies described in section 3.4.4. 𝜑𝑝𝑟𝑒 represents the
vector of coefficients of the pre-closing time dummies. 𝜑𝑝𝑜𝑠𝑡 represents the vector of coefficients of the post-closing dummies. The pre- and post-closing dummies are represented in vector Preclosing and Postclosing, respectively. The model takes the following form: 𝑅𝑖𝑑𝑒𝑠𝑗𝑡 = 𝛼 + 𝛽1∗ 𝑀𝑒𝑡𝑟𝑜𝐶𝑙𝑜𝑠𝑒𝑑𝑗𝑡+ 𝛽2∗ 𝑊𝑒𝑒𝑘𝑒𝑛𝑑𝑡+ 𝛽3∗ 𝐶𝑎𝑝𝑎𝑐𝑖𝑡𝑦𝑗𝑡+ 𝜃 ∗ 𝑊𝑒𝑎𝑡ℎ𝑒𝑟′𝑗𝑡+ 𝛾 ∗ 𝑌𝑒𝑎𝑟𝑠𝑡′+ 𝛿 ∗ 𝑊𝑒𝑒𝑘𝑠𝑡′+ 𝜑𝑝𝑟𝑒∗ 𝑃𝑟𝑒𝑐𝑙𝑜𝑠𝑖𝑛𝑔𝑗𝑡+ 𝜑𝑝𝑜𝑠𝑡∗
20
5. Results
This section describes the results of the regression models described in section 4. In these models, the correlation between bike-share ridership and metro ridership is estimated. A natural experiment set-up is used to create a dataset, and this dataset is used to perform OLS regression. In section 5.1 the results of the first and second models will be discussed. In section 5.2 the results of the third and fourth models will be discussed. In section 5.3, the results of the two regression methods will be compared. The results can be found in Table 7.
5.1 Model 1 & 2 results
21
Table 7: Regression results. (1) (2) (3) (4)
VARIABLES Model 1 Model 2 Model 3 Model 4
MetroClosed -0.643** -2.470*** 1.348*** 1.458*** (0.282) (0.277) (0.248) (0.248) Weekend -12.36*** -12.36*** -12.37*** -12.36*** (0.128) (0.124) (0.104) (0.104) Capacity 1.368*** 1.017*** 0.294*** 0.285*** (0.00775) (0.00942) (0.0208) (0.0209) Clear 0.357 0.408 0.615*** 0.628*** (0.278) (0.271) (0.228) (0.228) NotClear -2.151*** -2.066*** -1.911*** -1.989*** (0.675) (0.656) (0.553) (0.553) RainorSnow -5.132*** -5.095*** -4.913*** -4.889*** (0.203) (0.197) (0.166) (0.166) Tstorms -3.716*** -3.704*** -3.709*** -3.775*** (0.445) (0.432) (0.364) (0.365) preclosingweek5 2.328*** (0.772) preclosingweek4 4.572*** (0.772) preclosingweek3 0.520 (0.772) preclosingweek2 3.211*** (0.779) preclosingweek1 1.478* (0.779) postclosingweek1 0.720 (0.647) postclosingweek2 0.998 (0.803) postclosingweek3 0.319 (0.899) postclosingweek4 0.280 (0.741) postclosingweek5 1.619*** (0.570) citycentre 8.808*** (0.161) longwalk 0.803*** (0.127) bikelane 4.760*** (0.174) sharedlane 7.425*** (0.190) protectedbikelane 7.443*** (0.152) bufferedbikelane 4.038*** (0.235) Constant -16.15*** -16.21*** 2.169*** 2.084*** (0.485) (0.482) (0.547) (0.549) Observations 101,677 101,677 101,677 101,677 R-squared 0.409 0.442 0.374 0.374
Meteorological variables Yes Yes Yes Yes
Number of StationID 77 77 77 77
Post-/Pre-closing variables No No No Yes
Fixed Effects Time variables No Yes No Yes Yes Yes Yes Yes Standard errors in parentheses
22
In Column 2 of Table 7 the result of regression model 2 is described in section 4.1. The effect of Weekend has not changed in this model. The effect of an increase in capacity is still similar to the first model, 1 extra dock results in 1.017 more rides per day. The effect of the weather condition on the number of rides has not changed much either. The weather condition clear still doesn’t have a significant result. The relatively bad weather conditions not clear, thunderstorms, and rain or snow again decrease the estimated number of rides by 2.066, 3.716, and 5.132, respectively.
The constant is similar in both models. Model 1 has a constant of -16.15, model 2 has a constant of -16.21. In this set-up, a negative constant is not a problematic outcome. The prediction of rides in these models will always be compensated by positive variables like capacity. This will balance out the negative constant and prevent negative predictions for bike-sharing rides.
The added fixed effects all have a significant effect on the dependent variable. Bike-sharing stations located in the Loop are expected to have 8.808 rides per day more than stations located outside of the Loop. Surprisingly, the effect of the distance between the metro station and the bike-sharing station is positive in this model; if the metro station is located ½-mile or more away from the bike-sharing station, the rides from that station are expected to increase by 0.803 per day.
The nearby bike lane infrastructure also has a significant effect on rides in this model. Regular bike lanes and buffered bike lanes increase daily ridership by 4.760 and 4.038 rides than stations without bike lanes, respectively. The largest effects are found in the shared lanes and protected bike lanes, 7.425 more daily rides and 7.443 daily rides than stations without bike lanes, respectively.
23
5.2 Model 3 & 4 regression results
The results of model 3 can be found in column 3 of Table 6. This model uses the same variables as model 1. Nevertheless, not all results are similar as this model uses a Fixed Effects regression method. Once more, a significant decrease in rides is found on Saturdays, Sundays, and public holidays. The effect is essentially the same as in models 1 and 2; 12.37 fewer rides on one day compared to a weekday. The weather dummies find similar significant results as well. However, in this model the weather condition clear does have a significant result; .615 more rides on a clear day compared to a cloudy day. The negative, significant effects of the conditions not clear, thunderstorms, and rain or snow are almost identical to model 1 and 2: decreases of 1.911, 3.709, and 4.913 daily rides compared to cloudy weather respectively.
The effect of an increase in docking capacity by 1 results in an increase of .294 rides per day in this model. The effect is significant for p<0.1 and is considerably lower than the effect found in model 1 and 2.
The constant in model 3 is 2.169, which is exponentially different from the -16.15 and -16.21 constants from the previous two models. The constant of model 3 contains a lot of station fixed effects that were potentially omitted in the first two models. The additional time-invariant effects also affected the MetroClosed variable, to now a positive and still significant effect of 1.348 more rides on a day when the nearest metro station is closed. This result is very different in absolute value compared to the previous models. However, the biggest change compared to those models is the sign change from negative to positive.
In model 4, the postclosing and preclosing variables were added. For the first four postclosing weeks, no significant difference was found compared to other weeks. In the fifth week after closing, we see a significant increase in daily rides by 1.619. For 5, 4, and 2 weeks before closing the model also finds a positive significant result, the rest of the weeks have no significant impact on bike-sharing rides. These effects mostly took away some of the effect of the time variables (specified in the appendix), as all other variables and the R2 has not changed compared to model 3.
24
The effect on a capacity increase hasn’t changed much either, 0.285 more daily rides for every extra dock. The observed effect on MetroClosed is still significant, at 1.458 more rides on a day where the metro is closed.
6. Discussion
Through modeling daily bike-sharing ridership with several relevant variables, we find evidence that metro ridership and bike-sharing ridership are connected in Chicago in the period 2014-2017. In section 6.1 we will discuss the outcomes of the weekend variable, capacity variable, and weather variables. In section 6.2 we discuss the outcomes of models 1 and 2 for MetroClosed and the station fixed effects bike infrastructure, location, and distance. In section 6.3 we discuss the outcomes of the Fixed Effect models with respect to MetroClosed and the post- and pre-closing variables.
6.1 Weekend, capacity, and weather variables
For all models discussed, the variables weekend, capacity, and weather all had similar results. As expected through previous literature and transport statistics, less bike-sharing activity was found on Saturdays, Sundays, and public holidays. This effect can be explained through business activity; on business days there is more demand for short-distance travel because of commuting. The amount of people who commute is significantly less during the weekend or public holidays.
The effect of increasing capacity is positive in every model. This result in itself is not surprising, as it is fair to assume that stations with more capacity can handle more demand and Divvy bases bike station capacity on data, therefore the stations with the most traffic will have their capacity increased the most. The difference between models 1 & 2 and 3 & 4 is that the capacity of stations is party a fixed effect; every now and then they increase (hence the part that is not fixed), though most of the time they stay the same. Therefore, part of the FE models (3 & 4) accounts for station fixed effects size automatically, taking that effect away from the variable capacity, hence the lower coefficients.
25
than cloudy result in a negative effect (not clear, thunderstorms, rain/snow) and better than cloudy (clear weather) result in a positive effect. So it is clear that the choice to cycle is highly dependent on weather and temperature. This is explained by the results in the models and the seasonal trends of bike-sharing rides from figure 4 in section 3.4.1.
6.2 Metro station closing and fixed effects in model 1 & 2
In models 1 and 2, the effect of a closed metro station is the opposite of what was expected in the hypothesis in section 2.6; it is a negative result. This result implies a complementary effect between metro ridership and bike-sharing ridership. The effect gets stronger with the extra fixed effect variables in model 2.
The fixed effects will be discussed first. Most fixed effects have the expected effect. Bike-sharing stations in the Loop are expected to experience more traffic, and an improved bike lane infrastructure should encourage people to use bike-sharing, as it becomes a safer and more convenient mode of transport if the infrastructure is built around it. The effect between the different bike lanes is not very large, it seems that the marginal benefit from improving a bike lane is diminishing, as the jump from no bike lane to any form of bike lane is big, but improving a regular bike lane to a protected bike lane will not have that big of an effect. It is safe to assume from this model, that there exists a correlation between bike infrastructure and bike-sharing, and it is a relevant variable to control for.
A surprising find is that the location of the metro station compared to a bike-sharing station has a positive effect. This result could imply that if a bike-sharing station is more isolated, it gets used more. It is also possible that if a bike-sharing station is far away from a metro station but still the closest, that this station is located in a place with not many transportation alternatives. In any case, it is reasonable to assume a correlation between the location of bike-sharing station compared to transportation alternatives.
26
6.3 Fixed Effects Models
By using FE regression, station fixed effects are controlled for more accurately. The results from this regression are similar for all variables discussed in section 6.1. When comparing our main results to models 1 and 2, we observe that the effect of a closed metro station is now positive; as is in line with the hypothesis stated in section 2.6. This result implies a substitution effect between metro ridership and bike-sharing ridership. The results of these models are drastically different from the first two models and this means the suspicion of omitted variable bias is now even stronger. This is because the FE models control for all station fixed effects, and in the first model there are no fixed effects, and the second model only controls for a few fixed effects. Since the results are so different, it is safe to assume that the first two models are missing too many station fixed effects to draw relevant conclusions on our hypothesis.
The sign change from negative to positive for the most significant variable in these models shows that the first two models are lacking critical variables which skew the results. This can be explained through the fact that a lot of significant fixed effects are missing in the model. For example, information on other transport alternatives like taxi ridership and bus ridership could also play a role in the choice to use bike-sharing. The location of a bike-sharing station could be more significantly controlled for. For instance, bike-sharing stations located close to large public buildings, or busy restaurants, or schools could be used more than bike stations without those nearby facilities.
27
7. Conclusions
In this thesis, the effect that closing a metro station has on bike-sharing ridership was quantified by performing OLS and OLS FE regression on bike-sharing data in Chicago in the period 2014-2017. The results indicate a significant positive effect of closing a metro station on bike-sharing ridership in our preferred models (models 3 and 4). In model 3, closing a metro station increased daily bike-sharing rides by 1.348. In model 4, this effect was 1.458 daily rides. These models are the preferred models, as they control for all time-invariant effects. In model 1 no fixed effects were controlled for. In model 2, a few fixed effects were added, but the result was vastly different from models 3 and 4. This proves that model 2 was flawed, as too many time-invariant effects were missing.
For a lot of variables, the outcomes were in line with previous literature. As expected, fewer bike-sharing rides happen on the weekend or on holidays and the capacity of a station is positively correlated with bike-sharing rides. We can also conclude that the weather has a significant impact on bike-sharing ridership. All four models that have been used in this thesis find these results. As Li et al. (2015), Wang et al. (2013), Rixey (2013), and Noland (2014) concluded in their research, the better the weather conditions are the more bike-sharing rides can be expected. In raining, snowing, and storming conditions fewer rides will be taken. In clear and sunny conditions, more rides will be taken.
What can be concluded from model 2 is that time-invariant effects like bike infrastructure, location, and distance to alternative transport are all correlated with daily bike-sharing rides. This is an interesting find for local governments who are trying to encourage bike-sharing projects, as they can carefully plan the location of bike-sharing stations and improve bicycle infrastructure in the city. City planning can have a great effect on bike-sharing ridership, and future research could quantify this even more precisely.
28
about the metro closings. This information could be critical for this analysis and change the outcomes as a result. Hopefully, a better understanding of the relationship between these two modes of transportation and other variables can help planners and policymakers in their decision-making on urban planning and their mission to encourage people to use bike-sharing more.
8. Areas for improvement and future research
There are various areas for improvement and potential for further research that have emerged from this thesis. Firstly, this natural experiment could have benefitted from a larger sample size; mainly in the “treatment” group of stations experiencing a closing of the nearby metro station. Combining multiple cities to create the same experiment could create a more balanced dataset and potentially more evidence for certain conclusions.
Various fixed effects have the potential to be explored more. The meteorological analysis could be more detailed; the weather conditions could be split in more categories, with more indicators like wind and temperature. The locational variables could be worked out more as well. For example, instead of using a cut-off locational variable like we used in this thesis, one could control for absolute distance. That way the effect of walking distance can be observed in more detail.
The models also deal with the problem that some bike-sharing stations are located in places that have other transport alternatives like the bus, uber, or others. These alternatives could also affect the choice of transportation. Additionally, the problem of multiple metro stations could be an issue. Some bike-sharing stations have more than one metro station within 0.6 miles; that means that if the nearest metro station closes, the natural experiment assumes that the choice of metro transportation vanishes, even though people may just walk a few more minutes to go to another (still opened) metro station. Exploring this issue more could create an improved model.
29
9.
Appendix
9.1 FiguresFig 1. Locations of metro stations in Chicago.
30
Fig 3. The location of the 5 bike-sharing stations within 1 kilometer radius of metro station Montrose – O’Hare.
31
Fig 4. Seasonal trends of weekly bike-sharing rides for years 2014, 2015, 2016, 2017, and the aggregated weekly rides of 2014-2017.
Fig 5. Type of bike lane in % of total observations (N = 101,677). Fig 6. Type of bike lane in amount of bike stations (N = 77).
0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000 1 11 21 31 41 51 R id es Weeknumber
2016
0 5000 10000 15000 20000 25000 1 11 21 31 41 51 R id es Weeknumber2017
0 20000 40000 60000 80000 1 6 11 16 21 26 31 36 41 46 51 R id es Weeknumber2014-2017
35% 16% 13% 28% 8%32
Fig 7. Bike lanes in Chicago as of February 2020.
Fig. 8 Example of a barrier protected bike lane. Fig. 9 Example of a buffered bike lane.
33
9.2 Tables
Table 1A – 1I : Information on metro stations. Includes nearby bike-sharing stations, ID number, distance of the
bike-sharing station to the metro station, and the bike lane infrastructure near the bike-sharing station.
Holiday Name Type Date
MADISON / WABASH ID Distance to metro (in miles) Bike lane
Michigan Ave & Madison St 197 0 None Wabash Ave & Adams St 39 0,2 Protected
Millennium Park 90 0,3 None
Dearborn St & Monroe St 49 0,3 Protected Michigan Ave & Jackson Blvd 284 0,5 None Lake Shore Dr & Monroe St 76 0,6 None Michigan Ave & Washington St 43 0,3 Protected State St & Randolph St 44 0,3 Protected
RANDOLPH / WABASH ID Distance to metro (in miles) Bike lane
State St & Randolph St 44 0,1 Protected Michigan Ave & Washington St 43 0,1 Protected Wabash Ave & Wacker PI 194 0,1 Protected Michigan Ave & Lake St 52 0,2 None Stetson Ave & South Water St 264 0,4 None Columbus Dr & Randolph St 195 0,3 Protected Daley Center Plaza 81 0,2 Protected Clark St & Randolph St 51 0,3 Protected Clark St & Lake St 38 0,3 None
WASHINGTON / WABASH ID Distance to metro (in miles) Bike lane
Michigan Ave & Madison St 197 0,1 None Wabash Ave & Adams St 39 0,2 Protected
Millennium Park 90 0,4 None
34
MONTROSE / O’HARE ID Distance to metro (in miles) Bike lane
Knox Ave & Montrose Ave 592 0 None Kildare Ave & Montrose Ave 630 0,2 None Keystone Ave & Montrose Ave 495 0,3 None Kilbourn Ave & Irving Park Rd 590 0,5 None Milwaukee Ave & Cuyler Ave 589 0,5 None
STATE / LAKE ID Distance to metro (in miles) Bike lane
Wells St & Hubbard St 212 0,6 Shared State St & Randolph St 44 0,1 Protected Wabash Ave & Wacker PI 194 0,1 Protected Michigan Ave & Lake St 52 0,2 None Stetson Ave & South Water St 264 0,3 None Clark St & Lake St 38 0,2 None Clark St & Randolph St 51 0,4 Protected Daley Center Plaza 81 0,3 Protected Franklin St & Lake St 164 0,4 Buffered State St & Kinzie St 47 0,3 Protected Rush St & Hubbard St 125 0,5 Bike lane LaSalle St & Illinois St 181 0,6 Bike lane
CERMAK-MCCORMICK ID Distance to metro (in miles) Bike lane
Wabash Ave & Cermak Rd 42 0 Protected Wentworth Ave & Cermak Rd 120 0,3 None Wentworth Ave & 24th St 132 0,5 None Wells St & 19th St 218 0,6 Bike lane State St & 19th St 178 0,2 Protected Michigan Ave & 18th St 273 0,6 Buffered Calumet Ave & 21st St 370 0,5 Shared Calumet Ave & 18th St 338 0,6 Shared
35
CALIFORNIA/MILWAUKEE ID Distance to metro (in miles) Bike lane
California Ave & Milwaukee Ave 123 0 Bike lane Milwaukee Ave & Rockwell St 222 0,2 Bike lane California Ave & Francis PI 259 0,2 Shared Stave St & Armitage Ave 185 0,5 Shared Humboldt Blvd & Armitage Ave 507 0,5 Shared Campbell Ave & Fullerton Ave 504 0,6 None California Ave & Altgeld St 502 0,3 Shared Kedzie Ave & Palmer Ct 290 0,5 Bike lane Western Ave & Winnebago Ave 116 0,6 Bike lane Kedzie Ave & Milwaukee Ave 260 0,6 Bike lane
HARRISON ID Distance to metro (in miles) Bike lane
36
Table 2: Observations per year in the dataset.
Year 2014 2015 2016 2017 Total
Observations 20,938 25,550 27,084 28,105 101,677
Table 3: Bike-sharing rides per year, total and average.
Year 2014 2015 2016 2017 Total
Rides 423,060 524,068 625,329 672,692 2,245,149
Avg 20.21 20.51 23.09 23.94 22.08
Table 4: Weather conditions in observations and in % of total observations (N = 101,667). Tstorms Clear NotClear RainorSnow Cloudy Total
1,927 5,207 827 11,532 82,184 101,677 1,90% 5,12% 0,81% 11,34% 80,83% 100%
Table 5: Observations of longwalk and citycentre variables (N = 101,677).
Longwalk = 1 Longwalk = 0 Citycentre = 1 Citycentre = 0
33,524 (32,97%) 68,153 (67,03%) 28,057 (27,59%) 73,620 (72,41%)
Table 6: Observations of MetroClosed variables (N=101,677).
MetroClosed = 0 MetroClosed = 1
96,968 (95,38%) 4,709 (4,62%)
DAMEN/MILWAUKEE ID Distance to metro (in miles) Bike lane
37
Table 7: Regression results.
(1) (2) (3) (4)
VARIABLES Model 1 Model 2 Model 3 Model 4
MetroClosed -0.643** -2.470*** 1.348*** 1.458*** (0.282) (0.277) (0.248) (0.248) Weekend -12.36*** -12.36*** -12.37*** -12.36*** (0.128) (0.124) (0.104) (0.104) Capacity 1.368*** 1.017*** 0.294*** 0.285*** (0.00775) (0.00942) (0.0208) (0.0209) Clear 0.357 0.408 0.615*** 0.628*** (0.278) (0.271) (0.228) (0.228) NotClear -2.151*** -2.066*** -1.911*** -1.989*** (0.675) (0.656) (0.553) (0.553) RainorSnow -5.132*** -5.095*** -4.913*** -4.889*** (0.203) (0.197) (0.166) (0.166) Tstorms -3.716*** -3.704*** -3.709*** -3.775*** (0.445) (0.432) (0.364) (0.365) preclosingweek5 2.328*** (0.772) preclosingweek4 4.572*** (0.772) preclosingweek3 0.520 (0.772) preclosingweek2 3.211*** (0.779) preclosingweek1 1.478* (0.779) postclosingweek1 0.720 (0.647) postclosingweek2 0.998 (0.803) postclosingweek3 0.319 (0.899) postclosingweek4 0.280 (0.741) postclosingweek5 1.619*** (0.570) citycentre 8.808*** (0.161) longwalk 0.803*** (0.127) bikelane 4.760*** (0.174) sharedlane 7.425*** (0.190) protectedbikelane 7.443*** (0.152) bufferedbikelane 4.038*** (0.235) Constant -16.15*** -16.21*** 2.169*** 2.084*** (0.485) (0.482) (0.547) (0.549) Observations 101,677 101,677 101,677 101,677 R-squared 0.409 0.442 0.374 0.374
Meteorological variables Yes Yes Yes Yes
Number of StationID 77 77 77 77
Post-/Pre-closing variables No No No Yes
Fixed Effects Time variables No Yes No Yes Yes Yes Yes Yes Standard errors in parentheses
38
Table 8: National holidays in Chicago in period 2014-2017.
Name: Type: Date:
Martin Luther King Day Federal 3rd Monday of January
Lincoln’s Birthday Government 12/02
President’s day Federal 3rd Monday of February
Memorial day Federal Last Monday in May
Independence day Federal 04/07
Labor day Federal 1st Monday of September Columbus day Federal 2nd Monday of October
Thanksgiving Federal 4th Thursday of November
Christmas day Federal 25/12
New Year’s day Federal 01/01
Table 9: Temporary closing dates of metro stations in period 2014-2017.
Metro Station Dates Closed
Madison / Wabash Closed from 18/03/2015 to 31/12/2017 Washington / Wabash Closed from 01/08/2017 to 27/08/2017 Montrose / O'Hare Closed from 04/06/2016 to 05/06/2016 State / Lake Closed on 24/09/2016
Cermak-McCormick Closed from 01/01/2015 to 02/02/2015 Damen / Milwaukee Closed from 20/10/2014 to 19/12/2014 California / Milwaukee Closed from 05/09/2014 to 14/10/2014
39
10. Bibliography
10.1 Article References
Bachand-Marleau, J., Lee, B. H. Y., & El-Geneidy, A. M. (2012). Better Understanding of Factors Influencing Likelihood of Using Shared Bicycle Systems and Frequency of Use. Transportation Research Record: Journal of the Transportation Research Board, 2314(1), 66–71. https://doi.org/10.3141/2314-09
Barbour, N., Zhang, Y., & Mannering, F. (2019). A statistical analysis of bike-sharing usage and its potential as an auto-trip substitute. Journal of Transport & Health, 12, 253– 262. https://doi.org/10.1016/j.jth.2019.02.004
Bullock, C., Brereton, F., & Bailey, S. (2017). The economic contribution of public bike-share to the sustainability and efficient functioning of cities. Sustainable Cities and Society, 28, 76–87. https://doi.org/10.1016/j.scs.2016.08.024
Campbell, K. B., & Brakewood, C. (2017). Sharing riders: How bikesharing impacts bus ridership in New York City. Transportation Research Part A: Policy and Practice, 100, 264–282. https://doi.org/10.1016/j.tra.2017.04.017
Choi, Y., & Choi, E. J. (2020). Sustainable Governance of the Sharing Economy: The Chinese Bike-Sharing Industry. Sustainability, 12(3), 1195.
https://doi.org/10.3390/su12031195
40
Fishman, E., Washington, S., & Haworth, N. (2012). Barriers and facilitators to public bicycle scheme use: A qualitative approach. Transportation Research Part F: Traffic
Psychology and Behaviour, 15(6), 686–698. https://doi.org/10.1016/j.trf.2012.08.002
Fishman, E., Washington, S., & Haworth, N. (2013). Bike-share: A Synthesis of the Literature. Transport Reviews, 33(2), 148–165.
https://doi.org/10.1080/01441647.2013.775612
Fishman, E., Washington, S., & Haworth, N. (2014). Bike-share’s impact on car use: Evidence from the United States, Great Britain, and Australia. Transportation Research Part D: Transport and Environment, 31, 13–20.
https://doi.org/10.1016/j.trd.2014.05.013
Fishman, E., Washington, S., & Haworth, N. (2015). Bikeshare’s impact on active travel: Evidence from the United States, Great Britain, and Australia. Journal of Transport & Health, 2(2), 135–142. https://doi.org/10.1016/j.jth.2015.03.004
Gebhart, K., & Noland, R. B. (2014). The impact of weather conditions on bikeshare trips in Washington, DC. Transportation, 41(6), 1205–1225. https://doi.org/10.1007/s11116-014-9540-7
Goh, C. Y., Yan, C., & Jaillet, P. (2019). A Locational Demand Model for Bike-Sharing. SSRN Electronic Journal, 1–47. https://doi.org/10.2139/ssrn.3311371
41
Liu, Z., Jia, X., & Cheng, W. (2012). Solving the Last Mile Problem: Ensure the Success of Public Bicycle System in Beijing. Procedia - Social and Behavioral Sciences, 43, 73– 78. https://doi.org/10.1016/j.sbspro.2012.04.079
Ma, X., Ji, Y., Yang, M., Jin, Y., & Tan, X. (2018). Understanding bikeshare mode as a feeder to metro by isolating metro-bikeshare transfers from smart card data. Transport Policy, 71, 57–69. https://doi.org/10.1016/j.tranpol.2018.07.008
Martin, E. W., & Shaheen, S. A. (2014). Evaluating public transit modal shift dynamics in response to bikesharing: a tale of two U.S. cities. Journal of Transport Geography, 41, 315–324. https://doi.org/10.1016/j.jtrangeo.2014.06.026
Qiu, L.-Y., & He, L.-Y. (2018). Bike-sharing and the Economy, the Environment, and Health-Related Externalities. Sustainability, 10(4), 1145. https://doi.org/10.3390/su10041145
Rixey, R. A. (2013). Station-Level Forecasting of Bikesharing Ridership. Transportation Research Record: Journal of the Transportation Research Board, 2387(1), 46–55. https://doi.org/10.3141/2387-06
Rojas-Rueda, D., de Nazelle, A., Tainio, M., & Nieuwenhuijsen, M. J. (2011). The health risks and benefits of cycling in urban environments compared with car use: health impact assessment study. BMJ, 343(aug04 2), d4521.
https://doi.org/10.1136/bmj.d4521
Saracoglu, H. (2018). A Snapshot of Chicago Divvy Bike-sharing System. Northwestern University.
