Impact of the REV on Montreal's Bike Share Program

This project assesses how the development of the Reseau Express Velo (REV) - a recently built network of high-quality bike infrastructure in Montreal's central neighborhoods - impacted usage of the city's bikeshare program, Bixi. I employ a difference-in-differences research design, comparing outcomes for Bixi stations located alongside the REV to those located farther away, to assess the REV's impact on bikesharing use.

Context

In 2019, Valerie Plante, Mayor of Montreal and leader of Projet Montreal, announced that the city would be undertaking a project to build more higher-quality bike infrastructure, beginning in its central neighborhoods. This would include the construction of new, wide, well sign-posted, and protected bike lanes with synchronized street lights to help Montrealers more comfortably and safely traverse large distances. These paths would also be prioritized for snow clearing in winter, making them usable year-round. You can read more about the REV here.

Projet Montreal put forward plans to build 5 such axes:

• Axis 1: Berri-Lajeunesse-Saint-Denis: This axis is the longest in the REV network and provides North-South coverage in the center of the city. It links the boulevard Gouin with the rue Roy. The axis was inauguarateed on 11/07/2020.
• Axis 2: Viger-Saint-Antonine-Saint-Jacques: This axis serves a short East-West segment between rue Guy and rue De Coucelle, predominantly in the city's Ville-Marie borough.
• Axis 3: Souligny: This axis, running East-West between the rue Honore-Beaugrand and avenue Hector, is furthest from the city center.
• Axis 4: Peel: This axis serves a short North-South corridor on Peel, a major commercial shopping street in the city's downtown core, between avenue des Pins and rue Smith.
• Axis 5: Bellechasse: An East-West axis running on Bellechase between de Gaspe and Chatelain, predominantly in the Rosemont-La Petite-Patrie borough, and intersecting with Axis 1 of the REV at Saint-Denis/Bellechasse.

While the REV has been touted as a success by bike enthusiasts, this is the first attempt - as far as I know - to causally evaluate the its effect on rideshare utilization.

Data

The project primilarly relies on ride-level data made freely available by Bixi for the period April 2014 - July 2024. I also use data from the City of Montreal, who has geocoded all of the city's existing bike network, as well as daily weather data from Environment Canada.

Code

Code for the project is written entirely in Python and separated into 9 sections. I run the code primarily on a computing cluster, given that the complete raw dataset is too large to be saved in memory. To run the code without modification, begin by creating a project directory and specifying the filepath in the Preliminaries section of the script. Next, create a subdirectory called data. Store the Bixi ride-level data in year-specific folders in data/ridership/, geocoded bike network data from the City of Montreal in data/bike_network/, and weather data from Environment Canada in year-specific folders in data/weather/. In the same project directory, create empty folders called figures and output to collect results.

Those looking to use the script to generate a GIF of ridershare usage by Bixi station, as is done in the script, will need a chart_studio account, which is used to save the map images consituting the GIF.

1. Preliminaries

In this section, I simply import modules that I'll need to conduct the work. I take advantage of a number of widely used libraries for data science, spatial analysis, and econometrics. I also specifying a filepath, which is automatically selected based on whether I am working on my personal computer or computing cluster.

2. Importing and Cleaning Ridership Data

I begin by reading and appending Bixi's ride-level microdata. In some years, Bixi provides ride-level data by month, while in other years all of the ridership data is included in a single dataset. Variable names change somewhat through time, as do date formats. The code handles these intertemporal inconsistencies. The dataset includes all of the approximately 62 million rides completed on Bixi bikes between April 2014 and July 2024.

In order to generate aggregate statistics by station, it is important to have a unique and time-invariant station identifier. The ride-level data from Bixi provides two potentially useful identifiers: $\text{Station Name}$ and $\text{Station Code}$. However, both can be unreliable as the same station may use different names or different station codes, in the same year and through time.

To address this issue, I grouped Bixi stations together based on whether I believe docking stations with different names actually refer to the same station. I did so by retaining unique station names and sorting by coordinates. I create a crosswalk file, data/ridership/id_crosswalk.xlsx, which I merged with the ride-level microdata. I also updated station names and coordinates. For each value of Station ID, I replaced each station's name with its modal station name. For each Station ID-Year pair, I replaced the station's coordinates with its modal coordinates.

Here is an example with just six observations, which are sufficient for illustrating the process. Several trips taken just a few minutes apart on the same day in 2018 are found to originate from seemingly different stations. Grouping stations by either station name or station code would be erroneous. All of these trips should originate from the same Bixi station, installed at Vendome Metro since 2014. This becomes evident when manually verifying the location of the Bixi docking stations.

Year	Date	Station Name	Latitude	Longitude	Station Code
2018	2018-04-23 17:42	de Vendôme / de Maisonneuve	45.4744	-73.604	6418
2018	2018-04-23 18:00	de Vendôme / de Maisonneuve	45.4744	-73.604	6418
2018	2018-04-23 18:03	Marlowe / de Maisonneuve	45.4739	-73.6047	6080
2019	2019-06-10 17:45	Métro Vendôme (de Marlowe / de Maisonneuve)	45.4739	-73.6047	6080
2019	2019-06-10 17:49	Métro Vendôme (de Marlowe / de Maisonneuve)	45.4739	-73.6047	6080
2019	2019-06-10 17:50	Métro Vendôme (de Marlowe / de Maisonneuve)	45.4739	-73.6047	6080

After merging the ride-level data with the crosswalk file, these rides are assigned the same station ID. In addition, I choose to replace the station's name with its mode and the station's coordinates with their year-specific mode. This ensures that I capture changes to the docking station's precise location if, for instance, Bixi decides to move a docking station across the street from one year to another.

Year	Date	Station Name	Latitude	Longitude	Station ID
2018	2018-04-23 17:47	Métro Vendôme (de Marlowe / de Maisonneuve)	45.4744	-73.604	174
2018	2018-04-23 18:00	Métro Vendôme (de Marlowe / de Maisonneuve)	45.4744	-73.604	174
2018	2018-04-23 18:03	Métro Vendôme (de Marlowe / de Maisonneuve)	45.4744	-73.604	174
2019	2019-06-10 17:45	Métro Vendôme (de Marlowe / de Maisonneuve)	45.4739	-73.6047	174
2019	2019-06-10 17:49	Métro Vendôme (de Marlowe / de Maisonneuve)	45.4739	-73.6047	174
2019	2019-06-10 17:50	Métro Vendôme (de Marlowe / de Maisonneuve)	45.4739	-73.6047	174

3. Creating Outcome Variables of Interest

Here, I create three outcome variables: trip count, trip distance, and trip duration.

Trip count measures the number of trips undertaken, by Bixi station and date.

As Bixi does not provide trip-level GPS data, which would be needed to track the precise journey undertaken by a user, I instead measure trip distance as the Haversine distance between the starting and ending station. I recognize that this is a flawed measure and a lower bar for the actual distance traversed on any given trip. For all trips with a distance of 0 - that is, trips beginning and ending at the same docking station - I replace trip distance with a missing value. As such, these trips will not be included in regression analysis in instances where the outcome of interest is trip distance.

Trip duration is calculated as the difference between a journey's start time and end time.

I remove Bixi trips with implausible distances or journey times to reduce the impact of outliers on parameter estimates. This removes relatively few observations.

4. Identifying Treated Bixi Stations

Rather than consider all axes of the REV, I focus exclusively on Axis 1 because it provides the best case study for assessing the REV's impact. Other axes were rolled out in a more staggered fashion, and were subject to delays and additional works. Axis 1, on the other hand, was inaugurated in its entirety on the same day and has been subject to fewer disruptions in the years since.

I define treated Bixi stations as those located within 100 meters of the REV path and control stations as those located between 100 and 300 meters from the REV. These thresholds are informed by the existing literature, which finds that up to 500 meters is a reasonable distance to walk to bikeshare stations. I select a lower threshold, in large part because of high Bixi station density in the neighborhood served by the REV's Axis 1.

The City of Montreal maintains a database with information on all bike paths in the city. Each path is subdivided into segments and precisely geocoded. As such, I can simply assign stations to treatment by employing the following procedure:

1. Select a Bixi station.
2. For each Bixi station, iterate through every segment of the REV.
3. For each REV segment, create a polygon using the coordinates bounding the segment.
4. Calculate the distance between the Bixi station and each line segment formed from any pair of vertices associated with the REV segment.
5. If the distance calculated is less than 100 meters, assign the Bixi station to treatment.
6. If the distance calculated is beteween 100 meters and 300 meters, continue iterating through REV segments. If no other REV segment is found to be less than 100 meters from the Bixi station, assign the Bixi station to the control group.
7. If the distance calculated between the Bixi station and the REV is never found to be less than 300 meters, assign the station to neither the treatment nor the control groups.
8. Return the treatment status and the distance between Bixi station and REV path.

Here is a plot showing the location of the REV Axis 1. Bixi stations are classified as either Treated, Control, or Other, depending on how far they are located from the path of the REV's Axis 1.

5. Exploring Data

At this point, I have all of the variables needed to perform the econometric analysis. Before doing so, however, I produce descriptive statistics to verify that the data was correctly prepared.

I first plot average daily ridership by month, for every month between January 2014 and July 2024. Clearly, there is strong seasonality in bike ridershp, with usage of Bixi peaking in summer months. I verify that daily ridership calculated from the microdata lines up with Bixi's self-reported ridership statistics.

Next, I plot average daily ridership by day of the week in July 2024. In line with expectations, Saturday and Friday are the most popular days for bikesharing.

Finally, I plot the number of active Bixi stations over time, only in months in which the service is operating. Note that, in the winter of 2023, Bixi piloted a project whereby it operated around 150 stations in the city's core for the first. This explains the sharp and temporary decline in the number of active Bixi stations. In 2024, the rideshare service operated around 850 stations.

6. Mapping

Given that the data is spatial in nature, I choose to create maps as well. Each bubble represents a Bixi station. The bubble's color scales in accordance with the number of bikeshare trips originating from that station while the bubble's size scales with the total distance travelled by bikeshare users on trips originating from that station. In animating the data, it's clear to see the gradual expansion of the network into neighborhoods further from the city center, as well as increased rideshare usage, both at the extensive and intensive margins.

7. Preparing Data for Econometric Analysis

Before I can perform regressions, I make the following adjustments. First, I convert key variables for the difference-in-difference regression to binary type. Second, I seasonally adjust the outcome variables. Third, I calcualte the distance between each Bixi station and the city's central business district. Finally, I merge in additional covariates measuring daily mean temperature, precipitation, and amount of snow on ground in Montreal.

8. Assessing Parallel Trends

To ensure that outcomes evolved similarly prior to treatment for both treatment and control groups, and to verify that bikeshare activity at treated stations did not somehow frontrun the completion of the REV, I plot seasonally adjusted outcomes in event time. The event time variable measures time elapsed since the inauguration of the REV's Axis 1 on 11/07/2020. In an effort to better assess trends, I plot only a single month, November, for every year.

Outcomes evolved quite similarly for both treated and control groups prior to the construction of the REV's Axis 1. At the time of treatment, usage of Bixi stations in both treated and control groups notably increased and began to grow more quickly in November 2020, following the REV's completion. Ridership increases more for treated stations than for control stations, and remains more elevated through the post-treatment period. I attribute at least part of the decline in ridership observed in November 2019 to road work undertaken along rue St. Denis, both for the REV's construction and for the completion of other works.

9. Model Estimation

I estimate a standard difference-in-difference model with $\text{Post}$, $\text{Treated}$, and $\text{Post} \times \text{Treated}$ terms. In addition to the key difference-in-differences regressors, I include a control for the distance between the Bixi station and the REV path as well as distance between the Bixi station and the city's central business district. Finally, I include observable weather-related covariates that I think may impact outcomes, including temperature, precipitation, and the amount of snow on the ground, as well as a full set of monthly dummies. Robust standard errors are used. The regressions are performed at the weekly-station level as outcomes are much less noisy than at a daily frequency. The most comprehensive specification is shown below.

$Y_{it} = \alpha + \beta_{1}\text{Treated}_{i} + \beta_{2}\text{Post}_{t} + \beta_{3}(\text{Treated}_{i} \times \text{Post}_{t}) + \beta_{4}\text{Distance to REV}_{it} + \beta_{5}\text{Distance to CBD}_{it} + \sum_{n}\beta_{n}X_t + \epsilon_{it}$

Where:

• $( Y_{it} )$ is the seasonally adjusted outcome variable for Bixi station $i$ in week $t$.
• $( \alpha )$ is the intercept.
• $( \text{Treated}_i )$ is a binary variable indicating the treatment group (1 if treated, 0 if control).
• $( \text{Post}_t )$ is a binary variable indicating the post-treatment period (1 if after treatment, 0 if before). The treatment date is 11/07/2020.
• $( \text{Treated}_i \times \text{Post}_t )$ is the difference-in-difference estimator, and is calculated as the interaction of the post-treatment period and the treatment group.
• $( \text{Distance to REV}_{it} )$ is the distance, in hundreds of meters, between Bixi station $i$ and the nearest segment of the REV path in week $t$
• $( \text{Distance to CBD}_{it} )$ is the distance between Bixi station $i$ in week $t$ and the central business district of Montreal.
• $( X_t )$ is a vector of time-specific covariates, including mean weekly temperature (degrees celcius), precipitation (mm), and snow on ground (cm).
• $( \epsilon_{it} )$ is the error term.

Because I believe the REV may elicit heterogenous treatment effects - that is, the impact of the REV may vary across Bixi stations - I also choose to estimate a two-way fixed effects (TWFE) model. By absorbing all time-invariant, station specific factors, as well as date-specific shocks, this model is also better suited to absorb unsoberved factors that may act to confound parameter estimates in the standard difference-in-differences model. The TWFE model is shown below.

$Y_{it} = \alpha + \theta_{i} + \mu_{t} + \delta(\text{Treated}_{i} \times \text{Post}_{t}) + \epsilon_{it}$

Where:

• $( Y_{it} )$ is the seasonally adjusted outcome variable for Bixi station $i$ in week $t$.
• $( \alpha )$ is the intercept.
• $( \theta_{i} )$ are Bixi station fixed effects.
• $( \mu_{i} )$ are weekly date fixed effects.
• $( \text{Treated}_{i} \times \text{Post}_{t} )$ is the difference-in-difference estimator, and is calculated as the interaction of the post-treatment period and the treatment group.
• $( \epsilon_{it} )$ is the error term.

Estimation results are exported as a LaTeX file, which is then interpreted in Overleaf.

Discussion of Results

Here are the regression results from the standard difference-in-differences model. For each outcome, I estimate three models. Each successive specification includes additional covariates.

The difference-in-difference estimator captures the treatment effect. It is found to be positive, statistically significant, and economically meainngful for all three outcomes. In particular, results indicate that, holding all else equal, stations located in close proximity to the REV see around 60 more weekly trips than those located further away. Trips originating from treated stations end 43 meters farther and last around 13 seconds longer, on average. If the research design used is convincing, then improvements in Bixi usage are a direct result of the REV's completion, rather than some confounding factor.

Distance to the REV path is negatively and significantly related with the number of trips taken. Parameter estimates suggest that, for every 100 meters further a Bixi station is from the REV, the number of weekly trips taken falls by around 4, on average. Distance to the REV has a comparatively larger effect on trip distance and trip duration. Trips originating at Bixi docking stations located 100 meters from the REV are expected to travel 40 fewer meters and last 2 fewer minutes, on average, than Bixi stations located adjacent to the path.

Distance to the CBD is always found to enter statistically significantly, but with a sign that varies. My interpretation is that, the number of trips taken at any given Bixxi station declines as the distance to the CBD grows, on account of less central neighborhoods having fewer Bixi stations at which to dock or rent a bike, as well as poorer auxiliary bike infrastructure. On the other hand, the distance and duration of ridersharing trips grows longer as a function of distance to the CBD because urban amenities are more sparsely distributed.

Other control variables have the expected sign, with temperature being positiely associated with ridesharing while precipitation and snow on ground are found to be negatively related with ridesharing. However, only the amount of snow on the ground is found statistically significantly benefit or reduce bikesharing.

I find the sign, magnitude, and statistical significance of the key results to be robust to changes in the treated/control thresholds and the addition of monthly dummies.

To conclude, results indicate that the REV broadly improved ridership at stations located nearest to its path. However, without ride-level user IDs it is impossible for me to infer whether increased usage is the result of new riders being induced by the REV, or simply a spatial reallocation of existing users. That is, it is possible the treatment effect is being driven by existing Bixi users choosing to rent their bikes from a Bixi station closer to the newly completed REV path, rather than new users choosing to begin using the bikeshare program.