The Impact of a Single Bus Rapid Transit Corridor on Transit Ridership: The Winnipeg Example

Abstract

This research explores how the implementation of a single bus rapid transit (BRT) corridor affected transit ridership change in Winnipeg, Manitoba, Canada. Key issues in measuring ridership change resulting from BRT include (1) understanding stop-level- rather than corridor-level change; (2) examining the ridership impacts of new infrastructure where there is no comparable pre-BRT infrastructure; and (3) assessing piecemeal implementation of BRT. To address these issues, we undertook a quasi-experimental study using agglomerative hierarchical clustering (AHC), propensity score matching (PSM), and t-tests with Cohen’s d to determine BRT’s causal ridership impact. The use of AHC and PSM in what we refer to as cluster-level modeling provided an improved method for measuring causal ridership change at the stop cluster level in areas with no pre-BRT stations. The results revealed no statistical evidence that BRT caused increased transit ridership for stop clusters directly along the BRT corridor. However, the results did indicate that stop clusters for routes connecting to a BRT station experienced an increase in transit ridership. The importance of such findings is grounded in understanding that a limited number of stops along a single corridor may not be enough to affect transit ridership, yet BRT’s flexibility in being able to operate off the BRT corridor does enhance transit ridership.

Keywords

planning and analysis transportation demand forecasting ridership estimation modeling public transportation bus transit systems bus rapid transit

Bus rapid transit (BRT) provides a viable and effective method for municipalities worldwide to increase transit ridership and positively alter travel behavior. Research has found that enhancing bus service creates positive benefits for transit ridership (1 –3). Numerous studies have examined ridership changes along a single rapid transit corridor (4, 5) or examined the corridors as part of a larger rapid transit network (6, 7) and have generally found increased transit ridership as part of BRT transit upgrades.

Nonetheless, three issues remain in measuring ridership changes stemming from BRT. First, existing research largely examines change at the corridor level (2, 4). However, corridor-level analyses can overlook changes at the stop level and corridor-level variations can obscure important influencers of transit ridership, such as land use or population density ( 8 ). In this regard, direct ridership models (DRMs) examining changes at the stop level may be more effective at measuring change at exact BRT stations. DRMs provide direct estimates of the sum of stop-level rider on- and off-boardings and can be used to estimate transit ridership directly from station boardings or stop environments ( 9 ).

However, a second issue arises if analyzing ridership changes because of BRT and using DRMs—that is, how to measure change when no station existed before BRT implementation. This is a key issue when BRT infrastructure requires large right-of-ways (ROWs), making it less likely to be constructed in existing built-up areas. To overcome this issue, research has incorporated agglomerative hierarchical clustering (AHC) to create stop environments or stop cluster areas and therefore capture ridership needs surrounding a particular stop location instead of at an exact stop (3, 6). In addition to work by Brands et al. ( 6 ) and Stewart et al. ( 3 ) research incorporating AHC or measuring stop area level ridership changes resulting from new transit infrastructure has remained sparse.

Third, financial constraints, land availability and acquisition, and a lack of political will can stymie the buildout of an entire BRT network. Such issues can also limit space for new transit investments, calling into question the value added from a single rapid transit corridor. Cost savings and the least resistance in land acquisition can cause BRT to open in underutilized and often ignored locations ( 10 ), but serving the same termini as a more congested or developed route. A downside of this is that implementing BRT infrastructure in such previously underutilized spaces may mean limited initial ridership potential. Yet, it may also provide the opportunity to better shape land use with minimal disruption and thus increase transit ridership over the long term. Nevertheless, understanding value gain in relation to transit ridership can show the importance of initial transit investments.

This research addresses the above issues by examining the causal impact of an initial BRT corridor on transit ridership. Through Winnipeg, Manitoba as a case study, we explored the potential for BRT investments on underutilized sites to affect ridership, allowing for a more nuanced understanding of transit use. Specifically, we undertook a quasi-experimental study using AHC, propensity score matching (PSM), and t-tests with Cohen’s d to determine the causal impact of new BRT investments on Winnipeg’s transit ridership. This paper contributes to existing research in two regards. First, it presents a cluster-level modeling (CLM) method to measure ridership change at the stop cluster level in areas with no available comparators. Second, it highlights the impact of a single BRT corridor, in the absence of a substantial citywide network. These issues are increasingly important as cities struggle to implement rapid transit in built-out areas, and because infrastructure planning and investments can take decades for implementation.

Existing Literature

Rapid Transit and Transit Ridership

The majority of quantitative research determining or predicting transit ridership often includes similar internal and external factors (see Hoang-Tung et al. [ 1 ], Cervero et al. [ 7 ], Kerkman et al. [ 8 ], and Currie and Delbosc [ 11 ]; also see Chan and Miranda-Moreno [ 12 ] and Dill et al. [ 13 ] for summaries of factors included in ridership models). However, the magnitude or effects of the factors often vary substantially from model to model. For instance, researchers suggest that external factors have a greater impact on transit ridership than internal factors—as seen largely with increases in transit supply across the United States and Canada, but limited ridership increases ( 11 ). In such cases, land use and density generally have the greatest impacts on transit use ( 11 ). However, other research suggests that internal factors, like increased frequencies or connectivity to other routes, also substantially influence transit ridership increases ( 7 ). For BRT specifically, providing exclusive bus-only lanes has been shown to boost ridership and offer better overall transit service and operations (e.g., on-time performance, travel-time savings) (2, 7).

Ultimately, the introduction of BRT or BRT features increases ridership for locations receiving the transit upgrade (1, 2, 7). For example, Currie and Delbosc ( 11 ) examine the factors influencing ridership on conventional bus and BRT routes and found that service levels are the dominating factor of ridership. However, they also found that BRT measures like exclusive ROWs significantly influence ridership. Using survey data and based on a quasi-experimental design, Hoang-Tung, et al. ( 1 ) additionally found that public transportation use increased marginally more in areas receiving- compared to areas not receiving – BRT treatment.

Measuring Impacts of New Transit Infrastructure on Transit Ridership

Although BRT is generally shown to increase ridership, how to isolate the impacts of new rapid transit ridership is a common concern in existing research (12, 14). Isolating the impacts is important to be able to determine whether changes are specifically the result of new transit investments or stem from overall area changes. To isolate new transit impacts, researchers often compare ridership changes in the new locations (i.e., treated locations) with similar locations that did not receive new transit upgrades (i.e., control locations) (1, 8). Here, ridership changes in treated locations above-and-beyond those in control groups indicate changes in transit ridership for BRT compared with non-BRT locations.

Comparing treated with control locations is useful as it compares areas with similar profiles, which may include location context, built environment, bus service, and demographic characteristics. Research has shown PSM to be an effective method in selecting similar areas (1, 14, 15). PSM helps find almost identical locations in the control group to compare with locations that have received treatment based on a single index—the propensity score or PS (1, 14). Essentially, PS is the probability of a unit of analysis receiving a particular treatment provided a given set of determinants are present ( 16 ).

Research examining ridership impacts often aggregates ridership to the entire transit corridor (4, 5, 12). However, a corridor-level approach is limited because it can overlook fluctuations in stop-level characteristics like land use or population density that can affect transit ridership (8, 14) Conversely, a DRM provides a direct measure of transit ridership and a more fine-grained analysis of surrounding station area characteristics compared with corridor studies. Overall, DRMs use direct ridership at the stop level to analyze changes to or influences of transit ridership (6, 7, 17).

Whereas DRMs avoid the aggregation issue and present direct ridership, examining ridership before and after upgrades becomes an issue if no exact rapid transit or comparable bus stop existed before the upgrades. Researchers attempt to overcome this issue by measuring before–after changes at the stop cluster, rather than at the direct stop level (3, 6, 8, 11). For instance, Stewart et al. compared changes in bus ridership between corridors that received BRT upgrades and those that did not across King County, WA ( 3 ). To measure changes from before to after BRT operations, the authors first created a 1/8 mi or 660 ft network distance around BRT stops. They identified all stops within this buffer both before and after BRT implementation and then averaged ridership in this buffer area for both periods, creating comparable transit station areas with which to measure change. Brands et al. also undertook a before–after assessment of a new metro line’s impact on ridership, among other factors, in Amsterdam ( 6 ). They used smart card data from before (a 5-week period before) and after (a 6-week period 7 weeks after) the new line’s opening. They formed stop clusters through AHC and using a 750-m Euclidean distance threshold for the clustering. They found, among other benefits, (1) a 4% increase in transit ridership and (2) journeys shifted from clusters without metro stations to those with metro stations. Thus, clustering stops can account for ridership attraction or need within the vicinity of a rapid transit stop before actual implementation. This provides a comparable unit of analysis with which to measure change before-to-after rapid transit implementation.

The Winnipeg Context

After decades of planning, Winnipeg, Manitoba, Canada, opened its first BRT corridor in 2012. With a population of approximately 700,000 people (2016), Winnipeg is the seventh largest city in Canada, and the provincial capital. In comparison to other mid-sized Canadian cities, Winnipeg has similar shares of commuting by transit, but implemented rapid transit relatively recently (Table 1).

Table 1.

Comparison of Rapid Transit and Usage in Mid-Sized Canadian Cities

	Winnipeg	Edmonton	Ottawa
Population (2016)	711,925	1,062,643	989,657
Rapid transit type (year opened)	BRT (2012)	LRT (1978)	BRT (1983)
			LRT (2019)
Rapid transit extent (# of stations)	11 km (16)	24.3 km (18)	BRT—31 km (44)
			LRT—12.5 km (13)
Percentage of commute trips by transit (2016) ( 18 )	14.9	14.6	20.6

Note: BRT = bus rapid transit; LRT = light rail transit.

The first phase of the BRT corridor (Southwest Transitway, Phase 1) was 3.6 km, with four stations. This portion of the Southwest BRT corridor was the first of a proposed network of corridors expected to open in Winnipeg creating a comprehensive, dedicated ROW BRT network for the city. The Southwest Transitway, Phase 2 opened in 2020 completing the Southwest Transitway connecting downtown Winnipeg with the University of Manitoba, two major employment hubs (Figure 1). Service on the BRT was planned to be both of higher frequency (5 min peak/10 min off-peak, compared with 15 to 60 min on local or community routes) ( 19 ) and higher speed (up to 80 km/h) than regular bus routes. Portions of both phases were constructed within the ROW of an existing freight rail corridor. This route selection, although minimizing land acquisition costs and allowing for new development potential, resulted in multiple BRT station sites with little adjacent existing development ( 20 ).

Figure 1.

Winnipeg bus rapid transit (BRT) system.

The Winnipeg case demonstrates a piecemeal approach to BRT development, limited by the challenges of siting new infrastructure in existing built-up areas, as well as political and financial constraints. Similar to the existing line, disused rail corridor ROWs are also being considered for future BRT corridors ( 21 ). However, Winnipeg has yet to implement additional corridors of its proposed rapid transit network. Discussions and planning for the second corridor—the Eastern Transitway—have recently stalled with full BRT possibly scrapped in favor of bus-only lanes ( 22 ). Although Winnipeg Transit’s most recent Transportation Master Plan explicitly details new transit corridors, no firm dates for their implementation are noted—except that a preliminary design for rapid transit in the downtown portion of corridors should be completed by 2027 ( 19 ). The expense and time associated with implementing heavy BRT infrastructure highlight the importance of accurately understanding the potential of limited networks for ridership change.

Methods

The aim of this research is to measure bus ridership changes resulting from the opening of Winnipeg’s Southwest Transitway, Phase 1. Following Kim et al., our research strategy aimed to resemble a randomized or quasi-experimental study ( 23 ). The importance of undertaking a quasi-experimental study is that it assists in determining any causal impact of a treatment—that is, the opening of BRT stations. We established stop clusters that served as our units of analysis to be compared over different time periods with AHC. We next used PSM to create a matched sample of clusters to control for biases and establish similar likelihoods of treatments. We then use t-tests with Cohen’s d to analyze BRT’s possible causal effects on transit ridership. With AHC and PSM, we present a CLM approach as a method to investigate ridership changes resulting from new transit infrastructure.

Agglomerative Hierarchical Clustering

Using DRM to measure ridership changes was not feasible, as several stops did not exist before BRT implementation. As such, we performed AHC to establish stop clusters to measure change. The aim of AHC is to group together like entities, that is, bus stops. AHC starts by treating each bus stop as a cluster. Then, clusters are grouped together based on their similarity or closeness (6, 24). Their similarity in our case was their (Euclidean) distance to each other. Clustering continues until a certain threshold is reached. We used a threshold of 400 m—that is, bus stops are grouped together as clusters until there are no other independent clusters within a 400-m distance of each other. We used a threshold of 400 m based on Dunn Index (DI) values, in which higher values are better in evaluating clustering. DI quantifies the level of cluster compactness (see Ncir et al. [ 25 ] for an elaboration of DI). We tested several AHC techniques using different clustering methods (complete, ward.D2, and average) and thresholds (400 to 800 m at 50-m intervals). Using the average clustering method with a 400-m distance threshold provided the highest DI value (0.103 with the next highest being 0.0826). This AHC model also provided practical clusters in which clusters with BRT stops also included 2011 and 2019 bus stops—allowing for useful comparisons from before and after BRT implementation. We used the average linking function that links clusters by taking the average distance between clusters.

Propensity Score Matching

PSM helps to control for confounding, which is the distortion of the relationship between the independent (i.e., BRT stops) and dependent (i.e., transit ridership) variables because of a third (or more) variable(s) (i.e., socioeconomic and land-use characteristics) (16, 23). Thus, PSM establishes a comparison group of control units with statistically similar characteristics to the actually treated group. We conducted two sets of PSMs: one using a BRT stop dummy variable (BRT1) and the other using the BRT connection dummy variable (BRTx) as treatment variables. We included both for two reasons. First, one benefit of BRT is its flexibility: buses can use the fixed BRT corridor but also go onto roads without exclusive bus ROWs. As such, stops with routes connecting to the BRT corridor may also be affected by the upgraded transit service. Second, considering the small number of BRT stops opened in 2012 (four), the exact BRT stops may have limited direct impact on Winnipeg’s overall bus ridership. Nevertheless, we anticipated that BRT, whether directly (with BRT stops on the corridor) or indirectly (with bus stops on routes connecting to the BRT corridor), will cause an increase in transit ridership resulting from the use of the exclusive BRT corridor, which research suggests increases bus ridership (2, 7, 12).

For both, we developed binary logit models to determine the probability of a cluster having a BRT1 or BRTx stop. We included all available variables that could influence transit ridership and BRT as covariates. The residuals of the logit model are essentially the propensity scores that show the likelihood of a cluster having a BRT1 or BRTx stop. Clusters with similar scores are thus matched. We used the matchit function in R to undertake PSM ( 26 ). However, when using the BRT1 dummy variable, we also included a ratio command. Different methods usually take a 1:1 ratio in which every treatment cluster matches with a single control cluster. However, given the low number of treated clusters (four) and after consideration and multiple attempts to find a model with an acceptable covariate balance, we chose an 8:1 ratio in which every treatment cluster is matched with up to eight control clusters. This approach can increase bias, but also has better precision in matching like clusters. In determining the correct PSM models to use, we applied different rules in determining our final PSM models. Specifically, we tested different PSMs using A) glm and Mahalanobis distances with B) either nearest neighbor or optimal methods along with C) different matching ratios such as 1:1, 1:4, 1:8, or 1:12 (Tables S1 to S4 in Supplemental Material A display balance statistics for the different PSM models). For the BRT1 dummy variable, we ultimately used a PSM with a glm distance, optimal method, and a 1:8 matching ratio. For the BRTx dummy variable, we ultimately used a PSM with a glm distance, nearest neighbor method, and 1:1 matching. Additionally, with the BRTx PSM, we discarded both treated and control variables outside the propensity score range of support (see Ho et al. [ 26 ] for further details in relation to R’s matchit function).

We chose such PSM parameters due to the balance statistics of the different PSM models. Of key importance for PSM is establishing balance among the covariates. Standard mean difference (SMD) is a common and widely accepted statistic to report balance and ultimately indicate appropriate matching (27, 28). SMD shows the statistical similarity or difference between control and treated groups of the matched data. Research suggests a cutoff of 0.25 for the SMD ( 29 ). Others accept a stricter cutoff of less than 0.10 as a preferable balance ( 30 ). Yet, as Harder et al. suggest, researchers should assess their specific study to determine the best cutoff ( 29 ). For our study, we used 0.25 as a cutoff in light of our data and the lower sample size when using BRT as the treatment variable. The variance ratio (VR) is a secondary statistic to assess balance, with research suggesting that balance has been achieved when the VR is below 2.00 (28, 31). For BRT1 PSMs, only a PSM using glm distance, optimal method, and a 1:8 matching ratio command provided SMDs for all covariates below the 0.25 threshold. For BRTx PSMs, both PSMs using glm distance, a 1:1 matching ratio, and either nearest neighbor or optimal method commands provided SMDs for all covariates below the 0.25 threshold. However, the PSM using the nearest neighbor method provided slightly lower SMDs when using the optimal method.

T-tests and Cohen’s d

Using the matched samples, we next performed Welch’s two-sample t-test to compare the difference in the means of transit ridership between the treated and control clusters. T-tests show whether and how the average ridership differs between treated and control groups. If the t-tests were significant, we next performed Cohen’s d analyses to determine the effect size. Cohen’s d can help determine the relative difference in means between BRT and non-BRT stations allowing for comparisons across studies. Researchers note that effect size interpretations are relative (32, 33). As such, Lakens suggests that effect sizes can be interpreted as percentages of the standard deviation ( 33 ). For instance, a d of 0.5 can be interpreted as the difference in means equaling half a standard deviation.

Data and Variables

We list the variables used for AHC, PSM, t-tests, and Cohen’s d in Table 2. Unless otherwise noted, we used 2011 data for the covariates. Stop clusters served as the unit of analysis. AHC combined stops into distinct clusters. We then determined the centroid of each cluster and established a 400 m Euclidean buffer around each cluster. Thus, the stop clusters were a 400 m radius around the stop cluster centroid. We chose 400 m for two reasons: based on our AHC distance threshold (as indicated above) and as an average walking distance for riders to reach bus stops.

Table 2.

Variable Descriptions

Variable	Description	Source
Riders	Average cluster ridership change from 2011 to 2019 (natural log)	Winnipeg Transit, calculated by authors
RidersBPH	Average cluster ridership change by buses per hour (BPH) from 2011 to 2019 (natural log)
BRT1	Dummy variable: 1 if cluster contains a BRT station, 0 otherwise
BRTx	Dummy variable: 1 if cluster contains a stop with a route connecting to the Southwest Transitway, Phase 1, 0 otherwise
Residential area	Total residential area in cluster (km²)	Assessment and Taxation Department ( 34 )
Commercial area	Total commercial area in cluster (km²)
Institutional area	Total institutional area in cluster (km²)
Population density	Percentage: total population/cluster area	von Bergmann et al. ( 35 )
Income	Median household income (natural log) (in 2019 Canadian dollars)
Employed	Percentage of employed population (# employed/# in labor force)
Public transit commuters	Percentage of public transit commuters: public transit/total employed population aged 15 years and over
Rental occupancy	Percentage of renter-occupied housing
New housing	Percentage of housing built from 2006 to 2011
Rent	Median monthly shelter costs for rented dwellings (natural log) (in 2019 Canadian dollars)
Housing value	Median housing value (natural log) (in 2019 Canadian dollars)
Transit pass-ups	Average weekday bus pass-ups	Winnipeg Transit, calculated by authors
Buses per hour	Average buses per hour
Route connections	Average number of bus routes by bus stop
Active transit	Active transportation density: total length of active transit infrastructure (in km)/cluster area	Public Works, City of Winnipeg ( 36 )

Note: BRT = bus rapid transit.

We used two treatment variables for the t-tests and as the independent variables for our PSMs: BRT1 and BRTx. PSM covariates were selected based on literature highlighting the factors affecting bus ridership. For total population, total households, employed populations, rental occupancy, new households, and public transit ridership variables we used areal interpolation based on the cluster area—multiplying the cluster area within a census tract by the census tract’s total population, for instance. We then calculated cluster percentages of such (e.g., rental occupancy rates) using the areal modified figures. If clusters overlapped different census tracts, we added the area-weighted figures for all census tract portions within a cluster before calculating the percentages. For the remaining variables, we used the census tract figure in which the cluster resides. If a cluster overlapped multiple census tracts, we averaged these values for the given clusters.

Our main variable of interest is bus ridership change. We calculated this in two ways, based on (1) average bus ridership and (2) bus ridership by buses per hour (BPH). We used BPH to help control for possible endogeneity. Places with high ridership require high service levels, but high service levels may already indicate high ridership or high ridership demands, which could be reasons why certain corridors receive transit upgrades (8, 12, 14). Thus, controlling for BPH can help overcome this endogeneity issue (8, 12). For both, we first normalized ridership data per stop using the natural logarithm as transit ridership data are often skewed toward zero (13, 37). Based on AHC, we then calculated the average ridership and average ridership by BPH of stops within a given cluster for 2011 and 2019, separately.

Data Sources

Ridership data were derived from Winnipeg Transit (via direct email). This data included stop level on-boardings, off-boardings, and total (on- and off-boardings combined) ridership, the routes serving each stop, and the average number of BPH for each bus stop across Winnipeg from 2008 to 2020 for the fall (September to December). We averaged BPH per day per stop. Winnipeg Transit also provides publicly available transit pass-up data ( 38 ), which show how many times buses pass up potential riders owing to the bus being too full. We used the cancensus package in R to collect Canadian census data ( 35 ). Additional spatial data were derived from the City of Winnipeg and through its Assessment and Taxation and Public Works departments (city boundary, parcels, and active transportation shapefiles, respectively), from Manitoba Land Initiative (water boundaries), and Statistics Canada (census tracts) (34, 36, 39 –41). We used ArcGIS, R Studio, and specific R packages (clValid, cluster, geosphere, Matchit, rbounds, rgeos, stats) to transform variables and conduct our analyses (26, 42 –48).

Results

Tables 3 and 4 display the descriptive statistics for the variables used in this study. (We also include descriptive statistics for the variables before clustering in Table S5 in Supplemental Material.) A few interesting results are worth noting. The number of average, daily riders—overall and in relation to BPH—decreased slightly from 2011 to 2019. However, overall clusters without a BRT1 stop showed a slight increase in average ridership (0.42) but a decrease in ridership when considering BPH (−0.83) (Table 3). BRT1 clusters exhibited a relatively large decrease in ridership (−140) (Table 3). For the BRT1 PSM matched clusters (Table 4), all clusters experienced a decrease (−4.2) in average ridership whereas the non-BRT clusters experienced an increase (12.74). Meanwhile, BRTx PSM matched clusters overall and treated clusters experienced increases (2.14, 6.35, respectively), but control clusters experienced a decrease in average ridership (−2.07) (Table 4).

Table 3.

Descriptive Statistics: All Clusters

	All clusters
	All clusters (N = 700)		BRT1 (N = 4)		Non-BRT1 (N = 696)		BRTx (N = 127)		Non-BRTx (N = 573)
Statistic	Mean	SD	Mean	SD	Mean	SD	Mean	SD	Mean	SD
Riders19	62.37	164.98	201.51	181.22	61.57	164.69	133.56	351.05	46.59	68.85
Riders11	62.75	158.34	341.22	533.79	61.15	153.42	121.52	334.86	49.72	70.73
Riders change	−0.38	38.96	−139.71	352.85	0.42	29.6	12.04	85.82	−3.13	13.79
RidersBPH19	13.57	15.53	22.15	18.25	13.52	15.52	18.68	28.51	12.44	10.43
RidersBPH11	14.47	15.55	33.56	45.25	14.36	15.24	17.59	26.67	13.77	11.66
RidersBPH change	−0.89	4.93	−11.41	27.25	−0.83	4.54	1.09	8.71	−1.33	3.46
Residential area	0.21	0.1	0.14	0.04	0.21	0.1	0.2	0.1	0.21	0.1
Commercial area	0.04	0.06	0.07	0.03	0.04	0.06	0.05	0.07	0.04	0.06
Institutional area	0.04	0.05	0.02	0.01	0.04	0.05	0.05	0.07	0.03	0.04
Population density	2,352	1,401	4,213	1,648	2,341	1,394	2,016	1,551	2,427	1,356
Income (Canadian dollars)	69,820	21,619	50,860	7,959	69,929	21,627	80,173	24,398	67,525	20,272
Employed	0.94	0.02	0.93	0.01	0.94	0.02	0.94	0.02	0.94	0.03
Public transit commuters	0.14	0.06	0.21	0.04	0.14	0.06	0.14	0.06	0.14	0.06
Rental occupancy	0.29	0.2	0.58	0.24	0.29	0.2	0.28	0.24	0.29	0.19
New housing	0.06	0.12	0.02	0.01	0.06	0.12	0.07	0.12	0.06	0.11
Rent	889	281	709	49	890	282	1013	386	861	244
Housing value (Canadian dollars)	261,282	65,740	227,293	13,870	261,477	65,871	304,447	66,994	251,715	61,540
Transit pass-ups	0.01	0.05	0.05	0.06	0.01	0.05	0.03	0.1	0.01	0.02
Buses per hour 2019	3.25	2.02	7.11	1.59	3.23	2	4.43	3.25	2.99	1.51
Buses per hour 2011	3.12	1.86	6.24	3.45	3.1	1.84	4.03	2.91	2.92	1.47
Active transit	1.18	1.37	3.46	1.73	1.17	1.36	1.74	1.73	1.06	1.25
Route connections	1.88	1.08	2.93	1.79	1.87	1.08	2.69	1.53	1.7	0.86

Note: SD = standard deviation.

Table 4.

Descriptive Statistics: Matched Clusters

	BRT1 matched clusters						BRTx matched clusters
	All clusters (N = 36)		BRT1 (N = 4)		Non-BRT1 (N = 32)		All clusters (N = 242)		BRTx (N = 121)		Non-BRTx (N = 121)
Statistic	Mean	SD	Mean	SD	Mean	SD	Mean	SD	Mean	SD	Mean	SD
Riders19	184.78	285.71	201.51	181.22	182.69	298.24	69.4	107.61	76.57	100.05	62.22	114.63
Riders11	188.98	311.11	341.22	533.79	169.95	279.89	67.26	120.23	70.23	128.33	64.29	112.01
Riders change	−4.2	124.92	−139.71	352.85	12.74	53.88	2.14	53.07	6.35	71.83	−2.07	21.44
RidersBPH19	18.47	14.36	22.15	18.25	18.01	14.1	13.64	11.81	15.2	11.8	12.07	11.65
RidersBPH11	20.67	20.65	33.56	45.25	19.06	16.1	13.6	12.42	14.29	12.64	12.9	12.21
RidersBPH change	−2.2	9.63	−11.41	27.25	−1.05	4.53	0.04	6.21	0.91	8.05	−0.83	3.35
Residential area	0.14	0.09	0.14	0.04	0.14	0.1	0.21	0.09	0.21	0.09	0.21	0.1
Commercial area	0.08	0.08	0.07	0.03	0.08	0.08	0.05	0.07	0.05	0.06	0.05	0.08
Institutional area	0.01	0.01	0.02	0.01	0.01	0.01	0.04	0.05	0.04	0.05	0.03	0.04
Population density	4,287	2,163	4,213	1,648	4,297	2,241	2,049	1,373	1,962	1,490	2,136	1,245
Income (Canadian dollars)	52,233	17,755	50,860	7,959	52,405	18,695	82,088	23,398	81,782	23,605	82,394	23,283
Employed	0.93	0.02	0.93	0.01	0.93	0.02	0.95	0.02	0.94	0.02	0.95	0.02
Public transit commuters	0.21	0.08	0.21	0.04	0.21	0.08	0.13	0.05	0.13	0.05	0.13	0.05
Rental occupancy	0.58	0.29	0.58	0.24	0.58	0.29	0.25	0.21	0.26	0.23	0.24	0.2
New housing	0.01	0.02	0.02	0.01	0.01	0.02	0.07	0.11	0.07	0.13	0.06	0.09
Rent	722	125	709	49	724	132	1012	350	1023	391	1001	306
Housing value (Canadian dollars)	235,264	50,731	227,293	13,870	236,261	53,645	305,069	64,789	307,189	65,654	302,949	64,116
Transit pass-ups	0.04	0.05	0.05	0.06	0.04	0.05	0.01	0.03	0.02	0.03	0.01	0.03
Buses per hour 2019	6.04	3.9	7.11	1.59	5.91	4.1	3.82	2.25	3.97	2.39	3.66	2.09
Buses per hour 2011	5.51	3.7	6.24	3.45	5.42	3.77	3.61	2.13	3.64	2.42	3.58	2.01
Active transit	3.11	2.38	3.46	1.73	3.07	2.47	1.48	1.45	1.58	1.58	1.38	1.32
Route connections	2.79	2.09	2.93	1.79	2.78	2.15	2.42	1.17	2.52	1.2	2.32	1.14

Note: SD = standard deviation.

BRT1 clusters showed less built-up space compared with non-BRT1 clusters, as seen with residential- and institutional areas (Table 3). However, BRT1 clusters had higher population densities and rental occupancy rates than non-BRT1 clusters (Table 3). In 2011, median incomes and housing values were approximately Cdn$20,000 and $34,000 less, respectively, in BRT1 clusters compared with non-BRT1 clusters. On the other hand, median incomes and average housing values were approximately $12,000 and $45,000 more, respectively, for matched BRTx clusters than for the averages throughout Winnipeg. Additionally, the BRTx clusters had just half the population density of the BRT1 clusters. Furthermore, BRTx clusters had a lower population density than the Winnipeg average (2,016 versus 2,352 people per square kilometer or ppsq) and less than half the population density of the BRT1 clusters (4,213 ppsq).

Agglomerative Hierarchical Clustering

Figure 2 displays all the Winnipeg stops that existed in both 2011 and 2019 (Figure 2a) and the stop clusters after undertaking AHC (Figure 2b). We linked together all 2011 and 2019 stops, which provided 10,222 original stops. AHC established 709 stop clusters. We omitted clusters with only 2011 or 2019 stops, clusters with zero ridership for 2011 or 2019, and clusters with no applicable socioeconomic data. Thus, 700 total stop clusters remained. Each cluster had both 2011 and 2019 data, allowing us to measure change between the two time periods.

Figure 2.

Winnipeg total stops and clusters.

Propensity Score Matching

Figures 3 and 4 display all matched clusters (3a and 4a) and only treated clusters (3b and 4b) resulting from PSM using BRT1 and BRTx as the treatment variables, respectively. Tables 5 and 6 present the PSM results, or the balance statistics, with BRT1 and BRTx as the dependent or treatment variables in the binary logit models, respectively. For the BRT1 PSM, 32 control clusters were matched with the 4 treated clusters. All covariates showed improvements in SMD when matching and all covariates were under the 0.25 threshold. For the BRTx PSM, rental occupancy did not improve (−123.3), but all other covariates’ SMDs did improve. BRTx PSM showed relatively better balance among the covariates compared with the BRT1 PSM. This was possibly owing to the larger sample size: a total of 36 for the BRT1 PSM and 242 total for the BRTx PSM. With the exception of new housing for the BRTx PSM, matched data VRs were under 2.0, thus providing additional support for covariate balance. Overall, the covariate balance summaries indicated balance among the covariates and, thus, acceptable PSM and stop clustering.

Table 5.

BRT1 PSM Balance Statistics

	Summary of balance for unmatched data				Summary of balance for matched data
	Means treated	Means control	SMD	VR	Means treated	Means control	SMD	VR	SMD % balance improvement
Distance	0.12	0.01	0.74	39.30	0.12	0.08	0.28	3.00	61.5
Income	10.81	11.09	−1.83	0.24	10.81	10.79	0.10	0.20	94.4
Residential area	0.14	0.21	−1.96	0.14	0.14	0.14	−0.18	0.15	90.7
Commercial area	0.07	0.04	1.12	0.22	0.07	0.08	−0.24	0.13	78.6
Institutional area	0.02	0.04	−2.45	0.03	0.02	0.01	0.15	0.53	94
Employed	0.93	0.94	−1.20	0.10	0.93	0.93	−0.14	0.11	88.3
Population density	4213.09	2341.42	1.14	1.40	4213.09	4296.54	−0.05	0.54	95.5
Rental occupancy	0.58	0.29	1.20	1.45	0.58	0.58	−0.02	0.67	98.5
New housing	0.02	0.06	−4.50	0.01	0.02	0.01	0.24	0.19	94.8
Rent	6.55	6.75	−3.00	0.06	6.55	6.56	−0.04	0.13	98.7
Housing value	12.33	12.43	−1.94	0.05	12.33	12.33	−0.11	0.05	94.4
Transit pass-ups	0.05	0.01	0.61	1.81	0.05	0.04	0.20	1.46	66.3
Buses per hour	6.24	3.10	0.91	3.52	6.24	5.42	0.24	0.84	73.9
Public transit commuters	0.21	0.14	1.44	0.57	0.21	0.21	−0.05	0.28	96.7
Active transit	3.46	1.17	1.32	1.62	3.46	3.07	0.23	0.49	82.8
Route connections	2.93	1.87	0.59	2.77	2.93	2.78	0.08	0.70	85.6
Sample sizes:
	Control	Treated
All	696	4
Matched	32	4
Unmatched	664	0
Discarded	0	0

Note: PSM = propensity score matching; SMD = standard mean difference; VR = variance ratio.

Table 6.

BRTx PSM Balance Statistics

	Summary of balance for unmatched data				Summary of balance for matched data
	Means treated	Means control	SMD	VR	Means treated	Means control	SMD	VR	SMD % balance improvement
Distance	0.48	0.11	1.24	3.97	0.46	0.34	0.41	2.38	66.9
Income	11.23	11.06	0.50	1.11	11.25	11.26	−0.02	0.82	95.1
Residential area	0.20	0.21	−0.08	0.97	0.21	0.21	−0.02	0.93	78.3
Commercial area	0.05	0.04	0.21	1.20	0.05	0.05	−0.04	0.62	81.6
Institutional area	0.05	0.03	0.16	3.68	0.04	0.03	0.08	1.70	49.9
Employed	0.94	0.94	0.20	0.45	0.94	0.95	−0.17	0.47	13.9
Population density	2016.05	2426.61	−0.26	1.31	1961.79	2135.86	−0.11	1.43	57.6
Rental occupancy	0.28	0.30	−0.04	1.62	0.26	0.24	0.10	1.26	−123.3
New housing	0.07	0.06	0.07	1.14	0.07	0.06	0.03	2.01	55.6
Rent	6.89	6.72	0.56	1.41	6.90	6.85	0.15	1.13	72.7
Housing value	12.59	12.40	0.85	0.80	12.60	12.59	0.07	0.89	91.5
Transit pass-ups	0.03	0.01	0.26	22.73	0.02	0.01	0.05	0.83	79.1
Buses per hour	4.03	2.92	0.38	3.92	3.64	3.58	0.02	1.24	95.1
Public transit commuters	0.14	0.14	−0.12	1.30	0.13	0.13	0.07	1.00	46.5
Active transit	1.74	1.06	0.39	1.93	1.58	1.38	0.12	1.44	69.9
Route connections	2.69	1.70	0.64	3.19	2.52	2.32	0.13	1.12	79.2
Sample sizes:
	Control	Treated
All	573	127
Matched	121	121
Unmatched	423	0
Discarded	29	6

Note: PSM = propensity score matching; SMD = standard mean difference; VR = variance ratio.

Figure 3.

BRT1 matched clusters: (a) all and (b) treated only.

Figure 4.

BRTx matched clusters: (a) all and (b) treated only.

We additionally performed sensitivity analyses using the psens function of the rbounds package in R ( 47 ). Here, we tested how sensitive the results of PSM may be from biases or factors not identified with the observed covariates ( 49 ). These sensitivity analyses provide a sensitivity parameter or Gamma (Γ) that is the odds of hidden bias (see 47 ). Essentially, the closer the Gamma to 1, the greater the sensitivity to unmeasured confounding, or the greater the bias of unobserved factors, influencing the results. As Rosenbaum notes though, “studies vary markedly in their sensitivity to bias” ( 50 , p. 4). Table 7 displays results of our sensitivity analyses for both BRT1 and BRTx (using Gamma parameter increments of 0.1 and condensed in the table for brevity). The BRT1 results suggest that the lower bound p-value changed from not significant (0.102 at Γ = 2.5) to significant (0.095) when Γ = 2.6. Since we received balance with the observed covariates, a cluster may be 2.6 times more likely (when using a significance of 0.10) to receive treatment (i.e., a BRT stop) than not receiving treatment because of unobserved factors. For BRTx, the odds were 1.7 times higher where the upper bound p-value changed from not significant (0.072 at Γ = 1.6) to significant (0.118) when Γ = 1.7. Following explanations by Keele, the odds of one cluster receiving a BRT stop (BRT1) were 2.6 times higher, or a BRT connection (BRTx) 1.7 times higher because of different unobserved covariate values even with identically matched covariates ( 47 ). These sensitivity analyses indicated that although unobserved factors may play a role in the impact of BRT on bus ridership, a larger bias explained away the impacts of BRT on ridership for BRT1 compared with BRTx.

Table 7.

Sensitivity Analyses

BRT1			BRTx
Gamma	Lower bound	Upper bound	Gamma	Lower bound	Upper bound
1	0.358	0.358	1	0.00	0.000
2	0.151	0.602	1.1	0.00	0.001
2.1	0.139	0.619	1.2	0.00	0.003
2.2	0.129	0.635	1.3	0.00	0.008
2.3	0.119	0.650	1.4	0.00	0.019
2.4	0.110	0.664	1.5	0.00	0.039
2.5	0.102	0.678	1.6	0.00	0.072
2.6	0.095	0.691	1.7	0.00	0.118
2.7	0.088	0.703	1.8	0.00	0.177
2.8	0.081	0.715	1.9	0.00	0.248
2.9	0.075	0.726	2	0.00	0.327
3	0.070	0.737	3	0.00	0.919
4	0.033	0.819	4	0.00	0.997
5	0.017	0.874	5	0.00	1.000

Note: significance of 0.10

T-tests

T-tests measured the difference in the means of average bus ridership change between clusters with a BRT stop or connection and those without, from the matched samples. We ran two t-tests for each PSM sample using the (natural log of) change in average ridership per cluster (Riders) and (natural log of) change in average ridership by BPH per cluster (RidersBPH) as the outcome variables. Considering that we were running t-tests on our matched dataset, the difference between the means of the treated (i.e., clusters with a BRT station or BRT stop connection) and control clusters represented the estimated treatment effect within the matched sample. In this regard, the treatment effect would be equivalent to the average treatment effect on the treated rather than the average treatment effect for bus ridership. This was because the matched sample only represented the clusters most likely to receive treatment (i.e., a BRT station or bus stop with a route connecting to a BRT station) and not all of Winnipeg’s clusters.

The results with BRT1 as the treatment variable showed no statistically significant difference in means between the clusters that had a BRT stop and those that did not (Table 8). Here, the null hypothesis is that the true difference in means is not equal to zero—that is, there is a difference between the means of the treated and control groups. However, the p-values were not significant. Thus, BRT did not cause a significant difference in ridership change.

Table 8.

BRT1 T-Tests

Outcome variable	Mean treated	Mean control	ETE	t-Statistic	df	Cohen’s d
Riders	0.400	0.096	0.304	−0.446	3.098	NA
RidersBPH	−0.007	−0.025	0.018	−0.069	3.275	NA

Note : ETE = estimated treatment effect; NA = not available.

p < 0.1; ^**p < 0.05; ^***p < 0.01.

Results for the BRTx matched sample suggest a statistically significant difference in means between the treated and control groups (Table 9). Here, both change in average ridership (Riders) and change in average ridership by buses per hour or BPH (RidersBPH) were statistically significant. In both cases, if a cluster has a stop with a route serving stops along the BRT corridor, then these clusters are likely to have a greater amount of ridership than similar stops without BRT connections. Thus, having a BRT connection causes an increase in transit ridership.

Table 9.

BRTx T-tests

Outcome variable	Mean treated	Mean control	ETE	t-Statistic	df	Cohen’s d
Riders	0.171	−0.052	0.223	−3.476^***	220.09	0.447
RidersBPH	0.078	−0.063	0.142	−2.458^**	238.36	0.316

Note : ETE = estimated treatment effect.

p < 0.1; ^**p < 0.05; ^***p < 0.01.

Since the t-tests proved significant for the BRTx models, we performed Cohen’s d analyses to determine the magnitude of the difference. We found a Cohen’s d statistic of 0.447 for Riders and of 0.316 for RidersBPH. Clusters with a BRT connection had almost half a standard deviation increase in change in average transit ridership compared with the control clusters. Clusters with a BRT connection had about a third of a standard deviation increase in change in average transit ridership by BPH compared with the control clusters.

Discussion and Conclusions

As cities continue to implement BRT infrastructure with the hopes of increasing ridership and changing mode shares, it is increasingly important to understand the potential of these systems to achieve these goals. Although BRT can be successful in increasing ridership in certain conditions, systems may be implemented in areas with different types of land-use patterns and limited connectivity. Using AHC, PSM, and t-tests with Cohen’s d, we aimed to understand how the implementation of a single BRT corridor influenced transit ridership change across Winnipeg. We found no statistical evidence that BRT caused increased transit ridership when analyzing only the stop clusters with stations along the BRT corridor. However, the results indicated that stop clusters with routes connecting to the BRT corridor did experience an increase in transit ridership.

We identified two key implications from these results. First, a limited number of BRT stops along a single corridor may not be enough to increase or even shift transit ridership. Instead, planners should consider a larger BRT network and implementing the network in as short a timeframe as possible for any significant effect on ridership. Put simply, four stops in previously undeveloped areas may just not be enough to influence or cause changes in bus ridership. Given the political and environmental imperatives to demonstrate the effectiveness of BRT, this may mean favoring lighter infrastructure BRT that could be implemented more flexibly and allowed to more fully develop. Second, our results underscored the importance of embracing the flexibility of BRT and BRT’s positive impact beyond just the BRT corridor. Existing research often touts one of BRT’s benefits as allowing buses to operate on the fixed transitway and exit it to serve lower-density or spread out areas (51, 52). In this regard, if BRT stations open in previously under-served areas, having a wide array of routes connecting to the BRT infrastructure could assist in increasing overall transit ridership even when BRT stations themselves do not do so. Embracing BRT’s flexibility could also serve to show BRT’s overall ridership value while giving station areas time to more fully develop.

Given the significant funding, time commitment, and permanence of BRT, it is important to understand how ridership could change in areas without previous service. This research is key in introducing a model—CLM—that can account for ridership change in areas where exact transit stops previously did not exist, an important consideration as cost issues and land acquisition may require municipalities to build BRT in spaces without existing infrastructure. CLM can help show the impacts of the new infrastructure on transit ridership, or rather transit interest, within an area served by the new infrastructure. CLM does have its limits, though. Specifically, CLM may not be feasible in areas without transit stops within a specified distance—for example, greater than 400 or 800 m—of a new rapid transit station, as no transit stops would exist to form a transit cluster. Nevertheless, for Winnipeg and similar areas, CLM could provide a useful method to measure ridership change stemming from new BRT.

Although our results show that BRT—in the form of BRT stop connections—did cause an increase in transit ridership, there were a few study limitations. First, our results do not necessarily mean that BRT increases overall transit ridership. Rather, BRT increases ridership in comparison to similar stop cluster areas without BRT access. Indeed, total ridership stayed roughly the same for all stops examined in the study between 2011 (437,375 total riders) and 2019 (437,229 total riders) and the average cluster ridership actually decreased slightly (−0.38) from 2011 to 2019. As such, and similar to what Kim et al. ( 23 ) suggest, riders may have switched from non-BRT connected routes to BRT connected routes. Thus, even though having a BRT connection caused an increase in ridership in those cluster areas, Winnipeg overall did not see a significant increase in ridership, suggesting that BRT is causing a shift rather than a gain in transit ridership—which may point to transit-oriented development potential for BRT. Another limitation is the low number of BRT stops that actually opened during our period of study. Winnipeg only opened four BRT stations in 2012. Again, opening four stops in a previously undeveloped area may just not be enough to cause a significant increase in transit ridership, yet it is important for planners and policy makers to be able to understand preliminary impacts before further investments. Methodologically, PSMs may perform better with more treatment variables or greater sample sizes. Still, research has found that PSM can predict correct treatment effect estimations with low sample sizes ( 53 ). Our PSM models also revealed proper balance among the covariates—a key indicator of matching. Nevertheless, our method could be used with larger sample sizes to provide more robust results in the future. Additionally, data on vehicle ownership were not publicly available at a geographic scale smaller than the city level for Winnipeg. Research could include this variable in future analyses.

Our results suggested a shift in ridership to BRT connected routes, but not necessarily at the actual BRT infrastructure and not an increase in overall ridership. As existing research proposes, 10 years may not be enough time to truly measure the value of BRT in relation to investments that could trigger increased ridership or vice versa, especially with a limited number of BRT stations (10, 54). In this regard, the value of Winnipeg’s BRT resides in its flexibility or the ability of buses to serve both the BRT corridor and standard routes, at least in the short term. Despite our findings and the importance of our research overall, the question of BRT’s overall added value remains. Like other forms of rapid transit, BRT’s value may not just reside in public transit use. Research focusing on BRT station area property values or land-use change could provide additional support for the value of BRT.

Supplemental Material

sj-docx-1-trr-10.1177_03611981221085531 – Supplemental material for The Impact of a Single Bus Rapid Transit Corridor on Transit Ridership: The Winnipeg Example

Supplemental material, sj-docx-1-trr-10.1177_03611981221085531 for The Impact of a Single Bus Rapid Transit Corridor on Transit Ridership: The Winnipeg Example by Dwayne Marshall Baker and Orly Linovski in Transportation Research Record

Footnotes

Acknowledgements

The authors thank Winnipeg Transit for providing the transit data and Xiaoyu Li of Winnipeg Transit for assistance with interpreting the data.

Author Contributions

The authors confirm contribution to the paper as follows: study conception and design: D. Marshall Baker; data collection: D. Marshall Baker; analysis and interpretation of results: D. Marshall Baker; draft manuscript preparation: D. Marshall Baker, O. Linovski. All authors reviewed the results and approved the final version of the manuscript.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Dwayne Marshall Baker

Supplemental Material

Supplemental material for this article is available online.

References

Hoang-Tung

Kato

Huy

T. T.

Le Binh

Duy

Impacts of the Introduction of Bus Rapid Transit on Travel Behaviors of Commuters in Hanoi, Vietnam: A Quasi-Experimental Approach. Case Studies on Transport Policy, Vol. 9, No. 1, 2021, pp. 95–102.

Stewart

A. F.

Attanucci

J. P.

Wilson

N. M. H.

Ridership Response to Incremental Bus Rapid Transit Upgrades in North America: Demographic and Network Effects. Transportation Research Record: Journal of Transportation Research Board, 2015. 2538: 37–43.

Stewart

O. T.

Mouden

A. V.

Saelens

B. E.

The Causal Effect of Bus Rapid Transit on Changes in Transit Ridership. Journal of Public Transportation, Vol. 20, No. 1, 2017, pp. 91–103.

Guerra

Mexico City’s Suburban Land Use and Transit Connection: The Effects of the Line B Metro Expansion. Transport Policy, Vol. 32, 2014, pp. 105–114.

Lee

Boarnet

Houston

Nixon

Spears

Changes in Service and Associated Ridership Impacts Near a New Light Rail Transit Line. Sustainability, Vol. 9, No. 10, 2017, p. 1827. https://doi.org/10.3390/su9101827.

Brands

Dixit

van Oort

Impact of a New Metro Line in Amsterdam on Ridership, Travel Times, Reliability and Societal Costs and Benefits. European Journal of Transport and Infrastructure Research, Vol. 20, No. 4, 2020, pp. 335–353.

Cervero

Murakami

Miller

Direct Ridership Model of Bus Rapid Transit in Los Angeles County, California. Transportation Research Record: Journal of the Transportation Research Board, 2010. 2145: 1–7.

Kerkman

Martens

Meurs

Factors Influencing Stop-Level Transit Ridership in Arnhem-Nijmegen City Regions, Netherlands. Transportation Research Record: Journal of the Transportation Research Board, 2015. 2537: 23–32.

Mucci

R. A.

Erhardt

G. D.

Evaluating the Ability of Transit Direct Ridership Models to Forecast Medium-Term Ridership Changes: Evidence from San Francisco. Transportation Research Record: Journal of the Transportation Research Board, 2018. 2672: 21–30.

10.

Loukaitou-Sideris

Banerjee

The Blue Line Blues: Why the Vision of Transit Village May Not Materialize Despite Impressive Growth in Transit Ridership. Journal of Urban Design, Vol. 5, No. 2, 2000, pp. 101–125.

11.

Currie

Delbosc

Understanding Bus Rapid Transit Route Ridership Drivers: An Empirical Study of Australian BRT Systems. Transport Policy, Vol. 18, 2001, pp. 755–764.

12.

Chan

Miranda-Moreno

A Station-Level Ridership Model for the Metro Network in Montreal, Quebec. Canadian Journal of Civil Engineering, Vol. 40, No. 3, 2013, pp. 254–262.

13.

Dill

Schlossberg

Meyer

Predicting Transit Ridership at the Stop Level: The Role of Service and Urban Form. Presented at 92nd Annual Meeting of the Transportation Research Board, Washington, D.C., 2013.

14.

Cao

X. J.

Schoner

The Influence of Light Rail Transit on Transit Use: An Exploration of Station Area Residents along the Hiawatha Line in Minneapolis. Transportation Research Part A: Policy and Practice, Vol. 59, 2014, pp. 134–143.

15.

Wetwitoo

Kato

Regional and Local Economic Effects From Proximity of High-Speed Rail Stations in Japan: Difference-in-Differences and Propensity Score Matching Analysis. Transportation Research Record: Journal of Transportation Research Board, 2019. 2673: 323–333.

16.

Rosenbaum

P. R.

Rubin

D. B.

The Central Role of the Propensity Score in Observational Studies for Causal Effects. Biometrika, Vol. 70, No. 1, 1983, pp. 41–55.

17.

Liu

Erdogan

Ducca

F. W.

How to Increase Rail Ridership in Maryland: Direct Ridership Models for Policy Guidance. Journal of Urban Planning and Development, Vol. 142, No. 4, 2016, p. 04016017.

18.

Statistics Canada. Journey to Work (Main Mode of Commuting for the Employed Labour Force), by Census Subdivision. Statistics Canada, 2016.

19.

City of Winnipeg. Transit Master Plan, 2021. https://winnipegtransit.com/assets/2794/Winnipeg_Transit_Master_Plan-reduced.pdf.

20.

Linovski

Baker

D. M.

Manaugh

Equity in Practice? Evaluations of Equity in Planning for Bus Rapid Transit. Transportation Research Part A: Policy and Practice, Vol. 113, 2018, pp. 75–87. https://doi.org/10.1016/j.tra.2018.03.030.

21.

City of Winnipeg. Winnipeg Transportation Master Plan, 2011. https://winnipeg.ca/publicworks/transportation/pdf/transportationMasterPlan/2011-11-01-TTRWinnipegTMP-Final-Report.pdf.

22.

Kives

Winnipeg’s Dream of Rapid Transit Appears to be Dead. CBC News, 2020. https://www.cbc.ca/news/canada/manitoba/winnipeg-rapid-transit-analysis-1.5457171.

23.

Kim

J. Y.

Bartholomew

Ewing

Another One Rides the Bus? The Connections Between Bus Stop Amenities, Bus Ridership, and ADA Paratransit Demand. Transportation Research Part A: Policy and Practice, Vol. 135, 2020, pp. 280–288.

24.

Cats

Wang

Zhao

Identification and Classification of Public Transport Activity Centres in Stockholm using Passenger Flows Data. Journal of Transport Geography, Vol. 48, 2015, pp. 10–22.

25.

Ncir

C. -E.

Abdallah

Waad

Parallel and Scalable Dunn Index for the Validation of Big Data Clusters. Parallel Computing, Vol. 102, 2021, p. 102751.

26.

D. E.

Imai

King

Stuart

E. A.

MatchIt: Nonparametric Processing for Parametric Causal Inference. Journal of Statistical Software, Vol. 42, No. 8, 2011, pp. 1–28. https://www.jstatsoft.org/v42/i08/.

27.

Austin

P. C.

A Critical Appraisal of Propensity-Score Matching in the Medical Literature Between 1996 and 2003. Statistics in Medicine, Vol. 27, No. 12, 2008, pp. 2037–2049.

28.

Zhang

Kim

H. J.

Lonjon

Zhu

Balance Diagnostics after Propensity Score Matching. Annals of Translational Medicine, Vol. 7, No. 1, 2019, p. 16. https://dx-doi-org-s.web.bisu.edu.cn/10.21037%2Fatm.201.

29.

Harder

V. S.

Stuart

E. A.

Anthony

J. C.

Propensity Score Techniques and the Assessment of Measured Covariate Balance to Test Causal Associations in Psychological Research. Psychological Methods, Vol. 15, No. 3, 2010, p. 234.

30.

Normand

S. -L. T.

Landrum

M. B.

Guadagnoli

Ayanian

J. Z.

Ryan

T. J.

Cleary

P. D.

McNeil

B. J.

Validating Recommendations for Coronary Angiography following Acute Myocardial Infarction in the Elderly: A Matched Analysis Using Propensity Scores. Journal of Clinical Epidemiology, Vol. 54, No. 4, 2001, pp. 387–398.

31.

Linden

Samuels

S. J.

Using Balance Statistics to Determine the Optimal Number of Controls in Matching Studies. Journal of Evaluation in Clinical Practice, Vol. 19, No. 5, 2013, pp. 968–975.

32.

Cohen

Statistical Power Analysis for the Behavioral Sciences, 2nd ed. Lawrence Erlbaum Associates, NJ, 1988.

33.

Lakens

. Calculating and Reporting Effect Sizes to Facilitate Cumulative Science: A Practical Primer for T-tests and ANOVAs. Frontiers in Psychology, Vol. 4, 2013, pp. 863. https://doi.org/10.3389/fpsyg.2013.00863.

34.

Assessment and Taxation Department, City of Winnipeg. Map of Assessment Parcels. Shapefile. 2021. https://data.winnipeg.ca/Assessment-Taxation-Corporate/Map-of-Assessment-Parcels/rt7t-3m4m.

35.

von Bergmann

Shkolnik

Jacobs

cancensus: R Package to Access, Retrieve, and Work With Canadian Census Data and Geography. R Package Version 0.4.2, 2021.

36.

Public Works, City of Winnipeg. Active Transportation Network. Shapefile. 2017. https://data.winnipeg.ca/Transportation-Planning-Traffic-Management/Active-Transportation-Network/wm39-dswz.

37.

Chakour

Eluru

Examining the Influence of Stop Level Infrastructure and Built Environment on Bus Ridership in Montreal. Journal of Transport Geography, Vol. 51, 2016, pp. 205–217.

38.

Winnipeg Transit. Transit Pass-ups. Dataset. 2019. https://data.winnipeg.ca/Transit/Transit-Pass-ups/mer2-irmb.

39.

City of Winnipeg. City Limit. Shapefile. 2014. https://data.winnipeg.ca/City-Planning/City-Limit/jx93-sett.

40.

Manitoba Land Initiative. Hydrography. Shapefile. n.d. https://mli2.gov.mb.ca/water_resources/index.html. Accessed April 2021.

41.

Statistics Canada. 2011 Census – Boundary Files: Census Tracts. Shapefile. 2019. https://www12.statcan.gc.ca/census-recensement/2011/geo/bound-limit/bound-limit-2011-eng.cfm.

42.

Environmental Systems Research Institute (ESRI). ArcGIS Release 10.7.1. Redlands, CA, 2019.

43.

R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria, 2021. https://www.R-project.org/.

44.

Brock

Pihur

Datta

clValid: An R Package for Cluster Validation. Journal of Statistical Software, Vol. 25, No. 4, 2008, pp. 1–22. https://www.jstatsoft.org/v25/io4/.

45.

Maechler

Rousseeuw

Struyf

Hubert

Hornik

cluster: Cluster Analysis Basics and Extensions. R Package Version 2.1.1, 2021.

46.

Hijmans

R. J.

Geosphere: Spherical Trigonometry. R Package Version 1.5-10, 2019. https://CRAN.R-project.org/package=geosphere.

47.

Keele

L. J.

rbounds: Perform Rosenbaum Bounds Sensitivity Tests for Matched and Unmatched Data. R Package Version 2.1, 2014. https://CRAN.R-project.org/package=rbounds.

48.

Bivand

rundel

rgeos: Interface to Geometry Engine – Open Source (‘GEOS’). R Package Version 0.5-5, 2020. https://CRAN.R-project.org/package=rgeos.

49.

Olmos

Govindasamy

Propensity Scores: A Practical Introduction Using R. Journal of MultiDisciplinary Evaluation, Vol. 11, No. 25, 2015, pp. 68–88.

50.

Rosenbaum

P. R.

Sensitivity Analysis in Observational Studies. In Encyclopedia of Statistics in Behavioral Science, Vol. 4 ( Everitt

B. S.

Howell

D. C.

, eds.), John Wiley & Sons, Chichester, 2005, pp. 1809–1814.

51.

Cervero

Dai

BRT TOD: Leveraging Transit Oriented Development With Bus Rapid Transit Investments. Transport Policy, Vol. 36, 2014, pp. 127–138.

52.

Hecker

J. E.

Bus Rapid Transit Offers Communities a Flexible Mass Transit Option. Report for the Committee on Banking, Housing, and Urban Affairs, US Senate General Accounting Office, Washington, D.C., 2003.

53.

Pirracchio

Resche-Rigon

Chevret

Evaluation of the Propensity Score Methods for Estimating Marginal Odds Ratios in Case of Small Sample Size. BMC Medical Research Methodology, Vol. 12, No. 1, 2012, pp. 1–10.

54.

Landis

Cervero

Middle Age Sprawl: BART and Urban Development. Access Magazine, Vol. 1, No. 13, 1999, pp. 2–15.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.03 MB