You are on page 1of 16

Transportation Research Part D 50 (2017) 154–169

Contents lists available at ScienceDirect

Transportation Research Part D


journal homepage: www.elsevier.com/locate/trd

A sinusoidal model for seasonal bicycle demand estimation


Nicholas Fournier ⇑, Eleni Christofa, Michael A. Knodler Jr.
Department of Civil and Environmental Engineering, University of Massachusetts, Amherst, MA 01003, United States

a r t i c l e i n f o a b s t r a c t

Article history: As urban populations grow, there is a growing need for efficient and sustainable modes,
such as bicycling. Unfortunately, the lack of bicycle demand data stands as a barrier to
design, planning, and research efforts in bicycle transportation. Estimating bicycle demand
Keywords: is difficult not only due to limited count data, but to the fact that bicyclists are highly
Bicycle demand estimation responsive to a multitude of factors, particularly seasonal weather. Current estimation
Seasonal demand methods capable of accurately adjusting for seasonal demand change often require sub-
Sinusoidal model
stantial data for ongoing calibration. This makes it difficult or impossible to utilize those
Continuous bicycle counters
methods in locations with minimal continuous count data. This research aims to help mit-
igate this challenge by developing an estimation method using sinusoidal model to fit the
typical pattern of seasonal bicycle demand. This sinusoidal model requires only a single
calibration factor to adjust for scale of seasonal demand change and is capable of estimat-
ing monthly average daily bicycle counts ðMADBÞ and average annual daily bicycle counts
ðAADBÞ. This calibration factor can be established using a minimum of two short-term
counts to represent the maximum and minimum monthly MADB in summer and winter.
To develop the model, this research use data from bike-share systems in four cities and
47 permanent bicycle counters in six cities. Although this model is not suitable for loca-
tions with mild or atypical seasons, it successfully models MADB and serves as a useful
alternative or supportive estimation method in locations where minimal demand data
exist.
Published by Elsevier Ltd.

1. Introduction

The lack of bicycle volume data stands as a barrier to transportation agency efforts to evaluate and improve bicycle trans-
portation. The practical applications of extrapolating daily and annual bicycle counts are as boundless as would be for con-
ventional motorized count data, from high level planning efforts down to specific operational improvements. The lack of
demand information not only limits the ability for infrastructure investment decisions, but severely limits research efforts
attempting to improve infrastructure (Crites et al., 2014).
Safety or perceived safety is the greatest barrier preventing more people from cycling in the United States (Sanders,
2013). To resolve this issue, accurate bicycle demand information is necessary to perform crash rate analyses to evaluate
infrastructure and to determine the need for improvement (FHWA, 2011). With regards to bicycle infrastructure, demand
information can not only aid in the direct design of a bicycle facility, but can also offer data-supported reasoning to imple-
ment (or not implement) infrastructure or maintenance procedures (Thompson et al., 2013).

⇑ Corresponding author.
E-mail addresses: nfournie@umass.edu (N. Fournier), christofa@ecs.umass.edu (E. Christofa), mknodler@ecs.umass.edu (M.A. Knodler Jr.).

http://dx.doi.org/10.1016/j.trd.2016.10.021
1361-9209/Published by Elsevier Ltd.
N. Fournier et al. / Transportation Research Part D 50 (2017) 154–169 155

For automobile traffic there is already a standard practice and extensively studied method for estimating average annual
daily traffic (AADT) and average daily traffic (ADT) from only short-term traffic counts using only basic adjustment factors
(Gadda et al., 2007; AASHTO, 2009; Wang et al., 2013). However, this simple method is not directly applicable to bicycle
demand estimation. Where automobile demand is relatively consistent and predictable throughout the year, bicyclists are
substantially more responsive to external factors such as weather, trip purpose, and daylight hours (Ahmed et al., 2013;
Thomas and DeRobertis, 2013; Xing et al., 2010). This high variability in bicyclist requires that these additional factors be
accounted for when estimating bicycle demand.
To further gain an accurate measure of bicycle demand, cities have been counting bicycles manually for short periods dur-
ing the year. However, this leaves data for the vast majority of the year unknown. To resolve this, some cities have utilized
permanent bicycle counting devices to provide continuous counts throughout the year. These continuous counts have made
it possible for an adjustment factor method to be developed for bicycles. Although these adjustment factors are simple and
adaptable, they degrade over time and can be location specific (El Esawey et al., 2013). In order to continually update these
factors and expand the range of applicable locations, a continual feed of count data and additional counters are necessary.
Unfortunately, these permanent counters are expensive and stand as a major cost barrier to cities, particularly small cities
which may already be reluctant to invest in bicycle infrastructure.
As an alternative solution when sufficient continuous counters for conventional methods are not available throughout a
city, a model that exploits the seasonal pattern of bicycle demand would be highly beneficial. The objective of this research
was to provide an estimation method based on a mathematical model to minimize the necessary data. To accomplish this, a
sinusoidal model of seasonal bicycle demand was calibrated and validated using data from a limited number of continuous
counters and bike-share data in multiple cities. Although this model aids in minimizing the necessary data, it ultimately
relies on strong seasonal change in bicycle demand. Thus in locations with mild or atypical seasonal patterns, this model
breaks down and becomes inapplicable. Despite these weaknesses, this model offers potential benefit to locations that do
suffer from high seasonal variation in bicycle demand.

2. Background

In order for transportation engineers and planners to plan, design, and evaluate bicycle facilities, a demand estimate is
necessary. In an attempt to determine this, cities often conduct massive counting campaigns. For example, Portland, Oregon
used 100 volunteers to conduct counts at over 218 locations across the city (Portland Bureau of Transportation, 2012).
Despite the effort of such a laborious endeavor, these bicycle counts are only conducted on one day and provide a very small
window of information, leaving a large data gap for the remainder of the year. Considering that bicycle demand can fluctuate
dramatically based on weather changes (Gallop et al., 2012; Nkurunziza et al., 2012; Nordback et al., 2013; Thomas et al.,
2009), such a small window makes it difficult for planners and engineers to design for the ever-growing and changing
demand for bicycle facilities.
Bicycles have traditionally been far from center on the stage of transportation engineering in the United States, but due to
recent societal shifts, bicycles have experienced a resurgence in urban settings across the United States. This bicycle renais-
sance has also been reflected in the research arm of transportation, with an influx of new and increasingly advancing bicycle
demand models and estimation methods. This research has varied in complexity and applicability, from simple adjustment
factors to complex artificial intelligence-based network models.

2.1. Adjustment factors

Adjustment factors are one of the more fundamental approaches to demand modeling and are often incorporated into
other models. Adjustment factors have been used for automobile demand estimation for decades, but have since been
adapted for bicycles. The process essentially utilizes known traffic data to calibrate adjustment factors that can scale
short-term counts taken at different times or locations based on a variety of factors. Typical adjustment can account for com-
mon factor groups, such as time-of-day ðTODÞ, day-of-week ðDOWÞ, and month-of-year ðMOYÞ. Hankey et al. (2012) and
Nordback et al. (2013) performed extensive studies to develop practical procedures for establishing adjustment factors as
documented in the 2013 Traffic Monitoring Guide (Federal Highway Administration, 2013).
Adjustment factors are typically established using continuous counters in a city. Using the data gathered, the adjustment
factors are calculated for each traffic group as the ratio of that group to the annual average. For example, the MADB of July
may be 200 percent of the annual average AADB, and then an hourly traffic may be 13 percent of that daily count. Essentially,
adjustment factors are a ratio of the specific time to the average. The purpose of using ratios is that adjustment factors, which
are established using a minimum of three to five continuous counters, may then be transferred to different location. A short
duration count at a new location may be scaled using the adjustment factors. Nordback et al. (2013) suggested a short-term
count of one week is the most cost-effective. A major consideration that differs from automobile adjustment factoring is trip
purpose of cyclists. Recreational and commuter cyclists will cycle at greatly different times at nearly all time scales, hourly,
weekly, and yearly. For example, recreational cyclists will generally be far more sensitive to weather and are less likely to
cycle in winter months. Furthermore, recreational cyclists will typically cycle on weekends and at off-peak periods, whereas
156 N. Fournier et al. / Transportation Research Part D 50 (2017) 154–169

commuter cyclists will tend to follow the typical morning and afternoon peak-hour weekday patterns (Federal Highway
Administration, 2013). It is important when scaling to this level of detail to understand the cyclist population of the location.
Later research by Hankey et al. (2014) attempted to further improve accuracy by adjusting directly from day-of-year
ðDOYÞ, rather than grouping into months or weekday/weekends. This approach improved accuracy, but was limited by loca-
tion, less general than standard adjustments, and requires substantial data. Additional research has also investigated adjust-
ments based on other factors, such as weather conditions and road classifications (El Esawey et al., 2013). On a nationwide
scale, Lindsey et al. (2015) captured national trail data in an effort to create a suite of analytical spatial, demand, and eco-
nomic models called the Trail Modeling Assessment Platform (T-MAP). T-MAP built on previous work to create a modeling
platform that can estimate bicycle and pedestrian demand by dividing the country into different climate zones (Hankey
et al., 2014; Lindsey et al., 2015). This climate-zoned approach is an excellent contribution to the research and perhaps
enough data may become available that assumptions for adjustment factors can be made based on region.

2.2. Planning models

Bicycle demand research is often taken from a social, behavioral, and land use perspective. There is a growing number of
studies that employ statistical methods such as regression, to better understand the external factors that influence the deci-
sions to cycle (Akar and Clifton, 2009; Caulfield et al., 2012; Monsere et al., 2012). Although broader planning concepts
regarding mode choice may not directly estimate demand, it is important to understand underlying factors that influence
bicycle ridership.
Most researchers would agree that land use and topography are often major factors affecting bicycle demand (Cervero
and Kockelman, 1997; Pucher et al., 2010; Xing et al., 2010; Griswold et al., 2011). Researchers have often utilized available
census data and surveys to yield predictive patterns between bicycle demand and population demographics (Barnes and
Krizek, 2005; Sullivan et al., 2015). Faghih-Imani et al. (2014) utilized bike-share data with land use data to investigate
how land uses can affect bicycle flows in a bike-share system. Although bike-share users may differ from private cyclists,
some of the same factors affecting bike-share users may apply. Studies even extend to e-bikes and the effects that e-
bikes have on demand, whether e-bikes increase bicycle usage or merely augments existing users (Fyhri and Fearnley,
2015). In addition to these largely static factors is weather, which has a large impact on the daily decision to cycle. Studies
have targeted weather to understand its effect on cyclists (Tin Tin et al., 2012; Thomas et al., 2009). In general, these studies
alone do not form new bicycle demand models, but they do provide literary advances that are important when developing a
behaviorally influenced demand model.

2.3. Time series models

Time series modeling has also proven to be an effective method of estimating and predicting various demand types. Wang
et al. (2014) developed a short-term traffic speed prediction model for automobile traffic. Similarly adapted for bicycles,
work by Doorley et al. (2014) used weather as a predictive factor in a time series demand model. Other research has focused
on specific weather events, rather than overall seasonal weather, such as Gallop et al. (2012) who created a weather-based
bicycle demand of Vancouver, British Columbia. Although these methods present highly accurate and adaptive methods
where continuous count data are available, they do not develop a more generic seasonal function.

2.4. Network models

Although short-term counts are more easily applied to a greater number of locations than continuous counters, it is cur-
rently infeasible to physically count all locations across a city. Gosse et al. (2014) accounted for this issue by incorporating
these points into a network of bicycle demand. This network model utilized traffic generating zones, as used for automobile
modeling, but further incorporated temporal scaling factors, topography, distance, and weather to develop a network of bicy-
cle traffic flows. This is an impressive integration of methods that can yield extremely useful results. However, much of the
model relies upon adjustment factors for estimation, that requires complete data sets. El Esawey et al. (2015) took a very
different approach by utilizing a model called an ‘‘autoencoder neural network” to ‘‘backcast” for locations where data
are missing. This unique method is somewhat different to other network models in that the goal is to improve individual
counts, rather than the whole network. Modeling over a network allows for the ability to utilize mixed sources of bicycle
data, such as GPS, manual count, and continuous counts. Furthermore, it allows for spatially distributed crash data to be
incorporated for a spatial analysis of safety across the network. Strauss et al. (2015) had done this for the city of Montreal
to create a map of high risk locations as well as to compare risk for certain road types.
Although each method is highly generalizable and can achieve a high degree of accuracy, they ultimately rely on a large
amount of continually updated calibration data. Instead of attempting to further improve accuracy, the proposed sinusoidal
model attempts to reduce the amount of calibration data necessary by exploiting common seasonal trends in bicycle
demand.
N. Fournier et al. / Transportation Research Part D 50 (2017) 154–169 157

3. Methodology

The fundamental process for this research is to analyze and identify seasonal bicycling demand patterns for use in the
development of a mathematical model for bicycle demand. The research approach consists of four primary steps:

1. Step 1: Bicycle Counts and Bike-Share Data: collect bicycle count and bike-share data to be used for calibration and val-
idation of the model.
2. Step 2: Data Analysis: analyze and identify any patterns that may exist across the varying locations.
3. Step 3: Model Calibration: develop a general mathematical function to represent seasonal bicycle demand
4. Step 4: Model Validation: validate the function by testing it against observed data and bike-share data.

Each of the methodological tasks is described in further detail in the following sections.

3.1. Step 1: Bicycle counts and bike-share systems

For this research to be conducted, large quantities of accurate bicycle count data were needed. Continuous counters are
long-term automated devices that continuously count all vehicles that pass through a designated location over the course of
time, and in this case, over the span of years. Conversely, short duration counts are conducted over the span of days or weeks.
Although continuous bicycle counting is still relatively new, continuous bicycle count data were acquired from six cities with
a total of 47 count locations, shown in Table 1 along with associated temperature ranges to demonstrate local climate
(Environment Canada, 2015; National Oceanic and Atmospheric Administration, 2015). Continuous count data for Seattle,
Washington; Vancouver, British Columbia; Ottawa, Ontario; and Cambridge, Massachusetts; with 3, 4, 12, and 1 locations,
respectively were acquired using their respective open data portals (City of Cambridge, 2016; City of Ottawa, 2015; City
of Seattle, 2015; City of Vancouver, 2015). Continuous count data were obtained from Portland, Oregon and Arlington, Vir-
ginia; with 6 and 21 locations respectively were requested by the research team and generously offered through each city’s
department of transportation.
Of the 47 locations, 45 were used in the initial model development, and the data from two locations were reserved for
model validation. Bike-share data were also utilized separately for further comparative analysis. All of the counts had been
adjusted to account for the separation of bicycles, pedestrians, and vehicles by the data source provider. If any different
counting technologies were used at different locations, they were not expected to affect the model results.
The counter locations varied between sites that would expect to see more recreational cyclists, such as in parks and trails,
and commuters, such as in urban streets. In general, the expected bicyclist type (e.g., recreational or commuter) was mixed
across the cities and sites. Furthermore, many sites could expect to see a mixture of both commuters and recreational bicy-
clists at the same count location. Given the purpose of the study was to establish a generic function of demand, the use of
both commuter and recreational cyclists count data strengthened the model by further removing bias.
Bike-share data are an excellent alternative or supplemental source of bike demand data in a community. As opposed to
bike count data which are often limited, fragmented, or simply unavailable, bike-share data are usually meticulously
recorded and offer a much larger sample of a community with sometimes hundreds of bike-share stations. For this reason,
bike-share data were also analyzed in parallel to investigate the proposed model’s applicability to these systems. Data from
four bike-share systems used were the Hubway system in Boston, Massachusetts, the Nice Ride system in Saint Paul, Min-
nesota, the CitiBike system in New York, New York, and the Capital Bike-Share in Washington, D.C (Hubway, 2016; Nice Ride,
2016; Citi Bike, 2016; Capital Bike, 2016). The systems varied in size and usage but operate similarly with dedicated stations
and user fee options for yearly, monthly, or casual use memberships.

Table 1
Collected data.

City Number of counters


Ottawa, ON 12
Cambridge, MA 1
Arlington, VA 21
Portland, OR 6
Vancouver, BC 4
Seattle, WA 3
Years of available data Bike share name
Boston, MA 2011–2013 Hubway Bike-Share
Washington, D.C. 2013–2015 Capital Bike-Share
New York City, NY 2010–2015 Citi Bike-Share
Saint Paul, MN 2010–2015 Nice Ride Bike-Share
158 N. Fournier et al. / Transportation Research Part D 50 (2017) 154–169

3.2. Step 2: Data analysis

Continuous counters can frequently malfunction. For this reason, the data sets were thoroughly reviewed for errors by
researchers before any analysis could begin. Such malfunctions would appear as zero counts or unusually high or low counts.
Where a malfunction was encountered within a day, that entire day was removed from the data set. Furthermore, when
more than two weeks of counts were malfunctioning, the entire month was removed from the data. This ‘‘cleaning” of
the data ensured that only accurate MADB was calculated.
Most locations had only a single counting device, but some utilized multiple devices to count each direction or intersec-
tion approach. For example, an intersection might have separate counts for northbound, southbound, eastbound, and west-
bound. In this occurrence, the data were combined for a total bicycle count for the intersection in all directions. If one or
more of the devices were malfunctioning, the portion of the data with the malfunction was removed from the data such that
only periods where all devices were functioning were used.
Once the data had been reviewed for errors, they were summed into daily totals and then divided by the number of days
the device functioned during that month to obtain an average daily bicycle demand per month ðMADBÞ. A further averaged
value for annual demand was calculated as the average annual daily bicycle demand ðAADBÞ. The MADB was then plotted,
this allowed for further exploratory analysis and verification of the suitable function for describing seasonal bicycle demand.
An example of the MADB plot generated for Ottawa, Ontario, Canada is presented in Fig. 1.

3.3. Step 3: Model calibration

When comparing annual bicycle count data across cities, a common sinusoidal trend of seasonally based fluctuations can
be seen in nearly every city, as depicted in Fig. 2. One location, NW 58th Street in Seattle, Washington did not fit the model.
However, upon relative comparison to the other count locations the AADB may be so low that it appears flat. Furthermore,
the street was under construction during part of the time that counts were taken. Despite that, the sinusoidal pattern
appeared strong across multiple cities, exposing the potential to exploit that common pattern for demand estimation
purposes.
While it is not particularly surprising that bicycle demand fluctuates with season, it is surprising that nearly every loca-
tion followed very closely the form of a sinusoidal function. This allowed for developing a simplistic time series model of
annual bicycle demand that follows the general equation of a sinusoid, shown in Eq. (1). The variables of this equation
are further visually described in Fig. 3.
MADBt ¼ A  sinðx  ðt  /ÞÞ þ AADB ð1Þ
where
MADBt Monthly Average Daily Bicycle count ðMADBÞ, in bicycles/day
AADB Sinusoid centerline, or Average Annual Daily Bicycle count ðAADBÞ, in bicycles/day
t Time value, in months
A Amplitude of MADB sinusoid, or the seasonal change in average bicycles/day
x = 2pf Wavelength of MADB sinusoid, in months
1
f =12 Frequency of MADB sinusoid
/
/ Phase of sinusoid, x = horizontal shift in months

Fig. 1. Monthly average daily bicycle counts at Laurier Ave. and Metcalfe St. in Ottawa, ON from July 2011 to January 2015.
N. Fournier et al. / Transportation Research Part D 50 (2017) 154–169 159

Fig. 2. Plots of MADB counts and sinusoidal fit.


160 N. Fournier et al. / Transportation Research Part D 50 (2017) 154–169

1.0 , ADBMax
0.9
0.8 Amplitude,
A
0.7
0.6
, ,
0.5
0.4
0.3
Centerline Average,
0.2 AADB
0.1
, ADBMin
0.0
1 2 3 4 5 6 7 8 9 10 11 12

Fig. 3. Generic sinusoidal function.

1
An initial assumption for the model is that all cities will share the same wavelength of 12 months, or a frequency of 12
months, allowing us to simplify these constants to x ¼ 2pf ¼ p6 . The phase is a constant for the number of months that
the function is shifted such that the crest is aligned with the month of maximum demand. For example, a / value of 0 would
result in the crest being located at t ¼ 3, and / ¼ 4 would result in the crest being located at t ¼ 7. In this study July was
found to most typically be the month of maximum bicycle demand in all cases, and for this reason / was set at a constant
of 4 months. However, it is possible that the phase may require adjustment in some cases where a lag is experienced. For
example, cyclists continue to cycle later in the fall, but are late to resume cycling in the spring.
With the wavelength and phase being known, this yields Eq. (1), with MADB as a function of only three variables: seasonal
change, A, annual average, AADB, and the third being one known variable of time t, in months. Although the equation is sim-
ple, it still has two unknown variables that must be determined prior to the MADB estimation. For complete data sets with
continuous counts for at least one whole year, the values of A and AADB can be easily calculated, but the purpose of this
model is to calculate where data are not known. For incomplete data sets, a one-month long winter count and a one-
month-long summer count could also be used to calculate A and AADB. Nordback et al. (2013) found that only one week
was necessary for creating MADB estimates, but two weeks are preferred.
Seasonal change (A) is equal to half the height of the wave, which can be calculated by dividing the difference between
maximum daily average ðMADBMax Þ and the minimum daily average ðMADBMin Þ for the year in half, shown in Eq. (2). The max-
imum and minimum daily averages should be associated with summer and winter months, respectively.
MADBMax  MADBMin
A¼ ð2Þ
2
where
MADBMin Annual minimum MADB count, in bicycles per day
MADBMax Annual maximum MADB count, in bicycles per day

AADB is the average annual daily bicycles, which can be most directly calculated by dividing the total annual bicycle count
by 365, shown in Eq. (3).
P365
t¼1 ADBt
AADB ¼ ð3Þ
365
where
MADBt = Monthly MADB for month t, in bicycles per day.
Alternatively, AADB is also the centerline of the sinusoid, which can be calculated by averaging the maximum and min-
imum MADB values, i.e., summer and winter average daily bicycle demand, shown in Eq. (4).
MADBMax þ MADBMin
AADB ¼ ð4Þ
2

3.3.1. AADB versus seasonal change


It is important to investigate whether any relationships exist between demand at different locations in order to utilize
data from one location to other locations where less data are available. When visually inspecting the plots in Fig. 2, there
is a noticeable ‘‘stacking” of curves, shown more clearly in Fig. 4. This demonstrates that as AADB increases, A increases
N. Fournier et al. / Transportation Research Part D 50 (2017) 154–169 161

Fig. 4. 2009–2014 estimated bicycle counts from 21 locations in Arlington, VA.

as well. More simply stated, as the total number of bicyclists counted increases, the difference between the summer max-
imum and winter minimum increases as well. In order to describe this relationship between A and AADB, linear regression
analysis was used to investigate the mathematical equation relating the two variables.
Using the data available, simple linear regression was performed for all count locations in all cities, excluding bike-share
data, with A as the dependent variable and AADB as the independent variable. When looking at the distribution of points

Fig. 5. Seasonal change, A, versus AADB and linear regression functions for all city locations.
162 N. Fournier et al. / Transportation Research Part D 50 (2017) 154–169

shown Fig. 5, there appears to be individual alignments of points for each city. This is an extremely important as it suggests
that each city has a different level of seasonality.
Although the linear regression equation could contain a non-zero intercept, it realistically makes sense to have an inter-
cept of zero as it is impossible to have a fluctuation in bicycle demand when there are no bicyclists. Therefore, a linear regres-
sion analysis was performed with a zero intercept using the available points for each individual city. The linear regression
equation is shown as Eq. (5):
A ¼ a  AADB ð5Þ
where
a = regression coefficient.
Table 2 lists the results of the regression analysis performed for each city to determine the coefficient a and respective
statistical significance. Note that number of count locations for Ottawa and Arlington differ as two validation locations were
removed from this portion of analysis. It is interesting that in the regression results the cities with the greatest temperature
difference tended to also have the largest value for a. This corroborates the concept that each city has its own level of sea-
sonality dictated by climate.
All of the locations yielded a P-value well above the 95 percent confidence threshold and an R2 above 0.99 in all cities,
with the exception of Cambridge, Massachusetts having only one location. Looking beyond the numbers, the relationship
between seasonal change A and AADB in general is fairly weak due to the lack of data. Despite this lack of data, the fit is good
and the data that are available would suggest that there is a linear relationship between A and AADB. Ottawa, Ontario in par-
ticular is an excellent example of this, demonstrating a good fit between A and AADB. Further data gathering and investiga-
tion will be necessary to better verify this for other cities.
Solving Eq. (5) for a results in a ratio of A divided by AADB, which can be calculated using Eqs. (2) and (4) to yield Eq. (6)
below. Eq. (6) can directly calculate a based on minimum and maximum MADB monthly counts assuming that the yearly
bicycle demand follows a sinusoidal trend. For example, an MADB of 1,000 cyclists in the summer and an MADB of 100
cyclists in the winter, results in an a factor of 0.82.
MADBMax  MADBMin
a¼ ð6Þ
MADBMax þ MADBMin

3.3.2. Estimating AADB and MADB


Based upon the determination that A has a linear relationship with AADB, we can substitute Eq. (5) into Eq. (1) and solve
for AADB, which results to Eq. (7). We can also easily rearrange Eq. (7) to solve for MADB, as shown in Eq. (8).
MADBt
AADB ¼   ð7Þ
a  sin p6 ðt  /Þ þ 1
h p  i
MADBt ¼ AADB a  sin ðt  /Þ þ 1 ð8Þ
6
The importance of these equations is that once an a factor is established for a community, AADB can be estimated using
Eq. (7) based on the measured MADBt at a given time t. t is counted in months from the beginning of the year, e.g., t ¼ 6 cor-
responds to the month of June. Furthermore, once a constant value for AADB is estimated, it can be used in Eq. (8) to solve for
MADB for any other month, t. This means that given only two short-term counts, MADBMin and MADBMax ; AADB and all remain-
ing MADBs can be estimated using only those two counts.

3.3.3. Bike-share comparison


The fundamental assumption of this model is that a bicycle count from a location is simply a sample of an entire popu-
lation (i.e., a community) that has a proportional distribution of people with different willingness to cycle later into the win-
ter months of the year. If that assumption is true, it could also hold for large sample sizes, such as an entire bike-share
system. Other studies have noted a seasonal change in bicycle demand was seen in both e-bike and traditional bike-share
systems (Langford et al., 2013; Noland et al., 2016). To test this, the sinusoidal model was applied using data from Boston’s

Table 2
City a factors and regression results.

City Count locations Temp. difference (°F) (High  Low) a R2 P-value Standard error
Ottawa, ON 11 56.6 (70.2  13.6) 0.96 0.9661 2:12e13 0.0189
Cambridge, MA 1 44.4 (73.4  29.0) 0.58 – – –
Arlington, VA 21 43.8 (79.8  36.0) 0.61 0.9651 2:95e18 0.0242
Portland, OR 6 29.1 (69.5  40.4) 0.45 0.9664 7:13e05 0.0376
Vancouver, BC 4 25.9 (64.4  38.5) 0.62 0.9919 3:12e04 0.0326
Seattle, WA 3 24.1 (66.1  42.0) 0.55 0.9663 1:84e03 0.0238
N. Fournier et al. / Transportation Research Part D 50 (2017) 154–169 163

Hubway, New York’s CitiBike, Washington D.C.’s Capital Bike, and Saint Paul’s Nice Ride. Analyzing bike-share data allows for
a longitudinal cross section of one community over time, rather than a spatial cross section of multiple locations within a
community over the same time frame. Essentially, the availability of bike-share data makes it possible to determine whether
the a factor changes over time.
From the analysis results shown in Table 3, the value of a changes over time. However, in all cases the first year saw the
most change which can be explained by low initial ridership. Eventually, each of the bike-share systems’ a factor decreased
and stabilized. A decreasing and stabilizing a suggests that as the systems establish themselves and people become more
familiar with the systems, more riders become willing to ride more into the winter months. It is important to remember that
a is effectively the ratio of A and AADB, not the slope of AADB. This is demonstrated in Fig. 6 where both overall ridership
ðAADBÞ and seasonal change ðAÞ increased steadily, yet the a for those years stabilized at approximately 0.45. This shows that
as bike-share systems establish themselves in a community, the a factor stabilizes and becomes more reliable, even as rid-
ership increases. What this means is that a is not a static factor, but more a measure of how sensitive a bicycle population is
to seasonal change for that year, regardless of the size.

3.4. Step 4: Validation

The process of validating the model resulted of the development of the proposed model. For this reason it is presented in
Section 4.2 below.

4. Results and application

The proposed model is fundamentally based on three basic assumptions:

 A bicycle count is a sample of demand in the community as a whole.


 Average daily bicycle demand per month follows a seasonal sinusoidal wave function.
 For each city/region, the seasonal change, A, has a linear relationship with AADB.

If these assumptions are true, then a bicycle count at one location is essentially a small sample of the community pop-
ulation. These three assumptions are justified by the number of count locations that fit the sinusoidal model, shown in Fig. 3,
and the regression results shown in Table 2 and Fig. 5. The only requirement is that a calibration factor of a is established
from one winter month count (MADBMin ) and one summer month count (MADBMax ), or ideally using a year of continuous
count data. Once an a is established for a community, far less data are required for subsequent estimations of MADB and
AADB at multiple locations where the model fits, as demonstrated in Fig. 2.

Table 3
Bike-Share model results summary.

Year Bike rentals Months counted AADB MADBMin MADBMax A a Temp. difference (°F) (High  Low)
Hubway, Boston, Massachusetts
2011 4602 5 920 0 1313 656 0.71 44.4 (73.4  29.0)
2012 17,470 9 1941 0 2807 1404 0.72
2013 29,599 8 3700 0 4794 2397 0.65
Average 0.69
Citi Bike, New York City, New York
2013 206,500 8 25,812 1454 38,731 18,639 0.72 39.3 (74.4  35.1)
2014 291,411 12 24,284 8834 35,072 13,119 0.54
2015 335,205 12 27,934 7769 42,742 17,486 0.63
Average 0.63
Capital Bike, Washington, D.C.
2010 3853 4 963 140 1604 732 0.76 43.8 (79.8  36.0)
2011 40,687 12 3391 1230 4776 1773 0.52
2012 67,235 12 5603 3116 7286 2085 0.37
2013 84,768 12 7064 3971 9506 2768 0.39
2014 96,649 12 8054 3678 10,920 3621 0.45
2015 81,635 9 9071 3920 11,836 3958 0.44
Average 0.49
Nice Ride, Saint Paul, Minnesota
2010 3324 6 554 0 769 384 0.69 58.2 (73.8  15.6)
2011 7181 8 898 0 1343 672 0.75
2012 9023 8 1128 0 1748 874 0.77
2013 10,150 8 1269 0 2115 1058 0.83
2014 13,620 8 1702 0 2984 1492 0.88
2015 15,801 8 1975 0 3085 1543 0.78
Average 0.78
164 N. Fournier et al. / Transportation Research Part D 50 (2017) 154–169

Fig. 6. Capital bike-share demand & sinusoidal estimate, Washington, D.C.

4.1. Example calculation steps

The calculation steps for this method are described in the five steps below. These five steps are divided into two cate-
gories, establishment of the community-wide a factor, and target location calculations. The community-wide a factor only
needs to be calculated once to be usable by any other location in the community. However, the a factor can be improved with
more data.

1. Establish community-wide a factor


(a) Conduct preliminary bicycle counts for at least one year and at least one location to calculate A and AADB using Eqs.
(2) and (4), or more accurately using Eq. (3).
(b) Calculate a community-wide a factor using Eq. (6), or more accurately with regression analysis using Eq. (5).
2. Estimate bicycle demand at other locations
(a) Conduct short-term counts at target location to establish a monthly average, MADBt .
(b) Calculate target location AADB using Eq. (7), setting / to the expected peak bicycle demand month (July would be
/ ¼ 7  3 ¼ 4).
(c) Estimate any desired MADBt by plugging AADB into Eq. (8) for a month t.

4.2. Validation

To validate the model, two count locations were removed from the data set and reserved for validation; one from Arling-
ton, Virginia and one from Ottawa, Canada. The validation data for Ottawa and Arlington were continuous counts taken from
the Alexandra bridge from January 2010 to September 2012, and from Bon Air West from December 2012 to July 2015. The
multi-year continuous count data were then averaged per month to represent the monthly average MADB. The reason why
Arlington and Ottawa were used to validate the model was based on three criteria:

1. Climate difference: To test the model against a range of weather conditions, the two sites were chosen due to the dif-
ferent climates they have. Arlington has relatively mild seasons, while Ottawa much more severe.
2. Duration of data collection: Counter locations with at least two years of data were used in order to validate the model
using a large data sample.
3. Quantity of city count locations: Since the data were removed for validation, locations with the largest number of coun-
ters were selected to ensure that the city had enough remaining count locations for model development.

Validation was performed as if attempting to estimate bicycle demand without a whole year of continuous bicycle counts.
First calculating an a factor using Eq. (6) by taking month-long counts in July and January, which typically represent MADBMax
and MADBMin . The calculated a factors for Ottawa and Arlington were 0.99 and 0.57, respectively. These a factors for Ottawa
and Arlington are quite similar to the 0.96 and 0.61 calculated using regression in Table 2.
The measured MADB was then used with the a factor to estimate AADB using Eq. (7). This process was repeated for each
month as if it was only month with available data in the year to measure the relative accuracy that each month has in esti-
mating AADB. Results are shown in Table 4. The error varied dramatically between the two locations, with Arlington’s error
ranging from 1.0 to 18.2 percent and Ottawa from 0.2 to over 100 percent error. The high errors were associated with winters
months during which a small number of bicyclists counted in Ottawa. May to October, the error ranged from 0.2 to just 12.5
percent error at both locations. Given these results, it is recommended to avoid utilizing winter months when estimating
AADB using Eq. (7).
N. Fournier et al. / Transportation Research Part D 50 (2017) 154–169 165

Table 4
Validation results.

Location a / Measured MADB Month, t Measured AADB Estimated AADB Error (%)
Ottawa, ON 0.99 4 2b 1 934 39 2299.0
9 2 54 1636.0
319 3 615 56.9
789 4 789 22.4
1426 5 963 0.2
1819 6 993 2.8
1808a 7 922 4.7
1702 8 929 3.9
1550 9 1047 7.83
936 10 936 3.1
641 11 1234 21.8
204 12 1219 20.8
Arlington, VA 0.57 4 694b 1 1745 1776 1.7
701 2 1485 17.5
1111 3 1598 9.2
2132 4 2132 18.2
2601 5 1994 12.5
2514 6 1646 6.0
2507a 7 1557 12.0
2572 8 1683 3.7
2445 9 1874 6.9
1762 10 1762 1.0
1184 11 1702 2.5
717 12 1518 14.9
a
Used as MADBMax .
b
Used as MADBMin .

From the AADB estimated using one MADB, the remaining MADBs were then estimated using Eq. (8). This process was
repeated for each estimated AADB. These estimated MADBs were compared to actual MADBs in the data. Results from this
validation step are presented in Table 4 and in Figs. 7(a) and (b) to visually show how the estimated MADB compared with
the measured MADB.

4.2.1. Estimation accuracy


The estimation accuracy of this model was further investigated using all 47 available count locations to estimate both
AADB and MADB. The AADB estimations were done similarly to the methods used in the validation step, but instead used
the a factors previously determined by regression in Table 2. The AADB for each location was then calculated using each
month as if it were the sample month. The percent errors at each location were then averaged by city in order to determine
the relative estimation accuracies of each sample month for each city. The results of this AADB estimation are shown in
Table 5.
Overall the AADB estimation accuracy of the model yielded an average error of 15.58 percent, ranging from 8.64 percent in
August to 28.11 percent in January. In general the model achieved the most accurate estimates of AADB when the sample was
taken from between June and October. This matches the findings in both the validation step as well as findings by El Esawey
et al. (2013). For the cities the overall error ranged from 11.84 percent to 24.71 percent. However, the overall city averages
were conservative as they include inaccurate winter months, such as January and December. When looking at just June for
example (a summer month where all cities had data), the error was a much lower range from only 8.22 percent to 13.09
percent and an average of 11.41 percent.
The methods used to investigate MADB estimation accuracy were again similar to those used in the validation step. How-
ever, the sample month for the calculation was assumed to be July in all cases, rather than a rotating sample. July was used as
it was observed to be the typical maximum monthly MADB and also typically had a high degree of accuracy in estimation.
The one exception to this was Cambridge, Massachusetts where July data was unavailable and June was used in its place.
Using the one estimation month and the previously determined a values, the MADB was estimated for each month at all
47 locations in all cities. The results were again averaged per month and per city in order to determine the relative accuracy
of the model in each case. The results of this are shown in Table 6.
The model achieved an average overall accuracy of 21.14 percent error between the estimated and actual MADB. Similarly
with AADB, the accuracy varied between months, with the most accurate estimates being between June and September. It is
noticeable that cases of high error in AADB were reflected in MADB error. This is because MADB is calculated using the pre-
viously calculated AADB, thus carrying over any error. This may be refined through further investigation of the AADB calcu-
lation step, for example, by comparing the use of Eqs. (4)–(7) in AADB calculation.
This accuracy evaluation of the model is by no means absolute, but the initial findings show that the model was capable of
performing quite well. Nordback et al. (2013) used Average Annual Daily Traffic (AADT) of motorized vehicles as a benchmark
of comparison. The study compared motorized vehicle AADT estimates conducted in Florida and Minnesota, which found
166 N. Fournier et al. / Transportation Research Part D 50 (2017) 154–169

Fig. 7. Estimated MADB versus measured MADB values.

Table 5
Average percent error in AADB estimation.

Sampled City and number of count locations


month
Arlington, VA (21) Cambridge, MA (1) Ottawa, ON (12) Portland, OR (6) Seattle, WA (3) Vancouver, BC (4) Average
(%) (%) (%) (%) (%) (%) (%)
January 15.24 16.17 54.28 23.95 24.92 34.09 28.11
February 18.41 10.81 15.31 14.78 13.89 35.82 18.17
March 12.23 5.84 57.85 10.14 7.20 5.61 16.48
April 15.89 23.64 34.86 10.13 5.56 17.10 17.86
May 7.63 19.60 15.12 8.02 13.34 25.94 14.94
June 8.22 11.45 12.10 12.99 10.61 13.09 11.41
July 15.77 NA 9.52 14.43 5.91 10.46 11.22
August 12.33 NA 9.75 6.85 11.21 4.63 8.96
September 5.98 NA 10.26 6.35 17.29 3.34 8.64
October 5.58 NA 13.53 7.14 21.04 7.76 11.01
November 11.43 27.81 25.61 18.03 23.59 13.87 20.06
December 13.33 15.22 38.30 13.08 21.55 18.87 20.06
Average 11.84 16.32 24.71 12.16 14.67 15.88 15.58

errors ranging from 5 to 83 percent, with an average of 12 and 14 percent in Florida and Minnesota. The adjustment factors
in that study achieved AADB estimation errors of 15–30 percent (Nordback et al., 2013). In comparison, the sinusoidal model
had an average AADB error of 11.22 percent for July samples and 15.58 percent overall. The model performed quite well
considering the possible reduction of input data necessary to achieve that level of accuracy. However, where traditional
adjustment methods can continually be improved through accumulation of calibration data, the sinusoidal model will
not. This is because the sinusoidal model attempts to fit the seasonal pattern using an equation, additional data will only
improve the initial calibration, but not the shape of the model.
N. Fournier et al. / Transportation Research Part D 50 (2017) 154–169 167

Table 6
Average percent error in MADB estimation.

Estimated City and number of count locations


month
Arlington, VA (21) Cambridge, MA (1) Ottawa, ON (12) Portland, OR (6) Seattle, WA (3) Vancouver, BC (4) Average
(%) (%) (%) (%) (%) (%)
January 34.53 5.64 60.76 47.66 41.24 59.44 41.54
February 30.94 0.72 57.44 42.43 36.58 56.31 37.40
Marh 23.22 5.95 51.80 32.55 17.35 21.81 25.44
April 40.01 15.97 31.34 21.17 9.08 18.45 22.67
May 27.22 10.15 20.56 10.42 13.03 17.59 16.49
June 11.77 0.00 10.91 11.36 5.66 14.46 9.03
July 6.43 NA 10.41 2.79 0.57 6.61 5.36
August 10.59 NA 9.27 11.19 5.55 14.11 10.14
September 24.88 NA 14.86 14.99 13.10 15.18 16.60
October 22.49 NA 22.21 23.40 14.11 18.72 20.19
November 26.55 22.67 36.01 15.32 14.45 23.14 23.02
December 29.42 4.46 48.84 19.77 18.15 33.83 25.74
Average 8.19 24.00 31.20 21.09 15.74 24.97 21.14

5. Discussion and conclusions

5.1. Seasonality

The proposed model does not come without its caveats. There are several factors to consider that may cause this model to
break down. The main factor to consider when implementing this model is the seasonality of a location, or the lack thereof.
In locations with mild seasons, such as southern cities in the United States, it is likely that these places will not have as large
a change in bicycling counts throughout the year. In contrast, locations with extreme summers may actually have reduced
cycling due to the heat. The exact threshold at which the sinusoidal model becomes invalid has not been determined due to
the limited number of cities explored in this project. However, the sinusoidal model is expected to be less sensitive to overall
temperature difference, but rather the seasonal change from ‘‘comfortable” to ‘‘uncomfortable” cycling weather. With that
consideration, cities that remain ‘‘comfortable” all year, such as San Francisco, may not be applicable for this model and a
more in depth adjustment factor method may be more appropriate.
Although this study utilized a substantial amount of data from 47 locations in six cities across North America, all possible
climate zones are not represented. In order for seasonality to be fully investigated, sample cities in all climate zones must be
systematically analyzed to determine the applicable boundaries of the model. Future efforts should pursue this research to
determine the threshold of applicable cities and their seasonality.

5.2. Cyclist type

Cyclist type is also an important factor to consider. Recreational cyclists are expected to be more sensitive to adverse
weather than their daily commuting counterparts. Given that a is essentially a measure of how sensitive a cyclist population
is to seasonal weather, it is safe to expect that locations dominated by recreational cyclists will have a higher a factor than
locations dominated by commuters. This project studied both commuter-heavy urban locations as well as scenic bike trails,
in both cases a sinusoidal model was able to be modeled. The a values may have varied, but Fig. 5 has shown that there is
generally a prevailing a value that can be used more generically for a city in the absence of a localized a factor.
This consideration for cyclist type also extends to bike share systems as well. Although bike-sharing systems are an excel-
lent source of data, they potentially exclude an extremely important group from their data. Public bike-share systems are
complex systems and often have an even more complex relationship with their users. The low commitment systems often
function as a means of transitioning potential bicyclists into full time cyclists (Castillo-Manzano et al., 2015). With this tran-
sitional nature, it is possible that bike-share users are made up of part-time or introductory cyclists (Heinen et al., 2011).
A bike-share rider population that does not include full-time cyclists would mean excluding cyclists who are more likely
to cycle regularly in the winter. This means that a bike-share system may have a seasonal change greater than the total bicy-
clist population. This would result in a higher a factor. Thus a factors may not entirely be interchangeable between bike-
share usage and bicycle traffic counts, despite sharing the same location.

5.3. Bike counts

Low-volume bicycle count locations may present a challenge to the sinusoidal model. Minor changes in bicycle counts are
to be expected as weather and time are not the only factors that affect an individual bicyclist. However, these minor fluc-
tuations can be magnified as error when calculated. It was observed in this study that low-volume locations with an AADB
168 N. Fournier et al. / Transportation Research Part D 50 (2017) 154–169

of less than 100 showed signs of estimation error. However, the majority of count locations in this study were relatively high
volume and definitive conclusions are unknown. Further study should investigate the potential error of this model at low-
volume locations.
Although this study benefited from the luxury of continuous count data for a whole year to calculate MADB, work by
Nordback et al. (2013) and Hankey et al. (2012) found that counts of 30 days are unnecessarily long to establish accurate
factors. The same applies when using short-term counts with the sinusoidal model. The Traffic Monitoring Guide (TMG)
by the Federal Highway Administration (2013) recommends an ideal count duration of two weeks, or a minimum of seven
days when using an automatic counting device. It is recommended that the same minimums also apply to short-term counts
for the sinusoidal model. However, automated count devices may be challenging to implement during winter months in
snowy and icy locations. The devices may be damaged by plows on the road, or by the severe weather itself. Manual counts
may be a feasible alternative, but with a recommended count duration of at least four to six hours per day, cold weather
exposure is still a challenge. Emerging technologies, such as video and infrared may help alleviate this issue in the future
as costs are reduced.
It is recommended that at least one continuous counter be used in a high bicycle traffic location in order to establish the
presence of the sinusoidal pattern in a particular city. However, if other empirical evidence proves the existence of the sinu-
soidal trend, such as multiple short-term counts, then it may not be necessary to utilize continuous counters for a calibra-
tion. As discussed, a values may differ within cities depending on localized sensitivities, but in general there will be a
prevailing a value. It is difficult to set a minimum number of recommended count sites to establish a prevailing a due to
the varying seasonal sensitivities of cities. For example, Ottawa, Ontario likely requires fewer data points necessary for deter-
mining an a compared to the highly variable and milder seasoned Arlington, Virginia. Compared to establishing adjustment
factors, the TMG recommends a minimum of five to seven sites to establish factors. This threshold is consistent with results
in this study where statistical significance was severely weakened when below five count sites. For this reason, it is sug-
gested using at least five count sites in a city in order to determine a prevailing a value. However, the benefit of the sinu-
soidal model is that these five count sites do not need to be continuous counters if a sinusoidal pattern is known to exist,
only the two winter and summer MADB counts.

5.4. Implications

It is undeniable that bicycling is on the rise in cities across the United States and the need to accommodate and further
encourage safe bicycling is growing as well. Although accurate estimates of demand are a fundamental tool in transportation
planning and design, transportation agencies and designers currently are limited by the lack of available count data. Further-
more, estimating bicycle demand at locations with more severe seasonal weather are greatly disadvantaged compared to
cities with more mild climates. This makes it difficult to support any engineering, policy, or planning decisions behind bicy-
cle infrastructure and reinforces the need for bicycle demand models.
The sinusoidal bicycle demand model exploits a common seasonal trend that many North American cities follow, allow-
ing for a reduction in calibration data to two short-term MADB counts. This model is not applicable to all locations, and is not
a replacement for other more established methods of bicycle demand estimation. However, the model is an alternative that
could prove extremely useful in cases where data are limited and a strong seasonal pattern is known to exist. In many cases a
rough estimate is all that is needed for preliminary demand estimation or high-level safety analysis. In these cases where
only MADB and AADB are needed, the sinusoidal model is more than adequate. Furthermore, the model is not capable of esti-
mating at the day-of-week (DOW) or time-of-day (TOD) levels, but it can accurately estimate for AADB and monthly MADB.
Perhaps using the short-term counts taken for calibration, TOD and DOW factors could be established in conjunction with the
sinusoidal model. This mixed method approach may achieve the desired level of accuracy while still reducing the amount of
input data.
Recommended next steps include the collection and analysis of more data to improve the accuracy and understanding of
applicable locations of the model. Further research should be done to investigate the variability of a factors across cities in
different climate zones, not only to understand how climate affects the a factor and associated seasonality, but to understand
the threshold at which mild seasoned cities become inapplicable for this model. This presents only an introduction of this
new model, further research is necessary for refinement.

Acknowledgments

This research was funded in part through a research project awarded by the Safety Research Using Simulation (SAFER
SIM) University Transportation Center (UTC) at the University of Iowa. Partial funding was also provided by the New England
UTC and the Dwight D. Eisenhower Graduate Transportation Fellowship Program. Funding for the UTC Program is provided
by the Office of Assistant Secretary for Research and Innovation (OST-R) of the United States Department of Transportation
(USDOT). Sincere thanks to Andy Berthaume for his research support. Special recognition to the cities of Portland, Oregon;
Arlington, Virginia; Seattle, Washington; Ottawa, Ontario; Vancouver, British Columbia; and Cambridge, Massachusetts for
taking it upon themselves to invest in automated bicycle counters and to so willingly share the data.
N. Fournier et al. / Transportation Research Part D 50 (2017) 154–169 169

References

AASHTO, 2009. AASHTO Guidelines for Traffic Data Programs. American Association of State Highway and Transportation Officials.
Ahmed, F., Rose, G., Jakob, C., 2013. Commuter cyclist travel behavior: examination of the impact of changes in weather. Transport. Res. Rec.: J. Transport.
Res. Board (2387).
Akar, G., Clifton, K., 2009. Influence of individual perceptions and bicycle infrastructure on decision to bike. Transport. Res. Rec.: J. Transport. Res. Board
(2140), 165–172.
Barnes, G., Krizek, K., 2005. Estimating bicycling demand. Transport. Res. Rec.: J. Transport. Res. Board 1939 (1), 45–51.
Capital Bike, 2016. Capital Bikeshare System Data. <https://www.capitalbikeshare.com/system-data>.
Castillo-Manzano, J.I., Castro-Nuño, M., López-Valpuesta, L., 2015. Analyzing the transition from a public bicycle system to bicycle ownership: a complex
relationship. Transport. Res. Part D: Transport Environ. 38, 15–26.
Caulfield, B., Brick, E., McCarthy, O.T., 2012. Determining bicycle infrastructure preferences A case study of Dublin. Transport. Res. Part D: Transport Environ.
17 (5), 413–417.
Cervero, R., Kockelman, K., 1997. Travel demand and the 3Ds: density, diversity, and design. Transport. Res. Part D: Transport Environ. 2 (3), 199–219.
Citi Bike, 2016. Citi Bike System Data. <https://www.citibikenyc.com/system-data>.
City of Cambridge, 2016. Cambridge Open Data. <https://data.cambridgema.gov/>.
City of Ottawa, 2015. Ottawa Open Data. <http://data.ottawa.ca/>.
City of Seattle, 2015. Seattle Open Data. <https://data.seattle.gov/>.
City of Vancouver, 2015. Vancouver Open Data. <http://vancouver.ca/your-government/open-data-catalogue.aspx>.
Crites, J.M., Vice, E., Fort, D., International, W., 2014. NCHRP REPORT 797 Guidebook on Pedestrian and Bicycle Volume Data Collection, Vol. D.
Transportation Research Board.
Doorley, R., Pakrashi, V., Caulfield, B., Ghosh, B., 2014. Short-term forecasting of bicycle traffic using structural time series models. 17th International IEEE
Conference on Intelligent Transportation Systems (ITSC), vol. 4. IEEE, pp. 1764–1769.
El Esawey, M., Lim, C., Sayed, T., Mosa, A.I., 2013. Development of daily adjustment factors for bicycle traffic. J. Transport. Eng. 139 (8), 859–871.
El Esawey, M., Mosa, A.I., Nasr, K., 2015. Estimation of daily bicycle traffic volumes using sparse data. Comput. Environ. Urban Syst. 54, 195–203.
Environment Canada, 2015. Canadian Climate Normals. <http://climate.weather.gc.ca/>.
Faghih-Imani, A., Eluru, N., El-Geneidy, A.M., Rabbat, M., Haq, U., 2014. How land-use and urban form impact bicycle flows: evidence from the bicycle-
sharing system (BIXI) in Montreal. J. Transport Geogr. 41, 306–314.
Federal Highway Administration, 2013. Traffic Monitoring Guide.
FHWA, 2011. Roadway Safety Information Analysis: A Manual for Local Rural Road Owners. <http://safety.fhwa.dot.gov/local_rural/training/fhwasa1108/
ch3.cfm>.
Fyhri, A., Fearnley, N., 2015. Effects of e-bikes on bicycle use and mode share. Transport. Res. Part D: Transport Environ. 36, 45–52.
Gadda, S., Magoon, A., Kockelman, K.M., 2007. Quantifying the uncertainty in annual average daily traffic (AADT) count estimates. In: 11th World
Conference on Transport Research.
Gallop, C., Tse, C., Zhao, Jinhua, 2012. A seasonal autoregressive model of Vancouver bicycle traffic using weather variables. Transportation Research Board
91st Annual Meeting (12-2119), 1–17.
Gosse, C.A., Hall, T., Clarens, A., 2014. Estimating spatially and temporally continuous bicycle volumes using sparse data. Transport. Res. Rec.: J. Transport.
Res. Board 2443, 115–122.
Griswold, J.B., Medury, A., Schneider, R.J., 2011. Pilot models for estimating bicycle intersection volumes. Transport. Res. Rec.: J. Transport. Res. Board (2247),
1–7.
Hankey, S., Lindsey, G., Marshall, J., 2014. Day-of-year scaling factors and design considerations for non-motorized traffic monitoring programs. In:
Proceedings of the 93rd Annual Meeting of the Transportation Research Board, pp. 1–17.
Hankey, S., Lindsey, G., Wang, X., Borah, J., Hoff, K., Utecht, B., Xu, Z., 2012. Estimating use of non-motorized infrastructure: models of bicycle and pedestrian
traffic in Minneapolis, MN. Landsc. Urban Plann. 107 (3), 307–316.
Heinen, E., Maat, K., van Wee, B., 2011. The role of attitudes toward characteristics of bicycle commuting on the choice to cycle to work over various
distances. Transport. Res. Part D: Transport Environ. 16 (2), 102–109.
Hubway, 2016. Hubway Data Visualization Challenge. <http://hubwaydatachallenge.org/>.
Langford, B.C., Cherry, C., Yoon, T., Worley, S., Smith, D., 2013. North America’s first e-bikeshare: a year of experience. Transport. Res. Rec.: J. Transport. Res.
Board (2387).
Lindsey, G., Thorstensen, L., Hadden Loh, T., Lowry, M., 2015. Trail Modeling and Assessment Platform (T-MAP) – Rails-to-Trails Conservancy.
Monsere, C., McNeil, N., Dill, J., 2012. Multiuser perspectives on separated, on-street bicycle infrastructure. Transport. Res. Rec.: J. Transport. Res. Board
(2314), 22–30.
National Oceanic and Atmospheric Administration, 2015. National Centers for Environmental Information. <http://www.ncdc.noaa.gov/>.
Nice Ride, 2016. Nice Ride System Data. <https://www.niceridemn.org/data/>.
Nkurunziza, A., Zuidgeest, M., Brussel, M., Van Maarseveen, M., 2012. Examining the potential for modal change: motivators and barriers for bicycle
commuting in Dar-es-Salaam. Transport Policy 24, 249–259.
Noland, R.B., Smart, M.J., Guo, Z., 2016. Bikeshare trip generation in New York City. In: Transportation Research Board 95th Annual Meeting.
Nordback, K., Marshall, W.E., Janson, B.N., Stolz, E., 2013. Estimating annual average daily bicyclists. Transport. Res. Rec.: J. Transport. Res. Board (2339), 90–
97.
Portland Bureau of Transportation, 2012. Portland Bicycle Count Report 2012 (May 2013).
Pucher, J., Dill, J., Handy, S., 2010. Infrastructure, programs, and policies to increase bicycling: an international review. Prevent. Med. 50 (Suppl. 1), S106–
S125.
Sanders, R.L., 2013. Dissecting Perceived Traffic Risk as a Barrier to Adult Bicycling 2000 (January).
Strauss, J., Miranda-Moreno, L.F., Morency, P., 2015. Mapping cyclist activity and injury risk in a network combining smartphone GPS data and bicycle
counts. Accident Anal. Prevent. 83, 132–142.
Sullivan, J., Dowds, J., Aultman-Hall, L., 2015. Regional Models of Bicycle and Pedestrian Travel in Chittenden County, Vermont.
Thomas, B., DeRobertis, M., 2013. The safety of urban cycle tracks: a review of the literature. Accident Anal. Prevent. 52, 219–227.
Thomas, T., Jaarsma, R., Tutert, B., 2009. Temporal variations of bicycle demand in the Netherlands: the influence of weather on cycling. In: Transportation
Research Board 88th Annual Meeting.
Thompson, S., Monsere, C., Figliozzi, M., Koonce, P., Obery, G., 2013. Bicycle-specific traffic signals. Transport. Res. Rec.: J. Transport. Res. Board (2387), 1–9.
Tin Tin, S., Woodward, A., Robinson, E., Ameratunga, S., 2012. Temporal, seasonal and weather effects on cycle volume: an ecological study. Environ. Health:
Global Access Sci. Source 11 (1), 12.
Wang, H., Liu, L., Qian, Z.S., Wei, H., Dong, S., 2014. Empirical mode decomposition autoregressive integrated moving average: hybrid short-term traffic
speed prediction model. Transport. Res. Rec.: J. Transport. Res. Board (2460).
Wang, T., Gan, A., Alluri, P., 2013. Estimating annual average daily traffic for local roads for highway safety analysis. Transport. Res. Rec.: J. Transport. Res.
Board 2398, 60–66.
Xing, Y., Handy, S.L., Mokhtarian, P.L., 2010. Factors associated with proportions and miles of bicycling for transportation and recreation in six small US
cities. Transport. Res. Part D: Transport Environ. 15 (2), 73–81.

You might also like