Examining Travel Time Reliability-Based Performance

Indicators for Bus Routes Using GPS-Based

Bus Trajectory Data in India
Akhilesh Chepuri 1; Jairam Ramakrishnan 2; Shriniwas Arkatkar 3;
Gaurang Joshi 4; and Srinivas S. Pulugurtha, M.ASCE 5
Downloaded from ascelibrary.org by University Of Sydney on 03/11/18. Copyright ASCE. For personal use only; all rights reserved.

Abstract: This paper aims to evaluate travel time variability as well as reliability indexes using global positioning systems (GPS)–based
trajectory data of bus trips collected along a selected bus route of the city of Chennai in the southern part of India. Travel time reliability
indexes, such as planning time index (PTI), buffer time index (BTI), and buffer time (BT), along with other statistical measures over
different time periods are estimated. Generalized extreme value (GEV) distribution is found to be the best-fitted distribution for explaining
bus travel time variability reasonably well, using the Kolmogorov-Smirnov (KS) test. Buffer time and 95th percentile travel time are the
reliability measures with the most potential, the variation of which reasonably matches the variation in k-value (shape parameter of GEV
distribution) over time. The findings from the statistical distribution analysis indicate that travel times during peak hours can be better
described using normal distributions. The generic model is developed for predicting volumes based on bus journey speeds. Further, the
developed model is validated with the help of travel time data of the same route during a different time period. The study also attempts to
demonstrate a methodology for establishing level-of-service (LoS) criteria using reliability indicators. The classification of reliability
indicators, considering segment-level travel time data, coefficient of variation (COV) of travel time, and volume-to-capacity ratio (V=C),
is finally presented using the cluster technique. Finally, the study concludes that the most effective performance indicators for examining
travel time variability on a given bus route are 95th percentile travel time and BT. DOI: 10.1061/JTEPBS.0000109. © 2018 American
Society of Civil Engineers.
Author keywords: Travel time reliability; Buffer time index (BTI); Generalized extreme value (GEV) distribution; Clustering; Silhouette.

Introduction reliability is a better way of assessing the performance of the sys-

tem under varying traffic and other conditions over space and time.
Reliability is a key factor for capturing the disturbances that occur Reliability analysis should efficiently include the effect of vari-
on any road network system. These disturbances or variations may ous factors on travel time variability and, in turn, reliability. Major
be either recurrent or nonrecurrent in nature. Recurrent delays are factors that affect travel time reliability are driving behavior, road-
caused by the variations in traffic volumes and composition over way and traffic conditions, and occurrence of incidents or events.
space and time, whereas nonrecurrent delays are caused by the oc- Considering all these factors simultaneously in reliability analysis
currence of some unexpected events or incidents. For capturing is very complex in nature and requires a large amount of reliable
these disturbances or delays, the travel time variations under differ- and accurate data. This is only possible in the presence of intelli-
ent conditions are studied in detail. The consistency in travel time gent transportation system (ITS) infrastructure on a given road net-
variations is better measured in terms of reliability. Thus, travel work. In the case of a public transportation system, other factors
time reliability analysis for any roadway facility helps in assessing such as dwell times, number of passengers alighting and boarding
its performance for a wide range of possible extreme conditions that a bus at a given time, and bus schedules also affect the reliability
can occur. In particular, in the case of public transportation systems, and makes it even more complex. Reliability analysis can be carried
out at different levels, such as link level, segment level, or corridor-
1
level (especially for private transportation), route level (used gen-
Research Scholar, Dept. of Civil Engineering, Sardar Vallabhbhai erally in case of public transportation systems), and network level
National Institute of Technology, Surat 595007, India.
2
Formerly, Postgraduate Student, Dept. of Civil Engineering Sardar
(private as well as public transportation). The desired objectives
Vallabhbhai National Institute of Technology, Surat 595007, India. and purpose of a reliability study decide the level at which the
3
Assistant Professor, Dept. of Civil Engineering, Sardar Vallabhbhai analysis has to be carried out.
National Institute of Technology, Surat 595007, India (corresponding The present study is carried out using route-level and segment-
author). E-mail: [email protected]; [email protected] level reliability data pertaining to the public transportation system.
4
Associate Professor, Dept. of Civil Engineering, Sardar Vallabhbhai Segment-based reliability data are considered for the classification
National Institute of Technology, Surat 595007, India. technique. For calculating the travel time reliability indexes, the
5
Professor, Dept. of Civil and Environmental Engineering, Univ. of 95th percentile travel time, average travel time (ATT), and free-
North Carolina at Charlotte, Charlotte, NC 28223-0001.
flow travel times are considered from a data set. Free-flow travel
Note. This manuscript was submitted on March 16, 2017; approved on
July 24, 2017; published online on March 8, 2018. Discussion period open time in the present study implies the possible minimum travel time
until August 8, 2018; separate discussions must be submitted for individual that has been experienced by a bus during off-peak hours. In the
papers. This paper is part of the Journal of Transportation Engineering, case of public transportation, free-flow travel time implies the
Part A: Systems, © ASCE, ISSN 2473-2907. minimum possible impedance to the free movement of buses.

© ASCE 04018012-1 J. Transp. Eng., Part A: Syst.

J. Transp. Eng., Part A: Systems, 2018, 144(5): 04018012

 The impedance can be attributed to dwell times (boarding and The characteristics of various factors affecting the travel time
alighting passengers) or the traffic movement on a particular route. variability are generally aggregated for a given route or corridor.
This research aims to evaluate travel time variability as well as This may not be more appropriate because the reliability changes
reliability using indexes for the selected bus route under mixed traf- over space for a given time period (based on land use, traffic, and
fic conditions. The study of variation in generalized extreme value roadway conditions). Thus, link- or segment-level reliability analy-
(GEV) distribution parameters under different flow conditions is sis helps in quantifying the effect of individual links on overall
also attempted in this study. The study also aims at the development route reliability. Puvvala et al. (2015) researched and proposed
of a volume prediction model using bus speed along with the Cronbach’s coefficient α (a two-dimensional measure) as a perfor-
validation under mixed traffic conditions prevailing in India. In mance measure complementing the traditional indicators to assess
addition, the study also presents a discussion on the effective reli- link-level reliability.
ability performance measure for signifying the level-of-service A few researchers used archived ITS data of transit systems
(LoS)–based classification. This study also emphasizes methodol- for analyzing the travel time variability and developed reliability
ogy and level of data aggregation to be considered for reliability based LoS (Duddu et al. 2014). Kieu et al. (2014) established the
performance indicator classification and presents a case study high- key definitions of public transport travel time variability (PTTV)
Downloaded from ascelibrary.org by University Of Sydney on 03/11/18. Copyright ASCE. For personal use only; all rights reserved.

lighting practical applications of the concept and reliability as a and calculated probability indicators of PTTV. Pulugurtha et al.
performance measure. (2004) developed car travel time prediction models based on bus
travel time data obtained from global positioning system (GPS)
units. Travel time reliability may not be very complex in the case of
Literature Review dedicated transit systems because there is no interaction with the
mixed traffic stream. In this case, factors affecting travel time reli-
Most agencies have focused on volume-to-capacity ratio (V=C) ability may be largely attributed to variation in transit passenger de-
and LoS measures to describe the efficiency of a road network mand only, apart from any unplanned incident occurrence. However,
(Transportation Research Board 2000). The efficiency of any road it is very important to study the reliability of the bus transit system, in
network system can be effectively measured in terms of travel time which buses greatly interact with the mixed traffic stream most of the
reliability. Reliability of a transportation system is defined as the time. In this case, there are numerous factors that influence the travel
network, which can guarantee an acceptable LoS for road traffic even time variability and route performance.
if the functions of some links are physically damaged or a large Overall, very few reliability studies (Duddu et al. 2014; Kieu
amount of travel demand is occasionally generated (Asakura and et al. 2014; Pulugurtha et al. 2004) are carried out on public trans-
Kashiwadani 1991). In other words, travel time reliability is “an im- portation systems, particularly under mixed traffic conditions.
portant measure of service quality for travelers” (Chen et al. 2003). Also, the importance of segment-level reliability (in both private
The U.S. Department of Transportation proposed four different and public transportation systems) is not at all emphasized under
travel time reliability measures: planning time (PT), planning time mixed traffic heterogeneous conditions. Past reliability studies
index (PTI), buffer time (BT), and buffer time index (BTI) (Federal (Duddu et al. 2014; Kieu et al. 2014; Pulugurtha et al. 2004) have
Highway Administration 2006). Researchers have examined the not attempted the discussion of effective reliability measures for
fundamental characteristics of travel time reliability measures using capturing travel time variations of bus routes. These issues formed
the traffic flow data from archived databases on various roadway as the core basis of motivation for this research. However, avail-
facilities (Higatani et al. 2009). These results can be used for the ability of continuous vehicular trajectory data and large samples
improvement of real-time transportation management and traveler of travel time data plays a key role in achieving accurate and re-
information (Lyman and Bertini 2008). liable results. With increasing emphasis on sustainable transporta-
Further, researchers have highlighted the importance of analyz- tion systems and smart infrastructure, many public transportation
ing the effect of various factors such as traffic volume, number of operators in developing economies are also rigorously focusing on
vehicle breakdowns, number of road accidents, pedestrian cross- effective integration of public transportation system operations with
ings, effect of the intensity of rainfall (weather impacts), and falling ITS technologies.
objects from the vehicles on reliability (Chalumuri and Asakura Very little past research focused on the statistical distribution
2007; Chalumuri et al. 2010). Bharti et al. (2013) analyzed the analysis of travel time data. The characteristics of travel time var-
variation of PTI and BTI values at different V=C ratios. Sathya iations may vary significantly for a given hour of the day and day of
Prabha and Mathew (2013) developed a multivariate travel time the week. These variations can be effectively studied using statis-
regression model by considering independent variables such as tical distributions. In this study, travel time data are tested for a few
length, speed, and volume. Mehbub Anwar (2010) evaluated the selected potential statistical distributions and it is found that GEV
travel time variability as a function of delay, congestion index, is the best fit based on goodness-of-fit tests. It can be certainly ex-
and V=C ratio using four weekdays of travel time data of approx- pected that variations in travel time cause variations in the shape of
imately 30 days. GEV distribution. For studying this, it is attempted to correlate the
Clark and Watling (2005) aimed to reconstruct a full probability variation of different GEV parameters with travel time reliability
distribution for the network performance (total network travel time) measures. Segment-based travel time data are considered for stat-
to examine the effect of variable origin-destination demand flows istical analysis so the effect of characteristics of different individual
on the total network travel time. Travelers’ perception and expect- segments on reliability can be studied. This analysis will help in the
ation of travel time and travel time reliability vary across times better understanding of variations in travel time over space and time
of the day and route choice is based on a combination of travel by means of a statistical approach.
time, travel time reliability, and cost (Nam et al. 2005). Further,
prediction models were developed from statistical distributions
for analyzing travel time and operational reliability on arterial Study Area and Data Collection
routes (Polus 1979). Statistical distributions were also used previ-
ously for analyzing travel time variability, and thereby predicting The bus trajectory data of the 19B route in Chennai, India, is col-
route choice behavior (Gan and Bai 2014; Liu et al. 2004). lected from the ITS center established at the Indian Institute of

© ASCE 04018012-2 J. Transp. Eng., Part A: Syst.

J. Transp. Eng., Part A: Systems, 2018, 144(5): 04018012

 Before carrying out the reliability analysis, it is very important to
understand the travel time variability for different hours of the day,
periods of the day, and days of the week. These travel time vari-
ability patterns help in determining the peak and off-peak hours of
the day, peak day of the week, and peak season of the year. Because
the available data are limited, variations in hour of the day and day
of the week are considered in this study. For this, analysis is carried
out by plotting the variation in travel time throughout the bus opera-
tional hours in a day. Fig. 2(a) shows one such plot, where the box
plot over the 5 on the X-axis represents the hour of 5 to 6 a.m. and
the one over the 16 on the X-axis represents the hour of 4 to 5 p.m.,
and so on. It can be observed that travel times in the morning hours
of 8 to 11 a.m. and in the evening hours of 6 to 8 p.m. are relatively
higher than other hours. Therefore, they are identified as the peak
Downloaded from ascelibrary.org by University Of Sydney on 03/11/18. Copyright ASCE. For personal use only; all rights reserved.

hours for further analysis. Based on these results, the distribution

fitting is separately carried out for peak and off-peak hours. The
travel time data for the entire trip length are considered for further
analysis.
For a given route, the traffic is assumed to follow certain pat-
terns. The patterns can vary over space and time. To study the varia-
tion in travel patterns, trajectories are plotted from the collected
data of 1,195 trips. The temporal patterns are carried out at two
levels: hourly patterns (peak and off-peak) and daily patterns
(weekday and weekend). The hourly patterns explain the variation
in travel time for the selected route over different hours of the day,
mainly focusing on the peak and off-peak behavior of traffic as
shown in Fig. 2(a). It can be seen in Fig. 2(b) that peak trips take
more time to complete the trip, which can be seen in the upper
portion of the spectrum, while the off-peak trips constitute the
lower spectrum of the time-space plots. Plots of the daily patterns
help identify the difference in travel behavior during weekdays and
weekends for a given route. As expected, it can be seen in Fig. 2(c)
that weekday trips compose the upper portion of the spectrum,
showing high variance over the segments, while weekend trips
are shallow with lesser variance. Overall, the plots show that the
trajectories of peak and off-peak trips, and weekday and weekend
Fig. 1. Route 19B (map data © Google)
trips are different, and have to be analyzed separately.

Technology, Madras (IIT Madras). The travel time data are collected
for approximately 30 days (October 2013) and this entire data set Travel Time Variability Using Statistical
is considered for route-level reliability analysis and GEV distribu- Distributions
tion fitting. The generic model for volume prediction is developed
based on segment travel time data of off-peak (5 to 7 a.m.), morning Travel time data for 2 h (of all days), each in the morning and eve-
peak (9 to 11 a.m.) and evening peak (5 to 7 p.m.) data of two week- ning peak and off-peak hours, are considered for examining the
days and two weekend days. For the purposes of validation of the best-fit distribution. Generalized extreme value distribution is
volume prediction model, travel time data of another 30 days for found to fit well in all cases. The results obtained are shown in Fig. 3
similar times of day and days of the week (as considered in model (a). It can be observed that for the off-peak period, the k-value
development) are collected and used for the analysis. For clustering (shape parameter) is observed to be less (higher negative) than that
analysis, 15 days of segment data for selected time periods and days of the peak period (lesser negative or near to zero). For the off-peak
(out of these 30 days) are considered. The data are retrieved for period, the curve has a higher peak, which signifies that most of the
each day from the generated Sequential Query Language (SQL) travel times are lower than the mean travel time.
database. The route has a length of 30 km with an average journey From Fig. 3(b), it can be observed that for the weekend days of
time of 90 min for one trip in one direction. The selected bus route Saturday and Sunday, the curve is leftward and the tail extends to-
road is composed of six lanes of divided carriageway across the en- ward the right side. For Thursday and Friday, the curves have less
tire route, with 15 bus stops in between the source and destination. skewness and the shape resembles normal distribution. Land use
The selected city bus route represents a typical Indian urban arterial may also be a predominant factor, which affects flow characteristics
characterized by heterogeneous and non-lane-based traffic condi- during a given hour of a day for weekdays and weekend days. For
tions. The selected bus route is shown in Fig. 1. example, a bus route section with predominantly commercial land
use may have peak flow characteristics during the weekends and
holidays (for a given hour of a day). Similarly, a bus route section
Preliminary Data Analysis with predominantly industrial or institutional land use will have
peak conditions during the weekdays rather than weekend days
In this study, segregation of data is carried out at two levels: time of (for the same hour of the day). From the GEV results, the skew-
the day (hour of the day and period of the day) and day of the week. ness of the curve gradually decreases and approaches normal

© ASCE 04018012-3 J. Transp. Eng., Part A: Syst.

J. Transp. Eng., Part A: Systems, 2018, 144(5): 04018012

Downloaded from ascelibrary.org by University Of Sydney on 03/11/18. Copyright ASCE. For personal use only; all rights reserved.

Fig. 2. (a) Peak and off-peak hours; (b) peak and off-peak trip patterns; (c) weekday and weekend trip patterns

distribution (as k equals zero) as traffic flow conditions vary from shape of GEV density function). Therefore, ATT alone may not be
free-flow to extreme-peak-flow conditions. adequate for capturing the effect of variation and randomness in
Further, the variation of shape parameter (k) is plotted for differ- travel times, and hence also the randomness in the k parameter
ent hours of the day for similar days of the week, along with other of GEV.
measures. The measures considered for plotting the k variation are From Fig. 4(a), it can be observed that the k for a given period
BT, ATT, and coefficient of variation (COV). For this purpose, the of the day for Segment 1 follows the same trend as that of BT. Plots
travel time data for three selected segments (Segments 1–3) for overlap and are parallel to each other except for the duration of 1 to
three similar or indifferent days of the week are considered. The 3 p.m. This is because of the higher COV values for these hours of
entire route is decomposed into 56 individual segments, 500 m the day, arising because of smaller sample size or higher standard
in length each. Three such segments are considered based on deviation values. In the case of Segment 2, BT for the same hour
the minimum (Segment 1), medium (Segment 2), and higher seems to deviate from the k plot, whereas it is observed to be quite
(Segment 3) variances of travel times among all the segments. consistent for the rest of the hours. This implies that the variation in
For this purpose, three different weekdays (Tuesday, Thursday, the k parameter of GEV can be predicted based on the BT of any
and Friday) are selected based on the GEV curve as obtained in segment. From 5 a.m. to 9 p.m., an increase or decrease in BT val-
Fig. 3(b). Travel time data of all these weekdays from 5 to 9 p. ues resulted in a substantial increase or decrease of k-values. Sim-
m. for consecutive 2-h time intervals are aggregated and variations ilar results are also obtained in the case of Segment 2, as shown in
in k and BT are plotted as shown in Figs. 4(a and b) as examples. Fig. 4(b). In the case of Segment 1, which has lesser variance, there
Compared with other measures, BT is observed to follow the trend is no perfect distinction between peak and off-peak periods. It can
of travel time variation, and hence the trend of k variation more be presumed that variations in traffic volume across the day for this
consistently for different hours of the day. Here, hour of the day segment do not result in significant variability of travel times.
implies the combined effect of 2 consecutive hours of the day. The maximum BT for Segment 1 is approximately 1.5 min. The
For the same value of ATT, there may be different values of stan- absence of the effect of bus stops or intersections or any kind of
dard deviations resulting in different k-values (variation in the road-side frictions on Segment 1 makes it more reliable. On the

© ASCE 04018012-4 J. Transp. Eng., Part A: Syst.

J. Transp. Eng., Part A: Systems, 2018, 144(5): 04018012

 (a)
(a)
Downloaded from ascelibrary.org by University Of Sydney on 03/11/18. Copyright ASCE. For personal use only; all rights reserved.

(b)
(b) Fig. 4. Buffer time versus k for (a) Segment 1; (b) Segment 2
Fig. 3. Generalized extreme value probability density functions

day and also for different periods of the day. From the results, high
ATT (91.5 min) is observed during the evening hour of 6 to 7 p.m.,
other hand, for Segment 2, the BT value is found to range from 1.5
which can be considered as the peak hour. This is followed by the
to 3.5 min, indicating that it is relatively less reliable. Also, the
ATT for the evening hour of 5 to 6 p.m. The peak hours in the
k-value approaches zero in the peak hour (for 7 to 9 a.m. and 5
morning and evening are clearly depicted in box plots presented
to 7 p.m.), as shown in Fig. 4(b). Buffer time values were also
in Fig. 2(a). Higher ATTs may also occur because of higher travel
observed to be high for these two time segments on Segments 2
times, delays, and dwell times. On the other side, the lowest ATT
and 3, thereby confirming the higher travel time variability and
(47.3 min) is observed to occur from 5 to 6 a.m., which can be
lesser reliability. The same observation is also noticed in the
considered as the off-peak hour. Buffer time is calculated as the
GEV distribution analysis of route-level data for different hours
difference of the 95th percentile travel time and ATT. The highest
of the day and days of the week. In the case of Segment 3, the
BT is observed to be approximately 83.08 min from 8 to 9 a.m. This
BT trend is not exactly parallel to variation in k-value. This clearly indicates wide variations in travel time during this hour of
may be attributed to the delay-causing roadway elements, present the day. However, ATT is observed to be higher from 6 to 7 p.m.
in this segment, resulting in randomness of travel times. In the case This implies that higher ATTs may not always replicate higher BTs
of Segment 1, the BT value is observed to range from 0.3 to because the 95th percentile travel time is also involved in the cal-
1.5 min. For Segments 2 and 3, BT ranges from 1.5 to 3.5 min. culation of BT. For a given ATT, BT or BTI increases with an in-
During peak hours in the morning and evening (9 to 11 a.m. crease in the 95th percentile travel time. This clearly proves the
and 6 to 8 p.m., respectively), BT values are observed to be importance of travel time reliability measures in capturing the travel
approximately 1.5, 3.5, and 4.5 min for Segments 1, 2, and time variations as compared with ATT. The lowest BT is observed
3, respectively. Thus, the results obtained substantiate the signifi- to be 30.18 min from 7 to 8 p.m. A higher deviation (28.56 min) of
cance of k in predicting BT values and the BT values further can be travel time is observed to occur from 5 to 6 p.m. This deviation is
used as input for developing travel time variability or reliability high because of large random variations in the travel time during
models. this evening hour

Travel Time Reliability: Route Level 95th percentile travel time − ATT
BTI ¼ × 100 ð1Þ
ATT
The free-flow travel time is considered as the minimum travel time
observed from the available data sets of bus trips during off-peak 95th percentile travel time
hours. Travel time data for complete operational hours are consid- PTI ¼ ð2Þ
Free − flow travel time
ered for the analysis so that a wide range of variations in passenger
demand as well as traffic conditions are captured. The free-flow
travel time along the bus route is measured as 33.73 min for cal- ATT
TTI ¼ ð3Þ
culating PTI and travel time index (TTI) for different hours of the Free − flow travel time

© ASCE 04018012-5 J. Transp. Eng., Part A: Syst.

J. Transp. Eng., Part A: Systems, 2018, 144(5): 04018012

 Reliability-Based Classification: Segment Level
The present study also attempts to develop a methodology to pro-
pose a suitable reliability indicator as the LoS. For this purpose, the
entire route is segregated into different segments, each 1-km long.
The spatial as well as temporal variability in travel times may not be
efficiently captured based on the entire route travel time data under
varying flow conditions, particularly for longer routes (for example,
30 km in this case). In other words, route-level aggregation may
average out the effect of travel time variability. Hence, route-level
data aggregation may not help in capturing the actual travel time
variations, which can be used for classification purposes or
reliability-based LoS. On the contrary, the variation in travel times
for each segment (disaggregate level) under varying roadway and
Downloaded from ascelibrary.org by University Of Sydney on 03/11/18. Copyright ASCE. For personal use only; all rights reserved.

traffic conditions during different periods of the day can be more

appropriate. Hence, this classification technique can be used for
Fig. 5. Hourly variation of BTI for the 19B route further development of reliability-based LoS thresholds. Keeping
this in view, the present study attempts to capture the effect of
travel time variability through segment- and link-based reliability
The BTI, PTI, and TTI are calculated using Eqs. (1)–(3), as analysis.
defined by Federal Highway Administration (2006). Graphs are For this purpose, ideally the segment-level data should effec-
plotted to examine variations in BTI, PTI, and TTI across the day tively consider the variability based on hour of the day, day of
for operation hours of the bus route. The BTI variation (%) across the week, week of the month, and month of the year. However,
different hours of the day is presented in Fig. 5 as an example. as a preliminary experiment, the present study considers the
The highest BTI (103.5%) is observed for the hours of 8 to segment-based travel time data for 15 days. The hour of the day
9 a.m. and 10 to 11 a.m. during the morning period. In the evening and day of the week have been considered for capturing the varia-
period, BTI is observed to be highest (84.93%) for the hour of 5 to tion in travel times of various segments. In other words, the data for
6 p.m. The lowest BTI (36.71%) is estimated for 6 to 7 and 7 to the classification purpose should include the best possible mix of
8 p.m., as shown in Fig. 5. The characteristics of BT and BTI seem sources of fluctuations in various factors that affect reliability. For
to be similar in nature, in general. However, BT is cushion time or achieving this, travel times for the same segment for a given hour in
buffer time and BTI is an index (in percentage) calculated as the both peak and off-peak conditions for different days of week are
ratio of BT to ATT. For example, when ATT and the 95th percentile considered for clustering analysis. Morning and evening peak time
travel time are on the higher side, BT may be lower and BTI may be periods based on the primary survey are found as 9 to 11 a.m. and 5
higher (because of ATT in denominator). Individual values of the to 7 p.m., respectively. The off-peak time period is observed as 2 to
95th percentile travel time and ATT are lower comparatively for 4 p.m. The segment travel time data of two weekdays (Monday and
the hours of 5 to 7 a.m., but the difference of these values (BT) is Wednesday) and one weekend day (Saturday) for each of the weeks
observed to be slightly higher than that of 7 to 8 p.m. The BTI is (hence, six days for two weeks) in each of the time periods are
observed to be approximately 94% in the case of the morning hour considered for the analysis. Thus, selected travel time data for
of 5 to 6 a.m. This is because of the governance of lower ATT value all segments for different hours of the day and days of the week
(in the denominator) in the calculation of BTI. This implies that for are aggregated and considered for clustering analysis. All reliability
the same value of BT for any two hourly time intervals (with differ- performance measures such as ATT, COV, BT, PT, BTI, PTI, and
ent 95th percentile travel times and ATTs), there can be different TTI are calculated for each individual segment for different hours
values of BTI depending on the magnitude of ATT. Buffer time can of the day. Also, the effect of volume variations on reliability is
be lower in the case of off-peak hours. However, BTI can be higher analyzed by considering variation in the reliability indicators with
because of a lower ATT in the denominator. The PTI is highest respect to V=C ratio. Various steps followed for the development
(4.98) for the hour of 5 to 6 p.m. and lowest (2.72) for 5 to of a generic model for calculating V=C ratios are depicted in the
6 a.m., reflecting the effect of traffic volume levels during these flowchart, presented in Fig. 6.
hours of a day. Volumes for each segment over a time period are indirectly mea-
Similar plots are also generated for PTI and TTI to examine the sured as a function of the average journey speed of the bus. Average
variation for different hours of the day and days of the week. The journey speed of the bus was estimated using ATT. The estimated
PTI value is observed to be higher for the hour of 8 to 9 a.m. (4.85) speed was then correlated with the mixed traffic stream speed, re-
and 5 to 6 p.m. (4.93). The lowest PTI is observed for the hour of 5 sulting in the equation shown in Eq. (4). A linear relationship be-
to 6 p.m. (2.72) and 6 to 7 a.m. (2.78) in the morning and 7 to 8 p.m. tween traffic stream speed (as dependent variable) and bus speed
(3.01) in the evening. The highest TTI is also observed for the hours (as independent variable) is adopted based on variable significance,
of 8 to 9 a.m. (2.39) in the morning and 5 to 6 p.m. (2.69) in the F-value, and coefficient of determination value. Speed flow data,
evening. The lowest TTI values are observed for the morning hours and hence capacity of the same bus route section, are considered
of 5 to 6 a.m. (1.4) and 6 to 7 a.m. (1.7). Thus, the least reliability is from a study by Arasan and Vedagiri (2012). Few data sets (60%)
observed during the morning hour of 8 to 9 a.m. and 5 to 6 p.m., of bus journey speeds are used for calibrating this relationship with
while the highest reliability is observed in the case of the evening stream speed, whereas the remaining data sets (40%) are predicted
hour of 6 to 8 p.m. (in terms of BTI). In terms of PTI and TTI, from the model. In this case, the linear functional form is found to
reliability is observed to be higher during the morning period of be the best fit. After validation, stream speeds are calculated based
5 to 7 a.m. For the classification of peak and off-peak hours during on the developed linear model. Further, for volume and hence V=C
a day, BT, PTI, and TTI are more appropriate to consider compared estimation, available speed-flow relationship for a representative
with BTI. segment as shown in Fig. 6

© ASCE 04018012-6 J. Transp. Eng., Part A: Syst.

J. Transp. Eng., Part A: Systems, 2018, 144(5): 04018012

Downloaded from ascelibrary.org by University Of Sydney on 03/11/18. Copyright ASCE. For personal use only; all rights reserved.

Fig. 6. Volume-to-capacity calculation (image by Shriniwas Arkatkar)

as the dependent variable for forming clusters using different

Stream speed ¼ 1.1784 × Bus speed þ 6.0127 ðR2 ¼ 0.96Þ
reliability-based performance indicators such as ATT, BT, BTI,
ð4Þ PT, PTI, and TTI. The presented generic model has certain limita-
tions of not accounting for the effect of dwell times at bus stops
Using the available speed-flow relationship (parabolic func- directly. However, in the absence of traffic volume data for all
tion), corresponding volumes, and hence V=C ratios, are deter- segments and dwell time data at different bus stops, this model can
mined using stream speeds obtained by employing Eq. (4). Then, be used as a formative tool. The consideration of dwell times and
V=C ratios are calculated for each individual segment (total of quantification of various factors of delay for V=C estimation may
56 segments) using varying travel times. The V=C ratio is then used be treated as the future scope of the present study.

© ASCE 04018012-7 J. Transp. Eng., Part A: Syst.

J. Transp. Eng., Part A: Systems, 2018, 144(5): 04018012

 Validation of Reliability Models From Fig. 7, it can be observed that there is strong correlation
The generic model developed for predicting volume needs to be between 95th percentile travel time and V=C as well as BT and
validated by examining its efficiency. Travel time data of the se- V=C because the coefficient of determination (R2 ) is observed
lected period of the day and day of the week (considered during to be acceptable. The BTI has a weak correlation with V=C because
model development) for 1 month (January 2016) are collected. there can be the same set of BTI values for different values of
Segment-level travel time data are extracted and different reliability 95th percentile travel time and ATT. Though PTI and TTI are also
measures are calculated from travel time data sets. Then, developed observed to have good correlation with V=C, the definition of
generic model relationships are used for predicting volume based free-flow travel time is not appropriate in the case of bus routes
on bus speed. Based on best-fit relationships during model develop- in general.
ment as shown in Fig. 7, reliability measures (95th percentile travel The observed and predicted values of reliability measures are
time, BT, BTI, PTI, and TTI) for a given V=C ratio are predicted. compared, as presented in Fig. 8, and the root-mean-square error
These predicted reliability measures (using the generic model) are (RMSE) is estimated as shown in Table 1. It can be observed from
compared with those of actual or calculated reliability measures Fig. 8 that the 95th percentile travel time and BT have a strong
using travel time data. For developing the relationships, V=C data correlation because predicted values are close to observed values.
Downloaded from ascelibrary.org by University Of Sydney on 03/11/18. Copyright ASCE. For personal use only; all rights reserved.

sets of values greater than 0.6 are considered. The volume predic- Further, the calculated RMSE values are low. This further proves
tion models have a basic assumption that a bus is present in the the efficiency of these reliability measures in capturing the travel
traffic stream and there is some correlation of bus speed with stream time variability.
speed. Because the study route is located in one of the densest met-
ropolitan cities of India and passes through major activity areas, Scatterplots Using k-Means Clustering
a minimum V=C ratio of 0.5 can be reasonably expected at any Further, an entire data set is classified into various numbers of clus-
instant. These models may perform well during bus operational ters using the k-means clustering technique. The k-means cluster-
hours because the presence of a bus in the stream forms a basis for ing technique is based on the minimization of the Euclidean
volume prediction. distance (distance from a data point to the cluster mean) based

Fig. 7. Reliability measures versus V=C

© ASCE 04018012-8 J. Transp. Eng., Part A: Syst.

J. Transp. Eng., Part A: Systems, 2018, 144(5): 04018012

Downloaded from ascelibrary.org by University Of Sydney on 03/11/18. Copyright ASCE. For personal use only; all rights reserved.

Fig. 8. Predicted versus observed reliability measures

on some random iterations. In the present study, k-means clustering of route-based data, the maximum values of COV and V=C from
was performed for a total of 250 iterations using the Euclidean dis- the data set are found to be approximately 0.4 and 0.5, respectively.
tance method. Clustering analysis was performed for four different On the other hand, the same values are observed to be approxi-
cases of number of clusters, namely, two clusters, three clusters, mately 0.8 and 0.96 based on segment-based travel time data.
four clusters, and five clusters. The decision on consideration of Hence, due to the availability of a wider range of COV and V=C
the number of clusters mostly depends on the variation of different ratio, it can be considered adequate to develop LoS thresholds
variables available in the data set. If the range of variation is not based on disaggregated (segment-level) travel time data. Before-
sufficient, then a different number of possible clusters may be hand, it is imperative to study the effect of the individual reliability
attempted to classify data, targeting minimum or zero overlapping measure and also its ability to explain the variation in travel times.
data points between adjacent clusters. On the other hand, the de- This will enhance understanding for adopting the most effective
sired number of classifications (known number of clusters based on reliability indicator for classification of LoS thresholds.
literature review) may be adopted, if it can be confirmed that the Scatter diagrams are plotted considering the COV as a depen-
data set comprises a wide range of variation under varying roadway dent variable, classification number as a categorical variable, and
and traffic conditions. The optimum number of clusters for any all other potential reliability indicators (ATT, BT, PT, BTI, PTI, and
given data set can be determined using various widely used meth- TTI) as independent variables. The k-means clustering technique
ods such as the elbow method, silhouette analysis, and gap statistic is adopted for classifying potential travel time reliability indicators
method. Initially, route-based travel time data are considered for the such as BT, BTI, PTI, and TTI along with ATT in relation to the
clustering analysis. From the analysis, it is observed that in the case COV and V=C obtained using travel time data of each segment.
Scatterplots for different measures such as ATT, BT, and TTI with
respect to COV using different possible cluster numbers are de-
Table 1. RMSE Values for Different Reliability Measures picted in Figs. 9(a and b), respectively.
From Fig. 9(a), it can be observed that for the 2-cluster case,
Specimen number Reliability measure RMSE (min) ATT could not classify COV distinctly because the data are scat-
1 95th percentile travel time with V=C 0.14 tered, and hence, clear classification of clusters is not possible.
2 BT with V=C 0.10 Further, from COV versus the 95th percentile travel time, BT, and
3 BTI with V=C 2.73 PTI plots, the clusters can be distinguished sufficiently to classify
4 PTI with V=C 0.05 the data when compared with ATT. Similar trends are also observed
5 TTI with V=C 0.04
in the case of plots developed using three clusters, as shown in

© ASCE 04018012-9 J. Transp. Eng., Part A: Syst.

J. Transp. Eng., Part A: Systems, 2018, 144(5): 04018012

 0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

CV
CV

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1
Downloaded from ascelibrary.org by University Of Sydney on 03/11/18. Copyright ASCE. For personal use only; all rights reserved.

CLUSTER 2 CLUSTER 3
0.0 CLUSTER 1 0.0 CLUSTER 2
0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14 CLUSTER 1
(a) 95th % TT (b) 95th % TT

1.2 1.2

1.0 1.0

0.8 0.8

0.6
VC

0.6
VC

0.4 0.4

0.2 0.2

CLUSTER 2 CLUSTER 3
0.0 CLUSTER 1 0.0 CLUSTER 2
0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14 CLUSTER 1
(c) 95th % TT (d) 95th % TT

Fig. 9. (a) COV for 2-cluster case; (b) COV for 3-cluster case; (c) V=C for 2-cluster case; (d) V=C for 3-cluster case

Fig. 9(b). This further emphasizes the fact that ATT cannot capture appropriate in the case of private vehicles, but not appropriate for
the variation in travel time exclusively and thus indicates the transit-based systems. Similar results are also obtained in the cases
necessity and significance of travel time reliability indicators. Sim- of four and five clusters. Hence, it may be inferred that BTI and
ilarly, BT and BTI plots are found to classify data and depict clearly BT may be considered as the most suitable travel time reliability
distinguishable clusters with minimum overlapping or scattering indicators for classification, and therefore also for developing LoS
effect. This implies that these reliability measures are more suitable thresholds.
for capturing the variations in travel time with respect to COV. This The effect of variation in volumes over space and time on reli-
can be attributed to the fact that BT and BTI are sufficient because ability is a major concern in any reliability study. Keeping this in
the effects of both 95th percentile travel time and ATT are consid- view, scatterplots are plotted considering V=C ratio as a dependent
ered simultaneously. For the same value of ATT, there can be differ- variable in place of COV, as shown in Figs. 9(c and d). It can be
ent values of standard deviations, and hence different COV values. observed that the 95th percentile travel time and BT plots have less
This is the major cause of the overlapping effect in the case of COV scatter and hence these can be considered as potential reliability
versus ATT plots. Average travel time represents the 50th percentile measures for reliability analysis in the case of transit-based sys-
travel time only and so 50% of the trips may be accomplished tems. From Fig. 9(c), it can be observed that the first cluster has
within this travel time. On the other hand, reliability indicators such a boundary V=C value of approximately 0.8. This implies that the
as BT represent the additional travel time that can be planned in LoS does not deteriorate until the volume reaches 80% of the
addition to ATT so that 95% of the trips can be accomplished capacity. This value seems to be higher in the case of urban roads
within this planned travel time. The 95th percentile travel time is in India because design V=C is considered as 0.7. In the case of
significant in the calculation of BT, and hence both these measures three clusters, the threshold value of the first cluster is observed
are able to capture the variation of COV. In the case of PTI and TTI, to be approximately 0.5 as shown in Fig. 9(d). This value seems
the individual effect of 95th percentile travel time and ATT with to be logically acceptable because of the fact that reliability has
respect to free-flow travel time is better captured. The free-flow an effect on system performance as traffic reaches 50% of its capac-
travel time of a bus during the free-flow conditions also depends ity or the traffic state at which volume reaches design V=C of that
on dwell times, unlike private vehicles. Hence, PTI and TTI may be roadway facility.

© ASCE 04018012-10 J. Transp. Eng., Part A: Syst.

J. Transp. Eng., Part A: Systems, 2018, 144(5): 04018012

 Clearly, from both the cases of 2-cluster and 3-cluster analysis, Table 2. Silhouette Values for COV and V=C
the final cluster boundary V=C values are observed to be in the Serial Silhouette Silhouette
range of 0.8–0.85. When the V=C value is greater than 0.8, small number Cluster case value for COV value for V=C
variations in V=C can cause higher variations in BT and PTI. This is
1 Two clusters 0.57 0.38
the point at which reliability worsens and travel time variability is at 2 Three clusters 0.5 0.36
its maximum. In other words, this is the threshold V=C value (field 3 Four clusters 0.06 0.01
value) that is equal to or slightly greater than design V=C of any 4 Five clusters 0.23 0.02
roadway facility.

Silhouette Analysis
Silhouette analysis gauges the dissimilarity of each data set in its the optimum number of clusters are two or three because silhouette
own assigned cluster and other clusters, in which the corresponding values are greater. For the V=C case also, the optimum number of
data set is not present. The silhouette value is such a parameter that clusters can be either two or three because the silhouette value is
identifies this dissimilarity and is calculated for each individual higher comparatively
Downloaded from ascelibrary.org by University Of Sydney on 03/11/18. Copyright ASCE. For personal use only; all rights reserved.

data set. Thus, the silhouette value of each data set signifies how
well each data set is assigned in its own cluster and how different bðiÞ − aðiÞ
Silhouette value½sðiÞ ¼ ð5Þ
it is when compared with the data sets in other clusters (Rousseeuw max½aðiÞ; bðiÞ
1987). The silhouette value generally ranges from −1 to þ1.
Higher silhouette values (greater than 0 and close to þ1) imply that where i = each data set; aðiÞ = average dissimilarity of i with all
data clusters are well separated and lesser values (less than 0 other data within same cluster; and bðiÞ = lowest average dissimi-
and close to −1) imply that data clusters are extremely slackly larity of i to any other cluster of which i is not a member.
separated. Silhouette values are calculated for both COV and Hence, the optimum number of clusters can be considered as
V=C cases and graphs are plotted as shown in Figs. 10(a and b), three for the present study based on the available variation in the
respectively. Silhouette values greater than 0.5 may be consid- data sets considered. Based on the scatter plots of effective reliabil-
ered sufficient for deciding the optimum number of clusters ity measures with respect to COV and V=C, thresholds of different
(Rousseeuw 1987). clusters are identified as shown in Tables 3 and 4.
The silhouette value is calculated for every segment as shown in
Eq. (5). The average silhouette value is calculated for the different
cases of various numbers of clusters with respect to COV and V=C. Practical Application
The average silhouette values in the case of COVand V=C are given It is extremely important to understand the significance of reliabil-
in Table 2. From Table 2, it can be inferred that for the COV case, ity studies and evaluate the performance of bus routes based on the
results derived from those studies. Reliability studies enable the
operators for designing, planning, and evaluating operational per-
formance of bus routes over space and time. This study attempts
to provide an illustrative case study for evaluating the reliability
of a bus route based on GPS data collected from the same route
for a different time period. For this purpose, travel time data of all
Fridays and Saturdays of a 1-month period are considered for the
illustration of the case study. The overall methodology for the
reliability-based evaluation of the bus routes is presented in Fig. 11.
Travel time data for the selected bus routes can be collected from
field surveys or from GPS logged systems. The travel time data of
each segment can be segregated based on hour of the day and day
of the week. For a given hour of the day and day of the week, the

Table 3. Cluster Thresholds with Respect to COV

95th percentile
Coefficient travel time per BT per
Level of service of variation kilometer (min) kilometer (min)
Cluster 1 <0.3 <2.5 <1.5
Cluster 2 0.3–0.4 2.5–6 1.5–4.0
Cluster 3 >0.4 >6 >4.0

Table 4. Cluster Thresholds with Respect to V=C

95th percentile travel BT per
Level of service V=C time per kilometer (min) kilometer (min)
Cluster 1 <0.5 <2.0 <1
Cluster 2 0.5–0.8 2.0–4.5 1–2.5
Fig. 10. Silhouette plots for (a) COV; (b) V=C Cluster 3 >0.8 >4.5 >2.5

© ASCE 04018012-11 J. Transp. Eng., Part A: Syst.

J. Transp. Eng., Part A: Systems, 2018, 144(5): 04018012

Downloaded from ascelibrary.org by University Of Sydney on 03/11/18. Copyright ASCE. For personal use only; all rights reserved.

Fig. 11. Reliability-based evaluation: a case study

descriptive statistics and reliability indexes can be calculated for Concluding Remarks
a given segment of interest or consideration. The cluster in
From the results of GEV distribution, it can be presumed that peak-
which calculated reliability measures fall according to thresholds
hour travel times can be modeled using normal distribution and off-
presented in Table 3 can be identified, and hence LoS can be
peak travel times can be modeled using GEV distribution. These
defined.
distributions are to be further established by carrying out analysis
For illustration purposes, travel time data of the 19B route of
pertaining to different roadway and traffic conditions.
Chennai, India, pertaining to certain days of the week (Friday and
This research presents an original methodology of volume pre-
Saturday) for a duration of approximately 30 days are considered.
These data consisted of segmented travel time data samples for the diction using bus speed in the absence of automated (online or
bus trips extracted from GPS-enabled data loggers. Hourly travel offline) traffic volume data from the field. This method can be rea-
time variations of all the days of the week for the considered du- sonably expected to be applicable to the bus routes under varying
ration are plotted as shown in Fig. 12. A representative 1-km seg- roadway and recurrent traffic flow conditions. For a given set of
ment is considered and the travel time data of this segment are roadway and geometric conditions, unique relationships can be de-
segregated based on the hour of the day (2-h intervals). Descriptive veloped and validated with the help of travel time data collected
statistics as well as reliability indexes per unit distance for each over the same route for a different time period. This approach will
hour of the day are calculated based on the obtained data samples help in the prediction of reliability measures for a given volume
as shown in Table 5. Based on Table 4, the clusters under which or V=C, which can save an enormous amount of time and staff
these calculated reliability measures fall are identified for each time in travel time data collection and extraction. Regression models
interval (such as 6–8, 12–14, and so on). Thus, the bus route for a involving V=C as an independent variable for predicting reliability
given segment can be evaluated based on reliability measures. The measures are developed for V=C values greater than 0.6. This is
temporal variations are highlighted and presented in Table 5. In a more pragmatic because travel time variability or reliability be-
similar way, different segments for a given duration or hour of the comes a critical matter of concern once the flow approaches
day can also be evaluated so as to understand the spatial travel time V=C of approximately 0.6 in the case of urban arterial roadway
variations. conditions.

© ASCE 04018012-12 J. Transp. Eng., Part A: Syst.

J. Transp. Eng., Part A: Systems, 2018, 144(5): 04018012

Downloaded from ascelibrary.org by University Of Sydney on 03/11/18. Copyright ASCE. For personal use only; all rights reserved.

Fig. 12. Hourly variation of route-level travel time

Table 5. Reliability Measures and Classification for Selected Segment

ATT per 95th percentile Cluster based on BT per Cluster
Hour of the day kilometer (min) travel time per kilometer (min) 95th percentile travel time kilometer (min) based on BT
6 to 8 a.m. 1.98 2.42 Cluster 1 0.43 Cluster 1
8 to 10 a.m. 2.28 4.23 Cluster 2 1.94 Cluster 2
10 a.m. to 12 p.m. 2.22 3.91 Cluster 2 1.69 Cluster 2
12 to 2 p.m. 2.57 4.1 Cluster 2 1.52 Cluster 2
2 to 4 p.m. 2.68 4.78 Cluster 2 2.09 Cluster 2
4 to 6 p.m. 2.36 5.08 Cluster 2 2.72 Cluster 2
6 to 8 p.m. 3.72 8.49 Cluster 3 4.76 Cluster 3
8 to 10 p.m. 2.1 3.05 Cluster 2 0.94 Cluster 1

This study thus highlights the methodology for the identification includes the overall journey time. Depending on the data set type
of effective reliability measures from scattering technique. This and quantity, it may be difficult to consider all the factors affecting
adds sensible novelty to the current state of the art in the area reliability every time simultaneously. More studies are required in
of travel time reliability. Reliability-based LoS forms a trustworthy this direction to account for the effect of all these various factors for
basis for prioritizing the road links or road segments in a road or a better understanding of their role on reliability. Data sets of differ-
route network under prevailing traffic conditions. Thus, it can be ent routes (based on land use, road geometry, and side frictions) for
very helpful for operating agencies and also transport planners longer periods under varying roadway and traffic conditions can be
when allocating resources. Further, it also helps in the enhancement aggregated and disaggregated for unflawed classification purposes.
of the mobility and traveler convenience, and overall results in cost This will further lead to the better characterization of reliability-
savings. This study focuses on the variation of a few selected reli- based LoS thresholds in an efficient way.
ability measures with respect to parameters of the best-fitted stat-
istical distribution for travel time. Nevertheless, it provides valuable
insights for the bus travel time modeling using statistical distribu- Acknowledgments
tions under heterogeneous traffic conditions. The case study pre- The authors extend their heartiest thanks to the Center of Excel-
sented in this work is expected to be useful for transit operators lence in Urban Transport at IIT Madras, India, for sharing the travel
and planners so as to evaluate the performance of different links or time data.
segments along a given bus route in terms of reliability. This will
also form a strong basis for the optimization of schedules according
to the actual travel time variations occurring in the field. References
The effect of various roadway characteristics like abutting land
use, bus stop locational characteristics, and road-side impedance, Arasan, V. T., and Vedagiri, P. (2012). Planning for priority for public
and operations characteristics like headway irregularities, incident transport: Providing exclusive bus lanes on roads carrying hetero-
occurrence, and scheduled service reduction are not considered in geneous traffic, LAP Lambert Academic Publishing, Saarbrücken,
Germany.
the present scope of this research work. The individual effect of
Asakura, Y., and Kashiwadani, M. (1991). “Road network reliability caused
dwell times and intersections on travel time are not explicitly quan- by daily fluctuation of traffic flow.” Proc., 19th PTRC, Summer Annual
tified in the present study. However, their effect is considered indi- Meeting, Univ. of Sussex, Brighton, U.K., 73–84.
rectly (implicitly) because travel time data over space and time used Bharti, A. K., Chandra, S., and Chalumuri, R. S. (2013). “Performance
in this study are derived from GPS-based bus trajectory, which also evaluation of urban arterial in Delhi using travel time reliability.” Proc.,

© ASCE 04018012-13 J. Transp. Eng., Part A: Syst.

J. Transp. Eng., Part A: Systems, 2018, 144(5): 04018012

 Eastern Asia Society for Transportation Studies, Vol. 9, Springer, Liu, H. X., Recker, W., and Chen, A. (2004). “Uncovering the contribution
Berlin. of travel time reliability to dynamic route choice using real-time loop
Chalumuri, R. S., and Asakura, Y. (2007). “Measuring travel time reliability data.” Transp. Res. Part A, 38(6), 435–453.
of transportation system.” Indian Highways, 35(11), 61–72. Lyman, K., and Bertini, R. (2008). “Using travel time reliability measures
Chalumuri, R. S., Kitazawa, T., Tanabe, J., Suga, Y., and Asakura, Y. to improve regional transportation planning and operations.” Transp.
(2010). “Examining travel time reliability on Hanshin Expressway Res. Rec., 2046, 1–10.
network.” J. Eastern Asia Soc. Transp. Stud. (EASTS), 7, 2274–2288. Mehbub Anwar, A. (2010). “A study on factors for travel time variabil-
Chen, C., Skabardonis, A., and Varaiya, P. (2003). “Travel time reliability ity in Dhaka City corporation area.” J. Bangladesh Inst. Planners, 3,
as a measure of service.” Transp. Res. Rec., 1855, 74–79. 53–64.
Clark, S., and Watling, D. (2005). “Modelling network travel time reliabil- Nam, D., Park, D., and Khamkongkhun, A. (2005). “Estimation of value of
ity under stochastic demand.” Transp. Res. Board, Part B: Methodol., travel time reliability.” J. Adv. Transp., 39(1), 39–61.
39(2), 119–140. Polus, A. (1979). “A study of travel time and reliability on arterial routes.”
Duddu, V. R., Pulugurtha, S. S., and Najaf, P. (2014). “New transit perfor- Transportation, 8(2), 141–151.
mance measures and LOS criteria for link-level assessment.” 93rd Pulugurtha, S. S., Puvvala, R., Pinnamaneni, R. C., Duddu, V. R., and
Annual Meeting of the Transportation Research Board, Transportation Najaf, P. (2004). “Buses as probe vehicles for travel time data collection
Downloaded from ascelibrary.org by University Of Sydney on 03/11/18. Copyright ASCE. For personal use only; all rights reserved.

Research Board, Washington, DC. on urban arterials.” 83rd Annual Meeting of the Transportation
Federal Highway Administration. (2006). “Travel time reliability: Research Board, Transportation Research Board, Washington, DC.
Making it there on time, all the time.” U.S. Dept. of Transportation, Puvvala, R., Pulugurtha, S. S., and Duddu, V. R. (2015). “Cronbach’s
Washington, DC. coefficient as a performance measure to assess link-level reliability.”
Gan, H., and Bai, Y. (2014). “The effect of travel time variability on route 94th Annual Meeting of the Transportation Research Board, Transpor-
choice decision: A generalized linear mixed model based analysis.” tation Research Board, Washington, DC.
Transportation 41(2), 339–350. Rousseeuw, P. J. (1987). “Silhouettes: A graphical aid to the interpretation
Higatani, A., Kitazawa, T., Tanabe, J., Suga, Y., Sekhar, R., and and validation of cluster analysis.” J. Comput. Appl. Mathematics, 20,
Asakura, Y. (2009). “Empirical analysis of travel time reliability mea- 53–65.
sures in Hanshin Expressway network.” J. Intell. Transp. Syst., 13(1), Sathya Prabha, R., and Mathew, S. (2013). “Evaluation of operational ef-
28–38. ficiency of urban road network using travel time reliability measures.”
Kieu, L. M., Bhaskar, A., and Chung, E. (2014). “Public transport travel- Int. J. Res. Eng. Technol., 1(2), 155–160.
time variability definitions and monitoring.” J. Transp. Eng., 10.1061 Transportation Research Board. (2000). “Highway capacity manual.”
/(ASCE)TE.1943-5436.0000724, 04014068. Washington, DC.

© ASCE 04018012-14 J. Transp. Eng., Part A: Syst.

J. Transp. Eng., Part A: Systems, 2018, 144(5): 04018012