ALTERNATIVE SEGMENTATION SCHEMES FOR THE DESIGN OF
TRAFFIC MAPS
By
Joshua E. Stevens

A THESIS
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
MASTER OF SCIENCE
Geography
2011

ABSTRACT
ALTERNATIVE SEGMENTATION SCHEMES FOR THE DESIGN OF
TRAFFIC MAPS
By
Joshua E. Stevens
Drivers in urban environments constantly face the problem of congestion. As a geographic hindrance that is growing rapidly in both frequency and severity, congestion
represents a major burden to drivers and transportation systems alike. At the same
time, recent advancements in technology have given cartographers unprecedented access to high-quality, real-time traffic data that is readily employed in traffic maps from
various sources. Previous research has demonstrated that providing drivers with relevant, time-based information reduces environmental uncertainty, yielding benefits to
individuals as well as the transportation network as a whole. Unfortunately, many of
the available traffic maps are variations of a single design approach which is replete
with ambiguity that hinders the task of travel time estimation. This is largely due to
the irregular segmentation of the transportation network used in conventional traffic
map design. To address this issue, two prototype designs are proposed that present
regular, fixed segmentation schemes. The first approach employs segments that have
a fixed length of distance, removing the users’ need to estimate incomplete distances
along the route. The second approach represents travel time directly by fixing segment
length to a unit of travel time. Results of a user-based experiment strongly favor the
regularly-segmented approaches for the purpose of travel time estimation and suggest
that conventional traffic map designs are not adequate for this task. Traffic maps utilizing the fixed-minute segmentation were greatly preferred by users and were associated
with significant improvements in travel time estimation.

Copyright by
JOSHUA E. STEVENS
2011

ACKNOWLEDGMENTS
Foremost, I would like to thank Dr. Kirk Goldsberry for his guidance during the time
leading to this thesis. He continuously challenged me to think and to welcome obstacles
of all sorts. This experience has had a profound influence on me both as a geographer
and as a person. I would also like to thank Dr. Judy Olson for always asking the
tough questions and forcing me to be skeptical of even the most obvious things. Equal
appreciation is given to Dr. Ashton Shortridge for his advice and generous assistance.
The accomplishments I have made are all in great part due to the incredible amount
of support each of you have given me over the last two or more years. Sincerest thanks
are given to the Department of Geography for providing this incredible opportunity. I
also thank my friends and fellow students for all that they have done for me. Thanks
to Kristie for your help and advice. Thanks to Carolyn, being in your lab was a great
motivator for me. You were an awesome mentor, even if you didn’t realize you had
taken on that role. To Paul, David, and Mark, thank you all for the laughs and good
times. Nicole, I could never thank you enough for the encouragement you’ve given me
over the years. This thesis could not have been done without the unending support
from my family. I truly appreciate all of you!

iv

Table of Contents
List of Tables

vii

List of Figures

viii

Chapter 1: Introduction
1.1 Research Goals . .
1.2 Research Questions .
1.3 Hypotheses . . . .
1.4 Research Relevance

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

1
7
7
7
9

Chapter 2: Literature Review
11
2.1 The State of the Literature on Traffic Maps, Cognition, and Mobile
Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Sensor Placement and Transportation Networks . . . . . . . . . . 13
2.3 ATIS, ITS, and System Performance . . . . . . . . . . . . . . . 15
2.4 Network Data Collection and Reliability . . . . . . . . . . . . . 16
2.5 Spatial Knowledge Construction . . . . . . . . . . . . . . . . . 17
2.6 Classed vs Unclassed Data . . . . . . . . . . . . . . . . . . . 19
2.7 Human Vision and Image Properties . . . . . . . . . . . . . . . 19
2.8 Attention, Mental Workload, and Driver Performance . . . . . . . 23
2.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Chapter 3: Methodology
31
3.1 Overview and Goals . . . . . . . . . . . . . . . . . . . . . . 31
3.2 An Evaluation of Conventional Traffic Maps . . . . . . . . . . . 31
3.3 Data Acquisition and Creation of the Equal-interval Design . . . . . 33
3.4 Creation of the Fixed-minute Design . . . . . . . . . . . . . . . 37
3.5 Experimental Design . . . . . . . . . . . . . . . . . . . . . . 40
3.5.1 Travel time Arithmetic Questions . . . . . . . . . . . . . 42
3.5.2 Map-based Distance Estimation . . . . . . . . . . . . . . 43
3.5.3 Travel time Estimation . . . . . . . . . . . . . . . . . . 44
3.5.4 User Confidence . . . . . . . . . . . . . . . . . . . . 46
3.6 Expected Results . . . . . . . . . . . . . . . . . . . . . . . . 47
3.6.1 Arithmetic Ability . . . . . . . . . . . . . . . . . . . . 48
3.6.2 Equal-interval Segmentation and Distance Estimation Expectations . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.6.3 Travel time Estimation and Segmentation Schemes . . . . . 49
3.6.4 Map Preference . . . . . . . . . . . . . . . . . . . . . 50
3.7 Participants . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.8 Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.8.1 Hardware . . . . . . . . . . . . . . . . . . . . . . . 51
3.8.2 Software . . . . . . . . . . . . . . . . . . . . . . . . 52
3.9 Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

v

Chapter 4: Results and Analysis
4.1 Arithmetic . . . . . . . . . . . . .
4.2 Distance Estimation . . . . . . . . .
4.3 Travel Time Estimation . . . . . . . .
4.4 Map Preference and Subjective Feedback

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

Chapter 5: Discussion
5.1 Arithmetic Ability . . . . . . . . . . . . . .
5.1.1 Error and Accuracy . . . . . . . . . .
5.1.2 Confidence . . . . . . . . . . . . . .
5.1.3 Potential Bias . . . . . . . . . . . . .
5.2 Distance Estimation . . . . . . . . . . . . .
5.2.1 Error and Accuracy . . . . . . . . . .
5.2.2 Response Time . . . . . . . . . . . .
5.2.3 Confidence . . . . . . . . . . . . . .
5.3 Travel Time Estimation . . . . . . . . . . . .
5.3.1 Error and Accuracy . . . . . . . . . .
5.3.2 Response Time . . . . . . . . . . . .
5.3.3 Confidence . . . . . . . . . . . . . .
5.3.4 Map Preference and Subjective Feedback
5.4 Limitations . . . . . . . . . . . . . . . . .
Chapter 6: Conclusion

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

54
54
57
61
65

.
.
.
.
.
.
.
.
.
.
.
.
.
.

68
68
68
68
68
69
69
69
70
70
70
73
74
75
76
78

Appendices
82
Appendix A: Python Code for Identifying 1-minute Segments . . . . . . . 82
Appendix B: Advertisement for Soliciting Participants . . . . . . . . . . 84
Appendix C: Consent Form . . . . . . . . . . . . . . . . . . . . . 85
References

88

vi

List of Tables
1

The classification scheme employed by Google’s Traffic Maps. . . . .

36

2

Participants were asked a total of 60 questions, 10 from each category.

40

3

Map designs were grouped and compared based on the estimation task

54

4

Magnitude of percent error across all arithmetic questions.

. . . . .

55

5

Percent error from forty-nine participants . . . . . . . . . . . . .

55

6

A comparison of the results from the distance estimation portion . . .

59

7

ANOVA reveals that each result is significantly different . . . . . . .

62

8

The differences in response times . . . . . . . . . . . . . . . . .

64

vii

List of Figures
1

A typical traffic map, as provided by Google Maps. . . . . . . . . .

4

2

The ILD data for this thesis is comprised of 4,236 individual sensors .

14

3

Current traffic map designs from three popular vendors . . . . . . .

32

4

Estimating travel time with current traffic map designs . . . . . . .

33

5

The traffic and sensor configuration data are joined

. . . . . . . .

34

6

Each roadway is split into 5-mile segments

. . . . . . . . . . . .

35

7

The final equal-interval map . . . . . . . . . . . . . . . . . . .

36

8

Each roadway is split into segments with length equal to 52.8 ft

. . .

37

9

The final fixed-minute map . . . . . . . . . . . . . . . . . . .

39

10

Before and after the route fade effect. . . . . . . . . . . . . . . .

42

11

A typical question from the travel time Arithmetic category. . . . . .

43

12

The control and segmented map designs. . . . . . . . . . . . . .

45

13

The control traffic map, with pins indicating one of ten routes . . . .

45

14

User confidence and map error can reveal interesting properties . . .

47

15

Nearly 75% of all responses to arithmetic were within the error window. 56

16

Participants were overall very confident with the arithmetic portion . .

17

Overall, the equal-interval segmentation was associated with lower errors 58

18

The equal-interval map was associated with great improvements . . .

59

19

Almost 80% of responses were confident with the control (distance) . .

60

20

Confidence was dramatically higher with the segmented map . . . .

61

21

Fixed segmentation schemes greatly improve travel time estimations. .

62

22

The fixed-minute design improved the number of correct responses

.

63

23

Response times were also lower with the fixed-minute design . . . . .

64

24

Confidence was improved by both segmented designs . . . . . . . .

65

25

None of the participants chose the Google map as their favorite . . .

66

viii

57

26

Possible strategies for estimating travel time with the prototype designs

72

27

The flyer used to solicit participants. . . . . . . . . . . . . . . .

84

ix

Chapter 1
INTRODUCTION
A major problem facing the United States and indeed the world is the burden of
traffic. Whether it is delaying commuters on their way to work, disrupting the flow of
goods and services, or clogging major cities, congestion is problem that affects millions,
if not billions, of people every day. This burden is also growing rapidly. New drivers
and automobiles are produced by the hour and the number of vehicles on the roadways
increases continuously. Yet the transportation networks remain largely unchanged; they
cannot possibly be modified fast enough to provide a supply that keeps pace with the
ever-growing demand. The end result is congestion that is frequent in both time and
space.
Congestion affects both travelers and freight shippers who “must plan around traffic
jams for more of their trips, in more hours of the day and in more parts of town than
in 1982” (Schrank & Lomax, 2009, p. 10). Furthermore, congestion carries an annual
price tag that is more than $20 billion higher than it was just seven years ago. This
represents an average annual cost of $760, as well as 36 hours spent in delay, for every
driver in the United States. These figures are nearly three times higher than what they
were only three decades ago.
We often think of congestion as a personal inconvenience without much consideration to how large the problem of congestion really is. Schrank & Lomax noted that
congestion has led Americans to spend 4.2 billion additional hours on the road, using
2.8 billion gallons of fuel, which encumbered drivers with a total cost of $87.2 billion
in traffic-related expenses. A problem of this magnitude demands attention and the
drivers who are directly affected by it could benefit greatly from any improvements
made. Among the suite of possible solutions is the accurate and efficient delivery of
real-time traffic data. Al-Deek & Khattak (1998) note that much of the congestion

1

is caused by drivers making uninformed decisions about their routes. Poor decisions
become exacerbated in the presence of existing congestion as drivers become unsure
of when, where, or if they should divert or stay on course. As the number of unsure
drivers increases, system efficiency diminishes and congestion becomes likely to arise
or worsen.
It is clear that a need exists for the efficient and dynamic delivery of useful traffic
information that could help users make decisions while on the road or during pretrip planning. Time-based information is especially crucial. While the phenomenon
of congestion does have a range of monetary and physical costs, the cost of time is
perhaps the number one concern to many drivers. We live in a fast-paced world, full of
deadlines, meetings, and responsibilities. Anything that may prolong or delay us from
reaching our destination represents a cost in time that, if given the opportunity, would
be readily avoided.
Congestion particularly affects commuters in large urban environments who must
deal with traffic-related delays on a daily basis. Given the dynamic, unpredictable nature of congestion, it represents a constantly changing obstacle to commuters. This is
largely due to the uncertainty that is imposed whenever congestion arises. The cognitive maps people build during their everyday travels are full of various waypoints,
routes, and other spatial knowledge. Human beings have an almost instinctive familiarity with the trips they make often. We “know them like the back of our hands.”
Although the waypoints and routes rarely change, the temporal conditions of those
routes often do. The amount of time required to make a trip can change at a moment’s
notice, unknown and unannounced. In these cases, congestion clouds our cognitive
maps; the route remains familiar but the time-cost becomes a mystery, or foggy at best.
Although there are many possible solutions to the problem of congestion, and indeed
the problem is so large that no single solution will suffice (Schrank & Lomax, 2009),

2

providing travelers with high-quality and easily-used traffic information would be substantially beneficial. Even if the improvements were realized by only 1% of drivers in
the United States, the affected population would still number in the millions.
Current technology enables the real-time acquisition of traffic data that can be accessed by a wide array of devices, including radios, GPS units, and mobile phones. Despite these technological advances, millions of drivers still do not have reliable access to
relevant traffic information that could improve their decision-making and aid them in
avoiding congestion. By filling in the temporal gaps in drivers’ cognitive maps and reducing their environmental uncertainty, a greater number of individual drivers would
have the potential to avoid congestion and the resultant delays. At the macro scale, the
combined effect of an increase in well-informed drivers could also yield system-wide
improvements — a consequence that has already been noticed when a large portion of
drivers have enhanced information about their environment (Mahmassani & Jayakrishnan, 1991).
Presently, these enhancements often come in the form of broadcasts made possible by
Advanced Transportation Information Systems (ATIS) and Intelligent Transportation
Systems (ITS). These technologies enable the collection and dispersal of information
about the transportation network that can improve drivers’ decision making, reduce
system-wide congestion, and make travel a safer, more reliable experience (Al-Deek &
Khattak, 1998).
Maps are an increasingly common component of ATIS and in-vehicle navigation;
they also have a long history of attempting to provide information quickly and efficiently. Now, more than ever, maps can take advantage of mobile devices and cell
phones, which are becoming increasingly widespread (Ishigami & Klein, 2009). This
presents an opportunity to reach a large audience and provide real-time traffic information to drivers in an efficient and proven format.

3

Although mobile maps and traffic have been studied in great detail independently
(see Walker et al. (1991); Al-Deek & Khattak (1998); Clarke (2004); Fujii et al. (2009)),
little to no published research exists on traffic maps on mobile devices, or on the cognitive aspects associated with traffic maps alone. This is unfortunate, given the popularity
of mobile devices and the ease with which they can access visual information, such as
maps.

Figure 1: A typical traffic map, as provided by Google Maps. Red, yellow, and green are
used to indicate velocity, a proxy for the severity of congestion. For interpretation
of the references to color in this and all other figures, the reader is referred to
the electronic version of this thesis.

4

The purpose of this thesis is to examine traffic maps within the context of cognition
in a two-fold manner. First, existing research on the in-vehicle use of mobile devices
will be used to guide a critique of current traffic map designs. The results of the critique will provide an evaluation of traffic maps that have not commonly been subjected
to such research. This critique will then facilitate the creation of two new traffic map
designs that will attempt to alleviate some of the problems thus revealed. Secondly,
a conventional and widely used traffic map design (Figure 1) will be compared to the
new alternatives in a user-based experiment, focused on common traffic map usage.
Although there are several reasons one might use a traffic map, for many users, especially those accessing the maps from a mobile device, route selection and travel time
estimation are likely very common.
Traffic maps help even very experienced urban travelers reduce environmental uncertainty and the stresses associated with not knowing time-costs that separate their
origins and destinations. They enable better wayfinding decisions and potentially improve system-wide efficiency. Some of the most common tasks associated with traffic
maps include identifying the fastest route and avoiding congestion, both of which depend on travel time and velocity estimation. Such tasks are very difficult to perform
with conventional traffic maps, which provide ample opportunity for improvement. In
order to utilize a conventional traffic map for selecting the fastest route or identifying
an expected delay, the user must first solve a set of sub-tasks, which include:
1. Estimate the distance of the desired route.
2. Estimate the velocity based on the current traffic conditions.
3. Perform the arithmetic necessary to convert the determined velocity and distance into time.
The difficulty in performing these tasks is increased by the presence of ambiguous

5

segments representing various levels of congestion in the conventional traffic map. A
single route may be composed of several different segments, each representing a different velocity and distance. An effective strategy for computing all this information is
left for the user to figure out. Not only is such a task burdensome in its own right, but
conventional design practices have not done much to improve it. The current traffic
maps rarely reveal the velocities represented by a particular color and often utilize a
straight scale bar that is not congruent with curvy or irregular routes. This significantly
reduces the user’s ability to make accurate time-cost estimates.
In addition to providing insight into how traffic maps might be improved or designed in new ways, this research has implications beyond cartography. In their 2008
report for the National Highway Safety Administration (NHSA), Ascone et al. found
that nearly 6,000 drivers were killed in accidents related to distracted driving, which
accounted for 16% of all fatal crashes that year. The source of distractions in the report include in-vehicle mapping units and cell phones, devices which have been shown
to demand a great deal of the operator’s attention (Walker et al. (1991); Parkes et al.
(1991); Srinivasan (1997); Drews et al. (2009); Horrey et al. (2009); Ishigami & Klein
(2009)). Improved traffic map design will not solve this problem outright. It is however
unlikely that the habitual use of these devices will decrease, therefore any effort that
reduces the distraction they impose is a step in the right direction.
Increasingly, the categories of cell phones and in-vehicle mapping units used by
the NHSA need not be separate, as the technologies continue to converge. In 2009,
Garmin — a popular manufacturer of GPS units — released their first map-enabled
hybrid Smartphone. Google Maps has been a standard application on Smartphones
for years, and the online version is accessible to any mobile phone with an Internet
connection.
Due to the popularity of these mobile mapping applications and their inherent re-

6

lationship to vehicular travel, it is reasonable to assume that traffic maps are included
in the map types used by drivers. It is therefore necessary that an assessment of traffic
maps be done within the context of reducing the attention such maps require, while
maintaining continuity with previous studies on map use and mental workload.

1.1 Research Goals
The goals for this thesis are to:
1. Identify key limitations in conventional traffic map designs.
2. Develop prototype design strategies to reduce the limitations.
3. Test the performance of the prototype traffic map designs against the existing
standard in terms of travel time estimation.

1.2 Research Questions
To reach the goals set for this thesis, the following questions must be addressed:
1. In what ways can current traffic map designs be improved?
2. Can map readers quickly and accurately determine travel time from existing
map designs?
3. Does an improved design, guided by a critique of traffic map designs, improve
the accuracy of users’ travel time estimations and/or the time required to make
the estimations?

1.3 Hypotheses
Hypothesis 1: When provided with a distance and a velocity, most people cannot accurately perform the necessary arithmetic to determine travel

7

time. In order to determine travel time (t), a distance (d) and velocity (v) must be known
(or assumed), and the arithmetic t

=

d
v be performed. Mental arithmetic, especially

division, is a difficult task. If map readers cannot quickly and accurately perform it,
they cannot reasonably be expected to determine travel time from a map that requires
such cognitive ability.

Hypothesis 2: Map readers cannot adequately determine route distance
when using a conventional traffic map. The current traffic map designs in popular use employ a linear scale bar to assist with estimating distance along a nonlinear
transportation network. This incongruity is likely to be a burden to map readers, impeding their ability to accurately estimate the distance of a route. This would greatly
diminish a key component of an accurate travel time estimation.

Hypothesis 3: Map readers cannot accurately estimate travel times from
conventional traffic maps. Current traffic map designs employ hue as the main
identifier of traffic quality. These hues are allocated to segments along the route, each
of which is associated with a different velocity and an unknown length. The combined
task of estimating the length and velocity of these segments is likely to pose a significant
challenge to map readers. Failure to accurately determine a single component of the
travel time equation will result in error, and this calculation must be performed for every individual segment along the route and then averaged for a final value. This task
presents a large opportunity for error propagation, and poor performance is expected.

Hypothesis 4: Traffic maps that are designed to facilitate the common task
of travel time estimation will be associated with a faster, more accurate
estimations. Current visualization strategies are more of an artifact of data collection

8

strategies than they are an attempt to streamline the exchange of traffic information.
Instead of using segments that vary in their representation of distance and velocity, the
segmentation of the transportation network can be fixed to a standard unit the mapreader is more familiar with. A fixed segmentation scheme may take one of two forms:
1) An equal-interval scheme in which the segments are always the same length or 2)
A fixed-duration scheme, where segments represent a constant unit of time. In this
way, the task of distance estimation is made significantly easier, or in the case of the
fixed-duration scheme, it is removed entirely. This cartographic treatment is expected
to improve travel time estimations in terms of accuracy and time-to-completion.

1.4 Research Relevance
The empirical investigation of mapping techniques has been at the center of academic
cartography in the United States for nearly 60 years (see Robinson’s The Look of Maps
(1952)), and our understanding of the map-human relationship has grown tremendously since (MacEachren, 2004). Nevertheless, traffic maps, which are used by millions of daily users, have gone without much study, especially from a cognitive viewpoint. An examination of conventional traffic maps and the potential creation of a
superior alternative would help streamline the flow of much-needed information about
an increasingly troubling geographic phenomenon. Congestion is a dynamic, unpredictable geographic hindrance affecting millions of daily drivers around the world. The
resultant spatial uncertainty and suboptimal wayfinding decisions made by drivers reduces the overall performance of the transportation network, which has substantial
costs in both time and money. Efficient traffic maps offer a potential solution to this
problem.
The relationship between mobile devices and distracted driving is important. Traffic
map generation is not only a likely use for such devices, but also represents an area of

9

cartography that has not been given much empirical consideration. Although many
previous studies have demonstrated that maps affect our attention (Walker et al., 1991;
Parkes et al., 1991; Srinivasan, 1997; Horrey et al., 2009), none have evaluated traffic
maps or the task of travel time estimation.
Furthermore, the cognitive influence of traffic maps is largely unknown. We do not
yet know which aspects of their use are troublesome, or even if they perform perfectly
well. Without an empirical evaluation of traffic maps, cartographers can only speculate. The results from this thesis will shed light on these issues, and attempt to clarify
the source of the errors that are associated with travel time estimation. This would reveal where current maps are performing poorly, as well as provide new techniques and
approaches to map design. This research aims to better our understanding of traffic
map interpretation and address a topic that is long overdue for examination.
The remainder of this thesis will cover a review of relevant literature, which includes
the areas data collection methods, traffic sensor placement, spatial knowledge and cognition, human vision and lastly, mental workload and attention. Following the literature review, the experimental design will be described along with the expected results,
materials used, and procedures employed to both create and test the new prototypes.
Results and analyses will then be provided prior to a discussion on the meaning of the
findings. The last sections will cover the limitations of this thesis and provide a summary concluding the research done.

10

Chapter 2
LITERATURE REVIEW
The call for empirical cartography in the United States can be traced to The Look of
Maps, published in 1952 by Arthur Robinson. In this book, Robinson argued for an objective view towards cartography. In contrast to subjectivity, where maps are critiqued
based on their aesthetic or artistic nature, Robinson laid the foundation for an empirical cartography in which a map’s quality was determined by how well it performed a
given task. This was an important distinction to make, and it remains so today. If a
map is said to be superior to another, it must be demonstrably proven to be so. Before
this idea, cartographic techniques largely followed convention, a practice that laid its
defense on rhetoric and opinion (e.g, red is good for elevation maps because soil is also
red). Unjustified adherence to convention is a practice criticized by Robinson:
“Tradition and familiarity maintain a strong hold. On the other hand,
it is reasonable to expect that the developments in the science of vision and
the spreading appreciation of the importance of design may combine with
the marked increase of interest in cartography in recent years to promote
more critical examination of many of the long standing conventions” (pg
12).
This review attempts to outline the literature that is related to congestion, mobile
maps, and cognition from an objective viewpoint akin to Robinson’s.

2.1 The State of the Literature on Traffic Maps, Cognition, and
Mobile Devices
At present, maps can be found almost everywhere. Where they once hung from walls,
maps now light up electric displays regularly with unprecedented popularity. Recent

11

advancements in technology have brought maps into vehicles and onto mobile devices,
such as cell phones. While they were once luxury items or gadgets owned only by a
tech-savvy few, these devices are now ubiquitous (Ishigami & Klein, 2009). The growth
in the popularity of these devices can likely be attributed to the information age, the
pervasiveness of wireless networks (WiFi), lower costs, and an increased demand for
knowledge on the go.
Because of their habitual use while the car is in motion, these devices also pose a
danger for motorists and pedestrians alike (Walker et al., 1991; Parkes et al., 1991;
Srinivasan, 1997; Drews et al., 2009; Horrey et al., 2009; Ishigami & Klein, 2009). In
a study on the behavioral responses elicited by these devices, Golledge (2002) warns that
“distracting a driver from visually assessing traffic conditions (especially when traffic is
moving at speed) is potentially very dangerous.” As these devices saturate our markets
and become more universal, these safety concerns only grow. This suggests that an
understanding of traffic map design and its affect on human behavior is critical (Walker
et al., 1991).
Given the apparent need for such research and the proliferation of devices capable
of displaying traffic maps, we might expect the literature to be replete with study after
study outlining the effects of their use. Unfortunately, this is not the case and others
have noted that research in this area has been surprisingly scarce (Walker et al., 1991;
Fujii et al., 2009).
Traffic mapping represents the pairing of many technologies and fields. The rest
of this review will reflect these different components, starting with the transportation
networks themselves, Advanced Traveler Information Systems (ATIS) and Intelligent
Transport Systems (ITS) strategies that inform drivers, and data reliability. Lastly, human behavior and spatial knowledge construction will be discussed alongside the topics
of cognition, mental workload, and map interpretation.

12

2.2 Sensor Placement and Transportation Networks
Before addressing traffic maps specifically, it is important to understand how they are
created and how the data are collected. Liu et al. (2009) explain that in the U.S., traffic data are often collected from sensors beneath the roadway called Inductive Loop
Detectors, or IDLs (Liu & Danczyk, 2009). The State of California, which served as
the study area for this research, employs two basic types of IDLs: single loop detectors,
which only detect if a vehicle passes by, and double loop detectors which can measure
velocity directly. Both detectors monitor an electrical current that is triggered whenever a vehicle passes overhead. Originally designed as a form of counting mechanism,
the single loop detectors can also be used to report the speed of traffic. This is done
by assuming an average vehicle length and then dividing that value by the amount of
time the vehicle was detected. Even though the double loop detectors can record traffic speed directly, the California Department of Transportation Freeway Performance
Measurement System (PeMS) uses a more sophisticated algorithm to determine velocity
based on network flow in preference to the more error-prone but direct measurement.
Once a sensor is triggered, data are sent to a small cabinet on the side of the road
called a Front End Processor (FEP). Front End Processors control and regulate the
detectors and record when, and for how long, the detectors connected to it were triggered. From the FEP the data are sent to an off-site server where they are aggregated
by district (of which there are 12 in the State of California). A router connected to the
server then distributes the data over a wide-area network (WAN) that is accessed by the
PeMS servers to capture the data in 5-minute intervals. Once captured, the data can
be viewed through the web-based interface developed by PeMS, or downloaded over
a File Transfer Protocol (FTP) server in a raw format.
Ideally, such sensors would be placed at regular distance intervals along the highways, allowing for the collected data to represent an equal sampling of the area of

13

interest. Unfortunately, the sensors are not placed in such a way (see Figure 2), largely
due to budgetary constraints and the fact that their placement was originally devised
for other purposes (Liu & Danczyk, 2009).
In their research, Liu & Danczyk sought a way to determine optimal sensor position
that addressed the needs of several stakeholders, including construction companies and
transportation agencies. By likening road segments to chromosomes, they treated traffic sensors as genes within a genetic algorithm. In this way, biological concepts such as
gene crossover and offspring generation can be used to determine the ideal locations
for the sensors. The placement suggested by this method took into consideration the
existing sensor configuration, and was found to be superior to the placement currently
in use. The cartographic implication is that sensor placement is irregular and suboptimal, therefore some degree of interpolation and aggregation will be required, and
utmost care must be taken to ensure the data are collected and displayed accurately.

Figure 2: The ILD data for this thesis is comprised of 4,236 individual sensors for mainline highways around the Los Angeles area.

14

2.3 ATIS, ITS, and System Performance
The performance of a transportation network is a consequence of its users, the decisions they make, and the ability for the network to supply service equal to its demand (Golledge, 2002). These decisions can be influenced by the use of ATIS and
ITS (Mahmassani & Jayakrishnan, 1991; Al-Deek & Khattak, 1998; Golledge, 2002;
Chorus et al., 2007). These systems collect information about current traffic conditions via the sensors, which are then used to determine travel speeds and the amount of
congestion within a particular segment of the roadway. The derived information may
then be provided to drivers. How and if ATIS and ITS are used by drivers depends on
the particular driver and the availability of the information system, and will ultimately
affect network efficiency (Golledge, 2002).
A transportation network is said to be most efficient when the flow of its users is at an
equilibrium (Mahmassani & Jayakrishnan, 1991; Mahmassani & Chang, 1987). This
occurs when drivers enter and exit a particular transportation corridor independently.
Due to an accident, adverse weather, or numerous other reasons, congestion can arise
when many drivers attempt to make the same travel decisions. The choice to remain
on course or to seek an alternate route can also be influenced by information provided
by an in-vehicle display using ATIS (Parkes et al., 1991; Al-Deek & Khattak, 1998;
Golledge, 2002).
In their evaluation of the system-wide affect of ATIS, Al-Deek & Khattak (1998)
found that ATIS-equipped drivers made better decisions and fewer late route diversions, which increased network efficiency. Although ATIS is beneficial to the individual and to the network, the benefits are dependent on the market saturation of the
device, and are highest when saturation exists but is low (Al-Deek & Khattak, 1998).
Once ATIS-equipped drivers make up approximately fifty percent of the users of the
network, diminishing returns are introduced and system performance suffers (Mahmas-

15

sani & Jayakrishnan, 1991). The resultant equilibrium and efficiency then exist only
when drivers make decisions that are independent from one another, and this is best
achieved when ATIS and ITS are provided and used intelligently.

2.4 Network Data Collection and Reliability
Many different metrics for measuring and reporting travel time have been developed.
One such measurement is the Travel Time Index (TTI), used by Schrank and Lomax
(2009). This index produces a ratio of peak-period to free-flow travel time, resulting in
an intuitive scale that ranges from 1 (no difference between peak-period and free-flow)
to any greater number (for example, a TTI of 1.75 means there is a 75% increase in
travel time). In addition to its ease of use, TTI can be aggregated to produce citywide
averages, allowing for cities to be compared or ranked based on their average travel
times. This can be used to identify areas with particularly high travel times (such as Los
Angeles, Santa Ana, and Long Beach, TTI = 1.49), and reveal transportation systems
which could benefit most from improvement (Schrank & Lomax, 2009). As demonstrated by a number of studies from the previous decade, this improvement can be in
the form of ATIS- and ITS-based in-vehicle displays (Mahmassani & Jayakrishnan,
1991; Al-Deek & Khattak, 1998; Golledge, 2002; Clarmunt et al., 2000). At present,
these displays are likely to take the form of a traffic map.
Other methods for improving the efficiency of transportation networks have been
proposed. These methods are not based on enhanced information distribution (ATIS,
ITS, etc). Instead, they rely on economic principles or direct manipulation of the network to achieve the desired results. In an evaluation of these alternative methods,
Rosenbloom (1978) proposed shorter work weeks, a decentralized workplace, and artificially increased fuel and toll prices. Note that under these practices only the demand
for the system is reduced; its supply potential remains unchanged. For this reason, it

16

cannot be said that such practices are an improvement to the system. Further opposition to these claims is provided by Schrank & Lomax (2009), who state quite simply that
“higher fuel prices are not the answer.” Their reasons for such a response are that: (1)
modest fuel prices do not suffice; increases would have to be incredibly high and happen over a short period of time, (2) the new prices would have to remain in effect and
not be reduced cyclically, as they currently are, and (3) even under ideal circumstances,
this approach does not address the bulk of the congestion problem (Schrank & Lomax,
2009). Without superior alternatives, the goal of increasing the efficiency of transportation networks must therefore be focused on ATIS- and ITS-like systems. In addition to
the effectiveness these systems, a web-based survey performed by Chorus et al. (2007)
indicated that there is a high demand for time-based information to be provided to
drivers, a need that intuitively-designed traffic maps could satisfy.

2.5 Spatial Knowledge Construction
Whether they are aware of it or not, humans are constantly acquiring spatial information from their environment, which manifests itself in the brain as a cognitive map.
Gollege & Garling (2004) insist that these mental representations are often based on
landmarks that are encountered during routine commutes, past experiences, and the
spatial relationships that are noted by the traveler. This information then serves as the
spatial construct for an internal counterpart of the real world that is unique to each
traveler. Once created, cognitive maps are continuously updated as new information
is acquired, and many decisions rely on this information. This process of enhancement and revision allows for different levels of knowledge to exist. Leiser & Zilbershatz
(1989) argue that there are four such levels. Level one is reached when the traveler
recognizes basic landmarks and uses them as a framework for constructing their spatial understanding. Level two represents the ability to determine direction based on

17

intersections (a form of landmark). Upon reaching the third level, travelers are able to
memorize sequences of rules, which are derived from the previous level. At the fourth
and final level, the “survey level,” the traveler has created a network-like assemblage
that associates bearing changes with landmarks (Leiser & Zilbershatz, 1989). Due to the
utility this knowledge has when making travel decisions, it is likely that anything that
enhances or impedes the creation or understanding of cognitive maps will influence
travel behavior (Gollege & Garling, 2004).
The internal creation, recall, and use of spatial information also varies among individuals. In their research on how well various in-vehicle navigation systems can be
perceived, Fujii et al. (2009) noted that males were able to recall routes with greater
accuracy than females, and that the ability to interpret and transcribe mapped information decreased with age. Likewise, both the frequency with which one operated a
vehicle and confidence in reading maps were associated with a greater understanding
of the information being displayed by the in-vehicle device.
Without knowing the severity or duration of a traffic event, drivers’ spatial knowledge is incomplete. Successfully delivering traffic information to drivers through more
intelligently designed maps could largely alleviate this problem. Drivers with access
to additional information about their environment can make more informed decisions
than those without such access. This difference is likely to positively influence driver
behavior which, at the micro scale, grants drivers an opportunity to avoid congestion
by selecting a less congested route. At a more macro level, a system of well-informed
drivers making good decisions can expect improvements in performance that reduce
the frequency or severity of congestion. When considering the number of drivers and
the ubiquity of traffic, the benefits become difficult to overstate.

18

2.6 Classed vs Unclassed Data
How and if maps are classified is a subject of debate, and depends upon the circumstances of their use. In 1971, Jenks & Caspall presented a paper critical of choropleth
maps with several classes, and claimed that the difficulty of interpreting the map increased with the number of classes it contained. Waldo Tobler, on the other hand,
contended that as forms of visual information, maps should be treated much like pictures or remotely sensed imagery. Under this treatment, unclassed data will provide
a greater and more transitional range of values, thereby remaining most honest to the
data being mapped (Tobler, 1973). Still, concern remained about map readers’ inability to interpret the large amounts of data encoded in unclassed maps.
Indeed Jenks & Caspall were correct that such a difficulty exists. This is reported
in a study by Brewer & Pickle (2002), who asked users to compare several different
classification methods and assessed the readers’ ability to interpret maps based on these
methods. It was determined that a quantile classification was superior to the others,
which were based on hybrid equal area intervals, box-plots, standard deviations, natural
breaks (Jenks’ method), minimum-boundary-errors, and the shared area method. The
authors found the results somewhat surprising, since the quantile method is one of the
simplest known. The need to enhance the clarity of cartographic design is also made
clear by Arthur Robinson in The Look of Maps (1952). As was done in the research by
Fujii et al. (2009), the individual assessment of particular map features can reveal the
elements which perform poorly, iteratively converging towards a superior map design.

2.7 Human Vision and Image Properties
A map’s design cannot be critiqued, within the context of cognition, without a solid
understanding of how maps are interpreted by human beings, and how that interpretation is based on past experiences. Some believe that these experiences come early,

19

that map interpretation is a basic task, and even that children are well-prepared to
make inferences from maps. Liben & Downs (1989) dispute this claim, arguing that
map interpretation is a learned skill that is dependent on spatial knowledge construction that is procedurally enhanced over time. In their study that examined research
from several different fields in which maps were evaluated, Liben & Downs found that
maps are not so obviously seen as conceptualizations of the real world. Rather, maps
represent a collection of symbols that we individually link to referents, a process that is
only successful when paired with a spatial understanding of the relationship between
those referents in the real world.
The idea that humans interpret visual stimuli as collections of individual components
is mirrored in Irving Biederman’s research on human image understanding. He found
that objects in color photographs were identified more quickly than simpler line drawings, though error rates were higher with the photographs (Biederman, 1987). This
finding suggests that the difficulty of interpreting an image increases with the number of objects contained within that image, meaning that each object or component
is demanding its own, separate, interpretation — a concept which has been applied
to computerized visual search routines (Ullman, 1984; Itti & Koch, 2000; RodriguezSanchez et al., 2007). The conclusion drawn by Biederman was shared by Daniel L.
Schwartz (1995), who also found that subjects identified realistic images with higher error than a more abstract counterpart. In questions about reasoning and problem solving, Schwartz also found that humans assess situations differently when viewing stimuli
that closely resemble their referents compared to when the stimuli are of schematic
form. In other words, the logic used by subjects to determine relationships between
real objects is not the same as that used to make inferences about abstract ones. This
is an important realization for cartographers, as it suggests that we cannot expect map
readers to solve problems on a map in the same way they would in real space. More-

20

over, disparities may exist between map readers’ spatial knowledge and their ability to
interpret maps.
Other research has been conducted on the visual properties of objects in a visual display, and how these properties affect recognition and interpretation. Such properties
include brightness, color or chromaticity, shape, salience, and figure-ground relationships. The latter of these properties is the focus of the research by Kienker & Sejnowski
(1986) that attempted to model human figure recognition using various parameters,
including how the figure was drawn (outlines vs several techniques for a solid figure)
and where attention was allocated. Their results indicate that the visual discrimination
between figure and ground is dependent upon the entire composition of the image, and
which component receives attention. Results like these, when paired with previouslymentioned studies, suggest that humans seek to identify features of an image separately,
with accuracy favoring abstractions of those features. Therefore the successful interpretation of an image, or map, rests on the readers’ ability to rapidly and accurately
distinguish figure from ground.
Common examples of figure-ground discrepancies can be found in numerous optical illusions, images that trick us into seeing something that defies our expectations,
or cause us to see different shapes that swap back and forth without notice. A classic
example is the ’vase-face’ image, in which a white vase is made visible by the silhouettes
of two adjacent faces. The ease and reliability with which these illusions occur demonstrates a shortcoming in the human visual system that games and picture books have
made a household familiarity. Despite how many of these books have classified these
visual tricks as ’magical,’ science has shown that they are not beyond understanding.
In research involving the famous ’vase-face’ image and many like it, Lindauer &
Lindauer (1970) asked university students to identify which component (e.g., the faces
or the vase) was the figure, and which was the ground. Overwhelmingly, the students

21

identified the darker of the shapes to be the figure — a response that increased as
the brightness difference between the shapes increased. When reduced to simple outlines with no difference in brightness, students almost always saw a face. While easily
tricked by clever images, humans do show an adeptness at quickly identifying abstract
or schematic forms, as also noted in the research previously mentioned.
This is an encouraging fact because maps, as abstractions of the real world, are
almost always more schematic than realistic. Traffic maps in particular present many
opportunities to take advantage of difference in figure-ground relationships, since both
the transportation network and its various segments must be readily identified. This
can be done implicitly through the network’s schematic nature, or explicitly through
carefully chosen differences in brightness.
Even in the presence of color brightness plays a significant role. Stark differences
in hue are negated by similarities in brightness, a property that presents challenges
to cartographers and map readers alike. In Semiotics of Visual Language (Saint-Martin,
1990), the author argues:
“A more luminous color recedes in depth, and colors of the same luminosity regroup in perception. The contrast between two colored regions
depends more on their different luminosities than on their chromatic diversity: a red figure on a green ground becomes indistinct, if the two
regions have an equal luminous intensity. The luminous quality then dissolves their boundaries” (pg 37).
The properties of luminance and hue are made even more important in the context
of red-green colorblindness, which appears in scientific literature dating back to the
19th century. Although we have expanded our knowledge since, that at least 4% of
men are affected by colorblindness has been known for over 130 years (Jeffries, 1880).
Modern research extends the estimates of the prevalence of red-green colorblindness,

22

stating that it may affect up to 7-10% of males; the estimates for females are difficult to
make due to the rareness of the condition in women (Murray, 1943). Nonetheless, segments of red and green with similar luminosities and shared boundaries are a hallmark
of the typical traffic map design.

2.8 Attention, Mental Workload, and Driver Performance
Within cognitive psychology, there is a great deal of research on attention and vision.
Unlike studies that focus on how figures in a scene are interpreted, attention-based research gives particular emphasis to the concept of attending to several visual stimuli
in parallel. More precisely, the goal of many studies within this field is to categorize
sources of stimuli based on their influence to the users’ capacity to perceive them, as
well as quantifying that capacity and how it affects other tasks (?).
The capacity to perceive visual stimuli, especially from multiple sources, varies as
some stimuli require more effort than others (Horrey et al., 2009). This effort is sometimes referred to as cognitive investment or mental workload, and is experienced whenever
a mentally-engaging task is performed. Scientific literature is replete with studies on
the mental workload required to use maps or mobile devices, and how their use affects
driver performance.
During previous decades, the use of paper maps by drivers was not uncommon. This
method of acquiring route information was compared to an in-vehicle electronic LCD
display by Parkes et al. (1991). In their research, the LCD-based text display was found
to be superior to paper maps by requiring fewer glances from the driver, and a lower
duration of time spent per glance. This suggests that the information presented on the
LCD could be interpreted faster, and therefore required a lower mental workload to
process effectively.
Claiming that it is the mind, not the hands and feet that control a vehicle, Ishigami

23

& Klein (2009) assert that paying attention to one task necessitates withdrawing from
another. This has been demonstrated in several studies in the past, and has been shown
to adversely affect the performance and safety of the driver. In 1991, Walker et al. noted
that an estimated 12% of drivers would leave their lane when placing a call on a cellular
phone. Similarly, the visual demands of an in-vehicle navigation device can overload
the user and jeopardize safety. During the same study, it was reported that the use of
highly complex visual devices resulted in more missed gauge changes, longer reaction
times, and increased the time necessary to complete a route.
The severity of mental workload and its negative effects on driver performance also
varies among users, and appears to decrease with experience. In a study that evaluated peripheral task detection — that is, noticing one event while attending to another
— Patten et al. (2006) found that the performance of newer, low-mileage drivers was
particularly affected by a secondary task. The difference in reaction times between the
groups was on average 250 milliseconds. Requiring a quarter of a second more to respond while driving is not trivial; such a time allows for a vehicle at highway speed to
travel an additional 25 feet before the driver acknowledges the event.
Within the same year as Patten et al.’s research, Horberry et al. (2006) also performed
an experiment aimed at driving distraction and concurrent tasks. In this study, participants operated a driving simulator while exposed to one of two concurrent tasks, which
included manipulating the in-vehicle radio and using a hands-free phone. While in the
simulator, the participants encountered both simple and complex road environments,
and their performance was recorded alongside other factors such as age. Although
age did play a role in the speed at which drivers navigated the complex route, a reduction in performance associated with the additional stimuli was consistent across all
age groups. The authors note that while both forms of stimuli significantly affected
performance, it was the operation of the in-vehicle radio that was most detrimental,

24

further supporting the body of evidence that the requirement to look at something is a
cognitively demanding task.
Also during 2006, Wittmann et al. performed an experiment to uncover the relationship between the position of an in-vehicle display and its effect on driver performance.
By positioning a lighted display in one of seven different locations inside the vehicle,
and varying the distance between the driver and the display, the authors discovered two
major findings. First, the position of the display has a strong relationship with reaction
times while driving. Moreover, vertical displacement of the display showed a greater
influence on driver performance than a horizontal configuration. Secondly, the cognitive influence of a device increases in proportion to its distance from the driver. When
the display was located at great distances, causing the driver to spend additional time
without a view of the simulated road, performance decreased precipitously (Wittmann
et al., 2006). These findings suggest that a less imposing device will require less visual
dedication from its user, allowing for more focus to be allocated to driving or other
tasks.
Despite having comparatively lower mental demands, audio-based navigational systems also degrade driver performance, and have even been referred to as “annoying” by
test subjects (Srinivasan, 1997). This is largely due to the discrete nature of audio-based
navigational cues, which occur only at certain points in time that may not coincide with
the users’ desire to be informed. Srinivasan noted that in addition to audio messages
being related to increased reaction times while driving, the participants in the study
also expressed wishes to shut the devices off. Others have attempted to pair audio and
visual stimuli in an attempt to maximize the benefits of both, but ultimately found the
combination to be detrimental to user performance. In a study on a paired audio-visual
stimuli and users’ ability to pair these signals with events, Liu et al. (2009) found that the
audio and visual signals did not additively enhance the pairings as expected. Rather,

25

the signals competed and interfered with one another.
In a more recent study on navigational maps on mobile devices, Fujii et al. (2009)
evaluated the behavioral patterns that were exhibited while participants interpreted
a map while concurrently performing other tasks, such as answering questions and
pointing out locations. During this time, participants were navigating a route on a
personal computer, to approximate the experience of a more sophisticated simulator.
Over the course of the experiment, map design conditions were varied in their degree
of complexity. Not surprisingly, difficulty in the interpretation of the maps increased
with their visual complexity. In addition to a degradation in subjects’ ability to navigate
the route, the authors also found that subjects referred to the display on average 4 to 6
times per minute, with a duration of up to 4 seconds per glance. The same data in other
terms implies that drivers spend between 4 and 24 seconds per minute with their eyes
not on the road — enough time to travel over 300 ft, or over 20 car lengths, entirely
unaware of the upcoming route.
Yet even more worrying, drivers seem to be unable to detect increases in the cognitive demands imposed by these devices. In a comprehensive study on the effects of
mental workload on driver performance, Horrey et al. (2009) tested participants’ ability to perform mental tasks, before and during driving. The mental tasks included a
guessing game and an arithmetic problem. While performing a single task, participants reported that the guessing game required more mental effort than the arithmetic
problem. However, in a dual-task situation where participants experienced either the
guessing game or the arithmetic problem while driving, they reported no difference
in the mental demands between them. Although they were said to be equally engaging, the guessing game was associated with a significantly greater decrease in driver
performance.
While driving simulators present a wealth of opportunity to study the cognitive in-

26

fluence of concurrent tasks, they also reveal exciting ways in which mental workload
can be measured. An example exists in Just et al. (2008)’s use of functional Magnetic
Resonance Imaging (fMRI) to measure the activity in participants’ brains as they performed a driving task along with a listening task. Without the requirement to listen
and answer questions (a single-task condition), subjects completed a driving route with
high success. Those in the dual-task condition showed a marked decrease in driving
performance, as well as a 37% reduction in parietal lobe activity. The latter finding
has profound implications for cartographers: the parietal lobe is responsible for spatial
processing.
The effects of secondary tasks, especially those involving the use of mobile devices,
can be seen outside of the simulator as well. Acknowledging the growth of mobile technology, Drews et al. (2009) proposed that cellular phones and navigation devices provide “additional sources of distraction...readily available and widely used by drivers,” a
statement later supported by their research. In their analysis of automobile accidents,
it was found that almost a quarter of the drivers had used their mobile phone within the
ten minutes prior to the crash. Further, no difference between the use of a hands-free
phone and a hand-held phone were found; both types of phones, as well as the manipulation of radio and navigational devices by the drivers, were associated with slower
reaction times, more frequent deviations in lane position, and more time without a forward view of the road (Drews et al., 2009). Decreased reaction times witnessed under
the supervision of a scientist must be taken seriously when these lapses in performance
can manifest as real-world injury, or worse.
Links between concurrent tasks and increased mental workload are so well established that in addition to exploring the connections, emerging studies are focusing on
methods to detect mental distraction as it occurs. By pairing a driving experiment with
a support vector machine (SVM), which is a method for pattern recognition that can

27

be tuned towards a desired sensitivity over time, it was found that detecting cognitive
distraction is not only feasible, but shows a success rate up to 85% (Kutila et al., 2007).
In their research, Kutila et al. imposed cognitively-engaging tasks, such as arithmetic
problems, on drivers completing a 45-minute course. Using the amount of time the
driver held their lane and the number of eye and head movements made, the SVM
adaptively acquired the ability to recognize when the driver became distracted. It is
worthwhile to note that the calculation of arithmetic problems — a key requirement
to estimating travel time with current traffic maps — lead to a high enough loss in
performance that the SVM could be triggered.
Once distraction is detected, it then becomes a matter of how to handle or acknowledge it. One idea proposed is vehicle automation. If the task of driving, in whole or
part, is relegated to the vehicle then it might be expected that driver focus can shift elsewhere, thereby reducing the stress or mental workload experienced. To test this theory,
Funke et al. subjected college-aged drivers to various stress-inducing conditions within
a driving simulator. One group of students randomly lost control of the test vehicle
due to simulated wind or black ice, whereas the second did not encounter such hazards
even though they were told to expect them. Additionally, both groups experienced various driving conditions, which included a partially-automated vehicle (participants did
not have to control velocity). On the whole, participants in the automated condition
noticed pedestrians and responded to external stimuli frequently and with higher performance. However, workload indices calculated indicated that cognitive demand was
not reduced (Funke et al., 2007). Rather than reducing the amount of mental workload
experienced, it appears that the automation of a vehicle causes the driver to simply reallocate attention to other tasks or procedures. While this finding might mean that the
reallocated attention could be safely diverted to map-reading, vehicle automation is not
yet commonplace. In the contexts of reducing mental workload and cartography, this

28

leaves the creation of intuitive maps to be a more immediate endeavor.
Increasingly sophisticated methods for detecting distraction exist as well. Recall the
research that employed fMRI data to quantify workload in drivers, performed by Just
et al. (2008). In conjunction with determining an amount of workload, thresholds may
be set to allow devices measuring changes in the body to suggest when distraction is
occurring. The types of measurements are not limited to fMRI. Electrocardiograms
(EKG) may be used to measure elevated heart rates, and basic electrical activity of the
brain can be recored through electroencephalogram (EEG) devices, which may also
report overall alertness (Brookhuis & de Waard, 2010). This research demonstrates
the great strides that have been made in the study of mental workload, as well as the
rising pace with which technology is advancing science in this area. But Brookhuis &
de Waard temper their arguments with a discussion on ethics, which currently prevent
these kinds of studies outside of the simulator, especially those involving altered states of
mind — e.g., driving under the influence of alcohol. Their work not only furthers the
validity of simulated driving exercises, it reminds us that the use of simulators is often
necessary.

2.9 Summary
There are several studies linking increases in mental workload while reading a map
to decreases in performance in secondary tasks, like driving. Additionally, the mental
workload associated with map interpretation was found to increase as designs become
more complex. Further, simulators provide excellent opportunities to evaluate mental
workload and its effects. Like all other forms of information delivery, great care must be
taken when choosing what information will be provided and how it will be presented.
Together, these concepts suggest that the safest and most efficient map design will be
one based on simplicity. This may be achieved through a symbol based approach, the

29

carefully selected inclusion of secondary street information at an appropriate brightness, and expertise in proper data classification and presentation.

30

Chapter 3
METHODS

3.1 Overview and Goals
The methodology of this thesis is designed to evaluate conventional traffic map design
and potential improvements. This is done through comparisons made between two
new traffic map designs and those that are currently in use. Because the new traffic map
designs were derived from a critique of the existing approaches, it is necessary to provide
an overview of that critique. Additionally, the use of Geographic Information Systems
(GIS) and the techniques used to achieve the desired map designs will be discussed, as
such designs have not yet been employed by other cartographers. Lastly, the methods
used to create and perform the experiment that investigated the hypotheses introduced
in Chapter 1 will be discussed in detail. The goals of this thesis are:
1. To identify key limitations in conventional traffic map designs.
2. To develop prototype design strategies to reduce the limitations.
3. To test the performance of the prototype traffic map designs against the existing
standard in terms of travel time estimation.

3.2 An Evaluation of Conventional Traffic Maps
There are a number of commercial and government services providing traffic maps,
each of which employs its own design, but all are variations on a common approach.
Some of these services are illustrated in Figure 2. The main design choice common
to these and nearly all other current traffic maps is the use of red, yellow, and green
segments along the transportation network. It is no accident that these hue choices
mirror those used in stoplights. In addition to being one of three colors, the lengths of

31

the segments vary depending on the underlying velocity data, which is derived from
the Inductance Loop Detectors (ILDs) embedded in the roadways.

Google® Maps

SigAlert®

NavTeq®

Figure 3: Current traffic map designs from three popular vendors (captured at different
times). Each employs colored segments of variable length and velocity from
which the user must estimate the desired information. The arrangement of the
detectors is faintly preserved in the Sig-Alert map.

The task of travel time estimation requires that the user first make two estimations:
distance and velocity. Once this is done, the user must then perform the arithmetic
function

t =

d
v . Error may be introduced at any or all of these stages (see Figure

3). In the case of the current traffic map designs, this process of estimation and arithmetic must be done for every segment, as each has its own velocity category and a map
length. This presents an enormous opportunity for error propagation, which grows in
proportion to the number of segments to be estimated.
Due to the ambiguous segmentation used in current traffic map designs and their
requirement for the user to estimate both velocity and distance, two new approaches to
traffic map design are proposed. Both approaches discard ambiguous network segmentation in favor of a fixed segmentation based on units familiar to the map-reader. The
first approach attempts to remedy the problem of distance estimation by dividing the
transportation network into segments at fixed distance intervals. This will be referred

32

Segment 1... n-1

Stage 1
Estimate Distance
Stage 2
Estimate Velocity
Stage 3
Perform Arithmetic
Segment n

Stage 4
Compute Average
Total
Travel-time

Figure 4: Estimating travel time with current traffic map designs. Error can be introduced
at any stage. Stages 1-3 have additional opportunities for error equal to the
number of segments in the route.

to as the fixed-interval design. By fixing the segment length to a known unit of distance,
the map-reader need only count the segments to determine the total distance of the
route. In this way, the burden of distance estimation could be greatly reduced.
The second approach, which will be referred to as the fixed-minute design, divides
the transportation network into segments with a length equal to an amount of travel
time, which is unique to the current traffic conditions. This approach inoculates the
map-reader from estimation entirely. In the fixed-minute design, the travel time of a
route is equal to the number of segments in that route. No arithmetic or estimation is
required.
As the two approaches being proposed are new, the methods for their creation will
follow.

3.3 Data Acquisition and Creation of the Equal-interval Design
All traffic data was acquired via File Transfer Protocol (FTP) from the Performance
Measurement System (PeMS), provided by the California Department of Transporta-

33

tion. Within PeMS, the state of California is divided into twelve districts. Due to its
notoriety in terms of severe traffic conditions, the Los Angeles area (District 7) was
selected to provide adequate coverage of differing levels of congestion. The data provided by PeMS included: 1) A sensor configuration file (Extensible Markup Language,
or XML format), detailing the position, direction, and name for all ILDs in the state,
and 2) traffic data for the subset of sensors within District 7 (Comma-separated Values,
or CSV format), which comprise 4,923 sensors in total.
The sensor configuration file is state-wide while the traffic data are specific to an
individual district. For this reason, the appropriate sensors must be identified and selected from the configuration file. This was done using regular expressions (regex) for
basic string-matching. The results of the matching were then reformatted to the CSV
format and joined with the traffic data. The resultant file was then loaded into ESRI
ArcMap where it was saved as an ESRI shapefile.

XML

FTP

SHP

CSV

Figure 5: The traffic and sensor configuration data are joined to create an ESRI shapefile,
which makes the display and calculation of the data readily available to many
GIS operations.

34

Data for the transportation network was acquired from the National Atlas of the
United States from 2009. Once loaded into the GIS, the expressways and highways
were separated from the other roads, as the PeMS does not supply information for
smaller arterial streets. Together, these data representing the highways and expressways will be referred to as roadways in the remainder of this thesis. Next, each roadway
was duplicated so that there were appropriate eastbound-westbound and northboundsouthbound pairs. The northbound and eastbound roadways were visually shifted +3
display units in ArcMap to enable the map-reader to distinguish two lanes at a scale
equal to Google Maps zoom level 9 (approximately 1:1,155,581). Finally, each roadway
was split into 5-mile segments, as illustrated in 6. This was done for both directions, using the end opposite of the direction of travel as the starting point for the split operation
(e.g., the east-most end was the starting point for all westbound highways).

Before

After
5 mi

Figure 6: Each roadway is split into 5-mile segments, for both of its directions.

The shapefile containing the velocity data for each sensor was then overlaid, and
sensors for eastbound and northbound roadways were also shifted +3 map units. All
sensors within a 5-mile segment were then given an ID that matched the segment, allowing for the aggregation of their values to be applied to the segment. Each segment
then contained as its value the average velocity recorded by all sensors within that seg-

35

ment. The segments were then classified according to the classification scheme used by
Google’s traffic maps, seen in Table 1. To maintain further consistency with Google’s
design, the same hues and basemap were also used. The result is provided in Figure 7.

Hue
Red
Yellow
Velocity 0 - 24 mph 25 - 49 mph

Green
≥ 50 mph

Table 1: The classification scheme employed by Google’s Traffic Maps.

5 mi
50+ mph
25 - 49 mph
0 - 24 mph
5 mi

Figure 7: The final equal-interval map, adherent to the classification, hues, and base map
used by Google.

36

3.4 Creation of the Fixed-minute Design
The fixed-minute design utilized the same roadways and sensor data as the equalinterval map. The difference in the design is entirely due to the cartographic procedure
used. As with the previous map, the two datasets were overlaid and the necessary shifting of the northbound and eastbound data took place.
Before the roadways could be represented with segments at 1-minute intervals, they
were first prepared by being split into a large number of segments exactly 52.8 feet
1

( 100 mile) in length, seen in Figure 8. In this way, the roadways do not have to be resegmented for each new set of data; only those segments which represent 1-minute of
travel need to be identified and made apparent to the user. The specific length of 52.8
feet has two advantages. First, it is mathematically and computationally convenient,
given the velocity data are in miles per hour. Second, such a small length results in subpixel precision on a modern digital display. The map-reader, and indeed the display
device, would not be able to visually distinguish two adjacent segments. This enables
segments to be selected and symbolized at every perceivable location along a route.

Before

After

.01 mi

Figure 8: Each roadway is split into segments with length equal to 52.8 ft, for both directions.

A surface of velocity values was created by isolating a single roadway and direction
and performing inverse-distance weighted interpolation (IDW) to generate a surface of

37

travel time for every location along the route, with a cell size also equal to 52.8 feet.
The sensors are already influenced by other distant sensors in a complex way on the
actual transportation network. For this reason, the interpolation was performed using
the standard formula:
N
∑

w(x, y) =

i=1

where λi was calculated with p

λiwi

= 4 such that:

λi =

1
( d )4
i

N
∑
k=1

( d1 )4
k

This was done in order to reduce further influence of one sensor on another within
the IDW operation. Once created, map algebra was performed on the surface to convert the cell values into the time required to travel through that cell, given the velocity
contained therein. Calculating this value, and adjusting the time from hours to seconds,
was done in the following way:

tseconds

d
= =
v

1
100
cell value
3600

The travel time surface was then spatially joined with the segmented roadway. Each
of the numerous 52.8 ft segments then contained as its value the number of seconds
needed by a vehicle to travel through the segment. The final step involved identifying
those segments where a sequential cumulative sum of 60 seconds was reached, resetting
the sum to zero each time this was true.
In this way, each identified segment becomes an isochrone line representing 1 minute
of travel time. A Python script was created to perform this function. It operates on a

38

CSV file containing segment ID’s and time field (SegID and TimeSec, respectively).
The script is included in Appendix A.
Alternatively, this functionality can be added directly within ArcMap provided that a
new field is added to the roadway shapefile. Then, assuming a field name of ticMark,
the script in 5.4 can be modified to simply change the new field’s value to 1. The
symbology can then be applied on a conditional basis (e.g., where ticMark == 1).
Finally, the fixed-minute map is styled according to Google’s hue and base map.
Additionally, every fifth 1-minute segment of each roadway was made bold to enable
easier counting over long routes. The final map is depicted in Figure 9.

5 minutes

1 minute

Figure 9: The final fixed-minute map after applying a color scheme and base map that
matched Google’s design.

39

3.5 Experimental Design
University students were solicited to participate in a computer-based experiment as a
means to evaluate the newly created prototypes and compare them to the conventional
traffic map. A total of 49 students volunteered to answer 60 questions as well as preand post-test questionnaires. The main questions were categorized into six different
question types (Table 2).

Category
# Questions
Travel Time Arithmetic
10
Distance Estimation (Control)
10
Distance Estimation (Equal-interval Experimental Segmentation)
10
Travel time Estimation (Control)
10
Travel time Estimation (Equal-interval Experimental Segmentation)
10
Travel time Estimation (Fixed-minute Experimental Segmentation)
10
Table 2: Participants were asked a total of 60 questions, 10 from each category.

The order of questions and categories was randomized into four distinct sequences,
and each sequence had an equal probability of being selected by any participant. Upon
answering a question the participants were asked if they were confident in their answer,
with ’yes’ or ’no’ as the options to choose from. For all main questions, the time taken to
respond, or time-to-completion, was also recorded. Suggested by Walker et al. (1991),
response time can serve as a proxy for mental workload.
Prior to seeing the questions from the categories in Table 2, the subjects were asked
for basic background information, which included age, gender, and whether or not the
participant had a valid driver’s license. Upon completion of the 60 questions, a posttest questionnaire was given. The post-test asked if participants had used traffic maps
in the past, if they believe they’ll use traffic maps in the future, which traffic map design
was preferred, and why that design was preferred. Additionally a space for general

40

comments was provided so that participants could give their feedback on any aspect of
the experiment.
Questions that included maps and predefined routes utilized graphical pins to designate the starting and ending position for the route. Pin placement varied to indicate ten
possible routes for each of the five map-based question types. The routes also varied
in their length and composition of traffic conditions (if present), and none of the routes
required changing roadways more than once (e.g., if a route began on I-405 East a
single road change would allow the route to end on I-710 North. The route could not
then continue to a third roadway to end on I-105 West). Limiting the number of road
changes to zero or one prevents the introduction of an additional variable for analysis,
as the difficulty in estimating travel time or distance is very likely to increase with a high
number of road changes for any map design. In this way, only the cartographic representation of the network segmentation varies across map designs, keeping the focus on
the intended variables.
When each map first appeared to a participant, the route to be considered is highlighted brightly and the base map is given a darkened monochrome appearance for 1
second. After that time the highlight fades smoothly and the base map is given normal
coloration over a period of 500 milliseconds (Figure 10). At this point the map remains
unchanging, the timer begins, and the participant is free to answer the question using
all the time he or she requires. The maps in all of the map-based categories cover the
same extent and scale (Google zoom level = 9). Likewise, interactivity was disabled so
the maps could not be panned or zoomed further. Only the information provided by
the map at its present level of detail was available to each participant.
All of the maps featured a transportation network in the state of California. This
reduces the possibility for participants at a Michigan university to be familiar with the
study area. Additionally, all city labels and points of interest were removed from the

41

maps so that participants could not aid their responses with any previous knowledge
about those locations and the distances or travel times between them.

Figure 10: Before and after the route fade effect. It is particularly useful when a route
includes a destination that might be reached in multiple ways.

3.5.1 Travel time Arithmetic Questions
The travel time arithmetic questions were designed to assess participants’ ability to
perform the proper mathematical computation of travel time. Questions from this
section of the test provided participants with a velocity and a distance, and asked them
to determine how many minutes the trip would take (Figure 11).
The difficulty of the arithmetic questions varied from easy to moderate, and covered
a range of distances and velocities that one might encounter in a real-life setting (i.e.,
no provided velocity exceeded the maximum California speed limit, and no distance
was disproportionately short or long).
Questions of this type were added in order to evaluate the source of error in travel
time estimations. High accuracy in travel time calculations when provided a distance
and velocity paired with poor performance in a map-based setting would suggest that
map interpretation is the source of the error. Conversely, poor arithmetic performance
but high success in distance estimation would imply that error in travel time estimations

42

Figure 11: A typical question from the travel time Arithmetic category.

is due to the inability to perform arithmetic. Identifying the source of error is a key
element in determining which designs perform well, but also why certain designs might
perform as well as they do.
3.5.2 Map-based Distance Estimation
In order to evaluate participants’ performance with distance estimation, a control had
to be established for a baseline comparison. The control used in these map-based questions was an unaltered map of the Los Angeles area provided by Google. The purpose
of this question type was to assess the ability of participants to estimate distances along
an irregular and curving route using a map based on a conventional scalebar.
The distance questions provide the participant with a full map view, as would typically be seen from Google maps. Answers were recorded in miles.
To address the role of a fixed segmentation scheme, an experimental map design
was created. This design is identical to the Google map, with the exception that the

43

roadways are split into 5-mile segments. The scale bar for the alternative distance map
is reflective of this change, and emphasizes the non-euclidean nature of the transportation network, which can be seen in Figure 12.
Responses to questions from this category will allow a direct comparison to the previously mentioned maps that have not undergone a equal-interval segmentation treatment. Such a comparison will determine if a segmentation scheme based on fixed,
familiar units improves distance estimation. In conjunction with the arithmetic questions, the information acquired helps narrow the source of travel time estimation error
to a particular stage in the estimation process (recall Figure 4). Success with arithmetic and failure to correctly estimate distance from a map without a fixed segmentation scheme would indicate an inadequacy in distance estimation that is detrimental to
overall accuracy.
Beyond measures of accuracy, recording the time to respond also enables detailed
revelations. Perhaps non-fixed and fixed segmentation schemes have similar accuracies,
but one design type is associated with significantly faster response times. This would
suggest a more intuitive design that can be interpreted with greater ease, even if the
end result is the same; the implication then is that the intuitive design imposes lower
cognitive demands and therefore makes a better choice for in-vehicle use.
3.5.3 Travel time Estimation
Following arithmetic and distance estimation, the next category of questions exposed
participants to actual traffic maps. In this category, there were three design conditions:
the control, a traffic map based on equal-interval segmentation, and a map based on
fixed-minute segmentation. The control map used in this set of questions was taken
directly from Google Maps (Figure 13), the de facto standard and most popular online
mapping service.
The text paired with the travel time estimation maps read: “Using the route that was

44

Figure 12: The control and segmented map designs. The experimental map on the right
has been enhanced by the inclusion of segments at 5-mile increments.

Slow

Fast

5 mi
10 km

Figure 13: The control traffic map, with pins indicating one of ten routes shown to participants.

45

highlighted, what is the travel time between the pins?” Participants were instructed to
enter their responses using minutes as the units.
Experimental Design 1: Equal-interval Segmentation: Previous research on
map design and human vision suggests simpler designs outperform complex designs
in their ability to be interpreted (Walker et al., 1991; Parkes et al., 1991; Fujii et al.,
2009). One attempt to simplify traffic map design involved fixing segment lengths to
a standard and familiar distance. A length of 5 miles was chosen for the segments,
congruent with the experimental condition in the distance estimation questions (recall
Figure 7).
Experimental Design 2: Fixed-minute Segmentation: A further degree of
simplification is present in the third traffic map design condition — the fixed-minute
design (recall Figure 9). This experimental design uses time rather than distance as the
basis for segment length. Such a design is free from user error introduced in distance
estimation and arithmetic. By delivering the information in the same units as required
by the task for which the map is intended, the potential for error should be significantly
reduced.
These three design conditions sequentially reduce the number and complexity of the
tasks the map-reader must perform to acquire the intended information.
3.5.4 User Confidence
Not only does confidence reveal a participant’s expected outcome in their individual
success with a question or map design, it also allows for the discovery of relationships
between confidence and error that might otherwise go unnoticed. These relationships
can be particularly telling in terms of user interfaces and human behavior. Confidence
and error can be paired in four basic ways (Figure 14).
For example, if users are continually confident but also consistently wrong in their
answers, that might suggest that a map design is untrustworthy. It may be imbuing

46

more confidence than it is worthy of and misleading its users. This would be a troublesome scenario outside of an experiment and in a real setting, as map readers are most
often not informed when they’ve made an error; there is no researcher to turn to for
questions, no screen showing the correct answer, nor any indication that they’ve made
a mistake. Yet the confidence invoked by the map would instill a feeling of satisfaction
in the user, who would likely continue using the map oblivious to their own error.
The opposite situation is just as worrying. In this case, the map-reader would be accurately making conclusions while simultaneously having no confidence in those conclusions. Without a prompt that indicated a correct response, the user would be unaware of their own success. Despite their correctness, the map-reader wouldn’t feel
confident and would therefore be likely to abandon the use of such a design or prefer

ro
r
Er
gh
Hi

Lo
w

Er

ro
r

others in its place. Ideally, confidence and success would always be matched.

High Confidence

Ideal

Unknowingly
Misleading

Low Confidence

Uknowingly
Successful

Undesirable

Figure 14: User confidence and map error can reveal interesting properties about design
and map characteristics.

3.6 Expected Results
Due to the relationship between complexity and map interpretation, it is expected that
sequential improvements will be noticed in both accuracy and time-to completion as
the map designs become more simplified. The ambiguity previously identified in the

47

control traffic map is likely to significantly impair both distance and travel time estimations.
3.6.1 Arithmetic Ability
Although the formula for converting distances and velocities into times is not complicated when written or explained, it is challenging to perform mentally. This is especially
true when the values are not whole numbers. It is expected that participants will have
high rates of error in their calculations. Without a basis for comparison, no assumptions are made about the time to complete the arithmetic questions. However, it is
predicted that the response times will be sufficiently long to suggest that such problems
are not trivial or readily solvable with ease. Low to moderate levels of confidence are
expected, further suggesting that mental arithmetic is difficult and users are aware of
their inability to perform it well.
3.6.2 Equal-interval Segmentation and Distance Estimation Expectations
It is assumed that map readers will have great difficulty in applying a linear scale bar
present in traditional maps to non-linear transportation networks. For this reason, it is
expected that accuracy in distance estimations will be greatly improved with the equalinterval design.
Similarly, the time taken to respond to questions based on the equal-interval segmentation is expected to be much shorter than the times associated with the control condition, while participants’ confidence is likely to be higher with the equal-interval design
than the control. Two reasons for these predictions exist. First, the equal-interval segmentation embeds the scale bar directly into the transportation network. This takes
much of the guesswork out of the estimations and reduces the task to a counting procedure. While users will still have to estimate the distances where the end of a route
does not coincide with the end of a segment, these short distances are unlikely to affect

48

accuracy as much as an unsegmented design that requires this kind of guesswork the
entire route. Users will be able to determine, for example, that approximately half a
1

segment is about 2 2 miles and that a quarter of a segment is a little more than a mile.
Secondly, by embedding the scale bar, users do not have to continuously refer to a
legend to reassure themselves of the scale in use. The map and the legend are one,
reducing both the number of eye glances away from the focus of the map and the time
spent not looking at that focus. Together, the shorter time required and the increased
potential for accuracy are likely to encourage higher confidence in distance estimations
with the equal-interval map design.
3.6.3 Travel time Estimation and Segmentation Schemes
Since accurate distance measurements are a prerequisite to travel time estimations, the
difficulty with using a linear scale bar to estimate distances along an irregular route are
expected to be problematic in the control traffic map design. Inaccuracies in distance
estimations paired with the inability to perform arithmetic are expected to combine
and severely affect travel time estimations. The ambiguous velocities in the control
legend are also expected to exacerbate this problem, diminishing travel time estimation
accuracy even further. Due to the need to refer to a legend that is entirely separate from
transportation network, and the combined challenge of assuming correct velocities and
performing the proper arithmetic, it is also expected that the control condition will be
associated with the highest response times.
The equal-interval segmentation scheme is predicted to show an improvement in
travel time estimation. Reasoning for this prediction is based on the previous expectations that a equal-interval design will lead to distance estimations that are more accurate. Improved distance measurements, even when paired with an unchanged arithmetic ability, should yield more accurate travel times. With fewer opportunities for
error to propagate (one less task is required per segment), it is expected that the equal-

49

interval design will improve travel time estimation accuracy significantly. Although
response times are likely to decrease, the task of travel time estimation is still very similar between the control and equal-interval designs. For this reason, response times are
expected to be shorter with the equal-interval segmentation, but not substantially so.
The fixed-minute segmentation scheme is expected to be associated with the least
amount of error in travel time estimations. By reducing the need for users to estimate
distance or perform any arithmetic, the only task required of them is counting segments.
Since counting is arguably easier than arithmetic, and there is no need to estimate
distance, error is predicted to markedly low with this approach. Likewise, time-tocompletion is also expected to be significantly reduced with the fixed-minute design.
The less complicated nature of the task will likely allow users to determine travel times
very accurately. This reduced mental engagement and time involved to interpret the
fixed-minute map is also expected to increase user confidence well above that found in
the other design approaches.
3.6.4 Map Preference
The final measure of map performance in this experiment is how much users prefer using a particular map design. Expectations for the other measures of map performance
(accuracy and time-to-completion) favor the fixed-minute design, and it is predicted
that this will also be apparent in user preference. A map that is easier to use, requires
less time to interpret, and continuously rewards the user by reaffirming their correct
responses (i.e., elicits higher confidence) is indicative of a more positive user experience.
With the prediction that the previous expectations are met, the fixed-minute design is
likely to be strongly preferred by users.

50

3.7 Participants
A total of forty-nine students at Michigan State University volunteered for this study.
Subjects volunteered for this research by signing up during classes in which the study
was announced and briefly introduced. During the introduction, explanations of the
research were deliberately vague so that potential volunteers were aware that a study
was taking place without knowing what criteria or variables were going to be evaluated.
The gender composition among the subject pool was approximately equal, with twentyseven male participants and twenty-two female. The ages of the participants ranged
from 18 to 33 (mode: 20). All participants indicated that they had a driver’s license valid
in the United States. Each subject received $10 as compensation for participating in
the study, which required approximately 20 to 30 minutes of time to complete.
Note that the subject pool is composed of adults with at least some university education (for reference, only about a quarter of adults in the U.S. have a bachelor’s
degree). This may introduce bias and produce results that indicate performance that
is slightly higher than what could be expected from the general public. In some cases,
this could prove beneficial; if university students cannot perform a certain task well it is
unreasonable to expect those with less education to do as well or better. Certain map
interpretation exercises that prove to be problematic for the subjects in this study may
be even more problematic to others.

3.8 Materials
3.8.1 Hardware
Sixteen Dell Precision 690 computers equipped with Dell 1907 FPt displays were used
for the study. The computers were configured with Microsoft Windows XP (Microsoft,
Inc., 2001) and had identical settings, including 1280 x 1024 screen resolution and a
refresh rate of 60 Hertz.

51

3.8.2 Software
The study was designed in Adobe Flash CS4 (Adobe Systems, Inc., 2008) using the ActionScript 3 programming language. Animation for the fade effect was created with
the Tweener library, version 1.33.74 (MIT License). The study was accessed externally
through the Mozilla Firefox web browser (Mozilla Corporation, 2010), which ran in
full-screen mode. Data from the study was recorded anonymously and delivered by a
PHP script that arranged the data into a comma-separated values file.

3.9 Procedure
Participants completed the study in the computer lab of the Geography Building at
Michigan State University (Room 201). The computer lab was reserved for six different
one-hour time periods, from which volunteers chose the time they wished to participate.
The number of participants taking the study at any one time ranged from three to
twelve. The computers used were arranged at 90◦ angles so that participants were
unlikely to disturb one another and could not see others’ responses.
Once all participants arrived at their scheduled time, they were thanked for their
participation and were free to choose a seat of their liking. Two consent forms (see
Appendix A) indicating the researchers responsible for the study, the tasks involved,
and participants’ right to withdraw were handed out, as required by the Institutional
Review Board (IRB). Participants were instructed to read the form carefully and then
sign and return one copy, retaining the other for their records. While participants read
and signed the consent forms, the door to the computer lab was posted with a sign
requesting that others not knock or attempt to enter.
After each participant had returned a signed copy of the consent form, the format
for the test was explained and instructions for entering responses and navigating the
test were given. At this time, participants were free to begin the study, starting with the

52

pre-test that acquired basic subject background information. This was followed by a
brief on-screen tour of the test that familiarized participants with the test interface and
the question formats they would encounter.
At the completion of the tour, users were notified that the actual test would start
once they indicated that they were ready. Immediately after the participant chose to
continue, the experimental test began. In order to complete the study, all 60 questions had to be acknowledged, which included the option to skip a question without
providing a response, in compliance with IRB protocol (note: none of the forty-nine
participants chose to skip a response). During the experimental test, participants could
not use calculators, write anything down, or use any tools that might assist them in their
responses.
Completion of the experimental test lead to a brief post-test, in which participants
were asked about their experience and which map type they preferred. This allowed
participants the opportunity to supply their opinion about the map designs presented
to them, as well as their overall thoughts on the test itself. Once the post-test was
completed, participants were presented with a ‘submit’ button. Upon clicking the button, their data were combined with existing responses and stored anonymously in a
spreadsheet-style format.
When participants had completed the study, they were again thanked for their participation and given the $10.

53

Chapter 4
RESULTS AND ANALYSIS
Performance for each map design was evaluated in terms of accuracy and time-tocompletion. Accuracy, or conversely error, was measured in two ways depending on the
question type. For arithmetic questions it is unreasonable to expect perfect accuracy
or an absolute absence of error. Therefore percent error was used, which provides
the additional benefit of reducing error bias related to the magnitude of the values
being estimated. Accuracy for the map-based questions, which include the distance
and travel time estimation portions of the experiment, was measured in mean error.
For comparison between map types, absolute mean error was used to prevent the effect
of positive and negative errors canceling one another out, though the raw error was
examined for directional bias.
The time-to-completion, or response time, for all questions was measured by the
amount of time required for the user to submit a response after the question appeared.
Responses to the arithmetic questions were examined separately, while responses to
the map-based questions were compared within the following groupings:

Design Condition Compared
Group 1: Distance
Control
Equal-interval (no velocity)
Group 2: Travel Time Control Equal-interval Fixed-minute
Table 3: Map designs were grouped and compared based on the estimation task they were
associated with.

4.1 Arithmetic
Percent error was used to evaluate the accuracy of math questions. For the purpose
of this study, a window of

±15% was used for hypothesis testing.
54

This presents a

reasonable threshold for error without being overly generous or unfairly conservative
(for example, if the actual answer is 60 minutes, this threshold allows ‘correct’ responses
to be within ± 9 minutes of that value).
The overall results of the arithmetic portion are summarized in Tables 4 and 5.

Absolute Percent Error
n Min Max Median
¯
x
490 0.00 316.70
4.17
16.47
Table 4: Magnitude of percent error across all arithmetic questions.

Percent Error
n
Min
Max Median ¯
x
490 -90.08 316.70
0.00
3.07
Table 5: Percent error from forty-nine participants responding to ten arithmetic questions.The mean is slightly positive, but not significantly so.

Overall, participants exhibited adequate arithmetic ability and did not perform poorly
in converting distances and velocities into travel times. While the mean of absolute
percent error was higher than the predetermined threshold of 15, this difference is not
significant (t

= 1.01, df = 489, p = 0.31).

The per-subject results show that nearly 3 out of 4 responses demonstrate adequate
skill in calculating travel times (Figure 15). Out of 490 total responses to the arithmetic questions, 362 (74.08%) of them were within the pre-determined error window.
Additionally, only 11 subjects had an average percent error that exceeded 15%; the
remaining 38 subjects (77.55%) all had mean percent errors within ±15%.
These results were not expected and are contrary to the prediction made in Hypothesis 1. This suggests that map readers do not have innate difficulties with the mental

55

% Arithmetic Responses ≤ 15% Error
100
80

77.55%

74.08%

60
40
20
0

All Responses Subject Means

Figure 15: Nearly 75% of all responses to arithmetic were within the error window. The
percentage of participants with individual means within ±15% of the correct
value was also very high.

arithmetic involved in travel time estimation. It is therefore likely that the failure to adequately determine travel times from a traffic map derive from an inability to accurately
interpret the map, which is a direct consequence of the map’s design.
Additionally, a directional bias in travel time estimations was not revealed by the
arithmetic test (refer to Table 5). Participants did not significantly over- or underestimate travel times (t

= 1.88, df = 489, p = 0.06).

The number of subjects

with positive mean percent errors was 28, which accounts for 57.14% of the subject
pool. This is not significantly larger than half (t = 1, df = 48, p-value = 0.1612).
The participants’ ability to perform arithmetic at a satisfactory level this frequently,
and without a directional bias, suggests that travel time arithmetic is not a substantial
burden to map readers. A caveat is that the subjects in this study were university students, many of which were likely to have had recent courses in mathematics and more
familiarity with, if not only fresh exposure to, complex arithmetic.
Participant responses to the arithmetic portion of this study are also associated with
high levels of confidence. Out of 490 possible individual responses, 427 of them were

56

marked “confident.” Similarly, once shown the correct answer alongside the response
given, the majority of participants also felt their responses were close to the actual values
(Figure 16).

Subjective Responses Before and After Seeing Correct Arithmetic Values
100

87.14%

100

80
60

60

40

80.82%

80

40

20
0

20

12.86%
Not Confident

0

Confident

Confidence

19.18%

Not Close

Close

Answer was ‘Close’

Figure 16: Participants were overall very confident with the arithmetic portion of the test
(left), and later felt their responses were close to the actual values (right).

Not only did participants perform the arithmetic successfully based upon objective
criteria, participants also felt confident when doing so. It can therefore be determined
that the task of arithmetic for travel time estimation was not particularly difficult to the
map readers in this experiment.

4.2 Distance Estimation
Following the arithmetic portion, participants responded to twenty questions asking
them to estimate distances between two pins on a map. Ten of the questions were
paired with a map without a segmented network while the others where paired with a
map based on a fixed-minute segmentation scheme.

57

40
20
10

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

Control

30

Time (seconds)

Magnitude of Error

50

Distance Estimation: Error
.

Equal-interval

Distance Estimation: Response Time
.
.
60
.
.
50
.
.
.
.
.
40
.
.
.
.
.
.
.
.
30
.
.
20
10
Control

Equal-interval

Figure 17: Overall, the equal-interval segmentation was associated with lower errors and
lower response times.

The equal-interval segmentation scheme improved accuracy and response times
(Figure 17). In both cases, these differences were statistically significant (Table 6).
When evaluated based on the number of responses and subject means that were within

±15% of the correct value, the equal-interval segmentation was associated with great
improvements over the control (Figure 18). When using the equal-interval map, almost all participants (48 out of 49, or 97.96%) had a mean percent error within the
threshold considered correct. The improvement in distance estimation is also revealed
in an increase in total correct responses, which account for 375 out of the total 490
(76.53%).The control stimuli produced a subject mean error below or equal to 15%
for only 19 subjects (38.78%). In terms of total responses, 210 out of 490 (42.86%)
were within the 15% threshold.
Further, the increased accuracy with the equal-interval design did not come at an
additional cost to the user; such responses were made even faster than those using the
non-segmented, conventional map design.

58

x
¯
median
min
max

Error

x
¯
Response Time

median
min
max

Control Equal-interval
T-test
4.54
2.42
t
7.86
3.00
0.73
0.00
0.10 df
489
52.50
34.99 p < .001
14.86
13.19
t
3.93
12.94
11.47
4.07
4.25 df
489
68.35
56.02 p < .001

Table 6: A comparison of the results from the distance estimation portion of the experiment shows that the equal-interval segmentation scheme improves accuracy and
response time.

% Subject Responses < 15% Error

% Total Responses < 15% Error
100

100
76.53%

80
60
40

97.96%

80
60

42.86%

40

20

20

0

0

38.78%

Conventional

Equal-interval

Conventional

Equal-interval

Figure 18: The equal-interval map was associated with great improvements in distance
estimation.

59

Although confidence was moderately high with both map designs (Figures 19 and
20), the segmented design was associated with the highest amount of confident responses (79.80% vs 96.53%, respectively). This is particularly important as there were
differences in the relationships between confidence and error between the map types

¯
as well. In the case of the control map, confident responses (x
more accurate than those marked not confident (x
¯
is associated with a t-value of

error = 3.75) were

error = 7.66). This difference

−4.467 (p < .001).

Conversely, there was no dif-

ference in accuracy between confident and non-confident responses to the segmented
map design (x
¯

error = 2.37 vs x error = 3.85, t = −1.64, p = 0.12).
¯

The control map elicited fewer confident responses, likely due to the lack of accuracy
associated with the responses. This effect did not occur with the segmented design.

Responses Before and After Seeing Correct Distance Values (Control)
100

100
79.80%

80

80

60
40

40

61.02%

60

20
0

20.20%

Not Confident

38.98%

20
0

Confident

Confidence

Not Close

Close

Answer was ‘Close’

Figure 19: Almost 80% of responses were confident with the control (distance) map,
though the percentage of responses indicating a close response was moderately
low.

In addition to evoking responses that were almost always confident, the equal-interval
segmentation seems to inoculate the user from any loss in accuracy associated with a

60

Responses Before and After Seeing Correct Distance Values (Equal-interval)
100

96.53%

80

100

84.08%

80

60

60

40

40

20
0

20

3.47%
Not Confident

0

Confident

Confidence

15.92%
Not Close

Close

Answer was ‘Close’

Figure 20: Confidence was dramatically higher with the segmented map, and the vast
majority of participants felt their responses were close.

lack in confidence. In this way, the segmented design was both more rewarding and
more forgiving.
Both design types were evaluated for directional bias. The control map was strongly
associated with overstated distances, with 45 out of 49 responses overstating the correct
value. The experimental map was not associated with such a bias and the number of
negative and positive responses was approximately equal (25 vs 24, respectively).

4.3 Travel Time Estimation
Large differences were noted in travel time accuracy and response times across the traffic map design types (Figure 21). Participants estimated travel times with an average
error of ±19.77 minutes when using the Google traffic map. The equal-interval traffic map design improved accuracy significantly, reducing the average error to ±13.57
minutes. The largest improvements were seen with the fixed-minute traffic map, which
was associated with an average error of ±5.52 minutes. These travel time errors were
subjected to ANOVA, which indicates that the differences are significant (Table 7).

61

Magnitude of Error

Travel Time Estimation Error
by Map Type
80
60
40
20

.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

0

Fixed-minute
Equal-interval
Conventional

Figure 21: Fixed segmentation schemes greatly improve travel time estimations.

Out of the 490 possible responses to the travel time estimation questions for each map
type, the conventional traffic map was associated with only 135 responses (27.55%)
within the

±15% error window.

The equal-interval map produced similar results,

with 129 (26.33%) being within ±15%. The fixed-minute map produced 389 correct
responses, accounting for 79.39% of the total.

ANOVA p-Matrix of Error
Equal-interval Fixed-minute
Fixed-minute
0.011
Google
0.044
< 0.001
Table 7: ANOVA reveals that each result is significantly different (alpha = 0.05) from the
others.

Subject means reveal a similar trend. Roughly half of the subjects (26, or 53.06%)
had total means within the error window when using the conventional traffic map.
Similarly, only 25 (51.02%) subjects had total means at or below 15% with the equalinterval traffic map. The fixed-minute traffic map was again associated with the highest

62

number of correct responses; 43 subjects (87.76% of them) had average means within
the accepted window of error (Figure 22).

% Subject Means < 15% Error

% Total Responses < 15% Error
100

100
79.39%

80
60
40

80
60

27.55%

87.76%

53.06%

51.02%

40

26.33%

20

20

0

0

Equal-interval
Conventional
Fixed-minute

Equal-interval
Conventional
Fixed-minute

Figure 22: The fixed-minute design improved the number of correct responses substantially.

Errors were varied in their directions as well. Participants significantly underestimated travel times by an average of 11.77 minutes with the conventional traffic map
(t

= −16.28, p < 0.001).

Only two participants consistently overestimated

travel time with the conventional traffic map. Conversely, participants overestimated
travel times by an average of just 1.49 minutes with the equal-interval design (t

=

2.30, p = 0.01). This also occurred with the fixed-minute design, which was associated with an average overestimation of only 1.31 minutes (t

= 5.06, p < 0.001).

Similarly, response times also varied across design types (Figure 23). The fixedminute segmentation scheme was again associated with the best results (lowest response
time,

x = 18.64 seconds).
¯

Likely due to the similarities in the interpretation tasks

involved, the Google design and the equal-interval design share similar response times
(x
¯

= 26.03 seconds and x = 28.07 seconds, respectively).
¯
63

Response times as-

sociated with the fixed-minute map were the only results to be significantly different
from the others. ANOVA indicates that the slight increase in response times noted
with the equal-interval map were not significantly different than those associated with

Time (seconds)

the Google map (Table 8).

Travel Time Response Times
by Map Type
.
.
150
.
.
.
.
.
.
.
100
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
50
.
.
25
0

Fixed-minute
Equal-interval
Conventional

Figure 23: Response times were also lower with the fixed-minute design, while the Google
and equal-interval designs were very similar in this regard.

ANOVA p-Matrix of Response Times
Equal-interval Fixed-minute
Fixed-minute
0.007
Google
0.536
0.307
Table 8: The differences in response times between the Google design and the equalinterval segmentation were not significantly different, though those associated
with the fixed-minute segmentation were.

Relationships between confidence and map type were also noted. Out of all the individual responses to the Google map design, 40.82% of them were marked as confident.

64

The equal-interval map improved confidence be eliciting confident responses 56.69%
of the time, and the fixed-minute map was associated with the highest amount of confident responses at 94.56% (Figure 24). Confident responses were not more accurate
with the Google design (t

= 1.57, p = 0.12), nor did confidence affect accuracy

with the equal-interval design
(t

t = 0.39, p = 0.70) or the equal-interval design

= 1.87, p = 0.07).

% Responses Marked ‘Confident’
100

94.56%

80
56.69%

60
40

40.82%

% Responses Marked ‘Close’
100

90.41%

80
60
40

20

0

39.18%

20

0

33.47%

Equal-interval
Conventional
Fixed-minute

Equal-interval
Conventional
Fixed-minute

Figure 24: Confidence was improved by both segmented designs, though the fixed-minute
design was associated with very high confidence. Similar associations were seen
between the map designs and responses marked close.

4.4 Map Preference and Subjective Feedback
Participants overwhelmingly preferred the fixed-minute traffic map design, with 46 of
49 of them choosing it as the one they liked the most. Only three participants chose
the equal-interval design, and none preferred Google’s traffic map (Figure 25).
Out of the 49 participants, 33 (67.35%) responded that they had used a traffic map
in the past, and 44 of them (89.80%) said that they expect to use traffic maps sometime

65

% of Participants Preferring Each Map
100

93.88%

80
60
40
20
0

0.00%

6.12%

Equal-interval
Conventional
Fixed-minute

Figure 25: None of the participants chose the Google map as their favorite, while the vast
majority preferred the design based on a fixed-minute segmentation.

in the future.
Feedback also came in the form of comments which cannot be quantified but do
provide insight into user opinion. One participant commented:
“It was really easy to tell how long it would take to get somewhere, right
down to the minute, because you could count if you had to. I hated the
[Google and equal-interval maps] because you really had to think, and
do math in your head, which I did not enjoy.
I felt the only maps that were helpful were the ones which did not involve speed at all.”
Another participant shared similar thoughts regarding the fixed-minute map, “ This
map was easiest to tell how far away something was both in terms of distance and in
time. The other maps were confusing.”
Despite the popularity and the public availability of the Google traffic map, one user
added, “I really hope they never start using the type of maps that had 3 different colors
for slow, fast, etc. I’d never get anywhere on time.” This sentiment was shared by

66

another participant, who left the following:
“I found that it took me a long time to figure out the maps with the
different color indicators showing how long (fast, slow, etc.) it would take
to get through the area. I thought the range of speeds given for each
color were not specific enough, and it was difficult to determine the travel
time.”
These comments, and the others like them, suggest that the addition of colored
classes and ambiguous segmentation is a burden that affects travel time estimation.
While the empirical results allow for an objective comparison of the different design
variables, the subjective feedback sheds light onto why the results turned out the way
they did. Aside from representing a larger and more difficult set of tasks which results
in high amounts of error and long response times, interpreting the Google traffic map
and the equal-interval design is a task that, on the whole, users did not enjoy.

67

Chapter 5
DISCUSSION

5.1 Arithmetic Ability
5.1.1 Error and Accuracy
The results from this experiment suggest that map readers do not have difficulty performing the arithmetic to convert velocities and distances into times. Participants had
an average percent error of 16.47. This did not exceed the threshold of 15% at significant levels. Furthermore, a directional bias was not present and participants neither
under- or over-estimated travel times when performing the arithmetic.
5.1.2 Confidence
Confidence with arithmetic was also high, and the subjective response to whether or
not an answer was close to the actual value was also high. More than 87% of responses
indicated confidence and 80.82% of responses were identified as being close to the actual value after seeing it. This might suggest that responding to the arithmetic questions
was a rewarding experience. If on the other hand responses were confident but rarely
close, this would imply that the user feels positive about their ability initially but changes
his or her opinion later. A constant change in an opinion and an indication that despite one’s confidence he or she still cannot do well would likely be a very frustrating
scenario.
5.1.3 Potential Bias
It is important to recall that the participants in this study do have the advantage of at
least a partial University education with a tested mathematics aptitude and a recent
history of math coursework. It cannot be said that the same performance would be

68

enjoyed by the general public. In fact, it would be fair to hypothesize that arithmetic
performance would be lower — perhaps low enough to significantly exceed the 15%
threshold. This would be unlikely to change the relationship of the results for the map
interpretation exercises, but increased error propagation and lower overall accuracy
would likely be noticed. This caution also applies to the following sections of this chapter.

5.2 Distance Estimation
5.2.1 Error and Accuracy
On the whole, participants had better performance with distance estimation than was
expected. The control condition was associated with an average error of 4.54 miles,
and the equal-interval design demonstrates improvement by eliciting an average error
of only 2.42 miles. Additionally, the control condition lead to a much higher maximum
error (52.50 miles vs 34.99 miles), suggesting that the control map left a larger window
for participants to make a mistake.
5.2.2 Response Time
Participants were also able to make the more accurate distance estimations more than
one second faster with the equal-interval design. The average response time for the control condition was 14.86 seconds while the equal-interval design was associated with a
response time averaging 13.19 seconds. Allowing users to make more accurate estimations in less time represents a significant improvement in map design. Instead of having
increased accuracy come at a cost, the improved estimates came even faster to users.

69

5.2.3 Confidence
The success of the equal-interval segmentation for estimating distance is also revealed
by the participants’ confidence in it. Unlike the control condition which elicited confident responses only 79.80% of the time, 96.53% of responses to the equal-interval
design were marked as confident. Similarly, only 61.02% of responses were classified
as “close” by participants using the control map while 84.08% of responses were classified this way for the equal-interval design.
This suggests that participants were aware of the improvements in distance estimation associated with the equal-interval design. Participants felt confident using it and
their confidence was often justified.

5.3 Travel Time Estimation
5.3.1 Error and Accuracy
Participants performed significantly better with both alternative traffic map designs
than they did with the conventional design. The average error in travel time estimations
associated with the conventional traffic map was just short of 20 minutes (19.77). Actual
travel times for the conventional map questions ranged from about 19 minutes to nearly
74 minutes. An error of 20 minutes represents 25-100% of the actual travel times for
the routes. This is a substantial amount of time to be uncertain about. In contrast, the
equal-interval segmentation was associated with an average error of only 13.57 minutes.
Improvements over both designs were seen with the fixed-minute segmentation, which
produced an average error of just 5.52 minutes.
Such a large improvement associated with the fixed-minute segmentation demonstrates the benefit of reducing the complexity of a map’s design. By eliminating classed
data and the need to estimate distances and velocities, the fixed-minute design does not
involve any computation or estimating on the user’s end. This may also be due to the

70

symbology being more congruent with the actually phenomenon being mapped. In a
transportation network, congestion causes vehicles to become backed up, reducing the
space between one vehicle and the next. The symbology used in the fixed-minute map
does exactly the same thing. When congestion is high, travel times increase and it takes
a greater amount of minutes to travel a short distance. This reduces the distances between the symbols, effectively causing them to become congested as well. Fixed-minute
segmentation represents congestion with congestion.
A secondary benefit of the fixed-minute map is that it is not reliant on hues, and
therefore does not pose problems for individuals with red-green color blindness. A
colorblind-friendly traffic map can be created by making the red and green hues different in their luminosity, or by slightly increasing the amount of blue in the green
segments. Despite these relatively easy solutions, conventional traffic maps continue
to use hues that are troublesome for those with red-green colorblindness, providing yet
another advantage for the fixed-minute map.
Not only did the conventional map fail to allow users to make reasonably accurate
travel time estimations, it was also associated with estimated travel times that were significantly lower than the actual values. When using the conventional traffic map, participants most often underestimated travel times with an average of 11.77 minutes. It
is unclear why this may be. Participants performed better than expected at arithmetic
and distance estimation with a conventional map, and neither task revealed a directional bias. Yet when the tasks are combined, participants rarely provided accurate
responses and a very large directional bias occurs. For reasons that are not yet known,
it appears that the conventional traffic map not only conveys information in a more
complicated way, but its design has a mitigating effect on the perception of delays due
to congestion. A frequently assumed benefit of the conventional map is that its reliance
on hue alone enables the map-reader to quickly judge the severity of congestion at a

71

glance. However, the conventional traffic map significantly downplays the severity of
congestion, leading users to believe conditions are better than they actually are.
The sequential improvement in accuracy that was revealed across the traffic map
designs was predicted in Hypothesis 4. Each design represent a more simplified strategy
to distance estimation (Figure 26). In the case of the conventional traffic map, distance
is uncertain and velocity is ambiguous; the legend does not provide any numerical basis
to guide velocity estimates. The equal-interval design removes ambiguity as well as the
need to estimate full distances, but still leaves room for uncertainty when estimating
velocities. Only the fixed-minute design removed both ambiguity and uncertainty. This
simplified design was expected to be associated with the highest levels of accuracy, and

Segment 1... n-1

the hypothesized relationships appeared in the results of this study.

Stage 1
Estimate Distance
Stage 2
Estimate Velocity
Stage 3
Perform Arithmetic
Segment n

A

Stage 4
Compute Average
Total
Travel-time

B

Stage 1
Count Segments
Total
Travel-time

Figure 26: Possible strategies for estimating travel time with the prototype designs. The
equal-interval approach (A) only requires distance estimation once, whereas
the fixed-minute design (B) transforms the entire task of travel time estimation
into a counting exercise. See Figure 4 for comparison.

These findings suggest that velocity estimation is the largest source of error when estimating travel times with conventional traffic maps. It can be argued that improvements
in data classification may alleviate this problem and improve map readers’ ability to estimate travel times. Maps that are entirely free of velocity and distance classifications
are better yet, as demonstrated by the fixed-minute map which substantially improved

72

travel time estimations.
Both of the traffic maps based on a fixed segmentation scheme enabled participants
to make estimations that were closer to reality. This suggests that these maps facilitate spatial knowledge that is more congruent with the actual environment and the
phenomenon of congestion. An increased similarity between the interpreted data and
the real world allows users to make decisions based on a level of precision that did not
previously exist. Outside of a laboratory setting, the improvements in traffic-related
decisions would allow drivers to choose better routes or make better decisions about
avoiding congestion.
While the benefits to the individual would be immediately recognized, network performance would also be influenced, as noted by Mahmassani & Jayakrishnan (1991).
A transportation system utilized by drivers with access to accurate information about
their environment has the potential to become increasingly efficient. If this were to
occur on a system-wide scale across the entire network, congestion could be reduced
at the macro level, reducing the substantial costs associated with congestion that were
revealed by Schrank & Lomax (2009).
5.3.2 Response Time
Participants had the lowest response times when using the fixed-minute traffic map
(18.64 seconds). Response times associated with the conventional map and the equalinterval map were similar (26.03 and 28.07 seconds, respectively). Response times are
important measurements because they provide insight into the difficulty of a task and
can be used as a proxy for mental workload (Walker et al., 1991). Within this context,
it can be argued that the fixed-minute map is the most intuitive and easiest for map
readers to interpret.
While the conventional traffic map and the equal-interval map shared similar response times, the equal-interval design elicited higher accuracy in travel time estima-

73

tions. The similarity in response times is likely due to the similar tasks involved with
the interpretation of these map types. Both maps have a somewhat similar appearance
and both use hue as the primary visual variable. Furthermore, the addition of a fixed
segmentation scheme that was advantageous in pure distance estimation tasks did not
convey similar results in travel time estimation tasks. This indicates that it is perhaps the
combined arithmetic and velocity estimation that dictates the majority of time taken to
interpret these maps.
Research has demonstrated that lower response times are indicative of a lower mental workload associated with performing a cognitive task, such as map interpretation
(Walker et al., 1991; Parkes et al., 1991; Horberry et al., 2006; Horrey et al., 2009).
This suggests that the designs based on a fixed segmentation require less cognitive engagement from their users. While it cannot yet be said that this reduction in cognitive
engagement will manifest in safer, better-performing drivers or pedestrians that are
more aware while using these maps, it does nonetheless open the door for such a possibility. Noted by Ishigami & Klein (2009), paying attention to one thing requires the
withdrawal from another. If this is true, the converse must also be true: a facilitated
reduction in attention from one thing frees attention that can then be allocated elsewhere. It is therefore not unreasonable to hypothesize that maps associated with a
lower cognitive demand will also be associated with higher performance in concurrent
tasks.
5.3.3 Confidence
Fixed segmentation schemes elicited more frequent confident responses than conventional traffic maps. Nearly all responses to the fixed-minute traffic map were confident
(94.56%). This represents a frequency of confident responses that occurs more than
twice as often than those associated with the conventional traffic map (40.82% of responses were confident). The equal-interval traffic map improved confidence more

74

than 15% over the conventional map as well.
A similar relationship between the map types was observed with the number of responses marked “close.” The conventional map only caused participants to say their
answer was close 33.47% of the time. A slight improvement was associated with the
equal-interval map, with 39.18% of responses being close. Again, the fixed-minute
map proved far superior in comparison, eliciting close responses 90.41% of the time.
The largest difference between confident and close responses was noted with the
equal-interval traffic map. This suggests that the map initially invoked feelings of confidence that was not entirely warranted. On the other hand, only a small decreased was
observed between confident and close responses to questions about the fixed-minute
map. The implication then is that the fixed-minute map often rewarded participants;
they felt confident entering their response, and then they saw their response next to the
correct answer and the vast majority of the time were satisfied with their answer.
5.3.4 Map Preference and Subjective Feedback
Participants overwhelmingly preferred the fixed-minute traffic map (46 out of 49 participants chose it as their favorite), followed by the equal-interval map (chosen by only 3
participants). No participants preferred the conventional traffic map. Although it was
somewhat anticipated that participants would prefer the maps based on fixed segmentation schemes, the magnitude of the preference for the fixed-minute traffic map was
wholly unexpected. A review of the results indicates why this might have occurred. In
every metric used (error, response time, confidence, and whether or not the user felt
close), the fixed-minute traffic map outperformed the others. The fixed-minute map
represented a much simpler task that required less time and cognitive ability and did
not mislead participants by evoking undue confidence.
Moreover, participant responses as to why they chose the map they did reveals a
quality of map interpretation that is difficult to measure: participants simply enjoyed

75

using the map. Forty-six participants who chose the fixed-minute traffic map frequently
used the words “enjoyed” and “liked” in their feedback. The high levels of confidence
and preference associated with the fixed-minute traffic map suggest that users felt good
about making decisions with it, they were confident their decisions were correct, and
they remained satisfied with their responses after seeing the correct value. Previous
sections have demonstrated the empirical evidence that suggests a fixed-minute design
is superior for travel time estimation. However, these results provide insight into the
user experience, which on the whole, was extremely positive.
Much has been said regarding the improvements to both micro- and macro-scale
user behavior within a transportation system as a result of higher quality traffic information. But if nothing else, the fixed-minute design provided users with a map that
was not only enjoyable to use and strongly preferred, it was associated with very low
error and short response times. These alone are reasons that make a strong case for
improvements in traffic map design and a possible shift towards fixed segmentation
schemes — especially those based on units of time.

5.4 Limitations
The primary limitation of this study is that the prototype maps produced are not dynamic or readily deployable. Both the equal-interval and fixed-minute traffic map represent static snapshots in time. Although these maps were created successfully, the
methods used for their creation are not intended for real-time display of traffic conditions. Largely based on raster operations, the current methodology would be too
intensive to both computing power and time to be implemented for dynamic mapping
in their present form. Alternative methodologies might take advantage of technologies
such as Scalable Vector Graphics (SVG), Adobe Flash, or a raster-based overlay built
on the Google Maps Application Programming Interface (API).

76

This study also did not ask users to estimate velocities alone from the traffic maps
used. The decision to omit such questions was made deliberately. The classes used by
the conventional traffic map represent large ranges in velocities (e.g., 25 to 49.9 mph).
Any value submitted by a participant within those ranges would have to be considered
correct, assuming that the response fell within the correct class.
It is also not known which traffic map design enables users to quickly and most accurately determine the severity of congestion at a glance. Often times a precise amount
of time required to make a trip may not be the information a map-reader desires; he or
she may only be interested in where traffic is “good” or “bad.” It can by hypothesized
that a hue-based traffic map facilities this type of task the best, but empirical research
has not been done for this task. In either case, hue could be easily added to the fixedminute traffic map to enable users to determine accurate travel times or to acquire an
at-a-glance assessment of the traffic conditions.

77

Chapter 6
CONCLUSION
Congestion represents a geographic phenomenon that is pervasive, dynamic, and
detrimental to individuals and infrastructures alike. The size of the congestion problem
in the United States is difficult to overstate. Daily drivers in the U.S. number in the
millions, and annual costs — in both time and money — are represented in the billions.
Despite efforts, congestion in America is only getting worse (Schrank & Lomax, 2009).
At the same time, numerous studies have demonstrated that travelers have a need for
high quality traffic information (Chorus et al., 2007), and satisfying that need can lead
to improved driver behavior that affects the entire transportation system in a positive
way (Mahmassani & Jayakrishnan, 1991; Golledge, 2002). The problem of congestion
is one faced every single day and accounts for an enormous hindrance to travel, the
supply of goods, and less quantifiable quality of life factors.
Recent advancements in technology have enabled the real-time display of traffic
information on mobile devices of many types. Maps are commonly accessed in this
fashion, and traffic maps present a wealth of opportunity to provide drivers with up to
date information on current traffic conditions. Although other media has been used to
disseminate real-time traffic information, maps are preferable for two reasons.
First, a map depicts the information visually. Delivered in this way, the information
does not have to be read from top to bottom or left to right as text would; the map
reader need only glance at the area of interest. This is also preferable to audio delivery,
which requires the listener to attend to the message in its full duration in order interpret
its final meaning.
Secondly, traffic data presented in map form allows the user relate the information to
the spatial environment, fostering knowledge generation and a better understanding of
the severity of congestion and its extent. This facilities improved decision making that

78

is at the heart of ITS/ATIS strategies. These strategies are only effective when drivers
have the ability and information to make decisions, which often include re-routing or
altered trip plans. Traffic maps show many possible routes at once, making this possible.
Audio or text-based data delivery methods are route-specific, preventing users from
making such decisions or planning routes that may change at any time during travel.
It is clear that a segmentation scheme based on a fixed unit of measurement improves
distance estimation in maps, and in the case of traffic maps, travel time estimations as
well. These improvements were seen in the form of lower errors and often, lower response times which were sometimes substantially lower (refer to Figure 23 and Table
8). In no case did improvements in distance or travel time estimation come at the cost
of increased response times. Together, the pairing of lower error and quicker interpretation suggest that fixed segmentation schemes are superior to non- or ambiguouslysegmented designs for the purpose of travel time estimation.
Users also strongly preferred a fixed-minute segmentation for traffic maps. This type
of map represents a novel approach to symbolizing dynamic, time-based phenomena.
Although significantly high levels of success were seen with the fixed-minute traffic
map in this study, several questions remain. While the fixed-minute design improved
travel time estimations, it is not yet known if this design will hinder or facilitate more
coarse assessments of traffic conditions. Additionally, the performance of the fixedminute traffic map has not yet been evaluated on mobile displays, at multiple scales,
in oblique views, or in the presence of concurrent tasks. Future research could benefit
from additional analysis of this new map type, or attempts to create others like it.
This research has also shown the gap between cognitive cartography and traffic mapping can be easily bridged. Scientific literature is replete with studies on each topic
independently. Despite being seeming disparate, there are several overlaps between
traffic research, studies on mobile device use, and visual image interpretation.

79

Cartographers and the public alike have unprecedented access to both traffic and
geographic information, as well as technologies capable of displaying these data. As
technologies continue to emerge and evolve, it is essential that we understand how
these technologies are being used, how they may be used in the future, and to what
extent they can be improved. Above all, conventions must always be challenged and,
in the case of demonstrable proof, be replaced when superior alternatives exist.

80

APPENDICES

81

Appendix A: Python Code for Identifying 1-minute Segments
# Copyright 2011 Joshua E. Stevens
# Python code to identify 1−minute intervals from a prepared CSV
file
# This program is free software: you can redistribute it and/or
modify
# it under the terms of the GNU General Public License as published
by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
#
#
#

This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.

# You should have received a copy of the GNU General Public License
# along with this program. If not, see
<http://www.gnu.org/licenses/>.
#######################################################################
# Setup script to handle csv input/output
import csv
# Read in data
in_data = csv.reader(open(’segment_times.csv’, ’rb’))
# Get fields and data required
SegID = []
TimeSec = []
for row in in_data:
TimeSec.append(row[4]) # row 4 contains the times for each
segment
SegID.append(row[5]) # row 5 contains an ID unique to each
segment
# Search for 60−second markers and save their location
travelTime = 0
ticMark = []
i = 1
while(i < len(TimeSec)):
travelTime += float(TimeSec[i])
if(travelTime >= 60):
# Save locatation of tick mark
ticMark.append(SegID[i])
travelTime = 0 # Reset travel time to continue for next
segment
i += 1

82

# Write new CSV file to join later
outfile = ”ticMarks.csv”
output = open(outfile,”w”)
output.writelines(’”TicMark”’) # Make header
for each in ticMark:
output.writelines(’\n’)
output.writelines(each)
output.close()

83

Appendix B: Advertisement for Soliciting Participants

Figure 27: The flyer used to solicit participants.

84

Appendix C: Consent Form

Research Participant Information and Consent Form

You are being asked to participate in a research project. This study is being conducted by Joshua
Stevens, under the advising of Dr. Kirk Goldsberry. This document provides information about
the study, and the risks and benefits of participation to empower you to make an informed
decision. Participation in this study is completely voluntary. You should feel free to ask the
researchers any questions you may have.
Alternative Segmentation Schemes for the Design of Traffic Maps
Joshua Stevens, Graduate Student
Kirk Goldsberry, Assistant Professor
Department of Geography
Michigan State University
116 Geography Building, East Lansing, MI 48824
steve413@msu.edu, kg@msu.edu

1. PURPOSE OF THE RESEARCH:
You are being asked to participate in a research project about map-reading.
This study investigates the effectiveness of different techniques for mapping traffic data. Many
current maps are designed in such a way that they are easy to produce but do not perform well
for those who use them. This research aims to discover what aspects of these particular maps are
troublesome and to identify possible solutions for improving traffic maps.
2. WHAT YOU WILL DO:
You will be asked to submit answers to questions presented on a computer screen. These
questions will ask you to solve a basic arithmetic problem, estimate a distance between two
locations on a map, or to identify which of two routes is the shortest, longest, or fastest to travel.
Occasionally, questions will ask you to estimate travel speeds or the time necessary to complete
a route given the information provided. Some of these questions will be timed, though you will
not see the timer or know which question is being timed. After each response, you will be asked
how confident you are that your answer was correct. At times, you will be shown the value you
submitted along with the correct value and asked to if you believe whether your answer is similar
or ‘close.’ Use your best judgment for what you feel is similar based on the task being
performed.
Prior to the test you will answer basic questions about yourself. These questions include gender,
age group, and driving status. Once the test is complete you will have an opportunity to share
your opinions on the test and which maps you preferred.

85

3. YOUR RIGHTS TO PARTICIPATE, SAY NO, OR WITHDRAW:
Participation in this research project is completely voluntary. You have the right to say no. You
may change your mind at any time and withdraw. Your answers will then be destroyed and will
not be included in the results of this study. You may choose to stop participating at any time.
Should you choose to withdraw from the study, you will not receive the advertised
compensation. You will not, however, be penalized in any other way.
4. COSTS AND COMPENSATION FOR BEING IN THE STUDY:
You will not incur any costs for participating in this study. You will be compensated $10 for
completing the test in full. Withdrawal prior to the completion of the test will result in your
compensation being prorated for the time spent up to that point, to the nearest $1.
5. CONTACT INFORMATION FOR QUESTIONS AND CONCERNS:
If you have concerns or questions about this study, such as scientific issues or how to do any part
of it, please contact the researcher Joshua Stevens (Graduate Student, Department of Geography,
Michigan State University, 116 Geography Building, East Lansing, MI 48824,
steve413@msu.edu, 517-507-2292) or Dr. Kirk Goldsberry (Assistant Professor, Department of
Geography, Michigan State University, 116 Geography Building, East Lansing, MI 48824,
kg@msu.edu, 517-353-0308)
If you have questions or concerns about your role and rights as a research participant, would like
to obtain information or offer input, or would like to register a complaint about this study, you
may contact, anonymously if you wish, the Michigan State University’s Human Research
Protection Program at 517-355-2180, Fax 517-432-4503, or e-mail irb@msu.edu or regular mail
at 207 Olds Hall, MSU, East Lansing, MI 48824.
6. DOCUMENTATION OF INFORMED CONSENT.
Your signature below means that you voluntarily agree to participate in this research study.

________________________________________
Signature

You will be given a copy of this form to keep.

86

_____________________________
Date

REFERENCES

87

References
Al-Deek, H., & Khattak, A. (1998). A Combined Traveler Behavior and Transportation
System Performance Model with ATIS. Transportation Research A, 32, 479–493.
Ascone, D., Lindsey, T., & Varghese, C. (2010). An Examination of Driver Distraction as
Recorded in NHTSA Databases. Technical Report.
Biederman, I. (1987). Recognition-by-Components: A Theory of Human Image Understanding. Psychological Review, 94, 115–147.
Brewer, C. A., & Pickle, L. (2002). Evaluation of Methids for Classifying Epidemiological Data on Choropleth Maps in Series. Annals of the Association of American Geographers,
92, 662–681.
Brookhuis, K. A., & de Waard, D. (2010). Monitoring Drivers’ Mental Workload in
Driving Simulators Using Physiological Measures. Accident Analysis and Prevention, 42,
898–903.
Chorus, C. G., Arentze, T. A., Timmermans, J., Molin, E. J. E., & Wee, B. V. (2007).
Travelers’ Need for Information in Traffic and Transit: Results from a Web Survey.
Journal of Intelligent Transportation Systems, 11, 57–67.
Clarke, K. C. (2004). Mobile Mapping and Geographic Information Systems. Cartography and Geographic Information Science, 31, 131–136.
Clarmunt, C., Jiang, B., & Bargiela, A. (2000). A new framework for the integration,
analysis and visualisation of urban traffic data within geographic information systems. Transportation Research C, (pp. 167–184).
Drews, F. A., Yazdani, H., Godfrey, C. N., Cooper, J. M., & Strayer, D. L. (2009). Text
messaging during simulated driving. Human Factors, 51, 762–770.
Fujii, S., Kitamura, R., Nagao, M., & Doi, Y. (2009). An experimental analysis of intelligibility and efficiency of in-vehicle route guidance system displays. Transportation,
36, 779–786.
Funke, G., Matthews, G., Warm, J. S., & Emo, A. K. (2007). Vehicle automation: A
remedy for driver stress? Ergonomics, 50, 1302–1323.
Golledge, R. (2002). Dynamics and its: Behavioral responses to information available
from atis. In H. Mahmassani (Ed.), Perceptual Motion: Travel Behavior Research Opportunities and Application Challenges (pp. 81–126). Elsevier Science.
Gollege, R., & Garling, T. (2004). Cognitive Maps and Urban Travel. In K. J. H.
D. A. Hensher, K. J. Button, & P. R. Stopher (Eds.), Handbook of Transport Geography
and Spatial Systems (pp. 501–512). Oxford:Elsevier volume 5.

88

Horberry, T., Anderson, J., Regan, M. A., Triggs, T. J., & Brown, J. (2006). Driver
distraction: The effects of concurrent in-vehicle tasks, road environment complexity
and age on driving performance. Accident Analysis and Prevention, 38, 185–191.
Horrey, W., Lesch, M., & Garabet, A. (2009). Dissociation between driving performance and drivers’ subjective estimates of performance and workload in dual-task
conditions. Journal of Safety Research, 40, 7–12.
Ishigami, Y., & Klein, R. M. (2009). Is a hands-free phone safer than a handheld
phone? Journal of Safety Research, 40, 157–164.
Itti, L., & Koch, C. (2000). A saliency-based search mechanism for overt and covert
shifts of visual attention. Vision Research, 40, 1489–1506.
Jeffries, B. J. (1880). On Color Blindness. Science, 1, 197.
Jenks, G. F., & Caspall, F. C. (1971). Error on choropleth maps. Annals of the Association
of American Geographers, 61, 217–244.
Just, M. A., Keller, T. A., & Cynkar, J. (2008). A decrease in brain activation associated
with driving when listening to someone speak. Brain Research, 1205, 70–80.
Kienker, P., & Sejnowski, T. (1986). Separating figure from ground with a parallel
network. Perception, 15, 197–216.
Kutila, M. H., Jokela, M., Makinen, T., Viitanen, J., Markkula, G., & Victor, T. W.
(2007). Driver cognitive distraction detection: feature estimation and implementation. Proceedings of the Institution of Mechanical Engineers Part D-Journal of Automobile
Engineering, 221, 1027–1040.
Leiser, D., & Zilbershatz, A. (1989). The traveller: a computational model of spatial
network learning. Environment and Behavior, 21, 435–463.
Liben, L., & Downs, R. (1989). Understanding maps as symbols: The development of
map concepts in children. Advances in Child Development and Behavior, 22, 145–201.
Lindauer, M., & Lindauer, J. (1970). Brightness differences and the perception of figureground. Journal of Experimental Psychology, 84, 291–295.
Liu, B., Wang, Z., & Jin, Z. (2009). The integration processing of the visual and auditory
information in videos of real-world events: An ERP study. Neuroscience Letters, 461,
7–11.
Liu, H. X., & Danczyk, A. (2009). Optimal Sensor Locations for Freeway Bottleneck
Identification. Computer-Aided Civil and Infrastructure Engineering, 24, 535–550.
MacEachren, A. M. (2004). How Maps Work. New York: The Guilford Press.

89

Mahmassani, H. S., & Chang, G.-L. (1987). On boundedly-rational user equilibrium
in transportation systems. Transportation Science, 21, 89–99.
Mahmassani, H. S., & Jayakrishnan, R. (1991). System performance and user response
under real-time information in a congested traffic corridor. Transportation Research A,
25, 293–307.
Murray, E. (1943). Anomalies in Color Vision. The Scientific Monthly, 57, 322–331.
Parkes, A. M., Ashby, M. C., & Fairclough, S. H. (1991). The effects of different invehicle route information displays on driver behavior. In Vehicle Navigation and Information Systems Conference (pp. 61–70). Dearborn, MI: Society of Automotive Engineers,
Inc.
Patten, C. J. D., Kircher, A., Ostlund, J., Nilsson, L., & Svenson, O. (2006). Driver
experience and cognitive workload in different traffic environments. Accident Analysis
and Prevention, 38, 887–894.
Robinson, A. H. (1952). The Look of Maps. Madison, WI: Univsersity of Wisconsin
Press.
Rodriguez-Sanchez, A. J., Simine, E., & Tsotsos, J. K. (2007). Attention and Visual
Search. International Journal of Neural Systems, 17, 1–14.
Rosenbloom, S. (1978). Peak-period traffic congestion: a state-of-the-art analysis and
evaluation of effective solutions. Transportation, 7, 167–191.
Saint-Martin, F. (1990). Semiotics of Visual Language. Advances in Semiotics. Bloomington: Indiana University Press.
Schrank, D., & Lomax, T. (2009). 2009 Urban Mobility Report. Technical Report Austin,
TX.
Schwartz, D. L. (1995). Reasoning about the referent of a picture versus reasoning
about the picture as the referent: An effect of visual realism. Memory & Cognition, 23,
709–722.
Srinivasan, R. (1997). Effect of Selected In-Vehicle Route Guidance Systems on Driver
Reaction Times. Human Factors, 39, 200–215.
Tobler, W. (1973). Choropleth Maps Without Class Intervals. Geographical Analysis, 5,
262–265.
Ullman, S. (1984). Visual Routines. Cognition, 18, 97–159.
Walker, J., Alicandri, E., Sedney, C., & Roberts, K. (1991). In-vehicle navigation devices: effects on the safety of driver performance. In Vehicle Navigation and Information
Systems Conference (pp. 499–525). Dearborn, MI: Society of Automotive Engineers,
Inc.

90

Wittmann, M., Kiss, M., Gugg, P., Steffen, A., Fink, M., Poppel, E., & Kamiya, H.
(2006). Effects of display position of a visual in-vehicle task on simulated driving.
Applied Ergonomics, 37, 187–199.

91