Tilt Map: Interactive Transitions Between Choropleth Map, Prism Map and Bar Chart in Immersive Environments

Full text

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 1
Tilt Map: Interactive Transitions Between
Choropleth Map, Prism Map and Bar Chart
in Immersive Environments
Yalong Yang, Tim Dwyer, Kim Marriott, Bernhard Jenny and Sarah Goodwin
Abstract—We introduce Tilt Map, a novel interaction technique for intuitively transitioning between 2D and 3D map visualisations in
immersive environments. Our focus is visualising data associated with areal features on maps, for example, population density by state.
Tilt Map transitions from 2D choropleth maps to 3D prism maps to 2D bar charts to overcome the limitations of each. Our paper
includes two user studies. The first study compares subjects’ task performance interpreting population density data using 2D
choropleth maps and 3D prism maps in virtual reality (VR). We observed greater task accuracy with prism maps, but faster response
times with choropleth maps. The complementarity of these views inspired our hybrid Tilt Map design. Our second study compares Tilt
Map to: a side-by-side arrangement of the various views; and interactive toggling between views. The results indicate benefits for Tilt
Map in user preference; and accuracy (versus side-by-side) and time (versus toggle).
Index Terms—Immersive analytics, Mixed / augmented reality, Virtual reality, Geographic visualization, Interaction techniques.
F
1 INTRODUCTION
CHOROPLETH maps are arguably the most widely used
visualisation for showing data linked to geographic
areas, such as population density [1], [2]. These maps colour
or shade the areas on the map to indicate the associated
values. A less common way of showing such area-linked
data is the prism map [3], [4] where areas are extruded
into the third dimension so that the height of the “prism”
represents the associated value. With the arrival of commod-
ity head-mounted displays (HMDs) for virtual reality (VR)
and augmented reality (AR), we can expect to see choro-
pleth maps, prism maps, and other area-linked geographic
visualisations used in immersive applications. However,
area-linked geographic visualization is under explored in
immersive environments and the trade-off between the
two-dimensional choropleth map and the three-dimensional
prism map is currently unclear.
In this research we explore whether traditional 2D choro-
pleth maps are the best way to show area-linked data in
such immersive environments, or whether other visuali-
sations, such as a prism maps that make use of a third
dimension, or some combination of these may be better. This
paper makes three main contributions.
The first contribution is a controlled study comparing
choropleth, prism and coloured prism maps for the first
time in VR. We found that participants were more accurate
using the prism maps but faster using the choropleth maps.
Participants preferred the coloured prism map, but raised
some concern about occlusion. These results accord with
•Yalong Yang was with the Department of Human-Centred Computing,
Faculty of Information Technology, Monash University, Australia. He is
now with School of Engineering and Applied Sciences, Harvard Univer-
sity, Cambridge, MA, 02138. E-mail: yalongyang@g.harvard.edu
•Tim Dwyer, Kim Marriott, Bernhard Jenny and Sarah Goodwin are
with the Department of Human-Centred Computing, Faculty of Infor-
mation Technology, Monash University, Australia. E-mail: {tim.dwyer,
kim.marriott, bernie.jenny, sarah.goodwin}@ monash.edu
Manuscript received April 19, 2005; revised August 26, 2015.
previous studies comparing choropleth and prism maps
in non-VR settings, confirming a trade-off between faster
choropleth maps and more accurate prism maps.
Our second contribution is Tilt Map, a novel interactively-
controlled transition between three views: a choropleth, a
prism map and a bar chart. We incorporate a bar chart as
an additional view, as we thought this would aid compar-
ison tasks and alleviate the difficulties of 3D occlusion and
perspective foreshortening. In the Tilt Map, orientation de-
termines the view. With a vertical orientation, the choropleth
map is shown. As the viewer tilts the map it morphs into the
prism map, and when the map nears a horizontal orienta-
tion, it morphs into the bar chart (Fig. 1). We believe this use
of orientation to select the view that is appropriate to the
user’s view angle is new and opens up many possibilities
for the design of immersive visualisations.
Our third contribution is a controlled study comparing Tilt
Map with two other conditions: a Side-by-Side complemen-
tary arrangement (of choropleth, coloured prism, and bar
chart) and a Toggle representation, which switches between
the three views with a controller click. We evaluated user
preference and performance. The results indicate benefits
for Tilt Map in user preference and accuracy (versus side-
by-side) and time (versus toggle).
Our work contributes to the growing body of knowledge
of how to present data with a geographical embedding
in immersive environments [5], [6], [7]. It provides further
evidence that for geographically embedded data, there can
be benefits in utilising a third dimension to show the data
variable. We also introduce a new kind of interaction specif-
ically suited to immersive data visualisation. Just as large
tiled wall displays allow the use of proxemic interaction to
naturally control the choice of presentation [8], the ability
to hold and tilt virtual artefacts, such as maps, in immer-
sive environments provides an intuitive embodied method
for transitioning between different views, with the aim of
providing the view that is best suited to the viewing angle.

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 2
Fig. 1: Orientation-dependent visualisation with a Tilt Map: The user can tilt a prism map (left) to morph it to either a
vertical choropleth map or a bar chart (middle). The transition stages with angle intervals were labeled from ato g(right).
We demonstrate the transition stages with visualisations in Fig. 8.
2 RE LATED WORK
The choropleth map—where area-based values are encoded
via colour, shading or pattern—is one of the most commonly
used thematic map types [1], [2], [9]. Whilst using the
positions in the display space for representing geographic
information is fundamental to reveal spatial patterns, the
attribute values have to be represented by non-spatial visual
variables, such as colour and brightness. Raw-total values
are avoided on choropleth maps because large geographic
areas likely have large corresponding mapped quantities.
Choropleth maps with non-uniform areas should instead
display data in the form of density measures, proportions,
or ratios [1], [2], [4].
A prism map is a 3D choropleth map with extruded
height to encode a numerical attribute [3], [4]. Prism maps
are quite uncommon. Most cartographic textbooks mention
them only en passant when discussing choropleth maps, with
the recent exception of Field, who mentions that “prism maps
are predominantly used for visual impact”, and speculates that
the unfamiliarity with prism maps hampers their under-
standing [4]. Although less common, cartographers have
used prism maps for many years, especially since com-
puter software for the generation and animation of prism
maps became available [10], [11], [12], [13]. For static prism
maps, it has been shown that readers consistently associate
prism height, and not prism volume, with data value [14].
This is the case for both absolute values and proportional
values [14], which implies that converting quantities to
densities or ratios is not necessary for prism maps.
The question of whether it is preferable to use choropleth
maps or prism maps has no immediate answer. Height
(the visual variable used by prism maps) is far more ac-
curate for interpreting associated quantitative values than
brightness (the visual variable commonly used by choro-
pleth maps) [15], but perspective foreshortening and oblique
viewing angles can result in distortion and excessive oc-
clusion in prism maps. Few studies compare prism maps
to choropleth or other thematic map types. One study [16]
evaluated short-term learning benefits of choropleth maps
and prism maps, using uniform gridded areas for both
map types instead of the more common irregular areas.
It found that the choropleth map improved participants’
ability to correctly assess detailed information (ranking of
city populations), while the prism map appeared slightly
more useful for reading general patterns (overall population
distribution). Another study compared prism maps and
area-proportional circle maps, and found similar reading
accuracy for both [14]. Popelka [17] compared prism maps,
standard choropleth maps and illuminated choropleth maps
(where illumination effects give flat areas a subtle 3D prism-
like appearance [18]). The task involved comparing the
values of two marked areas. Prism maps resulted in more
accurate but slower responses than choropleth maps. Mean-
while, studies by Bleisch, Dykes & Nebiker [19] and Seipel
& Carvalho [20] indicate that reading accuracy of bar charts
heights in 2D and 3D maps is similar. Creating 3D maps
is also supported in many commercial products such as
ArcGIS, Mapbox, kepler.gl, rayshader. Therefore, although
3D mapping techniques are not as popular as 2D ones,
it seems reasonable to assume that encoding quantitative
values with prism heights is a practical design choice.
The mentioned studies have used static maps on paper
or standard 2D computer displays. Our research provides
the first comparison of choropleth and prism maps in an
immersive environment where the viewing angle can be
easily adjusted. In general, there has been surprisingly
little research into thematic cartography in immersive en-
vironments. Only recently have researchers begun to sys-
tematically investigate immersive geospatial data visuali-
sation [21]. Yang et al. explored immersive visualisation of
origin-destination flow maps [6] and of maps and globes
in virtual reality [5], [22]. They found clear benefits for
the use of 3D representations. Quang and Jenny placed
bar graphics in a virtual landscape and found that linking
the bars with bar charts and maps with bars improves
performance [23]. Wagner et al. compared space-time cubes
in virtual reality and on 2D displays [7]. They found sim-
ilar tasks performance but the immersive version received
higher subjective usability scores. Furthermore, immersive
environments support embodied interaction [24] and allow
the map reader to adjust position, size and scale in engaging
ways [25]. Our research combines embodied interaction,
animated transitions between data graphics [26] and inter-
active 3D geovisualisation [27], resulting in a novel type of
cartographic visualisation that adjusts to the tilt angle.
3 ST UDY 1: PRI SM MAP V S CHOROPLETH MAP
Our first user study compared three different visual rep-
resentations of areal population density data in VR: 2D
Choropleth map; 3D Monochrome Prism map; and 3D double-
encoded Coloured Prism map (Fig. 2). A previous study [17]
using 2D desktop displays found that task performance

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 3
Fig. 2: Study 1: Evaluated Choropleth,Monochrome Prism and Coloured Prism maps, demonstrating US examples.
with prism maps was more accurate but slower than with
choropleth maps. We wished to see if this changed with
the additional depth cues provided by head-tracking and
stereoscopic presentation in modern VR environments. We
evaluated the visualization conditions with one elementary
task and one synoptic task (see Section 3.2).
3.1 Visualisations and Interactions
We map areal population density data in different encodings
in our tested conditions.
Monochrome Prism: We mapped the density values linearly
to height.
Choropleth: In order to compare directly to height, we
used a continuous sequential colour scheme for density val-
ues, resulting in an unclassed choropleth map. The colour
scheme was derived through linear interpolation of the
YlOrBr palette from ColorBrewer [28]. This colour scheme
was chosen as it is colour-blind friendly [28] and popular in
cartography literature (e.g. in [29], [30]).
Coloured Prism: We double-encoded the density values
linearly with height (as in Monochrome Prism) and with
colour (as in Choropleth).
Encodings common to all conditions: (a) Borders were
added to the boundaries of geographic areas as pilot tests
revealed participants preferred them; (b) soft shadow was
enabled in both prism map conditions, but was not neces-
sary for the 2D Choropleth maps; (c) legends were placed
at top, bottom, left, and right sides of all maps. Legends
ranged from 0 to 100 with ticks every 5 and labelled ticks at
0, 25, 50, 75 and 100. For Choropleth, the legends lay in the
plane of the maps to show colour mapping. For Monochrome
Prism, the legends were cylinders standing perpendicularly
on the plane of the base map and extruded in parallel to
the prism heights. Coloured Prism legends were the same as
Monochrome Prism but also coloured as for Choropleth.
Interaction: We provided the same interactions for all three
visualisations. First, viewers could move in space to change
their viewpoint. Second, viewers could pick up the map
using a standard hand-held VR controller, and reposition
or rotate it in 3D space using clutched grasping with the
controller trigger. Users could not resize or scale the visuali-
sations but could physically zoom by moving closer. We did
not allow other interactions such as filtering as we wished
to focus on base-line readability of the representations.
3.2 Experiment
In this subsection, we first introduce the tasks of the user
study and the way we create task data. We then report details
of the user study including: experimental set-up,design and
procedure,participants and measures.
Tasks: Researchers [31], [32], [33] distinguish tasks in ge-
ographic data visualisation into two levels: elementary and
synoptic. Elementary tasks refer to single elements while
synoptic tasks involve a set of elements. Following this
taxonomy and previous related studies [16], [17], [18], we
designed one task of each type for the first study:
Area-Comparison Task: Compare the density values of two
given geographic areas. Initially, in our pilot study, we asked
participants to identify the area with the larger density
value in two given geographic areas. The same task was
tested by [17], [18] on 2D computer displays. We found
participants can answer this question easily with very high
accuracy in all conditions. Inspired by [34], [35], instead of a
binary result, we asked the participants to perform the more
difficult task of estimating the numeric difference between
two given geographic areas.
Region Task: Estimate the population density of a region con-
sisting of contiguous marked areas on the map. A similar task
was tested by Niedomysl et al. [16] on printed A4 size maps.
We randomly chose regions of 5 contiguous states in each
US map and 15–20 contiguous LADs in each UK map. The
correct answer is the area-weighted average of population
density across geographic areas of the region. We explained
this task in detail with examples to make sure participants
fully understood, emphasising that larger geographic areas
contribute more to the total region weight. After the ex-
planation all participants reported they understood. Again,
participants needed to provide a numeric answer.
In pilots, we highlighted the borders of target areas
(i.e. the two given areas in the area-comparison task and
the set of contiguous areas in the region task), following
Harrower [36]. However, participants reported extra effort
to visually search for the two targets in the area-comparison
task. As the intention of this study was not to examine the
time for visual search, we chose to further mark the two
areas with leader lines, which proved adequate for rapid
target identification. In region tasks, no participant reported
difficulty in identifying the target regions as the highlighted
borders of multiple adjacent geographic areas made the
region highly visually salient. Thus, we did not use leader
lines in region tasks.
For the area-comparison task, we anticipated that dis-
tance between the two targets was likely to affect the perfor-
mance. We randomly sampled pairs of areas and computed
the great-circle distance between areas. We then created two
categories for the area-comparison tasks: close and far. We
considered great-circle distances below 3° in the US and
below 0.5° in the UK as close and within 25°–28° in the US
and within 5°–5.5° in the UK as far. For the region task,

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 4
we identified that the degree of variation of density in the
region can affect performance. For example, in the extreme
case there is no variation, and the participant only needs to
ascertain the value of a single geographic area to answer the
region task. We used the coefficient of variation (CV) [37] to
measure the variation of density values. Again, we sampled
CVs from random regions. Based on the distribution, we
decided to use CVs within 40%–60% for all region tasks.
We mapped the data values linearly to colour and height,
indicated by legends as described in Sec. 3.1 in the range 0–
100. Piloting revealed that results below 20 and above 80
were too easy to distinguish. Thus, we controlled the results
of tasks to the range of 20–40, 40–60 and 60–80.
Data: We used 2018 population density data for the United
States (US) [38] and the United Kingdom (UK) [39]. For
the US, we used the conterminous 48 States and for the
UK, we used the 391 Local Authority Districts (LAD). The
original datasets had very skewed distributions, as they
contained a small number of areas with high population
density. This would have made it trivial for participants to
provide correct answers. We therefore applied a square root
transformation to make the tasks more challenging and to
ensure the map visualisations used the entire range of colour
or height variation.
To minimise the learning effect, we used different data
for each question. We generated data based on the trans-
formed population density data. As spatial autocorrelation
structure is important in spatial analysis, we used the tech-
nique from [29] to ensure the Moran’s Iof the generated
data was the same as that of the original data.
Experimental Set-up: We used an HTC Vive Pro with
110° field of view, 2160×1200 pixels resolution and 90Hz
refresh rate. The PC was equipped with an Intel i7-6700K
4.0GHz processor and NVIDIA GeForce GTX 1080 graphics
card. Only one hand-held VR controller was needed in the
experiment: participants could use this to reposition and
rotate the map in 3D space. The frame rate was around
110 FPS.
Visuals were positioned comfortably within users’ reach.
The map was created on top of a transparent quadrilateral
measuring 1×1m, and placed 0.6m in front of the partic-
ipants’ eye position and 0.1m below it, tilted to 45°. We
repositioned the map at the beginning of every question.
The height of the Monochrome Prism and Coloured Prism
maps was linearly mapped to the range of 0−20 cm.
Design and Procedure: The experiment was within-
subjects: 3 visualisations ×2 task ×2 datasets ×3 answer
ranges ×2 repetitions (1 close and 1 far for area-comparison
task) = 72 responses per participant and lasted 1.5 hours
on average. We used the Latin square design to balance the
order of the three tested visualisations, i.e. each visualisation
occurred in every position in the ordering for an equal
number of participants.
Participants were first given a brief introduction to the
experiment. After putting on the VR headset, we asked them
to adjust it so as to clearly see the sample text in front
of them. Two types of training were provided: interaction
training and task training. 1) Interaction training was con-
ducted when each visualisation was presented for the first
time. The participants were introduced to the visualisation
and encoding. Then we gave them sufficient time to famil-
Fig. 3: Study 1 — Accuracy and response time with 95%
confidence intervals (AbsDiff = absolute difference between
participants’ answers and the correct answers.)
iarise themselves with the VR headset, controller, visualisa-
tion, encoding, and repositioning and rotating interactions.
2) Task training was conducted when each condition (task ×
visualisation) was presented for the first time. Two sample
tasks, differing from the experimental tasks, were given to
participants with unlimited time. After participants finished
a training task, we displayed the correct answer and allowed
them to confirm the answer with the visualisation. We asked
the participants to check their strategies both when they
were doing the training tasks and after the correct answers
were shown.
Participants first completed the area-comparison task,
then the region task with a 5-minute break in between. For
each task, the three visualisations were presented in a coun-
terbalanced order. After completing the tasks, a post-hoc
questionnaire recorded feedback on: (1) preference ranking
of visualisations in terms of visual design and ease of use for
the tasks; (2) rated confidence with each visualisation with a
five-point Likert scale; (3) advantages and disadvantages of
each visualisation; (4) strategies for different visualisations;
and (5) background information about the participant. The
questionnaire listed visualisations in the same order as
presented in the experiment.
Participants: We recruited 12 participants (2 female and 10
male) from our university. All had normal or corrected-
to-normal vision and included university students and re-
searchers. All participants were within the age group 20–30.
VR experience varied: 7 participants had less than 5 hours
of prior VR experience, 2 had 6–20 h, and 3 had more than
20 h. While we used colour encoding in the study, we chose
a colour-blind safe colour scheme (see Sec. 3.1). Therefore,
we did not test participants for colour blindness.
Measures: We measured time from first render of the visu-
alisation to double-click of controller trigger. After double-
click, the visualisation was replaced by a slider to select an
integer answer from 0–100. Participants could precisely step
adjust their answer by tapping the sides of the touchpad. We
report error rate as absolute difference between participants’
answers and the correct answers. We also recorded the
position and rotation of the headset, controller and map
every 0.1 seconds.

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 5
Fig. 4: Study 1 — Graphic depiction of statistic comparisons.
3.3 Results
Since the distributions of dependent variables were skewed,
we analyzed the square root of both error rate (absolute
difference, abbreviated to AbsDiff) and time. Linear mixed
modeling was used to evaluate the effect of independent
variables on the dependent variables [40]. All independent
variables and their interactions were modeled as fixed ef-
fects. A within-subject design with random intercepts was
used for all models. The significance of the inclusion of an
independent variable or interaction terms were evaluated
using the log-likelihood ratio. Tukey’s HSD post-hoc tests
were then performed for pair-wise comparisons using the
least square means [41]. Homoskedasticity and normality
of the Pearson residuals were evaluated graphically using
predicted vs residual and Q—Q plots respectively. Degree
of freedom, χ2and pvalue for fixed effects were reported
following [42, p. 601]. Accuracy (AbsDiff) and time with 95%
confidence of different visualisations are shown in Fig. 3.
Area-Comparison Task: Three independent factors (visual-
isation or V, country or C and distance or D) and their inter-
actions (V×C, V×D, C×D and V×C×D) were modeled.
The type of visualisation had a statistically significant
effect on AbsDiff (χ2(2) = 9.2, p =.0101). Post-hoc tests
revealed that Monochrome Prism and Coloured Prism were
statistically more accurate than Choropleth (see Fig. 4).
The type of visualisation also had a statistically signifi-
cant effect on time (χ2(2) = 63.4, p < .0001). Post-hoc tests
revealed that Choropleth was statistically faster than both
Monochrome Prism and Coloured Prism (see Fig. 4).
Region Task: Two independent factors (V and C) and their
interaction (V×C) were modeled.
The type of visualisation had a statistically significant
effect on AbsDiff (χ2(2) = 7.6, p =.0223). Post-hoc tests
revealed that Monochrome Prism was statistically more accu-
rate than Choropleth (see Fig. 4). The type of country had a
marginally significant effect on AbsDiff (χ2(1) = 2.7, p =
.0977). Post-hoc tests revealed that participants tended to be
more accurate in US than UK (p=.0991).
The type of visualisation also had a statistically signif-
icant effect on time (χ2(2) = 41.5, p < .0001). Post-hoc
tests revealed that Choropleth was statistically faster than
both Monochrome Prism and Coloured Prism (see Fig. 4).
The type of country had a marginally significant effect on
time (χ2(1) = 3.4, p =.0661). Post-hoc tests revealed that
participants tended to spend more time on US than UK with
p=.0672. However, the difference was small (see Fig. 3).
Plotting the AbsDiff and time in the order of trials in the
study revealed no obvious learning effect.
Fig. 5: Time percentage in different movements.
Fig. 6: Tilt angle distribution for the different maps with
median and third quartile lines.
Interactions: In order to investigate user interaction, we
sampled every frame. If the head or the map moved more
than 1 cm or rotated more than 5°, we classified it as an
interaction, see Fig. 5. We accumulated the interaction time
for each question and then normalised with respect to the
total time spent on that question. The linear mixed model
showed no statistically significant effect of visualisations on
head movement, but on map movement (χ2(2) = 56.90, p <
.0001). Participants moved Monochrome Prism and Coloured
Prism more often than Choropleth (all statistically significant
at p < .0001). This is possibly because participants moved
the maps more to deal with the occlusion in Monochrome
Prism and Coloured Prism.
We also analyzed the tilt angle of the maps, i.e. the
angle between the normal vector of the map plane and
the horizontal plane (see Fig. 6). Due to the non-normally
distributed residuals, we used the Friedman test to compare
the percentage of time spent with a view angle larger
than 45°. The test revealed a statistically significant effect
(χ2(2) = 12.67, p =.0018). As one might expect, partic-
ipants spent statistically more time with a large tilt angle
in Monochrome Prism and Coloured Prism than Choropleth,
all p<.05. In Fig. 6, we see a peak of around 45° for all
visualisations, the probable reason is that maps were tilted
to 45° at the beginning of each question.
User preference: For visual design, in Fig. 7(a), the strongest
preference was for Coloured Prism. 75% participants ranked
it as the best. The Friedman test revealed a statistically
significant effect of visualisation on preference (χ2(2) =
13.5, p =.0012). The post-hoc tests found a statistically
stronger preference for Coloured Prism than Monochrome
Prism with p=.0007.
For readability, in Fig. 7(b), the strongest preference was
again for Coloured Prism. 75% participants ranked it as the
best. The Friedman test revealed a statistically significant ef-
fect of visualisation on preference (χ2(2) = 10.2, p =.0062).
The post-hoc tests found a statistically stronger preference
for Coloured Prism than Monochrome Prism (p=.0119) and
Choropleth (p=.0216).
For confidence, in Fig. 7(c), the strongest preference was
again for Coloured Prism with 83.3% rating it 4 or 5. The
Friedman test revealed a statistically significant effect of
visualisation on confidence (χ2(2) = 8.93, p =.0115).
The post-hoc tests found participants felt statistically more

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 6
Fig. 7: User preference: (a, b) ranking for each of the three vi-
sualisations; (c) confidence rating in five-point-Likert scale.
Arrows indicate statistic comparisons with p<.05.
confident with Coloured Prism than with Monochrome Prism
(p=.0261) and with Choropleth (p=.0261).
Strategies: Participant responses to the questionnaire pro-
vided insight into their strategies for the two tasks. They
did not mention using special strategies for Choropleth. For
Monochrome Prism and Coloured Prism, participants tended
to use similar strategies. For the area-comparison task, some
described how they computed the value of an occluded tar-
get. They searched for a geographic unit without occlusion
adjacent to the target and read its value from the legend.
Then they used this geographic unit as the “yardstick”
to determine the target’s value. For the region task, two
strategies were described: Cut and fill–imagine cutting the
tall ones and using the volume to fill the lower ones; Divide
and mix–divide the set into two halves with similar size and
“do the math”.
Key Findings: The results of this first comparison of 2D
choropleth maps with 3D prism maps in an immersive en-
vironment accord with those of [17], [18] using 2D monitors:
participants are more accurate but slower with prism maps
than choropleth maps. This is in line with early studies
showing that people are more able to discriminate differ-
ences in the size of marks than colour [15]. Specifically we
found:
•Monochrome Prism was more accurate than Choropleth.
•Choropleth was faster than both Monochrome and Coloured
Prism.
•Coloured Prism was more accurate than Choropleth for the
area-comparison task.
We also found that participants tended to move maps more
with Monochrome and Coloured Prism than Choropleth and
tended to look from the side more with Monochrome and
Coloured Prism than Choropleth. Comments such as “regions
that were blocked were hard to read” suggest that this may be
because participants sought to find viewing angles to reduce
occlusion and perspective distortion.
Participants felt more confident with Coloured Prism
than with Choropleth and Monochrome Prism and preferred
Coloured Prism to Choropleth and Monochrome Prism in terms
of both visual design and readability.
4 TI LT MAP:AHYBRID 2D A ND 3D METHOD
Study 1 found a trade-off between speed and accuracy for
choropleth and prism maps in immersive environments.
One could interpret this to suggest that choropleth pro-
vides a good “overview” for quick comparisons between
regions, however, an encoding that uses the third dimension
to provide a spatial mapping of value (such as prism) is
better for precise reading. We therefore felt that a hybrid
Fig. 8: Tilt Map combines (a) Choropleth, (c) Prism, and (g) Bar
Chart views with intermediate transitions. The current view
is controlled by the tilt angle.
combination of the two views might have advantages over
either on its own. Furthermore, while using height to encode
data in prism maps improved accuracy, their 3D nature
led to issues with occlusion and perspective distortion. We
therefore thought that it would be beneficial to provide a
complementary 2D bar chart view.
Complementary visualisations are commonly combined
in two ways. The Tableau data visualisation software, for
example, allows for the creation of tabbed displays where
users can switch between different views that take up the
whole screen. Such tools also support the creation of dash-
board displays with tiled visualisations positioned side-by-
side. Usually in such a tiled display each individual view

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 7
shows a different subset or different dimensions of the
data. It is also possible to use perceptually complementary
displays [43]. These show the same data using different
visualisation techniques, each of which is better suited to
specific analysis tasks.
A third option for combining views, however, is to
unify the best aspects of each view into a single conceptual
display. Indeed, the Coloured Prism can be considered as
a hybrid representation. Viewed directly from above, it
appears like a choropleth map, but not perfectly. Linear
perspective means that the sides of prisms at the periphery
of the view are visible and that areas that are closest to
the viewer will appear larger, exaggerating their quantity.
Furthermore, lighting effects can cause shadows to be cast
across lower areas. From the side (which Fig. 6 shows was a
common viewing angle in Study 1) the Prism map appears
like a bar chart. But it is a 3D bar chart that suffers from
occlusion and perspective distortion.
Using the “magic” of virtual reality, however, we can
correct for all of these issues. As illustrated in Fig. 8, we can
dynamically morph the 3D model in response to tilt angle.
When viewed from directly above we can completely flatten
the model so it is precisely a 2D choropleth map (Fig. 8a). As
the user rotates the model (a natural gesture using the 6DOF
tracked VR controller), we can scale the height of the prisms
directly with view angle (Fig. 8b), until the visualisation is a
true prism map at about 45◦(Fig. 8c). As the user continues
to rotate the model we can further adapt the display, flatten-
ing the prisms to eliminate foreshortening and sliding them
sideways to remove occlusion (Fig. 8d-f). Viewed from 90◦
the view that we show is precisely a 2D bar chart (Fig. 8g). It
is worth noting that a couple of the stages are reminiscent of
other thematic map visualisation techniques. In particular,
Fig. 8d resembles a non-contiguous cartogram and Fig. 8e
resembles bars on a map (recently popularised by the 3D
map feature in MS Office 365).
To differentiate this morphing hybrid from the other
conditions of our next study, we refer to it as a “Tilt Map”.
4.1 Tilt Map Design Considerations
Colouring: A difficult design decision was whether to
colour the Prism view. While our previous study suggested
possible benefits to using a monochrome prism map, we de-
cided that the visual continuity provided by preserving the
colouring of the choropleth view outweighed the potential
downside.
User Control: In all conditions, users were able to freely
move and rotate maps by moving the controller inside the
map boundary and “grasping” the controller’s trigger. From
the time it is grasped until the time it is released the map’s
position and angle relative to the ground is “latched” to the
controller. In the Tilt Map we chose to couple the angle of
the map relative to the ground, in the vertical plane of the
user’s view vector, to the type of visualisation shown.
Choice of transition angle intervals: Transitions from
Choropleth view, to Prism view, to Bar Chart view, occur at
key angles within a 90◦arc of angle relative to the ground.
The precise angles triggering the transitions, indicated in
the figure to the right, were informed by our observations
of angles at which users tended to view the different map
techniques in Study 1, see Fig. 6. Fig. 1 (right) shows
labels (a–g) that correspond to the viewing angles in Fig. 8.
Transitions are not sudden but continuously controlled by
tilt angle so that the user sees a smooth animated transitions
as they tilt the map.
Projection order of prisms to bars: A key design decision
for the transition from Prism view to Bar Chart was to order
the bars corresponding to the left-to-right order of the area
centres in the view plane of the Prism view. In other words,
a straightforward projection of the regions into the left-to-
right view axis. Thus, the order of the bars depends upon
the horizontal view angle. If the user was viewing the Prism
view with the southern edge of the map closest to them
at the time of transition to the bar chart, the bars would
be ordered west to east; if the eastern edge of the map
was closest to them, the bars would be ordered south to
north; and so on. This design choice minimised the relative
movement of the bars when transitioning, which we felt
was a relevant feature and was likely to minimize occlusion.
Other orderings, e.g. from smallest value to largest, are
also possible, but would lead to larger movements during
transition.
Legend/axis positions and orientation: Following immer-
sive visualisation best practice, text labels in all views are
billboarded. That is, they reorient dynamically as the view
angle changes such that they always face the user directly.
For the Choropleth view it is essential to show a legend
indicating the mapping of data value to area colour. For the
Prism and Bar Chart views it is also necessary to show axes
indicating mapping from data value to prism/bar height.
For Tilt Map we combine each colour legend and axis into
a single annotation. It is important that this annotation be
clearly visible regardless of the view angle, however, the
axes in the Prism view cannot be bill-boarded, since they
must always lie in the height axis. In the Choropleth view it
is natural that the legend lie in the plane of the map. Thus,
another transition occurs for these legend/axis annotations:
the axis rotates from the map plane to the height axis as
the Tilt Map transitions from the Choropleth to the Prism
view. In Choropleth and Prism views there are four of these
legend/axis annotations, positioned at the four sides of the
map. In the final transition to the Bar Chart view, the top and
bottom axes are removed leaving only axes on the left and
right of the bar chart.
5 ST UDY 2: TILT MAP EVALUATIO N
We conducted a second user study in order to evaluate Tilt
Map. We wished to compare it to more traditional ways
to combine a choropleth, prism and bar chart view. We
compared Tilt Map to:
Side-By-Side: We placed three views following the transi-
tion order of Tilt Map, i.e. from left to right: Choropleth view
(left), Prism view (centre) and Bar Chart view (right). All
three views were at the same distance to the viewer. The
Prism view was immediately in front of the viewer, while
the Choropleth and Bar Chart views were positioned in an
egocentric layout, 80° anticlockwise and clockwise respec-
tively. Participants were free to rearrange all three views
independently, but the layout was reset at the beginning of
each trial.

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 8
Toggle: Participants could switch the view by tapping the
left or right half of the touchpad on the controller to cycle
forward or backward. The initial view was Choropleth view,
and consistent with Tilt Map, the forward-toggle order was
Choropleth –Prism –Bar Chart.
5.1 Experiment
Following the same structure of Sec. 3.2, in this subsection,
we first introduce the tasks of the user study and the way
we create task data. We then report details of the user
study including: experimental set-up,design and procedure,
participants and measures.
Tasks: We evaluated Tilt Map,Side-By-Side and Toggle using
the Region Task from the first user study. We kept the
coefficient of variation (CV) within 40%−60% for all tasks.
The answers to the generated tasks evenly fell into three
different ranges: 20−40, 40−60 and 60−80. We did not use
the area-comparison task as the Bar Chart view made this
task too easy. We used 5 contiguous states as the target
region for the US and 10 contiguous areas for the EU.
Data: We again tested using two datasets. We used the same
US data as in the first user study. We chose not to use the
UK data because its large size made the bar chart view too
wide. Instead we generated data based on Europe (EU) data
[44]. We used the first level of Nomenclature of Territorial
Units for Statistics (NUTS). After removing a few areas
geographically distant from Europe, the dataset contained
116 geographic areas. Again, the original population density
data was highly skewed with only very few areas with high
population densities. To ensure the map visualisations used
the entire range of colour or height variation (see Sec. 3.2 —
Data), here we applied a fourth root transformation. We
again generated different data for each question and kept
Moran’s Ithe same as the original data.
Experimental Set-up: PC and headset were as per the first
study. All maps were created on top of a transparent quad-
rangle of 1×1m. For Tilt Map and Toggle, the initial views
were Choropleth view with 0° tilting. For Side-By-Side, the
Choropleth and Bar Chart views were placed perpendicularly
to the horizon plane and the Prism view was tilted to 75°.
The initial views of Tilt Map and Toggle were placed the same
as in the first study, i.e. 0.6 m in front of the participants’
eye position and 0.1 m below it. For Side-By-Side, to avoid
overlap among views, the distance was enlarged to 0.9 m.
Design and Procedure: The experiment was within-
subjects: 18 participants ×3 visualisations ×1 task ×2
datasets ×3 answer ranges ×2 repetitions = 648 responses
(36 responses per participant) and lasted one hour and
ten minutes on average. Latin square design was used to
balance the order of visualisations.
The study procedure was similar to the first study but
with two modifications:
(1) A pre-study training was added. Based on the participants
recruited in the first study, we expected most participants to
have limited experience with VR. Therefore, we recruited
a participant without VR experience to do a pilot study.
The pilot study revealed some initial difficulty with VR
interactions. Thus, we designed a pre-study training which
instructed participants in basic VR interactions, including
repositioning, rotating and tilting a 2D flat world map.
Fig. 9: Study 2 — Accuracy and response time with 95%
confidence intervals (AbsDiff = absolute difference between
participants’ answers and the correct answers.)
Fig. 10: Study 2 — Graphic depiction of statistic compar-
isons with p<.05.
Participants were given unlimited time to become familiar
with these interactions. The pre-study training usually took
5–10 minutes.
(2) Questions were added to the post-hoc questionnaire. We asked
participants to rate the usefulness of each view within the
combined visualisations with a five-point Likert scale, i.e.
Choropleth view, Prism view and Bar Chart view. We also
asked participants to rate the usefulness of the continuous
transition in Tilt Map with a five-point Likert scale and
explain their reasons for the rating.
Participants: We recruited 18 new participants (5 female,
13 male) from our university. All had normal or corrected-
to-normal vision and included students and researchers. 12
participants were aged between 20−30, 5 between 30−40,
and 1 over 40. VR experience varied: 10 participants had
less than 5 h of prior VR experience, 3 had 6−20 h, and 5
more than 20 h.
Measures: In addition to real-time recording of partici-
pant’s head, controller, map position and map rotation, we
recorded the view the participant was looking at.
5.2 Results
As in the first study, we used linear mixed modeling to check
for significance and applied Tukey’s HSD post-hoc tests to
conduct pairwise comparisons for the square root of Abs-
Diff and time. Two independent factors (visualisation and
country) and their interaction (visualisation×country) were
modeled. Accuracy (AbsDiff) and time with 95% confidence
of different visualisations were presented in Fig. 9.
The type of visualisation had a statistically significant
effect on AbsDiff (χ2(2) = 6.5, p =.0393). Post-hoc
tests revealed that Tilt Map was statistically more accurate
than Side-By-Side (see Fig. 10). The type of country also
had a marginally significant effect on AbsDiff (χ2(1) =
3.6, p =.0569). Post-hoc tests revealed that participants
were marginally more accurate in US tasks than EU tasks.

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 9
Fig. 11: Time percentage in different movements.
Fig. 12: Time percentage in each main view with 95% confi-
dence interval.
The type of visualisation had a statistically significant
effect on time (χ2(2) = 23.1, p < .0001). Post-hoc tests
revealed that Toggle was statistically slower than both Tilt
Map and Side-By-Side (see Fig. 10). The type of country
also had a statistically significant effect on time (χ2(2) =
8.3, p =.0040). Post-hoc tests revealed that participants
were statistically faster in US tasks than EU tasks (see
Fig. 10).
Since one task and one dataset were identical in the
two studies, we were able to compare the effectiveness
of the hybrid visualisations in Study 2 with the single
visualisations used in Study 1 for the US region task.
Following mixed designs modeling [42, Chapter 14], we
used linear mixed model to check statistical significance
and applied Tukey’s HSD post-hoc tests to conduct pairwise
comparisons. We found a statistically significant effect of
visualisations on AbsDiff (χ2(5) = 18.3, p =.0026). In
addition to the findings in Fig 10, we found that, Tilt Map
was statistically more accurate than Choropleth and Coloured
Prism, with p=.0024 and p=.0082 respectively. We
also found a statistically significant effect of visualisations
on time (χ2(5) = 53.2, p < .0001). Choropleth was faster
than all other visualisations and Toggle was slower than all
other visualisations, all p<.05. We found that Tilt Map had
very similar accuracy and time to Monochrome Prism. This
suggests that the choice to colour the prism view in Tilt Map
did not degrade performance.
Again, we plotted the AbsDiff and time against task
order: again we found no obvious learning effect.
Interactions: Linear mixed model was used to analyse
the differences in the way participants interacted with the
three visualisations. Unsurprisingly, participants spent a
statistically greater percentage of time moving their heads
in Side-By-Side than Tilt Map and Toggle. Participants also
spent statistically more time manipulating the views in Tilt
Map than Side-By-Side and Toggle. Within each visualisation,
participants spent statistically more time moving their head
than moving the visualisation (all p<.001).
We investigated the time spent in each main view (i.e.
Choropleth view, Prism view and Bar Chart view). In Fig. 12,
we can identify different patterns for different visualisa-
tions. For Tilt Map, participants spent significantly more
time in Bar Chart view than Choropleth view or Prism view.
For Side-By-Side, participants spent significantly more time
in Bar Chart view and Prism view than Choropleth view, while
time spent in Prism view was more than Bar Chart view. For
Toggle, participants spent time almost evenly in each of the
Fig. 13: User preference: (a, b) ranking for each of the
three visualisations; (c) confidence rating in five-point-
Likert scale. Arrows indicate statistic comparisons with
p<.05.
three views.
In the Tilt Map, participants could continuously morph
between the three main views. Participants spent on average
13.5% of time in the intermediate state between Choropleth
and Prism (see Fig. 8(b)), they also spent similar time (avg.
12.7%) in the intermediate state between Prism and Bar Chart
(see Fig. 8(d,e,f) and discussed further in Strategies).
To tilt the Tilt Map most participants grasped it hold-
ing the controller vertically as per Figures 1 and 8.
However, three of the 18 participants
grasped it holding the controller hor-
izontally (like a motorcycle throttle),
as per the figure to the right. This
was unexpected and interesting as it
allows for a greater range of rotation.
There were not enough participants us-
ing this strategy to determine with any
statistical significance if it improved
performance, but it made us reflect that
a handle affordance could be added to
the Tilt Map to encourage such a grasp.
User Preference: For visual design, in Fig. 13(a), the strongest
preference was for Tilt Map with 83.3% participants ranking
it as the best. The Friedman test revealed a significant effect
of visualisation on preference (χ2(2) = 25, p < .0001). The
post-hoc tests confirmed Tilt Map was statistically preferred
to Side-By-Side (p=.0333) and Toggle (p < .0001). Side-By-
Side was statistically preferred to Toggle (p=.0333).
For readability, in Fig. 13(b), Toggle was least preferred
with only 5.6% voting it as the best. The Friedman test
revealed a significant effect of visualisation on preference
(χ2(2) = 10.11, p < .0001). The post-hoc tests confirmed
Tilt Map and Side-By-Side were statistically more preferred
than Toggle with p=.0209 and p=.0128, while there is no
statistical difference between Tilt Map and Side-By-Side.
For confidence, in Fig. 13(c), participants felt confident
with Tilt Map and Side-By-Side with both 72.2% rating them 4
or 5. The Friedman test revealed a statistically significant ef-
fect of visualisation on confidence (χ2(2) = 13.3, p =.0013).
The post-hoc tests found participants felt statistically more
confident with Tilt Map and with Side-By-Side than with
Toggle with p=.0076 and p=.0029 respectively.
Additionally, we asked participants to rate the impor-
tance of each individual view (i.e. Choropleth,Prism and Bar
Chart views) with a five-point Likert scale. Participants rated
the three views as equally important: 61.1% rated Choropleth
view 4 or 5 (i.e. obviously useful), 72.2% rated Prism view
obviously useful and 61.1% rated Bar Chart view obviously
useful. The small difference was not statistically significant.

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 10
Fig. 14: Demonstration of representative strategies for dif-
ferent visualisations. (TM1, TM2, TM3, TM4) for Tilt Map,
(S1, S2) for Side-By-Side, (T1, T2, T3, T4) for Toggle.
We also asked participants to rate the usefulness of the
morphing (seeing intermediate states of view changing) in
the Tilt Map with a five-point-Likert scale. 77.8% rated the
usefulness of morphing obviously useful. We also asked
participants to explain their rating. Most participants com-
mented that the tilting interaction was “intuitive” in VR.
Some detailed comments revealed:
•The Prism Shrinking transition reduced occlusion and the
final stage, Bars on Map transition, was found to be very
useful (5 participants mentioned this).
•The transition from the Prism view to the Bar Chart view
helped them locate the position of areas in the Bar Chart
view (8 participants mentioned this).
Strategies: To better understand how participants interacted
with the visualisations, we performed an exploratory analy-
sis of the sequence of views used by participants in order to
provide an initial identification of common strategies used
for each visualisation (see Fig. 14).
For Tilt Map,
TM1 spent most time with Choropleth and Bar Chart views.
TM2 switched between views frequently.
TM3 spent a significant time on transitions.
TM4 spent similar time in each view.
For Side-By-Side,
S1 spent most time with the Prism and Bar Chart views.
S2 spent similar time in each view.
For Toggle,
T1 spent most time with the Choropleth and Bar Chart views.
T2 spent most time with the Prism view.
T3 spent most time with the Prism and Bar Chart views.
T4 spent similar time in each view.
From our observations and the questionnaire we further
identified a specific strategy in S1: 3 participants reposi-
tioned Prism and Bar Chart views to allow both of them
in their field of view (vertically aligned). For the other
participants who used S1, instead of changing the layout,
they rotated their heads to switch the view.
We found that the strategies used by participants were
relatively consistent in Side-By-Side and Toggle, while strate-
gies were much more diverse in Tilt Map.
Feedback: The final section of the study allowed partici-
pants to give feedback on the pros and cons of each design.
Qualitative analysis of these comments revealed (overall):
Tilt Map was found to be “intuitive and cool”. Most
participants “enjoyed using it as it ‘gratified’ in a way” and “it
is easy to control which view I want”. Some found the transition
and the intermediate states helpful. However, some also
mentioned they felt it “a bit tiring” and the transition from
the Bars on Map to Bar Chart to be “a bit too fast”.
Side-By-Side was found to be informative as “everything
is around you”. Some participants also liked to be able to
switch the view by simply rotating their heads, while some
also complained that this body movement to be “a bit tired”.
Toggle was found to be “efficient” to switch the view, but
“boring”. Many participants found difficulties in switching
to the desired view: “I eventually just slammed the button in
any direction until I found the one I was looking for”.
Key Findings: The main finding of the second study was
that participants were generally more accurate with Tilt Map
without paying a significant cost in time. Specifically:
•Tilt Map was more accurate than Side-By-Side and faster
than Toggle.
•Tilt Map was more accurate than the single Choropleth and
single Coloured Prism for (the region task) with the small
dataset (US).
•Side-By-Side was faster than Toggle.
We also found that Tilt Map was strongly preferred for
visual design while Toggle was least preferred for both visual
design and readability. Generally, participants felt more
confident with both Tilt Map and Side-By-Side than Toggle.
Analysis of user interaction revealed that participants
had different strategies of using different views in the three
visualisations. We were surprised to find that some partici-
pants used the transition views between the main views in
Tilt Map to answer the questions.
6 DISCUSSION AND CONCLUSION
We have investigated the presentation of area-linked data
in immersive visualisation environments. While data vi-
sualisation is currently rare in VR or AR environments,
we believe it will become much more common with the
commodification of HMD VR and AR displays [21], [45],
[46]. Our long term goal is not only VR but mixed reality
(MR), which includes VR and AR. We foresee MR eventually
supplanting 2D displays in many situations, just as cur-
rent mobile devices have displaced traditional computing
platforms. In particular, MR supports remote collaboration
and situated data analysis, which opens up exciting new
possibilities for visualisation away from the desktop. Our
research therefore provides a timely insight into how best to
visualise area-linked data in MR.
In our first study, we found that encoding quantita-
tive data with prism height was more accurate but slower
than with colour in more traditional 2D choropleth maps.
However, it was clear that occlusion and perspective distor-
tion still hampered understanding of the prism maps. Our
results suggest that participants had to spend extra time
with the prism map to find occlusion-free viewing points.
While occlusion is inevitable in prism maps, the embodied
interaction of changing viewing point and direction in the
immersive environment appears to alleviate this issue.
In our second study, we investigated hybrid representa-
tions that combined 2D choropleth, 3D prism map and 2D
bar chart views. We compared the Tilt Map—a new kind
of interactive visualisation whose choice of idiom depends
upon the user’s viewing angle—with side-by-side place-
ment and interactive toggling. We found benefits in time,
accuracy and user preference for the Tilt Map over these

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 11
alternatives. We believe, compared to our tested alterna-
tives, the continuous embodied transition from the Tilt Map
reduces the cognitive load of context-switching between
views. In side-by-side placement, the user has to visually
link separate display spaces; and in interactive toggling,
the immediate transition makes it difficult to track objects.
However, further research is needed to test how effective
the morphing is in the Tilt Map for object tracking. Due
to the limitations of visual working memory, the user can
only track a limited number of areas in the Tilt Map. This
potentially explains why we found that the advantage of
Tilt Map was larger in the US data (five areas to track) than
in the EU data (ten areas to track).
Although we have not seen them used before, we believe
that orientation-dependent visualisations like the Tilt Map
are widely applicable in immersive environments. Tilting
provides a natural, embodied interaction for view switching
that allows the choice of view to simply adapt to and take
advantage of the viewing angle. Another example would be
for an interactive map showing area-linked data. When held
vertically in front of the user this could show a traditional
2D proportional symbol map. When tilted it could transition
to a map with height-varying 3D bars, then transitioning
like the Tilt Map to a bar chart. This is something we plan to
explore further and also test applicability to non-geographic
and non-spatial data visualisation.
Limitations of our two studies are that only population
data was tested and only a limited number of tasks were
trialled. The available interactions were also deliberately
limited. Future work will be needed to add standard view
interactions such as filtering, selection, and zooming. Future
work should also investigate the impact of the participant’s
field of view. For example, in the area-comparison task of
the first study, the regions in the far and close conditions
in the UK map are roughly within the foveal field of
view while in the far condition of the US map they may
not be. Additionally, we used a sequential colour scheme
from ColorBrewer [47]. Colour schemes are well studied for
traditional cartography, but there are no clear guidelines
for using them in an immersive environment, where the
use of shading may interfere with color perception. In AR
with optical see-through rendering, other issues arise, for
example, colour contrast between virtual objects and the real
world backdrops is often too weak.
One limitation of the Tilt Map and other hybrid views
used in the second study was that the bar chart view does
not scale well to large numbers of areas. With the resolution
of current HMD displays, we were able to visualise charts
with approximately 100 bars. For larger numbers of bars,
one possible solution would be to split the bar chart into
multiple rows if it contains more bars than can be shown
within the field of view.
We would also like to explore if other ways of deter-
mining the order of the bar chart may bring extra benefits.
For example, in Fig. 8 there is a vertical sequence of areas
where Texas is followed by Oklahoma, Kansas, Nebraska,
South Dakota, and North Dakota. The bar chart currently
displays the corresponding bars in an apparently random
order, but it could possibly retain the vertical order of
the prism map. Alternatively, it would also make sense to
“gridify” the bar chart to reflect the geography, similar to
spatial treemaps [48] or grid maps [49]. Thus, if the map
was oriented with north at the top, the assigned row would
reflect the north-south position of an area, and the assigned
column would reflect its east-west position. These alterna-
tive arrangements might simplify object tracking when the
Tilt Map transitions to the bar chart.
ACKNOWLEDGMENTS
This work was supported by the Australian Research Coun-
cil through grants DP140100077 and DP180100755. Yalong
Yang was partially supported by a Harvard Physical Sci-
ences and Engineering Accelerator Award. We also wish to
thank all our participants for their time and our reviewers
for their comments and feedback.
REFERENCES
[1] B. D. Dent, J. Torguson, and T. W. Hodler, Cartography: Thematic
Map Design, 6th ed. New York: McGraw-Hill Education, Aug.
2008. ISBN 978-0-07-294382-5
[2] T. A. Slocum, R. M. McMaster, F. C. Kessler, H. H. Howard, and
R. B. Mc Master, Thematic cartography and geographic visualization,
3rd ed. Prentice Hall, 2009.
[3] M.-J. Kraak and F. Ormeling, Cartography : visualization of spatial
data, 3rd ed. Guilford Press, 2011.
[4] K. Field, Cartography. Esri Press, 2018.
[5] Y. Yang, B. Jenny, T. Dwyer, K. Marriott, H. Chen, and M. Cordeil,
“Maps and globes in virtual reality,” in Computer Graphics Forum,
vol. 37, no. 3, 2018, pp. 427–438. doi: 10.1111/cgf.13431
[6] Y. Yang, T. Dwyer, B. Jenny, K. Marriott, M. Cordeil, and H. Chen,
“Origin-destination flow maps in immersive environments,” IEEE
transactions on visualization and computer graphics, vol. 25, no. 1, pp.
693–703, 2019. doi: 10.1109/TVCG.2018.2865192
[7] J. A. Wagner Filho, W. Stuerzlinger, and L. Nedel, “Evaluat-
ing an Immersive Space-Time Cube Geovisualization for Intu-
itive Trajectory Data Exploration,” IEEE Transactions on Visual-
ization and Computer Graphics, vol. 26, no. 1, pp. 514–524, 2019.
doi: 10.1109/TVCG.2019.2934415
[8] M. R. Jakobsen, Y. S. Haile, S. Knudsen, and K. Hornbæk,
“Information visualization and proxemics: Design opportuni-
ties and empirical findings,” IEEE Transactions on Visualization
and Computer Graphics, vol. 19, no. 12, pp. 2386–2395, 2013.
doi: 10.1109/TVCG.2013.166
[9] J. S. Torguson, “Choropleth Map,” in International Encyclope-
dia of Geography: People, the Earth, Environment and Technology,
D. Richardson, N. Castree, M. F. Goodchild, A. Kobayashi, W. Liu,
and R. A. Marston, Eds. Oxford, UK: John Wiley & Sons, Ltd,
Mar. 2017, pp. 1–9. doi: 10.1002/9781118786352.wbieg0951
[10] W. Tobler, “A computer program to draw perspective views of
geographic data, Cartographic Lab Report, No. 1,” University of
Michigan, Department of Geography, Tech. Rep., 1973.
[11] R. Franklin and H. R. Lewis, “3-d graphic display of discrete spa-
tial data by prism maps,” in ACM SIGGRAPH Computer Graphics,
vol. 12, no. 3. ACM, 1978, pp. 70–75.
[12] H. Moellering, “The real-time animation of three-dimensional
maps,” The American Cartographer, vol. 7, no. 1, pp. 67–75, 1980.
[13] J. Nimmer, T. Nash, and D. Phillips, “Proportional prism maps: a
statistical mapping technique.” in the 5th International Symposium
on Computer-Assisted Cartography, 1982, pp. 555–564.
[14] D. J. Cuff and K. R. Bieri, “Ratios and absolute amounts conveyed
by a stepped statistical surface,” The American Cartographer, vol. 6,
no. 2, pp. 157–168, 1979. doi: 10.1559/152304079784023087
[15] W. S. Cleveland and R. McGill, “Graphical perception: Theory,
experimentation, and application to the development of graphical
methods,” Journal of the American Statistical Association, vol. 79, no.
387, pp. 531–554, 1984. doi: 10.2307/2288400
[16] T. Niedomysl, E. Elld´
er, A. Larsson, M. Thelin, and B. Jansund,
“Learning benefits of using 2d versus 3d maps: Evidence from a
randomized controlled experiment,” Journal of Geography, vol. 112,
no. 3, pp. 87–96, 2013. doi: 10.1080/00221341.2012.709876
[17] S. Popelka, “Eye-tracking evaluation of 3D thematic maps,” in the
3rd Workshop on Eye Tracking and Visualization. ACM Press, 2018,
pp. 1–5. doi: 10.1145/3205929.3205932
[18] J. Stewart and P. J. Kennelly, “Illuminated choropleth maps,”
Annals of the Association of American Geographers, vol. 100, no. 3,
pp. 513–534, 2010. doi: 10.1080/00045608.2010.485449

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 12
[19] S. Bleisch, J. Dykes, and S. Nebiker, “Evaluating the effectiveness
of representing numeric information through abstract graphics in
3d desktop virtual environments,” The Cartographic Journal, vol. 45,
no. 3, pp. 216–226, 2008. doi: 10.1179/000870408X311404
[20] S. Seipel and L. Carvalho, “Solving combined geospatial tasks
using 2d and 3d bar charts,” in 2012 16th International Conference on
Information Visualisation, 2012, pp. 157–163. doi: 10.1109/IV.2012.36
[21] K. Marriott, J. Chen, M. Hlawatsch, T. Itoh, M. A. Nacenta,
G. Reina, and W. Stuerzlinger, Immersive Analytics: Time to Recon-
sider the Value of 3D for Information Visualisation. Cham: Springer
International Publishing, 2018, pp. 25–55.
[22] Y. Yang, “Visualising Geographically-Embedded Origin-
Destination Flows: in 2D and immersive environments,” Ph.D.
dissertation, Monash, 2018.
[23] Q. Quach and B. Jenny, “Immersive visualization with bar
graphics,” Cartography and Geographic Information Science, 2020.
doi: 10.1080/15230406.2020.1771771
[24] M. Cordeil, A. Cunningham, T. Dwyer, B. H. Thomas, and K. Mar-
riott, “Imaxes: Immersive axes as embodied affordances for inter-
active multivariate data visualisation,” in Proceedings of the 30th
Annual ACM Symposium on User Interface Software and Technology.
ACM, 2017, pp. 71–83. doi: 10.1145/3126594.3126613
[25] K. A. Satriadi, B. Ens, M. Cordeil, B. Jenny, T. Czauderna, and
W. Willett, “Augmented Reality Map Navigation with Freehand
Gestures,” 2019 IEEE Conference on Virtual Reality and 3D User
Interfaces (VR), pp. 593–603, 2019. doi: 10.1109/vr.2019.8798340
[26] J. Heer and G. Robertson, “Animated transitions in statistical data
graphics,” IEEE transactions on visualization and computer graphics,
vol. 13, no. 6, pp. 1240–1247, 2007. doi: 10.1109/TVCG.2007.70539
[27] L. Herman, V. Juˇ
r´
ık, Z. Stachoˇ
n, D. Vrb´
ık, J. Russn´
ak, and
T. ˇ
Rezn´
ık, “Evaluation of user performance in interactive and
static 3d maps,” ISPRS International Journal of Geo-Information,
vol. 7, no. 11, p. 415, 2018. doi: 10.3390/ijgi7110415
[28] M. Harrower and C. A. Brewer, “ColorBrewer.org: An online tool
for selecting colour schemes for maps,” The Cartographic Journal,
vol. 40, no. 1, pp. 27–37, 2003. doi: 10.1179/000870403235002042
[29] R. Beecham, J. Dykes, W. Meulemans, A. Slingsby, C. Turkay,
and J. Wood, “Map LineUps: Effects of spatial structure
on graphical inference,” IEEE Transactions on Visualization
and Computer Graphics, vol. 23, no. 1, pp. 391–400, 2017.
doi: 10.1109/TVCG.2016.2598862
[30] Y. Yang, T. Dwyer, S. Goodwin, and K. Marriott, “Many-to-
many geographically-embedded flow visualisation: An evalua-
tion,” IEEE Transactions on Visualization and Computer Graphics,
vol. 23, no. 1, pp. 411–420, 2017. doi: 10.1109/TVCG.2016.2598885
[31] N. Andrienko and G. Andrienko, “Tasks,” in Exploratory Anal-
ysis of Spatial and Temporal Data. Springer, 2006, pp. 47–161.
doi: 10.1007/3-540-31190-4˙3
[32] R. E. Roth, “An empirically-derived taxonomy of interaction
primitives for interactive cartography and geovisualization,” IEEE
Transactions Visualization and Computer Graphics, vol. 19, no. 12, pp.
2356–2365, 2013. doi: 10.1109/TVCG.2013.130
[33] Y. Yang and S. Goodwin, “What-why analysis of expert in-
terviews: Analysing geographically-embedded flow data,” in
IEEE Pacific Visualization Symposium. IEEE, 2019, pp. 122–126.
doi: 10.1109/PacificVis.2019.00022
[34] Y. Jansen and K. Hornbak, “A psychophysical investigation of
size as a physical variable,” IEEE Transactions on Visualization and
Computer Graphics, vol. 22, no. 1, pp. 479–488, 2015.
[35] I. Spence, “The apparent and effective dimensionality of represen-
tations of objects,” Human Factors, vol. 46, no. 4, pp. 738–747, 2016.
doi: 10.1518/hfes.46.4.738.56809
[36] M. Harrower, “Unclassed animated choropleth maps,” The
Cartographic Journal, vol. 44, no. 4, pp. 313–320, Dec. 2007.
doi: 10.1179/000870407X241863
[37] C. E. Brown, “Coefficient of Variation,” in Applied Multivariate
Statistics in Geohydrology and Related Sciences. Berlin, Heidelberg:
Springer, 1998, pp. 155–157. doi: 10.1007/978-3-642-80328-4˙13
[38] The U.S. Census Bureau, “U.S. census data,” https://www.census.
gov/, 2019, accessed: 2019-09-05.
[39] T. U. O. for National Statistics, “Open geography portal,” https:
//geoportal.statistics.gov.uk, 2019, accessed: 2019-09-05.
[40] D. Bates, M. M ¨
achler, B. Bolker, and S. Walker, “Fitting linear
mixed-effects models using lme4,” Journal of Statistical Software,
vol. 67, no. 1, 2015. doi: 10.18637/jss.v067.i01
[41] R. V. Lenth, “Least-squares means: The R Package lsmeans,”
Journal of Statistical Software, vol. 69, no. 1, 2016.
[42] A. Field, J. Miles, and Z. Field, Discovering statistics using R. Sage
publications, 2012.
[43] C. Chang, T. Dwyer, and K. Marriott, “An evaluation of percep-
tually complementary views for multivariate data,” in 2018 IEEE
Pacific Visualization Symposium (PacificVis). IEEE, 2018, pp. 195–
204. doi: 10.1109/PacificVis.2018.00033
[44] The Publications Office of the European Union, “EU open data
portal,” https://data.europa.eu/euodp/en/home, 2019, accessed:
2019-09-05.
[45] T. Chandler, M. Cordeil, T. Czauderna, T. Dwyer, J. Glowacki,
C. Goncu, M. Klapperstueck, K. Klein, K. Marriott, F. Schreiber
et al., “Immersive analytics,” in 2015 Big Data Visual Analytics
(BDVA). IEEE, 2015, pp. 1–8. doi: 10.1109/BDVA.2015.7314296
[46] T. Dwyer, K. Marriott, T. Isenberg, K. Klein, N. Riche, F. Schreiber,
W. Stuerzlinger, and B. H. Thomas, “Immersive analytics: An
introduction,” in Immersive Analytics. Springer, 2018, pp. 1–23.
doi: 10.1007/978-3-030-01388-2˙1
[47] C. A. Brewer, A. M. MacEachren, L. W. Pickle, and D. Herrmann,
“Mapping Mortality: Evaluating Color Schemes for Choropleth
Maps,” Annals of the Association of American Geographers, vol. 87,
no. 3, pp. 411–438, Sep. 1997. doi: 10.1111/1467-8306.00061
[48] J. Wood and J. Dykes, “Spatially ordered treemaps,” IEEE Trans-
actions on Visualization and Computer Graphics, vol. 14, no. 6, pp.
1348–1355, Nov. 2008. doi: 10.1109/TVCG.2008.165
[49] G. McNeill and S. A. Hale, “Generating Tile Maps,” Computer
Graphics Forum, vol. 36, no. 3, pp. 435–445, Jul. 2017.
Yalong Yang is a Postdoctoral Fellow at the
John A. Paulson School of Engineering and Ap-
plied Sciences, Harvard University. He obtained
his PhD from Monash University, Australia in Jan
2019 and continued as a Postdoc at Monash
before joining Harvard in Oct 2019. His research
designs and investigates interactive information
visualisations on both conventional 2D screens
and in 3D immersive environments. He received
an Honorable Mention Award at InfoVis 2016.
Tim Dwyer is a Professor at Monash University,
Australia, where he leads the Immersive Analyt-
ics and Data Visualisation research group. He
received his PhD from the University of Sydney
in 2005, was a post-doctoral Research Fellow at
Monash, then from 2008 he moved to Microsoft,
Redmond, USA as a Visiting Researcher and
then Senior Software Development Engineer. He
returned to Monash in 2012 as a Larkins Fellow.
Kim Marriott is a Professor in Computer Sci-
ence at Monash University, Australia. His re-
search is in data visualisation, human-in-the-
loop analytics, assistive technologies and im-
mersive analytics. After obtaining his PhD from
the University of Melbourne in 1989, he worked
at the IBM TJ Watson Research Center until
joining Monash in 1993.
Bernhard Jenny is an Associate Professor at
Monash University, Australia. He obtained a PhD
in cartography from ETH Zurich, and worked
in cartography and geovisualisation at Oregon
State University and RMIT Melbourne. His re-
search focuses on immersive maps for visualis-
ing and interacting with geographic data in virtual
reality and augmented reality.
Sarah Goodwin is a Lecturer at Monash Uni-
versity, Australia. She received her PhD in Geo-
graphical Information Science from City, Univer-
sity of London in 2015. Her research explores
novel geovisualisation techniques and method-
ologies for visualisation design studies. Her pro-
fessional and academic background is in geog-
raphy, spatial analysis and geovisualisation.