Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
Emulation of epidemics via Bluetooth-based virtual safe virus
spread: experimental setup, software, and data
Azam Asanjarani
1Y*
, Aminath Shausan
2Y
, Keng Chew
3‡
, Thomas Graham
2‡
, Shane G.
Henderson4‡, Hermanus M. Jansen5‡, Kirsty R. Short3‡, Peter G. Taylor6‡, Aapeli
Vuorinen7‡, Yuvraj Yadav8‡, Ilze Ziedins1‡, Yoni Nazarathy2‡
1Faculty of Science, The University of Auckland, Auckland, State, New Zealand
2School of Mathematics and Physics, The University of Queensland, Brisbane,
Queensland, Australia
3School of Chemistry and Molecular Biosciences, The University of Queensland,
Brisbane, Queensland, Australia
4School of Operations Research and Information Engineering, Cornell University,
Ithaca, NY, USA
5Department of Applied Mathematics, Delft University of Technology, Mekelweg 4,
2628CD Delft, The Netherlands
6Mathematics and Statistics, The University of Melbourne, Melbourne , Victoria,
Australia
7Department of Industrial Engineering and Operations Research, Columbia University,
New York, NY, USA
8Mechanical Engineering Department, Indian Institute of Technology Delhi, New Delhi,
Delhi, India
YThese authors contributed equally to this work.
‡These authors also contributed equally to this work.
* azam.asanjarani@auckland.ac.nz
Abstract
We describe an experimental setup and a currently running experiment for evaluating
how physical interactions over time and between individuals affect the spread of
epidemics. Our experiment involves the voluntary use of the Safe Blues Android app by
participants at The University of Auckland (UoA) City Campus in New Zealand. The
app spreads multiple virtual safe virus strands via Bluetooth depending on the social
and physical proximity of the subjects. The evolution of the virtual epidemics is
recorded as they spread through the population. The data is presented as a real-time
(and historical) dashboard. A simulation model is applied to calibrate strand
parameters. Participants’ locations are not recorded, but participants are rewarded
based on the duration of participation within a geofenced area, and aggregate
participation numbers serve as part of the data. Once the experiment is complete, the
data will be made available as an open-source anonymized dataset.
This paper outlines the experimental setup, software, subject-recruitment practices,
ethical considerations, and dataset description. The paper also highlights current
experimental results in view of the lockdown that started in New Zealand at 23:59 on
August 17, 2021. The experiment was initially planned in the New Zealand environment,
expected to be free of COVID and lockdowns after 2020. However, a COVID Delta
strain lockdown shuffled the cards and the experiment is currently extended into 2022.
March 29, 2022 1/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
NOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice.
Author summary
In this paper, we describe the Safe Blues Android app experimental setup and a
currently running experiment at the University of Auckland City Campus. This
experiment is designed to evaluate how physical interactions over time and between
individuals affect the spread of epidemics.
The Safe Blues app spreads multiple virtual safe virus strands via Bluetooth based
on the subjects’ unobserved social and physical proximity. The app does not record the
participants’ locations, but participants are rewarded based on the duration of
participation within a geofenced area, and aggregate participation numbers serve as part
of the data. When the experiment is finished, the data will be released as an
open-source anonymized dataset.
The experimental setup, software, subject recruitment practices, ethical 1
considerations, and dataset description are all described in this paper. In addition, we 2
present our current experimental results in view of the lockdown that started in New 3
Zealand at 23:59 on August 17, 2021. The information we provide here may be useful to
4
other teams planning similar experiments in the future. 5
Introduction 6
The COVID-19 pandemic is the most significant global event of the 21st century to date.
7
In response to the pandemic, multiple solutions have been and are still being developed
8
and deployed, including vaccines and contact tracing technologies. As part of this effort,
9
various initiatives that integrate digital health and “AI systems” (artificial intelligence 10
for pandemics) are being thought out. A key initiative includes measuring the spread of
11
pathogens as well as the level of physical human contact. The Safe Blues project is one
12
such idea, where virtual safe virus-like tokens are spread between cellular phones in an
13
attempt to mimic biological virus spread for purposes of measurement and analysis, 14
while respecting the privacy and safety of the population. 15
Much COVID-19 data is being gathered by contact-tracing apps to aid in identifying
16
infected people or their contacts. However, there can be a time lag of 1 to 2 weeks 17
between being infected and being diagnosed as positive with the result that data 18
obtained in this way is always lagging and biased. Also, asymptomatic cases who may 19
have already spread the virus to others are frequently missed by such methods. Data 20
delays and bias make it difficult for public health officials and others who want to use 21
the data to implement timely mitigation measures. Our approach, on the other hand, is
22
specifically designed to make inferences about the global characteristics of an epidemic
23
in real-time, allowing governments to implement relevant mitigation measures in a 24
timely fashion. 25
Safe Blues, introduced in [1, 2], works by spreading virtual ‘virus-like’ tokens, which
26
we call strands. The strands can be of Susceptible-Exposed-Infectious-Removed (SEIR),
27
Susceptible-Infectious-Removed (SIR), Susceptible-Exposed-Infectious (SEI), or 28
Susceptible-Infectious (SI) type. Each strand is artificially seeded into the system at 29
chosen times and can then spread between phones of users. At any given time, a phone
30
can be infected with many strands, and the phone reports its strand infections to the 31
server periodically. Individuals’ identities and social contacts are not recorded in this 32
reporting, ensuring anonymity. A key aim of the Safe Blues idea is to give policymakers
33
another tool that they can use in their effort to visualize the real-time spread of an 34
epidemic. In contrast to those systems that model population contact and implement 35
agent-based simulations, Safe Blues is an emulation of a group of epidemics based upon
36
a contact process that takes place in the population itself. 37
We devised a campus-wide experiment at The University of Auckland City Campus.
38
March 29, 2022 2/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Table 1. The timeline of the experiment at The University of Auckland.
Phase
Dates Study
period Comments
1 May 1 - Jun. 10, 2021 S1, 2021 Used for debugging
2 Jul. 18 - Sep. 16, 2021 S2, 2021 Lockdown on Aug. 17, 2021
3 Sep. 17 - Nov. 4, 2021 S2, 2021 Cancelled due to lockdown
4 Jun. 10 - Sep. 8, 2022 S2, 2022 Currently planned
5 Sep. 9 - Nov. 3, 2022 S2, 2022 Currently planned
This is the first attempt to implement such a system. An outcome of this experiment is
39
an open-source (virtual) epidemic spread dataset which can be used for further 40
modeling, training, and analysis. Our initial plan was to conclude the experiment 41
during November 2021, with the release of data afterwards. However, due to an 42
extensive lockdown in Auckland, the experiment will now run through the second half of
43
2022. In this paper, our primary focus is on the experiment’s methods and the 44
experience gained. Also, we illustrate general outcomes and results to date. The details
45
we present may be valuable to other teams planning similar experiments in the future. 46
Table 1 describes the phases of the experiment, their timelines, and the period at the 47
University of Auckland during which these phases run. 48
As an illustration of the experiment and some of the collected results, consider Fig 1
49
where we depict the timeline July 28 – September 9, 2021. Phones of participants were
50
“infected” with strands on July 29 and the figure presents the trajectories of the ensuing
51
epidemics along with the number of participants who attended the campus during that
52
period. There are multiple Safe Blues strand trajectories, the (artificial) infection on 53
July 29 included multiple repeats of the same type of strand and multiple types of 54
strands. In fact, not displayed in this figure, about 600 strands were seeded into the 55
participating population. The black trajectory depicts the daily count of campus 56
participants. The weekly attendance pattern, with lower attendance at weekends, can 57
be seen clearly. The green and red trajectories represent the hourly counts of 58
participants whose phones were in the states of exposed (infected but not infectious) and
59
infected, respectively. As is apparent from the plot, Safe Blues infections continued until
60
the week of August 17 at which point the campus was closed due to a (real) government
61
lockdown. At that point, the number of participants who attended the campus 62
immediately dropped to fewer than 5 per day. As a result, the number of new infections
63
(exposed participants) immediately decreased and within several weeks the number of 64
infected participants also decreased to zero. 65
The Safe Blues experiment was not intended to interact with actual COVID 66
numbers or lockdowns. In fact, we chose New Zealand as a destination because it was 67
essentially COVID free for the second half of 2020 and the first half of 2021 and we 68
believed that a university campus could serve as a good first testbed for Safe Blues. In
69
making this decision, we were aware that the university campus did not directly mimic
70
the population dynamics in all of New Zealand. For instance, during the Auckland 71
lockdown in Phase 2, the campus was completely shut down, while in contrast, people in
72
greater New Zealand still interacted, for example, to go shopping. We did not foresee 73
this lockdown in planning the experiment. Nevertheless, the closure of the campus due
74
to the actual physical lockdown served to illustrate the key point of Safe Blues: safe 75
virtual virus strands that are measured in real-time can give an indication of how actual
76
viruses are spreading, and with enough data and proper machine learning techniques, 77
prediction can take place as well. The Safe Blues system could thus be applied to 78
predict the spread of viral diseases within a subgroup of the population. 79
This paper outlines the experimental setup, software, subject recruitment practices,
80
ethical considerations, and dataset description of the experiment. It also presents our 81
March 29, 2022 3/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Fig 1. The effect of campus closure, due to actual lockdown in New Zealand on August 17, on
the virtual Safe Blues epidemic. Red and green trajectories show the daily counts of
participants whose phones were infected (red) and exposed (green) respectively. The black
trajectory shows the daily count of participants who attended the campus.
current experimental results. Our goal in doing so is to showcase the methodologies and
82
experience gained from the experiment. The source code for the project is freely and 83
openly available at [3]. Further, at the conclusion of the experiment, data will be made
84
available via [4]. 85
Background 86
We now present an overview of current practices and specific non-clinical experimental 87
studies that share similar concepts with Safe Blues. Most COVID-19 data are gathered
88
by public health authorities from testing, hospitalizations, and deaths. Various 89
non-government organizations, such as the World Health Organization (WHO) [5], the 90
Center for Systems Science and Engineering at Johns Hopkins University [6], and 91
nCOV2019 [7], collect this data on a global scale and provide daily trend updates. 92
However, such data are prone to lags, biases and inconsistencies, and may not reveal the
93
true characteristics of the disease in real-time. Hence, alternative surveillance methods
94
are needed. 95
Participatory syndromic-surveillance is one such approach that collects self-reported
96
data on COVID-19 symptoms, test results, and other risk factors for COVID-19 via 97
mobile applications or web-based surveys. Examples of such web surveys include 98
InfluenzaNet [8], FluTracking [9], Outbreaks Near Me [10], CoronaSurveys [11], and the
99
Global COVID Trends and Impact Survey [12]. Among the most recent mobile apps are
100
the COVID Symptom Study [13] and Beat COVID-19 Now [14]. 101
To identify potential COVID-19 hot spots, artificial intelligence is being used in 102
conjunction with information obtained from informal sources, such as Google News, 103
eyewitness reports, social media, and validated official alerts. HealthMap [15,16], 104
BlueDot [17], and Metabiota [18] are such tools. Early detection of outbreak regions 105
through wastewater examination is also used in some countries [19]. Remote patient 106
monitoring devices, such as continuous wearable sensors (e.g. smartwatches, Fitbit, 107
Oura Ring, WHOOP strap) and smart thermometers, are also being tested as potential
108
tools for tracking COVID-19 [20–25]. These tools measure some of the physiological 109
March 29, 2022 4/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
indicators of an individual’s health, such as temperature, heart rate, blood oxygen level,
110
pulse rate, sleep performance, and step counts, on a daily basis. The device can identify
111
deviations from an individual’s baseline level which may indicate the possibility of an 112
illness developing. 113
Contact tracing is a popular method for identifying infected but asymptomatic 114
individuals. Under this approach, people who have a history of exposure to a positive 115
case are identified and tested as soon as possible. Various mobile applications and 116
web-based surveys are used for contact tracing [11, 12, 26–33]. Many Bluetooth-based 117
apps, however, were abandoned after their initial release in 2020 [34,35]. Many 118
jurisdictions have QR code scanning systems in place to track and manage COVID-19. 119
However, data from such apps are typically useless when prevalence increases. For 120
example the Check In Qld app [36] in the Australian state of Queensland was 121
successfully applied to trace contacts of positive cases during 2020 to mid-December 122
2021 when the Queensland state border was closed and case numbers were single or 123
double digit at most. However, once the border opened and daily new cases grew to 124
three or four digits, the Check In Qld app was generally abandoned. 125
Moving on to experimental studies, we mention two major non-clinical citizen 126
science experiments conducted prior to the pandemic for disease surveillance purposes. 127
The first is the FluPhone experiment which took place in the United Kingdom between
128
2009 and 2011 [37]. In this experiment, participants reported their influenza like illness
129
symptoms using the FluPhone app, which also recorded the proximity of participants’ 130
devices via Bluetooth and their location via GPS. The number of people encountered by
131
each participant was then estimated and published on the study website [38]. The 132
FluPhone app, like the Safe Blues app, modeled the spread of virtual SEIR type 133
diseases, allowing participants to see real-time profiles of disease propagation in their 134
contact network [39]. However, unlike the Safe Blues app, which is designed to simulate
135
hundreds or thousands of strands, the FluPhone software was designed specifically to 136
mimic the spread of SARS, flu, and the common cold. FluPhone was a unique 137
experiment at that time, but in contrast to Safe Blues, it was designed with less of a 138
focus on capturing physical social interactions in a privacy preserving manner, and more
139
of a focus on mimicking real disease. The second study is “Contagion! The BBC Four 140
Pandemic experiment”, which also took place in the UK, but this time in 2018-2019. 141
The BBC Pandemic mobile phone app was used in the experiment to record 142
participants’ locations and self-reported contacts. A subset of this dataset was used to 143
simulate various non-pharmaceutical intervention (NPI) strategies, such as case 144
isolation, tracing, contact quarantining, and social distancing, to investigate their 145
effectiveness in limiting the spread of COVID-19 [40]. 146
Since the middle of 2020, many countries have been investigating the risk factors 147
involved in opening their society through mass-gathering experiments. Two well-known
148
examples are the RESTART-19 [41] experiment, which took place in Germany in August
149
2020, and a study which took place in Spain in December 2020 [42]. Both assessed the 150
risk of COVID transmission during an indoor live concert, using a variety of seating, 151
standing, and hygiene measures, as well as maintaining optimal air ventilation inside the
152
venue. In both studies, contact tracing devices were used to measure contacts made 153
during the event, and PCR tests were performed a few days later. The RESTART-19 154
study showed that when moderate physical distancing was applied in conjunction with
155
mask-wearing and the conditions for good ventilation were met, indoor mass-gathering
156
events could be held safely. Also, the trial in Spain demonstrated that with 157
comprehensive safety measures, such as face masks and adequate ventilation, indoor 158
mass events could be held without the need for physical distancing. 159
Some experiments have included a series of mass-gathering events with a variety of 160
indoor and outdoor settings, seated and standing audience styles, structured and 161
March 29, 2022 5/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
unstructured audience styles, and participant numbers. Two such examples are the 162
Fieldlab Events [43] which took place in the Netherlands in February and March 2021, 163
and the Events Research Program [44], which took place in the United Kingdom from 164
April to July 2021. In both experiments, comprehensive public health measures, such as
165
face mask use, hand sanitizing, social distancing, and adequate ventilation at indoor 166
events were observed. Following the events, contact tracing and PCR testing were 167
carried out. According to the Fieldlab Events, outdoor events with 50-75% of the 168
normal visitor capacity could be held provided that strict non-pharmaceutical 169
intervention measures are followed. A robust result from the Events Research Program
170
is yet to be published 171
Other relevant experiments include a health workers protest [45] in South Korea in 172
August 2020 and a martial arts competition [46] in the UAE in July 2020. During both
173
events, participants were required to wear face masks, practice hand hygiene, and 174
maintain physical distance. COVID-19 symptoms were self-reported by protesters in 175
South Korea after the rally. All PCR tests performed on a subset of rally participants 176
returned negative results. PCR tests were conducted twice weekly during the UAE 177
event, and none of the contestants had positive results, indicating that mass-gathering 178
events with restrictive measures could be held safely. 179
Materials and methods 180
We now describe the experimental setup, ethics, software, participant management, and
181
data collection aspects of the experiment as well as supporting tools such as a 182
simulation model. 183
Experimental setup and the Safe Blues system 184
As stated in the introduction, the overarching purpose of the campus experiment is to 185
test the performance of the Safe Blues system. In doing so, we are interested in 186
assessing the ability to use virtual safe virus-like tokens to predict the spread of 187
pathogens. However, an experiment involving actual biological pathogens, or relying on
188
the actual spread of disease is infeasible and hence our experiment uses measurements 189
from the digital domain. The key question is then to test if the spread of some Safe 190
Blues strands can be detected and predicted by measuring the spread of other strands. 191
In its most basic form, our purpose is to treat a single strand as a red strand which is
192
assumed not to be measurable in real time. Further, we treat all other Safe Blues 193
strands as real-time measurable virtual viruses, namely blue strands. The statistical goal
194
is then to benchmark predictions of the future evolution of the red strand based on 195
either, 196
(I) only past measurements of the red strand (a proxy for estimation in the absence 197
of blue strands), or 198
(II) the combination of past measurements of the red strand, past measurements of 199
blue strands, and current measurements of blue strands. 200
For example, Fig 2, which also appeared in [1], presents a simulation run where blue 201
strands are measured in real time, but the red strand is only measurable with a 202
two-week delay. Here the Safe Blues machine learning framework was used to predict 203
the current unobserved state of the red strand. Similarly, it can be used for near future
204
predictions. However, this figure is taken from Monte Carlo simulations of physical 205
contact processes, and not from actual experimental measurements. The Safe Blues 206
experiment attempts to improve upon this by using the actual physical mobility of 207
March 29, 2022 6/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Fig 2. Estimation via simulated epidemics. At day 115, we only have red strand information
up to day 100. Nevertheless, current blue strand measurements allow us to estimate the current
state of the epidemic during days 101-115.
individuals. The experiment aims to test whether (II) can yield much better predictions
208
than (I). 209
An additional salient feature of Safe Blues is the interaction with social distancing 210
measures. For this, we would ideally like to ask participants to group together or stay 211
apart similarly to the way that government social distancing measures work. However, 212
this is clearly not feasible with real-life participants and hence the experiment creates 213
virtual social distancing to mimic social-distancing measures. The details of how this is
214
done are described below in the subsection Virtual social distancing. 215
As a first attempt for such an experiment, we chose the University of Auckland City
216
Campus due to the fact that the campus was open to students and staff during 2021 217
(up until the unexpected lockdown of Aug 17, 2021). The experiment consists of 5 218
phases. Table 1 provides the timeline of the experiment including the time period of the
219
year, the study period and a brief description of each phase. The target population of 220
the experiment is the student body, but participation is open to any regular attendee or
221
visitor of the UoA City Campus who is at least 16 years of age and uses an Android 222
mobile phone. All participation is voluntary, and at any time, participants could opt-out
223
of the experiment and uninstall the Safe Blues app. By default, participants are invited
224
to join prize draws which we carefully designed to maximize participation (see details in
225
subsection Participant management and ethical considerations below). However, 226
participants are allowed to take part in the experiment without joining the prize draws.
227
The Safe Blues system is made out of four components: (1) the Safe Blues app, (2) 228
the campus simulation dashboard, (3) the campus experiment leader dashboard, and (4)
229
the Safe Blues data dashboard. All four components are available online at [47], [4], [48],
230
and [49] respectively. Fig 3 displays a snapshot of these four components. 231
Strand and device management 232
Participants run the Safe Blues app [47] (see also Fig 3 (top left) for an illustration of 233
the app) as they go about their normal day to day activities on campus while enabling
234
Bluetooth and location services. Location services are only needed for prize based 235
rewards as described below. A participant’s app then communicates with the apps of 236
other participants via Bluetooth to pass on digital ‘virus-like’ tokens, namely Safe Blues
237
‘strands’. This simulates an epidemic spreading through the community. There are 238
many types of strands of such virtual safe epidemics, and the live emulation of all of the
239
epidemics happens in parallel driven by the actual physical contact processes of 240
March 29, 2022 7/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Fig 3. The Safe Blues system: the Safe Blues app (top left), the simulation dashboard (top
right), the campus experiment (campus hours leader board on bottom left), and the data
dashboard (bottom right).
participants. 241
The app is not malicious and does not interact with any other app that users may be
242
running on their phone. Open-source code is available on GitHub via the Safe Blues 243
website [50]. Nevertheless, the app, like any other mobile app, consumes the battery of
244
the phone. It is the participants’ responsibility to manage their phone battery usage, 245
and our experience has shown that some participants turn off the app while away from
246
the campus. The app is only available for Android due to the fact that iOS phones 247
cannot run such an app in the background. This clearly limits the participating 248
population. We discuss the technical details of the app software in S2 Appendix. 249
With the exception of participant reward information, described in 250
subsection Participant management and ethical considerations below, information 251
recorded by the Safe Blues system is limited to the aggregated counts of each strand. 252
Every 15 minutes a phone uploads the status of its infections in terms of ‘exposed’, 253
‘infected’, and ‘recovered’ for each strand. The total number of ‘susceptibles’ is then 254
inferred based on the total number of phones participating at any given time. This 255
uploading occurs via a temporary anonymous 256 bit ID which changes every 24 hours
256
on the phone. Thus the Safe Blues server does not keep track of the individual infections
257
of phones and it cannot uniquely identify a phone beyond a 24 hour period. The 258
temporary ID is still useful for correct counting of infections on the server side, since 259
messages are sometimes lost or not sent if the phone is without connectivity (see S4 260
Appendix where we describe the algorithm for interpolation and imputation of counts to
261
handle this). The individual phone strand information is never cross-referenced with 262
private participant information, further preserving the anonymity of participants. 263
March 29, 2022 8/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
The injection of new infections into the participating phone population is carried out
264
via an API (Application Program Interface) available to the phones. In each phase of 265
the experiment we inject multiple strands with each batch containing a collection of 266
individual strands. For example, in phase 1, where we focused on testing and tweaking
267
the system, there were 7 batches in total labeled 1.01 to 1.07. Similarly, in phase 3 268
there were 22 batches in total, labeled
3.01
to
3.22
. We discuss the number of strands
269
in each batch, their parameters, and their purposes in the experiment in S5 Appendix. 270
When a batch of strands is ‘injected’, all participating phones become aware of the 271
strands of the batch and each strand has a pre-specified seeding probability which is 272
typically set to 0.05, 0.1, or 0.2. Then at a specified start time, each phone is 273
independently infected by the new strand in accordance with the seeding probability. 274
This ‘seeding’ of new strands thus emulates the arrival of new outbreaks of the epidemic
275
into the population. 276
We applied four types of epidemic models for the phone population: SEIR, SIR, SEI,
277
and SI. In both SEIR and SEI models, when a susceptible phone receives a strand, it 278
first becomes exposed and remains in this state for a random time period known as the
279
incubation period. During the incubation period, the phone is unable to infect other 280
phones. After the incubation period, the phone becomes infected and remains in this 281
state for a random time period known as the infection period. During this time, the 282
phone has the ability to infect nearby phones by exchanging a Bluetooth token. The 283
distribution of the incubation and infection periods is explained further below. The 284
infection period in the SEIR type epidemic is finite, and the phone stops infecting other
285
phones at the end of this period. Consequently, its state is labeled as recovered. The 286
infection period in the SEI epidemic is infinite, and the phone cannot recover. There is
287
no incubation period in the SIR and SI epidemics, and a susceptible phone becomes 288
infected immediately after receiving the strand. The SIR epidemic has a finite infection
289
period and the phone recovers at the end of it. The infection period of the SI epidemic
290
is infinite, and the phone remains in the infected state throughout the epidemic. 291
During phases 1-3, and including the intervening period between phase 3 and 292
phase 4, we injected 4155 strands in total into the system. Of these, 28% are of the SI 293
or SEI type (not involving removal/recovery) and the remaining 72% are of the SIR or
294
SEIR type (allowing recovery). Fig 4 depicts the cumulative counts of strands released
295
over time during phases 1-3. 296
Beyond the classification of SI, SEI, SIR, and SEIR, each strand, uniquely identified
297
by a strand id, which has specific parameters that influence its spread. A full 298
specification of the protocol for these parameters is in Appendix A1 of [2]. However, the
299
protocol there does not deal with specific distributional information and the infection 300
probability mechanism. Hence we now outline these details. 301
We use gamma distributions for both the incubation and infection times and 302
parameterize them by a mean
µ
and a shape parameter
k
. That is, for each
x∈
(0
,∞
),
303
the probability density function is 304
f(x;µ, κ) = 1
Γ(κ)µ
κκxκ−1e−κx
µ,(1)
where Γ(·) is the gamma function. In this case, the ratio between the variance and the 305
square of the mean is 1/k. In many cases we used k= 3; in other cases to create 306
(nearly) deterministic times we set k= 10,000. 307
The other important strand information deals with the probability of infections of 308
nearby phones. At every time where two participating phones are near each other, 309
Bluetooth messages are exchanged during a session and throughout this session, 310
distance measurements are carried out. In principle using individual Bluetooth messages
311
may appear to be preferable to creating sessions. However the nature of the Bluetooth
312
March 29, 2022 9/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Fig 4. The cumulative number of strands over time broken up into SI, SEI, SIR, and
SIER types and phases of the experiment.
protocol and the underlying software implies that sessions are the preferable technique;
313
see S2 Appendix. 314
At the end of the session (capped at 30 minutes), the median distance from all 315
messages is computed and denoted by d. The duration of the session is denoted by t.316
We expect that with closer distances and longer duration, infection is more likely. We 317
chose the probability of infection parameterized by the strand’s strength, σ, and the 318
maximal infection distance, ρ, to be 319
p(d, t;σ, ρ) = (1−e−σt(1−d/ρ), d < ρ,
0d≥ρ. (2)
In general, we expect strands with higher ρor higher σto be more infectious. 320
In setting the strand parameters σand ρwe initially used a simulation model (see 321
subsection A campus simulation model for details). Subsequently, we adjusted the 322
parameters based on field experience (see section Results and discussion section for 323
details). 324
Participant management and ethical considerations 325
Our goal in participant management is to motivate participants to run the Safe Blues 326
app while on campus. A first decision was whether to couple participation data with 327
strand data (number of virtual infections). We chose not to do so. While such coupled 328
data could be useful, our primary goal is the strand-count time series for which coupling
329
is not needed. 330
A second decision was whether to pay participants a ‘flat rate’ for participation, for
331
example with coffee vouchers proportional to their participation hours or to use prizes. 332
As the total budget was limited, and in accordance with other experimental 333
research [51,52], we opted for prizes. As this is a digital experiment we chose iPad, 334
Android phone, and Fitbit prizes, with 9 prizes per prize draw. See S1 Appendix for 335
details of the prize draw rules. 336
Participants were recruited directly via online flyers, posters and videos. To take 337
part in the experiment, participants first needed to install the Safe Blues Android app 338
March 29, 2022 10/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
on their mobile phones. This gave them a random 10 digit ID which identifies them 339
only for purposes of experiment participation and prizes but is not associated with their
340
strand infections. With this ID, participants can then register their email address which
341
is used for communicating experiment messages and prize winners. 342
As participants enter the city campus, an Android geofencing mechanism spawns an
343
event on the app, and then when they leave the campus (leave the geofence) an 344
additional event is spawned. A message of participation hours is recorded on a server. 345
The participation hours contribute to the chance of winning a prize. In general, the 346
more hours a participant runs the app on campus, the higher the chance of winning a 347
prize (see S1 Appendix). The app does not track the location of participants with the 348
exception of indicating whether or not the participant is within the campus geofence 349
area. Fig 5 (left) displays a snapshot map of the UoA City Campus with the geofenced
350
area marked on it.
north
W
S
E
00200200
5050 100100 150150
0033
1122
SCALE 1:4 000 (1cm – 40m)SCALE 1:4 000 (1cm – 40m)
minutes (at 4km/h walking pace)minutes (at 4km/h walking pace)
metresmetres
CG0053
15
3
18
29
15B
15A
16A
7
7A
16
1V
4 4A
2
26
21
14
20A
20B
20
6
24
38
37
Authorised
vehicles only
Authorised
vehicles only
P
Pay and Display
1
1616
1616
16
1616 1616
16
11
11
1
1
1
1
16
1
314314
109109
104104
106106
315315
322322 201E201E
206206
207207
219219
220220
250250
240240
241241
810810
820820
801801
802802
803803
804804
805805
135135
132132
123123
107107
122122
121121
120120
102102
110110
124124
125125
114114
113113
103103
118118
119119
273273
274274
275275
226226
225225 253253
251251
252252
255255
312312
260260
408408
405405
402
402
401401
423423
410410
421E421E
422422
409409 400400
200200
212212
215215
216216
105105
800800
100 sector100 sector
311311
301301
303303
300300
813813 814814
502
501
503
504 505
628
643 629
523 507
528
529
532532
534534
599
614614
620620
600600
619619
439439
431431
432432
433433
440440
436436
435435 400400
601601
602602
603603
600600
500 sector
600
Grafton Campus
Newmarket
Campus
City CampusCity Campus
901
902
903
904
907
900 sector
831831
832832
833833
834834
Newmarket
Railway Station
Grafton
Railway Station
Mount Eden
Railway Station
Britomart
Railway Station
(underground)
The Strand
Railway Station
(Britomart backup station)
Auckland
Ferry Terminal
Pier 1
Pier 3-4
Pier 2
911
908
910
302302
906
835835
836836
837837
909
621
623
632
Parnell
Railway Station
303S303S
110N110N
105S105S
201N201N
421W421W
434434
511
509
616616
203203
204204
316316
501C
130130
441441
442442
645
644
450450
840840
646
310310
(under construction)(under construction)
242242
(under
construction)
(under
construction)
818818
Grafton
Campus
Grafton
Campus
Recreation
Centre
Recreation
Centre
General
Library
General
Library
Old Choral
Hall
Old Choral
Hall
BiologyBiology
Student
Comm
Student
Comm
Kate Edger
Info Comm
Kate Edger
Info Comm
Social
Sciences
Social
Sciences
MusicMusic
Kenneth Myers
Centre
Kenneth Myers
Centre
Newman
Hall
Newman
Hall
Fisher
Building
Fisher
Building
Maclaurin
Chapel
Maclaurin
Chapel
Old Government
House
Old Government
House
University
House
University
House
George Fraser
Gallery
George Fraser
Gallery
ThomasBldgThomasBldg
Cultures,
Languages
& Linguistics
Cultures,
Languages
& Linguistics
MaraeMarae
Fale Pasifika
Complex
Fale Pasifika
Complex
High CourtHigh Court
Pullman HotelPullman Hotel
Copthorne
Anzac Ave
Copthorne
Anzac Ave
UniLodge
Anzac / Beach
UniLodge
Anzac / Beach
Sir Owen G Glenn
Building
Sir Owen G Glenn
Building
Conference
Centre
Conference
Centre
Alfred Nathan
House
Alfred Nathan
House
LodgeLodge
Science
Centre
Science
Centre
Science
Centre
Science
Centre
Starship Children’s
Hospital
Starship Children’s
Hospital
O’Rorke
Hall
O’Rorke
Hall
Auckland
Bioengineering
House
Auckland
Bioengineering
House
Elam
School
Elam
School
University
Hall Towers
University
Hall Towers
University Hall
Apartments
University Hall
Apartments
WaipārūrūHall
- Whitaker
WaipārūrūHall
- Whitaker
Auckland City HospitalAuckland City Hospital
Grafton
Student Flats
Main Building Auckland
War Memorial
Museum
Auckland
Domain
Auckland
Domain
Auckland
Domain
The ClockTowerThe ClockTower
Albert
Park
Albert
Park
PORT OF AU CKLA ND
Mahuhukiterangi
Reserve
Mahuhukiterangi
Reserve
Te Taou
Reserve
Te Taou
Reserve
Fraser ParkFrase
Scarborough
Reserve
Tahaki Reserve
WintergardenWintergarden
Auckland
Town Hall
Aotea Centre
Myers
Park
Carlaw Park
Student Village
Carlaw Park
Student Village
Emily Place
Park
Emily Place
Park
Alten
Reserve
Alten
Reserve
Grafton
Cemetery
East
Grafton
Cemetery
East
Grandstand
Ferry Building
Auckland
Public Library
Auckland
Public Library
Newmarket Park
Endeavour
Park
Olympic
Corner
Lumsden
Green
Station
Square
Outhwaite
Park
Mountain
Reserve
Khyber Pass
Reserve
Basque
Park
Grafton Cemetery
West
Alberon
Reserve
Bassett
Reserve
MonaMona
OkaretaOkareta
BelgraveBelgrave
BayreuthBayreuth
John Hood
Plaza
John Hood
Plaza
Beach Road CyclewayBeachRoad Cycleway
CyclewayCycleway
GullyGully
GraftonGrafton
Grafton Gully CyclewayGrafton Gully Cycleway
HumanitiesHumanities
55 Symonds55 Symonds
Pembridge
House
Pembridge
House
Waikohanga
House
Waikohanga
House
UniLodge
Whitaker
UniLodge
Whitaker
Grafton Hall
Ko te Ako
o te Tui
Waipārūrū
Hall
Waipārūrū
Hall
Recreation
Centre
Recreation
Centre
Ray Meyer
Building
Court
Te Tirohanga
o te Tōangaroa
Te Tirohanga
o te Tōangaroa
ArchitectureArchitecture
Auckland
Domain
Auckland
Domain
88554422
77553A3A 33
1818 1616 1414 1212 1010 66
33
2525 1515 1111 9911
1616
1818
20-2620-26
1616 1818
138138
1616
1818
19A19A
2525
7474
2323
881010
1133557799
1212
1-111-11
1515
1717
99
2222
2424
2626
3434
3636
3838
2323 1717 1111
2020
2222
2424
2626
4040
133133
1616
4949
7070 5858
1414 3030 2020
6767
2727
1616
9393
89-9189-91
85
40 50 62 70
2-6 828
6
19-2619-26
38
10A
10
4444
9-11
9
5555
25-3525-35
262 368
3131
7070
128128
Alfred StreetAlfred Street
StreetStreet
AltenAlten
RoadRoad
Anzac AvenueAnzac Avenue
Waterloo Quadr antWaterloo Quadra nt
Parliament StreetParliament Street
BowenLaneBowenLane
Bowen AvenueBowen Avenue
Wynyard StWynyard St
Princes StreetPrinces Street
CharlesNalden LaneCharlesNalden Lane
Churchill StreetChurchill Street
StreetStreet
StanleyStanley
NichollsNicholls
StanleyStanley
StreetStreet
Lower DomainLower Domain
DriveDrive
CarlawCarlaw
ParkPark
AvenueAvenue
Grafton RoadGrafton Road
Short StreetShort Street
EdenEden
CrescentCrescent
PlacePlace
Emily PlaceEmily Place
EmilyEmily
BanksideBankside StreetStreet
ShortlandShortland
StreetStreet
StreetStreet
PrincesPrinces
SymondsSymonds
Claybr
Road
o
ok
Mount StreetMount Street
SaintSaint
Martins LaneMartins Lane
Whitaker PlaceWhitaker Place
PlacePlace
WhitakerWhitaker
SymondsSymonds StreetStreet
Park Road
Park Avenue
Seafield View Road
Boyle
Cres
Glasgow Terrace
Park Road
Park Road
Domain
Grafton RoadGrafton Road
GraftonGrafton
RoadRoad
KariKari
StreetStreet
Fencroft StFencroft St
Carlton Gore Road Carlton Gore Road
Grafton
Road
Beckham Place
Wellesley Str eetWellesley Street EastEast
St Paul StreetSt Paul Street
St Paul StreetStPaul Street
Wakefield StWakefield St
Lyndock StLyndock St
Airdale StreetAirdale Street
City Road
LaneLane
TennisTennis
Grafton BridgeGrafton Bridge
gahape
Road
Chancery StreetChancery Street
Bacon LaneBacon Lane
Freyberg PlFreyberg Pl
FieldsFields
LaneLane
O’Connell StreetO’Connell Street
rthouse Lanerthouse Lane
CouCou
HighHigh
StreetStreet
Halsey Street
Pakenham Street East
Viaduct
Harbour
Avenue
Lucy
Lane
Market Place
Market Lane
Sturdee Street
Lower
Hobson Street
Customs Street
West
Customs
Street West
Fanshawe Street
Fanshawe
Street
Quay Street
Customs Street East
East
Queen Street
Quay Street
Tyler Street
Tyler Street
Galway Street Galway Street
Commerce
St
Commerce
St
Street
Gore StGore St
Street
Fort LaneFort Lane
Lower Albert Street
Wolfe Street
Federal StreetF
StreetStreet
SwansonSwanson
MillsMills
Wyndham Street
Wyndham StreetWyndham Street
LaneLane
Mills Lane
Lane
Federal Street
Kingston Street
Durham LaneDurham Lane
Durham Street WestDurham Street West
Albert StreetAlbert Street
BrandorLane
Bouzaid Way
Gorst Lane
Vogel Lane
Graham
Stre
Hardinge
Street
et
Victoria Street West
Nelson Street
Victoria Street West
West
DurhamStreet EDurham Street E
Vulcan
Ln
Vulcan
Ln
GoreStreet LaneGore Street Lane
FortFort
StreetStreet
BritomartBritoma
Place
TeA ra Tuhuhu Wa lkway
BeachBeach
AnzacAnzac
AvenueAvenue
StreetStreet
KitchKitch
Kitchener StreetKitchener Street
Lorne StreetLorne Street
Lorne StreetLorne Street
St Patricks
Square
Hobson Street
Queen StreetQueen Street
Wellesley Str eet West
Wellesley Street West
West
Sale St
Cook Street
Nelson Street
enerener
Tinley Street
Plumer Street
Tangihua StTangihua St
Te Taou CrescentTe Taou Crescent
Te Taou CrescentTe Taou Crescent
CrescentCrescent
MahuhuMahuhu
DocksideLaneDockside Lane
RoadRoad
RoadRoad
BeachBeach
RonayneRonayne
StreetStreet
PlaceP
NgahoeNgahoe
Shipwright LnShipwright Ln
TheThe
StrandStrand
The Strand
The
Strand
Quay Street
Gladstone Road
Faraday Street
Kenwyn Street
Watt Street
Augustus Terrace
Eglon Street
Marston Street
Fox Street
York
Street
Earle Street Earle Street
Parnell Road
Tamaki Dr
Churton
Street
Bath Street
Parnell RiseParnell Rise
Bradford Street
Street
Bath
Garfield
Street
Garfield Street
Bradford
Street
Windsor Street
Farnham
Street
Cleveland Road
Saint Georges Bay
Road
Ruskin
Scarborough
Street
Ruskin Street
Terrace
Denby Street
BedfordBedford
Street
Cracroft Street
Mutu
St
Mutu
St
StreetStreet
Heather
Street
Akaroa Street
TildenSt
Falcon Street
Tika St
Cheshire
Street
Gibraltar
Crescent
Aorere Street
Parnell Road
Parnell Road
ScarboroughLane
Saint Georges Bay Road
Saint Stephens Avenue
Birdwood
Cres
cent
Cath
edral Place
Ayr Street
Laurie Avenue
Dom
ain
Drive
Parnell Road
Cowie Street
Maunsell Road
Titoki Street
Maunsell Road Extension
Cenotaph RoadCenotaph
Cenotaph Road
MuseumMuseum Circuit
LittleGeorge St
Football
Road
Football Road
Wintergarden RoadWintergarden Road
RoadRoad
Kiosk
osk
CresCres
TheThe
centcent
DomainDomain
DriveDrive
George
Street
George Street
Morgan Street
Clayton Street
Alma Street
Sarawia Street
Railway Street
YoungsLane
Laxton
Terrace
Broadway
Crescent
Davis
Broadway
Symonds Street
Exmouth Street
Basque Road
Dundonald Street
Burleigh Street
Nugent Street
New
North
Road
Boston Road
Water Street
Lauder Road
Auburn Street
Road
Boston
Severn Rd
Arawa Street
EdwardWayte
Place
Khyber Pass Road
Claremont Street
Parkfield Terrace
Huntly Avenue
Suiter Street
Kingdon Street
Short Street
Leek Street
eaux Way
John
Stokes
Terrace
Maui
Grove
Joseph
Banks Terrace
Broadway
Teed Street
Kent Street
Osborne Street
York Street
Eden Street
Bourke Street
Morrow Street
Mortimer Pass
Coventry Lane
Balm Street
Nuffield Lane
Nuffield
Street
Crowhurst Street
Seccombes Road
Roxburgh Street
Melrose Street
McColl Street
Maungawhau Road
Lion Place
Almorah Place
Almorah
Gillies Avenue
Remuera Road
Belmont Terrace
Middleton
Mamie Street
Lauriston Avenue
Ada Street
Ely Avenue
Nikau
Street
Shaddock Street
Korari Street
Flower Street
Ruru Street
Akiraho Street
Sylvan Avenue East
We
Street
Edenvale Cres
Edenvale Cres
GraysonsLane
Puka St
Mount Eden
Road
Mary Street
Harold Street
Kohekohe Street
Edwin
Street
Enfield
Street
Normanby
Road
Normanby
Road
Road
Clive
Clive Road
Barn
ett Crescent
Mountain Road
Gilgit
Road
Rockwood
Place
Esplanade Road
Bassett Road
Middleton Road
Southern
Motorway
Road
Bassett
Furn
Road
James
Cook
Crescent
Road
Seccombes Road
Mountain
Road
Khyber Pass Road
Khyber Pass Road
Mount Eden Road
Kelly
ArotauPlace
Madeira Place
Burton St
MadeiraLane
Nor
Street
Newton Road
Symonds Street
LowLow
Domain DriveDomain Drive
erer
LoversLovers
LnLn
Drive
Road
Grandstand Road North
Grandstand
South
Hobson Street
Albert Street
Federal Street
ElliottElliott
Darby StreetDarby Street
StreetStreet
Queen Streett
Rutland StreetRutl and Street
Airdale Street
dale Street
GovernorFitzroy PlGovernorFitzroy Pl
Mayoral
Drive
Lorne StreetLorne Street
Mayoral
Drive
DriveDrive
MayoralMayoral
Queen Street
Upper
Randolph Street
Saint
Benedicts Street
Street
Karaka
Dacre Street
West Street
St Sth
France
Stable Lane
Hohipere Street
Glenside Cres
Alex Evans Street
Ian
Mckinnon
Drive
Canada
Street
Cross Street
South St
East Street
cury Lane
r
Me
Karan
Queen Street
Upper
Scotia Place
Liverpool Street
Turner Street
Queen Street
Waverley Street
MarmionStreet
t
WhiteWhite
StreetStreet
Pitt Street
Union
St
Hopetoun St
Vincent Street
Greys Avenue Greys Avenue
Nicholas
Street
PoyntonTerrace
Beresford Sq
LaneLane
GraftonGraf ton
MewsMews
st
wich
R
Q
P
O
N
M
L
K
J
I
H
G
F
E
D
C
B
59 000m18 N
182
184
186
188
190
192
194
196
198
200
202
204
206
208
59 000m
21 N
17 800m
56 E
57
0
57
2
57
4
57
6
57
8
58
0
58
2
58
4
58
6
58
8
17 000m
59 E
56789101112131415164
R
Q
P
O
N
M
L
K
J
I
H
G
F
E
D
C
B
59 000m18 N
182
184
186
188
190
192
194
196
198
200
202
204
206
208
59 000m
21 N
17 800m
56 E
57
0
57
2
57
4
57
6
57
8
58
0
58
2
58
4
58
6
58
8
17 000m
59 E
5678910
11 12 13 14 15 16
4
KK
JJ
II
HH
GG
FF
EE
DD
7788991010
1111 1212
KK
JJ
II
HH
GG
FF
EE
DD
88991010 1111
1212
201 ArtsG10
260 Business SchoolH10
421E Creative Arts & IndustriesH9
402 EngineeringH9
810 Auckland Law SchoolE10
505 Medical & Health SciencesM9
301 ScienceH9
Security
Emergency Phone
Hospital
Information Desk
Childcare Centre
Disability Services
Food & Drink Outlet
Health & Counselling
Library
Pharmacy
Playground
Postal Agency
Public Toilet
Recreation Centre
University Bookshop
Access Parking
Bike Rack, locked
Bus Stop (city service)
Cycleway
Traffic Lights
Underpass
Visitor Parking
Pedestrian Crossing
Staff Parking, Area No
Railway Crossing
Railway Station
0 200
50 100 150
03
12
SCALE 1:4 000 (1cm – 40m)
minutes (at 4km/h walking pace)
metres
Faculty Student Centres
Security
Information Desk
Hospital
Access Parking
Bus Stop (city service)
Cycleway
Security
Information Desk
Hospital
Access Parking
Bus Stop (city service)
Cycleway
Traffic Lights
Underpass
Visitor Parking
Traffic Lights
Underpass
Visitor Parking
Pedestrian CrossingPedestrian Crossing
Staff Parking, Area NoStaff Parking, Area No
26
P
CITY CAMPUS
GRAFTON CAMPUS
CLINICS
LIBRARIES
CG0042
NEWMARKET CAMPUS
49 Symonds Street 620I8
55 Symonds Street 616I7
Academic es 620Programm I8
Academic Services 105F9
Accommodation Solutions 315G9
Accounting & Finance, Dept of 260G,H10
Acoustics Research & Testing Service 422H9
Alfred Nathan House 103F9
Alten Road Early Childhood Centre 241F11
Alumni Relations 135& Development E9
Applications & Admissions 105F9
Architecture & Planning, School of 421E, 421WH9
Arts, Faculty Office 215G10
Asia-Pacific Excellence, Centre for (CAPE) 130F9
Asian Studies 207G10
AskAuckland Central 103F9
Auckland Bioengineering e 439Hous I8
Auckland Bioengineering Institute 439I8
Auckland UniServices Ltd620 I8
Auckland University Press 810E10
Auckland University Students’Association,AUSA 322 G9
Bayreuth House 220G10
Belgrave House 212G10
Biological Sciences, School of 106, 110104, G10, F10
Biology Building 106G10
Business & Economics, Faculty Office 260 G,H10
Business School Computer Laboratories 260G,H10
Campus Life 315G9
Campus Store311 G9
Careers Services 531 G9
Carlaw Park Student Village 831-837H12
Catholic Tertiary Centre 805E10
Centre of Methods & Policy Application in the Social Sciences
(COMPASS) 260G,H10
Chemical & Materials Engineering, Dept of 401201E, 201N,
G10, H9
Chemical Sciences, School of 301-302H9
Civil & Environmental Engineering, Dept of 401201E, 201N,
G10, H9
Commercial Law, Dept of 260G,H10
Communications103 F9
Comparative Literature, Centre for 207G10
Computer Science, Dept of 303SH9
Conference Centre 423H9
Confucius Institute 013 F9
Conservation International 302H9
Counselling 315G9
Creative Arts & Industries, Faculty Office 423 H9
Cultures, Languages & Linguistics, School of 207G10
Cultures, Languages & Linguistics (CLL) Building 207G10
Dance Studies 113, 820F10, D9
DELNA315 G9
DigitalServices 435 I8
Disability Services 105F9
Drama Studio 206G10
Economics, Dept of 260G,H10
Education (Liberal Arts Programme) 201EG10
Elam School of Fine Arts 113, 431-433F10, I8, I9
Electrical & Computer Engineering, Dept of 301, 303, 401H9,
G9
Engineering, Faculty Office 401H9
Engineering Science, Dept of 439I8
English Language Academy 619I,J7
English Language Enrichment (ELE) 315G9
Environment, School of301-302 H9
Equity Office 105 F9
European Languages & Literatures 207G10
Examinations& Timetable Services Office (ETSO) 620 I8
Examinations Office 105F9
Facilities Management620 I8
Fale Pasifika Complex 273-275G10
Financial Services 620I8
Fine Arts, Elam School of 113, 431-433F10, I8, I9
Fisher Building 804E10
Food Science Programme 302H9
Forensic Science Programme 301H9
George Fraser Gallery 132F9
Gender EStudies 201 G10
Graduate School of Management 260G,H10
Graduation 620I8
Grounds Maintenance 120-122F10
Gus Fisher Gallery 820D9
Health, Safety & Wellness Manager620 I8
Health Services 315G9
Human Resources 620I8
Humanities, School of 206G10
Humanities Building 206G10
ID Card Centre 103F9
Information Systems & Operations Management, Dept of 260
G,H10
Innovation & Entrepreneurship, Centre for 402H9
International Office 103F9
International Relations Office 103F9
iSPACE (for international students) 315G9
James Henare Mori Research Centre 225āG11
John Hood Plaza 260H10
Kate Edger Information Commons 315G9
Kenneth Myers Centre 820D9
Key, Access & Parking Control 409H9
Language Laboratories207 G10
Law, Faculty Office 801E10
Learning & Research in Higher Education, Centre for (CLeaR)
804E10
Learning & Teaching Development 315G9
Learning Environment Support Unit (LESU) 260G,H10
Legal Research Foundation 803E10
Lippincott Cottage 118F10
Lodge, Old Government House 123E10
Maclaurin Chapel & Chaplains 107E10
Management & International Business, Dept of 260G,H10
Mori Material Culture Workshop 226ā, Te Ruawaihanga G11
Mori & Pacific Studies - Te W nanga o Waipapa 253ā ā 226, G11
Marae 251-252G11
Marketing103 F9
Marketing, Dept of 260G,H10
Mathematics, Dept of 303G9
Maurice Wilkins Centre for Molecular Biodiscovery106, 110
G10, F10
Mechanical Engineering, Dept of 401201E, 201N, G10, H9
Media Productions804 E10
Mona House 216G10
Music, School of 250, 804, 820G10, E10, D9
Muslim Prayer Room 301H9
National eScience Infrastructure (NeSI) 439I8
Newman Hall 805E10
New Start206 G10
New Zealand Asia Institute 260G,H10
New Zealand Leadership Institute 260G,H10
Ng Pae o te M ramatanga (National Institute of Researchāā
Excellence for Mori Development & Advancement) 253āG11
Office of Research Strategy & Integrity 105, 620F9, I8
Okareta House 215G10
Old Choral Hall 104G10
Old Government House 102F10
Old Government House Bike Cage 124F10
Organisational Performance & Improvement 620I8
O’Rorke Hall 614I8
Pacific Studies, Centre for 273G10
Pembridge House 130F9
People & Organisational Development (POD) RoomsTraining
620I8
Photographers 804E10
Physics, Dept of 303G9
Planning & Quality Office 105F9
Prime Minister’s Chief Science Advisor (PMCSA) 810E10
Property, Dept of 260G,H10
Property Services 409, 620 H9, I8
Psychology, ofSchool 302G10
Records, Enrolment & Fees 620I8
Recreation Centre 314, 840G9, H11
Research Operations Centre 620I8
Retail Facilities 311, 315G9
Scholarships 620I8
School of Graduate Studies 103F9
Schools Partnership Office 103 F9
Science, Faculty Office 302 H9
Science Centre 301303- , 303SG9, H9
Security 409H9
Shared Services 620I8
Short Courses 260G,H10
Short Street 810E10
SirOwen G Glenn Building 260 G,H10
Social Sciences, School of 201EG10
Social E, 201NSciences Building 201 G10
Staff Common Room 102F10
Statistics, Dept of 303G9
Strata, Postgraduate Commons 315G9
Student Commons 315311, 312, G9
Student Contact & Support 620I8
Student Information Centre 103F9
Student Union 311-312G9
Symonds Street Early Childhood Centre 410H9
Te Khanga Reo 255ö G11
Te Tirohanga o te Tōangaroa 818E11
Te Wänanga o Waipapa - Mäori Studies & Pacific Studies,
School of 253, 273-275G11, G10
Tertiary Foundation Certificate Programme 206G10
The ClockTower 105, 105S, 119F9
Thomas Building 110F10
Translation Studies, Centre for 207G10
Tuition Fees103 F9
UniLodge Anzac / Beach 813-814F11
UniLodge Whitaker 450J8
University Careers Services 531 G9
University Hall Apartments 436 I8
University Hall 440Towers I8
University House 135E9
University of Auckland Foundation 135E9
University of Auckland Society 135E9
Vice-Chancellor’s Office 105F9
Waikohanga House 434I8
Waipärürü Hall 441-442J8
Waipärürü Hall–Whitaker 601-603 J,K7, J,K8
DIRECTORYDIRECTORY
General Library109 G9
Law (Davis) 802E10
LIBRARIESLIBRARIES
Arts FO: 215 , SC: 201EG10 G10
Business & Economics FO: 260 , SC: 260G10 H10
Creative Arts & Industries FO: 423 , SC: 421EH9 H9
Engineering FO: 402 , SC: 402G10 H9
Law FO: 801 , SC: 810E10 E10
Science FO: 302 , SC: 301I8 H9
Research Services, Libraries and Learning Services
© The University of Auckland, January 2020
Research Services, Libraries and Learning Services
© The University of Auckland, January 2020
FACULTY OFFICES (FO)
STUDENT CENTRES (SC)
FACULTY OFFICES (FO)
STUDENT CENTRES (SC)
Fig 5. The University of Auckland City campus with the geofenced area supporting the
experiment marked as a circle (left). A heatmap representation of buildings used in the
simulation prior to the experiment (right).
351
As an additional side-benefit of the experiment, we provide the aggregated visits to
352
campus and duration statistics as part of the dataset. The left side plot in Fig 6 depicts
353
the daily number of participants who were registered, reporting, and attending the 354
campus in the experiment during phases 1–3. The right side plot in this figure displays
355
the distribution of the means for the daily campus hours collected by participants, over
356
weekdays and weekends, during phases 1–3. As per the prize draw rules (S1 Appendix)
357
the maximum daily campus hours that each participant can collect is capped at 10. 358
By the end of phase 2, about 20% of the registered participants were not running the
359
app. In an attempt to enroll more participants in the experiment, we upgraded the 360
reward scheme during phase 2. This included an ‘invite-a-friend’ option, which 361
increased a participant’s chance to win a prize if new participants joined the experiment
362
through their invitation. Those who joined the experiment through the invite-a-friend 363
mechanism were also rewarded with bonus eligible hours. See S1 Appendix for further 364
details about rules and the invite-a-friend mechanism. Joining the prize draws was not
365
compulsory for those taking part in the experiment. 366
March 29, 2022 11/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Fig 6. Evolution of the daily number of participants who were on campus (Attending),
reporting, and the cumulative number of participants who were registered in the experiment,
during phases 1 to 3 (left). Box plot of the means from the 5 number summaries for daily
campus hours (right).
Although we were not conducting any clinical or health research including human 367
data, we required approval from the University of Auckland Human Participants Ethical
368
Committee (UAHPEC) before doing any form of research involving university 369
volunteers. The study was approved by UAHPEC under ethics number 22143 in March
370
2021. In this application we addressed ethics considerations, including naming all 371
researchers, description of the study, location of study, methodology, participants and 372
recruitment process, data management, funding, M¯aori-focused consultation and 373
engagement, and consistency with the principles of Te Tiriti o Waitangi. 374
We also provided the ethics committee with a copy of all the Safe Blues website 375
pages, a permission letter from course directors (for big statistics courses where the 376
project was advertised) and head of the Department of Statistics, participant 377
information sheet, poster, the consent form, the data management plan, and the 378
observation schedule. 379
Data management 380
The experiment is managed through two distinct databases. A Participant Management
381
System (PMS) is used to store the email addresses and consent agreements, as well as a
382
record of the campus hours. The PMS is hosted at the UoA, and data are used only for
383
the purposes of managing the prize draws and the list of participants. Data in the PMS
384
is completely disconnected from the experiment data and will not be publicized in any 385
way, with the exception of analysis of aggregated participation counts over time (for 386
example, see Fig 6). 387
The second database, called the Anonymous Data Server (ADS), is managed in the
388
cloud and contains an aggregate, anonymized, time-stamped record of the number of 389
phones with each strand. For each strand, we record this data on an hourly and daily 390
basis, and indicate the aggregate number of phones in each epidemiological state 391
(susceptible, exposed, infected, recovered) over time. As this database follows the Safe 392
Blues protocol, phone (app) identities are not revealed during communication, and 393
phones (apps) only have temporary IDs that are replaced on a daily basis. We spell out
394
the technical details of how we record data in the PMS and ADS in S3 Appendix. 395
We record aggregated participation counts from the PMS as daily and hourly 396
measurements, along with the daily and hourly means and five number summaries–that
397
is, the minimum, first quartile, second quartile, third quartile, and maximum–of campus
398
hours in a CSV file. On days where there are fewer than 5 participants these numbers 399
are omitted for privacy reasons. Similarly, the aggregate Safe Blues data from the ADS
400
March 29, 2022 12/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
are also stored in several CSV files, one for each strand. See S6 Appendix for specific 401
details of the Safe Blues data repository. The Safe Blues data will be made publicly 402
available in the Safe Blues data repository after the experiment concludes. Plots of the
403
data are currently available as a web-based dashboard at [49]. See Fig 3, bottom right,
404
for a snapshot of the dashboard. 405
By agreeing to take part in the experiment, a participant agrees to share the Safe 406
Blues data of their Safe Blues app. At any point in time, a participant may choose to 407
withdraw from the experiment and this will result in deletion of their personal 408
information from the PMS. However, their aggregated anonymized data already 409
recorded on the ADS will remain in the database and will potentially contribute to the
410
scientific findings of the experiment. 411
A campus simulation model 412
We used a simple simulation model to approximately capture the expected behavior of 413
the participants, and to aid with our initial choice of strand parameters to be used in 414
the experiment. This discrete-time stochastic spatial compartmental SEIR model was 415
used as an initial guide for ranges of the maximal infection distance and infection 416
strength parameters in Eq (2). 417
A Safe Blues strand is characterized by its seeding probability,
π
, infection strength,
418
σ, maximal infection distance, ρ, incubation time distribution, and infection time 419
distribution. In the simulation, the initial infections were determined by Bernoulli 420
random variables with each simulated participant independently having a chance πof 421
becoming infected when the strand was activated. The remaining participants could 422
only become infected by being a ‘close contact’ of an already infectious person. After 423
each time step, the positions of participants were independently drawn from a heat map
424
designed to resemble likely locations attended by participants in the real-world 425
experiment; see Fig 5 (right). The time step considered here was 1 hour, which 426
corresponds to the duration of a lecture. Each susceptible individual who was
d
meters
427
away from an infectious individual for
t
minutes was infected with the probability given
428
by Eq (2). 429
After a strand was successfully transmitted to a susceptible individual, they became
430
exposed and remained in this compartment for a random amount of time (incubation 431
time) drawn from a gamma distribution with mean µEand shape κEas in Eq (1). 432
Subsequently, once their exposure time elapsed, they became infected and were able to
433
infect further individuals with this strand. The duration of their infection (infection 434
time) was again gamma distributed with mean µIand shape κI. A strand without 435
incubation (SI or SIR) can be described by setting µE≈0 and 1
κE≈0 and a strand 436
without recovery (SI or SEI) can be described by setting 1
µI≈0 and 1
κI≈0. Further 437
details are in the code repository within the Safe Blues GitHub repository [3]. 438
We developed a web-based interactive simulation dashboard; see Fig 3 (top right) 439
and [48]. The interactive dashboard has a variety of control features that enable a user
440
to set the model and its parameters. We used the simulation to focus on exploring 441
parameters for a strand’s infection strength σand maximal infection distance ρ. In 442
particular, we explored σin the range [0, 0.1] and ρin the range [0, 20], and fixed the 443
remaining parameters at a specific values. We chose
π
= 0
.
1,
µE
= 24 hours (or a single
444
day), κE= 5, µI= 168 hours (or a single week), κI= 5. 445
The simulation indicated that with a population of 50 or more participants, some 446
level of Safe Blues spread is possible. Further we simulated epidemics on population 447
sizes 100, 200, and 500. Based on 1000 simulation runs of the model, we observed that,
448
1. there existed a minimum value for ρfor sustained transmission, which decreased 449
as population size increased, and 450
March 29, 2022 13/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
2. there existed a region of transitional parameters, which narrowed as population 451
size increased. 452
These observations allowed us to conclude that
σ∈
[0
,
0
.
05] and
ρ∈
[10
,
20] were the
453
ideal parameter ranges for a population size of 100,
σ∈
[0
,
0
.
05] and
ρ∈
[5
,
15] the ideal
454
parameter ranges for a population size of 200, and σ∈[0,0.04] and ρ∈[2,12] the 455
optimal parameter ranges for a population size of 500. These observations then guided
456
our initial strand parameter choices. 457
Virtual social distancing 458
Apart from using Safe Blues as a tool for collecting data on virtual epidemic spread, we
459
also tested it as a means to explore how ‘virtual social distancing’ affects these 460
epidemics. Our goal in implementing ‘virtual social distancing’ was to provide a rich 461
dataset which could be used by researchers or public health bodies to explore future 462
intervention strategies through social distancing. 463
In order to implement the ‘virtual social distancing’, we tweaked the measured 464
(observed) distance, d, in Eq (2) by a ‘social distancing factor’. Thus, if the measured 465
distance was 4m and the social distancing factor was 1.5, then the distance, d, for 466
infection computation was 4 ∗1.5 = 6m. See the Results and discussion section below 467
for results of this testing mechanism. 468
Results and discussion 469
Calibration of the maximal infection distance parameter 470
At the start of the experiment we used the parameter ranges determined from the 471
campus simulation study as an initial guide for choosing strand parameters,
σ
and
ρ
as
472
in
(2)
. Our initial purpose was to find dynamic ranges of both the infection strength,
σ
,
473
and maximal infection distance,
ρ
, that affect the spread of strands. Initial results from
474
the campus experiment immediately confirmed the effect of the maximal infection 475
distance, while the effect of infection strength was not apparent in the dynamic range of
476
values that we used. We then further explored the range for the maximal infection 477
distance parameter using the strands released in batches
1.05
and
1.06
, during phase 1
478
trials. In particular, we experimented with the maximal infection distance parameter 479
within the range [7.5,500] while choosing the strength parameter within the range 480
[0.1, .24]. 481
In batch 1.05, we released 30 SI type strands with the strength parameter fixed at 482
0.16 for all strands and the maximal infection distance parameter chosen from the set 483
{7.5, 15, 30, 60, 120, 500}. We observed that, in general, there were 3 ranges for the 484
distance parameter that produced distinct epidemics. Specifically, the epidemics 485
established when the maximal infection distance parameter was greater than 30. 486
Increasing the maximal distance parameter beyond 120 did not necessarily produce 487
more severe epidemics. Further, the epidemics did not propagate for most of the strands
488
when the maximal distance parameter was less than 30. With this observation in mind,
489
we fine tuned the search grid for the maximal distance parameter using the strands 490
released in batch
1.06
. We released 90 SI type strands in that batch, with the maximal
491
distance parameter chosen from the set {20, 26, 34, 44, 57, 74, 97, 125, 163, 212}and 492
the strength parameter chosen from the set {0.1, 0.16, 0.24}.493
The plots in Fig 7 depict the effect of varying the maximal infection distance 494
parameter on the propagation of epidemics. The left side plot shows the daily infection
495
trends of strands in batch
1.06
over 5 days categorized by 3 maximal infection distance
496
ranges. The right side plot displays the difference in infection count between the 1st day
497
March 29, 2022 14/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
and 5th day for the strands in both batches as the distance parameter varies. In both 498
these plots we ignored the effect of infection strength since its confounding effect was 499
negligible. 500
We fitted a two parameter scaled, and shifted sigmoidal curve to the data, plotted in
501
red. The curve clearly indicates an upward infection effect of the maximal infection 502
distance with saturation at distances over 80 meters, and less than 30 meters. This is 503
not unexpected based on our design of the infection formula (2), yet we initially found 504
the magnitude of the distances puzzling. One may expect Bluetooth transmission to be
505
effective at distances that are significantly shorter. Towards that end, we believe the 506
observed distance,
d
in
(2)
, may be skewed in our app measurements which are based on
507
averaging of RSSI Bluetooth signal strengths. Such bias between the actual distance of
508
devices and the observed distance, d, may be further investigated via direct phone to 509
phone measurements of the app. We have yet to carry out such measurements to 510
completion, but initial tests indicated of a mismatch of the order of 30 meters, meaning
511
that phones that are xmeters apart perceive a distance in the order of d=x+ 30.
Fig 7. Infection trends over 5 days for strands in batch 1.06, categorized by 3
maximal infection distance ranges (left). Effect of varying maximal infection distance on
the difference of infection (day 5 - day 1) for strands in batches 1.05 and 1.06 (right).
The red curve on the right side plot is a fitted sigmoid function.
512
Herd immunity 513
Herd immunity occurs when a significant proportion of a population become immune to
514
an infectious disease through either vaccination or previous infection, making the 515
disease unlikely to spread within the population. The ‘herd immunity threshold’ is the 516
minimum proportion of the population that must be immune in order to achieve herd 517
immunity. In the simplest SIR model, this quantity can be calculated as 1 −1/R0,518
where R0is the basic reproduction number of the disease [53]. The basic reproduction 519
number is the expected number of secondary infections caused by a single infectious 520
person in an otherwise susceptible population [54, 55]. The basic reproduction number 521
for the delta variant of COVID-19 is estimated as 5.1 [56], and thus the herd immunity
522
threshold for the delta variant is approximately 80%. 523
In our experiment, we observed the herd immunity phenomenon from some of the 524
strand’s epidemics. For example, Fig 8 displays the epidemics of two SEIR type Safe 525
Blues strands with different mean infection periods. It is evident from the two plots 526
that about 80% of the participating population recovered and about 20% remained 527
susceptible after the disease died off, depicting the herd immunity phenomenon. 528
March 29, 2022 15/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Fig 8. Epidemic trajectories of an SEIR type Safe Blues strand with mean infection
period 10 days (left), and 5 days (right). The remaining strand parameters are common
to both strands.
Testing the virtual social distancing mechanism 529
Social distancing is intended to increase spatial separation. It could be implemented by
530
putting a lower bound on dor scaling dup. As mentioned previously, we implemented 531
virtual social distancing by multiplying the observed distance parameter,
d
, in Eq 2 by a
532
given social distancing factor. We considered 4 social distancing factors; 1 (no social 533
distancing) 1.25 (low), 1.5 (medium), and 3.0 (high). We tested virtual social distancing
534
for the strands released in batch 1.07, starting from the 3rd day of their release. This 535
batch comprised 60 SI type strands. 536
We compared each strand’s infection counts on the days prior to and after virtual 537
social distancing was imposed. Fig 9 displays the effect of social distancing and the 538
maximal infection distance on infection counts of three strands. Each data point is 539
comprised of number of infections one day before the virtual lockdown and number of 540
infections one day after the virtual lockdown. We saw three distinct patterns for the 541
counts one day prior to implementing virtual social distancing. When maximum 542
distance was 40, the counts were less than 2, when it was 60, the counts were between 2
543
and 6, and when it was above 80, the counts had similar values. This distinction was 544
not visible for the infection counts after imposing virtual social distancing. 545
In terms of the effect of the social distance factor, we can observe that infections on
546
the day after the virtual lockdown were, in general, lower for the high (3.0 SD) and 547
medium (1.25 SD) factors. This result highlights the fact that our initial trials on 548
virtual social distancing had an impact in reducing the severity of the epidemics. Higher
549
participant numbers and using more strands would probably strengthen this calculation.
550
We intend to explore virtual social distancing in our future trials once the experiment 551
restarts in 2022. 552
The effect of an actual lockdown on the experiment 553
In the previous section we highlighted that we were able to observe reduction of strand
554
infection counts based on artificially imposed social distancing. However, we were able 555
to observe the same phenomena after the actual lockdown that occurred in New 556
Zealand during phase 2 trial of our experiment. 557
The actual lockdown in Auckland, took place on August 17, 2021 at the time when 558
we released batch 2.01 strands, and the lockdown was later extended until the end of 559
2021. There were 600 strands in total in this batch. These were the first set of 560
experimental strands that were released after we determined the maximal infection 561
distance parameter ranges and tested virtual social distancing. The campus was shut 562
down due to the lockdown, and the number of attendees on campus immediately 563
March 29, 2022 16/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Fig 9. Infection counts of three Safe blues strands on the day prior to and after implementing
virtual social distancing, categorized by their maximal infection distance. Each point is the
centroid of the triangle formed from the infection counts of the three strands. Social Distancing
(SD) is categorized as; 1.0 (no SD), 1.25 (low SD), 1.5 (medium SD), and 3.0 (high SD).
dropped. Consequently, we saw an immediate reduction in the number of exposed 564
participants, and within weeks the number of infected participants reduced to zero. 565
Thus, our data showcased the effectiveness of the actual lockdown. 566
In Fig 10 we depict the infection trajectories of all strands in batch 2.01,567
categorized into their model type (SEIR, SIR, SEI, SI). For all model types we can see a
568
clear effect of the actual lockdown on their strand’s trajectories. As expected, for both
569
the SEIR and SIR models (top two plots in Fig 10), the infection trends gradually 570
reduced or remained close to zero within weeks after the lockdown, and for both the SEI
571
and SI models (bottom two plots in Fig 10), the infection trends stabilized. We also see
572
that the effect of the lockdown was immediate, showcasing the real-time nature of Safe
573
Blues information. 574
Conclusion 575
We have described the design of a Safe Blues experiment in Auckland, New Zealand. 576
The experiment was interrupted by a lockdown in August 2021, so we have extended 577
the experiment to include two more phases once the second semester begins in June 578
2022. The partial data we collected while calibrating the experiment in early phases, 579
especially after the lockdown, suggests that Safe Blues data will be a valuable tool in 580
the fight against pandemics. In particular, we saw the effect of the lockdown 581
immediately in the strand data, even though in reality the effect of the lockdown would
582
not be observed for several days, since infection counts lag true infections by the 583
incubation period together with the time it takes to get tested and recorded the test 584
result which can be several days. The value of Safe Blues real-time data is even greater
585
in the presence of under-reporting of cases, which arises in the presence of 586
asymptomatic infections, or skepticism or ignorance of the value of reporting or where 587
the efficiency of other methods (contact-tracing, wastewater monitoring or even PCR 588
with group testing) is reduced when prevalence is high. 589
Our focus in this paper was on the design of the experiment. We discussed the 590
March 29, 2022 17/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Fig 10. Infection trajectories for all strands in batch
2.01
, categorized into SEIR (top left),
SIR (top right), SEI (bottom left), and SI (bottom right).
databases, strand management, measures taken to ensure participant privacy, and 591
participation incentives that are essential in such an effort. A simulation tool was useful
592
in initially calibrating strand parameters, but data from initial phases became more 593
valuable than simulation for full calibration. In the early stages of the experiment we 594
were able to approximately model the effect of social-distancing mandates and the 595
results suggest the potential for the full experiment to showcase the potential 596
effectiveness of such measures, as did the data from the true lockdown. As part of the 597
experiment we have built a number of visualizations and these, along with all necessary
598
source code, are available on an open repository. We look forward to providing data 599
from Phases 4 and 5 of the experiment once these are complete. 600
Supporting information 601
S1 Appendix. Prize draw rules. Details of Prize draw rules. 602
S2 Appendix. App software. Details of the Safe Blues app software 603
S3 Appendix. ADS and PMS servers. The ADS & PMS technical details 604
S4 Appendix. Interpolation and imputation algorithms. Details of the 605
interpolation and imputation algorithms, 606
S5 Appendix. Strand details. Details of strands and their purposes 607
S6 Appendix. Data structure. Structure of Safe Blues dataset 608
March 29, 2022 18/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
References
1. Dandekar R, Henderson SG, Jansen HM, McDonald J, Moka S, Nazarathy Y,
et al. Safe Blues: The case for virtual safe virus spread in the long-term fight
against epidemics. Patterns. 2021;2(3):100220. doi:10.1016/j.patter.2021.100220.
2. Dandekar RA, Henderson SG, Jansen M, Moka S, Nazarathy Y, Rackauckas C,
et al.. Safe Blues: a method for estimation and control in the fight against
COVID-19; medRxiv:20090258v1 [preprint]. 2020 [cited 2021 December 20].
Available from:
https://www.medrxiv.org/content/10.1101/2020.05.04.20090258v1.
3.
Safe Blues. Safe Blues GitHub organization; 2021 May [cited 21 December 2021].
Available from: https://github.com/SafeBlues.
4.
Safe Blues. Safe Blues Data; 2021 July [cited 21 December 2021]. Available from:
https://safeblues.org/data.
5. WHO. WHO Coronavirus (COVID-19) Dashboard; 2020 April [cited 21
December 2021]. Available from: https://covid19.who.int.
6. Dong E, Du H, Gardner L. An interactive web-based dashboard to track
COVID-19 in real time. The Lancet, Infectious Diseases. 2020;20(5):553–534.
doi:10.1016/S1473-3099(20)30120-1.
7. nCOV2019. ncov2019 epidemiological data; 2020 January [cited 21 December
2021]. Available from: https://ncov2019.live.
8. Influenzanet. Influenzanet; 2009 January [cited 21 December 2021]. Available
from: http://influenzanet.info.
9.
FluTraking. The surveillance system for monitoring influenza; 2006 June [cited 21
December 2021]. Available from: https://info.flutracking.net.
10. Boston Children’s Hospital. Outbreaks Near Me; 2020 January [cited 21
December 2021]. Available from: https://outbreaksnearme.org/us/en-US.
11. CoronaSurveys. Estimating global Covid-19 trend; 2020 January [cited 21
December 2021]. Available from: https://coronasurveys.org.
12. Astley CM, Tuli G, Mc Cord KA, Cohn EL, Rader B, Varrelman TJ, et al.
Global monitoring of the impact of the COVID-19 pandemic through online
surveys sampled from the Facebook user base. Proceedings of the National
Academy of Sciences. 2021;118(51):e2111455118. doi:10.1073/pnas.2111455118.
13. Chan AT, Drew DA, Nguyen LH, Joshi AD, Ma W, Guo CG, et al. The
COronavirus Pandemic Epidemiology (COPE) consortium: a call to action.
Cancer Epidemiology and Prevention Biomarkers. 2020;29(7):1283–1289.
doi:10.1158/1055-9965.EPI-20-0606.
14. Swinburne University Technology. Beat COVID–19 Now; 2020 April [cited 21
December 2021]. Available from: https://www.swinburne.edu.au/research/
global-health-equity/our-research-projects/beatcovid19now.
15. Freifeld CC, Mandl KD, Reis BY, Brownstein JS. HealthMap: global infectious
disease monitoring through automated classification and visualization of Internet
media reports. Journal of the American Medical Informatics Association.
2008;15(2):150–157. doi:10.1197/jamia.M2544.
March 29, 2022 19/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
16.
HealthMap. Novel Coronavirus 2019-nCoV (interactive map); 2020 January [cited
21 December 2021]. Available from: https://www.healthmap.org/covid-19.
17. BlueDot. BlueDot Products: AI-driven infectious disease surveillance; 2013
January [cited 21 December 2021]. Available from: https://bluedot.global.
18. Metabiota. The Epidemic Tracker; 2009 January [cited 21 December 2021].
Available from: https://www.metabiota.com/epidemic-tracker.
19. Randazzo W, Truchado P, Cuevas-Ferrando E, Sim´on P, Allende A, S´anchez G.
SARS-CoV-2 RNA in wastewater anticipated COVID-19 occurrence in a low
prevalence area. Water Research. 2020;181:115942.
doi:10.1016/j.watres.2020.115942.
20. Robert Koch Institut. Corona-Dadatenspende; 2020 April [cited 21 December
2021]. Available from: https://corona-datenspende.de/science/en.
21. Quer G, Radin JM, Gadaleta M, Baca-Motes K, Ariniello L, Ramos E, et al.
Wearable sensor data and self-reported symptoms for COVID-19 detection.
Nature Medicine. 2021;27(1):73–77. doi:10.1038/s41591-020-1123-x.
22. Shandhi MMH, Cho PJ, Hseih J, Kalodzitsa A, Lu X, Wang WK, et al.
CovIdentify: Using commercial wearable devices and smartphones to detect and
monitor COVID–19. Presented at IEEE-EMBS International Conference on
Biomedical and Health Informatics (BHI’21). Athens, Greece;.
23. Chamberlain SD, Singh I, Ariza CA, Daitch AL, Philips PB, Dalziel BD.
Real-time detection of COVID-19 epicenters within the United States using a
network of smart thermometers; medRxiv:20039909v1 [preprint]. 2020 [cited 2021
December 21]. Available from:
https://www.medrxiv.org/content/10.1101/2020.04.06.20039909v1.
24.
Smarr BL, Aschbacher K, Fisher SM, Chowdhary A, Dilchert S, Puldon K, et al.
Feasibility of continuous fever monitoring using wearable devices. Scientific
Reports. 2020;10(1):1–11. doi:10.1038/s41598-020-78355-6.
25. Miller DJ, Capodilupo JV, Lastella M, Sargent C, Roach GD, Lee VH, et al.
Analyzing changes in respiratory rate to predict the risk of COVID-19 infection.
PLoS One. 2020;15(12):e0243693. doi:10.1371/journal.pone.0243693.
26. The National Health Service. The NHS COVID-19 app; 2020 January [cited 21
December 2021]. Available from: https://www.nhs.uk/nhs-app.
27. Singapore Government Agency Website. Trace Together; 2020 April [cited 21
December 2021]. Available from: https://www.tracetogether.gov.sg.
28. Apple Inc, Google Inc. Privacy-preserving contact tracing; 2020 April [cited 21
December 2021]. Available from:
https://www.apple.com/covid19/contacttracing.
29.
Australian Government. Protect yourself and the community; 2020 April [cited 21
December 2021]. Available from: https://www.covidsafe.gov.au.
30. Boulos MNK, Geraghty EM. Geographical tracking and mapping of coronavirus
disease COVID-19/severe acute respiratory syndrome coronavirus 2
(SARS-CoV-2) epidemic and associated events around the world: how 21st
century GIS technologies are supporting the global fight against outbreaks and
epidemics. International Journal of Health Geographics. 2020;19:8.
doi:10.1186/s12942-020-00202-8.
March 29, 2022 20/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
31.
Zhou C, Su F, Pei T, Zhang A, Du Y, Luo B, et al. COVID-19: challenges to GIS
with big data. Geography and Sustainability. 2020;1(1):77–87.
doi:10.1016/j.geosus.2020.03.005.
32. Jhunjhunwala A. Role of telecom network to manage COVID-19 in India:
Aarogya Setu. Transactions of the Indian National Academy of Engineering.
2020;5:157–161. doi:10.1007/s41403-020-00109-7.
33. State of Israel Ministry of Health. HaMagen: the Ministry of Health app for
fighting the COVID-19 outbreak; 2020 May [cited 21 December 2021]. Available
from:
https://govextra.gov.il/ministry-of-health/hamagen-app/download-en.
34.
White L, Van Basshuysen P. Without a trace: Why did corona apps fail? Journal
of Medical Ethics. 2021;47(12):e83–e83. doi:10.1136/medethics-2020-107061.
35.
Garrett PM, White JP, Lewandowsky S, Kashima Y, Perfors A, Little DR, et al.
The acceptability and uptake of smartphone tracking for COVID-19 in Australia.
Plos One. 2021;16(1):e0244827. doi:10.1371/journal.pone.0244827.
36.
Queensland Government. Check In Qld; 2021 February [cited 21 December 2021].
Available from: https://www.covid19.qld.gov.au/check-in-qld/
about-the-app-and-when-to-use-it.
37. Yoneki E, Crowcroft J. EpiMap: Towards quantifying contact networks for
understanding epidemiology in developing countries. Ad Hoc Networks.
2014;13:83–93. doi:10.1016/j.adhoc.2012.06.003.
38.
University of Cambridge. FluPhone Project: Understanding Spread of Infectious
Disease and Behavioural Responses; 2009 January [cited 21 December 2021].
Available from: https:
//www.cl.cam.ac.uk/research/srg/netos/projects/archive/fluphone2.
39.
Yoneki E. Fluphone study: Virtual disease spread using haggle. In: Proceedings
of the 6th ACM Workshop on Challenged Networks; 2011. p. 65–66.
40. Firth JA, Hellewell J, Klepac P, Kissler SM, Kucharski AJ, Spurgin LG, et al..
Combining fine-scale social contact data with epidemic modelling reveals
interactions between contact tracing, quarantine, testing and physical distancing
for controlling COVID-19; medRxiv:26.20113720v2 [preprint]. 2020 [cited 2021
December 21]. Available from:
https://www.medrxiv.org/content/10.1101/2020.05.26.20113720v2.
41. Moritz S, Gottschick C, Horn J, Popp M, Langer S, Klee B, et al. The risk of
indoor sports and culture events for the transmission of COVID-19. Nature
Communications. 2021;12(1):1–9. doi:10.1038/s41467-021-25317-9.
42.
Revollo B, Blanco I, Soler P, Toro J, Izquierdo-Useros N, Puig J, et al. Same-day
SARS-CoV-2 antigen test screening in an indoor mass-gathering live music event:
a randomised controlled trial. The Lancet Infectious Diseases.
2021;21(10):1365–72. doi:10.1016/S1473-3099(21)00268-1.
43. Fieldlab Events program. Fieldlab Events; 2021 February [cited 21 December
2021]. Available from: https://fieldlabevenementen.nl/fieldlab-englis.
March 29, 2022 21/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
44.
Department for Digital, Culture, Media & Sport. Events Research Program; 2021
April [cited 21 December 2021]. Available from:
https://www.gov.uk/government/publications/
events-research-programme-phases-i-ii-and-iii-data-release.
45. Kim SK, Kim EO, Kim SH, Jung J. Universal screening of severe acute
respiratory syndrome coronavirus 2 with polymerase chain reaction testing after
rally of trainee doctors. Journal of Korean Medical Science. 2020;35(42):e380.
doi:10.3346/jkms.2020.35.e380.
46.
Hagemann G, Hu C, Al Hassani N, Kahil N. Infographic. Successful hosting of a
mass sporting event during the COVID-19 pandemic. British Journal of Sports
Medicine. 2021;55(10):570–571. doi:10.1136/bjsports-2020-103511.
47.
Safe Blues. Safe Blues App; 2020 May [cited 21 December 2021]. Available from:
https://play.google.com/store/apps/details?id=org.safeblues.mobile.
48. Safe Blues. Safe Blues Simulation; 2021 November [cited 21 December 2021].
Available from: https://simulation.safeblues.org.
49. Safe Blues. Safe Blues Dashboard; 2021 November [cited 21 December 2021].
Available from: https://analytics.safeblues.org.
50.
Safe Blues. Safe Blues homepage; 2020 May [cited 21 December 2021]. Available
from: https://safeblues.org.
51. Petry NM, Martin B, Cooney JL, Kranzler HR. Give them prizes and they will
come: Contingency management for treatment of alcohol dependence. Journal of
Consulting and Clinical Psychology. 2000;68(2):250.
doi:10.1037/0022-006X.68.2.250.
52. Petry NM, Peirce JM, Stitzer ML, Blaine J, Roll JM, Cohen A, et al. Effect of
prize-based incentives on outcomes in stimulant abusers in outpatient
psychosocial treatment programs: a national drug abuse treatment clinical trials
network study. Archives of General Psychiatry. 2005;62(10):1148–1156.
doi:10.1001/archpsyc.62.10.1148.
53. Fine P, Eames K, Heymann DL. ”Herd immunity“: a rough guide. Clinical
infectious diseases. 2011;52(7):911–916.
54. Anderson RM, May RM. Infectious diseases of humans: dynamics and control.
Oxford University Press; 1991.
55. Kermack WO, McKendrick AG. A contribution to the mathematical theory of
epidemics. Proceedings of the Royal Society of London Series A, Containing
Papers of a Mathematical and Physical Character. 1927;115(772):700–721.
56.
Liu Y, Rockl¨ov J. The reproductive number of the Delta variant of SARS-CoV-2
is far higher compared to the ancestral SARS-CoV-2 virus. Journal of Travel
Medicine. 2021;28(7):1–3. doi:10.1093/jtm/taab124.
March 29, 2022 22/22
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Appendix 1: Prize draw rules
As participants run the Safe Blues app within the geofence area (See Fig 5), they collect hours which increase
their chance of winning prizes. This is done by accumulating campus hours which are then translated into eligible
hours via the rules described below. During Phase 1 we used a simpler set of rules, and introduced the ‘invite
a friend’ mechanism from Phase 2 onward intending to increase participant numbers. See also the prizes web
page, https://safeblues.org/prizes/.
This is an overview of the rules:
•For every hour running the app on campus, one campus hour is collected.
•Campus hours are accumulated since joining, or since the start of the current phase (campus hours are
not collected during paused periods (see Table 1).
•A participant may collect up to 200 campus hours per phase.
•A maximum of 10 campus hours per day can be collected.
•Campus hours are converted to eligible hours as follows:
–Each of the first 20 campus hours counts for 2 eligible hours.
–After the first 20 campus hours, each additional campus hour counts as 1 eligible hour.
•Participants may obtain more eligible hours via the ‘invite a friend’ mechanism.
–After accumulating 20 campus hours, participants may invite up to 10 friends to join the experiment
and thereby receive bonus eligible hours.
–For every friend they invite, after the friend collected 20 campus hours, the inviting participant
receives 5 additional eligible hours.
–The mechanism for inviting a friend is by generating a 6 digit invite code and asking the friend to
enter that code.
–Invited friends will receive 5 eligible hours when they sign up.
•Each phase with the exception of phase 3, has a prize draw at the end of the phase. Additionally, in place
of the phase 3 prize draw, which was voided due to lockdown, there is a special prize draw (see below)
•In each prize draw (for phases 1, 2, 4, and 5), prizes are drawn as follows:
–The chance of winning a prize is based on the eligible hours divided by the total number of eligible
hours of all participants.
–There are 9 prizes in each draw. A single top prize (iPad Pro), 3 second-tier prizes (Android mobile
phones), and 5 third-tier prizes (FitBit tracker).
–A participant may win at most one prize in a draw and this works as follows. First, everyone competes
for the top prize and the winner is removed from the pool. Then all remaining participants compete
for the second-tier prizes, each time removing the winner. Similarly for the third-tier.
–A participant is eligible to win at most a single prize from each tier over the course of the experiment.
For instance, a winner of the top prize in the first draw will be excluded from winning the top prize
in draws that follow (but may still win other prizes).
•Participants can track both their collected campus hours and eligible hours per phase and see how many
hours they have collected relative to the distribution of hours collected by other participants (see Fig 3,
bottom left plot for a snapshot of the participants leader board showing this distribution).
•All participants are emailed a full report from the prize draw and winners collect directly from experiment
staff.
1
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
•Winners are asked if they wish to have their picture and short bio posted. This is optional.
•There is a special prize draw in place of phase 3 during the period between phase 3 and phase 4. In this
prize draw 10 participants have a chance to win a Fitbit tracker. The prize draw is performed by uniformly
selecting 10 winners among all those participants who have accumulated 5 or more campus hours during
the period of phase 2.
2
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Appendix 2: App software
The Safe Blues Android app is based on the Trace Together Android App (Open Trace) [1] and was forked
from the Open Trace Android version in April, 2020. The Open Trace software is designed for contact tracing,
however we modified it to support the Safe Blues protocols. The app is written in the Kotlin language and has
been made available on Google Play since March 2021.
The app is now specifically tailored for the Safe Blues campus experiment and includes on–boarding screens
that assign the app instance a unique 10 digit ID. There is no email or other authentication information queried
in the app. With the exception of the basic on–boarding screens, there is no user engagement in the app as
it runs in the background. The app requires the user to enable location services and Bluetooth. Bluetooth is
clearly needed for strand propagation. Location services are not directly part of the Safe Blues system but are
needed to recognize that participants are in the geofenced area for the purpose of allocating prizes.
The initial phase of the experiment (see Table 1 in Appendix 6) included the release of several versions of
the app that were automatically updated by participants. The sequence of these versions fixed several initial
bugs. The most notable bug was a non-random initialization of the random number generator in the app,
which caused all participating phones to seed Safe Blues strands in unison. Specifically, in strand batch 1.01
all participating phones decided together whether to seed a strand or not. Once this bug was fixed, a new
app version was deployed. This deployment process included the strand batch 1.02 which did not include any
strands per-se.
Due to limitations on control over the underlying Bluetooth hardware, the app transfers information between
two phones by pairing them together for a brief period of time. We call each such Bluetooth interaction a “ping”.
Each time two phones ping each other, they transmit their set of infectious strands as well as the strength of their
transmitter. While the app is running, it continuously tries to ping other phones that are part of the experiment,
including phones that have been pinged recently. These pings are then combined together into longer Safe Blues
sessions (capped at 30 minutes) after which the Safe Blues simulation step runs. The original rationale for the
session concept was to allow more flexibility in setting the infection mechanics of the simulation: our system
could, for instance, be used to investigate the effect of a perfect contact tracing regime in tandem with these
virtual pandemics (where a contact is traced if they spend a minimum duration at a minimum distance with
a contact). Furthermore, the underlying Bluetooth systems are far from ideal for this, exhibiting congestion
in large groups of participants and being unreliable even when there are few participants. The session system
accommodates for some of this unreliability while also giving more accurate estimates of distance.
References
[1] BlueTrace. OpenTrace; 2020 April [cited 21 December 2021]. Available from: https://github.com/opent
race-community.
1
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Appendix 3: ADS and PMS servers technical details
The Safe Blues experiment system separates the anonymous data server (ADS) and the participant manage-
ment system (PMS). The ADS is part of the operational Safe Blues system and is used to push new strands to
the client apps and record phone strand information through anonymized client IDs generated within the app.
The PMS is a system designed specifically for the experiment and is primarily aimed at participant management
and prizes. It records how long the app was running in the background while the participant was within the
experiment geofence. The PMS accomplishes this by using a separate “experiment ID” that is Completely
isolated from the client ID in the ADS. Both systems run in parallel on the phone: the ADS is part of the
experiment, while the PMS simply checks whether the participant is on campus with the required permissions
and Bluetooth turned on, and relays summary counts to the PMS server.
We chose to physically separate these two servers in order to provide an extra layer of privacy protection for
participants as well as to enable future Safe Blues deployments (perhaps operational) to use the ADS while not
the PMS. In this separation, the PMS is hosted on a Nectar Research Cloud server managed by The University
of Auckland, whilst the ADS is hosted on the commercial Amazon Web Services (AWS) Cloud.
The ADS uses a PostgreSQL database and exposes a gRPC API for phones indicating strand (and virtual
social distancing) information, as well as a RESTful admin API. Phones running the app send messages to the
ADS every 15 minutes, informing it of the status of the strands. The following are the protocol buffer messages
about strands:
message Strand {
string name = 13;
int64 strand_id = 1;
google.protobuf.Timestamp start_time = 2;
google.protobuf.Timestamp end_time = 3;
double seeding_probability = 4;
// the two parameters of the infection probability map
double infection_probability_map_p = 5; // strength
double infection_probability_map_k = 6; // radius
double infection_probability_map_l = 7; // unused
// mean and shape of the gamma distribution for incubation period
double incubation_period_mean_sec = 8;
double incubation_period_shape = 9;
// mean and shape of the gamma distribution for infectious period
double infectious_period_mean_sec = 10;
double infectious_period_shape = 11;
uint32 minimum_app_version = 12;
}
Messages from phones to the ADS are encoded as follows:
message InfectionReport {
string client_id = 1;
int32 version_code = 5;
repeated int64 current_incubating_strands = 2;
repeated int64 current_infected_strands = 3;
repeated int64 current_removed_strands = 4;
}
All such incoming messages are stored on the ADS Postgres database. Phones are only identified by a
temporary 256-bit client ID that changes every 24 hours. Appendix 4 describes the algorithm running on the
ADS for interpolation and imputation. Note that the ADS is not aware of the 10 digit participant ID which
was only created for the purposes of the experiment.
1
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
The PMS stores a list of email addresses associated with each participant ID but does not store any further
personal participant information (that is, we do not keep home addresses, names, afflictions, or other private
information). The PMS receives messages from the phones via a restful JSON API indicating time spent
on campus. An example of such a message is below (where duration and count active are in units of 15
minute intervals, and the truncated entry time is the UNIX timestamp of the entry into campus. As can be
seen, these messages indicate the time on campus. The phone generates such a message whenever leaving the
geofenced area.
{
"participant_id": 1234567890,
"version_code": 60,
"statuses": [
{
"status_id": 112,
"truncated_entry_time": 18731,
"duration": 12,
"count_active": 11
},
{
"status_id": 113,
"truncated_entry_time": 18944,
"duration": 17,
"count_active": 17
},
...
]
}
As shown in the bottom left image of Fig 3, the PMS aggregates these messages in a MySQL database,
which is then queried for prize information and for presenting users with their current leader-board standings.
The PMS also acts as a web server for a React-based website used for experiment registration, the ‘invite a
friend’ mechanism, and the aforementioned leader-board.
Some specific PMS information is made available in accordance with the ethics approval. This is the total
number of participants on campus and the total number of registered participants, as shown in Fig 6 (left plot),
as well as summary of the distribution of daily campus hours of participants using the mean and a 5-number
summary (see the right plot in Fig 6).
2
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Appendix 4: Interpolation and imputation algorithm
The epidemiological status of strands is stored in the ADS as a collection of reports received from participants’
smartphones. The interpolation and imputation problem is to use these reports to reconstruct each participant’s
infection status at arbitrary times throughout the entire day. Specifically, the way in which the experimental
data is collected and stored results in two issues that must be resolved: (a) a participant’s status before any
reports were received and (b) a participant’s status during the time between reports.
Here, in describing the method used to solve the aforementioned problems, we fix a 24-hour UTC interval,
an active strand, and a 256-bit ID. Note that, because a participant’s anonymity 256-bit ID changes every
24 hours, our procedure cannot make use of any information before this window. We collect from the ADS
database the following information: the time at which the participant sends their first report, TF; the time at
which the participant first reported a state of “exposed”, TE; the time at which the participant first reported
a state of “infected”, TI; and the time at which the participant first reported a state of “recovered”, TR. We
choose TR=∞if a “recovered” report is never received, TI=TRif an “infected” report is never received, and
TE=TIif an “exposed” report is never received. Moreover, given any T∈ {TE, TI, TR}with T=TF, we set
T= 0. The participant’s state at an arbitrary time tduring the day is given by
“susceptible”, t < TE,
“expected”, TE≤t<TI,
“infected”, TI≤t < TR,
“recovered”, TR≤t.
(1)
1
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Appendix 5: Strand details
Here we overview the main purpose of the batches of experiment strands at different stages of the experiment.
Specific strands in each batch can be identified from the strand id parameter listed in the strand.csv file in
the Safe Blues Experiment Data repository (see also Appendix 6). Table 1 details the total number of strands
released for each batch during phases 1–3 of the experiment, as well as the strand IDs associated with them.
Phase Batch No. of
strands
Strand IDs Type Main Purpose
00.01 50 1–50 SEIR, SIR, SEI The debug version, which is the first
test of system in a workshop involving
10 users.
11.01 162 51–212 SEIR, SEI, SI First time testing the system on the
experiment population. A bug (“same
seed”) was detected.
11.02 1 213 SI To test the system after fixing the
“same seed” bug.
11.03 10 214–223 SEI After the “same seed” bug, this was an
intermediate debug batch while waiting
for Google Play app update.
11.04 162 224–385 SEIR, SEI, SI Same structure of strands as 1.01 re-
leased after fixing the “same seed” bug.
11.05 30 386–415 SI Experimentation of the effect of
strength and maximal infection dis-
tance on infection via SI strands.
11.06 90 416–505 SI Further experimentation of the effect of
maximal infection distance on infection
via SI strands.
11.07 60 506–565 SI Testing the virtual social distancing
feature.
22.01 600 566-1165 SEIR, SIR, SEI, SI Main release of all types of strands to
begin Phase 2 of the experiment. These
strands were planned to live until the
end of Phase 3.
22.02 –2.06 600
(per
batch)
1166–4023 SEIR, SIR, SEI, SI Weekly releases identical to 2.01 dur-
ing the first month of the lockdown.
33.01 –3.22 6 (per
batch)
4024–4155 SI As Auckland is still in lockdown, releas-
ing 6 strands per week until the planned
start of Phase 4 in case of any changes
in behavior over this period
4 & 5 - - - - Strands for phases 4 and 5 will be sim-
ilar to 2.01, currently planned with a
release of new strands every two weeks
during those phases.
Table 1: Batches of strands and their main purpose in the campus experiment.
We now overview the batches of Table 1 starting with the debug batch, 0.01, and up to the Phase 3 final
batch, 3.22. Batch 0.01 or the debug batch was used as an initial test of the app prior to deploying it to
the experimental population. This test trial involved 10 people who attended a Julia language meetup at The
University of Queensland on April 21, 2021. This batch contains 50 strands in total, with a combination of SEIR,
SIR, and SEI types. Following this, batch 1.01 was the first experimental test of the system as part of The
University of Auckland experiment. Immediately after release of this batch, we observed that all participating
1
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
devices were initialized with the same random number generator seed. This bug, named as the “same seed”,
caused all phones to make the same decision about seeding a given strand or not. The bug was fixed and a single
SI type strand was released as batch 1.02 to test this. While waiting for a Google Play update for the new app
version, strands from batch 1.03 were released as an intermediate step for further testing. These were 10 SEI
type strands. Once the bulk of the experiment participants updated their app to the new version, batch 1.04
was released with the same composition of strands initially intended in 1.01. This batch, 1.04, constitutes the
main experimental batch within Phase 1.
The remaining batches of Phase 1, 1.05 –1.07, were aimed at further calibration of the strength and
maximal infection distance parameters (as described in Eq (2)), as well as testing the virtual social distancing
feature (see Materials and methods section in the main document). Specifically, 1.05 tested a range of distance
and strength parameters and their effects on virtual viral spread, and 1.06 refined the search grid in the maximal
distance parameters. We were able to see a clear effect of the maximal infection distance parameter on viral
spread in both cases (see Fig 9). However, we were unable to detect a significant effect of the strength parameter
within this parameter search. We released strands in batch 1.07 at the beginning of the week (Monday morning
New Zealand time), and by Wednesday we had inflicted various levels of virtual social distancing, with factors
ranging from 1.25 to 3.0. Measurements from this batch clearly indicated that the virtual social distancing
mechanism has a significant impact on strand transmission (see Fig 9).
Moving onto the batches of phases 2 and Phase 3, the major batch providing data to date is 2.01. This
batch was released one week into Phase 2 (on Thursday, July 29) and included an extensive variety of strand
types based on experience gained in Phase 1. Three weeks after the release, the major New Zealand lockdown
took place and this immediately affected strand propagation (see Fig 10 presenting all strands from this batch).
With anticipation of the lockdown potentially lifting, we released batches 2.02 –2.06 weekly where each such
batch contains the same types of strands as 2.01. These batches have not yielded meaningful infections due to
the continued lockdown. Finally, throughout Phase 3, and the interim period between Phase 3 and Phase 4,
we released the weekly batches 3.01 –3.22. These batches, each with only 6 SI strands, are intended to “keep
alive” the Safe Blues system.
2
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Appendix 6: Data structure
The Safe Blues dataset is organized in the data repository as shown in Fig 1. The strand parameters
determine the epidemiological behavior of each virtual virus–like token, including its transmissibility, incubation
duration, and infection duration. A table of these parameters is stored in data/strands.csv. This table
contains a row for each strand circulated during the campus experiment and contains columns described in
Table 1. The transmission data provides aggregate measurements of the spread of strands throughout their
strands
strand1.csv
strand2.csv
...
data
daily
strands
strand1.csv
strand2.csv
...
participants.csv
hourly
participants.csv
strands.csv
Figure 1: Safe Blues Data Structure.
reporting periods. We store the progression of the ith strand (strand id = i) as a daily and hourly time
series in strand(i).csv file. These tables have a row for each time point (either daily or hourly) and have
the columns described in Table 2. The participant data presents aggregate information on the engagement of
participants throughout the course of the experiment. This is also available as either a daily or hourly time
series in participants.csv files. Again, these tables contain a row for each time point (either daily or hourly)
and contain the columns described in Table 3.
1
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Name Description
strand id The unique numerical identifier given to each strand
batch The release batch or group of strands within which each strand
model The epidemiological model used by each strand (either SI, SEI,
SIR, or SEIR)
start utc The time at which each strand’s reporting begins in UTC timezone
start nzt The time at which each strand’s reporting begins in NZST/NZDT
timezone
seed utc The time at which each strand’s initial infections occur in UTC
timezone
seed nzt The time at which each strand’s initial infections occur in
NZST/NZDT timezone
stop utc The time at which each strand’s reporting ends in UTC timezone
stop nzt The time at which each strand’s reporting ends in NZST/NZDT
timezone
initial A participant’s probability of initial infection for each strand
strength The strength of transmission for each strand
radius The maximal infection distance of transmission for each strand
incubation mean The mean parameter for each strand’s gamma distributed incu-
bation duration
incubation shape The shape parameter for each strand’s gamma distributed incu-
bation duration
infection mean The mean parameters of each strand’s gamma distributed infec-
tion duration
infection shape The shape parameters of each strand’s gamma distributed infec-
tion duration
Table 1: Description of the columns in strands.csv. The incubation mean and incubation shape values
are missing when the model is either SI or SIR type. The infection mean and infection shape values are
missing when the model is either SI or SEI type.
Name Description
strand id The unique numerical identifier given to each strand
time utc The time of each measurement in UTC timezone
time nzt The time of each measurement in NZST/NZDT timezone
susceptible The current number of participants who are susceptible to the strand
exposed The current number of participants who are incubating the strand
infected The current number of participants who are infected with the strand
recovered The current number of participants who have recovered from the strand
distance factor The artificial distance multiplier used to emulate social distancing
Table 2: Description of the columns in the strand i.csv file, where iis the ith strand. The values for exposed
column are missing when the model is either SI or SIR type. The values for recovered column are missing
when the model is either SI or SEI type.
2
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
Name Description
time utc The time of each measurement in UTC timezone
time nzt The time of each measurement in NZST/NZDT timezone
count campus The number of participants on campus during the current day
count reporting The number of participants whose phones are sending reports
during the current UTC day
count registered The number of participants who have registered for the experiment
by the current day
hours mean The mean number of campus hours
hours min The minimum number of campus hours collected
hours q1 The lower quartile number of campus hours collected
hours q2 The median number of campus hours collected
hours q3 The upper quartile number of campus hours collected
hours max The maximum number of campus hours collected
Table 3: Description of the columns in the participants.csv files. All statistics reported in the two
participants.csv files are in reference to NZST/NZDT days except for count reporting, which is in reference
to UTC days. The mean and 5 number summary correspond to the campus hours collected by participants
who attended campus on the current day. Values for the mean and 5 number summary are missing when
count campus is less than or equal to 5.
3
. CC-BY-NC-ND 4.0 International licenseIt is made available under a
is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 1, 2022. ; https://doi.org/10.1101/2022.03.31.22273262doi: medRxiv preprint
