Modelling SARS-CoV-2 transmission in a UK university setting

Around 40% of school leavers in the UK attend university and individual universities generally host thousands of students each academic year. Bringing together these student communities during the COVID-19 pandemic may require strong interventions to control transmission. Prior modelling analyses of SARS-CoV-2 transmission within universities using compartmental modelling approaches suggest that outbreaks are almost inevitable. We constructed a network-based model to capture the interactions of a student population in different settings (housing, social and study). For a single academic term of a representative campus-based university, we ran a susceptible–latent–infectious–recovered type epidemic process, parameterised according to available estimates for SARS-CoV-2. We investigated the impact of: adherence to (or effectiveness of) isolation and test and trace measures; room isolation of symptomatic students; and supplementary mass testing. With all adhering to test, trace and isolation measures, we found that 22% (7%–41%) of the student population could be infected during the autumn term, compared to 69% (56%–76%) when assuming zero adherence to such measures. Irrespective of the adherence to isolation measures, on average a higher proportion of students resident on-campus became infected compared to students resident off-campus. Room isolation generated minimal benefits. Regular mass testing, together with high adherence to isolation and test and trace measures, could substantially reduce the proportion infected during the term compared to having no testing. Our findings suggest SARS-CoV-2 may readily transmit in a university setting if there is limited adherence to nonpharmaceutical interventions and/or there are delays in receiving test results. Following isolation guidance and effective contact tracing curbed transmission and reduced the expected time an adhering student would spend in isolation.


Methods:
We constructed a network-based model to capture the interactions of a student population in different settings (housing, social and study). For a representative campus-based university, we ran a susceptible-latent-infectious-recovered type epidemic process, parameterised according to available estimates for SARS-CoV-2. Over the course of a single academic term, we investigated the impact on infection control of adherence to (or effectiveness of) isolation, test and trace measures, the additional use of room isolation as an intervention and supplementary mass testing.
Results: Incorporating uncertainty in the fraction of cases that are asymptomatic and their associated infectivity, in the absence of interventions our model estimated that 16% (2% -38%) of the student population could be infected during the autumn term. In contrast, with full adherence to isolation measures and engagement with test-and-trace, predictions of the cumulative infection count were lower, 1.4% (0.4% -5%). Irrespective of the adherence to isolation measures, on average a higher proportion of students resident on-campus became infected compared with students resident off-campus. Widespread adherence of interventions led to reductions in the average fraction of time those individuals adhering to measures were expected to be isolated, with room isolation as an additional intervention generating minimal benefits. The model found that a one-off instance of mass testing would not drastically reduce the term-long case load or end-of-term prevalence, but regular weekly or fortnightly testing could reduce both measures by more than 50% (compared to having no mass testing).
Conclusions: Our findings suggest SARS-CoV-2 may readily transmit amongst a student population within a university setting if there is limited adherence to nonpharmaceutical interventions and there are delays present in receiving test results. Following isolation guidance and effective contact tracing both curbed transmission and reduced the expected time an adhering student would spend in isolation. Additionally, widespread adherence throughout the term suppresses the amount of unwitting asymptomatic transmission to family and community members in the students' domicile regions at the end of term.

Introduction 1
Globally, many countries have employed social distancing measures and nonpharmaceutical interven-2 tions (NPIs) to curb the spread of SARS-CoV-2 [1]. In the UK, the enaction of lockdown on 23rd 3 March 2020 saw the closure of workplaces, pubs and restaurants, and the restriction of a range of 4 leisure activities. The education sector was also impacted, with schools closed (with the exception of 5 children of key workers) and higher education establishments, such as universities, delivering the end 6 of the 2019/2020 academic year via online means. 7 In the summer months, the national implementation of strict measures transitioned to a localised 8 approach, targeting regions experiencing the highest burden of transmission. As the number of SARS- 9 CoV-2 confirmed hospitalised cases and deaths began to decline, many sectors of society cautiously 10 reopened, with measures in place to reduce transmission. Universities began to develop plans to 11 reopen, with several adopting a blended learning strategy of limited face-to-face teaching combined 12 with online lectures. Higher education in the UK comprises a sizeable population of students, with 13 over 2.3 million higher education students enrolled in the 2018/2019 academic year across over 160 14 higher education providers [2]. This results in a sizeable movement of students nationwide at the 15 beginning and end of academic terms (in addition to international student travel). The migration 16 of students would contribute to increased population mobility, which had already grown since the 17 easement of lockdown measures occurred [3,4], with an associated need for careful management in 18 order to minimise the risk of seeding outbreaks in low prevalence locations. term in 2020 [17], and a working paper looking at how mathematical approaches may help inform the 47 reopening of higher education spaces to students whilst minimising risk [18]. 48 Many of the previous studies of SARS-CoV-2 outbreaks in a university setting have adopted com-49 partmental modelling approaches, in which individual behaviour and interventions such as contact 50 tracing cannot be readily captured. In this paper, we present an individual-level network-based model 51 framework for transmission of SARS-CoV-2 amongst a student university population, which includes 52 test, trace and isolation interventions. Contacts occur across household, study and social settings, 53 underpinned by empirical data where possible. We find that maintaining strong adherence to isola-54 tion guidance and engagement in test and trace could both curb the amount of infection throughout 55 the academic term and limit SARS-CoV-2 prevalence at the beginning of the winter break. Use of 56 room isolation and a single mass testing instance may offer marginal benefits, though the underlying 57 adherence to interventions remains crucial. These results show the possible impact of SARS-CoV-2 58 transmission intervention measures that may be enacted within a university population during the 59 forthcoming academic year.

61
To enable a modelling analysis of the transmission of SARS-CoV-2 within a university population, we 62 adopted a network approach to capture the interactions between students in different settings, upon 63 which we ran an epidemic process. In this section we provide in some detail: (i) a description of 64 the network model; (ii) the data sources used to parameterise the network contact structure; (iii) the 65 model for SARS-CoV-2 transmission and COVID-19 disease progression; (iv) the simulation protocol 66 employed to assess the scenarios of interest.

67
Network model description 68 We used a multi-layered network model to encapsulate identifiable groupings of contacts. Our model 69 was comprised of four layers: (i) households, (ii) study groups/cohorts, (iii) organised societies and 70 sports clubs, and (iv) dynamic social contacts.

71
Household contact layer 72 In our model we considered contact networks on campus and off campus separately. The network 73 for on-campus accommodation contained a hierarchical structure, from the smallest scale to largest, 74 of a household (typically based around a shared kitchen), floor, block, hall (comprised of multiple 75 blocks). We constructed the on-campus accommodation units to match that of a representative campus 76 based university. We assigned students resident off-campus to households with sizes sampled from an 77 estimated student household size distribution (see Supporting Text S1: Off-campus student household 78 size).

79
Within a household, irrespective of on-campus or off-campus location, we assumed each individual 80 to have the potential to transmit infection to each other person within their household (i.e. a fully 81 connected network). In addition, on any given day, on-campus students could randomly make contact 82 with any other individual outside their direct household, though situated on the same floor or contained 83 within the same accommodation block.
Note that we assumed these cohort contacts occurred between study friendship groups outside of any 90 face-to-face classes. In other words, we have presumed for teaching spaces the enforcement of COVID-91 secure measures sufficiently minimises the transmission risk to prevent the onward spread of infection 92 within that setting. 93 Contacts in organised societies and sports clubs 94 A prominent aspect of the university experience is the presence of societies and sports clubs. For 95 constructing contacts resulting from involvement in such groups, we applied a uniform probability of 96 forming a contact with each other individual in the group, with that probability differing based on 97 whether the group was a society or sports club. As simplifying assumptions, these links did not alter 98 during the course of a simulation and we set each social group to meet three times per week with 99 assigned members attending all sessions. The final contact layer sought to capture random, dynamic, contacts made each day with any other in-102 dividuals in the student population. For each timestep, random connections were selected for each stu-103 dent according to a specified cohort-dependent distribution. We carried out the contact network gen-104 eration using the Erdös-Rényi model [20], which assumes a Poisson distribution degree sequence.

105
Contact parameterisation 106 We characterised the network structure across the various contact layers by applying two differing 107 approaches.

108
The first method was a data-driven approach, using data from the Social Contact Survey [9,21]. 109 The Social Contact Survey was a paper-based and online survey of 5,388 participants in the United 110 Kingdom conducted in 2010. We extracted records provided by 347 students, with a total of 10,302 111 contacts. These data informed the network construction parameters for the cohort and dynamic 112 social contact layers, with stratification according to the student's level of study (undergraduate or 113 postgraduate). We fit parameters for these contact distributions using maximum likelihood estimation 114 via the fitdistrplus package in R.

115
The second method was a subjective approach, used when we did not have relevant empirical mea-116 sures available to enable the parameterisation of the given contact layer. This was applied to the 117 formulation of random contacts within on-campus accommodation blocks (though outside the direct 118 household) and organised social club contacts. We provide a summary of the network parameterisation 119 in Table 1.

121
To estimate the number of contacts occurring with those in the same study cohort, we used the 122 student contact survey records listed as occurring in the work or school setting. We kept entries 123 specifying a duration of 60 minutes or more and that occurred more than once per week, assuming 124 the retained contacts with these characteristics would be reflective of a university study cohort. We 125 independently fit using maximum likelihood estimation lognormal distributions for undergraduates 126 and postgraduates, using a mean and standard deviation parameterisation, acquiring distributions of 127 Lognormal(1.646,1.590) and Lognormal(1.211,1.128), respectively (Fig. 1). For estimating a distribution for dynamic social contacts, we considered contacts reported within the 130 Social Contact Survey occurring in all locations except home and all non-'first-time' contacts. We 131 limited valid contacts to those recorded as either involving touch or lasting longer than 10 minutes. 132 Valid contacts also had to last less than 60 minutes, with a view that longer duration contacts would 133 be captured by the cohort and society contact layers. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

Dynamic social contacts
The copyright holder for this this version posted October 18, 2020. .

138
Given the limited available data to be able to parameterise a degree distribution for those contacts, 139 we took a pragmatic approach and assumed a low constant probability of contacts occurring in the 140 broader accommodation unit. We additionally assumed these contact probabilities lessened for higher 141 levels of accommodation hierarchy. Specifically, we attributed a higher chance of interacting with 142 someone on the same floor (daily chance of contact of 10%) than someone on another floor within 143 the same block (daily chance of contact of 5%). We assumed no random accommodation associated 144 contacts with other students living in different blocks in the same hall.

145
Contacts in organised society and sports club activities 146 We also did not have available the necessary information to parameterise contacts within organised 147 social groups using a data-driven approach. Therefore, we stress that the values stated here are 148 subjective and alternative proposals would add to result variability. 149 We considered a community of 335 organised social groups, comprising 270 societies and 65 sports 150 clubs. In the absence of a membership size distribution, we allowed a breadth of group sizes by ran-151 5 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. . domly assigning each group a membership size of 10 to 100 (in increments of 10). We also subjectively 152 chose a monotonically decreasing probability mass function for the number of organised social groups 153 students actively participated in: 50% of students not in any group; 40% involved in a single group; 154 2.5% a piece in two, three, four or five groups.

155
Following group assignment, within each group we established contacts between students with a fixed 156 probability of each link existing. These probabilities we set at 0.05 for societies and 0.1 for sports 157 clubs. Accordingly, a student was likely to make more contacts in groups with large membership. For 158 simplicity, we retained the same arrangement of contacts for each meeting. We ran a susceptible-latent-infectious-recovered (SEIR) type disease process on the network structure. 162 Once infected, we assumed infectiousness could not start immediately (i.e. on the same day), with the 163 earliest permitted moment being the following day. We assumed an Erlang-distributed incubation 164 period, with shape parameter 6 and scale parameter 0.88 [22].

165
The distribution of infectiousness had a four day pre-symptomatic phase, followed by a seven day 166 symptomatic phase. This gave a total of 11 days of infectivity and a minimum 12 day infection duration 167 (for the full temporal profile, see Table 2). It was based on a Gamma(97.2,0.2689) distribution, with 168 shape and scale parameterisation, shifted by 25.6 days [23,24]. Following completion of the infectious 169 period, the individual entered the recovered state. Infected individuals could be either asymptomatic or symptomatic, with an ascribed probability deter-172 mining the chance of each individual being asymptomatic. There remains uncertainty in the fraction of 173 COVID-19 cases that are asymptomatic and how that statistic may vary with age, however community 174 surveillance studies have been performed to help diminish this uncertainty. The REal-time Assessment 175 of Community Transmission-1 (REACT-1) study found approximately 70% of swab-positive adults and 176 80% of swab-positive children were asymptomatic at the time of swab and in the week prior [25]. To 177 6 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. . reflect the uncertainty in this value, in each replicate simulation we sampled the asymptomatic case 178 probability from a uniform distribution within the interval 0.6 and 0.8.

179
There is currently limited data available to provide a robust quantitative estimate of the relative 180 infectiousness of asymptomatic and symptomatic individuals infected with SARS-CoV-2, though there 181 are some indications that asymptomatic individuals could be considered to be less infectious than 182 symptomatic individuals [26]. Therefore, we set an asymptomatic individual to have a lower risk of 183 transmitting infection compared to a symptomatic individual, with the current uncertainty reflected 184 by sampling the value for the relative infectiousness of an asymptomatic in each simulation replicate 185 from a Uniform(0.4,0.7) distribution. We applied the scaling consistently throughout the duration of 186 infectiousness for asymptomatics, meaning there was no time dependence on the scaling term over the 187 course of infectiousness.

188
Setting transmission risk 189 Attributing risk of transmission to any particular contact in a particular setting is complex. This 190 is partly due to the huge heterogeneity in contact types, and partly due to the different scales of 191 data: contact information is by its nature individual-based, whereas transmission rates are generally 192 measured at the population level. Therefore, whilst we can attribute a relative risk to each contact 193 type (home, social, study), there is an arbitrary scaling to translate these relative risks to an absolute 194 growth rate of infection in the population.

195
For household transmission, we attributed a household secondary attack rate to each individual based 196 on their household size. We sampled from a normal distribution whose mean value depended on the 197 household size, based on estimates of adjusted household secondary attack rates from a UK based 198 surveillance study [27]. The mean values used were: 0.48 for a household size of two, 0.40 for for a 199 household size of three, 0.33 for a household size of four, 0.22 for a household size of five or above. 200 The standard deviation of the normal distribution for households of size two or three was 0.06, and 201 for households of four or above was 0.05.

202
For transmission risk in other settings, we performed a mapping from the Social Contact Survey [9] 203 to obtain a relative transmission risk compared to the central estimate of adjusted secondary attack 204 rate in the household setting against those aged 18-34 of 0.34 [27] (further details in Supporting 205 Text S3: Parameterisation of contact risk). The relative magnitude of those means when compared 206 to the household transmission risk were used to scale the standard deviation. Transmission risks 207 were consistent across all non-household settings, with the exception being in student societies where 208 we assigned a lower transmission risk to reflect the implementation of COVID-secure measures that 209 would be required to permit these meetings to take place. We also reiterate that we attributed zero 210 transmission risk to face-to-face study.

211
To calibrate the relative transmission risks to achieve an uncontrolled reproductive number, R t , in the 212 expected range of 2 − 4, we apply a universal scaling of 0.4 to all of the above rates (see Supporting 213 Text S4: Non-intervention scenario calibration). Upon symptom onset, students adhering to guidance enter isolation for ten days. At that moment, 217 fellow household members of the symptomatic case that adhere to guidance enter self-isolation for 14 218 days [28]. Students that are symptomatic and that will engage with the test and trace process take a 219 test upon symptom onset. We included a two day delay before receiving the test result.

220
Once isolation periods are begun they are seen out in full unless the test result was negative (false 221 negative probability of 0.13 [29]). On occasions where a negative result was given, household members 222 7 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. . Students identified as a contact of a confirmed case, and that were adhering to self-isolation guidance, 228 spent 14 days in self-isolation [30]. The modelled tracing scheme looked up contacts for an index case 229 up to two days before onset of symptoms. We assumed that the probability of an individual being 230 able to recall their 'dynamic' contacts diminishes with time, from 0.5 one day previously, reducing 231 in increments of 0.1, such that the probability of successfully tracing a contact five days prior to 232 the tracing occurring is 0.1. Once again, other assumptions could be explored and a wider range 233 of assumptions, collectively, would generate more variation in the results. We give an overview of 234 isolation, test and trace related parameters in Table 3. We used this model framework to evaluate the transmission dynamics of SARS-CoV-2 amongst a 237 university student population during the autumn term of the 2020/2021 academic year, and the 238 potential impact of both adherence to the guidance and additional interventions. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

235
The copyright holder for this this version posted October 18, 2020. . https://doi.org/10.1101/2020.10.15.20208454 doi: medRxiv preprint the length of welcome week plus the ten week academic autumn term. 242 We seeded the number of latent, asymptomatic and recovered individuals based on UK regional preva-243 lence estimates for 26th September 2020 and student flow data (we provide further methodological 244 details in Supporting Text S5: Initial seeding of infected and recovered individuals). We assumed there 245 were no symptomatic infected (ill) students present at the beginning of each simulation replicate. 246 Our assessment comprised of three strands. First, we analysed how the strength of adherence to 247 guidance on isolation and engagement with test and trace affected case burden and accumulated 248 isolation time (total number of student isolation days over the term). Second, we considered adoption 249 of a policy of strict room isolation for on-campus residents displaying COVID-like symptoms. Third, 250 we analysed a collection of scenarios involving mass testing of students to study the impact on overall 251 case load, the expected time spent in isolation per adhering student and the prevalence of infection at 252 the conclusion of the autumn academic term. 253 We outline each of the three assessments in further detail below. Unless stated otherwise, for each 254 parameter configuration we ran 1,000 simulations, amalgamating 50 batches of 20 replicates; each 255 batch of 20 replicates was obtained using a distinct network realisation. We performed the model 256 simulations in Julia v1.5.

257
Adherence to isolation, test and trace 258 We sampled adherence to isolation from zero compliance (value 0) to full compliance (value 1) in 259 increments of 0.1. We assumed an identical adherence to isolation restrictions independent of the 260 cause (presence of symptoms, household member displaying symptoms, identified as a close contact 261 of an infected by contact tracing). Additionally, we assumed those that would engage with isolation 262 measures would also engage with testing and tracing.

264
For those resident in on-campus accommodation and suffering from COVID-like symptoms, another 265 applicable intervention may be to bolster household isolation by mandating quarantine of those indi-266 viduals in en-suite rooms (with meals and essentials delivered). Those residing in accommodation with 267 communal bathrooms would be re-housed. Those already living in en-suite accommodation would be 268 isolated in their rooms. 269 We modelled this intervention by assuming those rehoused or put into room isolation had no contacts. 270 We applied these measures on the same timestep the student reported being symptomatic.

271
Mass testing 272 We explored altering the timing and frequency mass testing was carried out: a single instance on day 273 21 (end of week 2 of the academic term); a single instance on day 63 (end of week 8 of the academic 274 term); regular mass testing every two weeks ('fortnightly', on day 1, then day 14, day 28,. . ., day 70); 275 regular mass testing on a weekly basis (on day 1, then day 7, day 14,. . ., day 70).

276
Additionally, we varied coverage amongst the eligible student population: all students, on-campus 277 resident students only, off-campus resident students only. We carried out sensitivity to the underlying 278 adherence to isolation measures by performing the analysis for adherence probabilities of 0.2 (low), 279 0.5 (moderate) and 0.8 (high), respectively. 280 We did not include with the mass testing procedure students that had previously reported infection 281 and then subsequently received a positive test. We also assumed that all tests were performed on 282 9 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. . the same day, those latently infected (thus not yet infectious) would always return a negative test 283 result, and contact tracing was performed rapidly such that those contacts who were both traceable 284 and adhered to isolation guidance were isolated from the next timestep.

286
Trade-offs between case load and isolation 287 With high adherence to isolation measures and engagement with test and trace, a lower number of 288 infections (the sum of both identified cases and undiagnosed infections) would be expected to arise 289 during the course of the autumn term (Table S3). Specifically, with no interventions we estimated 290 a median proportion of 0.16 (95% prediction interval: 0.02-0.38) of the entire student population to 291 be infected during the autumn term. In contrast, with full adherence the median proportion infected 292 was 0.014 (95% prediction interval: 0.004-0.05). Irrespective of the adherence level, we found that (on 293 average) a higher proportion of students resident on-campus would become infected versus students 294 resident off-campus ( Fig. 3(a)).

295
In addition to strong adherence leading to suppression of case numbers, it also delivered benefits 296 from the perspective of time spent in isolation. Out of those students that would adhere to isolation 297 guidance, those resident on-campus were expected to spend a greater proportion of time in isolation 298 compared to adhering students living off-campus. Further, we witnessed greater variability in the 299 predicted time an adhering individual spent in isolation for those resident on-campus versus those 300 resident off-campus ( Fig. 3(b), Table S3).

301
Inspecting temporal patterns of infection prevalence and proportion of the student community in isola-302 tion demonstrated trade-offs between case numbers and the requirement for portions of the population 303 to isolate (Fig. 4). Averaging across all simulation replicates, zero adherence led to a growing epidemic 304 throughout the course of the term, although no one enters isolation at any time. With half of the 305 student population adhering to control measures, we observed a slower growth of the epidemic concur-306 rent with a steady rise in the proportion of students entering isolation. Under complete adherence the 307 expected prevalence of infection was kept low, although this generated an initial surge in the amount 308 of students isolating which declined slowly from day 20. Ultimately, with full adherence the number of 309 students expected to be isolated at the end of term was less than if only half of the student community 310 adhered to isolation and tracing guidance.

311
The effect of strong adherence in suppressing case numbers is reflected in the reduction in the amount 312 of tests carried out on those that are infected with SARS-CoV-2, compared to moderate adherence 313 (Fig. 5(a)). Median estimates for the maximum proportion of students isolated at any given time 314 plateaued at about 3% once the fraction of the population adhering surpassed 0.4 ( Fig. 5(b)). There-315 fore, for high levels of adherence, the substantial decline in cases just compensates for the greater 316 number of individuals who adhere to the rules and isolate as a consequence of the cases identified.

317
Prevalence at the end of term may be as high as 15% of the population under no interventions (median 318 6%), whilst for full adherence it is unlikely to be above 2% (Fig. 5(c), Table S4). Furthermore, the 319 results display potential for there being a sizeable number of non-symptomatic infecteds (comprising 320 latent, asymptomatic and presymptomatic) at the end of the term (Fig. 5(d)). At low adherence, 321 central estimates of the proportion of the population being in a non-symptomatic infected state at the 322 end of the term exceeded 5%. For strong adherence, central estimates of this measure were below 1%, 323 substantially reducing the risk posed by students returning home for the winter break. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. .  Outputs summarised from 1,000 simulations (with 20 runs per network, for 50 network realisations) for various levels of adherence to NPIs. Over the duration of the autumn term, distributions relative to students resident on-campus only (green violin plots), students resident off-campus only (orange violin plots) and to the overall student population (purple violin plots) for (a) proportion infected, and (b) proportion of time adhering students spend in isolation. The white markers denote medians and solid black lines span the 25th to 75th percentiles. For percentile summary statistics, see Table S4. Maintenance of nonpharmaceutical interventions and effective contact tracing curbed transmission, with the expected time an adhering student would spend in isolation also reduced. On-campus residents were more likely to become infected and spend a greater proportion of time in isolation compared to students living off-campus.

11
. CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020.  Table S4. 13 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. . https://doi.org/10.1101/2020. 10.15.20208454 doi: medRxiv preprint Use of room isolation 325 Isolation rooms, where on-campus students reporting symptoms and testing positive are rehoused to 326 prevent further transmission to housemates, has been postulated as an additional measure of social 327 distancing and control. For all tested adherence levels, we found median estimates for the number of 328 on-campus SARS-CoV-2 infected students being rehoused were below 40. In addition, the majority 329 of simulations returned counts less than 100 (Fig. 6(a)). For low adherence, despite the large number 330 of symptomatic cases present in the student population, under-reporting results in fewer rehousing 331 instances compared to situations where there is stronger adherence to NPIs. Similarly to observations 332 for tests used and cumulative isolation time, with high adherence the curbing of spread of infection 333 results in a reduction in the expected number of SARS-CoV-2 infected on-campus students being 334 rehoused.

335
When including room isolation as an extra intervention measure, in addition to social distancing, 336 isolation guidance and contact tracing, we observe a slight reduction in the median for estimated case 337 load over the autumn term ( Fig. 6(b)). There was also a reduced probability of a large number of 338 cases over the course of the autumn term. The number of tests administered on COVID-19 infected 339 individuals was generally lower (Fig. 6(c)).

340
When considering the maximum number of students isolated at a single time and cumulative isolation 341 days, central estimates and variability were also reduced when mandating room isolation for students 342 experiencing COVID-like symptoms who were resident on-campus (Figs. 6(d) and 6(e)). is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint Our investigation of the utility of mass testing during the course of the academic term involved three 345 variables: (i) the timing and frequency of mass testing that was carried out: a single instance on day 346 21 (end of week 2 of the academic term); a single instance on day 63 (end of week 8 of the academic 347 term); regular mass testing on a fortnightly basis (on day 1, then day 14, day 28,. . ., day 70); or 348 regular mass testing on a weekly basis (on day 1, then day 7, day 14,. . ., day 70); (ii) the coverage 349 amongst the eligible student population: all students; on-campus resident students only; or off-campus 350 resident students only; (iii) underlying adherence to isolation (tested values of 0.2 (low) 0.5 (moderate) 351 and 0.8 (high), respectively). These were compared relative to a baseline scenario that had identical 352 parameters with the exception of no mass testing being performed (Fig. 7).

353
Under the considered mass testing options, more intense testing (greater coverage) and earlier testing 354 on average led to a smaller outbreak. However, intense testing resulted in a greater amount of time 355 spent in isolation for those students adhering to isolation measures (Fig. 7). These relationships remain 356 contingent on the proportion of students arriving infected.

357
In more detail, for the single instance mass testing options and for the high adherence setting, an 358 early mass test (day 21) covering all students resulted in the lowest relative median estimate (0.83) 359 for the proportion infected and caused only a minor increase in the time each adhering student would 360 be estimated to be isolated (1.05, see Fig. 7(a) and Table S5). In contrast, a late date mass test (day 361 63) of all students, combined with a low adherence to isolation measures, led to only a minor drop 362 in the proportion infected but a near doubling (1.81) in the time spent in isolation for those students 363 that do adhere to isolation measures ( Fig. 7(b), Table S5).

364
Given low adherence circumstances, regular testing (either weekly or fortnightly) amplified the charac-365 teristic that adhering students are likely to be in isolation for a greater portion of term-time (Figs. 7(c) 366 and 7(d), Table S5). In particular, having weekly mass testing compared to no mass testing led to 367 a doubling in the expected median time that adhering students spent in isolation, and more than 368 trebling in the case of all students being mass tested each round. In a similar manner to the one-off 369 mass test strategies, we found testing that covered all students (combined with high adherence to 370 other control measures) returned the lowest relative median estimate (0.57 and 0.43 for fortnightly 371 and weekly testing, respectively) in the proportion of the student population infected over the course 372 of the academic term. Targeting testing at students resident off-campus only was comparably ineffec-373 tive; the median estimates for proportion of students infected during the term was comparable to a 374 scenario with no supplementary mass testing (proportional size of median estimates across adherence 375 settings ranged from 0.90-0.99). 376 We stress that across these findings there was a large amount of variability in the proportion of infected 377 students at the end of term (Figs. 8(a) to 8(c), Table S6). A single mass test event encompassing all 378 students slightly reduced the median expected prevalence at the conclusion of the academic term, with 379 greater reductions under a fortnightly or weekly mass testing programme (Figs. 8(d) to 8(f), Table 380 S7).

381
Across all considered mass test strategies, we saw the greatest reductions in median outcomes when 382 mass testing rounds covered all students and adherence to other intervention measures was high 383 (Fig. 8(f)). The top-performing strategy of weekly mass testing of all students and high adherence 384 to other control measures resulted in an end of term prevalence of just 8% of the no-mass testing 385 prevalence. On the other hand, with only a single mass test instance covering off-campus resident 386 students, there was no noticeable impact on the expected proportion of students infected at the end 387 of term. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. . sures, having regular mass testing increased the likelihood of there being fewer infected students at 390 the end of term compared to the baseline scenario (Fig. S7, Table S7). Mass testing was either not used (baseline scenario), a single instance took place at the end of week two or week eight of the academic term (corresponding to simulation day numbers 21 and 63, respectively), or regular mass testing was performed on a fortnightly or weekly basis. We present in each panel outputs from 1000 simulations for mass testing covering all eligible students (red), on-campus only (blue), off-campus only (grey). The left hand side of each panel corresponds to the relative proportion (compared to the baseline scenario) of the student population infected over the duration of the autumn academic term under low, moderate and high adherence. In a similar way, the right hand side of each panel presents data on the relative time adhering students spend in isolation. The one-off mass testing was performed on: (a) day 21; (b) day 63. Frequent mass testing was performed (c) fortnightly; (d) Weekly. Full estimates are given in Table S5. 17 . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020.  Tables S6-S7. On average, end of term prevalence was minimised with more frequent testing and if mass testing rounds included students resident on-campus.

18
. CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. In this paper, we have described the construction and application of a network model to characterise 393 the transmission of SARS-CoV-2 amongst a student population in a UK campus-based university. Our 394 findings suggest SARS-CoV-2 could readily transmit amongst a student population within a univer-395 sity setting over the course of a single academic term. Maintaining nonpharmaceutical interventions 396 and effective contact tracing curbed transmission, while also reducing the expected time an adhering 397 student would spend in isolation.

398
Our findings demonstrate the efficacy of isolation and tracing measures in controlling the spread of 399 SARS-CoV-2 if they are broadly adhered to. Reducing the quantity and riskiness of contacts breaks 400 chains of transmission, with the projected potency in the use of nonpharmaceutical interventions to 401 control spread of SARS-CoV-2 at a national scale previously documented [31][32][33].

402
Irrespective of adherence probability, we predicted that a higher proportion of the on-campus popu-403 lation would typically be infected compared to those living off-campus. In general, household sizes 404 within on-campus halls of residence are larger than those living in households off-campus. As a con-405 sequence, a higher level of mixing is expected, with an associated increased risk of infection; halls 406 of residence have been identified as environments conducive to the transmission of other respiratory 407 illnesses [34]. This outcome reinforces the importance of monitoring the situation in halls of residence, 408 in agreement with prior studies [16]. 409 We also analysed what impact separating on-campus residents who were confirmed infected from 410 household members (for the duration of the infected individual's isolation period) could have as a 411 potential extra barrier to disease spread. Though we saw marginal improvements compared to not 412 including the intervention, practicalities of the strategy and the outlay on required resources may 413 prohibit it as an implementable option. In particular, there would need to be the spare housing 414 capacity with suitable facilities to accommodate those confirmed infecteds that are living in households 415 with communal bathrooms, and a safe way of moving infectious individuals to the new rooms.

416
Whilst we found the absolute impacts of running a single mass test across the student population 417 on a range of epidemiological measures to usually be small, it does illustrate the importance the 418 stated objective can have on what is ascertained as the preferred strategy. Based on our modelling 419 framework, if one was looking to minimise the proportion of students infected, then earlier testing with 420 large coverage would be selected. However, an additional concern is the potential risks of asymptomatic 421 students returning home for the winter break and unwittingly spreading infection to their domiciled 422 community. Given an objective of minimising the prevalence of infection at the end of term, then 423 performing the mass test later in the term would be preferable.

424
Our findings from computational simulations of frequent mass testing strategies are in agreement with 425 prior modelling works indicating that mass testing of students would need to take place at regular 426 intervals, such as fortnightly or weekly, to suppress SARS-CoV-2 transmission [11,16]. There have 427 been calls that, before universities allow students to return home, community transmission must first 428 be curbed and frequent testing subsequently provided [35]. As an additional aid to help track and 429 monitor the spread of COVID-19 in their student and staff communities, several universities have set 430 up public-facing data dashboards in both the USA [36] and the UK [37].

431
Where possible, we have taken a data-driven approach to parameterise the system and instruct hetero-432 geneities we expect to be present, such as in student contact patterns. Nevertheless, this work has made 433 simplifying assumptions and our results therefore have limitations. Student numbers and estimates of 434 regional movements between term-time and out-of-term time addresses were taken from pre-pandemic 435 academic years; these movements may not accurately reflect the situation for the 2020/2021 academic 436 year during the COVID-19 pandemic. Additionally, we assumed there would be no students beginning 437

19
. CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. . term with COVID-like symptoms, and there was no transmission to students from the wider commu-438 nity. Relaxation of either of these assumptions is likely to generate a larger outbreak throughout the 439 term.

440
When constructing the contact networks, for simplicity we assumed each student maintained consistent 441 contacts throughout the entire term with others in their household, and selected others from their 442 cohort and the organised societies and sports clubs they were members of. While the assumption 443 for households may reasonably hold, given shared use of communal spaces, one would expect less 444 rigidity in the study and organised social group related contacts. We also used a fixed distribution 445 for drawing random daily social contacts throughout the term, whereas in reality it may be expected 446 the distribution of such contacts to vary temporally. A set of distributions could instead be used 447 to capture these temporal heterogeneities, were the necessary data available to initially discern the 448 amount of time periods warranting a distinct distribution and then subsequently parameterise each 449 distribution. Finally, the level of transmission through this network is contingent on the behaviour 450 of students and their compliance with social distancing measures. We have assumed an uncontrolled 451 reproduction number in the range 2-4 (dependent on the proportion of students that are asymptomatic 452 and the relative transmission rate from asymptomatic infections); unfortunately, the precise value can 453 only be estimated once students return and any emerging outbreak can be measured. In the event of 454 student populations at universities suffering outbreaks, there is scope for the network model framework 455 presented here to be used for real-time parameter estimation. Larger values of R are likely to result 456 in a higher number of cases and greater pressure being exerted on test and trace services earlier in the 457 term.

458
Multiple refinements of the model structure are still possible and may yield a better understanding 459 of the outbreak impact on the broader university community. We have not included university staff 460 members, or infection to and from the local community. Students with asymptomatic infection inter-461 acting with elder individuals in non-COVID secure environments may result in silent transmission of 462 SARS-CoV-2 into more vulnerable groups at risk of severe outcomes from COVID-19. Similarly, given 463 the observed rise with age of the likelihood of severe health outcomes due to COVID-19 disease [7], in 464 the event of widespread community transmission staff and surrounding communities would be likely 465 to experience higher levels of morbidity than students. Another aspect we have not included here is 466 the presence of other respiratory infections. Such an extension would permit the study of test capacity 467 requirements when levels of cough and fever are high due to non-COVID-19 causes, especially of con-468 cern in the winter period; were such a scenario to arise it would apply significant stress to the national 469 test and trace system [38]. 470 In the context of the COVID-19 pandemic the movement of students to attend universities, creating 471 large communities of predominately young adults, poses specific challenges in controlling transmission. 472 Infectious disease models may be a useful part of the public health decision-making process: deter-473 mining the most appropriate interventions to be applied in a university setting. Our work highlights 474 a network modelling approach to capture heterogeneities in contact structure that are particular to 475 the university student population and its projected impact on transmission of SARS-CoV-2. This 476 model suggests that encouraging student adherence with test-trace-and-isolate rules (as well as good 477 social-distancing, mask-use and hygiene practices) is likely to lead to the greatest reduction in cases 478 both during and at the end of term; mass testing is also found to produce strong benefits in terms of 479 reducing infection, but generally leads to a greater number of cases being found and isolated. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. . https://doi.org/10.1101/2020. 10.15.20208454 doi: medRxiv preprint

Supporting information items
Supporting Text S1: Off-campus student household size Estimation of an off-campus student household size distribution from the Social Contact Survey data.

Supporting Text S2: Cohort data
Outline of data underpinning the partitioning of the student population into different study cohorts.

Supporting Text S3: Parameterisation of contact risk
Summary of the use of contact survey data to estimate the relative transmission risk across contacts between students occurring in household versus non-household settings.

Supporting Text S4: Non-intervention scenario calibration
Overview of simulation outputs in the absence of any interventions.

Supporting Text S5: Initial seeding of infected and recovered individuals
Description of the method applied to estimate the expected number of students in each COVID-19 disease state at the beginning of the 2020/2021 academic year.

Additional tables
Summary statistics to support the displayed distribution plots.

Additional figures
Supplementary mass testing scenario results.

25
. CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. . https://doi.org/10.1101/2020. 10.15.20208454 doi: medRxiv preprint