The IT management facility at Georgia Tech (GT) accumulates WiFi access point logs over time. This is common in most universities with managed WiFi infrastructure. We actively collaborated with IT management to define safety and security safeguards that allow us to obtain a de-identified version of these raw logs. Before accessing the data we established a data-use agreement and an ethics protocol that was approved by the Institutional Review Board (IRB) at Georgia Institute of Technology (Protocol H20208). For the WiFi data, we were provided access to logs from Fall 2019 and Fall 2020. We processed these logs to characterize mobility (WiMob) and it encompasses all 40,000 unique visitors that connected to the network via 6,959 different access points (41). The logs did not contain any personally identifiable information and locations are also coded. The logs indicated the WiFi access point (AP) a device associates with and can therefore be used to infer dwelling locations of users across the entire campus. This is limited to indoor spaces where APs are located and the scope of this localization is at the granularity of a room or suite (39,43)). For En we only used aggregate insights for enrollment, which were derived from course registration transcripts. Note, we did not cross-identify any students. We used publicly accessible course schedules to approximate schedules of de-identified nodes and infer if they were students or staff, and non-residential or residential. We elaborate on our data in Supplementary Material, Data.

Note. Like most universities, GT’s managed WiFi network is not equipped with any Real-Time Location System (RTLS) (46,47). RTLS systems use Received Signal Strength Indicator (RSSI) values from multiple neighboring APs to provide high precise localization of individuals in terms of time and space. However, deploying such systems requires surveying the entire network. Additionally, precision localization raises more privacy concerns. These factors together make it challenging for universities to justify the deployment of RTLS, unlike small retail settings that can monetize RTLS insights directly (e.g., insights on footfall can be tied to improving revenue).

WiMob leverages the logs to create bipartite graphs Kt, for each day t, which connect P users to L access point locations (Figure 1A). Any edge, {p,l}i indicates the ith instance when a p was dwelling at l. These edges describe the time period of dwelling. Subsequently, by comparing all edges in Kt we can infer if different individuals are collocated near an AP to create a contact network, Gt, for each day t—between any collocated pi,pj∈P. These networks define the contact structure for an epidemiological agent-based model at every time-step. Similarly, by inspecting the sequence of dwelling locations for any p in graph K, we compute a mobility network, Ht—between locations l∈L. In our approach, we considered collocation as a form of proximate contact—people in the same room—and therefore established collocation only when this occurred for “an extended period” (23). By varying this threshold between 30 and 40 minutes we found the contact networks to be structurally similar as their clustering coefficients (over the semester) were highly correlated (r=0.97). In our experiments, we used the 40 minute threshold as it was more computationally less expensive. We provide more details of our approach in Supplementary Material, Data Processing and in Supplementary Material, Modeling Contact and Movement.

We compared the disease outcomes and burdens of 2 policies, Remote Instruction (RI) and Localized Closure (LC), both of which are modeled with WiMob.

Remote Instruction (RI): The status quo for data-driven policies offers strictly online instruction for large class enrollment, while continuing the other classes in person. When using En to model contacts, we implemented RI by removing connections between students who were only in contact through courses of size ≥30. When using WiMob to model contacts, we removed connections between students if they were only connected because of collocations during scheduled lectures of such courses.

To further elaborate, for RI we inferred enrollment size of each course in Fall 2019 by determining the number of unique individuals that visited lecture locations during scheduled times. After the first week, we applied the RI by removing all visiting edges in Kt for any lc∈LRI  if visits were during lecture times of course c with an enrollment ≥30. This helped create counterfactual contact networks Gt′. The removal of edges from K described the mobility budget of RI and the structure of Gt′ indicated the risk of exposure budget. We designed LC with these budgets by identifying locations for closure (LLC ) with different algorithms, such as PageRank (48), Eigenvector Centrality (49), Load Centrality (50), and Betweenness Centrality (51). When a location was closed, we removed all edges in Kt connected to any lx∈LLC . We aggregated the movement graph Ht over a week and apply the algorithms to identify locations. Subsequently, we identified the number of top-ranked locations to remove such that the resultant counterfactual contact network Gt″ has is within 1% of the budget. More details for closure policies have been expanded in Supplementary Material, Identifying Locations for Closure.

To make the comparisons between the closure policies, we established fixed budgets to design LC based on the resource utilization on RI. We considered 2 kinds of budgets, (i) mobility reduction—to depict space use on campus, and (ii) risk of exposure—to reflect testing capacity. Also note, response to closure policies can lead to unpredictable side-effects in campus behavior, particularly when a student’s schedule is entirely online. Therefore, we design policies within three behavioral scenarios (each with a varying budget):

[S1] Persistence: Irrespective of the locations closed or classes restricted, individuals continue their other visiting behaviors.

The budgets varied for different behavioral scenarios and we only compared policies within the same scenario. Similar to other works that model closure (11,52), we assume that when a location is shutdown, the individuals who ought to have visited that location isolated during the time. This is further elaborated in Supplementary Material, Modeling Policies and Scenarios.

We constructed an agent-based model (ABM) that captures the spread of COVID-19 between individuals active on campus. This ABM leveraged the contact networks produced by WiMob. The simulation iterated a time-step each day with the underlying contact networks i.e., Gt for no interventions, Gt′ for RI, and Gt″ for LC. Each agent in our ABM follows a modified version of susceptible–exposed–infectious–removed (SEIR) template that disambiguates the infectious compartment into asymptomatic and symptomatic. New infections were introduced to the model either externally or internally. External transmission arose because individuals could contract the virus outside campus and bring the infection back for local spread (7,53). We adopted data of positive cases from Fulton county (25) with a scaling factor α to estimate the probability that a susceptible individual, who is active on campus, was infected from interactions that take place outside campus. Internal transmissions are determined by p, as the probability of susceptible individuals in contact with an infectious one. We calibrated the parameters related to disease transmission by training and validating our models on the positivity rate reported by GT surveillance testing (24). Supplementary Material, Agent-Based Model details the disease progression and describes the various parameters.

In our study, we estimated three key parameters: (i) infectious individuals at day 0, (ii) transmission probability between infectious and susceptible individuals, and (iii) the probability of infection transmission from contacts outside the network. We estimated the range of optimal parameters for disease transmission by minimizing the root means square error (r.m.s.e) between the Georgia Tech surveillance testing positive rates (24,54) and the observed positivity rate of the model every week—percentage of new asymptomatic cases out of the total testable population. The surveillance testing conducted by Georgia Tech was designed for detecting individuals who contracted Covid-19 without showing Flu-like symptoms within the community (54). We calibrated the model on the positivity rates on the first 5 weeks of Fall 2020. To attain a point estimation of the optimal parameters, we fitted the model to predict trends in the remaining weeks by running a numerical optimization algorithm, Nelder-Mead (55). To account for quantitative uncertainty, we estimated a range of parameters, within 40% of optimum r.m.s.e (30). For other model parameters, we adopted values proposed by previous studies on similar populations (56–58). Table 1 shows a full list of our parameters.

Note that our calibration characterized latent factors associated with pandemic-related cautious behaviors, including the relationship with external transmission. And these factors could be related to “county characteristics, partisanship, media consumption, and racial and ethnic composition” (8). To account for the effect of these varying latent factors on disease outcomes, we performed additional calibrations for hypothetical variations in disease spread. For these analyses we kept the GT mobility behavior constant while calibrating the model on different time periods of surveillance testing and on positivity rates of different U.S. universities—University of Illinois at Urbana-Champaign (59) and University of California (60), Berkeley. We evaluated RI and LC on these variations and describe the design of these complementary experiments in Supplementary Material, Sensitivity Analyses. See Supplementary Material, Calibration for details on the calibration process and results of all variations are in Table S3.

Variation in contact over the semester would naturally impact the severity of disease spread. However, En describes a static network that does not capture such dynamics (Figure 2A). Instead, we found that WiMob shows contacts got sparser over the semester. Figure 2C presents a notable decline in contacts after the first two weeks, which coincides with multiple orientation seminars and the so-called “course shopping” period of Fall 2019. In fact, contact decreased considerably in classrooms, with a steeper slope possibly because of reduction in attendance. WiMob was able to reveal other observable changes, such as drop in contacts during exam period (week 15) and increase after fall recess (week 10). En rendered a highly connected static network, which can miscalculate the speed at which a disease spreads. By contrast, the longitudinal behavior represented by WiMob can help universities anticipate disease spread more accurately.

Campuses can assess risk of an outbreak by characterizing the number of individuals that would be at risk of infection through contact. In our study, En indicated 99% of the individuals on campus were clustered in a single component—if any of them would have been infected in Fall 2019, the entire component would be at risk. From the lens of En a virus can exhaust an entire population with infection very early. However, WiMob showed that only 69% of the population was connected in a single component (Table S2). This difference is because WiMob can distinguish how many individuals are active on campus. Therefore, WiMob provides a pragmatic estimate of risk by grounding it in local occupancy and helps campuses budget for resources better.

Reports suggest that a key contributor to cases in the pandemic is actually clustering of individuals in non-academic spaces (7). However, En does not depict a holistic view of campus contact. It is limited to classrooms and, therefore, fixates on contacts in lectures, while ignoring other spaces. In fact, WiMob showed that in the first 6 weeks of Fall 2019, the shortest path among individuals was smaller than that approximated by En (Figure 2B). With WiMob, we observed new paths in the contact network from situations outside classes. On a given week, WiMob showed the average shortest path with contact is 3.26(±0.5) when only considering lectures, whereas capturing all contexts reduced the average shortest path to 2.67(±0.28). Characterizing shorter pathways is crucial for policymakers as closure policies by design aim to disconnect these pathways.

Prior work uses En to posit that RI reduces contact and in turn significantly fragments the network for disease spread in universities (10,11). We evaluated the effectiveness of such a policy if it were applied in Fall 2019, with both WiMob and En. Figure 2D shows that RI with En reduced contact by 94% and increases shortest path by 50%. However, the same intervention with WiMob showed a relatively milder impact (contact reduction 45%; shortest path increase 11%). This reinforces that contact outside courses are significant and remain unaffected by enrollment-oriented policies like RI. WiMob provides a more encompassing view of the structural effects to a network and motivates design of more impactful closure policies.

Prior works show a few locations are responsible for majority spread (30) and restricting movement between them leads to greater control (62). We found that, if COVID-19 spread through Fall 2019 (a regular semester), the cases rose after 7 days (Figure 3A). Therefore, we applied both RI and LC interventions after the first week. To devise interventions, WiMob estimated how RI uses the budget and then designed LC to match this budget under every behavioral scenario. Table 2 describes how the budget for each policy varies. Additional details are present in Supplementary Material, Modeling Policy and Scenarios.

We present differences between LC and RI based on three infection reduction outcomes; peak infections (maximum active cases on a given day), internal transmission (exposure from infected individuals on campus), and total infections (cumulative cases at the end of the semester). Additionally, we measured the burden of policy interventions with the number of locations closed—requires resources to monitor and maintain super-spreader locations, the percentage of students that avoid campus—disruption to learning outcomes (13,20), and the percentage of individuals completely isolated—worsens mental wellbeing (64).

Controlling peak infections relaxes the burden on a university to support positive cases for any given day, and allows resources to be distributed over time. In all behavioral scenarios of our simulation of Fall 2019, we observed that the peak reduction was significantly better in LC (Figure 3) than RI. While RI impacted 58 different locations (classrooms and lecture halls), in S1 and S2, LC achieved better outcomes by closing fewer locations. For example, in S2, RI achieved a 28.9% peak reduction, but LC showed reductions of 49.3% (mobility budget) and 48.1% (exposure risk budget). This was attained by closing 38 or 50 locations respectively. Therefore, with such policies, policymakers need to restrict fewer locations to remarkably minimize the pressure of active infections on campus (e.g., diagnoses, treatment, quarantining).

Universities want to minimize the number of infected cases while ensuring majority of the population remains active on campus to continue successful operation. In Scenario S1, the total number of infections reduced by both LC was more than the reduction shown by RI. were similar. For other behavioral scenarios the total infection reduction between policies was similar (Table S2). In contrast, the impact the policies had on the student schedules was remarkably different. RI forced multiple students to adapt to fully online schedules. In Scenario S2, 9% of students did not visit campus and in S3, 27% of students did not visit campus. On the other hand, in LC, the number of students expected to avoid campus could be as low as 0 and never exceeded 12%. Besides sustaining economic loss to the campus, remote instruction can increase anxiety among students and hinder learning outcomes (20,21). Compared to RI, LC offers policymakers a way to defend against turnover in the student population, without compromising overall control of disease spread (Table 2). Limiting the number of students that avoid campus helps preserve on-campus businesses (18,19) and minimally disrupts the student wellbeing.

Universities are responsible for limiting spread on campus, but they must also ensure that aggressive policies do not worsen mental wellbeing of the community. In terms of internal transmission the reduction is significantly larger with LC (Table 2). However, when LC restricted the infections early in Fall 2019, it left more individuals susceptible to external transmission. College student behavior outside campus on weekends and breaks is known to impact local transmission (65). When policymakers consider LC they should also consider policies on re-entry or required testing based on off-campus activities. In terms of isolating individuals on campus, it’s notable that LC and RI were similar in S2. Interestingly, in S3, where LC closed more than 100 locations, the percentage of isolated individuals per week was less than that of RI. This finding implies that LC can keep individuals on campus without forcing them into complete isolation. Here “isolation” refers to no form of proximate contact with any individual on campus—extreme social distancing where individuals are not even collocated in the same suite or hall. While social distancing is a recommended countermeasure for COVID-19 (8), complete isolation can have adverse effects on psychological wellbeing (64,66,67). LC can help alleviate concerns of closure interventions that increase loneliness and limit social connectedness (66).

By using WiMob to design LC we were able to identify locations for closure at the granularity level of rooms, including unbound spaces such as lobbies and work areas. As policy design budgets changed with every behavioral scenario we found that LC identified different types of locations for closure. First, in S1, we found that most locations that LC targeted are a subset of the auditoriums–like rooms where large classes would take place in Fall 2019. Note, LC needs to restrict only a few such spaces to utilize the same budget as RI. This is because, under S1, RI policies only altered visits to lectures, while these spaces are used for other purposes during other times (e.g., club activities and seminars). We also noted that LC targeted “high traffic” locations like conference center lobbies which are typically used as waiting areas or for networking events. Next, in Scenario S2, we saw that in addition to spaces mentioned earlier, interestingly LC further restricted the use of smaller rooms (occupancy 13–35) which would not be affected by RI (as only classes of size ≥30 are offered online). LC also targeted areas in the recreation center (which includes locker rooms and indoor courts for 4–20 people). This insight indicates that our methodology WiMob accounts for a diverse set of student activities. Moreover, we also found a selection of spaces that would not be frequented by the undergraduate population, such as lab areas and facility buildings like the police station. Lastly, in Scenario S3, LC targeted closure of activity in far more spaces than RI. However, the better outcomes can be attributed to the fact that LC diversified the potential restriction areas. LC restricted heavily used small study rooms or breakout rooms (for 1–6 people). Furthermore, it restricts use of spaces where multiple small groups of people can organically assemble, such as cafes, dining halls, and reading areas. We also observed that LC restricted activity in about 10 Greek Houses but does not target other housing areas—demonstrating its ability to restrict social behavior that could amplify disease spread. Figure S18D shows the diversity in locations for various LC policies.

The results above use an ABM calibrated on the positivity rate of the first 5 weeks of Fall 2020. This rate can be influenced by many latent factors (e.g., mask-wearing, hand washing, distancing, and compliance). To study any effect of these variations, we also calibrated on different time windows throughout the semester. We calibrate on weeks 5–9 and 10–14 in Fall 2020, and validate on the remaining semester. In both cases, compared to RI, we found that LC still exhibits better reduction in peak infections (up to 90%) and internal transmission (up to 77%). In the original calibration, LC maintained the same level of total infections as RI, but with the new periods we found total infections were substantially less than RI (Tables S8 and S9). Another important variable for positivity is the wider context of the campus e.g. urban/rural, the surrounding county, city, etc. To investigate this, we also calibrated our ABM on the positivity rate of different universities in the US in Fall 2020 (along with information from their county to seed external cases). Consider this as a hypothetical where the mobility of the GT community remains the same but disease outcomes resemble a different campus. We calibrated on data from University of Illinois at Urbana-Champaign and University of California, Berkeley. We found no remarkable differences from our findings with GT (Tables S10 and S11).