Optimizing COVID-19 testing strategies on college campuses: evaluation of the health and economic costs

Colleges and universities in the US struggled to provide safe in-person education throughout the COVID-19 pandemic. Testing coupled with isolation is a nimble intervention strategy that can be tailored to mitigate health and economic costs, as the virus and our arsenal of medical countermeasures continue to evolve. We developed a decision-support tool to aid in the design of university-based testing strategies using a mathematical model of SARS-CoV-2 transmission. Applying this framework to a large public university reopening in the fall of 2021 with a 60% student vaccination rate, we find that the optimal strategy, in terms of health and economic costs, is twice weekly antigen testing of all students. This strategy provides a 95% guarantee that, throughout the fall semester, case counts would not exceed the CDC’s original high transmission threshold of 100 cases per 100k persons over 7 days. As the virus and our medical armament continue to evolve, testing will remain a flexible tool for managing risks and keeping campuses open. We have implemented this model as an online tool to facilitate the design of testing strategies that adjust for COVID-19 conditions, university-specific parameters, and institutional goals.

The model structure is diagrammed in Figure S1 and described in the equations below. In short, the population is divided into four groups based on vaccination status (subscripts and ) and quarantine status (subscript ). Within these groups, individuals can transition between disease states: susceptible ( ), exposed ( ), infectious pre/asymptomatic ( ), infectious symptomatic ( ), and recovered ( ). The symbols , , , , and denote the number of people in that state in the given vaccination/quarantine group. Individuals can transition from the active (in contact with others) state to the quarantine state and back based on receiving a positive test result and being released from quarantine, respectively. The total size of the active population (anyone not in quarantine) at any time is given by: . The force of infection at time t, denoted as is given by: Where and are the transmission rate of asymptomatic and symptomatic individuals respectively, and is the reduction in transmissibility of vaccinated infected individuals.
The model equations governing transition from one state to the next are given by: Where is the relative susceptibility to infection if vaccinated, and are the sensitivity and specificity of surveillance tests, and are the willingness to test amongst the vaccinated and unvaccinated individuals respectively, and are the surveillance testing frequencies among vaccinated and unvaccinated individuals respectively, is the probability that an individual isolates after receiving a positive test result, is the probability that an individual isolates and seeks testing after developing symptoms, is the rate of confirmatory testing, is the transition rate from exposed to infectious calculated from the latent period duration, and are the proportion of infected individuals who eventually show symptoms in the unvaccinated and vaccinated groups respectively, and are the transition rates out of the asymptomatic compartment for the vaccinated and unvaccinated individuals, respectively, and and are the transition rates out of the symptomatic compartments for the active and quarantined individuals respectively. The values of the transition rates are estimated from the average time spent in each compartment. See Table S2 for all parameter values. The model is implemented as a system of differential equations that are solved in R using the deSolve package [1]. It is assumed that the entire population of active individuals, including vaccinated and unvaccinated individuals are well mixed. The initial conditions are provided in Table S1, the values for the parameters in the transmission model are provided in Table S2, and the cost assumptions are provided in Table S3. Uncertainty was incorporated into the model via parameter uncertainty only. In short, we ran 100 simulations for each scenario we were interested in. At each simulation, we sampled from the parameter distributions prescribed, and ran the deterministic model forward. Summary statistics were calculated based on the results either at each time point or as an endpoint from eac simulation. S1 Figure. Compartmental model of COVID-19 transmission incorporating testing and vaccination. Each group (defined by vaccination ( , ) and quarantine states ( )) is modeled with a set of compartments. Upon infection, susceptible individuals ( ) progress to the exposed compartment ( ) and then to asymptomatic infectious compartment ( ). Some of those progress to symptomatic infections ( ) and some go directly to recovered ( ). Testing frequencies dictate the rate at which individuals move into their corresponding disease states in the quarantine group. S1 S2 Text. Results for vaccination coverage ranging from 50% to 90% with all students tested In this section, we provide epidemiological projections for vaccination scenarios ranging from 50% to 90% vaccination rate of the student body, with testing policies either targeted to all students equally (main text results), half as much in vaccinated students, and in unvaccinated only.
S2 Figure. Projected COVID-19 cases among students under different levels of proactive testing, assuming 50%, 60%, 70%, 80%, and 90% vaccination coverage amongst students. Graphs project the daily prevalence of symptomatic infections detected through a combination of symptomatic and proactive testing. Colors indicate the testing frequency for all students, assuming 75% compliance. Shading indicates the 90% prediction intervals. The red dashed horizontal line represents the very high risk threshold.
Focusing on the main text scenario where all students are tested at equal rates regardless of vaccination status, we show the full breakdown of cost and infection projections for vaccination rates ranging from 50% to 90%. S3 Text. Results for vaccine coverage ranging from 50% to 90% and different populations tested The main text results focus on the testing policy in which vaccinated and unvaccinated are tested at the same rate. However, it is reasonable that decision-makers would also want to explore alternative policies, such as testing only the unvaccinated or testing vaccinated half as much as the unvaccinated. In this section, we present the results of the projected symptomatic detected cases for all three of these testing policies side by side ( Figure S4). S4 Figure. Projected COVID-19 cases among students under different levels of proactive testing, assuming 50%, 60%, 70%, 80%, and 90% vaccination coverage and in testing policies for all students, testing vaccinated at half the rate, and testing in the unvaccinated only. Graphs project the daily prevalence of symptomatic infections detected through a combination of symptomatic and proactive testing. Colors indicate the testing frequency for all students, assuming 75% compliance. Shading indicates the 90% prediction intervals. The red dashed horizontal line represents the very high risk threshold. The top row corresponds to the testing of all students at an equal rate, the middle row corresponds to testing vaccinated at half the rate of unvaccinated, and the bottom row corresponds to testing of the unvaccinated only.
In the main text, we provide a table with the testing recommendation, total proactive tests, total cost to the university and cost (of testing) per infection averted assuming we test all students at the same rate. We provide the companion tables for the testing policies in which vaccinated are tested at half the rate as unvaccinated and vaccinated only are tested. *indicates that even daily testing in the unvaccinated isn't sufficient to keep symptomatic cases below the very high risk threshold with 95% certainty. This is because at high vaccination levels, this corresponds to a very small percent of the student body.

S3 Text. Sensitivity analysis: vaccine efficacy against infection and transmission
In the main results, we used the most up-to-date estimates of vaccine efficacy against infection, symptomatic disease, and transmission to make the health and economic projections. However, these estimates changed significantly from the time we initiated this project to when we finished it, and we expect that estimates of vaccine effectiveness will continue to be updated with new data and potentially with the impact of new variants. To better understand how our results hinge on this uncertainty, we performed a sensitivity analysis on vaccine effectiveness against infections and symptomatic disease. For each vaccination rate ranging from 50 to 90% coverage, we assumed weekly testing was present at 75% participation in all students regardless of vaccination status. We then varied vaccine efficacy against infection from 40% to 70% in 10% increments, and vaccine efficacy against symptomatic disease from 55% to 85% also in 10% increments. For each condition, uncertainty on vaccine efficacy against infection ranged from 5% below the median value to 22% above the median value, consistent with what is present in our main results where vaccines are assumed to be 53% effective against infection at median [5], with an uncertainty range from 50% effective to 63% effective [6]. The results of the sensitivity analysis reveal the large degree to which the results depend upon vaccine efficacy. For example, in the 70% vaccination scenario, weekly testing is more than sufficient to prevent exceeding the very high risk threshold if vaccines are 70% effective against infection and 85% effective against symptoms. However, if vaccines are only 50% effective against infection, as has recently been indicated [5], then weekly testing is projected to cause symptomatic cases to exceed the very high risk threshold. This critical uncertainty explains why analysis performed in the summer of 2021 [2] presents recommendations that differ from the main results in this report. This is one of the motivating factors for releasing the Rshiny app alongside this report, allowing for the various parameters such as vaccine efficacy to be updated as the situation continues to evolve.
Another key uncertainty in vaccine efficacy is the degree to which vaccinated individuals, if infected, go on to transmit to others. Early reports indicated that breakthrough infections reduced transmissibility by 50% [20], but more recent reports have indicated that vaccinated infected individuals have similar viral loads to unvaccinated infected individuals [4]. We performed a sensitivity analysis varying the effect of vaccines on transmissibility from breakthrough infections having the same transmissibility to being 50% less transmissible.
The results of the sensitivity analysis indicate that the degree to which vaccinated individuals transmit to others has a large effect on the resulting outbreak. For example, if vaccinated infected individuals have the exact same transmissibility as unvaccinated infected individuals, then even at 90% vaccination rate, weekly testing does not guarantee that symptomatic cases will not exceed the very high risk threshold. This result again underscores the need to be adaptive in testing strategies and willing to adjust input assumptions as more data about vaccine efficacy becomes available.

S4 Text. Modifications to framework/Rshiny app for future variants
The default setting of the partner Rshiny app is parameterized to a 50,000 student university, originally intended for use for university administrators for the Fall of 2021. At this time, the Delta variant was the dominant circulating variant, and "fully vaccinated" was defined as 2 doses of the mRNA vaccines or 1 dose of an adenovirus vaccine. In order to adapt this tool for use with different variants, for example the Omicron variant, which is the dominant variant circulating at the time of submission in early 2022, we suggest the following changes: S6 We acknowledge that the parameter options presented to adjust in the app do not represent the complexity of all things to be considered in the face of a novel variant, for example, multiple different levels of immunity (i.e. fully vaccinated but not boosted, recovered but greater than 6 months ago, etc). Our hope is that the framework is flexible enough to allow a user to make reasoned estimates of the range of some of the parameters, and to be able to adjust them as new data emerges and specific to their university's demographics and epidemic scenario. Additionally, all code to build the app and generate all figures, along with code to estimate some campus specific parameters, is available at https://github.com/kej1993johnson/university_testing_vax_proj.git .