Simulation-Based Estimation of SARS-CoV-2 Infections Associated With School Closures and Community-Based Nonpharmaceutical Interventions in Ontario, Canada

Key Points Question What is the association of school reopening or closure with incident and cumulative COVID-19 case numbers compared with other community-based nonpharmaceutical interventions? Findings In this decision analytical modelling study of a synthetic population of 1 000 000 individuals in Ontario, Canada, compared with community-based nonpharmaceutical interventions, school closure was associated with a small change in estimated COVID-19 incidence trajectories and cumulative case counts. Meaning These findings suggest that community-based interventions to reduce COVID-19 case counts should take precedence over school closure.


Introduction
During the first wave of the COVID-19 pandemic, school closures were a component of nonpharmaceutical interventions (NPIs) enacted to mitigate the transmission of SARS-CoV-2, largely based on the rationale that it had been effective in delaying or reducing the peak of the 2009 H1N1 influenza epidemic. 1 It was speculated that school-aged children infected with SARS-CoV-2, who are less likely to manifest symptoms, 2 may unknowingly pass on infections acquired in school to the members of their household and/or individuals at high risk in the community, thereby accelerating the increase in cases and subsequent strain on health care resources. In contrast, modeling studies conducted early in the course of the COVID-19 pandemic suggested that school closure was effective when combined with other NPIs, but that the effect of closure per se was modest in delaying peak case numbers or reducing the size of the peak. [3][4][5] However, there is a high degree of uncertainty in these results, primarily owing to inability to decouple school closure from other concurrent NPIs.
School closures may affect students adversely in terms of loss of access to high-quality instruction, loss of school-based health and social services, and negative effects on physical and emotional well-being. [6][7][8][9] Closures have been shown to result in adverse economic consequences for families, including loss of work hours to care for children and/or increased child-care expenses, which predominantly affect lower-income households. [6][7][8][9][10] In Ontario, Canada, after closure on March 15, 2020, schools reopened on September 15, 2020.
As in many jurisdictions, the second wave of COVID-19 in the province required the reinstitution of community-based NPIs and consideration of school closures. Given the negative consequences of school closure, determining the magnitude of the benefit of this measure in terms of case numbers is critical. The objective of this study is to determine the relative size of the increase in COVID-19 case numbers attributable to school reopening relative to community-based NPIs in Ontario, Canada, using an agent-based modeling approach informed by observed Ontario data.

This study was conducted under the blanket approval of the University of Toronto Research Ethics
Board for studies concerning the COVID-19 pandemic and making use of government registries. This study adheres to the Consolidated Health Economic Evaluation Reporting Standards (CHEERS) reporting guideline for modeling studies (eMethods in the Supplement).
Households were then randomly assigned to cities or a rural region according to the SPSD/M urban/ rural designation of the associated household type. Within urban settings, households were further randomly assigned to neighborhoods while rural household were assigned to districts.
All members of the synthetic population spent time each day in their households, neighborhoods or districts, and their city or rural region. Additionally, children ages 2 to 3 years were assigned to daycare settings (capped at 10 children), children ages 4 to 13 years were assigned to primary or elementary schools (capped at 23 children per classroom and 150 children per school), and children ages 14 to 17 were assigned to high schools (capped at 15 children per classroom and 150 children per school). Each daycare or classroom was assigned a teacher who was randomly sampled from adults in the region whose SPSD/M industry designation was educational. Adults ages 18 to 34 years could join the workforce or potentially go to college or university (a postsecondary institution).
Adults ages 18 to 34 years not enrolled in a postsecondary institution and those ages 35 to 64 years could be members of the workforce. Workplaces were characterized by industry type, region, and workplace size (ie, 1-20, 21-99, 100-499, and Ն500 workers). Working-age adults were then randomly assigned to workplaces in their region according to their SPSD/M industry designator.
Adults ages 65 years and older were assumed to be retired from the workforce.
All other individuals were initially susceptible to COVID-19 and could contact and be exposed to individuals who were infectious within their household, at school, at college or university, at workplaces, in neighborhoods and cities, or in rural districts. Exposed individuals may not have become infected and therefore remained susceptible, or they may have become infected. Individuals who were infected were unable to transmit the virus during a 4-day latent period, after which, individuals who had been infected became infectious and able to transmit the virus. Individuals who were infectious may or may not have developed symptoms. Individuals who were symptomatic entered a 1-day presymptomatic stage followed by a symptomatic stage. The duration of infectiousness for individuals with or without symptoms was modeled to be 15 days. The proportion of symptomatic cases by 10-year age group was obtained from Ontario's Case and Contact Management Plus (CCM Plus) database (eMethods and eTable 2 in the Supplement).
Symptomatic individuals could seek health care, be confirmed as cases, and either be admitted to hospital or sent home to quarantine until recovered. Symptomatic individuals who did not seek health care could self-isolate until recovery. Self-isolated or quarantined individuals could still transmit SARS-CoV-2 to their household members. A proportion of asymptomatic cases would be detected through contact tracing, reported as confirmed cases, and quarantined until recovery.
Individuals who were symptomatic but sought neither health care nor testing and undetected individuals who were asymptomatic would be free to transmit virus in household and nonhousehold settings. The probability of detection (ie, of an individual who was infectious being a confirmed case) during the first wave of COVID-19 in Ontario was estimated separately for symptomatic and asymptomatic cases via the cumulative number of confirmed cases until June 9, 2020, and the seroprevalence of antibodies for SARS-CoV-2 reported by Public Health Ontario until that date. 12 Death outside of hospitals, such as in a long-term care facility, was not considered in the current iteration of the model, nor was nosocomial transmission within hospitals.
Susceptible individuals were modeled to be potentially infected via close contact with an individual who was infectious. Mixing of individuals was assumed to occur randomly within the included settings based on the mean number of close contacts per day prior to the institution of NPIs in March and April 2020 from a contact survey, the CONNECT study 13  Children aged >10 y (first wave) 0. 13 Li et al (2020), 18 Kucharski et al (2020), 16 Bi et al (2020), 14 Cheng et al (2020) 14

Daycare and School Opening Scenarios
We simulated 2 opening scenarios, (eTable 4 in the Supplement). First, we modeled a counterfactual scenario in which daycares and schools did not reopen on September 15, 2020 (scenario A). Second, we modeled a scenario in which schools and daycares reopened on September 15, 2020, (scenario B) but with several measures in place to limit within-school transmission of COVID-19: daycare centers were capped at 10 children, primary and elementary class sizes were capped at 23 students, and high school classes were capped at 15 students; students remained in their assigned classrooms for the school day rather than moving among classrooms; universal masking was in place; in designated high schools in urban areas, students attended school only on alternate weekdays; and if more than 2 confirmed cases of COVID-19 occurred in a daycare or classroom less than 2 weeks apart, the daycare or classroom was closed for 14 days, with the children in the class excluded from school rather than moved to another classroom (eTable 4 in the Supplement).

NPI Scenarios
We modeled 3 community-based NPI scenarios (eTable 4 in the Supplement) in Ontario at the beginning of October 2020 in response to increasing confirmed daily case incidence from 185 cases

Statistical Analysis
The 2 school reopening and 3 community-based NPI scenarios created 6 combinations, 1A to 3B (eTable 4 in the Supplement). For each combination, we ran 100 repetitions of the ABMCT. In addition, 2 deterministic 1-way sensitivity analyses were conducted using the 2A and 2B parameter sets and either allowing the effectiveness of within-school mitigation to vary or the relative effectiveness of community NPIs to vary. All analyses were conducted with TreeAge Pro statistical software version 2020 R2 (TreeAge). Data were analyzed from May 5 to October 20, 2020.

Results
All results are presented on a provincial scale: case numbers derived from the 1 000 000-person synthetic population were multiplied by 14.5). The simulated population had a mean (SD) age of 41.5 (23.4) years, and 507 304 (50.7%) were women.

Classroom Closures
The

Cumulative Confirmed COVID-19 Cases
When community-based NPIs were not implemented (scenarios 1A and 1B) the mean, cumulative

Sensitivity Analyses
When NPIs were implemented and their effectiveness held at the base case value, as the effectiveness of mitigation efforts within schools diminished, the difference in mean estimated cumulative case numbers by October 31, 2020, between keeping schools closed or reopening them increased (eFigure 23 in the Supplement). When school mitigation effectiveness was held at the base case value, as the effectiveness of community-based NPIs decreased, the difference in mean estimated cumulative case numbers between keeping schools closed vs reopening them did not increase (eFigure 24 in the Supplement). Modeled counts subjected to Gaussian kernel smoothing with a bandwidth of 2 days.

Discussion
The findings of our decision analytical modelling study indicate that school reopening, compared with the counterfactual situation in which they remained closed on September 15, 2020, was associated with an increase in incident and cumulative COVID-19 cases across 3 community-based NPI implementation scenarios in Ontario, Canada, between September 1 and October 31, 2020, but most infections among students and staff were estimated to be acquired in the community rather than within schools. In our simulations, we show that community-based NPIs directed at reducing contacts, such as restricting gatherings, limiting workplaces to essential workers, and reducing transmission between contacts (eg, by requiring mask-wearing), had a much larger effect in our simulations on reducing incident or cumulative COVID-19 cases than keeping schools closed vs reopening them. It is important to note that our model assumed that measures would be taken in schools to mitigate spread of the virus by reducing high school class sizes, having students remain in the same classroom rather than moving among them, requiring universal mask-wearing, and closing classrooms when more than 2 confirmed SARS-CoV2 infections occurred within any 2-week period.
The results of our simulations are in agreement with a meta-analysis of COVID-19 pandemic data in Hong Kong, Singapore, and mainland China conducted by Viner at al, 24  Our results were also similar to several prior modeling studies. [3][4][5] However, these studies [3][4][5] were conducted early in the course of the pandemic when precise information regarding transmission probability between cases and when observational data required for model calibration were unavailable or did not distinguish between infections acquired in school vs those acquired in the community and imported into schools. Our results bolster the prior studies [3][4][5]  correlation between the frequency of school outbreaks and regional COVID-19 incidence. 39 Our simulation results suggest that the relatively modest estimated reduction in COVID-19 cases derived from keeping schools closed should be balanced against the adverse effects on children and families. School closure may interfere with educational advancement, prevent access to schoolbased health and social programs, reduce physical activity among children, and negatively affect household finances, particularly among low income households. 10

Limitations and Strengths
Our study has limitations. Like all models, our ABMCT is a simplification of complex, dynamic patterns of interactions and viral transmission in real populations. For example, although close contacts may occur among students, particularly for adolescents, just before and after school hours, this was not explicitly modeled. However, this mixing was partially captured in the close contacts among individuals in neighborhoods or rural districts. Modeled NPI scenarios may not precisely reflect the actual implementation of public health policy in Ontario but have the advantage of quantitatively determining the outcomes associated with school reopenings across a range of community-based NPI possibilities with robust qualitative results. Also, by not being overly jurisdiction-specific, the results of our ABMCT may be broadly applicable to regions similar to Ontario in terms of school systems, populations, and economy. Latent, incubation, and infectious periods were assumed to be fixed rather than drawn from distributions, which may have caused an underestimation of the uncertainty of pandemic trajectories but would not have affected the mean results. The ABMCT did not explicitly consider COVID-19 testing volumes as a potential driver of confirmed case numbers but rather assumed that, over time, testing rates, and therefore case detection probability, would stabilize.
Our study also has several strengths. The use of a synthetic population allows for realistic linkages of individuals into discrete clusters, such as households, classrooms, and workplaces, unlike other approaches that treat mixing of individuals among age strata in a probabilistic fashion. The agent-based modeling approach allows for so-called super-spreading events, in which a small number of individuals have a disproportionately large number of close contacts, and for representation of relatively complicated measures, such as classroom closure after a certain number of confirmed cases over a particular interval, which would be difficult with other transmission model designs, such as compartmental models. Unlike previous transmission models, we had the advantage of being able to calibrate our model to observational data on case numbers in Ontario from March through September 2020. Agent-based modeling allows the source of infections to be determined allowing the disentanglement of infections transmitted in schools vs those acquired in the community and brought into the school setting.

Conclusions
The results of this decision analytical modeling estimated that the magnitude of the effect of schools being open on COVID-19 cases was substantially lower than the effect of community-based NPIs.
These findings suggest that school closure be considered the last resort in the face of a resurgence of COVID-19, and that efforts should instead focus on widespread reduction of community transmission.