Study population and setting
This study estimated the county-level effects of fall term school reopening on COVID-19 hospitalization in the US through October, 2020. The authors used a difference-in-differences approach, whereby trends in hospitalizations over time were compared between counties with different reopening modalities. A supplementary analysis used teacher bargaining power as an instrumental variable, as it was hypothesized to affect COVID-19 hospitalization only through its effect on school reopening. The degree of reopening in a county was defined as the proportion of students allowed to attend in-person classes at a given time, which could take on three values at the district level (0 for fully remote, 0.5 for hybrid, and 1 for fully in-person); alternately, counties with any in-person learning were compared to those without. Three sources of school reopening data were employed, separately and in combination: Burbio, a private company that collected data from the websites of 1,200 mostly large school districts (~9% of all US school districts), aggregated to the county level, and imputed results from nearby districts for the least populous counties; MCH, a private company that collected data by telephoning nearly all school districts in the US (8,283 districts were included out of >13,000); and Education Week, a trade publication that collected data on 907 of the largest school districts (~7% of all districts). The outcome was defined as hospitalization with a diagnosis of COVID-19 or COVID-19-related symptoms. The primary source of information on COVID-19 hospitalization was medical claims data from Change Healthcare, which processes 55% of all medical claims in the US; sensitivity analyses were conducted using facility-level data from the US Department of Health and Human Services. Teacher unionization data, used in the instrumental variable analysis, were obtained from the US Department of Education. Potential confounding variables that were considered included state-level transmission control policies and college/university reopenings; county-level fixed effects and time effects were also included in models.
Summary of Main Findings
Across data sources, the proportion of districts that initially opened either fully in-person or with a hybrid approach ranged from 51% to 74%. The results indicate that school reopenings during the late summer and autumn of 2020 did not increase COVID-19 hospitalizations in counties that had lower hospitalization rates from March to July 2020 before reopening (below the 75th percentile). Results were similar when using different analytical approaches (e.g., using different data sources, treating reopening as continuous or as binary with propensity score matching, using teacher bargaining power as an instrumental variable, adjustment for college/university reopenings and state non-pharmaceutical interventions). For counties with higher hospitalization rates before reopening (>36-44 per 100,000 per week), some model variants showed increased hospitalization in counties with in-person learning, but results were inconsistent across methods. Because hospitalization trends in counties with and without in-person learning were not parallel before reopening, the primary results presented were from propensity-matched models and instrumental variable analysis.
The outcome variable of COVID-19 hospitalization is less subject to ascertainment bias than outcomes related to SARS-CoV-2 test positivity (due to variability in testing policy, availability, and accuracy). The authors performed multiple variants of their analysis using different data sources and statistical methods, providing several robustness checks on their results and focusing their conclusions on those results that were robust across approaches. The instrumental variable approach was a reasonable and unique approach to address possible confounding, and its limitations were discussed. There was generally careful framing of the results with sources of uncertainty highlighted.
A primary source of uncertainty is the quality and completeness of the school reopening data. Data from Burbio included <10% of all school districts and imputed reopening policies for 25% of the student population; data from MCH included approximately 60% of all school districts. The data sources displayed generally poor agreement regarding the proportion of districts in each reopening category. The three categories of school reopening also hide considerable variation in the actual educational environment: for example, a district with crowded classrooms and few infection control procedures could be classified as “hybrid” if it offered remote options for some students, while another district with small groups of students taught in strict pods and the majority learning remotely would also receive the “hybrid” designation. Hospitalization data covered approximately half of the population; if those not included (for example, the uninsured) were affected differently by school reopenings, results would be biased. Few potential confounding variables were considered, and those omitted from models (for example, population density, mask use, or local transmission control policies) may have been associated with both reopening and COVID-19 hospitalizations. Finally, the data are ecological in nature and may mask effects on subgroups of concern (for example, staff members or racial minorities).
This is the most comprehensive study to date, in the US or elsewhere, regarding school reopening and possible effects on SARS-CoV-2 transmission and/or COVID-19 outcomes.
This review was posted on: 11 February 2021