Association of In-person vs Virtual Education With Community COVID-19 Case Incidence Following School Reopenings in the First Year of the COVID-19 Pandemic

Key Points Question Was resuming in-person vs virtual instruction for middle and high school students in the fall of 2020 associated with differences in COVID-19 community incidence? Findings In a cohort study of 51 matched pairs of counties that reopened with in-person vs virtual instruction, those that reopened with in-person schooling experienced slightly higher county-level COVID-19 incidence at 6 and 8 weeks after in-person reopening compared with those that reopened virtually. Meaning These findings of this cohort study should be interpreted in the context of competing concerns of reduced school engagement and social well-being of children in prolonged virtual learning environments.


Matching variable
Data source Threshold for comparability Date of reopening for in-person schooling for students in 6th grade or above Data on school reopening plans were obtained from a systematic scan of official school district website contents, including announcements, COVID-19 reopening plan documents, and school district press releases. In some instances, official school district-run Twitter or Facebook accounts were used as sources for up-todate announcements on school reopening dates and policies.
Unexposed counties were eligible for matching if they did not reopen for inperson schooling for students in 6th grade or above before January 2021 or reopened at least 4 weeks after the matched exposed county's reopening date.

Geographic area
US. Bureau of Economic Analysis (BEA) US Census metropolitan statistical area (MSA) designations The matched unexposed county must be within the same BEA region or MSA as the exposed county, or within 200 miles of straight-line distance of the exposed county.

Total population
American Community Survey After matching on other variables, an eligible unexposed county that was closest to the exposed county on total population was chosen.

Population density
American Community Survey After matching on other variables, an eligible unexposed county that was closest to the exposed county on population density was chosen. Weekly case incidence per 100,000 county residents for the week prior to exposed county's school reopening

USAFacts
The difference in weekly cases per 100,000 residents at the match date between the exposed county and the matched unexposed county must be below 108, a threshold chosen based on the interquartile range of weekly cases across the sample for August 1 to December 31, 2020. Weekly median reproduction number (Rt) for the week prior to exposed county's school reopening The daily Rt in each county was estimated using the method of Cori et al. a with a moving average window of 3 days and an average of generation time of SARS-CoV-2 assumed as 7.5 (SD 3.4) days The difference in weekly median Rt at the match date between the exposed county and the matched unexposed county must be below 0.30, a threshold informed by research conducted by Huang et al. b

School sports activity
School sports activity data were obtained from a systematic scan of official school district website content, including athletics calendars, COVID-19 reopening plans, district announcements, or linked official Twitter and Facebook announcements. Schools were considered to have active fall sports activities if school sanctioned sports teams were holding practices or competing in games between August 1 and October 31, 2021.
The matched unexposed county must have the same plan for school-district-level fall sports activities as the exposed county.

eMethods. Sensitivity Analyses
We conducted multiple sensitivity analyses to evaluate the robustness of our effect estimates. In sensitivity analysis #1, we used a stricter constraint on the school reopening date of the unexposed counties, by only including matched pairs in which the unexposed county did not reopen schools within 5 weeks (primary analysis used 4 weeks) after the exposed county's reopening date. In this way, we had less overlapping of school-reopened days between matched exposed and unexposed counties in the observation period. In sensitivity analysis #2, we only included matched pairs in which the exposed county reopened schools in the August of 2020. In this analysis, the calendar time of the observation periods (2 weeks prior -8 weeks post school reopening date) was similar for all pairs of counties and was before the holiday season of 2020. The sensitivity analysis #3 only included matched pairs in which the intercounty geographic distance is >50km and <500km. Matched pairs that are too close to each other (<=50 km) were excluded to account for the possible between-county correlation in community COVID-19 transmission, and matched pairs that are too far away from each other (>500km) to account for residual confounding in temperature and unmeasured confounding due to heterogeneity between counties. In sensitivity analysis #4, we considered the potential bias of over-representing case incidence from matched control counties due to our sampling with replacement matching algorithm. Because we used matching with replacement in selecting the matched controls, one unexposed county could be matched with multiple exposed counties. In this analysis, to avoid our results being heavily impacted by a few unexposed counties that were over-represented among the controls, we only included matched pairs in which unexposed counties were only permitted to match to 3 or fewer exposed counties.
Finally, in sensitivity analysis #5, we examined whether the effect of school reopening was confounded by the change in social distancing in communities over time. A series of models were fitted with and without adjustment for the variables that encoded daily county-level measures of social distancing, including community and workplace mobility, over time. Several social distancing and mobility measures from two data sources were considered as potential confounders: 1) percent change in visitation to non-essential venues compared to a pre-COVID-19 period; 2) visitation to retail and recreation places; 3) visitation to parks; and 4) mobility to workplaces. The first, percent change in visitation to non-essential venues was measured using daily cellphone movement provided by Unacast (COVID-19 Social Distancing Scoreboard -Unacast. Accessed May 26, 2021. https://www.unacast.com/covid19/social-distancing-scoreboard). Essential venues (e.g., food stores, pharmacies) and non-essential venues (e.g. cinemas and theaters, spas and hair salons, hotels, restaurants) were categorized based on guidelines issued by various state governments and policy makers (A Guide to State Coronavirus Reopenings and Lockdowns -WSJ. Accessed May 26, 2021. https://www.wsj.com/articles/a-state-by-state-guide-to-coronaviruslockdowns-11584749351). The change in visitation to non-essential venues was calculated by comparing the daily visitation during observation period compared to the average visitations in a four-week pre-pandemic baseline period (February 10th and March 8th, 2020) in each county. The additional measures of social distancing related to visitation to community and workplace settings and were obtained from Google COVID-19 Community Mobility Report. These measures characterize how visits and length of stay at different places change based on data from users who have opted-in to Location History for their Google Account, so the data represents a sample of Google users in each county. The measures were defined as percent change in daily visits compared to pre-COVID baseline, which was the median value, for the corresponding day of the week, during the 5-week period of January 3 rd to February 6th, 2020. 0.0021 a The effect estimates were from generalized linear mixed effect models (GLMM). The models adjusted for population density, percent of population under age 18, rate of poverty, use of masks in public (percent of people reporting to always wear mask in public when expecting to be within six feet of another person, in a survey during July 2-14 of 2020), and daily wet-bulb-temperature (3-14 days lagged rolling averages). The GLMM used a log link function to account for the skewed distribution of daily cases in each county. Hierarchical random intercept and slope were used to account for correlations within matched pairs of counties and the correlations within longitudinal measures of individual county, separately. b The inter-county geographic straight-line distance was calculated using Esri ArcGIS Pro (Redlands, CA). c Because we used matching with replacement in selecting the matched controls, one unexposed county can be matched with multiple exposed counties. d P values obtained from testing for null hypothesis of no significant difference in trajectory of incidence case over time between exposed and unexposed counties, using Wald test on the regression coefficients of three interactions between school reopening and polynomials of time (linear, quadratic, and cubic time The effect estimates were from generalized linear mixed effect models (GLMM). The GLMM used a log link function to account for the skewed distribution of daily cases in each county. Hierarchical random intercept and slope were used in all models (Models 0-5) to account for correlations within matched pairs of counties and the correlations within longitudinal measures of individual county, separately. b Model 1 adjusted for population density, percent of population under age 18, rate of poverty, use of masks in public (percent of people reporting to always wear mask in public when expecting to be within six feet of another person, in a survey during July 2-14 of 2020), and daily wet-bulb-temperature (3-14 days lagged rolling averages). c Model 2 further adjusted for a social distancing measure, defined as the percent change in visitation to non-essential venues compared to a pre-COVID-19 period. Essential venues (e.g. food stores, pharmacies) and non-essential venues (e.g. cinemas and theaters, spas and hair salons, hotels, restaurants) were categories based on guidelines issued by various state governments and policy makers, and visitation to non-essential venues was measured using daily cellphone movement provided by Unacast. The change in visitation to non-essential venues was calculated by comparing the daily visitation during observation period compared to the average visitations in a four-week pre-pandemic baseline period (February 10th and March 8th, 2020) in each county. 3-14 days lagged rolling averages of this measure was used in the model. d In Models 3-5, visitation to retail & recreation places, visitation to parks, and mobility to workplaces were obtained from Google COVID-19 Community Mobility Report. These measures show how visits and length of stay at different places change based on data from users who have opted-in to Location History for their Google Account, so the data represents a sample of Google users in each county. The measures were defined as percent change in daily visits compared to pre-COVID baseline, which was the median value, for the corresponding day of the week, during the 5-week period Jan 3-Feb 6, 2020.