key: cord-0790213-ap7q3axi authors: Godoy, A. A.; Hart, R. K.; Groetting, M. W. title: Re-opening schools in a context of low COVID-19 contagion: Consequences for teachers, students and their parents date: 2021-03-26 journal: nan DOI: 10.1101/2021.03.25.21254219 sha: 799dbe6e2c870afae07929f4824cb4da3a9dd5c4 doc_id: 790213 cord_uid: ap7q3axi To balance children's right to schooling with contagion management, knowing how school re-openings affect the spread of SARS-CoV-2 is crucial. This paper considers effects on testing and positive tests for SARS-CoV-2 of re-opening Norwegian schools after a six-week closure to reduce contagion. We estimate the effect of school re-opening for teachers, parents and students using an event study/difference-in-differences design with comparison groups with minimal exposure to in-person schooling. We find no evidence that incidence increased following re-opening for either students, parents, or teachers pooled across grade levels. We find some suggestive evidence that infection rates among upper secondary school teachers increased; however, quantitatively, the effects are small and transitory. At low levels of contagion, schools can safely be re-opened when other social distancing policies remain in place. During the spring of 2020 schools across the world were closed as part of an effort to curb the first wave of the SARS-CoV-2 pandemic. A school closure of this dimension is unprecedented in modern history (Insights For Education 2020). When instruction is given online, lack of inperson schooling can have negative impacts on children's development, health and well-being (see, e.g., Donohue & Miller 2020 , Dooley et al. 2020 . 1 Due to their potentially detrimental consequences, school closures have been among the most controversial virus containment policies. To weigh children's burdens from school closures against increased contagion risk, policymakers need precise knowledge of how school closures and re-openings affect contagion rates for teachers, students and parents. In this paper, we assess the effects of school re-openings on the number of tests and confirmed incidence of SARS-CoV-2 for teachers, students, and parents. While the initial school closures were implemented as part of a broader set of non-pharmaceutical interventions, including the closure of many in-person businesses, the timing of school re-opening did not coincide with any other major policy shifts. School re-openings thus provide an attractive context for isolating the impact of school closure policies. To identify the effects of re-openings, we implement a difference-in-differences research design, comparing these outcomes before and after schools re-opened across groups with different levels of exposure to in-person schooling. 2 We rely on rich register data covering the universe of Norwegian residents. These data include individual-level data on SARS-CoV-2 testing and test results, occupation, and industry as well as other demographic information. We can connect children to parents, enabling comparisons of infection rates and testing among parents with (high school) school children to parents of young adult children who recently graduated high school. Similarly, the granular data allow us to identify teachers in elementary 1 While it is still too early to assess the long-term effects of pandemic-related school closures on student outcomes, early evidence suggests a "COVID learning gap", with increased social disparities in learning outcomes (Engzell et al. 2020 ). This effect is potentially driven by parents with higher social background giving more extensive support to online learning (Bacher-Hicks et al. 2020). and secondary schools as well as comparison groups. Our main findings can be summarised along the following lines. First, we show that confirmed incidence rates for students, parents, and teachers track closely with rates in their respective comparison groups throughout the first wave of the pandemic. Consistent with this, event study models fail to find evidence that the timing of school re-openings corresponds with a significant shift in confirmed incidence for either students, parents, or teachers. This result holds both in the country as a whole as well as in separate analyses of the capital region with its relatively higher infection incidence. The precision of our estimates allows us to rule out large increases in confirmed incidence following re-opening, e.g. the estimated 95 percent confidence interval for teachers allows us to rule out increases in confirmed incidence larger than 2.1 weekly cases per 100,000 teachers. Overall, our findings suggests that in a context with low contagion rates, where social distancing policies are maintained both in schools and in society as a whole, school reopenings do not necessarily lead to increased incidence among affected groups. Our findings contribute to the sparse body of knowledge on the impact of school closures on SARS-CoV-2 infections. In a recent systematic review, Walsh et al. (2021) identified only ten empirical studies of the effect of school closures on the spread of COVID-19. Effects ranging from no impact to substantial reductions in incidence and mortality were reported. Importantly, for most studies, the effect of school closures could not be isolated. In addition, the studies with the least issues from confounding were among the studies that reported no effects. To our knowledge, only a few examples of quasi-experimental studies aiming at identifying the causal effect of school closures on SARS-CoV-2 infection exists. Vlachos et al. (2020) exploit the fact that upper secondary schools in Sweden moved to online teaching, while lower secondary schools did not. They compare teachers and parents of students in the last year of lower secondary school to teachers and parents in the first year of upper secondary school and find somewhat higher infection rates among teachers and parents of the former group. Finally, exploiting variations in school openings (and closures) due to staggered summer holidays across German regions, von Bismarck-Osten et al. (2020) and Isphording 3 . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 https://doi.org/10. .03.25.21254219 doi: medRxiv preprint et al. (2020 find little evidence for school openings increasing the spread of SARS-CoV-2 across all German municipalities. More generally, our findings shed light on patterns of SARS-CoV-2 transmission among teachers and students. Current research suggests that Norwegian teachers had relatively low contagion levels compared to other professions during the spring of 2020, and somewhat elevated levels during the fall of 2020 (Magnusson et al. 2021) . Our approach adds to the understanding of contagion in schools by analysing whether changes in contagion and testing around the time of reopening are larger for teachers than for comparable groups. Current evidence suggests that severe consequences of COVID-19 disease are rare in children. However, children's role in the transmission of the virus is still subject to some debate and while symptomatic children are found to shed virus in similar quantities and to infect others in a similar way as adults, it is unclear how infectious asymptomatic children are (ECDC 2020). In addition, because children are more often asymptomatic, they are less likely to be tested, thus causing the prevalence among children to be potentially under-reported. However, a study comprising systematic testing of contacts of confirmed COVID-19 cases in 13 outbreaks in schools in the Norwegian capital region during August-November 2020, found minimal transmission from children (below age 14) to peers and adults (Brandal et al. 2021) . While the main analysis in this paper concerns high school students, we also present some evidence on effects of school re-openings for fourth graders (age 10) compared to fifth graders (age 11). Because schools re-opened two weeks earlier for the former group, we also compare changes in SARS-CoV-2 infection and testing rates among these two groups (and their parents). In accordance with our main results, we also find no evidence of increased infection of school re-openings for these two groups either. The paper proceeds as follows. Section 2 presents relevant background on SARS-CoV-2 infection rates and testing in general, and in Norway. Section 3 presents the empirical strategy in this paper, and Section 4 describes the data and presents some summary statistics. In Section 5, we present the results together with a set of robustness checks. Section 6 presents extensions and mechanisms. Section 7 concludes. 4 2 Background on lockdown and testing in Norway The first confirmed case of SARS-CoV-2 in Norway was registered on February 26, 2020. On March 12, 2020, the government initiated a lockdown to curb the spread of the disease, which included closing all preschools and primary and secondary schools. People were advised to limit the number of close contacts, to meet outside while maintaining physical distance, and to intensify basic hygiene measures. Moreover, all cultural or sporting events and all organised sports were prohibited. Mandated closures of recreational facilities (fitness studios, gyms, swimming pools, etc.), beauty salons (hairdressers, spas, etc.), bars, and restaurants 3 were enforced. People were advised to avoid all non-essential travel and to limit public transportation, and mandatory quarantine after international travel was enforced. In addition, all health care workers were prohibited from international travel (The Norwegian Directorate of Health 2020). With very few exceptions, and for the majority of students, schools remained closed from March 13. In-person attendance resumed for children of workers of critical importance for managing the pandemic and basic services, as well as for particularly vulnerable students (2.5 percent of students) or students granted special education (2.5 percent of students). 4 Due to concerns about the consequences of lockdown, particularly for vulnerable children, and to accumulated knowledge of lower COVID-19 infection and morbidity rates among children in general, school closures were among the first measures to be lifted. Preschools and elementary students in grades 1-4 were allowed back in schools starting April 27, while inperson schooling for older students resumed two weeks later (May 11). Schools reopened in a context of low spread of SARS-CoV-2 and relatively few other channels for transmission (Telle et al. 2021) . While local authorities were allowed some flexibility as to when the reopening was carried out, the majority of schools reopened in accordance with the legislation. About 36 percent of elementary schools reported a few days' delay in order to accommodate the strict infection 3 An exception was given to restaurants where physical distance in excess of 1 meter could be maintained. 4 These numbers are from elementary schools (The Norwegian Directorate for Education and Training 2020). Similar numbers from high schools are not available. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) When schools were opened after lockdown, extensive public guidelines to curb transmission were given by the health authorities. These measures included self-isolation of sick children and staff, intensification of basic hygiene (hand-washing, frequent cleaning of facilities) and physical distancing. Mask wearing was neither mandated nor encouraged for children. Importantly, a "cohort" system was introduced as the key physical distancing measure. The cohorts consisted of fixed groups of students and employees with limited contact between cohorts while allowing children to socialise within the cohorts (see Johansen et al. (2020) for more details on the guidelines). Cohorts were not recommended for students in high schools because it would excessively limit teaching. Because the current public guidelines were too strict to allow normal teaching, a "traffic light" system was introduced on May 29. Under this system, the intensity of the infection control measures reflected the local infection rates so that the infection control measures were to increase with the infection rates in society. Prior to the lockdown, SARS-CoV-2 tests were in short supply, and people were encouraged to self-quarantine if they developed symptoms of COVID-19 disease. Testing was limited to those with severe symptoms who had also been to affected areas or had been in close contact with someone with confirmed SARS-CoV-2 infection. On March 13, the first day of lockdown, test criteria were updated to include those with severe symptoms of respiratory infection who were either admitted to or working at a health care facility, and, from March 20, the recommendation was that all persons with symptoms should be tested. However, because test capacity was still limited, testing was conducted among a list of prioritised 5 In week 24, this number had fallen to 0.5 percent. 6 The figures presented here are from elementary school re-openings presented in a report by The Norwegian Directorate for Education and Training (2020). There are no such public reports for high schools, however, correspondence with local authorities supports that high schools opened in accordance with legislation and that the majority of students were present, at least to some degree, each week. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; groups. School staff and students were added to the list of prioritised groups on April 20. 7 Thus, when schools (and kindergartens) were re-opened, frequent testing upon development of symptoms was among the measures to keep the pandemic under control. In short, while a large proportion of cases were always expected to go undetected, positive tests are a much stronger proxy of actual cases at re-opening than at lockdown. 8 In this study, we seek to estimate effects of school openings on three groups that were directly impacted by school openings, namely students, teachers, and parents of school-age children. We implement a difference-in-differences approach comparing infection rates of groups that were directly affected by school openings (students, teachers, and parents of high school students, referred to as "in school") with outcomes of groups that were not directly affected (young adults above school-leaving age, non-teacher professionals, and parents of young adults) before and after the re-opening of Norwegian schools. Schools re-opened in two steps; in-person instructions for the youngest students (pre-school through grade 4) resumed on April 27th, while older students (grades 5-13) were allowed back in school on May 11th. In principle, it is possible to leverage the staggered implementation to identify effects of re-opening on testing and incidence rates. However, the fact that these two dates are so close together leaves us with only a 14 day period to evaluate differential changes in outcomes. Our primary analysis thus focuses on the second step in the re-opening of schools, i.e. May 11th, after which all teachers and students were allowed to resume in-person instruction. 9 We do not observe the degree to which individual students and teachers were physically present at schools following re-opening. As a consequence, our estimates should not be inter-7 The description of testing is based on logs of testing guidelines provided to the researchers by the Norwegian Institute of Public Health. 8 Teachers were among the groups most frequently tested during the first wave of infection (spring 2020) (Magnusson et al. 2021). 9 For completeness, we have analyzed the impact of the first stage re-opening on 4th grade students and their parents, using 5th graders as a comparison group. Results from these models, presented in section 6, are consistent with our main findings. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101/2021.03.25.21254219 doi: medRxiv preprint preted as the effects of physical attendance. Rather, we estimate an intention-to-treat (ITT) effect of re-opening. Let y igt denote the outcomes of individual i (belonging to group g) in week t. Our basic difference-in-differences regression specification can be written: where θ t and θ g denote a set of week and group fixed effects. 10 Our primary parameter of interest β post is attached to the interaction term between two indicator variables: treated g , equal to 1 if g is directly affected by the school re-opening, and post t , equal to 1 after schools reopen. We estimate equation (1) effects on parents, we compare parents whose youngest child is age 17-19 to those whose youngest child is aged 20-22. Thus, any effects of school re-opening from younger siblings will not contaminate our estimates. In the final sample, teachers employed at primary or secondary schools are compared to other workers in "professional" occupationsexcessively according to the ISCO-88 classification. See further details on the sample definitions in section 4. Our empirical strategy allows for teachers, students, and parents to have different COVID-19 incidence levels from their comparison groups. For instance, 17-19 year olds are more likely to still live at home relative to young adults aged 20-22. Our key identifying assumption is that each affected group would have trended in parallel with their comparison group, in the absence of schools re-opening. Under this assumption, β post will capture the causal effect of school opening on COVID-19 outcomes for teachers, students and parents. While the parallel trends assumption is inherently untestable, we can test for parallel pre-trends using the event-study specification: 10 For students, the group fixed effects are a set of age dummies; for parents, the group fixed effects are a set of dummies for age of youngest child; for teachers, the model includes occupation and industry fixed effects. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101/2021.03.25.21254219 doi: medRxiv preprint In this specification, the primary parameters of interest are the β k coefficients. For k < 0, the coefficient path should be close to zero if parallel pre-trends hold and a causal effect of reopening on our outcomes of interest should manifest as a discontinuous shift at or shortly after k = 0. Our baseline specification does not include covariates. However, to assess robustness with respect to differing trends, we estimate an augmented specification of equation (1) that includes time trends interacted with baseline characteristics. It should be noted that re-opening schools could increase teachers' and students' COVID-19 risks (relative to comparable professions) in other ways as well, e.g. if they are exposed when commuting to work while comparable professionals to a larger extent continue working from home. However, we emphasise that the re-opening of schools happened in a context of continued strict restrictions (e.g. on travel and gatherings). The re-opening thus pertained both to a change in exposure to social contacts for teachers and students alike and, indirectly, for the student's parents. The data applied in this paper are from the Emergency preparedness register for COVID-19 (Beredt C19 . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101/2021.03.25.21254219 doi: medRxiv preprint and the National Population Register. Individuals are linked across the registers and to family members using unique (de-identified) personal identifiers. Consequently, detailed information on demographics (age, gender, county of residence and family members) and employment (industry, occupation, and county of employment) for each individual are merged with data from COVID-19 PCR tests. Our main outcome variables are (i) the incidence of SARS-CoV-2, defined as incidence per 100,000 capita per week, and (ii) test rates, defined as tests per 100,000 capita per week. The number of confirmed cases is an imperfect proxy for the true incidence of SARS-CoV-2, because it is likely to reflect variation in the availability of testing as well as variation in underlying incidence. This is potentially a problem for our analysis if the re-opening of schools shifts the likelihood of being tested. To assess this possibility, we also analyse the total number of tests. We construct three separate regression samples for this study. First, students includes all residents aged 17-22, excluding those working as teachers or employed in primary or secondary schools. This sample consists of students in the final years of high school (aged 17-19) and a comparison group consisting of those recently graduated (aged 20-22). Second, parents includes the parents of the former sample (students). In this sample, outcomes for those where the youngest child is still of school age (17-19) will be compared to those where the youngest child has recently graduated (age 20-22). Finally, professionals, includes teachers working at schools defined according to occupation and industry (ISCO-08 codes 23 and NACE codes 85.1-85.3, respectively) and a comparison group of other "professional" workers. Compared to a cross-section of the working-age population, teachers tend to have higher educational attainment. To account for this and other differences in socioeconomic status, we restrict the comparison group to workers in other ISCO-08 category 2 occupations, excluding medical occupations (ISCO-08 code 22). The resulting comparison group, thus, includes science and engineering professionals (ISCO-08 codes 21), business and administration profession-10 . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101/2021.03.25.21254219 doi: medRxiv preprint als (ISCO-08 codes 24), information and communications technology professionals (ISCO-08 codes 25), and legal, social, and cultural professionals (ISCO-08 codes 26). 12 We exclude healthcare workers from all samples because they have a substantially different testing regime and exposure to COVID-19 during the sample period. To reduce computational demands, we collapse the data to the demographic cell-countyweek level. 13 In 2020, there were 11 counties in Norway. Table 1 presents summary statistics of our main estimation samples. Compared to other professionals, teachers are more likely to be female and less likely to live in the capital region (Oslo and Viken counties). Teachers and non-teachers have similar confirmed incidence of SARS-CoV-2 during the 12 week window around school re-opening. The two groups differ when it comes to testing, and teachers have an average of 83% more weekly tests around the time of school re-opening. There are relatively small differences in SARS-CoV-2 incidence between the two groups of parents, but test rates are higher for parents of older children. For students, infection and test rates are substantially higher for those aged 20-22 compared to those aged 17-19. Figure 1 plots trends for students and teachers. During most of the period of school closures, teenagers age 17-19 have lower rates of confirmed COVID-19 relative to older young adults (Panel (a)). However, incidence in the two groups converges and tracks closely through the reopening week. Incidence among parents (Panel (b)) and among teachers and other professionals (Panel (c)) tracks very closely before, during and after the period of school closures. In addition, incidence continues to trend very closely throughout the period for which we have data. The graph shows no divergence around the beginning/end of summer vacations (week 25/35), and we see no evidence of divergence during the fall when confirmed 12 In one of the robustness specifications, a wider range of workers are applied as the comparison group to a wider definition of school workers. 13 For students, data are collapsed by week, age, county, and gender. For parents, data are collapsed by age of youngest child, week, own age (in 5-year bins), county, and gender. For teachers, data are collapsed by occupation, industry, week, own age (in 5-year bins), county, and gender. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) Note: Table presents summary statistics of our three main estimation samples: Teenagers and young adults, collapsed at the age-gender-county-week level (columns 1 and 2), parents of teenagers and young adults, data collapsed at the age of youngest-age-gender-county-week level (columns 3 and 4), and teachers and non-teacher professionals -data collapsed at the industry-occupation-age-gender-county level (columns 5 and 6). Observations are weighted with the population in each cell. "C19 pos" refers to positive covid 19 tests, excluding cases contracted abroad. Weeks 14-25 correspond to the 10 week window around the time of school re-opening. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101/2021.03.25.21254219 doi: medRxiv preprint incidence rates are higher. Figure 2 plots the estimated event study models from equation (2) for students, parents, and teachers. For students (Panel a), the figure indicates non-parallel pre-trends. As shown in Figure 1 , incidence rates for young adults aged 20-22 and high school age teenagers aged 17-19 tended to converge in the last weeks of school closures. Consistent with these trends, the event study figure finds differential pre-trends that converge several weeks before reopening. Differential pre-trends are less pronounced for parents, and for teachers the pre-trends are close to zero and not statistically significant. The figure gives no indication of significant treatment effects. The estimated event study coefficients do not shift significantly after reopening for either high school students, parents, or teachers (pooled across all grade levels). Due to the age gradient in COVID-19 infections, high school students likely pose a larger infection risk than primary and secondary school students. Furthermore, they shift between specialisation groups throughout the week, giving more contact points and potentially facilitating transmissions. To test empirically if contagion is stronger in high schools, we estimate effects from teachers in high school and teachers in primary/lower secondary school separately. 14 Figure 3 shows these separate event study estimates. Among teachers in primary and lower secondary school (left panel), the pattern is much as in the main sample: no significant pre-trends and no infection effects from the re-opening. Among high school teachers, we see a relative increase in infection rates concentrated 2-3 weeks after reopening, as one would expect from the incubation period of COVID-19. However, the effect is estimated with low precision, and the event study coefficients are not statistically significantly different from zero in the post-period. The size of the effect peaks at about 10 infected teachers per 100,000. Reassuringly, the pre-trend does not deviate substantially or significantly from the pre-trend in the comparison group. 14 The data do not allow us to distinguish between primary and lower secondary schools. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101/2021.03.25.21254219 doi: medRxiv preprint Note: Figure shows weekly rates of positive COVID-19 tests for students by age group (high school age versus above high school age), for parents by age groups of youngest child (high school age versus above high school age) and for teachers versus other category 2 ISCO-08 occupations. 14 . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101/2021.03.25.21254219 doi: medRxiv preprint (2) with 95% confidence intervals. The dependent variable is the number of confirmed positive SARS-CoV-2 tests per 100,000 population. Week 19 and the comparison groups as reference. All models include week fixed effects. In Panel (a), models additionally include age fixed effects. In Panel (b), models also include controls for age of youngest child. For Panel (c), models also include controls for industry and occupation. Standard errors are clustered at the county level. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101/2021.03.25.21254219 doi: medRxiv preprint (2) with 95% confidence intervals. The dependent variable is the number of confirmed positive SARS-CoV-2 tests per 100,000 population. Week 19 and other ISCO-08 category 2 occupations as reference. Models include week, industry and occupation fixed effects. Standard errors are clustered at the county level. Table 2 presents estimates from the difference-in-differences specification of equation (1). For students, there is a significant increase in confirmed cases following reopening. Given the differential pre-trends presented in Figure 2 , this estimate should, however, not be given a causal interpretation. We find no statistically significant effects for parents. For teachers, we see no significant effects when grade levels are analysed jointly. The point estimate is close to zero, and the associated 95 percent confidence interval allows us to rule out increases in contagion larger than 2.1 weekly cases per 100,000 teachers. Consistent with the results in our event study models, we find that the effect of re-openings on teachers varies with grade level taught. For teachers at elementary and lower secondary schools (grades 1-10), we estimate a negative, but not statistically significant, effect of reopening. For high school teachers, school re-opening increased SARS-CoV-2 infection rates by 4.3 per 100,000 on average in the first 6 weeks following the re-opening. The effect is significant at the 5 percent level. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; Note: Difference-in-differences estimates from the specification in equation (1). All models include week, industry, and occupation fixed effects. In Panel (B), models additionally include group linear trends. Standard errors are clustered at the county level. *p < 0.10, **p < 0.05, ***p < 0.01 17 . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 To assess the robustness of the difference-in-differences estimates with respect to differential pre-trends, we estimate an augmented specification with a linear time trend interacted with treatment status. Results from this exercise are shown in Panel B of Table 2 . Including group-specific linear time trends substantially reduces the precision of the estimates. 15 The point estimates for students and parents switch signs, and the effect on students is no longer statistically significant. With group-specific linear time trends, we now estimate a significant negative effect of re-opening on teachers in grades 1-10. Taken at face value, this would suggest that re-opening schools reduced incidence in this group. However, the event study plots in Figure 3 suggest that this negative effect likely reflects diverging trends, as there was no corresponding drop in incidence. Adding group linear time trends, the estimated effect on high school teachers is no longer significantly different from zero. Our models find no significant effects of school closures on teachers pooled across elementary and secondary school levels. However, there could still be effects on people working in nonteacher occupations, e.g. support staff working in schools. Similarly, there may be effects on kindergarten workers. To assess this, we estimate additional event study models estimating treatment effects on two additional groups: (1) all workers employed in schools and (2) all workers employed in schools or kindergartens, in teacher or non-teacher occupations. In these models, the comparison group consists of all workers. Results from these models, presented in Appendix Figure A1 , find no indication that the reopening of schools and kindergartens increased the number confirmed COVID-19 cases among treated workers more broadly defined. 15 Though this specification is known to bias estimates in the presence of time-varying treatment effects, see, e.g. Borusyak & Jaravel (2017) . . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 6 Extensions and mechanisms 6.1 Effects on primary school students Our main models estimate effects of re-opening on upper secondary school students and their parents, using young adults just above school-leaving age as a control group. The re-opening of schools was implemented step-wise, with the youngest students (preschool through grade 4) returning to school 2 weeks before older students (grades 5-13). To assess whether the reopening of schools for the youngest students affected COVID-19 infections, we have estimated a set of event study models of confirmed COVID-19 incidence among fourth grade students and their parents, using 5th graders and their parents as a natural comparison group. 16 Results from this exercise are presented in Figure 4 . Due to the brief interval between the reopening of schools for the two age groups, we only have a 2-week period to assess the effects of reopening. With that caveat, we note that the results in Figure 4 are consistent with the conclusions from our main specifications. We find no indication that the earlier re-opening of schools for the younger children increased contagion rates among primary school students or their parents. One reasons for our modest effects, could be that infection levels among students are very low, or simply zero. Due to the exponential nature of infection, effects may be substantially stronger in contexts with even modestly higher infection. For relevance to contexts with higher overall infection, we explore effects in two contexts of relatively higher infection within our sample. (2) with 95% confidence intervals. The dependent variable is the number of confirmed positive SARS-CoV-2 tests per 100,000 population. Week 17 and fifth graders as reference. All models include week fixed effects. In Panel (a), models additionally include age fixed effects. In Panel (b), models include controls for age of youngest child. Standard errors are clustered at the county level. During the first wave of the pandemic, the confirmed incidence of SARS-CoV-2 was significantly higher in the Oslo region relative to the rest of the country. 17 In the 12-week window around re-opening, average weekly incidence was more than five times higher among high school age teenagers residing in the Oslo region relative to teenagers in the same age group in other counties (10.6 vs 2.2 per 100,000). Figure 5 presents event study models estimated separately for the Oslo region versus other counties. The confidence intervals are wider for the Oslo sample, likely reflecting the smaller sample size. 18 Meanwhile, we find similar point estimates for the Oslo region sample and the other counties. This pattern holds for both students, parents, and teachers. In other words, we find no indication that COVID-19 rates increase at re-opening in the relatively high incidence Oslo region. 17 We define the Oslo region as Oslo and Viken counties. 18 Note the change of scale in Figure 5 compared to the other figures. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; (2) with 95% confidence intervals. The dependent variable is the number of confirmed positive SARS-CoV-2 tests per 100,000 population. Week 19 and the comparison groups as reference. All models include week fixed effects. In Panel (a), models additionally include age fixed effects. In Panel (b), models include controls for age of youngest child. In Panel (c), models include controls for industry and occupation. Standard errors are clustered at the county level. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101/2021.03.25.21254219 doi: medRxiv preprint In our main analysis, we assess changes in COVID-19 rates around the re-opening of schools in week 20. We also estimate corresponding event study models for the initial school closure. As shown in Figure 1 , the level of SARS-CoV-2 transmission in society was higher in Norway at lock-down in March 2020 than at the reopening 2 months later. If our lack of significant effects of school closures are driven by low incidence rates, we might still expect to see a fall in confirmed COVID-19 rates after the initial lock-down. In contrast to reopening, the school closure was implemented simultaneously with several other measures, and we cannot rule out that some of these measures impacted teachers and students differently than the comparison groups. For instance, all organised sports and activities for children were closed with immediate effect, and parents were encouraged to limit the number of peers their children met. The initial lock-down was announced on March 12, only 15 days after the first confirmed case of COVID-19 in Norway. As a consequence, we have a limited period with data to estimate pre-trends. With that caveat, the estimated event study models, presented in Figure 6 , indicate parallel pre-trends for students, parents, and teachers. For students, the figure shows a drop in confirmed cases following school closures. However, this reduction in confirmed infections could be driven by changes in the likelihood of testing, as discussed below. For parents and teachers, we are unable to detect any significant shift in the estimated event study coefficients around the time of the policy change, that is, there is no appreciable differential reduction in COVID-19 rates for affected parents or teachers following school closures. The lack of effect also holds when we split teachers by grade level taught, i.e. we find no effects for high school teachers (see Appendix Figure A2 ). The weekly rate of confirmed SARS-CoV-2 is an imperfect proxy for the true underlying incidence of the virus, as we only capture cases among patients who are tested. Put differently, there is an unknown number of undiagnosed infections that are missing from our data. This . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 (2) with 95% confidence intervals. The dependent variable is the number of confirmed positive SARS-CoV-2 tests per 100,000 population. Week 11 and the comparison groups as reference. All models include week fixed effects. In Panel (a), models additionally include age fixed effects. In Panel (b), models include controls for age of youngest child. In Panel (c), models include controls for industry and occupation. Standard errors are clustered at the county level. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 missing data issue has two consequences for our analyses. First, incomplete testing may have implications for interpretation of our effect sizes. While our baseline estimates allow us to rule out post-reopening increases in confirmed cases among teachers greater than 4 weekly cases per 100,000 teachers, we cannot rule out additional increases in undiagnosed cases. We note that our measure is somewhat more robust for symptomatic cases, because persons with symptoms of SARS-CoV-2 were encouraged to get tested. Second, school closures and reopenings may have effects on the probability that affected groups get tested for the virus. If students or teachers are more likely to get tested when schools are open, this could show up in our models as an increase in confirmed cases even if there is no effect of reopening on underlying prevalence. Plotting the trends in testing indicate that testing patterns differ between treatment and comparison groups (see Appendix Figure A3 ). In particular, the rate of testing rose sharply for young adults aged 20-22 relative to 17-19 year olds at the timing of school closures and remained higher throughout the lock-down and re-opening (Panel A). Starting mid-lock-down, test rates among teachers rose sharply compared to other professionals (Panel C). Parents have similar trends in testing throughout lock-down and the weeks that followed (Panel B). Estimated event study models of testing fail to show a discontinuous shift at the time of school reopening for parents (see Appendix Figure A5 , Panel B). In other words, the reopening of schools did not increase testing among parents, the only group for which we have credible causal evidence. The findings from this paper add to an extremely limited literature on the safety of re-opening schools -and, consequently, keeping schools open -in a context of relatively low SARS-CoV-2 contagion. Using detailed register data covering the universe of Norwegian residents, we compare infection rates for groups with different exposure to in-person schooling across the (closing and) re-opening of schools during the 2020 spring lock-down in Norway. We show that infection rates among teachers in general are similar to those of comparable professions both 24 . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 before and after the re-opening of schools. The same holds for adolescents attending the last 2 years of high school compared to their recently graduated counterparts and for the parents of these two groups of teenagers and young adults. We do find a slight increase in infection rates for high school teachers compared to other professionals. However, this increase was not statistically significant across specifications, and the number of extra cases was small (about 4 per 100,000). Our results show that at low levels of contagion, about 5 cases per 100,000, reopening of schools can be done safely. Our particular quasi-experimental conditions suggests that it is possible to maintain low contagion if open schools are combined with other social distance measures. For policy makers weighing the pros and cons of school re-opening, a pressing question would be what the threshold for "sufficiently low" contagion is. We see no disproportionally stronger effects in the region with the highest contagion (Oslo region). It should be noted that infection rates are relatively low also in this region, at about 20 cases per 100,000. Beyond that level, we cannot give answers based on data. The exponential nature of contagion makes extrapolation difficult. Some care must be taken in the interpretation of our results as people in the non-teacher professional group and both young adults who have recently graduated and their parents, which are used to contrast infections among teachers, students and exposed parents, might also be infected due to exposure to their own children or friends attending school (or to a household member working at a school). This would cause a downward bias in our estimates. However, as these people would be secondary infected, the primary effect of schooling is still evident in the event study graphs. Hence, our results for teachers should be interpreted as the added risk of being a teacher, rather than infection level overall from opening schools. So far, the limited evidence in this field has found that in-person schooling contributes little to the spread of SARS-CoV-2. The results in this paper are in line with those findings. The majority of studies are based on infection numbers from spring or summer 2020. As the fall and winter of 2020-2021 has seen a surge in infections across Europe and the US, with a larger share of infections among adolescents and young adults, an important next step is to 25 . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101/2021.03.25.21254219 doi: medRxiv preprint assess infections from in-person schooling under these higher infection rates. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) Evidence from Sweden's partial school closure', MedRxiv . 28 . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 Appendix A 29 . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (2) with 95% confidence intervals. The dependent variable is the number of confirmed positive SARS-CoV-2 tests per 100,000 population. Models include occupation, industry and week fixed effects. The comparison group includes all workers not employed in schools or kindergartens. Week 19 and the comparison groups as reference. Standard errors are clustered at the county level. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (2) with 95% confidence intervals. The dependent variable is the number of confirmed positive SARS-CoV-2 tests per 100,000 population. Week 11 and the comparison groups as reference. The models include week, industry and occupation fixed effects. Standard errors are clustered at the county level. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Note: Figure shows trends in weekly rate of SARS-CoV-2 tests for students by age (high school age versus above high school age), parents (youngest child in high school age versus youngest child above high school age) and teachers versus other professionals (category 2 ISCO-08 occupations). . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (2) with 95% confidence intervals. The dependent variable is the number of SARS-CoV-2 tests per 100,000 population. Week 11 and the comparison groups as reference. All models include week fixed effects. In Panel (a), models additionally include age fixed effects. In Panel (b), models include controls for age of youngest child. In Panel (c), models include controls for industry and occupation. Standard errors are clustered at the county level. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (2) with 95% confidence intervals. The dependent variable is the number of SARS-CoV-2 tests per 100,000 population. Week 19 and the comparison groups as reference. All models include week fixed effects. In Panel (a), models additionally include age fixed effects. In Panel (b), models include controls for age of youngest child. In Panel (c), models include controls for industry and occupation. Standard errors are clustered at the county level. . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 26, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 Inequality in household adaptation to schooling shocks: Covid-induced online learning engagement in real time Revisiting event study designs Minimal transmission of sars-cov-2 from paediatric covid-19 cases in primary schools, norway Covid-19 and school closures Low-income children and coronavirus disease 2019 (covid-19) in the us Covid-19 in children and the role of school settings in covid-19 transmission Learning inequality during the covid-19 pandemic Covid-19 and schools: What we can learn from six months of closures and reopening School re-openings after summer breaks in germany did not increase sars-cov-2 cases Infection prevention guidelines and considerations for 27 the author/funder, who has granted medRxiv a license to display the preprint in perpetuity