School closures are among the most controversial policies used to fight the pandemic, as their costs loom large. Prolonged teaching disruptions affect not only children’s learning outcomes but also their psychological and emotional development, with children from low socio-economic backgrounds particularly hurt (Engzell et al. 2020, Grenewig et al. 2020). School closures are also detrimental for the careers of parents who take on more educational responsibilities and have to reduce the number of hours worked (Fuchs-Schündeln et al. 2020). Since women typically shoulder most of the childcare responsibilities, school closures may also widen the gender wage gap (Alon et al. 2020).
These costs of school closures should be weighed against the potential benefits of reduced transmission of the virus. In Germany, as in other European countries, the share of children among COVID-19 cases increased around the time the new school year started, and those events have often been linked to each other in the press. The fear that schools constitute ‘hotspots’ of transmission has been fueled by reports of outbreaks in schools. But to guide policy, it is important to zoom out from the anecdotal or small-sample evidence and identify the causal effect of school closures and reopenings on aggregate infections.
Causal inference is challenging here because the decision to close schools is typically made in response to rising infections in the local area or country. Moreover, in many cases, school closures were enforced at the same time as other containment measures (e.g. Dehning et al. 2020), making it difficult to isolate the impact of schools.
Quasi-experimental studies address both drawbacks by leveraging a source of variation unrelated to the course of the pandemic. In von Bismarck-Osten et al. (2021), we compare the evolution of infection rates from 415 districts across Germany that entered school holidays at different times in the summer and fall of 2020, in what is called an ‘event study’. This strategy makes a causal interpretation of the estimates possible, as the timing of holidays was set years in advance, unaltered by the pandemic, and did not coincide with other containment measures.
More specifically, we implement the event study on the country-wide data by applying the novel ‘imputation’ estimator by Borusyak et al. (2021). Recent literature has shown that conventional event study estimation – i.e. by Ordinary Least Squares (OLS) with unit and period fixed effects and some lags and leads of treatment – produces estimates that are not reliable in the presence of effect heterogeneity, and potentially even have the wrong sign (e.g. Goodman-Bacon 2020). The imputation estimator produces a desired average of treatment effects and has attractive efficiency properties relative to other robust estimators (e.g. de Chaisemartin and D’Haultfoeuille 2020, Sun and Abraham forthcoming).
We consider three events separately: the summer and fall closures and the summer re-openings. Comparing the dynamics of infection rates around each event yields a consistent picture, suggesting there was little impact made by schools.
The summer school closures do not appear to have had a containing effect on infections in either the school population or older generations in a period of low infection rates. A similar finding for the fall holiday closures suggests that school closures are no more effective at more advanced periods of the pandemic (with the raw data and our estimates shown in Figure 1).
Figure 1 The impact of the fall school closures on children aged 5–14
A) Raw data
Notes: Panel A displays smoothed daily cases per 100,000 in the age bracket 5–14 years. Districts are grouped by the fall-holiday starting date. Panel B displays the corresponding estimates of the effects of school closures on the same outcome (blue dots), together with a test for parallel trends between districts where schools closed at different times (red squares). Districts in Bavaria are excluded in Panel B.
In line with our results on school closures, we find concerns about the return to full-schooling capacity after the summer holidays to be unsubstantiated. Infections among children and adults did not rise with the start of the new academic year (a result also shown in a complementary paper by Isphording et al. 2021). Instead, infections appear to have increased in the last weeks of the summer holidays and declined in the days after reopening (see Figure 2). We consider this to be best explained by the higher risk of infection among families returning home from their travels shortly before the summer holidays ended, and the increased testing of those families.
Figure 2 The impact of the summer school reopenings on children aged 5–14
A) Raw data
Notes: Panel A displays the smoothed daily cases per 100,000 in the age bracket 5–14 years, grouping districts by the school reopening date after the summer holiday. Panel B displays the corresponding estimates of the effects of school reopening on the same outcome, including anticipation effects (blue dots), together with a test for parallel trends between districts where schools reopened at different times (red squares). Districts in Bavaria and Baden Wuerttemberg are excluded in Panel B.
We acknowledge that our study period precedes the emergence of allegedly more transmissible variants of the virus (such as the British B.1.1.7 strain) and precedes vaccination efforts, both of which are likely to affect the transmission risk in schools, though in opposite directions.
While it is not our domain of expertise to explain why schools appear to play a subordinate role in the spread of SARS-CoV-2, epidemiological studies hint at several potential explanations. One possibility is that the measures introduced in German schools to avoid contagion have been effective. Alternatively, children could be less susceptible to infection (Davies et al. 2020), or less contagious than adults (Jones et al. 2020). Further epidemiological evidence on the relative role of such mechanisms would complement the policy implications of our paper.
Alon, T, M Doepke, J Olmstead-Rumsey and M Tertilt (2020), “The shecession (she-recession) of 2020: Causes and consequences”, VoxEU.org, 22 September.
Bismarck-Osten, C v, K Borusyak and U Schönberg (2021), “The Role of Schools in Transmission of the SARS-CoV-2 Virus: Quasi-Experimental Evidence from Germany”, CReAM Discussion Paper 22/20, Economic Policy, forthcoming.
Borusyak, K, X Jaravel and J Spiess (2021), “Revisiting Event Study Designs: Robust and Efficient Estimation”, Working Paper.
Chaisemartin, C de and X D’Haultfœuille (2020), “Two-Way Fixed Effects Estimators with Heterogeneous Treatment Effects”, American Economic Review 110 (9): 2964–2996.
Davies, N, P Klepac, Y Liu, K Prem and M Jit et al. (2020), “Age-Dependent Effects in the Transmission and Control of COVID-19 Epidemics”, Nature Medicine 26(8): 1205–1211.
Dehning, J, J Zierenberg, F P Spitzner, M Wibral, J P Neto, M Wilczek and V Priesemann (2020), “Inferring Change Points in the Spread of COVID-19 Reveals the Effectiveness of Interventions”, Science 369(6500).
Engzell, P, A Frey and M D Verhagen (2020), “The collateral damage to children’s education during lockdown”, VoxEU.org, 9 November.
Fuchs-Schündeln, N, M Kuhn and M Tertilt (2020), “The short-run macro implications of school and childcare closures”, VoxEU.org, 30 May.
Grewenig, E, P Lergetporer, K Werner, L Woessmann and L Zierow (2020), “COVID-19 school closures hit low-achieving students particularly hard”, VoxEU.org, 15 November.
Goodman-Bacon, A (2020), “Difference-in-Differences with Variation in Treatment Timing”, NBER Working Paper No. 25018.
Jones, T, B Mühlemann, T Veith, G Biele, M Zuchowski, J Hoffmann, A Stein, A Edelmann, V Corman and C Drosten (2020), “An Analysis of SARS-CoV-2 Viral Load by Patient Age”, medRxiv, 1101.
Isphording, I, M Lipfert and N Pestel (2020), “School Reopenings after Summer Breaks in Germany Did Not Increase SARS-CoV-2 Cases”, IZA Discussion Paper No. 13790.
Sun, L and S Abraham (forthcoming), “Estimating Dynamic Treatment Effects in Event Studies with Heterogeneous Treatment Effects”, The Journal of Econometrics.