Estimating the Effect of Social Distancing Interventions on COVID-19 in the United States

Andrew M. Olney; Jesse Smith; Saunak Sen; Fridtjof Thomas; H. Juliette T. Unwin


Am J Epidemiol. 2021;190(8):1504-1509. 

In This Article


Social-distancing interventions are important for limiting the spread of SARS-CoV-2. To our knowledge, we are the first to apply an established, semimechanistic Bayesian hierarchical model of these interventions to the spread of SARS-CoV-2 from Europe to the United States. We estimated the effect of interventions across all states, contrasted the estimated Rt values before and after lockdown for each state, and contrasted predicted numbers of fatalities with actual numbers of fatalities as a check on the model's validity. Overall, school closures and lockdown were the only interventions modeled that had an estimated effect with a 95% credible interval that was not close to zero (i.e., no effect). No state had an estimated Rt below 1.0 before lockdown, but 29 states reached an Rt below 1.0 after lockdown. The model's ability to successfully predict deaths supports the validity of estimated intervention effects. These results suggest that reversal of lockdown without implementation of additional, equally effective interventions will enable continued, sustained transmission of SARS-CoV-2 in the United States.

Our study has several limitations. First, the assumption that all interventions have the same implementation and effect in all states is a simplifying assumption with clear exceptions. For example, the public events intervention that banned gatherings of 100 persons or more could include bans on, for example, 10 persons or more or 50 persons or more; however, it is unlikely that such bans are truly equivalent. The schools or universities closing metric treats primary, secondary, and higher education the same, though emerging evidence suggests that younger children may be less effective at spreading the virus than adults.[17] This limitation has since been partially addressed in the European model[8] by allowing random effects for lockdown only. Second, the assumptions that interventions are binary, instantaneous, and nonharmful are oversimplifications that do not account for time-varying compliance with interventions or unintended consequences. Using mobility data as a measure of population mixing[14,18,19] partially addresses this. Third, the parameters of the model were estimated using reasonable, but still uncertain, assumptions about prior distributions. We have used the same assumptions as in the European model, but these assumptions may be contradicted by future empirical work.

Modeling of SARS-CoV-2 transmission is emerging and rapidly diversifying; it includes classical susceptible-exposed-infectious-removed models and derivatives,[20] deep learning,[21] and piecewise models for subexponential growth.[22] State and local governments are likewise rapidly adjusting policy decisions regarding interventions based on case data and economic concerns. As the United States adopts an increasingly fragmented response to SARS-CoV-2, modeling approaches like ours that focus on shared interventions may not be tenable. Although our results give valuable insights into which interventions did and which did not change the transmission rate substantially, we recommend that future studies measure the change in behaviors resulting from interventions and then strengthen the predictive relationships between these behaviors and disease transmission.