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


States implemented the 6 interventions at different rates. The mean period between the first and last intervention of a state was 18.64 (standard deviation (SD), 6.51) days (range, 4–31). The mean number of directives (e.g., executive orders) issued to implement interventions in a state was 4.32 (SD, 0.94; range, 2–6), and the mean number of interventions per order was 1.40 (SD, 0.35; range, 1–3). Some interventions were more likely to co-occur in a single directive than others, with banning of sports events (mean = 1.08 (SD, 0.83)) and of public events (mean = 1.04 (SD, 0.81)) occurring the most frequently with other interventions and schools or universities closing (mean = 0.40 (SD, 0.81)) and lockdown (mean = 0.14 (SD = 0.50)) occurring the least frequently with other interventions. Despite these differences, 96.33% of the interventions were implemented across states, with lockdown being the least implemented (n = 43). The decision to implement lockdown was not clearly data-driven across states: On the date of the last intervention, there was no significant difference in the cumulative case rate (P = 0.052) or the cumulative death rate (P = 0.059) using 2-sided rank-sum tests when comparing states that implemented lockdown with those that did not.

The mean IFR across states was 1.11% (SD, 0.12%; range, 0.76%–1.35%). Because confirmed case fatality data across states increased dramatically over the time period examined, similar statistics are not reported for these data.

The estimated national intervention effects on Rt are shown in Table 1. It is evident that only the closing schools or universities and lockdown had a nontrivial impact on Rt, with mean relative reductions of 23.7% and 54.4%, respectively. Moreover, schools or universities closing and lockdown were the only interventions with 95% credible intervals were not close to zero.

State-level measures and estimates of the models are shown in Table 2 (see also Web Figure 1). Of primary interest are Rt estimates before and after lockdown and the corresponding death counts forecasted 2 weeks into the future. Across states, the mean Rt before lockdown was 1.86 (SD, 0.56; range, 1.00–3.37), and the mean Rt after lockdown was 0.88 (SD, 0.25; range, 0.50–1.41). Notably, no state had a mean Rt below 1.0 before lockdown, but 29 states had a Rt below 1.0 after lockdown. Although lockdown was associated with a reduced Rt in all states that implemented it (a 54.4% reduction; see Table 1), in these specific 29 states, lockdown appears to have been the single critical intervention that led to containment of the disease. In the remaining states, the prelockdown Rt was too high (i.e., > 2.2) for lockdown to bring the Rt below 1.0.

Comparing the predicted number of deaths 2 weeks into the future with the actual number of deaths in each state served as a validity check on the model's estimates of intervention effects (see also Web Figure 2). Forty-five states (90%) had actual death counts that were within the 95% credible interval of predicted deaths. Notably, the mean numbers of predicted deaths were well above the actual number (>100 deaths) for Connecticut, New Jersey, Massachusetts, and New York. The mean absolute error of mean predicted deaths was 50.80, and without these 4 states included, the mean absolute error was 10.08. As expected, the model fit to actual deaths was even closer on the observed data, with a mean absolute error of 5.90 (n = 2,951).