Social distancing to slow the U.S. COVID-19 epidemic: longitudinal pretest-posttest comparison group study

Introduction

Modeling estimates suggest that up to 80% of Americans will be infected with severe acute respiratory syndrome coronavirus 2 (SARS CoV-2) if no preventive interventions are implemented [1]. In response, U.S. state governments have applied social distancing measures in an attempt to limit its transmission and reduce morbidity and mortality from coronavirus disease 2019 (COVID-19). Such measures have been applied during prior pandemics, with moderate success [2-5], and are predicted to prevent a rapid, overwhelming epidemic in modeling studies [6]. Data on social distancing and its associations with the course of the COVID-19 pandemic are now beginning to emerge, although no studies have examined changes in COVID-19-attributed mortality as an outcome. Recent studies have shown slowed COVID-19 epidemic growth coinciding with reductions in internal movement [7,8], as well as after implementation of shelter-in-place orders [9-11]. Because of the economic and social costs associated with social distancing measures [12,13], there is immense value in using different modalities to quantify the extent to which they have benefits for epidemic control. Herein, we contribute to this body of evidence, using a longitudinal pretest-posttest comparison group study design to compare the daily growth rate of COVID-19 cases, and the daily growth rate of COVID-19-attributed deaths, before and after implementation of social distancing measures in the U.S. Our primary aim was to empirically estimate the public health impact of government mandated non-pharmacological interventions in the period after their initial implementation and prior to their recently staged relaxation.

Methods

Results

A complete list of dates of statewide social distancing measures, by type of measure and state, is contained in the Supplement (Supplement Table 1). During March 10–25, all 50 states and the District of Columbia implemented at least one statewide social distancing measure (Supplement Figure 1). The most widely enacted measures on the first date of implementation were cancellations of public events (34/51 [67%]) and closures of schools (26/51 [51%]). The first social distancing measures were implemented when the median epidemic size was 35 cases (interquartile range [IQR], 17-72).

Figure 1A shows the mean COVID-19 daily case growth rate mapped against the date of the first statewide distancing measures. At the date of implementation of the first social distancing measure, states had a mean daily case growth rate of 30.8% (95% confidence interval [CI], 29.1-32.6; Table 1), corresponding to a doubling of total cases every 3.3 days. From fourteen days prior, through three days after, implementation of the first social distancing measure, the mean daily case growth rate did not change (-0.2% per day; 95% CI, −0.6% to 0.3%; P=0.51). Beginning four days after implementation of the first statewide social distancing measure, the mean daily case growth rate decreased by an additional 0.9% per day (95% CI, −0.4% to −1.4%; P<0.001). This estimate corresponds to a mean daily case growth rate that had declined to 26.5% (doubling of total cases every 3.8 days) by day 7 after enactment of the first statewide social distancing measures, to 19.6% (doubling time of 5.1 days) by day 14, and to 12.7% (doubling time of 7.9 days) by day 21.

• Download figure
• Open in new tab

Figure 1A-B.

Scatter plots and predictive margins with 95% confidence interval derived from regression models of the daily COVID-19 growth rate pre vs. post-implementation of the first statewide social distancing measures (1A) and statewide restrictions on internal movement (1B). The red line indicates the date of implementation in each state. The green dashed line is 4 days after implementation of the social distancing measure.

As of May 1, nearly all (45 [90%]) states had implemented a statewide restriction on internal movement. These restrictions on internal movement were implemented a median of 11 days after the first statewide social distancing measure (IQR 8-15) was implemented in the respective states, when the median epidemic size was 937 cases (IQR, 225-1414). The mean daily case growth rate had already been declining at a mean rate of −0.8% per day during the 14 days prior to implementation of statewide restrictions on internal movement (95% CI, −0.9% to −0.7%; P<0.001) (Table 1, Figure 1B). There was a drop detected three days after statewide restrictions on internal movement were implemented (-3.1%; 95% CI, −4.7% to −1.5%, P<0.001), but no statistically significant difference in the rate of change from before versus after their implementation (0.1% per day; 95% CI, −0.04% to 0.3%, P=0.14). As discussed in more detail below, there is substantial difficulty in disentangling the unique associations with statewide restrictions on internal movement from the unique associations with the first social distancing measures.

In the analysis of the secondary outcome, change in daily COVID-19-attributed deaths, given the uncertainty in the hypothesized lag between implementation of social distancing and observed changes (if any) in daily COVID-19-attributed deaths, we explored a range of thresholds. As shown in Table 2, by seven days after implementation of the first statewide social distancing measure, the mean daily growth rate in COVID-19-attributed deaths decreased by 2.0% per day (95% CI, −3.0% to −0.9%; P<0.001). By 14 days the estimated association was no longer statistically significant (-1.0% per day; 95% CI, −0.2% to 0.1%; P=0.09). No additional statistically significant benefit was found after 7 days of implementation of statewide restrictions on internal movement.

Sensitivity analyses suggested our estimates were not sensitive to inclusion of additional covariates, did not differ by size of the epidemic at implementation, and were consistent with the known incubation period (Supplement Tables 2, 3, and 4). In the event study specification (Supplement Figure 2), mean daily case growth was negative by day 4, and the estimates were statistically significant by day 8, consistent with the primary analysis. The event study analysis for change in daily COVID-19-attributed deaths also produced estimates qualitatively similar to the primary analysis, although with slightly larger confidence intervals given the smaller number of events (Supplement Figure 3).

Discussion

In this longitudinal pretest-posttest comparison group study, we found that implementation of social distancing measures was associated with a reduction in the mean daily growth rate of COVID-19 cases and in the mean daily growth rate of COVID-19-attributed deaths. Our estimates imply a more than doubling in the doubling time (from 3.8 days to 8.0 days) by three weeks following the implementation of social distancing measures. Assuming a cumulative epidemic size of 4,125 reported cases (equivalent to the cumulative number of cases in the U.S. at the time of implementation in each state), the reduction in growth rate we estimated corresponds to a difference between 26,281 reported cases with no social distancing versus 24,625 reported cases at 7 days after implementation, a difference between 158,518 reported cases with no social distancing versus 102,223 reported cases at 14 days after implementation, and a difference between 904,773 reported cases with no social distancing versus 283,161 reported cases at 21 days after implementation. Stated differently, our model implies that social distancing reduced the total number of reported COVID-19 cases by approximately 1,600 cases at 7 days after implementation, by approximately 56,000 reported cases at 14 days after implementation, and by approximately 621,000 reported cases at 21 days after implementation.

These results are consistent with both the theoretical effect of social distancing on epidemic spread [6] and the historical benefit observed with the implementation of such interventions during prior epidemics of communicable diseases [28]. They also are largely in keeping with recent data on the impacts of social distancing measures in the U.S. on both reduced mobility [7,8], and case growth rates [9-11], with generally similar effect sizes. Our study extends this literature by further examining COVID-19-attributed mortality as an outcome. The associations between social distancing and case growth rates were most apparent at the lower end of the incubation period that has been estimated based on publicly available data, with some evidence that the change in growth rates may have started even earlier. We suspect that this observation may have resulted from self-imposed social distancing that has reportedly occurred prior to government-issued mandates [33].

Our findings should be interpreted with the following limitations in mind. Our estimates would be biased toward the null if: 1) state and local governments had intensified social distancing measures in response to a worsening epidemic, 2) there were substantial violations of the stable unit treatment value assumption (e.g., workplace closures of large employers that had spillover effects across state lines), or 3) surveillance and testing intensified during the study period (thereby resulting in increased case reporting). Moreover, statewide restrictions on internal movement were often implemented after other social distancing measures had already been applied, further biasing our estimate toward the null. Estimates of cases and deaths in our model include those that are both laboratory-confirmed and suspected by health departments, but they are both likely to be under-estimates, due to limitations in testing, the presence of asymptomatic cases, and the occurrence of deaths that are not attributed to COVID-19 [34,35]. Nonetheless, our analyses focus on day-to-day changes in the growth rate of cases and deaths, so underreporting would only bias our results if reported versus true outcomes systematically differed from prior to versus after the implementation of social distancing measures. In contrast, our projected estimates of cases prevented are likely to be highly conservative, because they are modeled based on reported cases. While some studies have suggested that cases have been under-reported in the U.S. by 1-2 orders of magnitude [36], deaths appear to be under-counted by approximately 25% [34]. Finally, our estimates of the associations between social distancing and changes in mean daily case growth rates (and the corresponding number of cases averted) cannot be extrapolated linearly beyond the 21-day post-implementation analysis window. At the time of this analysis, states had begun to enact social relaxation measures, which necessarily prevents drawing conclusions about the long-term associational effects of these interventions in isolation.

Our study does not ascertain which of the statewide social distancing measures were most effective in reducing mean daily COVID-19 case growth. A variety of social distancing measures were used, often simultaneously, making it difficult to disentangle their independent associations. It is also possible that some state residents changed their behaviors in response to local (e.g., county-level) social distancing measures enacted prior to the statewide measures, or that some state residents changed their behaviors independently of social distancing measures enacted at any level of jurisdiction. The latter phenomenon has been observed empirically in other settings [37,38]. Such behavior change would be consistent with our sensitivity analysis suggesting a signal that extends back to a 0-day incubation period.

Finally, our analysis cannot answer questions about the appropriate time for rescinding social distancing measures. While our findings demonstrate that early mitigation efforts have yielded a substantial population health benefit, these benefits should be weighed against their costs. The costs of social distancing are likely to exacerbate the confluence of longstanding economic, social, and health decline that is already occurring in the U.S. [39-41], brought into even sharper relief given emerging data about racial, ethnic, and socioeconomic disparities in the incidence of COVID-19-related burden [42,43]. Moreover, there is currently robust debate about the extent to which legal authority to initiate or rescind social distancing measures resides with the federal or local government [44]. Additional study may be possible in the coming weeks to months if some statewide social distancing measures are relaxed while others are retained, or if measures are relaxed in some local jurisdictions while being retained in others. The use of cross-country data may also be helpful in this regard, albeit limited by any substantial within-country heterogeneity (such as has been observed in the diversity of uncoordinated U.S. federal and state responses to the epidemic).

Our finding that implementation of statewide social distancing was associated with a reduction in the mean daily growth rate of COVID-19-attributed deaths should be interpreted with more caution, given the uncertainty in published estimates about the median time from symptom onset to death [19-21]. The strongest associations we estimated, in terms of both statistical significance and magnitude, occurred at seven days after implementation. While this change could be reflective of what is known about time to death for the median hospitalized patient, other plausible thresholds were not associated with a statistically significant reduction in the mean daily growth rate of COVID-19-attributed deaths.

The finding that implementation of statewide restrictions on internal movement was associated neither with a statistically significant reduction in the mean daily growth rate of COVID-19 cases nor with a statistically significant reduction in the mean daily growth rate of COVID-19-attributed deaths warrants additional scrutiny. Identification of the unique effect of statewide restrictions on internal movement, separate from the initial social distancing measures, is a difficult undertaking given that they were implemented a median of 10 days later. There were no states in which statewide restrictions on internal movement were implemented without the prior implementation of other social distancing measures (e.g., cancellations of public events and closures of schools). As a result, we could not test hypotheses about the independent causal effects of statewide restrictions on internal movement, but rather could only ascertain the extent to which they were associated with additional reductions in the COVID-19 case growth rate, i.e., beyond those generated by the initial statewide social distancing measures. Moreover, the null finding should be interpreted in light of the other likely biases toward the null previously discussed.

In summary we demonstrate that the U.S. COVID-19 epidemic growth rate began to decline within approximately one incubation period following the initiation of statewide social distancing measures. Exploratory findings suggest that social distancing may also have had lagged mortality benefits. Future work should consider the potential public health benefits of relaxing these measures and, specifically, to what extent and how differential relaxation of certain measures promotes persistent epidemic control.