Study population and setting
The authors considered per-day growth rates in reported SARS-CoV-2 infections at the subnational level within China, the United States, Italy, France, South Korea, and Iran from the beginning of the epidemic to April 6, 2020. The authors used panel regression with fixed effects for each region to examine how daily growth in cases, modeled as exponential, changed in response to 1,717 region-specific policy deployments, the timing of which was assumed to be exogenous. These non-pharmaceutical interventions included travel restrictions, social distancing edicts, lockdowns, and other policies such as expansion of sick leave benefits. Authors considered changes in case definitions within regions. The magnitude of estimated policy effects was expressed in confirmed cases averted by comparing observed cases to those that would have been observed under a counterfactual scenario of no intervention (calculated by using region-specific pre-intervention growth rates). Projections are made beyond the exponential growth phase by incorporating effect estimates into an SIR (Susceptible, Infected, Recovered) mathematical model to account for depletion of susceptible individuals.
Summary of Main Findings
New infections across the six countries before any interventions grew at an average rate of 43% per day, corresponding to a doubling time of ~2 days. The combined effect of all interventions was estimated to substantially lower infection growth rates within each country: the estimated reduction in daily growth rates that would occur in response to the simultaneous implementation of all interventions enacted by a given country ranged from 16% in France to 49% in South Korea. Effects estimated for individual country-specific interventions were smaller in magnitude and had greater uncertainty: 22 of 29 interventions had point estimates suggesting effectiveness in reducing infection growth rates, though many of these confidence intervals included no effect. The actual implementation of all policies across all six countries, compared to a counterfactual scenario of no intervention, was estimated to have prevented ~62 million confirmed cases (corresponding to ~530 million total infections, accounting for under-ascertainment) by April 6, 2020. Results were generally similar under a number of validity checks and sensitivity analyses, including withholding regional blocks of data and inclusion of a lagged effect. In sensitivity analyses, varying the infection fatality rate and incorporating an SEIR (Susceptible, Exposed, Infected, Recovered) mathematical model led to higher estimates of cases averted.
The study incorporated a large amount of epidemiologic and policy-related data at the sub-national level, including dates of implementation for 1,717 specific policies, which are presented clearly in supplementary materials. Models took advantage of exponential growth exhibited during the initial phase of the pandemic, and used a reasonable approach in comparing pre-intervention to post-intervention growth rates in each region. Changes in testing protocols and therefore in case detection are explicitly addressed via terms in models. The authors performed a broad array of robustness checks that included varying epidemiologic parameters and policy groupings in their model. Assumptions are clearly described and the presentation of models and results is thorough and transparent.
To infer that policies affected infection growth rates, the authors made the crucial assumption that policy implementation dates were not determined by infection growth rates themselves (i.e., they are exogenous). While the authors make a reasonable argument for why this may be so, violation of this assumption cannot be ruled out. The analysis also assumes that individual behavior changes did not confound the results (for example, if people in a region increased social distancing around the time of policy enactment, but not because of these policies per se). Partial confounding by this individual risk-avoiding behavior also cannot be ruled out, though it should be noted that there is little published evidence to date of such an effect. Although the assumption of exponential growth is reasonable early in the epidemic, if there was considerable heterogeneity in susceptibility in a given region, slower growth in cases would occur earlier than might be expected under homogeneous susceptibility (for example, if infections quickly reached high prevalence in particularly susceptible groups such as essential workers), introducing the possibility of some confounding. Unmeasured changes in case detection also could have biased results, though sensitivity analyses performed by the authors suggested that this bias was likely small. Finally, the simultaneity of many of these interventions, and the high degree of heterogeneity in content, context, and implementation, made it difficult to attribute effects to specific interventions.
This study incorporated data from six countries and over 1,700 national and regional policy interventions, and represents one of the most thorough and credible estimates of the effects of large-scale non-pharmaceutical interventions on the spread of SARS-CoV-2.
This review was posted on: 10 July 2020