Data
COVID-19 cases. Data on cases and deaths come from the Johns Hopkins Coronavirus Resource Center (https://coronavirus.jhu.edu/). The data are published daily since January 1st, 2020. The data allow us to compute the cumulative deaths and cases on a daily basis. To compare across countries of different population sizes, we also count cases per million inhabitants or total cases per million inhabitants to build our main dependent variables.
In the analysis of short-term NPI effects on COVID-19, we account for the real trend of new cases and fewer cases reported during weekends by measuring the mean of cases per million inhabitants reported in the last seven days. For a given day t and a country c, the number of new cases per million inhabitants is thus computed as
with
l the number of lags and
Cases the reported cases per million inhabitants. We apply the same smoothing procedure for total cases. We use these smoothed variables to compute the new case rate as
The new case rate is a good proxy for the short-term pandemic scale and is used as a dependent variable.
In the long-term analysis, we assess NPI effects by focusing on total cases per million inhabitants, which are good proxies for government capacity to contain the pandemic during a period. In the model in which we use country-level cross-sectional data, we use the average total cases per million between the 90th and 120th days after the first reported case as the dependent variable; this is to control for the differences in pandemic trends across countries. Our assessment thus compares countries within the same time interval.
Non-pharmaceutical interventions and the timing of their implementation. Data on NPIs are obtained from the Response2covid19 dataset.22 We particularly consider six NPIs that have strong potential effects on virus containment. The six NPIs considered are mask mandates, international travel restrictions, domestic lockdowns, mass gathering bans, restaurant closures, and school closures. Mask mandates are obligations to wear masks in public spaces. International travel restrictions are bans on international flights or border closures, except for cargo flights or repatriation. Domestic lockdowns are stay-at-home orders. Mass gathering bans refer to limitations on public or private gatherings. Restaurant and school closures refer to these facilities’ closures.
The NPIs are coded in a three-scale format. NPIs can be mandated for the entire population (“strict” intervention), for only a subpopulation (“partial” intervention), e.g., in a localized area or for a given category of the population, or not implemented at all. This differentiation is essential because strict NPIs will be useless in some cases, for example, if the virus is contained in a given region. The categorization allows us to compare different clusters of countries—being strict vs. being partial in implementing NPIs vs. not implementing. NPIs are coded 0 if not implemented, 0.5 if partially implemented, and 1 if strictly implemented.
We consider different timeframes to analyze short-term NPI effects. Because of the incubation period, NPIs often do not have immediate effects. To consider their short-term effects, we use different lags – 5, 9, 12, 21, and 30 days. The first lags – between 5 and 12 days – are selected because most patients experience symptoms between the 5th and 12th days.23 The other lags, 21 and 30 days, are used to see if the NPIs have more lasting effects.
In the long-term analysis, we focus on NPI timing, an essential factor in virus containment.1,9 To operationalize this, we include dummies capturing whether a country has implemented a given NPI early on. We define early implementation as having the NPI in place within 14 days after the first reported case, as this is the global median for all six NPIs (see Fig. 1 in the main document).
Control variables. In the short-term analysis, we use various controls capturing the pandemic’s dynamics. Some controls are standard in SIR (Susceptible, Infected and Recovered) epidemiological models. For example, we account for the smoothed cumulative total cases and deaths reported at date t in a given country c. We also add two extra controls to improve model estimation as they indicate where a country stands at a particular time during the pandemic: We first include the lagged value of the dependent variable, which is the new case rate. We also include the logged number of days since January 1st, 2020, a time trend accounting for the timing of the virus’s worldwide spread, and the importance of knowledge and experience in containing the virus.
In the long-term analysis with country-level cross-sectional data, different socio-economic variables from the World Bank are included. Following recent studies,24–27 we include several variables capturing national healthcare capacities —hospital beds per 1,000 inhabitants and health expenditures as a percentage of GDP in 2018. We also use controls capturing the population’s health risk—the percentage of the population with diabetes and the percentage being overweighed. We also add the median age of the population to indicate its sensitivity to COVID-19. Note that this variable is correlated strongly with % of older population aged 65 or more, with r = 0.913 in our main estimation reported in Fig. 3. We thus do not include percent of older population aged 65 or more. We finally include controls capturing a country’s developmental level—GDP per capita and government effectiveness.
Regression techniques
The short-term NPI effect on COVID-19 infections is assessed using fixed-effects regressions. As we use longitudinal panel data, the within estimator allows us to model each country’s mean as a group-specific fixed quantity. This specification has two advantages. First, the fixed-effects model controls for all the time-invariant country-specific features such as experience in handling pandemics or the overall quality of government and health systems. Second, the within estimator relates changes in the dependent variable to effects from time-variant characteristics such as NPI adoptions.
The long-term NPI effect on COVID-19 infections is first assessed using country-level cross-sectional data with ordinary least squared regressions. To account for unobservable, beyond-country factors that are associated with both the independent and dependent variables, the model includes continent fixed-effects (Africa, America, Asia, Europe and Pacific). While this method is useful for comparing cross-country variations, it still does not control for other unobservable confounding factors specific to each country. For instance, countries adopting mask mandates early on may have characteristics different from others, such as culture, prior pandemic experience, domestic mask production capacity, and public information campaign capacity, among others, which are not captured by the five continent-fixed effects. If so, the estimated relationship could be due to such unobservable factors rather than early mask mandates. Besides, since a cross-sectional analysis only focuses on differences between countries, it does not tell how COVID-19 infections would be different if a country adopted the mask mandate earlier than other NPIs. Hence we assess long-term NPI effects on infections by using panel data with longitudinal least squared regressions methods. Here, the models include country fixed-effects or random-effects.
In this panel data structure using within-country variations (i.e., country fixed-effects models), each NPI adoption captures a shock at only one point in time, namely, the date the mandate came into effect nationally. Yet the dependent variable “total cumulative infections per million” evolves daily for all data points. Given that the two sets of variables were measured at different times, and the country-level factors had remained constant in the fixed-effects model, it is challenging to estimate the long-term effect of early mandate adoption. In other words, the model estimates the pure effect of NPI adoption timing using the variations within countries while taking into account differences in their inherent capacities for implementing such an instrument.
All models are estimated with robust standard errors clustered at the subcontinent level, which is at a greater geographical scale than a country, to avoid potential downward biases.28 The rationale is that the error terms may not be perfectly independent of other neighboring countries in the same subcontinent given that there may be more significant virus transmission and policy diffusion among proximate countries.29,30 We thus use 19 subcontinents: (1) North America, (2) the Caribbean and Central America, (3) South America, (4) East Asia, (5) Southeast Asia, (6) South Asia, (7) Central Asia, (8) West Asia, (9) Oceania, (10) Northern Europe, (11) Eastern Europe, (12) Southern Europe, (13) Western Europe, (14) Eastern Africa, (15) Southern Africa, (16) Northern Africa, (17) Central Africa, (18) Western Africa, and (19) the Middle East. This subcontinent-level clustering generates more conservative standard errors than country-level clustering while still allowing for considerable degrees of freedom.