Comparing the impact on COVID‐19 mortality of self‐imposed behavior change and of government regulations across 13 countries

Abstract Objective Countries have adopted different approaches, at different times, to reduce the transmission of coronavirus disease 2019 (COVID‐19). Cross‐country comparison could indicate the relative efficacy of these approaches. We assess various nonpharmaceutical interventions (NPIs), comparing the effects of voluntary behavior change and of changes enforced via official regulations, by examining their impacts on subsequent death rates. Data Sources Secondary data on COVID‐19 deaths from 13 European countries, over March–May 2020. Study Design We examine two types of NPI: the introduction of government‐enforced closure policies and self‐imposed alteration of individual behaviors in the period prior to regulations. Our proxy for the latter is Google mobility data, which captures voluntary behavior change when disease salience is sufficiently high. The primary outcome variable is the rate of change in COVID‐19 fatalities per day, 16–20 days after interventions take place. Linear multivariate regression analysis is used to evaluate impacts. Data collection/extraction methods: publicly available. Principal Findings Voluntarily reduced mobility, occurring prior to government policies, decreases the percent change in deaths per day by 9.2 percentage points (pp) (95% confidence interval [CI] 4.5–14.0 pp). Government closure policies decrease the percent change in deaths per day by 14.0 pp (95% CI 10.8–17.2 pp). Disaggregating government policies, the most beneficial for reducing fatality, are intercity travel restrictions, canceling public events, requiring face masks in some situations, and closing nonessential workplaces. Other sub‐components, such as closing schools and imposing stay‐at‐home rules, show smaller and statistically insignificant impacts. Conclusions NPIs have substantially reduced fatalities arising from COVID‐19. Importantly, the effect of voluntary behavior change is of the same order of magnitude as government‐mandated regulations. These findings, including the substantial variation across dimensions of closure, have implications for the optimal targeted mix of government policies as the pandemic waxes and wanes, especially given the economic and human welfare consequences of strict regulations.

of government policies as the pandemic waxes and wanes, especially given the economic and human welfare consequences of strict regulations.

lockdown, nonpharmaceutical interventions, salience, SARS-CoV-2, voluntary behavior change, Western Europe
What is known on this topic?
• Along with epidemiological data, analysts have tracked and published accounts of the nature, timing, and magnitude of government-mandated nonpharmaceutical interventions (NPIs) for many countries.
• A substantial literature provides initial evidence on which NPIs do and which do not constructively affect the course of the pandemic, for example, typically international travel restrictions appear to do so but stay-at-home orders do not as much.
• Much less analysis has addressed the extent to which voluntary behavior change also has an important role to play in the response to the pandemic.
What this study adds?
• The pandemic in Europe led people to substantially reduce their own risky behavior, resulting in reduction of COVID-19 mortality by an amount close to that of mandated NPIs.
• This suggests the value of government policies that enable or encourage voluntary NPIs (e.g., provision of free masks), as opposed to mandated NPIs (e.g., strict stay-at-home orders) which have a smaller benefit-cost ratio. being. This is much more than a global health crisis. 3 After the "first wave" of the epidemic receded in Western Europe, countries began to retrospectively examine their NPI policies, partly to assess when and how to reverse the school closure and movement restriction policies that have such substantial developmental and economic consequences, and partly to plan for subsequent epidemic waves.
The challenge, however, is that the method used to originally select the NPIs may be less helpful for actual evaluation. In the absence of real data or prior experience, the evidence base supporting the rollout of such unprecedented NPIs relied on mathematical forecasting models 4-9 drawing on input parameters for epidemiologic quantities like severity and attack rate, risk factors, and timing of transmission, for which empirical validation remains nascent. 10 These assumptions may have been inadvertently misleading, hence needing careful reassessment before being used as the basis for future decisions. For instance, with respect to school closures, a review of evidence from before COVID-19 11 as well as preliminary findings from Australia, 12 France, 13,14 and Ireland 15 suggest that school children-especially at primary level-may not be important drivers of coronavirus epidemics, in contrast to influenza, and school closure might play a substantially smaller role than the models had projected.
The need now is to retrospectively assess the true impact of NPIs on COVID-related morbidity and mortality, in order to optimize their implementation (or lack thereof) going forward, using empirical evidence.

| METHODS
We conduct a statistical analysis of the potential impact of NPIs, either government-imposed policies or voluntary behavior changes (before introduction of government policies), on COVID-19 deaths over March-May 2020 among 13 Western European countries.

| COVID-19 mortality data
Daily figures for new confirmed COVID-19 deaths by country were accessed through the European Centre for Disease Prevention and Control. 27 We used data for the 13 Western European countries with greater than 500 COVID deaths as of 16 May (Table 1) COVID mortality data were used because death constitutes a significant event; death certifications are less likely (than case notifications) to suffer from misclassifications; and the completeness of death data is far greater than that of case notification data due to varying testing capacity and accuracy across countries. However, (i) actual death tolls are still likely to differ from currently reported figures due to reporting issues, (ii) recording protocols can affect total numbers (e.g., whether deaths in nursing care homes are included), and (iii) reported date of death can be delayed from the actual date of death. Issues (i) and (ii) are mitigated here by focusing on relative changes in deaths, which also allow us to abstract away from total population size. Issue (iii) is mitigated in part by taking a 5-day moving average of deaths. Hence, as our dependent variable, we study the evolution over time of the following percentage change in smoothed daily deaths: where d i,t is the daily reported number of deaths in country i on day t.
To get a sense for the behavior of this variable Δ i,t , note that early in the pandemic, the number of deaths per day is typically rising, corresponding to the number of new infections having been growing a few weeks earlier, which implies that Δ i,t > 0. Late in the pandemic, when the number of daily deaths is declining, this percent change will be negative. In between, each day will yield approximately the same number of deaths, and hence, our dependent variable will be around zero. Smaller values are always better, since they imply a slower rise in fatalities (if positive) or a more rapid decline (if negative). Table A1 in the Supplementary appendix shows the distribution of values of this variable in our data, week by week.

| Nonpharmaceutical intervention data
For the interventions, we focus on two broad categories: government-imposed policies and regulations vs self-imposed and voluntary actions.
First, the Oxford COVID-19 Government Response Tracker provides dates and intensities for multiple categories of government policies across the globe. 28 Here, we focus on their "containment and closure" categories: school closing; workplace closing; canceling public events; restricting public events and gathering sizes; closing public transport; stay-at-home (or "shelter-in-place") requirements; and restrictions on internal movement and international travel. Separately, we add information on facial coverings (including formal regulations) from their "health measures" category. We define two alternate independent variables of the government closure measure: (i) an easy-to-interpret binary closure measure (i.e., 0 or 1) that occurs whenever broad stay-at-home restrictions are first promulgated and (ii) a continuous closure measure which is the sum of scores across all included categories, normalized by dividing by the maximum such score in the database. That is, each country was given a score (0, 1, 2, 3, sometimes up to 4 or 5 depending on the category) at each point in time, reflecting the stringency of any regulations in effect. We add those scores across all of the categories listed above and then standardize so that the maximum possible value is 1, in order for interpretation to be comparable to the binary measure.
Second, we also look at self-imposed restrictions on behavior which arose prior to the introduction of governmental interventions. Our primary measure, mobility decline, is based on Google's Community Mobility Reports, 29 which assess geographic mobility along different dimensions, as compared to a pre-crisis baseline within each country. The aggregated anonymized data come from every mobile device for which a user has signed in to a Google account and turned on their "Location History" setting. We construct an independent variable (dummy indicator) that switches from 0 to 1 in a given country when the mobility index is negative (representing activity being below baseline levels) and remains so thereafter, for all of the following three mobility categories: workplaces, transit stations, and retail and recreation (see, for instance, Figure 1 which presents mobility data for three illustrative countries, aggregated across these three categories). We do not consider residential mobility (defined as time spent at one's primary location) nor grocery and pharmacy activity, since that involves essential activity. Similar changes were observed in China early in the pandemic, where regional air pollution, indicative of reduced traffic and production, decreased after cases were reported locally but before any government restrictions had been imposed. 30 The mobility dummy indicator switches back from 1 to 0 when the government binary closure indicator turns on in that country because our goal is to evaluate the differential effect of unregulated behavior change. If binary closure takes place before self-imposed mobility decline (as in Austria and Germany), then the mobility variable remains equal to 0 throughout the study period. As with closure, we also define a continuous version of this mobility-independent variable equal to the normalized sum of mobility decline across the three relevant categories, on a given date. That is, we add up the percentage decreases across workplace, transit, and retail/recreation and then standardize so that the maximum value equals 1 and is comparable to the binary measure.

| Statistical modeling approach
First of all, evidently, none of these interventions, either regulatory or voluntary, will have an immediate effect on fatalities due to COVID-19.
Rather, we hypothesize that they will change the rate of new infections, leading to a change in deaths some time later. In order to model that delay, we assume that it is the sum of the incubation period, estimated to be 5 days, 5 and the period from symptom onset to death (for those who die), which has an observed median of 13 days. 31 Note that the overall typical time to death will be different from 13 days because in a growing epidemic, proportionately, more observations are from recent infections (some of whom will die later). We are modeling the observed data in the mid of a growing epidemic; hence, it is precisely the raw data that we need to match. Thus, we assume a median lag of 18 days from time of intervention to time of death. There is naturally some distribution for this lag, thus we employ a 5-day moving average of deaths, corresponding to lags from 16 to 20 days (Table 1).
Behavior i, tÀ18 To do this, we fixed t = 7, predicted the expected country-average growth rate Δ t¼7 ð Þ in each of the three scenarios respectively, and applied the following formula for doubling time: Finally, we estimated the number of COVID deaths that would have occurred in the first 7 weeks from the local starting date (t 0 ),  the simple salience indicator is necessarily coarser and more ad hoc than the mobility data. Alternative (iv) is again very similar to the primary specification, but by construction has fewer observations (so was not preferred). Note: 95% Confidence intervals are presented in square brackets. Specifications also included controls for t, t-squared, the percentage of population older than 65, the population density, and number of acute care beds per 100,000 people, and the date when the 5-day moving average of daily deaths is first equal to at least five. Standard errors are clustered at the country level. The unit of observation is a country-day: 1 day of data for a specific country. All indicators of government restrictions are as defined in the Oxford tracker and are normalized across an interval [0,1] for the 13 countries. N lower than in Models I-III in Table 2 because the lagged mobility data are only available for Italy from the third day of the epidemic. *Significant at 5% level. **Significant at 1% level. ***Significant at 0.1% level. Furthermore, we use the coefficients in Model I (see Table 2 deaths, for example, in Brazil, 42 France, 16 Sweden, 17  As one of a few studies to explore these issues empirically using substantial data, our analysis inevitably involves a number of limitations. Without randomization or other exogenous variation in the treatments, evidently, we cannot fully ascertain a causal link between the NPIs and the resulting changes in death rates. We do not expect any direct reverse causality, since future deaths will not change current behaviors. Current case rates could impact both variables, although as long as that effect is similar across dimensions it will not change the relative performance between voluntary versus government NPIs, or within the latter, which is our main outcome of interest.

| RESULTS
It is conceivable that some third variable, for instance, heightened media attention and scrutiny, could directly influence both government policies and individual behaviors. Future studies, using data on this and similar potential confounders, may be able to fully disentangle the various mechanisms at play.
Beyond endogeneity concerns, the quality of the fatality data may be subject to variation in reporting standards across countries, although this will be mitigated for the most part by focusing on rates of change rather than levels. Similarly, the quality of the government closure data is, although compiled independently without any apparent bias, somewhat subjective in nature as to the precise degree of severity in each category at each point in time. Meanwhile the mobility data, while more objective, does not capture the full range of voluntary self-protective behaviors (such as hand washing and maintaining personal distance). Our supposition is that these are all highly correlated with one another, but if this relationship differs substantially over time, then it could fail to be a good proxy for overall voluntary changes; it is not a priori clear in which direction this would affect the current results.
Our main messages are that NPIs can have significant impacts in reducing COVID-19 mortality and that almost half of this effect arises from simpler and more flexible voluntary interventions such as microlevel behavioral change, working remotely to the extent possible, and reducing discretionary travel, as opposed to stricter officially imposed regulations. Precisely, why that is we cannot say from our analysisother research has examined, for example, sociodemographic differences 46 -but the distinction is clearly important to countries at any stage of responding to the pandemic. Indeed this was suggested in a paper as early as March 2020: "Personal, rather than government action, in western democracies might be the most important issue." 47 These lessons are relevant around the globe, although cost-effective targeting and evidence-based policy are likely even more important for resourceconstrained countries with a weaker health and financial safety net.
ENDNOTES a We thank a reviewer for suggesting this.
b Recall t 0 is defined when the 5-day moving average of daily deaths is first equal to at least 5; for this counterfactual analysis, we normalized all countries to start at exactly 5 so that small changes in initial conditions (driven solely by the discrete nature of the threshold) would not arbitrarily affect the results.