key: cord-1043948-vn5pn1uh
authors: Loiacono, Luisa; Puglisi, Riccardo; Rizzo, Leonzio; Secomandi, Riccardo
title: Pandemic knowledge and regulation effectiveness: Evidence from COVID-19
date: 2022-02-23
journal: J Comp Econ
DOI: 10.1016/j.jce.2022.02.004
sha: c065ceab91b0b751931a80d2d6f35d63b340a4ac
doc_id: 1043948
cord_uid: vn5pn1uh

The spread of COVID-19 led countries around the world to adopt lockdown measures of varying stringency, with the purpose of restricting the movement of people. However, the effectiveness of these measures on mobility has been markedly different. Employing a difference-in-differences design, we analyse the effectiveness of movement restrictions across different countries. We disentangle the role of regulation (stringency measures) from the role of people's knowledge about the spread of COVID-19. We proxy COVID-19 knowledge by using Google Trends data on the term “Covid”. We find that lockdown measures have a higher impact on mobility the more people learn about COVID-19. This finding is driven by countries with low levels of trust in institutions and low levels of education.

JEL Codes: D7; E7; I18.

According to the latest data from the World Health Organization (February, 2022) , more than 396 million of COVID-19 infected cases have been reported, with more than 5.7 million deaths. 1 The pandemic has had a devastating impact on population health and well-being, and on the economy of countries across the globe (Levy Yeyati and Filippini, 2021) .

The World Health Organization announced the international outbreak of the COVID-19 infection on January 30, 2020, and declared the COVID-19 outbreak a global pandemic on March 11, 2020 (Cucinotta and Canelli, 2020) . Since then, the COVID-19 pandemic has reached nearly all countries around the world. However, the pandemic had largely heterogeneous effects, since countries have differed in their exposure to the virus, in the public and private response to it, and in the overall level of preparedness.

National governments have been implementing measures which restrict the movement of individuals (referred to, colloquially, as 'lockdown', a term we will also adopt throughout the paper) and impose social distancing on them. Interestingly, these measures display significant variation in their intensity, with some countries adopting stringency measures very early in the pandemic cycle, whereas others taking a less restrictive approach (Ferraresi et al., 2020) .

Of course, the purpose of these measures that restrict mobility and impose social distancing is to strongly reduce the spread of the virus, in order to contain the number of severe cases and deaths. From this point of view, policy makers and experts typically aim at avoiding an excessive pressure on hospitals and intensive care units, which would lead to a dramatic increase in the mortality of the disease. However, the accomplishment of this purpose not only depends on the design and timeliness of those coercive measures, but also on how citizens react to those measures, strengthening or weakening them with their individual course of action.

Interestingly, the lockdown measures have also been the subject of some controversy amongst political, legal scholars and the public. 2 Several demonstrations against lockdown have taken place in many countries in Europe 3 , in the US 4 , and elsewhere. It is unclear whether those protests are driven by impatience, a genuine knowledge that the lockdown measures are disproportionate to the pandemic threat, or simply an instance of aversion against an authoritarian turn in the actions of democratic and non-democratic governments alike.

Individual level reactions might be more compliant with government restrictions the more citizens are worried about the risks of contagion and of severe health outcomes. In turn, those perceived risks are affected by the information that citizens have about the pandemic, which they obtain by personal contacts and by being exposed to the mass media, both traditional and internet-based ones (namely, websites, and social networks). Recent literature has widely covered this topic across different domains. Mastrorocco and Minale (2018) find an effect of news media on crime perceptions. They use a difference-in-differences approach that compares individual perceptions of those with a wide range of available TV channels to those with limited choice. Scholars have analysed how perception and knowledge affect individual behavior in the case of political elections (Martinand and Yurukoglu, 2017) and crime (Shi, 2009; Spenkuch, 2018; Velásquez et al., 2020) . 2 In the UK, for example, the restrictions that underpin the COVID-19 lockdown measures have been challenged as being unlawful and disproportionate, breaching freedoms protected by the European Convention of Human Rights (Keene, 2020) . In this paper we investigate, at country level, the effects of stringency policies on citizens' daily mobility, taking into account a daily and country-specific measure of citizens' knowledge about the pandemic, i.e., the relative amount of Google searches about itself.

We exploit the staggered implementation of stringency measures adopted by countries over time, while controlling for country and daily fixed effects. We find that stricter lockdowns are significantly associated with lower mobility, and that this effect is larger the more people get information about the spread of COVID-19. These results survive a set of robustness tests, including the traditional event-study test à la Autor (2003) .

The remainder of the article is organized as follows: in Section 2 we lay down the empirical framework; in Section 3 we present the data, while in Sections 4 and 5 we discuss the results and perform some robustness tests, respectively. Finally, Section 6 summarises and concludes.

Our baseline empirical model builds on the large and expanding literature that makes use of the DiD method to investigate the net impact of a policy or a program on given outcomes.

The standard case for applying DiD is when an exogenous shock such as a lockdown measure (treatment) affects only a group of units (treated), in the presence of another group (control) which is similar in all respects but not affected by the intervention.

As noted in the introduction, while all countries eventually adopted lockdown measures in the year 2020 due to the COVID-19 outbreak, they differ in the timing of this adoption. This allows us to compare the change in the mobility index in the treatment group before and after the adoption of the policy with the corresponding changes in mobility that take place in the control group.

The estimated difference-in-differences (DiD) model is the following:

where is the Google mobility index for country c in day d; is the Stringency Index in country c and day d, ranging from 0-when lockdown measures have not been adopted yet-to 100, with 100 denoting the maximum level of lockdown; are daily variables at country level, such as temperatures, seven-days moving average of the pandemic confirmed cases per capita and the intensity of searches on Google of the term "Covid" for each country a week before 5 ; are country fixed effects that control for unobserved crosscountry heterogeneity 6 ; are daily fixed effects that capture time-specific shocks common to every country, such as Covid-related information that becomes available worldwide in a given day; is the error term, clustered at country level. In some specifications, we also control for country specific trends. Within this specification, γ is the DiD estimate of the (average) effect of the stringency on mobility.

To investigate whether there has been a heterogeneous response to containment measures as a function of the knowledge about COVID-19 on a given day in each country, we interact weekly Covid searches with the stringency measures. Covid searches ranges in each country from 0 -when there is no search in Google of the term "Covid" -to 100, with 100 denoting the maximum level of Covid searches.

The estimated model is a generalised version of Equation (1), taking the following form:

where our coefficient of interest which accounts for the impact of the interaction term ℎ , which is the indicator of Covid searches for country c and week w, and .

To measure the daily movement of people during the spread of COVID-19, we use the COVID-19 Community Mobility Reports provided by Google. 7 The mobility indicators measure the relative value of each weekday mobility, compared to the baseline value for that weekday, which in turn is calculated as the median value recorded during the 5-week period from January 3 to February 6, 2020, i.e., before the start of the pandemic. So, the indicator takes on a value of 100 if mobility in given day during the pandemic, say on a Helsingen et al. (2020) , we use observed data on mobility because they are more reliable than individual surveys due to the potential confounding role of individual biases in the way respondents self-report their behavior.

In order to deal with the COVID-19 outbreak, governments around the world adopted many and very different containment measures. We take into account the heterogeneity of governments' responses by making use of the Government Response Stringency Index (Stringency Index) developed by Hale et al. (2020) . The Stringency Index is calculated using the mean of nine metrics: school closures, workplace closures, cancellation of public events, restrictions on public gatherings, closures of public transport, stay-at-home requirements, public information campaigns, restrictions on internal movements, and international travel controls. Each of these variables is rescaled by its maximum value to create an overall score between 0 and 100. A higher score indicates a stricter response (i.e., 100 is equal to the strictest response). The index simply records the strictness of government policies, and it does not measure the effectiveness of a country's response.

We use data from the Google Trends tool to measure the time-varying pursuit of information about the pandemic by citizens. 9 As in previous works that use Google Trends to predict disease outbreaks (Carneiro and Mylonakis, 2009) , trading behavior in financial markets (Preis et al., 2013) and concern of public opinion about pension systems (Fornero, Oggero, and Puglisi, 2019) , we assume that Google search indicators provide reliable information about citizens' (search for) knowledge. The tool provides an index for online search intensity of a specific term (and its components) over the time period under consideration within a specific area. The index is a weekly measure of intensity, which is 9 We thank an anonymous referee for pushing us towards this interpretation of the Google Trends data.

computed as the number of weekly searches for the term divided by the maximum number of its weekly searches over the whole time period, in a given country. The result is scaled from 0 to 100, where 100 is the peak popularity and 0 means that there was not enough search volume for that specific term during that week. For our purposes, we collect searches related to the term "Covid" for the period from February to December 2020. In order to conduct a falsification test, we also collect searches related to the terms that were most searched worldwide on Google from February to December 2020, i.e. "translate", "porn", and "maps".

Notice that people's knowledge about COVID-19 might be strictly related to the amount of media coverage devoted to the issue. The link between media coverage and Google Trends searches has been emphasized by the literature, with specific reference to the pandemic: for example, Sousa-Pinto et al. (2020) show that Google Trends for COVID-19 symptoms such as cough, anosmia (loss of smell) and ageusia (loss of taste) are more strongly related to media coverage than to the underlying pandemic trends. 10 Interestingly, the authors find that peaks for the Google searches on the various symptoms occurred simultaneously, irrespective of the country's pandemic stage.

Summary statistics for all variables used in the analysis are reported in Table A2 of the Appendix. countries are averaged by day. The Stringency Index and Covid searches vary from 0 to 100 inside each country. The Mobility Index is equal to baseline value of 100 for each country, if on a given weekday it exactly equals its median value recorded during the first 5 weeks of 2020, i.e., before the start of the pandemic. Covid cases correspond to new active cases and are calculated as the difference in per capita cumulated cases between day t and day t-1. For more details, see Section 3.

The baseline specification, which includes Stringency Index, per capita confirmed cases 13 and country and time fixed effects, is reported in Column 1. Column 2 adds to the previous specification the temperature variable, which captures weather-related drivers of mobility. 14 In Column 3 we include as control variable Covid searches. The last three columns show the results that are based on different versions of Equation (2). In Column 4 we add the interaction between Covid searches and the Stringency Index, while in Column 5 we also 13 In Table A4 , we replicate the regressions in Table 1 by replacing the 7-days moving average of per capita confirmed cases with the 14 days lag of per capita confirmed deaths. The correlation coefficient between the 7-day moving average of confirmed cases per capita and the 14-day lagged confirmed deaths per capita is equal to 0.46. We do not observe any relevant difference in the coefficients of interest vis a vis the main specification. 14 Temperatures are retrieved from Global Historical Climate Network Daily (National Oceanic and Atmospheric Administration, 2020).

include country-specific linear trends. In Column 6, in order to check whether the potentially heterogeneous reaction to the Stringency Index depends on real world events rather than on citizens' knowledge about those events, we add the interaction between confirmed per capita cases and the Stringency Index itself. Finally, in Column 7, we include a variable capturing the so called "lockdown fatigue". Following Goldstein et al. (2021) , starting from the first day of the Stringency Index greater or equal to 74.5 (which corresponds to the 75th percentile), we build the lockdown fatigue variable by counting the number of days for which the Stringency Index was at least equal to 74.5. We include this variable in the regression both linearly and as a squared term.

In the first three specifications we find a negative and statistically significant relationship between mobility and stringency. The point estimates range from -0.489 to -0.398. This implies that, during the COVID-19 outbreak, the mobility in countries with stronger stringency measures decreases more than in those with weaker measures. Since the stringency variable measures the treatment intensity, we capture the impact of being treated by comparing the effect on mobility level when the stringency and Covid searches are at extreme values of their joint distribution. For instance, following the point estimates of Column 3, the mobility is reduced by approximately 11.41 percentage points when considering a shift from Uruguay, whose level of both the stringency measure and Covid searches are the closest to the 25th percentile value, to Dominican Republic, whose level of both the stringency measure and Covid searches are the closest to the 75th percentile value. 15

In Column 4 the coefficient on the interaction term Stringency Index*Covid searches is negative and statistically significant at the 1% confidence level, with a point estimate of -0.003, while it is 5% statistically significant in Columns 5 and 6. This implies that the magnitude of the effect of the stringency measures on mobility is stronger for higher level of COVID-19 knowledge, i.e., the effectiveness of stringency is amplified by the knowledge of the severity of the pandemic. On the other hand, the interaction of the stringency measure with the number of confirmed cases (Column 6) is not significant at ordinary confidence levels, while the interaction of stringency with Covid searches remains significant and with the same magnitude. This suggests that the role of Covid searches in determining the impact of stringency on mobility appears to be relevant and the real world events that are connected with the evolution of the pandemic by itself do not matter.

In Figure 2 

How to explain the fact that the interest in the pandemic -as proxied by Google searchesappears to affect the compliance with the stringency measures? Our intuition is that people comply with these regulations when they get to know more about the pandemic. When the pandemic becomes more relevant to them, people likely feel more pressure to comply with stringency measures themselves. In turn, the knowledge of the pandemic might matter more which are equal to one if a given country is above the median level in our sample for that variable, and zero otherwise. We also create a Low Education dummy which is equal to one if a given country is below the median level of Education in our sample. We interact these dummies with our baseline interaction between Stringency Index and Covid searches, to obtain a triple interaction term.

To compute the implied interaction term for countries whose level of Rule of Law is above the median, we sum the coefficient on the triple interaction ( Repression are equal to one if a given country is above the median level in our sample, and zero otherwise. Low Education is equal to one if a given country is below the median level of Education in our sample, and zero otherwise. The dataset includes 35 countries and 315 days. Robust standard errors are clustered at country level (and shown in parentheses). *** p<0.01, ** p<0.05, * p<0.1

These results indicate that transparency of institutions, citizens' confidence in the rules of society, low level of media repression and high level of education narrow down the effect of citizens' knowledge, as measured by the volume of Covid searches: when people are more likely to trust institutions, abide by the law, receive fair information and be properly informed because of their high level of education, Covid searches do not amplify or diminish the effects of stringency measures.

The key identifying assumption for DiD estimates is that the variation in mobility in countries belonging to the control group is an unbiased estimate of the counterfactual. While we cannot directly test this assumption, we can test whether the time trends in the control and treatment countries were the same in the pre-intervention periods. If the trends are the same in the preintervention periods, then it is likely that they would have been the same in the postintervention period, had the treated countries not adopted any lockdown measure. An eventstudy analysis can shed some light on the validity of the research design. In line with Autor (2003), we create a dummy variable which takes on the value of one on the first day of the stringency index greater than zero, and zero otherwise. We do not introduce this dummy variable directly in our specification, but we interact it with the mean of the Stringency Index adopted by each country, in order to account for the overall intensity of the government measures. Hence, starting from this variable, we create its leads (one for each day prior the day of the lockdown) and lags variables 23 (one for each day after the lockdown measure was introduced). If the trends in the mobility measure in adopting versus non-adopting countries are the same, then the leads should not be statistically significant. An attractive feature of this test is that the lags are informative and can show whether the effect changes over time. We estimate the following specification:

As the number of countries with more 282 lags sharply decreases after the 283 rd day from the stringency adoption, we replace each individual lag for the remaining 13 days with a single dummy variable interacted with the mean stringency. to 1 in country c and day d+, with π going from -32 to -2: those dummies stand for the leads of the variable . We also include the lags of the by building the dummies ( + ) equal to 1 in country c and day d+, with  going from 1 to 283.

Finally, we have which is the mean of the Stringency Index in country c.

All the other variables and fixed effects are defined as in Equation (1). This specification allows for testing parallel trends in the pre-treatment period, namely, whether the coefficients associated with the lead ( π , with π going from -32 to -2) are not statistically different from zero. This approach also helps understand whether the treatment effect fades, increases, or stays constant over time, depending on the estimated coefficients on the lags ( τ, with τ going from 1 to 283). The omitted day is the day before the lockdown, which (given the staggered time of the adoption) differs by country. For example, in Sweden the lockdown started on March 9, 2020, therefore there are 13 leads and 270 lags, and omitted day is March 8, 2020.

The estimates, together with their 90% confidence intervals, are plotted in Figure 3 3.

According to the point estimates, in the pre-treatment period there is no difference in the movement until around the 10th day after the adoption of the lockdown. Turning now to the lag coefficients, we find that the lockdown measures contribute to a reduction in mobility, but it takes some days for the effects to materialise. The coefficient associated with the lags turns out to be negative and statistically significant at the 5% after 11 days since the first day of the lockdown. From the 11 th day after the introduction of the stringency measures, we get a steep decrease in mobility for the following two weeks, followed by a milder decrease up to the 120th day after the introduction of the lockdown. Afterwards, the estimated coefficient starts increasing and reaches a plateau after the 160th day until the end.

Notes: Plots of estimates from Equation (3), with their respective pointwise 90% confidence intervals. The plotted estimated coefficient is the interaction between the leads and lags and the mean of the stringency index for each country during the entire time period. The dependent variable is the Mobility Index. The day before the start of the lockdown is omitted, so the estimates are normalized to zero in that day. The model also includes country and daily fixed effects, temperatures, and per capita confirmed cases as covariates. Errors are clustered at country level. The sample include 35 countries observed over 315 days.

To explain and better understand the panel variation of Covid searches we implement a fixed effects regression analysis.

First, in Column (1) of Table 3 , we show that the difference in the number of confirmed cases between country i and the average of its four closest neighbouring countries is positively and significantly correlated with Covid searches in country i, with a coefficient of 0.43 (1% confidence level). The intuition is that the search of information by citizens of a given country appears to be driven by the excess of country's own cases vis a vis its neighbouringand comparable-countries. The difference in number of confirmed cases with neighbors is the difference between the 7-days moving average of confirmed cases in country i and the mean of the 7-days moving average of confirmed cases in the four closest neighbouring countries. Low education is a dummy equal one when the Education is below the median. The dataset includes 35 countries and 315 days. Robust standard errors are clustered at country level (and shown in parentheses). *** p<0.01, ** p<0.05, * p<0.1 Second, to identify other observable factors that are significantly associated with Covid searches, we include additional variables in our regressions. In Column (2) we add the Stringency Index and find that it is positively and significantly correlated with Covid searches. Citizens are significantly more interested in searching about COVID-19 when governments implement stricter containment measures. In Column (3) we also include temperature, per capita confirmed cases and the interaction between the Stringency Index and the country-specific level of education.

We find that the coefficient on the interaction of the Stringency Index and the countryspecific low education level is negative and statistically significant offsetting the significant positive coefficient not interacted. In other terms, in countries with low levels of education, the Stringency Index is not significantly correlated with the outcome variable, while in countries with high levels of education the Stringency Index has a positive and statistically significant correlation with Covid searches.

In this section, we use a battery of robustness checks to address possible issues related to the research design that could bias our baseline estimates. First, we replace the main dependent variable by excluding one by one each component of the Mobility Index; then we move to a country-sensitive test to show that the estimated effects do not depend on a specific country, and lastly, we run a falsification test, replacing the Covid searches with other relevant terms searched on Google during the same timespan.

The dependent variable used in the main regression (Table 1) Table 4 we exclude one component at a time from the dependent variable. 24 The coefficient on the interaction term Stringency Index*Covid searches remains negative and statistically significant in all specifications, which is consistent with our results not essentially depending on a particular component of the Mobility Index. 

We also test whether our main findings are sensitive to the exclusion of a single country. For this reason, we estimate Equation (2), by dropping one country at a time. The estimated coefficients of the interaction term Stringency Index*Covid searches and their 95% confidence interval (Figure 4) are very similar to those obtained in our baseline specification.

Hence, it can be concluded that our main result is not driven by a particular country. Notes: Estimates of the coefficient from Equation (2) with its 95% confidence interval, excluding from the original set of 35 countries one country at a time (reported on the x-axis). We include country, daily fixed effects and country specific trends. The dataset includes therefore 34 countries and 315 days. Robust standard errors are clustered at country level.

Within our DiD analysis we conduct a placebo test to simulate how alternative Google searches that are unrelated to the pandemic might impact mobility. This test arises from the concern that Covid related searches could be endogenous to mobility, e.g. the week by week volume of Google searches can be correlated with the fact of staying at home, i.e. with lower mobility.

If the relationship between Covid searches and mobility were spurious, namely due to the stay-at-home order which cause more searching activity on Google, using our placebo variables we would get similar results to the ones obtained in the baseline specification which makes use of "Covid" searches. Specifically, we replicate the main analysis in Equation (2) by replacing Covid searches with the main three terms searched in Google in the year 2020 (Translate, Porn, and Maps). Notice, moreover, that these terms are most likely unrelated with the term Covid. The graphical analysis in Figure 5 shows that the searches for Translate, Porn, and Maps are not correlated with the Covid searches in the timespan of our analysis.:

the Pearson correlation index is respectively equal to 0.16, 0.17, and -0.09. In Table 5 , we use as explanatory variable Google searches for Translate (Column 1), Porn (Column 2), and Maps (Column 3). In all specifications we find that the coefficients on the interaction terms are statistically indistinguishable from zero: thus, Google searches different from Covid apparently do not affect the impact of stringency on mobility. Notes: The dataset includes 35 countries and 315 days. Robust standard errors are clustered at country level (and shown in parentheses). *** p<0.01, ** p<0.05, * p<0.1.

The variable Covid searches we use in our main analysis (Table 1) is a weekly intensity, which is measured as the number of weekly searches for the term, divided by the maximum number of its weekly searches over the whole time period, within each country, and scaled to 100 for easier readability.

On the other hand, Covid searches can be re-scaled at the aggregate level, i.e. jointly considering all sampled countries. To rescale the variable, we proceed as follows. First, we find the country with the maximum number of searches (i.e. Chile) in our sample of 35 countries. Then we collect the data from the other countries in groups of five from Google Trends, always including the leading country (Chile). Afterwards, we use the ratio between the leading country and the remaining observations of different groups to re-scale the variable. Eventually, we come up with a dataset with variables from 0 to 100, where the maximum value of 100 is only reached by Chile on April 2020 .

We replicate our baseline specifications by replacing the original Google searches with the scaled version (Table A5) : we find very similar results, and a negative and significant coefficient on the interaction term Stringency Index*Scaled Covid searches.

This paper has empirically shown that implementing lockdown measures has a significant and sizeable impact on individual mobility, as required to control the spread of the virus. 

Google COVID-19 Community Mobility Reports

Variation in Government Responses to COVID-19

Global Historical Climate Network Daily

Polarization and public health: Partisan differences in social distancing during the coronavirus pandemic

A Study on the Quality of Novel Coronavirus (COVID-19) Official Datasets

Outsourcing at will: The contribution of unjust dismissal doctrine to the growth of employment outsourcing

Trust and compliance to public health policies in times of COVID-19

Compliance with COVID-19 social-distancing measures in Italy: The role of expectations and duration

COVID-19, lockdowns and well-being: Evidence from Google Trends

Stay-at-home orders, social distancing and trust

Google Trends: a web-based tool for real-time surveillance of disease outbreaks

WHO Declares COVID-19 a Pandemic

The effectiveness of moral messages on public health behavioral intentions during the COVID-19 pandemic

A Matter of Trust? Political Trust and the Covid-19 Pandemic

The 'Great Lockdown' and its determinants

Information and Financial Literacy for Socially Sustainable NDC Pension Schemes

Lockdown Fatigue: The Diminishing Effects of Quarantines on the Spread of COVID-19

The COVID-19 Pandemic in Norway and Sweden -Threats, Trust, and Impact on Daily Life: A Comparative Survey

Toward effective government communication strategies in the era of COVID-19

Leviathan challenged -is the lockdown EHCR compliant?

The social and economic impact of Covid-19

Seasonal Adjustment of Economic Time Series and Multiple Regression Analysis

Bias in Cable News: Persuasion and Polarization

News Media and Crime Perceptions: Evidence from a Natural Experiment

Psychological Correlates of News Monitoring, Social Distancing, Disinfecting, and Hoarding Behaviors Among US Adolescents During the COVID-19 Pandemic

Political Beliefs affect compliance with COVID19 social distancing orders

Quantifying trading behavior in financial markets using Google Trends

The Limit of Oversight in Policing: Evidence from the 2001 Cincinnati Riot

Assessment of the Impact of Media Coverage on COVID-19

Related Google Trends Data: Infodemiology Study

Political Advertising and Election Results

COVID-19 and excess mortality in the United States: A countylevel analysis

A global panel database of pandemic policies

Government Response Tracker)

2020. I Read the News Today, Oh Boy: The Effect of Crime News Coverage on Crime Perception

Estimating the impact of physical distancing measures in containing COVID-19: an empirical analysis