key: cord-1026682-7ozmay3f
authors: Furceri, Davide; Kothari, Siddharth; Zhang, Longmei
title: The effects of COVID‐19 containment measures on the Asia‐Pacific region
date: 2021-08-21
journal: Pacific Economic Review
DOI: 10.1111/1468-0106.12369
sha: 4e4c3058147e6b526a1132b42c9450e257eebd8d
doc_id: 1026682
cord_uid: 7ozmay3f

As the COVID‐19 pandemic spread across the world, governments introduced significant containment measures to control the spread of the virus. In this paper, we leverage inputs from IMF desk economists to construct a novel narrative index of containment measures for 11 countries in the Asia‐Pacific region. A key innovation in our index is that it distinguishes between economic sectors (services, industry, retail), thus providing a more granular view of restrictions related to economic activity. Using this index, along with other high‐frequency data, we find that containment measures have been successful in reducing the spread of the virus (though with some heterogeneity) but have also been associated with large short‐term economic costs. Furthermore, exploiting the granularity of our index, we find that differences in strategies across countries regarding the closing of the industrial sector have mattered: imposing less severe restrictions on industry has been associated with lower economic costs without leading to worse health outcomes, possibly reflecting the less contact intensive nature of industrial activity.

nomic activity. Using this index, along with other highfrequency data, we find that containment measures have been successful in reducing the spread of the virus (though with some heterogeneity) but have also been associated with large short-term economic costs.

Furthermore, exploiting the granularity of our index, we find that differences in strategies across countries regarding the closing of the industrial sector have mattered: imposing less severe restrictions on industry has been associated with lower economic costs without leading to worse health outcomes, possibly reflecting the less contact intensive nature of industrial activity.

As the COVID-19 pandemic spread across the world, governments introduced various nonpharmaceutical measures to contain the virus, ranging from the introduction of testing and tracing systems to nation-wide lockdowns covering all non-essential sectors. While these containment measures were deemed essential to limit the health costs of the virus, they have widely been associated with large economic costs.

There is a growing literature that estimates the impact of containment measures on health and economic outcomes (e.g., Deb et al., 2020a Deb et al., , 2020b IMF, 2020a) , usually using the de jure measure of containment constructed by Hale et al. (2020) . 1 In this paper, we leverage inputs from IMF desk economists to develop an alternate de jure narrative measure of containment for 11 countries in the Asia-Pacific region. Our stringency index captures government-imposed restrictions related to six sectors, namely schools, retail, services, industry, gatherings and public events, and international travel. Our new index has two key advantages relative to other publicly available indices. First, it provides a more granular view of containment measures related to economic activity as it distinguishes between key economic sectors (services, industry, retail) . This greater granularity is important for Asian countries, as several of them imposed less severe restrictions on industry (Australia, Korea, Thailand, Vietnam) as compared to services and retail, which in turn impacted economic outcomes. Second, our index for China is far more granular as it is first constructed at the province level (Zhang, forthcoming) , and therefore is able to better capture the easing of containment measures that aggregate indices miss.

We use these indices to assess the impact of containment measures on health (COVID-19 cases) and economic (Google mobility data) outcomes in the Asia-Pacific region. Establishing causality is difficult in this context for several reasons, including: (a) concerns of reverse causality as containment measures are implemented in response to rising infections; (b) omitted variable bias as containment measures are implemented alongside several other nonpharmaceutical measures, like testing and tracing systems; and (c) possible anticipation of the implementation of measures due to pre-announcements. To address these concerns, we use high frequency daily data which allows us to control for recent virus developments, lags of the stringency index, and other non-pharmaceutical containment measures.

Our baseline results for Asia are qualitatively similar to that found in the recent literature on a broader sample of countries. In particular, 30 days after the imposition of a complete lockdown (shut down of all non-essential activities in all six sectors-equivalent to a change of 1 in the index), the number of confirmed cases is over 50% lower than the baseline without any restrictions. We also confirm the finding of Deb et al. (2020a) , that the effectiveness of containment measures in reducing infections depends crucially on country characteristics, including population density and the quality of the health system. As to the economic impact, we find that the tightening of containment measures-equivalent to a complete lockdown-leads to a decline of about 40% points in mobility (a proxy of economic activity) about a week into a lockdown. 2 Furthermore, exploiting the granularity of our index with respect to different economic sectors, we find that closing the industrial sector increases the economic costs of lockdowns (possibly because of challenges with working from home in the sector) without leading to significantly worse health outcomes (possibly reflecting the less contact intensive nature of industrial activity).

This paper contributes to a fast-growing literature on the effect of containment measures on health and economic outcomes. Several papers use cross-country or sub-national data to show the role played by containment measures, voluntary social distancing, and other factors in slowing the spread of the virus (e.g., Bretschger et al., 2020; Chernozhukov, Kasahara, and Schrimpf 2020; Dave et al. 2020; Deb et al. 2020a; Demirgüç-Kunt, Lokshin, and Torre 2020; Fang, Wang, and Yang 2020; Gapen et al., 2020; Glaeser, Gorback, and Redding 2020; Imai et al. 2020; Jinjarak et al. 2020; Luong and Nguyen, 2020; Zimmermann et al., 2020) . Another set of papers look at the economic impact of containment measures, either using high frequency data or recent GDP releases (Carvalho et al. 2020; Chronopoulos, Lukas, and Wilson 2020; Deb et al. 2020a; Baek et al. 2020; Baker et al. 2020; Béland, Brodeur, and Wright 2020; Coibion, Gorodnichenko, and Weber 2020; Goolsbee and Syverson, 2020; Gupta et al. 2020; Konig and Winkler, 2020a and . 3 Our contribution to this literature is threefold. First, we develop a novel narrative index of containment measures which provides a more granular view of containment measures, especially related to economic activity, distinguishing between restrictions placed on services, retail and the industrial sector. Second, we use this granularity to analyze how differences in strategies regarding industrial closure impacts economic and health outcomes, which, to the best of our knowledge, has not been looked at before. Third, we focus on the Asia-Pacific region. While this limits our sample size, we believe this is an interesting region to focus on as it was the first region affected by the pandemic, and countries in the region have undertaken significantly different strategies to fighting the pandemic and with varied results, providing important lessons on the effectiveness of containment measures.

The rest of the paper is structured as follows. Section 2 describes how we construct our containment indices. Section 3 documents stylized facts regarding the strategies adopted by Asian countries regarding lockdowns and exits. Section 4 describes our data and methodology, and Section 5 discusses our main results and provide several robustness checks. Section 6 concludes.

This paper develops a new narrative de jure index of containment measures at the daily frequency for 11 advanced and emerging economies in the Asia-Pacific region. The chosen countries include the largest economies in the region, as well as some countries which have had noted success in controlling the pandemic. The countries covered are Australia, China, Japan, Korea, Malaysia, New Zealand, the Philippines, India, Indonesia, Thailand, and Vietnam.

The construction of the index is based on the methodology of Franks et al. (2020) and IMF (2020b), which developed similar indices for 22 European countries. The aggregate index is composed of six sub-indices, which cover government-imposed restrictions related to different types of activities and sectors. The sectors covered include:

1. Schools: captures restrictions related to closures of schools and universities. 2. Retail: captures restrictions on the operation of retail outlets including malls (excludes essential activities such as grocery stores). 3. Services: captures restrictions related to all non-essential services other than retail such as restaurants, movie-theatres, spas and so on. 4. Industry: captures restrictions related specifically to the mining, manufacturing and construction sectors. 5. Public gatherings and events: captures restrictions related to the holding of large public gatherings and events 6. International travel: captures the closure of international borders.

Each sub-index takes four values ranging from 0 to 3, reflecting different levels of government restrictions: a value of 0 implies no restrictions are placed on the sector; 1 indicates the sector is operating throughout the country but with some restrictions or enhanced health protocols (e.g. 50% capacity in restaurant); 2 indicates the sector is partially closed, either geographically (i.e. some states or major cities are still closed) or in terms of some sub-sectors (e.g. high risk bars remain closed but other restaurants are open); and a value of 3 indicates the sector is completely closed nationally. Appendix 1 provides details regarding the coding of each subindex.

Various sources are used to identify government policy actions (both the imposition of restrictions and the subsequent withdrawal) related to each sector. These include a survey of IMF economists working on each country, and various publicly available sources such as the ACAPS Government Measures Dataset, government press releases, and other news sources. Each identified policy action is scored on the 0-3 scale described above to create daily indices for each sector-country. The final set of policy measures used to construct the index is then reviewed by IMF economists working on each country.

For most of the analysis in the paper, we use an aggregate stringency index at the countryday level. To compute this aggregate index for each country, the sub-indices are normalized to lie between 0 and 1 (by dividing by 3), and then the simple average of the six sub-indices is taken. The dataset covers the period from January 1, 2020 to September 9, 2020.

Except for China, all indices are constructed at the national level. For China, we use the index constructed by Zhang (forthcoming), which uses the same methodology as this paper but constructs the index at the province level. This granular index better captures the large heterogeneity in the imposition and lifting of restrictions at the province level. We average across provinces to construct a national level index which we use in the analysis for this paper.

Compared to other publicly available containment indices-most prominently the Coronavirus Government Response Tracker of the University of Oxford (Hale et al., 2020) -the new index developed in this paper has two important advantages. First, it provides a more granular measurement of containment measures related to economic activity as it distinguishes between key economic sectors (services, industry, retail). This greater granularity is important for Asian countries, as several of them had less severe restrictions for industry as compared to services and retail. Moreover, as shown in Section 5, different strategies with respect to closing of the industrial sector matters for health and economic outcomes. Second, our index for China is far more granular as it is first constructed at the province level, and therefore is able to better capture the easing of containment measures that aggregate indices miss.

To illustrate the difference between our index and the one developed by the University of Oxford (OxCGRT), Figure 1 plots both stringency indices for China and Korea. For China, our index picks up the gradual reopening since February which is missed by Oxford. For Korea, our index generally shows lower levels of restriction than Oxford, partly reflecting the fact that Korea placed less stringent restrictions on industry compared to services. Despite these differences, the two indices also have considerable overlap. As shown in Figure 2 , when averaged across the 11 countries, the two indices comove to a significant degree. In particular, the peak is similar across the two indices-both in terms of timing and magnitude-although the degree of containment before and after the peak is different, with our index being less stringent, especially in the exit phase. Figure A1 compares the two indices for all countries. Figure A2 provides details regarding each sub-index.

As with all de jure measures of containment, it is important to note some caveats regarding the index. While we go at length to try and ensure cross-country comparability of the indices, restrictions imposed in countries often take different forms which cannot be captured in full granularity in our relatively coarse ordinal indices. Furthermore, standardizing across countries inevitably leads to some subjective decisions when assigning stringency scores to particular policy actions. 4 3 | STYLIZED FACTS: CONTAINMENT STRATEGIES ADOPTED BY ASIAN COUNTRIES Before turning to more formal econometric analysis in the next section, this section provides some stylized facts regarding the lockdowns and reopening strategies adopted by countries in the Asia-Pacific region.

All countries in our sample introduced containment measures early in 2020, aimed at controlling the spread of the COVID-19 virus. 5 Usually, international travel restrictions were imposed first, followed by school closures. 6 However, the overall stringency and duration of the initial lockdowns differed markedly across countries.

F I G U R E 1 Overall stringency index, selected countries, 2020 (index). Plots the overall stringency index for China and Korea. IMF stands for the index used in this paper. For China, we aggregate the province level index constructed in Zhang (forthcoming) . Oxford is the overall stringency index constructed by Hale et al. (2020) . Figure A1 compares the Oxford and IMF indices for all countries [Color figure can be viewed at wileyonlinelibrary.com] F I G U R E 2 Overall stringency index, Asia, 2020 (index). Plots the overall stringency index averaged across 11 Asian countries. IMF stands for the index constructed in this paper. Oxford is the overall stringency index constructed by Hale et al. (2020) [Color figure can be viewed at wileyonlinelibrary.com] As shown in Figure 3 , several countries imposed near complete lockdowns (India, Malaysia, New Zealand, Philippines), with the overall stringency index reaching the maximum value of 1. On the other hand, some countries never imposed national lockdowns (Indonesia) while others only closed nonessential services and allowed industrial sectors to continue operating (Australia, Thailand, Vietnam) resulting in peak values of the index staying below the maximum value. Korea and Japan did not implement mandatory shutdowns, instead issuing strong recommendations regarding business closures.

The effectiveness of lockdowns in reducing infection rates also varied, with countries starting the process of reopening their economies at different stages of the epidemic curve. We define the reopening date for each country as the first time that the overall stringency index declines from its peak value. 7 India, China, Indonesia and the Philippines started reopening their economies before new cases had declined significantly from their peak. For China, some low risk provinces were partially opened in the second week of February while national level cases were still rising, driven mainly by the outbreak in Hubei. By March, the number of new cases in China were down to single digits. 8 In India, Indonesia and the Philippines, new cases continued to rise for a significant period after the start of the reopening process, with infections rates staying high through August and September.

By contrast, most other countries in the region started easing restrictions when new cases were more than 80% below peak pre-reopening levels ( Figure 4 ). In this group, several countries have seen second waves or localized outbreaks since reopening, and several countries retightened containment measures because of the reemergence of new infections, often at the sub-national level (Australia, New Zealand, Korea and Vietnam). See Figure A1 for country details.

The speed of reopening also differed across countries. To construct a metric for speed of reopening, we define the concept of effective days open to calendar days open for sector s, in country i as: F I G U R E 3 Peak lockdown stringency and duration (index, days [right-axis]). Plots the peak level of the overall stringency index for each country as well as the number of days that the index stayed at its peak value [Color figure can be viewed at wileyonlinelibrary.com]

where C s i,t is the containment index for sector s, in country i, on day t, and the summation is taken from the reopening day of country i (R i ) to the end of our dataset (day T). Therefore, if the containment index for a sector is at 3 on a particular day, then that is counted as zero effective days open, while if the index is at 2 then it is counted as a third of an effective day open, and so on. Figure A3 plots the speed of reopening for each country (averaging across sectors). While Indonesia, India, Malaysia and the Philippines have relaxed their harshest containment measures, the speed of reopening has been slow as many sectors remained partially closed. On the other hand, other countries that started easing restrictions after virus cases had subsided have continued easing restrictions over time and have reopened at a faster rate. 9 In terms of sequencing of sectors, the industrial sector was generally reopened first (or did not close), while schools and international borders have been the last to reopen (or still remain closed for some countries).

In the rest of the paper, we assess the impact of containment measures on the path of the epidemic and economic outcomes. For this exercise, we compile a comprehensive dataset of health F I G U R E 4 Reopening timing and latest infection rates (percentage change, new cases per million). Reopening date is defined as the first time the stringency index declines from its peak. 1 Excludes Hubei as the number of cases reported in Hubei shows a structural break on February 12, 2020 when clinically diagnosed cases were added as opposed to only laboratory confirmed cases, leading to a one time spike in case counts [Color figure can be viewed at wileyonlinelibrary.com] and economic variables at the daily frequency to complement the containment indices that we have constructed (see Table 1 for descriptive statistics).

COVID-19 cases: The number of confirmed COVID-19 cases is the main variable used to measure success in containing the pandemic. Data on number of infections for our 11 Asian countries is taken from the COVID-19 Dashboard of the Coronavirus Resource Center of Johns Hopkins University.

Google mobility: We use data on mobility from Google Mobility Reports as our primary proxy for economic activity. The report shows trends in visits and length of stay at different types of locations. The data for each day and location is reported as the change relative to a baseline value computed as the median for the corresponding day of the week during the 5-week period January 3, 2020 to February 6, 2020. To construct our aggregate measure of economic activity, we average across four indices, which capture movements in retail and recreation places, groceries and pharmacies, transit stations, and workplaces. China is excluded from the analysis on economic impact as Google mobility data for China is unavailable.

Other variables: We use several other variables as controls in robustness checks and also explore the heterogeneity of the impact of containment measures with respect to structural features of countries. The main variables used include daily temperature and humidity (from the Air Quality Open Data Platform); testing and tracing policies and public information campaigns (from OxCGRT); population density (latest available year from the World Development Indicators); and quality and preparedness of healthcare systems (health index from the World Economic Forum.

This section describes the empirical methodology used to examine the effect of containment measures on the evolution of the epidemic and on economic outcomes.

Establishing causality is difficult in this context for several reasons. When looking at health outcomes, reverse causality is a major challenge as countries introduced containment measures in response to the spread of the epidemic itself. Failure to control for possible reverse causality would result in estimates of the effect of containment measures on infections being upward biased (that is, toward not finding significant effects). The correlation between the implementation of containment measures and the spread of the virus can also bias estimates of the economic impact of containment measures as the spread of the virus can directly impact economic activity by altering the behavior of individuals (e.g. increased voluntary social distancing).

Another concern is that containment measures were announced before being implemented and, therefore, were anticipated. This may have resulted in reduced mobility ahead of the implementation of some containment measures, impacting economic activity and the spread of the virus, thus biasing estimates. Another important empirical challenge is omitted variable bias, as containment measures have been introduced as parts of broader non-pharmaceutical interventions (NPIs) which include enhanced testing, contact tracing, and public information campaigns aimed towards increasing social awareness. Not controlling for these factors may overestimate the effect of containment measures.

As discussed in more detail below, our use of daily data allows us to control for lagged values of infection rates and containment measures, alleviating some of the endogeneity concerns. We also conduct several robustness checks aimed at allaying concerns about omitted variable bias, including controlling for NPIs. We also constructed an Instrumental Variable (IV) using the international travel restrictions component of the containment measure, which is driven by COVID developments outside the country (Hsiang et al. 2020 ).

Following the recent literature, we use a reduced form econometric approach to estimate the impact of containment measures on health and economic outcomes (Hsiang et al., 2020; Deb et al., 2020a Deb et al., , 2020b IMF 2020b) . In particular, we use local projection methods (Jorda, 2005) where the baseline regressions take the form:

where Δy i,tþh ¼ y i,tþh À y i,tÀ1 is the cumulative change in various health and economic measures over a horizon of h days for country i, c i,t is the aggregate stringency index for country i at time t, X i,t is a vector of control variables including lagged values of the dependent variable and the stringency index, and time trends, and α h i are country fixed effects. β h is the coefficient of interest and will trace out the impulse response function to a unitary change-equivalent to a full lockdown-in the stringency index. Standard errors are clustered at the country level.

When analyzing the impact of containment measures on health outcomes, our dependent variable is the change in the log of confirmed cases, that is, Δy i,tþh ¼ ln cases i,tþh ð ÞÀln cases i,tÀ1 ð Þ , where cases i,t is the total number of confirmed COVID-19 cases in country i as of time t.

In our baseline specification, we control for several variables to try and address endogeneity concerns. As discussed above, the imposition and removal of containment restrictions are likely to be driven by previous infection trends, raising concerns about reverse causality. To address this, we control for lagged values of changes in the infection rates (i.e., lags of the dependent variable). We also control for lags of the containment index itself to allow for the fact that containment measures take time to impact detected cases and as some containment measures may have been announced before they were implemented. Finally, we also include country specific cubic time trends to allow for different virus dynamics for each country. As a robustness check, we expand the set of controls to include temperature, humidity and proxies for other non-pharmaceutical interventions (testing, tracing, and public information campaigns).

When analyzing the impact of containment measures on economic activity, our primary dependent variable is the change in mobility rate, that is, Δy i,tþh ¼ mob i,tþh À mob i,tÀ1 . Fern andez-Villaverde and Jones (2020) and Sampi and Jooste (2020) find a high correlation between GDP and Google mobility data, indicating that mobility is a good proxy for economic activity.

As in the health regressions, we control for several variables to try and address endogeneity concerns. We once again control for the evolution of the pandemic by including the change in log of confirmed cases (and its lags) as the imposition and removal of containment restrictions are likely to be driven by previous infection trends, and as higher infection rates can directly impact economic outcomes by changing individual behavior. As with the health regressions, we also control for lags of the containment index, and country specific cubic time trends.

In addition to the baseline, we explore how country specific characteristics can shape the effects of containment measures. We use two specifications for this purpose.

When testing for heterogeneity with respect to a continuous variable (e.g., population density), we run regressions of the form:

where F z i,t ð Þ¼ exp Àz i,t ð Þ 1þexp Àz i,t ð Þ , and z i,t is the country specific characteristic normalized to have mean 0 and standard deviation 1. The coefficients θ h L and θ h H capture the impact of containment measures for very low and high levels of z, respectively. In particular, when z is low then F z ð Þ is close to 1 and the impact of containment measures is picked up by the coefficient θ h L -F z ð Þ tends to 1 when z tends to z minus infinity. On the other hand, when z is high then F z ð Þ is close to 0 and the impact of containment measures is picked up by the coefficient θ h H -F z ð Þ tends to 1 when z tends to plus infinity. This approach is equivalent to the smooth transition autoregressive model developed by Granger and Terävistra (1993) . 10 When testing for heterogeneity with respect to a categorical variable (e.g., whether countries closed the industrial sector or not), we use the below specification:

where D i,t is a dummy variable which takes value 0 or 1. The coefficient γ h captures whether the response to containment measures differs by the country specific characteristic captured by the dummy variable D i,t . 11

In this section, we present our main results regarding the impact of the containment measures on health and economic outcomes. Figure 5 shows the impulse response for a unitary increase in the stringency index estimated from Equation (2). The log differences in confirmed COVID-19 cases is the dependent variable. As the chart shows, tightening of containment measures leads to a significant decline in the number of confirmed cases relative to a baseline in which restrictions are not tightened. In particular, the number of confirmed cases are over 50% lower than the baseline 30 days after the imposition of lockdowns. 12 Also interesting to note is that the negative effect only becomes significant after about 2 weeks, highlighting the lags with which containment measures impact infections, likely due to the incubation period of the virus. 13 The baseline results are robust to adding further control variables and other changes to the specification. As containment measures were implemented at the same time as other nonpharmaceutical measures (testing, tracing, public information campaigns), not controlling for these may overestimate the effectiveness of containment measures. Figure 6a controls for these non-pharmaceutical interventions as well as weather-related variables (temperature and humidity), giving very similar results to the baseline. Including time fixed effects to control for common factors across countries also does not change the results significantly ( Figure 6b ). Finally, excluding China from the regression (as our index for China is constructed slightly differently) also has little effect on the results (Figure 6c ).

F I G U R E 5 Impact of containment measures on COVID-19 cases (cumulative deviation from baseline, log difference). Impulse response function estimated using Equation (2) with changes over different horizons in the log of the number of confirmed cases used as the dependent variable. Grey shaded region is the 90% confidence interval. X-axis corresponds to different horizons, while the y-axis shows log differences in number of cases relative to baseline [Color figure can be viewed at wileyonlinelibrary.com]

In the baseline specification, our empirical strategy exploits the high-frequency nature of our data and the timing of imposition of restrictions to identify the effect of containment measures on health outcomes. As an additional robustness check, we also apply an instrumental variable (IV) strategy in which we use the international travel restrictions component of the containment measure as an instrument for overall stringency, the idea being that international travel restrictions are more likely to be imposed due to health developments in other countries than in the home country (Hsiang et al. 2020) . As shown in Figure 6d , the coefficient estimates using the IV strategy are similar to the baseline, although less precisely estimated. 14 As shown in Deb et al. (2020a) , the impact of containment measures also depends crucially on country specific characteristics. Figure 7 shows how the impact of containment measures differ depending on a countryʼs population density. Imposing restrictions tends to be very (2) with changes over different horizons in the log of the number of confirmed cases used as the dependent variable. In panel D, the international travel restrictions are used as an instrument for overall stringency. Grey shaded region is the 90% confidence interval. X-axis corresponds to different horizons, while the y-axis shows log differences in number of cases relative to baseline [Color figure can be viewed at wileyonlinelibrary.com] effective in countries with low population densities, reducing infection rates by about 67% compared to a no intervention baseline in the 90th percentile country in terms of population density in our sample (New Zealand). By contrast, for the 10th percentile country in our sample, containment measures reduce infection rates by only about 20%, potentially reflecting the fact that effective social distancing can be hard to achieve when population densities are high. These heterogeneous effects, however, should be interpreted with caution as variables such as population density may be correlated with other country characteristics. For example, as Figure 7 shows, in high population density countries, the imposition of lockdowns is followed by continued increase in infection rates for about 2 weeks before the infection curve starts bending downwards. This may reflect the fact that poorer countries in our sample tend to have higher population density and also worse health systems, which may be contributing to containment measures taking longer to become effective.

Effectiveness of containment measures also depends on the quality of the health system. As shown in Figure 8 , countries with higher scores for the WEF Health Index tend to see a bigger decline in cases after the imposition of restrictions, suggesting strong complementarities between health systems and containment measures; that is, lockdowns alone are likely to be less effective if they are not complemented with strong healthcare facilities. Figure 9 shows the impulse response for a unitary increase in the stringency index estimated from Equation (2) using change in mobility over different horizons as the dependent variable. As the chart shows, tightening of containment measures leads to a significant decline of about F I G U R E 7 Impact of containment measures on COVID-19 cases: Role of population density (cumulative deviation from baseline, log difference). Impulse response function estimated using Equation (3) with changes over different horizons in the log of the number of confirmed cases used as the dependent variable. The left panel plots the coefficient θ h L , which is the estimate for the impact of containment measures on COVID-19 cases for countries with low population densities, while the right panel plots the coefficient θ h H the estimate for countries with high population density. Grey shaded region is the 90% confidence interval. X-axis corresponds to different horizons, while the y-axis shows log differences in number of cases relative to baseline [Color figure can be viewed at wileyonlinelibrary.com] 40% points in mobility about a week into a lockdown relative to a baseline in which restrictions are not tightened. And while the effect declines and becomes insignificant at longer horizons, the cumulative effect of reduced mobility for several weeks on economic activity is large. Furthermore, as expected and unlike what we observed for infections, the impact on mobility does not occur with a lag, as de jure restrictions lead to an immediate decline in mobility and economic activity. 15 As with the health regressions, the baseline results are qualitatively similar when we control for weather and other non-pharmaceuticals variables as controls (Figure 10a ) and when time fixed effects are included (Figure 10b ). 16 As with the health index, we conduct an additional robustness check in which we instrument for the aggregate stringency measure using international travel restrictions. Results are again qualitatively similar to the baseline though standard errors are larger (Figure 10c) .

The response of mobility to containment measures also differs between tightening and loosening phases, where the tightening phase is defined as the period from the beginning of the sample to when the aggregate stringency index declines for the first time from its peak while the loosening phase is the period after the tightening phase-that is, the period after the first time that restrictions were eased. Figure 11 plots the interaction term γ h from Equation (4) where the D i,t dummy takes the value 1 for the tightening phase only. 17 The interaction term is negative, indicating a stronger negative effect of containment measures on mobility during the tightening phase. This indicates that a complete reversal of containment measures may not lead to economic activity recovering to pre-pandemic levels, potentially reflecting continued voluntary social distancing despite the removal of formal restrictions, or scarring effects from the pandemic. F I G U R E 8 Impact of containment measures on COVID-19 cases: Role of the healthcare system (cumulative deviation from baseline, log difference). Impulse response function estimated using Equation (3) with changes over different horizons in the log of the number of confirmed cases used as the dependent variable. Uses the WEF Health Index. The left charts plots the coefficient θ h L , which is the estimate for the impact of containment measures on COVID-19 cases for countries with low values of the health indices, while the right panel plots the coefficient θ h H the estimate for countries with high values of the health indices. Grey shaded region is the 90% confidence interval. X-axis corresponds to different horizons, while the y-axis shows log differences in number of cases relative to baseline [Color figure can be viewed at wileyonlinelibrary.com]

One key innovation in our stringency index compared to the earlier literature is that it distinguishes between restrictions imposed on different economic sectors (services versus industry). Asian countries have adopted different strategies when it comes to closing industry. Australia, Thailand, Korea, and Vietnam imposed less stringent restrictions on the industrial sector compared to non-essential services, while other countries generally had similar peak levels of restrictions for services and industry (though the industrial sector did reopen before services for some of these countries).

To assess the impact of keeping the industrial sector open, we estimate Equation (4) where the D i,t dummy takes value 1 for countries that did not close the industrial sector, that is, where the sub-index for industry remained at 0 (indicating no closure) or 1 (open with enhanced health protocols). 18 Figure 12 plots the coefficient on the interaction term γ h À Á that captures the extent to which confirmed cases and mobility reacts differently to containment measures for countries that did not close the industrial sector.

For confirmed cases (Figure 12, left panel) , the interaction coefficient is insignificant suggesting that containment measures had a similar impact in reducing cases across countries, irrespective of whether they closed the industrial sector or not; that is, keeping the industrial sector open did not lead to containment measures becoming ineffective in reducing virus spread. For mobility (Figure 12, right panel) , the interaction term is positive and significant, indicating that the negative effect of containment measures was smaller for countries that did not close the industrial sector. While these findings cannot be necessarily generalized to other regions, the data for Asia tends to suggest that imposing less severe restrictions on the industrial F I G U R E 9 Impact of containment measures on mobility (deviation from baseline, percentage points). Impulse response function estimated using Equation (2) with change in mobility over different horizons used as the dependent variable. Grey shaded region is the 90% confidence interval. X-axis corresponds to different horizons, while the y-axis shows log differences in number of cases relative to baseline [Color figure can be viewed at wileyonlinelibrary.com] F I G U R E 1 0 Legend on next page. sector can help reduce the economic cost without leading to significantly worse health outcomes.

A possible concern regarding our results on the differential impact of closing the industrial sector is that the countries which kept the sector open were also countries which depend more on industry. However, data from the World Development Indicators for 2017 (the last year for which data is available for all our countries) shows that the share of the industrial sector in GDP in the countries that closed industry is very similar to that in countries which did not close the industrial sector (31.9 versus 31.7% respectively).

While the above analysis focused on the industrial sector, a more general question is whether having greater differentiation in containment policies across sectors leads to different tradeoffs when it comes to health and economic outcomes. To test if this is the case, we use the standard deviation across the six sub-indices (for each country, day) as a measure of differentiation in containment policies. In addition to the overall stringency index, we then also include this standard deviation measure and its interaction with the overall stringency in our baseline regressions. Figure A4 plots the coefficient on this interaction terms for total cases (left panel) F I G U R E 1 0 Robustness: Impact of containment measures on mobility (cumulative deviation from baseline, log difference). (a) Include controls for weather and other non-pharmaceutical interventions. (b) Include time fixed effects. and (c) Instrumental variable approach. Impulse response function estimated using Equation (2) with change in mobility over different horizons used as the dependent variable. Grey shaded region is the 90% confidence interval. X-axis corresponds to different horizons, while the y-axis shows log differences in number of cases relative to baseline [Color figure can be viewed at wileyonlinelibrary.com] F I G U R E 1 1 Asymmetric impact of containment measures on mobility during tightening (deviation from baseline, percentage points). Impulse response function estimated using Equation (4) with change in mobility over different horizons used as the dependent variable. The interaction dummy D i,t takes value 1 from the beginning of the sample to the day before restrictions were loosened, thereby capturing the tightening phase of containment measures. Chart plots the interaction term γ h which captures the extent to which mobility reacts more to containment measures during the tightening phase. Grey shaded region is the 90% confidence interval. X-axis corresponds to different horizons, while the y-axis shows log differences in number of cases relative to baseline [Color figure can be viewed at wileyonlinelibrary.com] and mobility (right panel). The interaction term is positive and significant for mobility, indicating that for a given level of aggregate stringency, having greater differentiation across sectors can lead to lower economic costs, possibly because more damaging containment measures can be avoided. On the other hand, the coefficient for total cases is generally insignificant though positive, indicating neutral-to-potentially worse health outcomes in the case of greater differentiation.

In this paper, we develop a new narrative index of containment measures for 11 countries in the Asia-Pacific region. This index provides a more granular view of containment measures related to economic activity as it distinguishes between key economic sectors (services, industry, retail). We use these indices to assess the impact of containment measures on health (COVID-19 cases) and economic (Google mobility data) outcomes in the Asia-Pacific region. The analysis confirms the results of previous studies suggesting that containment measures have played a key role in limiting the spread of the COVID-19 virus but have also extracted a heavy economic toll.

Asian countries have adopted different strategies when it comes to closing industry, with some countries imposing less severe restrictions on the industrial sector. The analysis exploits this heterogeneity and provides a key novel finding compared to the existing literature: not closing the industrial sector can help reduce the economic costs of lockdowns without leading to significantly worse health outcomes, possibly reflecting the less contact intensive nature of industrial activity. Moreover, we also provide new evidence indicating that for a given level F I G U R E 1 2 Impact of containment measures: differential effect of not closing industry (deviation from baseline, log difference). Impulse response function estimated using Equation (4). Left panel uses changes over different horizons in the log of the number of confirmed cases as the dependent variable. Right panel uses change in mobility over different horizons as the dependent variable. The interaction dummy D i,t takes value 1 for countries which did not close the industrial sector with the same stringency as services (Australia, Korea, Thailand, and Vietnam). The stringency variable c i,t is the average over all sub-indices excluding industry. Both panels plot the interaction term γ h which captures the extent to which confirmed cases and mobility reacts differently to containment measures for countries that did not close the industrial sector. Therefore, a negative value of the interaction term in the left panel indicates that countries that did not close industry saw a bigger decline in cases following an increase in stringency of containment measures, while a positive value in the right panel indicates a smaller negative impact on activity for countries that did not close the industrial sector. Grey shaded region is the 90% confidence interval [Color figure can be viewed at wileyonlinelibrary.com] of aggregate stringency, having greater differentiation across sectors can lead to lower economic costs, possibly because more damaging containment measures can be avoided.

The views expressed in this paper are those of the author(s) and do not necessarily represent the views of the IMF, its Executive Board, or IMF management. The authors thank the editors and two anonymous referees for useful comments. The authors are also grateful to Alison Stuart, Cian Ruane, Diego Cerdeiro, Fan Zhang, Helge Berger, Nour Tawk, and Pragyan Deb for helpful discussions and to Chenqi Zhou and Shihui Liu for excellent research assistance. The authors are also very grateful to the IMF economists and research assistants working on country teams who provided inputs for the construction of the new containment indices, including Ana Kristina Fuentes, Andrea Felice Quinial, Bhayu Purnomo, Bo Hyun Chang, Francois Painchaud, Gee Hee Hong, Geoffrey Bannister, Han Teng Chua, Honey Karun, Rani Setyodewanti, Robin Koepke, Stella Kaendera, Sung Jin Kim, TengTeng Xu, and Yosuke Kido.

ENDNOTES 1 A related literature assesses the trade-off between the health benefits and economic costs based on calibrated models (see Dorn et al., 2020; Krueger et al., 2020; Born et al., 2020) . 2 We do not translate mobility changes into estimates for GDP growth as the relation between mobility and economic activity is likely to vary across countries and evolve over time, as lockdowns continue for extended periods of time and as economic agents adapt to new operating procedures. Furthermore, China is excluded from the analysis on economic impact as Google mobility data for China is unavailable. 3 A related literature looks at the impact of past pandemics on economic outcomes, usually finding large negative effects (see Jord a, Singh and Taylor 2020, Barro, Ursúa and Weng 2020, Carillo and Jappelli 2020). 4 Despite these limitation, the use of a de jure index instead of a de facto index is appropriate to assess the effectiveness of policy interventions because the de jure index captures actual policy announcements regarding containment measures whereas a de facto index will also pick up behavioral responses to conjunctural developments, including on infection rates themselves. 5 Table A1 provides key dates related to the imposition and removal of restrictions for each country in our sample. 6 Most countries introduced some form of international border restrictions as early as the first week of February, usually imposing restrictions on travelers arriving from China. 7 For most countries, the reopening process started between the middle of April and the middle of May. China and Korea started reopening earlier, reflecting the early initial start of the epidemic in these countries. Indonesia started reopening late, in early June. 8 The comparison of China to other countries in terms of number of days at peak stringency and timing of reopening should be interpreted with caution. As the China index is constructed by aggregating province level data, a single province reopening one sector will cause a decline in the index from its peak value. To make the reopening timing more comparable to other countries, we impose a minimum threshold of at least one of the sub-indices declining by 1 unit to consider it to be a reopening for China. Note that this threshold is the same as the minimum amount by which the sub-indices of the other countries can decline. 9 The speed of reopening metric is impacted by the fact that some countries (e.g. Australia) re-tightened restrictions, thus reducing their ratio of effective to actual days open. 10 The lagged dependent variables and the lags of the stringency index are also interacted with F z i,t ð Þ and 1 À F z i,t ð Þ in this specification. 12 The figures show log differences. These are translated to percent changes in the text using the formula exp β h 100 À 1 Ã 100. 13 The magnitude of the decline in cases relative to baseline is smaller than in Deb et al. (2020a) but larger than that found in IMF (2020a). The differences likely reflect the smaller sample of countries used in this paper, the difference in stringency index, and slightly different time periods used. 14 The test that the instrument is weak is rejected, with the Kleibergen-Paap rk Wald F statistic-which is equivalent to the F-effective statistic for non-homoscedastic error in the case of one endogenous variable (Andrews et al., 2019) -always being higher than the associated Stock-Yogo critical value (for a 15% relative bias). 15 The results for mobility found in this paper are qualitatively similar to that in IMF (2020a) which uses a broader sample. 16 China is automatically excluded from our mobility regressions as Google Mobility Reports does not have data for China. 17 We exclude any retightening of restrictions due to a second wave from the sample. F I G U R E A 3 Speed of reopening (effective days open/total days since reopening). Speed of reopening for each country-sector is defined to be the ratio of effective days open to total calendar days since reopening, where effective days adjusts for the extent to which the sector has been reopened (i.e. a sector with intensity score 3 is considered as fully closed, a score of 2 implies two-third closed etc.). For countries that re-tightened restrictions, the speed of reopening is effectively lowered [Color figure can be viewed at wileyonlinelibrary.com] F I G U R E A 4 Impact of containment measures: Interaction with standard deviation across sectors (deviation from baseline, log difference). Left panel uses changes over different horizons in the log of the number of confirmed cases as the dependent variable. Right panel uses change in mobility over different horizons as the dependent variable. Plotted values are the coefficient on the interaction between the level of aggregate stringency and the standard deviation across the six sub-indices which captures the extent to which confirmed cases and mobility reacts differently to containment measures for countries that did not close the industrial sector. Therefore, a negative value of the interaction term in the left panel indicates that countries that did not close industry saw a bigger decline in cases following an increase in stringency of containment measures, while a positive value in the right panel indicates a smaller negative impact on activity for countries that did not close the industrial sector. Grey shaded region is the 90% confidence interval [Color figure can be viewed at wileyonlinelibrary.com] T A B L E A 1 Key dates for first imposition and removal of restrictions 

Weak instruments in IV regression: Theory and practice

Unemployment effects of stay-at-home orders: Evidence from high frequency claims data

How does household spending respond to an epidemic? Consumption during the 2020 COVID-19 pandemic (NBER Working Paper 26949)

The coronavirus and the great influenza pandemic: Lessons from the "spanish flu" for the coronavirus's potential effects on mortality and economic activity (No. w26866)

COVID-19, stay-at-home orders, and employment: Evidence from CPS data (IZA Discussion Paper 13282)

Do lockdowns work? A counterfactual for Sweden

COVID-19 infections and fatalities developments: Empirical evidence for OECD countries and newly industrialized economies

Pandemic and Local Economic Growth: Evidence from the Great Influenza in Italy

Tracking the COVID-19 crisis with high-resolution transaction data (CEPR Discussion Paper 14642)

Causal impact of masks, policies, behavior on early COVID-19 pandemic in the US (CEPR COVID Economics Vetted and Real-Time Papers 35)

Consumer spending responses to the COVID-19 pandemic: An assessment of Great Britain

The cost of the COVID-19 crisis: Lockdowns, macroeconomic expectations, and consumer spending (NBER Working Paper 27141)

When do shelter-in-place orders fight COVID-19 best? Policy heterogeneity across states and adoption time (NBER Working Paper 27091)

The economic effects of COVID-19 containment measures (IMF Working Paper 20/158)

The effect of containment measures on the COVID-19 pandemic (IMF Working Paper 20/159)

The sooner, the better: The early economic impact of nonpharmaceutical interventions during the COVID-19 pandemic. (World Bank Policy Research Working Paper 9257)

The common interests of health protection and the economy: Evidence from scenario calculations of COVID-19 containment policies

Human mobility restrictions and the spread of the novel coronavirus (2019-nCoV) in China (NBER Working Paper 26906)

Exiting from lockdowns: Early evidence from reopenings in Europe (IMF Working Paper)

Assessing the effectiveness of alternative measures to slow the spread of COVID-19 in the United States

How much does COVID-19 increase with mobility? Evidence from New York and four other US cities (NBER Working Paper 27519)

Fear, lockdown, and diversion: Comparing drivers of pandemic economic decline 2020

Modelling nonlinear economic relationships

Effects of social distancing policy on labor market outcomes (NBER Working Paper 27280)

Oxford COVID-19 government response tracker

The effect of large-scale anticontagion policies on the COVID-19 pandemic

Adoption and impact of non-pharmaceutical interventions for COVID-19

The great lockdown: Dissecting the economic effects, World Economic Outlook

Accounting for global COVID-19 diffusion patterns

Longer-run economic consequences of pandemics

Estimation and inference of impulse responses by local projections

COVID-19 and economic growth: Does good government performance pay off? Intereconomics

Monitoring in real time: Cross-country evidence on the COVID-19 impact on GDP growth in the first half of 2020

Macroeconomic dynamics and reallocation in an epidemic (No. w27047)

Nowcasting economic activity in times of COVID-19: An approximation from the Google Community Mobility Report (World Bank Policy Research Working Paper 9247)

Globalization and infectious diseases: Evidence on the reproduction rate of the COVID-19 pandemic

China, the great lockdown and recovery: Some stylized facts

Inter-country distancing, globalisation and the coronavirus pandemic

The effects of COVID-19 containment measures on the Asia-Pacific region

The table below describes the coding used for each sub-index. School International borders are completely closed except to allow citizens or permanent residents to re-enter