key: cord-0855492-i3312hbi
authors: Asahi, K.; Undurraga, E. A.; Valdes, R.; Wagner, R.
title: The effect of COVID-19 on the economy: evidence from an early adopter of localized lockdowns
date: 2020-09-23
journal: nan
DOI: 10.1101/2020.09.21.20198887
sha: 09d7df41dae0b875f305bae830e456280b5a6c70
doc_id: 855492
cord_uid: i3312hbi

Background. Governments worldwide have implemented large-scale non-pharmaceutical interventions, such as social distancing or school closures, to prevent and control the growth of the COVID-19 pandemic. These strategies, implemented with varying stringency, have imposed substantial social and economic costs to society. As some countries begin to reopen and ease mobility restrictions, lockdowns in smaller geographic areas are increasingly being considered as an attractive policy intervention to mitigate societal costs while controlling epidemic growth. However, there is a lack of empirical evidence to support these decisions. Methods. Drawing from a rich dataset of localized lockdowns in Chile, we used econometric methods to measure the reduction in local economic activity from lockdowns when applied to smaller or larger geographical areas. We measured economic activity by tax collection at the municipality-level. Findings. Results show lockdowns were associated with a 10-15% drop in local economic activity, a two-fold reduction compared to municipalities not under lockdown. A three-to-four-month lockdown had a similar effect on economic activity than the year of the 2009 great recession. We found that costs are proportional to the population under lockdown, without differences when lockdowns were measured at the municipality or city-wide levels. Conclusions. Our findings suggest that localized lockdowns have a large effect on local economic activity, but these effects are proportional to the population under lockdown. Our results suggest that epidemiological criteria should guide decisions about the optimal size of lockdown areas; the proportional effects of lockdowns on the economy seem to be unchanged by scale.

of testing and isolation of infected people [6] [7] [8] [9] [10] [11] [12] . However, there is a substantial risk of a resurgence of 46 the epidemic [13] [14] [15] . Understanding the impacts of these interventions is critical because they will most 47 likely continue until an effective vaccine becomes available for a substantial proportion of the population 48 [16] . There is still limited empirical evidence of the effects of interventions to prevent viral transmission 49 [1, 17] ; most impact has been estimated using mathematical models [16, 18, 19] . The COVID-19 pandemic 50 has already imposed an enormous global burden, with about 30 million cases and one million deaths 51 reported so far [20] , and substantial social and economic costs from epidemic control measures [21] [22] [23] [24] [25] [26] [27] . 52 53 As countries begin to reopen and ease mobility restrictions, localized lockdowns are increasingly 54 considered a critical element of a non-pharmaceutical toolkit to control COVID-19 resurgence 55 [6, 7, 19, 28, 29] . In contrast to nation-wide lockdowns, localized lockdowns are implemented over a limited 56 geographical area, ranging from a neighborhood to a city, including suburbs, districts, or towns. Localized 57 lockdowns are appealing to policy-makers because, in principle, they would allow countries to reopen and 58 reclose specific jurisdictions based on local virus transmission dynamics. Large-scale lockdowns are 59 unsustainable because of the high costs they impose on the population [12] . Thus, compared to large-scale 60 interventions, localized lockdowns may control transmission hotspots while mitigating some social and 61 economic costs. Policy-makers need to make decisions about COVID-19 control strategies, considering 62 their epidemiological, social, and economic effects. 63 64 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted September 23, 2020. . https://doi.org/10.1101/2020.09.21.20198887 doi: medRxiv preprint Epidemiological evidence is one amongst several criteria for decision-making regarding non-65 pharmaceutical interventions. For example, a policy-maker would want to understand if costs of foregone 66 economic activity are proportional to the population under lockdown, or whether costs are mitigated or 67 amplified when lockdowns are implemented at different administrative levels (e.g., municipality, city, 68 state, country). On the one hand, demand spillovers would suggest that people in a municipality could buy 69 in stores of the neighboring municipality and, through substitution, limit the economic fallout in the city 70 as a whole. On the other hand, the fall in economic activity could be more than proportional if a lockdown 71 affects supply chains, such as when workers cannot work in a neighboring municipality because of 72 mobility restrictions in their municipality of residence. The answer to this question is mostly empirical, as 73 there are good arguments to both sides. However, there is limited and non-conclusive evidence on the 74 economic costs of non-pharmaceutical interventions. Researchers 

Drawing on a rich integrated dataset, including value-added tax (VAT) revenues, population data, and 104 daily incidence of lab-confirmed COVID-19 cases, we use econometric methods to empirically estimate 105 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted September 23, 2020. We limited our analysis of the 170 municipalities with above-median total VAT in 2018, excluding 128 mostly small and rural municipalities. This preferred sample of municipalities includes 97% of Chile's 129 2018 VAT and 89% of the population (Figure 2 ). Our sample also excluded the three municipalities that 130 concentrate large-company headquarters (Santiago, Las Condes, and Providencia), such as banks and 131 mining companies, because VAT data in these municipalities do not reflect local economic activity. 

Our main empirical specification is a two-way fixed-effects model: 143

where ∆%VATit corresponds to the percent variation of total VAT collected in municipality at month in 145 2020 relative to the same month in 2019.

is our variable of interest and represents the 146 proportion of days in a month that a municipality was under lockdown. and correspond to 147 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. in a conurbation may differ from more relatively isolated municipalities with no neighboring urban areas 156 ("standalone" municipalities). To examine whether the impact of lockdowns on economic activity is 157 heterogeneous depending on whether municipalities belong to a conurbation or are a standalone 158 municipality, we used the following regression specification: 159

where standalonei takes a value of one for standalone municipalities and zero otherwise. 161

The economic effects of localized lockdowns may also differ depending on the area under lockdown-for 163 example, at the municipality or conurbation level. To examine this question, we also ran our analysis 164 comparing all municipalities within a conurbation with standalone municipalities. We weighted the 165 number of days in quarantine in month of each municipality belonging to the conurbation according 166 to the total 2018 VAT: 167

CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted September 23, 2020. . https://doi.org/10.1101/2020.09.21.20198887 doi: medRxiv preprint

We estimated deaths and COVID-19 cases as a weighted average of deaths in municipalities within the 169 conurbation, using the municipality's population as the weight. Hence, the equation describing per capita 170 COVID deaths in each conurbation is as follows: 171

Last, we investigate how mobility at the municipality-level affects economic activity. We used a mobility 173 index based on cellphone data. 174 175

Descriptive statistics 177 Table 1 shows the main descriptive statistics of our sample. Figure CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted September 23, 2020. . https://doi.org/10.1101/2020.09.21.20198887 doi: medRxiv preprint 188 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. 

. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted September 23, 2020. (1) shows that one month of lockdown decreases monthly VAT around 12.5% (β:-0.125; 95% CI:-0.220,-206 0.031; p<0.01). The coefficient or the effect of lockdowns has about the same magnitude when restricting 207 the sample to municipalities with at least 50% of the urban population ( 

. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted September 23, 2020. . https://doi.org/10.1101/2020.09.21.20198887 doi: medRxiv preprint .We then limited our sample to urban municipalities (n=72) that are part of a conurbation ( We added an interaction term to examine whether lockdowns had a different effect on VAT in 232 municipalities that are part of a conurbation or in standalone municipalities. The results in Table 2 , column 233 (6) show a 23% decline in monthly VAT collection due to a one-month lockdown (β=-0.230; 95% CI:-234 0.345,-0.115; p<0.01). However, we did not find evidence of a differential effect for standalone 235 municipalities relative to municipalities in conurbations. 236 237 Last, we examined whether perceived threat or risk from COVID-19 deaths or cases could be an omitted 238 variable bias in the effect of local lockdowns on economic activity. Table 2 , Column (7) includes the 239 municipality's one-month lagged per-capita COVID deaths per 100,000 population as control. The effect 240 of lockdown is roughly the same as in column 1 (β=-0.125; 95% CI:-0.265,0.013; p<0.10). Controlling 241 for COVID-19 monthly incidence per 100,000 population Table 2 , Column (8) shows that one month of 242 lockdown results in a thirteen percent decrease in VAT collection (β=-0.135; 95% CI:-0.237,-0.033; 243 p=0.01). Results are robust to using one-month lagged COVID-19 deaths and cases. 244 245 Overall, Table 2 suggests that one month of lockdown would reduce economic activity by 10-15%, robust 246 to several model specifications. Notably, the effect size is not affected when controlling for COVID-19 247 deaths or case incidence, suggesting that the lockdown effect in this sample goes over and above the 248 impact of perceived threat or risk of contagion. 249 250 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this this version posted September 23, 2020. . https://doi.org/10.1101/2020.09.21.20198887 doi: medRxiv preprint

Next, we examined the effects of lockdowns on VAT when analyzed for conurbations or standalone 252 municipalities (Table 3 ). The objective was to test whether the effects of lockdowns were different when 253 there were no spillovers from closely interdependent neighboring areas. For the analysis, we collapsed 254 municipalities into conurbations. Our sample now had eighteen conurbations and seventeen standalone 255 municipalities in our baseline sample. 256 257 Table 3 , column (1), shows a statistically significant decline in monthly VAT collection of around 18% 258 (β=-0.184; 95% CI:-0.360,-0.009; p<0,050). Because Greater Santiago had the largest number of 259 municipalities with lockdown, we dropped Greater Santiago from the sample to test our results (Table 3,  260 column 2). The magnitude of the effect remained but was significant only at the 90% level (β=-0.188; 261 95% CI:-0.382, 0.051; p=0,056). In Table 3 , column 3, we examined whether there was a differential 262 effect for standalone municipalities. The results show that one month of lockdown results in a significant 263 decrease of 24% of VAT collection (β=-0.243; 95% CI:-0.370,-0.117; p<0,001); we did not find evidence 264 for a differential effect in standalone municipalities. However, the coefficient in Table 3 , column 3 was 265 not statistically different from the coefficient in Table 3 , columns (1) and (2). 266 267 Last, we examined whether the lockdown effect was different from the perceived threat or risk from 268 COVID-19. In Table 3 , columns (4) and (5) show lockdowns were no longer statistically significant at 269 conventional levels (p=0.240 and p=0.175, respectively). However, the coefficient's sign was still 270 negative and about the same magnitude as the coefficient in Table 3 , columns (7) and (8). The joint 271 significance test for the proportion of the month under lockdown and lagged per capita COVID-19 deaths 272 and incidence was significant (F=3.84, p<0.05; F=2.81, p=0.064, respectively). Thus, working with data 273 at the conurbation-level instead of the municipality-level makes it harder to disentangle the effect of 274 lockdowns. This is partly explained by insufficient statistical power and, from limited variation in the 275 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this this version posted September 23, 2020. . https://doi.org/10.1101/2020.09.21.20198887 doi: medRxiv preprint lockdown variable. The last columns of Table 3 reinforce the advantage of our baseline setting at the  276 municipal level, with more sizable variation in the lockdown (key) variable. 277 

. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this this version posted September 23, 2020. . https://doi.org/10.1101/2020.09.21.20198887 doi: medRxiv preprint 283 Table A1 in the Supplemental material also shows that our baseline results are robust to controlling for a 284 measure of cellphone-based mobility. However, we also argue that it might be misleading to control for 285 mobility since it is one of the main mediating channels by which lockdown affects economic activity (see 286

Supplemental material for further discussion). 287 Table 2 by an average short-run labor-to-economic activity elasticity of 0.4. 304 305 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this this version posted September 23, 2020. . https://doi.org/10.1101/2020.09.21.20198887 doi: medRxiv preprint It is also useful to contrast our results with the polar case of South Korea, without lockdowns. Aum et al. 306

[24] found that a one-per-thousand increase in the infection rate was associated with an employment loss 307 of 2-3%. Extrapolating this result to the United States and the United Kingdom, which had large-scale 308 lockdowns, Aum et al. [24] argue that only half of the 5-6% drop in job losses in these Western 309 economies might be attributable to lockdowns. The rest would be from social panic, some other large-310 scale non-pharmaceutical intervention, such as school closures, or demand effects. This similar effect of 311 areas with and without lockdown seems consistent with our findings. Importantly, we obtained our results Our estimates have limitations. First, we used a tax payment as a proxy for economic activity. 340

Nonetheless, we also have VAT and survey-based employment at the regional level in Chile. We found a 341 statistically significant elasticity of 0.3 between the drop in VAT and the decline in total employment 342 (including self-employed), consistent with short-run output-employment elasticities in the literature (34). 343

Another limitation is that informal economic activity is, by its very nature, not directly captured in our 344 measures of VAT. This is a more general limitation globally. However, compared to Latin America, Chile 345 has relatively low informality [53] . 346

Our study may also have other confounders. For instance, the government gave some leeway on when to 347 pay taxes, and we could only examine monthly-level observations. Nevertheless, there are no apparent 348 reasons why these confounders may interact with lockdowns. These confounders may have also 349 introduced measurement error in our tax measures. This would have increased our standard errors, 350 making it more difficult to get statistical significance. Nevertheless, we did get relevant and robust 351 estimates across various specifications, which mitigates the concerns related to measurement error. 352 353 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted September 23, 2020. . https://doi.org/10.1101/2020.09.21.20198887 doi: medRxiv preprint

We used a rich dataset of localized lockdowns in Chile to measure their effect on economic activity. We 355 find sizeable impacts of lockdowns, doubling the drop in economic activity compared to non-treated 356 municipalities, and robust to several model specifications and controls. As many countries are beginning 357 to reopen and ease mobility restrictions, localized lockdowns can be a critical tool to control COVID-19 358 resurgence while minimizing economic impact. We found no evidence that localized lockdowns generate 359 a proportionally larger or smaller effect in the economy when applied to areas of different sizes. 360

Critically, our results suggest that epidemiological criteria should guide decisions about the optimal size 361 of lockdown areas; the proportional effects of lockdowns on the economy seem to be unchanged by scale. 362

We thank Sebastian Piña (Universidad de Chile) for research assistance, Paula Aguirre (Pontificia 364

Universidad Católica de Chile) for help with Figure 1 , and the Servicio de Impuestos Internos (Chilean 365

Tax Authority) for the local Value-added-tax data. 366 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted September 23, 2020. . https://doi.org/10.1101/2020.09.21.20198887 doi: medRxiv preprint Mobility 516 We investigated whether mobility affects economic activity (Table S1 ). In a simple regression, with and 517 without controlling for lockdowns, the mobility index had a non-significant effect on economic activity. 518 Table S1 , column (1) suggests that lockdowns continue having sizable effects even after controlling for 519 mobility. We found no significant effects of mobility on economic activity (Table S1, column 2). 520

Since lockdowns and mobility could work in the same mechanism, in column (3) of Table S1 , we use the 521 method of instrumental variables. This is a method to analyze how lockdown-induced shocks to mobility 522 impact economic activity. For expositional purposes, this is done in two stages. In the bottom panel of 523 column (3), the so-called first-stage has a good fit, meaning that lockdowns impact mobility. On the top 524 panel, the second stage regresses VAT on the lockdown-induced changes in mobility which were 525 calculated in the first stage above. This second stage has a large and significant coefficient of 0.79. To get 526 correct standard errors, these two stages are jointly estimated. 527

Importantly, this method of instrumental variables tries to decompose the effect on lockdowns on 528 mobility and the subsequent impact of mobility on economic activity. The first coefficient means that a 529 month of lockdown changes monthly mobility by minus 15%. The second coefficient means that 530 lockdown-induced mobility changes VAT by +79%. The multiplication of these effects gives a sense of 531 the net impact of lockdowns on VAT. The multiplication (-0.15 × 0.79) yields a minus 0.11. This is 532 reassuring, since it falls within the range of our baseline estimates in Table 2 . Notably, while there could 533 be transmission mechanisms by which lockdowns affect economic activity beyond mobility, these results 534 suggest that mobility is the leading mechanism. 535 536 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted September 23, 2020. . https://doi.org/10.1101/2020.09.21.20198887 doi: medRxiv preprint

The effect of large-scale 385 anti-contagion policies on the COVID-19 pandemic

COVID-19 and strategies for mitigation and suppression in low-and middle-income countries

COVID-19): A Review

Estimating the 393 effects of non-pharmaceutical interventions on COVID-19 in Europe

Effects of non-395 pharmaceutical interventions on COVID-19 cases, deaths, and demand for hospital services in the 396 UK: a modelling study. The Lancet Public Health

Suppression of a SARS-CoV-2 outbreak in the Italian municipality of Vo'

Reconstruction of the full transmission dynamics 400 of COVID-19 in Wuhan

New Zealand eliminates COVID-19

Effect of non-pharmaceutical 408 interventions to contain COVID-19 in China

Answering 20 more questions on COVID-19

Resurgence of covid-19 in Japan

Assessing the impact of coordinated COVID-19 exit strategies across Europe

Covid-19: Risk of second wave is very real, say researchers

Projecting the transmission 417 dynamics of SARS-CoV-2 through the postpandemic period

Effectiveness of Localized Lockdowns in the SARS-CoV-2

Harnessing multiple 421 models for outbreak management

Local lockdowns outperform global lockdown on the far 423 side of the COVID-19 epidemic curve

An interactive web-based dashboard to track COVID-19 in real time

The Economy in the Time of CoVID-19

Unemployment effects of stay-at-home orders: Evidence 429 from high frequency claims data. Institute for Research on labor and employment working paper

Employment to output elasticities & reforms 457 towards flexicurity: Evidence from OECD countries

How South America became the new centre of the coronavirus pandemic

COVID-19 in Latin America

Epidemiological 462 and clinical characteristics of the COVID-19 epidemic in Brazil

How prepared is Brazil to tackle the COVID-19 disease?

How Latin America is fighting covid-19, for better and worse

Plan de acción por coronavirus

Impact of small-470 area lockdowns for the control of the COVID-19 pandemic. medRxiv

All Things Equal? Heterogeneity in Policy Effectiveness against COVID-19 Spread 472 in Chile

CoVID-19 has exposed how 'the other half' (still) lives

Medium-run Local Economic Effects 476 of a Major Earthquake. Unpublished Manuscript

Base de datos CoVID-19

Using social and 484 behavioural science to support COVID-19 pandemic response