ABSTRACT
Non-pharmaceutical interventions, such as social distancing and lockdowns, remain the most viable measure to control the COVID-19 pandemic. Localized lockdowns in small geographic areas have become an important policy intervention to prevent viral spread in cases of resurgence. These localized lockdowns can result in lower social and economic costs compared to larger-scale suppression strategies. In this paper, we adopt a novel approach from the causal inference literature and estimate the effect of localized lockdowns using augmented synthetic controls. Utilizing a rich integrated dataset from Chile, we estimated the direct and indirect causal effects of localized lockdowns on SARS-CoV-2 transmission. Our results show that the effects of localized lockdowns are strongly modulated by their duration and are affected by indirect effects from neighboring geographic areas that are not under lockdown. Our estimations suggest that extending localized lockdowns will slow down the epidemic. However, by themselves, localized lockdowns will be insufficient to control epidemic growth due to indirect effects from neighboring areas unless those contiguous areas also implement lockdowns. These results provide critical empirical evidence about the effectiveness of localized lockdowns in interconnected geographic areas.
Despite encouraging news about the effectiveness of COVID-19 vaccines and their approval by regulatory agencies (1-3), behavioral non-pharmaceutical interventions will remain an essential strategy to prevent and control the transmission of the severe acute respiratory syndrome-coronavirus 2 (SARS-CoV-2) until a substantial proportion of the population has immunity (4, 5). Non-pharmaceutical interventions range from simple individual-level recommended behaviors, such as wearing face masks, frequent hand-washing, or maintaining physical distance, to society-level regulatory actions, such as school closures, quarantines, or lockdowns (6, 7). Efforts to control epidemic growth based on these interventions have been successful in some countries (8-11). To date, the effects of non-pharmaceutical interventions have mainly been described using compartmental models (6, 7, 10, 12, 13) and have informed policies around the world since the beginning of the pandemic (14, 15). In this paper, we adopt a complementary approach from the causal inference literature (16, 17). Arguably, health policy impact evaluations require a variety of study designs, data sets, and analytic approaches that support each other to provide stronger evidence (18). In general terms, this approach seeks to estimate the causal effect of localized lockdowns by emulating the target randomized experiment (trial) that would have been conducted under ideal circumstances to evaluate the implemented policy (19-22).
As the COVID-19 pandemic develops across countries, policymakers need evidence to help them decide when and how to ease mobility restrictions or strengthen these restrictions in cases of resurgence. Even now that some countries have started vaccinating, large-scale non-pharmaceutical interventions continue to be important, particularly in low- and middle-income countries, to avoid large increases in the number of cases (5). In this context, localized lockdowns are becoming an increasingly relevant policy option (23-28). Localized lockdowns are typically implemented in transmission hotspots and can be applied to populations or areas large and small to suppress an outbreak. Localized lockdowns had not been widely used as a public health response to contain outbreaks until the current pandemic (23-28). In principle, localized lockdowns impose fewer social and economic costs compared to larger-scale SARS-CoV-2 suppression strategies and are thus more sustainable. They can also provide a gradual exit from nationwide lockdowns. Early in the pandemic, for example, the Chinese government imposed a localized lockdown and other strict non-pharmaceutical interventions in the city of Wuhan (10), effectively suppressing SARS-CoV-2 transmission (29). Subsequently, governments have implemented targeted lockdowns in neighborhoods (e.g., Beijing, China), suburbs (e.g., Melbourne, Australia), towns (e.g., Vo, Italy), districts (e.g., North Rhine-Westphalia, Germany), and at the city level in Leicester, England (8, 24). Despite the increasing importance of localized lockdowns, there is limited empirical evidence of their effectiveness.
In this study, we use data from Chile to estimate the effect of localized lockdowns on COVID-19 transmission. Our data set combines information from administrative COVID-19 surveillance records (30), a nationally representative household survey (31), and census data (32). We use a synthetic control approach (33, 34) to build control intervention units (municipalities) with similar sociodemographic features, trajectories of contagion, and histories of lockdowns until the time of the policy intervention, taking into account the spatio-temporal structure of the data. In other words, we assess whether the effectiveness of a localized lockdown implemented at the municipality level was affected by the lockdown status of its neighboring municipalities. This indirect effect may play an important role in municipalities within cities or urban areas where social and economic interdependencies exist. Allowing for such indirect effects or interference between municipalities (35, 36), we estimate the direct effects of extending the duration of localized lockdowns and the total (sum of direct and indirect) effects of maintaining lockdowns in neighboring municipalities.
METHODS
Lockdowns in Chile
The Ministry of Health reported the first COVID-19 case in Chile on March 3, 2020 (37). By the end of March, the government had restricted large gatherings (March 13), closed schools and universities (March 16), increased controls on national borders (March 18), enforced night-time curfews (March 22), and imposed mandatory use of facemasks in public at the national level (April 8) (37) (Figure 1). In Chile, the Ministry of Health implemented all epidemic interventions at the national level, including individual-level recommendations and regulatory actions (37). During the study period (March 3 through June 15, 2020), the exceptions were localized lockdowns implemented at the municipality level. Starting on July 19, 2020, the government initiated a gradual reopening scheme of five gradual steps, also implemented at the municipality level (38).
In Chile, policymakers implemented localized lockdowns at the municipality level, the smallest administrative subdivision in the country, at various points in time (Figure 2A, 2B). The government loosely defined the criteria used to impose lockdowns as a function of the number and density (per km2) of infectious COVID-19 cases, increases in case incidence, and health system capacity (37). Most of the time, lockdowns were implemented at the municipality level. However, in some cases, the lockdowns included only a proportion of the population within a municipality. Across the country, there was substantial variation in the duration of these municipality-level localized lockdowns and, for each municipality under lockdown, in the lockdown status of neighboring municipalities. As a result, the effectiveness of lockdowns varied geographically. We used this policy variation as a natural experiment to evaluate the effectiveness of localized lockdowns on SARS-CoV-2 transmission.
Study overview
Based on the potential outcomes framework for causal inference (16, 17, 39, 40), we used the augmented synthetic control method to analyze the epidemic’s progression in comparable municipalities that underwent different lockdown interventions (33, 34). For an individual municipality, we varied the duration of the intervention and the proportion of the population under lockdown in the neighboring municipalities at each time point, controlling for relevant covariates that could modify these effects (Figure 3B). We estimated the counterfactual progression that the disease would have exhibited had an alternative lockdown policy taken place (Figure 3B). We provide open-source code with step-by-step explanations to replicate the analyses and implement them in related settings (see the Supplementary Material).
Integrated surveillance records, survey measurements, and census data
Our data set combines information from administrative COVID-19 surveillance records, a nationally representative household survey, and census data. Specifically, we use epidemiologic surveillance records from the Department of Epidemiology of the Chilean Ministry of Health (37). COVID-19 cases are defined in our dataset as symptomatic and asymptomatic SARS-CoV-2 infections confirmed by a positive PCR test. Chile has tested for COVID-19 at a higher rate than any other Latin American country (41, 42) and counts with a high level of effective universal health coverage (43). We characterized municipalities based on Chile’s National Socioeconomic Characterization Survey (CASEN), a nationally representative household survey of the Chilean Ministry of Social Development that collects data on education, employment, income, health, and housing (31). Finally, we employed population data from the 2017 National Census (32). All data are publicly available.
We adjusted the COVID-19 case incidence series to correct the lag in reporting and some incomplete municipality-level data. First, we imputed the incomplete data by interpolating between the closest dates with complete data. In the data, the number of cumulative cases was reported typically every 2-4 days. Second, we estimated the lag in reporting using the PELT algorithm (44). Third, we adjusted the incidence series for space- and time-varying reporting lags using the approach by Zhao et al. (45). These adjustments considered that the lag between symptom onset date and report date could vary across municipalities and over time. We estimated the instantaneous reproduction number following Cori et al. (46) using the adjusted COVID-19 series. See the Supplementary Material for details.
Instantaneous reproduction number
We characterized transmission by the instantaneous reproduction number (Rt); that is, the average number of secondary cases per primary infected case (46). We did this using the daily series of COVID-19 cases reported by the Ministry of Health (37), adjusted for the time-lag between onset of symptoms and case report (Figure 1) (45, 46). Following Cori et al. (46), the instantaneous reproduction number can be estimated by, where It is the incidence at time t and ws is the infectivity function or density of the serial interval at time s. Cori et al. (46) propose estimating Rt over a window of time τ as See Gostic et al.(47) for a discussion of the method by Cori et al.(46) and other related approaches to estimating the instantaneous reproduction number. Based on data from the SARS-CoV-2 outbreak in Wuhan, Li et al. (48) estimated that the serial interval distribution had a mean of 7.5 days, a standard deviation of 3.4 days, and a 95% confidence interval of [5.3, 19] days. In conformity with Cori et al. (46), we assumed that the serial interval has a Gamma distribution with a mean of 7.5 days and a standard deviation of 3.4 days and τ = 5. Assuming homogeneous mixing of the population within a municipality, we calculated Rit for each municipality i (Figure 3A).
Potential outcomes
We base our analysis on the potential outcomes framework for causal inference (17, 39, 40) under interference (35, 36, 49). Let denote the lockdown indicator for municipality i = 1, …, I in cluster ji = 1, …, Ji at time t = 1, …, T, with if the municipality is under lockdown at time t, and otherwise. In our analysis, since the cluster of the municipality i is defined as the union of i and its adjacent municipalities, we can omit the index ji for simplicity in the notation, but our approach is more general. Write Lit for the lockdown history of municipality i until time t. Analogously, define P(i)t as the proportion of the population in the cluster of municipality i under lockdown at time t, excluding municipality i, and P(i)t as the corresponding history until time t. In this framework, ℓit can also denote the proportion of the population under lockdown in municipality i, beyond whether or not the municipality is under lockdown. Following the frameworks for causal inference under interference by Sobel (35) and Hudgens and Halloran (36), we put Rit (Lit = ℓit, P(i)t = p(i)t) for the potential instantaneous reproduction number for municipality i in its cluster at time t under lockdown histories ℓit and p(i)t for the municipality and its neighbors until time t, respectively. Finally, we designate Iit (Iit−1(*), Rit (Lit = ℓit, P(i)t = p(i)t)) as the potential incidence in municipality i at time t, which is a function of its potential incidence history Iit−1(*) and potential instantaneous reproduction number Rit (Lit = ℓit, P(i)t = p(i)t).
Direct, indirect, and total effects of lockdowns
For any given municipality i, we want to estimate the individual effect of lockdowns on the instantaneous reproduction number Rit when intervening both on the duration of the lockdown Lit and on the proportion of the population under lockdown in the neighboring municipalities P(i)t. We are particularly interested in the municipality-level direct, indirect, and total effects of lockdowns across time (36). See the Supplementary Material for precise definitions of these estimands.
Synthetic controls
We used the augmented synthetic control method (33, 34) to estimate the potential instantaneous reproduction number Rit(*) and calculate potential incidence Iit(*) from Rit(*) according to (see the Supplementary Material for details). The intuition behind this method is, for a given lockdown intervention in municipality i, to build a synthetic control municipality with a very similar covariate, intervention, and outcome histories until the time of the intervention by appropriately weighting in time control municipalities. We adjusted for or balanced several municipality-level characteristics that may affect virus transmission (see Table S2 in the Supplementary Material). Adjustments included the proportion of the population that is rural, female, older than 65 years of age, living in poverty, living in overcrowded households (≥2.5 people per room), lacks adequate sanitation infrastructure (access to potable water and sewage), average monthly income, and municipality area. We also adjusted for the seven-day history of lockdown interventions in the municipality Li[t−7,t−1] and its neighbors Pi[t−7,t−1], and of the instantaneous reproduction number Ri[t−7,t−1]. We analyzed all the municipalities in Greater Santiago that started their first lockdown after March 15, 2020, and completed it by May 15, 2020 (Figure 2A); that is, the first period of confinement that arguably shaped the evolution of the epidemic (Figure 1, Figure 2B).
RESULTS
Duration and indirect effects of localized lockdowns
Overall, our results suggest that the effectiveness of localized lockdowns is strongly modulated by the duration of the intervention and the magnitude of the indirect effects from neighboring geographic areas. The larger the proportion of neighbors under lockdown, the higher the effectiveness to control transmission.
We illustrate our findings with three representative municipalities in Greater Santiago that were put under lockdown on March 26: Lo Barnechea, Providencia, and Santiago (Figure 4; see the Supplementary Information for additional results). We chose these municipalities because they were among the first municipalities to be put under lockdown, and they have a substantial interdependence with other municipalities in Greater Santiago. The municipality of Santiago concentrates the country’s financial, commercial, and political activity and all major government infrastructure. Providencia is an upper-middle-class urban municipality with the highest population over 65 years of age, high-rise apartment buildings, and substantial commercial activity. Lo Barnechea is largely a residential area, with limited commercial activity, few buildings, and a heterogeneous population. In the Supplementary Material (Figures S3-S6) we present additional results for other municipalities in Greater Santiago (Ñuñoa, Independencia, Las Condes, Vitacura) and elsewhere in Chile (Arica and Punta Arenas, the northernmost and southernmost cities in Chile, respectively); the findings for these municipalities are consistent with our main results.
Figure 5 shows a large reduction in Rt (Fig. 5A) and COVID-19 cases (Figure 5B) with an extended lockdown. Had the lockdown been extended for three additional weeks, maintaining Pt constant, we estimate that the reduction in Rt would have been larger. The average Rt would have decreased from 1.83 to 1.27 (difference: −0.56, 95% confidence interval [CI]: [-0.63,-0.50]) in Lo Barnechea, from 1.82 to 1.34 (difference: −0.47, 95%CI: [-0.59,-0.36]) in Providencia, and from 1.95 to 1.23 (difference: −0.72, 95%CI: [-0.85,-0.58]) in Santiago. These reductions in Rt are equivalent to 177 (95%CI: [167,188]; or 143 per 100,000 population) averted COVID-19 cases over three weeks in Lo Barnechea, 94 (95%CI: [76,111]; or 59 per 100,000 population) averted cases in Providencia, and 1343 (95%CI: [1245,1441]; 267 per 100,000 population) averted cases in Santiago, which would represent 33-62 percent reductions in reported cases in that timeframe.
The reductions in transmission would have been even larger if it was possible to control lockdowns in neighboring municipalities to reduce indirect effects. Assuming neighboring municipalities of Lo Barnechea, Providencia, and Santiago maintained their lockdown status (Pt =53.0%, Pt =80.3%, and Pt =35.8%) for three additional weeks, we estimate that the average Rt would have decreased to 1.19 (95%CI: 1.13, 1.25), 1.25 (95%CI: 1.14, 1.37), and 1.21 (95%CI: 1.08, 1.34), respectively (Figure 5A). Figures 6A and 6B show the relationship between daily COVID-19 incidence and days of extended lockdown as a function of changes in Pt, after adjusting for observed covariates. The larger Pt, the greater the number of averted cases. Overall, results in Greater Santiago suggest that the decision to reopen these municipalities was premature, especially when lockdowns were brief because the effectiveness of lockdowns strongly depends on the duration of the intervention and the magnitude of indirect effects (findings for other municipalities with lockdowns are consistent with these results; Figures S3-S6).
Lockdowns without indirect effects
Figure S5 in the Supplementary Material reaffirms the results from a different angle. As happened with Lo Barnechea, Providencia, and Santiago, the municipality of Punta Arenas in the south of Chile was placed under lockdown early on in the pandemic, from April 1 to May 7. It initiated lockdown with one of the highest case incidence per 100,000 population in the country. Notably, Punta Arenas is geographically isolated and has few local interdependencies that could result in active transmission networks during a localized lockdown. Our estimates show negligible indirect effects: increasing Pt from 0 to 1 would only result in a reduction of Rt =0.02 (Figure S5 and Table S13), probably due to its geographical isolation and minor interdependencies with neighboring municipalities.
Varying the area under lockdown
Having assessed the role of duration and indirect effects, we evaluate the impact of lockdowns in geographic areas of increasing size. We considered three target lockdown areas (Figure 7A): the municipality of Ñuñoa (red), a cluster of six municipalities (orange), and Greater Santiago (green). We extended the study period to encompass the mandatory lockdown for Greater Santiago that began on May 15. We varied the population under lockdown in the targeted area and the proportion of the population under lockdown in neighboring municipalities (Pt). Note that the municipality of Ñuñoa had only about half its population under lockdown between April 14 and May 7, as indicated by the blue line in Figure 7B. Figure 7B shows the estimated Rt from March 15 to June 15. In general, an epidemic will continue to grow as long as Rt is greater than one. Figure 7 shows that the epidemic kept expanding in all three target areas until a city-wide lockdown was implemented on May 15. These results highlight the challenges of suppressing virus transmission in areas with a high degree of economic and social interdependencies, such as Chile’s capital, when there is a substantial proportion of neighbors that are not under lockdown.
DISCUSSION
Using augmented synthetic control methods, we estimated the effects of localized lockdowns on COVID-19 transmission, incorporating the effects of lockdown duration as well as lockdowns on neighboring areas. We found that localized lockdowns can help contain the transmission of the virus. However, their effectiveness is dependent on the duration of the intervention and potential indirect effects from neighboring geographic areas. For instance, the effectiveness of localized lockdowns within Greater Santiago, where there is high economic and social interdependency between municipalities, was strongly affected by suppression measures that were in place in neighboring municipalities. As expected, the larger the proportion of neighbors under lockdown, the higher the effectiveness to control epidemic growth. However, our estimates also show that in Greater Santiago, the epidemic is only controlled (i.e., Rt < 1) when generalized lockdowns are in place. In contrast, localized lockdowns showed promising results in municipalities such as Punta Arenas, which are geographically isolated and thus have transmission networks that are relatively unaffected by neighboring areas.
It is possible that the effectiveness of localized lockdowns is reduced over time, as suggested by the results in Figure S7 (Supplementary Material). On the one hand, lockdowns may be particularly challenging for socially vulnerable individuals who depend on daily wages, have limited savings, or do not receive external support, as their need to secure income may be at odds with policy (50). On the other hand, lockdowns may not immediately decrease Rt to their full potential because transmission within households or other residences, such as nursing homes, account for a substantial proportion of new cases (10, 51). There are sustained close contacts within households, particularly where people cannot isolate themselves in a separate room or where members share spaces such as restrooms (52-54). It is also possible that lockdowns may not be enough to contain an epidemic. For instance, Rt was significantly reduced in Wuhan with the implementation of a city-wide lockdown. However, control of the epidemic (Rt < 1) was achieved through a system of centralized quarantines and treatment for COVID-19 patients in field hospitals (10, 29). Our results suggest that the effectiveness of lockdowns is dependent on the duration of the intervention and of potential indirect effects from neighboring geographic areas so that extending lockdowns until Rt < 1 may not be possible if interdependencies exist.
Both our analysis and the currently available data have limitations. First, our analysis is based on reported COVID-19 cases by the Ministry of Health (30), which may be affected by underreporting. SARS-CoV-2 infection can result in a broad spectrum of clinical outcomes, including asymptomatic infection, mild symptoms, hospitalization, or death (55-57). Mild cases without apparent symptoms may be undetected (29, 58, 59). However, Chile has conducted tests at a substantially higher rate than any other country in Latin America, with approximately 130 total RT-PCR tests per 1,000 people by August 31 2020 (42). Closely related, it is possible that differences in health-seeking behavior and the severity of illness, for example, by age group, may bias the COVID-19 case data. However, since the population characteristics hardly changed during the study period, this possible bias should be stable and thus not significantly affect the results as long as reporting rates are consistent in time. Second, in Chile the Ministry of Health does not report COVID-19 cases by the onset of symptoms but by the day of reporting to the health system. We addressed this limitation by adjusting the time series, according to Zhao et al. (45). Third, we assumed that the entire population is susceptible to SARS-CoV-2. While heterogeneity in population immunity across municipalities could affect infection rates, our model uses data from the early stage of the epidemic in Chile (March 3–June 15), so this assumption should not affect our results. Please recall the first COVID-19 case in Chile was reported on March 3, and lockdowns began three weeks later. Fourth, the causal validity of our estimates is predicated on identification assumptions (such as direct interference and ignorable treatment assignment) often invoked in the causal inference literature. While strong, these assumptions are necessary to identify the direct and indirect effects of lockdowns with the available data. See the Supplementary Material for further details.
Last, to reduce the complexity and increase the generalizability of our model, we assumed that Rt in a municipality is affected only by its neighbors. There is commuting beyond neighboring municipalities that may be non-negligible. However, there are two reasons why we think that the main transmission route is from people moving within the municipality or from neighboring municipalities. The most recent mobility survey before the pandemic suggests that about 40% of trips in Greater Santiago were walking or in bicycle, followed by car (26%) and public transport (24%) (60). About half the car trips in Greater Santiago were less than five kilometers, 10% of which were less than one kilometer (60). Furthermore, mobility was greatly reduced during the pandemic. Estimates from the Inter-American Development Bank suggest traffic decreased by 54% between March 2-8 and the first week of June in Greater Santiago (61). The Ministry of
Transport estimated a 59% decrease in the same period (62). The second reason is that lockdowns were enforced at the municipality level (37). When under lockdown, individuals need a valid police-issued permit to be on the street. We lack data on mobility between specific municipalities to test this hypothesis. Our estimation approach can be modified to control for a network of influence that does not necessarily correspond to the nearest geographic neighbors but to the most influential entities through P(i)t in the potential instantaneous reproduction number.
It is well-known that the only way to stop an epidemic is to break the transmission chain. Today, until an effective vaccine becomes available for a substantial proportion of the population, strategies to suppress SARS-CoV-2 transmission are mostly limited to non-pharmaceutical interventions (4). Large-scale interventions have imposed high social and economic costs to societies worldwide (50, 63-65). By being less disruptive than large-scale interventions, localized lockdowns could help reduce those costs and provide a more sustainable strategy in time (66). In principle, localized lockdowns can break transmission chains by limiting contact between infectious and susceptible individuals, and this goal could be achieved at household, neighborhood, municipality, county, or state levels. However, the social distancing imposed by a lockdown must be maintained and enforced until adequate control of transmission is achieved. The effectiveness of non-pharmaceutical interventions depends on the willingness and capacity of the population to comply. Compliance may be particularly challenging in low- and middle-income countries, where a substantial proportion of the population works informally (67). Hence, the willingness to comply with stay-at-home restrictions is at odds with the need to secure income on a daily basis (50, 68). Another advantage is that localized lockdowns may provide a gradual, more controlled exit to larger-scale strategies at a lower cost if effectively implemented. This study shows that effective implementation of localized lockdowns is challenging and is affected by indirect effects from neighboring areas where transmission networks exist, such as in a city.
Data Availability
Datasets used and/or analyzed during the current study are available from Base de Datos CoVID-19 repository, Ministerio de Ciencia, Tecnologia, Conocimiento, e Innovacion.
Grants and financial support
This work was supported by a grant from the Alfred P. Sloan Foundation (G-2018-10118), an award from the Patient-Centered Outcomes Research Institute (PCORI; ME-2019C1-16172), Agencia Nacional de Investigación y Desarrollo (ANID) Millennium Science Initiative MICROB-R, (NCN17_081), and ANID/FONDAP CIGIDEN (15110017).
Conflict of interest
Authors declare no competing interests.
Data and material availability
The data sets used in this study are available from Base de Datos COVID-19 repository at http://www.minciencia.gob.cl/covid19. The R code used for analysis is available at https://scholar.harvard.edu/files/zubizarreta/files/code_v1.0.zip.
ACKNOWLEDGEMENTS
The authors thank Eric Cohn (Harvard School of Public Health), Dr. Eduardo Engel (Universidad de Chile), Dr. Catterina Ferreccio (Pontificia Universidad Católica de Chile), Dr. César Hidalgo (University of Toulouse), Dr. Ronald Kessler (Harvard Medical School), Dr. Mauricio Lima (Pontificia Universidad Católica de Chile), Dr. Guillermo Marshall (Pontificia Universidad Católica de Chile), Dr. Gonzalo Mena (University of Oxford), Dr. Bijan Niknam (Harvard University), Dr. Paul Rosenbaum (University of Pennsylvania), and Dr. Zirui Song (Harvard Medical School) for helpful comments and suggestions.
Footnotes
Updated abstract and introduction for added clarity. Updated the organization of figures (separated panels), and extended the supplementary material for added clarity in the methods.