Abstract
After the COVID-19 outbreak, China immediately adopted stringent lockdown policies to contain the virus. Using comprehensive death records covering around 300 million Chinese people, we estimate the impacts of city and community lockdowns on non-COVID-19 mortality outside of Wuhan. Employing a difference-in-differences method, we find that lockdowns reduced the number of non-COVID-19 deaths by 4.9% (cardiovascular deaths by 6.2%, injuries by 9.2%, and non-COVID-19 pneumonia deaths by 14.3%). The health benefits are likely driven by significant reductions in air pollution, traffic, and human interactions. A back-of-the-envelope calculation shows that more than 32,000 lives could have been saved from non-COVID-19 diseases/causes during the 40 days of the lockdown on which we focus. The results suggest that the rapid and strict virus countermeasures not only effectively controlled the spread of COVID-19 but also brought about massive unintended public health benefits. These findings can help better inform policymakers around the world about the benefits and costs of city and community lockdowns policies in dealing with the COVID-19 pandemic.
Introduction
By the end of June 2020, COVID-19 had affected more than 216 countries and caused more than 550, 000 deaths worldwide1. Facing this unprecedented crisis, different countries adopted various measures to mitigate its impacts, ranging from one extreme, where governments imposed draconian measures to restrict human mobility immediately after the outbreak, to the other extreme, where governments were reluctant to adopt any serious disease preventive measures and explicitly resorted to herd immunity. Effective policies not only depend on the social preferences of people and the capacity of government but also depend on our accurate understanding of the costs and benefits of different counter-COVID-19 measures. However, relatively little is known about the broader impacts of these policies.
A key component when evaluating the welfare implications of the anti-contagion policies is their overall public health consequences. Multiple studies have shown that strict social distancing and human mobility restrictions can effectively control the spread of COVID-19 and thus save lives from the virus2-6. However, it remains unknown to researchers and policymakers how such interventions affect disease patterns and deaths from other causes. On the one hand, hospitals may decline nonurgent service requests (especially when the system is overburdened by COVID-19)7 and the fear of getting infected by COVID-19 may make patients reluctant to visit hospitals. This could impact the quality of health services and delay medical treatment, which would negatively impact the population health. Additionally, in many countries, the strict virus containment policies led to sudden and sharp economic disruption, causing massive layoffs8. As documented in the previous literature, such economic downturns and high unemployment could also damage population healtht9-15. All these factors would increase the mortality when strict counter-virus measures were enforced. On the other hand, because the virus containment policies significantly improved air quality, restricted human-to-human interactions, and reduced traffic volume, it is also possible that a large number of people could be saved from dying from air pollution, other types of communicable diseases, and traffic accidents17-19. Therefore, it is of great scientific and policy relevance to assess whether the counter-virus measures bring about additional public health gains or additional public health losses.
Using data from China, we examine how city and community lockdown policies affect nonCOVID-19 mortality. We focus on China because the country mandated strict social distancing and lockdown policies to control the virus. Within a few weeks after the COVID-19 outbreak in Wuhan, a large number of cities enforced strict quarantines, traced close contacts, prohibited public gatherings, mandated social distancing, and limited human mobility. A large number of cities were locked down even though they had less than 100 confirmed cases (Supplementary Fig 1 and Supplementary Fig 2). Exploiting the staggered introduction of city and community lockdowns in different cities of China, we estimate the impacts of lockdowns on the number of deaths from various causes and explore the channels through which these impacts are manifested. These results will help policymakers around the world design effective measures to mitigate the damages from the pandemic.
The core of our empirical analysis uses the comprehensive deaths record from China’s Disease Surveillance Points (DSPs) system, covering more than 324 million people in 605 DSP’s districts/counties in 321 cities, which accounts for 24.3% of the country’s population20, 21. This dataset includes cause-specific deaths, which allows us to examine the mechanism of lockdowns’ impacts on non-COVID-19 mortality. Each city’s lockdown information is collected from news media and government announcements (Fig 1). There are two types of lockdowns: city lockdown and community lockdown. The former is defined as mobility being restricted across different cities, and the latter is defined as restriction of mobility within a city. Matching these datasets, we construct a daily DSP site-level panel dataset from January 1 to March 14, 2020, which is the period largely overlapping with the coronavirus outbreak in China. Our dataset includes 393,133 death records that were reported to the DSPs system by May 15, 2020 (Supplementary Table 1). Note that we exclude 3 DSPs in Wuhan from the baseline analysis because the city is the epicenter of the outbreak in China, and we are concerned that its death reporting process could have been affected during the study period22.
To quantify the impacts of lockdowns on mortality, we employ a difference-in-differences (DiD) approach, which is an econometric approach and is widely used to infer causal impacts of various policies and events using observational data23. An advantage of this approach is that it compares the policy effects relative to the plausible counterfactuals. While the results from a before-and-after comparison could be driven by different mortality trends or other unobserved confounders, DiD compares the changes in mortality between the locked-down DSPs (treatment group) and the non-locked-down DSPs (control group) before and after the enforcement of lockdown policies. In other words, the control group can serve as a counterfactual, mimicking what would have happened in locked down DSPs in the absence of the lockdown, which essentially allows us to compare the policy effects relative to business as usual. Note that a key assumption of the DiD is that the treatment and the control group follow parallel trends in the number of deaths in the absence of the lockdown policies. We examine whether this assumption is likely to hold using an event-study test. We describe the model in more detail in the Materials and Methods.
Results
Impacts of City and Community Lockdowns on Non-COVID-19 Deaths
Fig. 1 summarizes the baseline regression results by fitting the DiD model (Materials and Methods: Equation A1; full results are in Supplementary Table 2). Panel A reports the effects on the number of deaths, while Panel B reports the percentage change. In row (1), we find that lockdowns overall have a negative impact on non-COVID-19 mortality. After human mobility is restricted, the DSP-level daily number of deaths decreased by 0.429 (or 4.92%), as compared to the control group.
In rows (2) to (7), motivated by several factors that could potentially affect population health during the lockdown period, we separately examine the effects on different causes of death. We are especially interested in the following three outcome variables: cardiovascular diseases (CVD), injuries, and (non-COVID-19) pneumonia deaths. Existing literature on the acute effects of air pollution suggests that elevated air pollution levels can significantly increase deaths from strokes, myocardial infarction, and other types of cardiovascular diseases24-25. We thus expect the number of deaths from CVDs may decrease due to the improved air quality27. As shown in row (2) of Fig. 1, we find that cardiovascular deaths were reduced by 6.2% (0.27 in levels) after lockdown. Relatedly, as the lockdown policies restrict production, social activities, and traffic, we expect the number of deaths from injuries (which include workplace injuries, traffic accidents, etc.) to also drop. The result in row (3) of Fig. 1 confirms this conjecture; we observe that the number of deaths caused by injuries decreased by 9.2% (0.044 in levels). In addition, as human mobility is greatly restricted during the lockdown period, this should reduce the likelihood of people getting infected by and dying from other types of bacteria and viruses that cause pneumonia. The result in column (4) shows that deaths from non-COVID pneumonia were reduced by a large margin of 14.7% (0.022 in levels) during the lockdown period.
In rows (5) to (7), we report the findings on several other causes of death that are less likely to be affected by short-term restrictions on human activities. They include deaths from neoplasms, chronic respiratory diseases, and other diseases. While the coefficients for these causes of death are also negative, they are all statistically significant. We thus conclude the temporary human mobility restrictions during China’s lockdowns primarily reduce the deaths caused by acute diseases and accidents and have a weaker impact on people with chronic diseases and cancers.
Some additional analyses complement our main findings. A key assumption of the DiD is that the treatment and the control group follow parallel trends in the number of deaths in the absence of the lockdown policies. Using an event-study approach, we show that this assumption is likely to be held (Fig. 2 and Supplementary Note 1 and Supplementary Table 3). Also, we find that our results are robust to the inclusion of additional controls, adoption of different weighting, and sampling (Supplementary Note 2, Supplementary Table 4, and Supplementary Fig. 3). Finally, we further disaggregate the data into more specific causes/diseases (Supplementary Table 5). For example, in the cardiovascular disease category, we observe that deaths from myocardial infarction, strokes, and other types of cardiovascular diseases all significantly decreased after the lockdown.
Heterogeneity
In Fig 3, we examine the heterogeneous impacts of lockdowns on mortality. Here we report our findings on the total number of non-COVID-19 deaths and explore the following dimensions: baseline income (measured by per capita GDP in 2018), healthcare resources (measured by hospital beds per thousand people in 2018), air pollution levels (measured by average PM2.5 concentrations in 2019), industrial structure (measured by the share of employment in manufacturing industries in 2015), and initial health status (measured by mortality rate in 2019).
To do so, we interact the lockdown indicator separately with each of the heterogeneity dimensions in the regression (Supplementary Table 6), and then plot the predicted impacts and their 95% confidence intervals in Fig 4. We observe significant heterogeneities with respect to the air pollution level, the employment shares in the manufacturing industries, and the baseline mortality level. Specifically, the health benefit of lockdowns on mortality is greater when a DSP is more polluted and more industrialized, and when the initial health status is worse.
We also repeat this exercise separately for deaths from specific causes: cardiovascular diseases, injuries, and non-COVID-19 pneumonia (Supplementary Fig 4). Several patterns stand out: (1) for cardiovascular diseases, there exist significant heterogeneities for air pollution and industrial structure, with more polluted and more industrialized cities seeing fewer deaths from cardiovascular diseases during lockdowns relative to other cities (Panel a); (2) for injuries, the more industrialized the DSP, the higher its initial injury mortality, and, as expected, the greater the impact of the lockdown (Panel b); (3) for pneumonia, we only observe significant heterogeneity with respect to initial mortality rate, i.e., cities with a higher initial pneumonia mortality rate are more strongly affected by lockdowns (Panel c). Across all the causes of death, per capita GDP and availability of healthcare resources do not seem to play an important role in terms of magnitude, although occasionally they are statistically significant. The corresponding regression results are reported in Supplementary Tables 7-9. As a side note, we also examined many other dimensions of heterogeneity, including the severity of the COVID-19 outbreak, alternative measures of health care resources, other measures of economic structure, etc. However, we do not observe strong heterogeneities along these dimensions and thus do not report them in the paper.
Finally, we investigate which age group(s) are driving the overall reduction in mortality. We expect older people and younger people to be sensitive to the overall lockdown policies, while we expect adults to be vulnerable to injuries and accidents. Supplementary Fig 5 summarizes the results. We find that children (−10.6% in row 1) and the elderly (−5.5% in row 5) are indeed more likely than adults (−2.5% in row 2) to be saved by the lockdown policies. If we further examine different causes of death, we find that the elderly is saved both from air pollution-related disease (−6.6% in row 6) and infectious disease (−17.0% in row 8), and younger adults are protected from injuries (−14.7% in row 4). These results are generally consistent with our understanding of the threats of various diseases to different age groups. More detailed results are represented in Supplementary Table 10.
Back-of-the-envelope calculation
In Fig 5, using the estimates in our analyses, we calculate the averted non-COVID-19 deaths in the whole nation due to the lockdown policies during our study period. In Panel a, we plot the predicted average daily deaths. The red and blue lines respectively represent the predicted deaths with and without lockdown policies. Therefore, the differences between these lines can be regarded as the lockdown effects. We see that these two lines start to diverge as more cities implement lockdown policies, and the difference remains stable throughout mid-March.
Because our dataset includes around a quarter of the Chinese population, we apply our estimates to the entire Chinese population, in Panel b. During our study period, 486 DSPs (80.7%) eventually implemented lockdowns, with an average of 38.5 days. We apply our estimates to all the cities that implemented the lockdown policies and calculate the number of averted deaths during our study period. We find that the lockdown policies brought about considerable health benefits: as many as 32,023 lives may have been saved. If we look at the cause-specific effects, we find that cardiovascular diseases account for 62.9% (20,129) of overall averted deaths. Deaths from injury also declined by 10.2% (3,261), pneumonia by 5.0% (1,607), respiratory by 7.4% (2,373), and cancer by 8.5% (2,726).
Discussion
When COVID-19 spread across the globe, we observed a large variation in the public responses in mitigating its impacts: some countries immediately adopted harsh counter-virus measures while others delayed the launch of the policies. As an example of prompt and stringent responses to the COVID-19 outbreak, we investigate the mortality consequences of community and city lockdowns using data from China (excluding Wuhan). Here, we discuss several important implications of our findings.
First and foremost, our findings demonstrate that the China’s unprecedented lockdowns not only effectively controlled the spread of COVID-19, but also brought about substantial unintended benefits to population health during this period. We find that such policies reduced non-COVID19 deaths by 4.92%, which corresponds to 32,000 averted deaths in the nation during 40 days of lockdown. Given the increasingly heated cost-benefit debates regarding different counter-COVID19 policy choices across the world, our results provide a benchmark to understand the health consequences of the lockdown policies. Besides China, several other countries have managed to take the COVID-19 threat under control after one to two months’ strict social distancing, largely because they dealt with the COVID-19 seriously and decisively. We believe these stringent measures should also be better appreciated by policymakers around the world, particularly in countries where the COVID-19 is out of control.
Second, our research points out the directions to improve population health after the pandemic. In particular, we observe a significant reduction in the number of cardiovascular deaths during the lockdown periods, and the effect is larger in cities with higher levels of initial air pollution. A back-of-the-envelope calculation suggests that the total number of averted premature deaths from cardiovascular diseases in the locked-down DSPs alone has far exceeded the total number of deaths caused by COVID-19 in China. This result suggests that air pollution imposed a significant health risk to the Chinese population and it is critically important for the government to continue to improve the environmental quality even when the lockdown is lifted28-29. Besides, the finding on pneumonia mortality confirms that reducing human contacts and raising awareness of preventive measures (such as wearing masks) not only helps control the spread of COVID-19, but also other infectious diseases. These measures should be more appreciated by both public health practitioners and governments.
Third, our results also serve as corroborating evidence that China’s data on the number of COVID-19 deaths are largely reliable, especially those outside of Wuhan. The logic is the following: if the deaths from COVID-19 were intentionally classified as other causes, such as pneumonia or other unclassified diseases, we might observe an unexplainable hike in those causes of death in the locked-down cities (presumably, there were few cases of COVID-19 in the control group). Our results suggest this is not the case; we find that the lockdown reduces all these causes of death in the locked-down cities (using data outside Wuhan), suggesting that COVID-19 deaths are unlikely to be misreported in a substantial way. For Wuhan, however, we do have suggestive evidence of potential misclassification of COVID-19 deaths, as including Wuhan in the regression reverses the sign for deaths from non-COVID-19 pneumonia.
Finally, while the literature has emphasized that economic downturns are usually associated with increased mortality (particularly in less affluent countries), our analyses show that the negative health effects of income shocks during China’s lockdowns were offset by unintended benefits to population health, at least in the short run. While economic collapse is likely to seriously harm public health in the long run, we believe that countries currently affected by COVID-19 can maintain overall population health for a short time by containing the virus as quickly as possible through strict social distancing/mobility restrictions. Future research is needed to understand the long-term welfare implications of different ways to handle the COVID-19 pandemic.
Materials and methods
Materials
DSP Data
The Disease Surveillance Points (DSPs) system is managed by the Chinese Center for Disease Control and Prevention. The system collects death records from the surveillance locations to understand death and disease patterns in China. The system was established in 1978 and has gradually increased its geographical coverage over the past four decades. In 2013, the system got a major upgrade and expanded its coverage from 161 points to 605 points, making the data representative both at the provincial and national levels. Each surveillance point represents a district (if in urban areas) or a county (if in rural areas). For each surveillance point, deaths that occurred in both hospitals and homes are reported, and the causes of death are determined according to a standard protocol by trained staff located in local hospitals or CDC branches. The DSPs system covers more than 324 million people in China, which accounts for 24.3% of the country’s population (18, 19). The quality control procedures include annual training of standard workflow, random checking of the accuracy of disease classification and duplication, retrospective surveys on underreporting, and logic checks on the completeness and accuracy of disease codes. These quality checks are required to be done at the county, province, and national levels. The causes of death are coded in accordance with the International Classification of Diseases-10th revision (ICD-10). We classified the main underlying causes of deaths into 6 categories: I00-I99 for cardiovascular diseases (CVD), V01-Y89 for injuries, J12-J15, J18.9 and J98.4 for pneumonia (excluding COVID-19), C00-C97 for neoplasms, J30-J98 for chronic respiratory diseases, and other causes (remaining ICD-10 codes for all other causes). We further disaggregate cardiovascular diseases, injuries, and pneumonia deaths into specific diseases/causes. Cardiovascular diseases include stroke (I60-I62, I67, and I69), myocardial infarction (I20-I25), and other cardiovascular diseases. Injuries include traffic accidents (V01-V04, V06, V09, V87, V89, and V99), suicide (X60-X84 and Y87), and other injuries. Pneumonia includes mycoplasma pneumonia (J18.9), viral and bacterial pneumonia (J12-J15), and pulmonary infection (J98.4). We also divide the daily number of deaths into three age groups (0-15, 15-64, and ≥65). All death data are analyzed at the aggregated level. In our baseline analysis, we exclude three points (districts) in Wuhan due to concerns that the data might be unrepresentative because the pandemic started there.
In addition to the DSPs system, the Chinese CDC also manages the Communicable Disease Surveillance (CDS) system. The DSPs system is used for death registration, while the CDS system is used to monitor the development of infectious diseases, including COVID-19. These two systems are managed by different subsidiaries in the Chinese CDC and have separate administrative structures. Because the purpose of the DSPs system is to understand the overall cause-of-death patterns in China, and because the death registration process carries no political stake, there is no incentive for local hospitals or CDC branches to hide regular death information from the central government.
Lockdown Data
We collected local governments’ lockdown information city by city from news media and government announcements. Most of the cities’ lockdown policies were directly issued by the city-level governments, while a few were promulgated by the provincial governments. There are two types of lockdowns: city lockdown and community lockdown. The former is defined as human mobility being restricted across different cities, and the latter is defined as mobility being restricted within a city. At the early stage of the outbreak, to prevent the virus from spreading outside Hubei province, city lockdowns were adopted in Wuhan and its neighboring cities. The purpose of city lockdowns was to restrict people in the epicenter of coronavirus from traveling to other cities. Later, as more cases were identified in other cities, community lockdowns were implemented to further control the spread of the coronavirus within cities. The time lag between city lockdowns and community lockdowns was typically one to two weeks. The evolution of different DSPs’ lockdown status is presented in Supplementary Fig 1, 2. In Supplementary Table 11, we further provide a complete list of cities that adopted different lockdown policies at different times. The lockdowns gradually spread to different surveillance districts/counties between January 23 and February 20. By the end of February, 486 out of 602 surveillance points had lockdown policies.
Weather Data
Weather variables include daily temperature, atmospheric pressure, relative humidity, wind speed, and precipitation. The data are obtained from the China Meteorology Administration (CMA). We aggregate station-level air pollution data to city-level data using the inverse squared distance (to city centers) as the weights. Stations closer to the population center are given higher weights so that city-level weather data can be representative of people dwelling in the city.
Air Pollution Data
We obtain air pollution data from the Ministry of Ecology and Environment. The original dataset includes hourly air quality readings from over 2,000 monitoring stations covering 338 prefectural cities in China. We follow the same procedure to aggregate station level air pollution data to the city level. As an omnibus measure of the overall air quality, we use PM2.5 concentration in our regressions. Our results are quantitatively unchanged if we use the Air Quality Index or PM10.
Socio-Economic Conditions
We assemble the socio-economic data at the city or county level from the 2018 China City Statistical Yearbook and 2018 China County Statistical Yearbook, including GDP, population, and the number of hospital beds per 1,000 people. We also obtain data on the employment share of the manufacturing and service industries using the 10% sample of the 2015 1% Population Sampling Survey in China.
Summary Statistics
We report the summary statistics of mortality, lockdown status, and other covariates for 602 DSP counties in Supplementary Table 1. In Panel A, we report the summary statistics of the DSPs data. The average daily total number of deaths at the county level is 8.7, with a standard deviation of 0.025. We use DSPs data from January 1 to March 14, 2020, which include 393,133 death records that were reported to the DSPs system by May 15, 2020.The leading cause of death during this period is cardiovascular diseases, which account for 49.7% of all deaths. The second leading cause of death is neoplasms (22.3%), followed by chronic respiratory diseases (8.7%), and injuries (5.5%). In Panel B, we report the summary statistics of several other variables. The PM2.5 concentration during our study period is 50 µg/m3, five times higher than the WHO standard (10 µg/m3 for annual mean, and 25 µg/m3 for a daily mean). The average share of employment in the manufacturing industries was 24.2% as of 2015.
Statistical Analysis
We use a generalized Difference-in-Differences (DiD) model to identify the impact of counter-COVID-19 measures on mortality. First, in our baseline regression, we estimate the relative change in the number of deaths between the treated and control DSPs using the following model: where Dijt denotes the daily number of deaths in DSP i in city j on date t, and lockdownjt is a dummy variable indicating whether a city/community lockdown is in place in city j on date t. The lockdown dummy takes the value one if either city lockdown or community lockdown was implemented, and zero otherwise. Thus, the coefficient β measures the average effect of three types of lockdown policies: mobility restrictions across cities (city lockdown), mobility restrictions within a city (community lockdown), and both restrictions (city lockdown + community lockdown). To understand how the city and community lockdowns affect health outcomes differently, we separately estimate these effects (Supplementary Table 5). λi are DSP-fixed effects and πt indicate date fixed effects. ℰijt is the error term.
The county fixed effects, λi, which are a set of DSP-specific dummy variables, can control for time-invariant confounders specific to each DSP. For example, the DSP’s geographical conditions, short-term industrial and economic structure, income, and natural endowment can be controlled by introducing the DSP fixed effects. The date fixed effects, πt, are a set of dummy variables that account for shocks that are common to all DSPs in a given day, such as the nationwide holiday policies, macroeconomic conditions, and the national time trend for mortality. As both location and time fixed effects are included in the regression, the coefficient β estimates the difference in the number of deaths between the treated (locked down) and the control cities before and after the enforcement of the lockdown policy. We also add a set of control variables in the regressions to check the robustness of the results (Supplementary Fig. 3).
The underlying assumption for the DiD estimator is that lockdown and control cities would have parallel trends in the number of deaths in the absence of the event. Even if the results show that mortality declines in the treatment counties after the lockdown, the results may not be driven by the lockdown policy, but by systematic differences in treatment and control cities. This assumption is untestable because we cannot observe the counterfactual: what would happen to the mortality levels in the locked-down counties if such policies were not enforced. Nevertheless, we can still examine the trends in mortality for both groups before the lockdown and investigate whether the two groups are indeed comparable. To do so, we conduct the event study and fit the following equation: where lockdownjt,k are a set of dummy variables indicating the treatment status at different periods. Here, we put 7 days (one week) into one bin (bin m ∈ M), so that the trend test is not affected by the high volatility of the daily number of deaths.
The dummy for m = −1 is omitted in Equation (A2) so that the post-lockdown effects are relative to the period one week before the launch of the policy. The parameter of interest βk estimates the effect of lockdown m weeks after the implementation. We include leads of the treatment dummy in the equation, testing whether the treatment affects the air pollution levels before the launch of the policy. Intuitively, the coefficient βk measures the difference in the number of deaths between cities under lockdown and otherwise in period k relative to the difference two weeks before the lockdown. If lockdown reduces mortality, βk would be negative when k ≥ −1. If the pre-treatment trends are parallel, βk would be close to zero when k ≤ −1.
We feel confident in using the estimates from our main results to calculate the averted deaths in the entire country, because our dataset includes around one-quarter of the Chinese population and are representative. To do so, we predict the number of deaths in two scenarios: with/without lockdown policies. Taking the difference between these two predicted deaths, we can calculate the number of saved lives from the lockdown policies. To do so, we first predict the number of deaths with lockdown policies in each DSP county/district in each day by fitting the following model: where denotes the predicted deaths with lockdown policies in each DSP county/district i in city j. , , , and are the fitted values from Equation (A1). In this function, predicted deaths in each DSP, denoted by , can be affected by the lockdown status (represented by lockdownjt).
We then predict the counterfactual, i.e., the number of deaths that would have occurred without lockdowns in any DSP, by fitting the following equation: where denotes the predicted averted deaths without any lockdown policies. lockdownjt (0) always takes a value of zero so that this function is not affected by the policies. Taking the differences between and , we can calculate how many non-COVID-19 deaths are saved from the lockdown policies in each DSP in each day.
Because lockdowns were implemented for 38.5 days on average, we estimate the following model to obtain the averted deaths in the whole country during our study period: where denotes the averted deaths in the entire county during our study period, ChpopLD denotes the total Chinese population in locked-down cities (around 1,161 million), and DSPpopLD represents the total population in locked-down DSPs counties/districts in our dataset (around 291 million in 486 DSPs). The difference between the scenarios with and without lockdowns, denoted by , is totaled from January 1 to March 14, which is our study period (i ϵ I). Note that, in our main text, we repeat these steps to estimate the averted deaths from each cause and disease to understand how many averted deaths can be attributed to different diseases/causes.
Data Availability
The DSPs data are proprietary data owned by the Chinese CDC. They can be accessed through application to the National Center for Chronic and Noncommunicable Disease Control and Prevention (a subsidiary of the Chinese CDC). The codes necessary to re-produce all the tables in the paper are ready to submit to the journal as supplementary materials or to post on a public repository.
Funding
The project is funded by the National Key R&D Program of China (2018YFC1315301), Peking University Research Grant (7100602966) and HKUST School-Based Initiative (SBI17HS02).
Supplemental Materials
Supplementary Notes
Supplementary Note 1: Tests for Pre-Treatment Prallel Trends
The underlying assumption for the DiD estimates in Fig. 1 is that lockdown and non-lockdown DSPs have parallel trends in mortality without the lockdown policies. To test how likely this assumption is to hold, we conduct an event study and investigate whether the two groups of cities have parallel pre-treatment mortality (Materials and Methods: Equation A2). Fig. 2 plots our findings. Here we focus on the four outcomes (total deaths, and deaths from cardiovascular diseases, injuries, and pneumonia) that are statistically significant in Fig. 1.
In Panel a, we compare the total number of non-COVID-19 deaths between the treatment and control groups before and after lockdowns. The difference between the two groups one week before the lockdown is set as the reference group (i.e., the zero coefficient for week −1), so the post-lockdown effects are relative to the period one-week before the launch of the policy. We do not observe systematic difference in the trends of mortality between the two groups one week before the city/community lockdown, i.e., the estimated coefficients for the lead terms (k ≤ −2) are positive or close to zero and statistically insignificant. This finding implies that the parallel trend assumption is likely to hold in our setting. In comparison, the trends break after the lockdown policies were enforced, i.e., the lagged terms (k ≥ 1) become negative and statistically significant. In addition, we observe that the difference becomes larger as more lags are included, suggesting an accumulating health benefit of city/community lockdowns.
In Panels b, c, and d of Fig. 2, we repeat this exercise to investigate the trends in deaths from cardiovascular diseases, injuries, and non-COVID-19 pneumonia before and after lockdowns. The results suggest the parallel trend assumption holds for all these outcomes as well. The corresponding regression results are reported in Supplementary Table 2.
Supplementary Note 2: Robustness Checks and Placebo Tests
We conduct a variety of robustness checks and show that our results are not qualitatively affected by several decisions we make in the baseline analysis (Supplementary. Fig 5 and Supplementary Table 6). First, we add weather variables into the baseline regressions, including daily average temperature, humidity, wind speed, and air pressure (R1), and find that the results are quantitatively unchanged. To further control for the differences in time trends between the treatment and control groups, we also include interactions of time-invariant variables (i.e., per capita GDP, number of hospital beds per thousand people, and total population) with a third-order polynomial function of time in the regressions (R2). The estimates remain similar. These results suggest that the lockdown policies are uncorrelated with these factors and lend additional credibility to our baseline findings.
Second, we weight regressions by population in each DSP (R3). Intuitively, this allows us to estimate the lockdown effects on an average individual, instead of an average DSP. Without weighting, cities with smaller populations could drive the baseline results. We find that the results remain very similar, suggesting that this is unlikely to be the case.
Third, we exclude 22 DSPs in Hubei province from the regression (R4). COVID-19 was first identified in Wuhan city in Hubei province. Thus, the DSPs in the province could be very different. We find that the results are quantitatively similar to the baseline. Next, we include the three DSPs (districts) located in Wuhan in the regressions and find the effects generally become weaker (R5). In addition, the sign for non-COVID-19 pneumonia is reversed and becomes positive (statistically insignificant) (Panel d). This finding suggests that some COVID-19 deaths could be misclassified as deaths caused by other types of pneumonia in Wuhan. Wuhan was the epicenter of COVID-19 in China and contributed to nearly half of the country’s COVID-19 cases. During the first few weeks after the outbreak, the city faced severe medical resource shortages, and many patients could not get immediate diagnoses and treatments. As a result, it would not be surprising to see that some people dying from COVID-19 in the city had been misclassified as dying from other types of pneumonia. In fact, a retrospective survey was recently conducted by the Chinese government, aiming to better classify causes of death in Wuhan. We thus exclude Wuhan from our baseline analysis.
Next, we separately estimate the effects of two types of lockdowns, i.e., city lockdowns and community lockdowns, on the number of deaths (Panel f in Supplementary Table 6). Shortly after the COVID-19 outbreak, a dozen cities around Wuhan (covering 65 DSPs) launched city lockdowns (mostly in late January). Later, as the virus started to spread outside Hubei province, more cities began to enforce community lockdown policies (including those that initially implemented city lockdowns). We include two policy dummies, i.e., city lockdowns and “city+community” lockdowns, into the regressions. The results show that “city+community” lockdowns play a more important role in reducing the number of total deaths, as well as deaths from cardiovascular diseases and non-COVID-19 pneumonia. In other words, restricting human mobility within cities seems critical to explaining our results. City lockdowns have an immediate effect on deaths from injuries, likely because people were not allowed to travel to other cities after city lockdowns.
Finally, one potential threat to our empirical results is the large-scale travel across different cities during the study period. The COVID-19 outbreak coincided with China’s Spring Festival, during which many people leave the cities where they work and travel to their hometowns. If the death patterns are somehow correlated with this travel pattern, our results may be confounded. To address this issue, we conduct a placebo test using data from 2019 (Panel g in Supplementary Table 6). We assign the lockdown status to the same DSP in 2019 and compare changes in the number of deaths between the treatment and control DSPs before and after the “placebo” lockdowns. We find the “placebo” lockdowns do not have any impact on mortality in 2019, suggesting that our findings are not confounded by different mortality trends between the treatment and control groups related to the spring holiday travel.
Supplementary Figures
Supplementary Tables
Acknowledgments
We thank all the staff who work in the primary health facilities, hospitals, and Center for Disease Control and Prevention for death reporting at county/district, city, province, and national levels. We also thank Yun Qiu for sharing the community lockdown data with us. Wei Wang and Yaxuan Liu provided for research assistance.