Air Quality Response in China Linked to the 2019 Novel Coronavirus (COVID‐19) Lockdown

Abstract Efforts to stem the spread of COVID‐19 in China hinged on severe restrictions to human movement starting 23 January 2020 in Wuhan and subsequently to other provinces. Here, we quantify the ancillary impacts on air pollution and human health using inverse emissions estimates based on multiple satellite observations. We find that Chinese NOx emissions were reduced by 36% from early January to mid‐February, with more than 80% of reductions occurring after their respective lockdown in most provinces. The reduced precursor emissions increased surface ozone by up to 16 ppb over northern China but decreased PM2.5 by up to 23 μg m−3 nationwide. Changes in human exposure are associated with about 2,100 more ozone‐related and at least 60,000 fewer PM2.5‐related morbidity incidences, primarily from asthma cases, thereby augmenting efforts to reduce hospital admissions and alleviate negative impacts from potential delayed treatments.


Introduction
This supporting information provides descriptions of the data and methods used in this study.
Text S1. Chemical data assimilation An extended calculation of the Tropospheric Chemistry Reanalysis version 2 (TCR-2) (Miyazaki et al., 2020a) is used to evaluate emission and concentration changes. The reanalysis products used in this study have been obtained from the assimilation of OMI and TROPOMI NO 2 , MOPITT CO, OMI SO 2 , and MLS ozone and HNO 3 . The reanalysis calculation was conducted at 1.1 • (for OMI assimilation) and 0.56 • (for TROPOMI assimilation) horizontal resolution using a global chemical transport model MIROC-CHASER and an ensemble Kalman filter technique that optimizes both chemical concentrations of various species and emissions of several precursors. This approach was efficient for the correction of the entire tropospheric profile of various species and its year-to-year X -4 : mass factor (tropospheric air mass factor/geometric air mass factor> 0.2). We employed a super-observation approach (Miyazaki et al., 2012) to generate representative data with a horizontal resolution of the forecast model (0.56 • for TROPOMI and 1.1 • for OMI). The OMI SO 2 data used were the planetary boundary layer vertical column SO 2 L2 product obtained with the principal component analysis algorithm (Krotkov et al., 2016). Only clear-sky OMI SO 2 data (cloud radiance fraction< 20%) with solar zenith angles less than 70 • were used, and the first and last ten cross-track positions were excluded to limit the across-track pixels. The MOPITT total column CO data used were the version 7 L2 TIR/NIR product (Deeter et al., 2017). The version 4.2 ozone and HNO 3 L2 products from MLS (Livesey et al., 2018) were used to constrain the chemical concentrations in the upper troposphere and lower troposphere, which were important in correcting the influences of stratospheric intrusions on tropospheric ozone.
Text S4. Chemical transport model, MIROC-CHASER X -6 : ∆ Health impacts= y 0 × (1 − exp −β∆Exposure ) × P opulation where health impacts include respiratory hospital admissions (HAs) and asthma-related emergency room visits (ERVs) for short-term ozone exposure, and children asthma symptom days, children bronchitis, respiratory HAs, and cardiovascular HAs for short-term PM2.5 exposure. y 0 is the baseline rates of each health impact mentioned above in China; β is the exposure-response function for unit change of exposure; population is the number of people in the susceptible age-group living in the same 0.5 • ×0.5 • grid box with exposure estimates. For asthma ERV, y 0 is the baseline asthma-related ERVs, which is calculated as the product of baseline asthma prevalence and the fraction of ERVs among asthma patients (Anenberg et al., 2018).
Exposure-response functions are derived from existing studies. We first examined studies based in China. We found several cohort and time-series studies (Lu et al., 2020, : X -7 eral" estimates based on global scale multi-national results. The "China-specific" case may more accurately reflect the health effects in China, while the "general" case covers more health endpoints and is also better established as it relies upon meta-analyses of multiple epidemiological studies. We developed the "China-specific" estimates based on China-based epidemiological studies. In addition to the aforementioned individual studies, we found a systematic review focused on effects of PM2.5 related to hospital utilization in China (Lu et al. 2015).
It concluded that the China-specific effects can be close to the lower-end of estimates for European countries and the United States, even though overall magnitudes of effects are comparable to the global estimates. As such, we used a relative risk (RR) (2018)), because our goal is to assess the total burdens. For asthma, we calculated both HAs and asthmatic symptom days to address the total burdens that include different levels of severity. However, we thus did not include outpatient visits for asthma (as presented in Lu et al. (2020)), because of the likely overlap with HAs and asthmatic symptom days.
In addition, we develop our "general" estimates based on established global exposureresponse functions. Priority is given to studies that are either systematic reviews of multiple studies, or large-scale multi-country studies. For respiratory HAs, we used RR  (Table 1, S5, and S6) using the provided RR and baseline rate ranges.
Because the China-specific information is still very limited for our exposure estimates as discussed above, we present the "general" estimates based on established global exposureresponse functions in the main text and SI (Table 1 and S5), while also discussing the August 20, 2020, 8:47pm : X -9 "China-specific" case results for the PM2.5 cardiovascular HAs in the SI (Table S5). The comparison between the "China-specific" and "general" results demonstrates that the China-specific values are close to the lower-end of the global estimates for the PM2.5 cardiovascular HAs (Table S5). Impacts of using China-specific values for other endpoints need to be further explored as comparisons between China-specific values and those from studies elsewhere do not show a systematic difference (Burnett et al., 2018).
Before the health impact assessments, a time-constant bias correction was applied to MDA8 based on the validation against the in-situ observations (Table S4), while the observed temporal changes were already broadly reproduced by the model for MDA8. As summarized in Table S6, the time-constant concentration bias correction had only slight impacts on the respiratory HA changes. We also evaluated the impact of uncertainty in the estimated concentration bias by applying random numbers (n=1000) to the standard deviation of model bias within each province using daily concentrations based on the Monte Carlo approach. The uncertainty in the exposure-response functions had larger impacts than those in the model concentration bias uncertainty. We also tested a timeconstant bias correction for PM2.5, and obtained larger impacts in the exposure-response function uncertainty for both ozone and PM2.5-related HA estimates. As discussed in the main text, the underestimated temporal changes in the model PM2.5 concentration is likely due to the lack of observational constraints on direct aerosol emissions. Because we attempted to demonstrate the impacts of NOx and SO2 emission changes on human health through the secondary formation processes, a temporally-varying bias correction X -10 : was not tested, which would provide larger impacts on the estimated exposure changes than the time-constant bias correction.
As noted in the main text, hospital visits during the lockdown were often limited by regulations and avoided by personal preference, hence the results should be seen an indication of the extent of health impacts due to changing exposures to ozone and PM2.5, instead of the actual numbers of visits.
Text S7. Ozone production efficiency estimates using multi-model multiconstituent chemical data assimilation Data assimilation that relies on a single model may lead to biased estimation of emissions and model response. We used the multi-model multi-constituent chemical data assimilation (MOMO-Chem) framework (Miyazaki et al., 2020b) to estimate response of surface ozone concentration to NOx emissions and its uncertainty. This system integrates a portfolio of data assimilation analyses obtained using four forward CTMs (GEOS-Chem, AGCM-CHASER, MIROC-Chem, MIROC-Chem-H) in a state-of-the-art ensemble Kalman filter data assimilation system. The framework was used to demonstrate the importance of the performance of forecast models for tropospheric chemistry data assimilation and to provide multi-model integrated information on the tropospheric chemistry system. By applying linear regressions to the ozone increments with respect to the NOx emission analysis increments using the daily mean data assimilation outputs at each grid Text S8. Model and observation comparisons Data assimilation improved agreements with the assimilated satellite NO 2 retrievals ( Fig S5), with large reductions in model positive biases over polluted areas. In the OMI (TROPOMI) assimilation, the regional mean bias compared to the assimilated measurements are reduced by 91-94% (85-99%) by data assimilation, while root-mean-squareerror (RMSE) are reduced by 68-93% (78-90%). The agreements against the assimilated measurements confirm that the observational constraints were sufficient to reproduce the observed variability through emission optimization. The linear regression slope against the NAQMS surface NO 2 measurements was 0.88-0.95 (0.41-0.50) for OMI (TROPOMI) assimilation over northeastern and southeastern China. The TROPOMI standard product shows negative biases against the OMI product (Ialongo et al., 2020).

Text S9. Hospital admissions (HAs) due to COVID-19
We evaluated HAs due to COVID-19 during the analysis period (February 15-25, 2020) based on COVID-19 cases and the hospitalization fraction of COVID-19 cases. The total reported number of COVID-19 new cases for the analysis period is 11,769 with 10,976 cases in Hubei (Table 1). The proportion of all infections that would lead to hospitalisation is estimated at 1.04 % for 20-20 years, 3.43 % for 30-39 years, 4.25 % for 40-49 years, 8.16 % for 50-59 years, 11.8 % for 60-69 years, 16.6 % for 70-79 years, and 18.4 % over 80 years (Verity et al., 2020). By applying a fraction of 18.4 %, an upper limit of HAs due to COVID-19 for the analysis period is estimated at 2,165 for country-total and 2,019 for Hubei. X -12 : The temporal changes in ozone during the analysis period are overall consistent between the data assimilation analysis and in-situ observation at the province scale (Table S4)  Table S1. The observed concentrations of NO 2 (in µg m −3 ), O 3 (in ppb), and PM2.5 (in µg m −3 ) from the national air quality monitoring stations (NAQMS) stations averaged during two weeks before, during, and two weeks after the CNY holidays in 2019 and 2020.
The relative changes (∆[%]) between before and during and between during and after the CNY holiday are also shown. The estimated random and sampling errors are shown as the measurement uncertainty.     Table S6. Total changes of respiratory HAs for short-term ozone exposure in post-65 population (Ozone HA), respiratory and cardiovascular HAs for short-term PM2.5 exposure (PM2.5 HA), and their uncertainty due to exposure-response functions (E(exposure), ±2-σ) and model concentration bias and its uncertainty (E(conc), bias±2-σ).
The data used in this paper are available for download at https://ebcrpa.jamstec .go.jp/~miyazaki/data GRL2020/