key: cord-0889771-huo96w6q
authors: Shi, Peng; Dong, Yinqiao; Yan, Huanchang; Zhao, Chenkai; Li, Xiaoyang; Liu, Wei; He, Miao; Tang, Shixing; Xi, Shuhua
title: Impact of temperature on the dynamics of the COVID-19 outbreak in China
date: 2020-04-23
journal: Sci Total Environ
DOI: 10.1016/j.scitotenv.2020.138890
sha: 2b517920b4c978971337d286fec37b5454ad70aa
doc_id: 889771
cord_uid: huo96w6q

Abstract A COVID-19 outbreak emerged in Wuhan, China at the end of 2019 and developed into a global pandemic during March 2020. The effects of temperature on the dynamics of the COVID-19 epidemic in China are unknown. Data on COVID-19 daily confirmed cases and daily mean temperatures were collected from 31 provincial-level regions in mainland China between Jan. 20 and Feb. 29, 2020. Locally weighted regression and smoothing scatterplot (LOESS), distributed lag nonlinear models (DLNMs), and random-effects meta-analysis were used to examine the relationship between daily confirmed cases rate of COVID-19 and temperature conditions. The daily number of new cases peaked on Feb. 12, and then decreased. The daily confirmed cases rate of COVID-19 had a biphasic relationship with temperature (with a peak at 10 °C), and the daily incidence of COVID-19 decreased at values below and above these values. The overall epidemic intensity of COVID-19 reduced slightly following days with higher temperatures with a relative risk (RR) was 0.96 (95% CI: 0.93, 0.99). A random-effect meta-analysis including 28 provinces in mainland China, we confirmed the statistically significant association between temperature and RR during the study period (Coefficient = −0.0100, 95% CI: −0.0125, −0.0074). The DLNMs in Hubei Province (outside of Wuhan) and Wuhan showed similar patterns of temperature. Additionally, a modified susceptible-exposed-infectious-recovered (M-SEIR) model, with adjustment for climatic factors, was used to provide a complete characterization of the impact of climate on the dynamics of the COVID-19 epidemic.

During Dec 2019, an outbreak of a novel coronavirus pneumonia occurred in Wuhan, Hubei Province, China. On Jan 30, 2020, the World Health Organization (WHO) declared an international public health emergency due to infections by this virus. On Feb 20, the WHO officially named this condition coronavirus disease as coronavirus disease 2019 and the causative virus as SARS-CoV-2 (Wu and McGoogan, 2020; Zu et al., 2020) .

Initial studies of disease severity in early cases showed that COVID-19 had a 2.3% case-fatality rate (She et al., 2020) , much lower than in other diseases caused by other coronaviruses, such as Middle East Respiratory Syndrome (MERS, 34.4%) and

Severe Acute Respiratory Syndrome (SARS, 9.2%) (Ceccarelli et al., 2020; Wu and McGoogan, 2020) . However, Wu et al. reported that the number of COVID-19 cases doubled every 6.4 days from Dec 2019 to Jan 2020, indicating COVID-19 was much more infectious than SARS and MERS . In March 2020, the WHO declared that COVID-19 was a global pandemic. At that time, SARS-CoV-2 had spread rapidly throughout China and was present in 116 other countries and territories worldwide.

Environmental factors can affect the epidemiological dynamics of many infectious diseases. In particular, several studies of climate and weather conditions J o u r n a l P r e -p r o o f found that these environmental factors affected the spatial distribution and timing of infections (Bedford et al., 2015; Sooryanarain and Elankumaran, 2015; Lemaitre et al., 2019) . Based on analysis on climatic variables, there is evidence that temperature affect influenza epidemics in tropical regions (Tamerius et al., 2013) . Temperate regions of the Northern and Southern Hemispheres experience highly synchronized annual influenza epidemics during their winter months (Tamerius et al., 2013; Bedford et al., 2015; Sooryanarain and Elankumaran, 2015) . The seasonality of influenza in temperate monsoon climate regions may result from the meteorological factors that affect the environmental and physical stability of virus particles and human social behaviors, both of which contribute to virus epidemiological dynamics.

SARS-CoV-2 can be transmitted through aerosols, large droplets, or direct contact with secretions or fomites, similar to the influenza virus (Li et al., 2005) . However, the effects of different environmental factors on the incidence of COVID-19 remain to be elucidated. Based on dynamical equations, previous researchers developed susceptible-exposed-infectious-recovered (SEIR) modeling to estimate key epidemic parameters to better characterize the mechanisms underlying the dynamics of epidemics (Chanprasopchai et al., 2017; Liu et al., 2017; Niakan et al., 2019) .

We examined the association of the daily confirmed cases rate of COVID-19 with temperature using locally weighted regression and smoothing scatterplot (LOESS) and distributed lag nonlinear models (DLNMs), based on weather and epidemiological data from 31 provincial-level regions in mainland China between Jan J o u r n a l P r e -p r o o f 20 and Feb 29, 2020. We also considered environmental factors using a SEIR model, and developed a modified (M-SEIR) model to characterize the effect of climate on the dynamics of the COVID-19 epidemic in China.

Data on COVID-19, including the number of newly confirmed and probable cases, were retrieved from the China National Health Commission (CNHC, http://www.nhc.gov.cn/) and the CoV2019 package . COVID-19 data were collected from all 31 provincial-level regions in mainland China between Jan 20 and Feb 29, 2020. Data from Hong Kong, Macao, and Taiwan were not included in the study because these areas had major differences in the methods used for data collection. COVID-19 emerged in Wuhan at the end of 2019, and rapidly spread across mainland China. Thus, population dynamic factors, including birth rate and death rate, were not considered. Finally, the daily confirmed cases rate of in each of the 31 provincial-level regions and Wuhan were calculated by dividing the number of newly confirmed cases by the population size as of the end of 2018. These results were reported as cases per 100,000 people.

The daily mean temperature of 344 cities of during the same time period were collected from the meteorological authority in mainland China (http://data.cma.cn).

These city-wide data were consolidated into 31 provincial regions, and were calculated as means. Data on climate conditions and population were from official J o u r n a l P r e -p r o o f reports previously released in mainland China. Therefore, ethical review was not required.

Changes of temperature and daily COVID-19 incidence, including the rate and the common logarithm of newly confirmed cases (Log[N]), were analyzed using a LOESS in the 31 provincial-level regions of mainland China from Jan 20 to Feb 29, 2020.

DLNMs, based on a quasi-Poisson distribution generalized additive model (GAM), were used to infer the exposure-lag-response associations between daily mean temperatures and daily confirmed cases of COVID-19 (Gasparrini et al., 2010; Gasparrini, 2011) . Artificial distortion occurred in Hubei Province on February 12, 2020 and in other provinces on February 20, 2020. To deal with this artificial distortion, we used a 5 day moving average of confirmed COVID-19 cases number to replace case number on the day. Separate DLNMs were constructed for mainland China outside of Hubei Province, Hubei Province outside of Wuhan, and Wuhan.

Additionally, mean temperature of sites in Hubei Province outside of Wuhan were calculated as a representative of Hubei Province overall. To assess the exposure-lag-response relationship, a cross-basis function was used for temperature.

The resulting model is:

where t is the day of observation; E(y t ) is the expected value of the observed number of COVID-19 cases on day t; α is the intercept, cb is cross-basis matrix used to J o u r n a l P r e -p r o o f estimate the non-linear relationship between temperature and COVID-19 incidence and also describe lag effects of temperature. In the cross basis, T is the daily mean tempreature with 2 degree of freedom (df) and the lag is up to 5 days. time is the indicator variable constructed using natural spline with 1df to control long-term trends.

The df for each variable was determined by the quasi-Akaike Information Criterion (qAIC). Among the confirmed cases reported in Qinghai Province and Tibet Province, imported cases accounted for the majority, but those cases were not related to the temperature of each province. Therefore, after completing the modeling of each province, we used a random-effects meta-analysis to summarize the relationship between temperature and exposure-lag-response associations for 28 provinces in mainland China, with exclusion of Hubei Province, Qinghai Province, and Tibet Province.

To better understand the impact of temperature on the COVID-19 epidemic, temperature was considered based on an SEIR model, and an M-SEIR model was used to simulate the COVID-19 outbreak dynamics in Wuhan after implementation of travel restrictions. Sensitivity analysis was performed for quantitative risk assessment to evaluate the relationships between temperature and COVID-19 incidence.

The equations of M-SEIR model were:

where S(t), E(t), I(t), and R(t) are the number of susceptible, exposed, infectious, and removed individuals at time t,  1 and  1 are the mean latent and infectious periods, β t is a time-dependent rate of infectious contact, and β 1 and β 2 are coefficients.

The simulations of the dynamics of the COVID-19 epidemic and sensitivity analysis were conducted using the system dynamics section in AnyLogic software (version 8.5.2). Supplementary Table 1 provides the specific parameter values used in the modified model and basic model and further details.

There were 80,981 confirmed cases of COVID-19 in the 31 provincial-level regions of mainland China between Jan 20 and Feb 29, 2020. Due to the change in the diagnostic criteria used in Hubei Province, some patients with confirmed clinical diagnoses were considered healthy, so these data were removed and not considered in this study. A total of 68,034 of these cases (84.01%) were diagnosed in Hubei

Province. Analysis of newly confirmed cases and daily confirmed cases rate in mainland China (Supplementary From Jan 20 to Feb 29, 2020, the temperature varied in the 31 provincial-level J o u r n a l P r e -p r o o f regions in mainland China (Figure 1 A) . The highest temperature (26 ℃) was in Hainan Province (south-eastern China), and the lowest temperature (-22 ℃) was in Jilin Province (north-eastern China). There was a biphasic relationship of daily confirmed cases rate with temperature (with a peak at about 10 ℃) (Figure 1 B and C) . This is likely because some cases were clinically diagnosed without nucleic acid testing in Hubei Province prior to Feb 12.

The association between average cumulative relative risk over lags 0-5 and temperature across 28 provinces in China was shown in Figure 2A and

The overall epidemic intensity of COVID-19 reduced slightly following days with higher temperatures associated with the relative risk (RR) was 0.96 (95% CI: 0.93, 0.99). A random-effect meta-analysis including 28 provinces in mainland China, we found the statistically significant association between temperature and RR during the study period (Coefficient = -0.0100, 95% CI: -0.0125, -0.0074) by the use of meta-regression ( Figure 2B ). The result of Figure 2B confirms the negative relationship between temperature and RR among the 28 provinces in China once again. In order to further explore the impact of daily mean temperature on the confirmed cases of COVID-19 with different lag, 0-5 days lag were explored. We find that with the following lag days, RR gradually increases, reaching a peak when lagging by 2 days, and then decreases with the number of lag days increases ( Figure 2C ).

We analyzed the association between temperature and RR of COVID-19 in Hubei Province (outside of Wuhan), and Wuhan ( Considering the impact of temperature, we constructed an M-SEIR model to simulate the dynamics of the COVID-19 epidemic using the system dynamic sections in AnyLogic software. The SEIR dynamic transmission model compartmentalized the population into four disease states (susceptible, exposed, infected, and recovered) and

analyzed the relationships and interconnections using stock and set parameters, flows, and table functions ( Figure 4A and Supplemental Video 1). We set the initial values and functional relationships. The results indicated that the transmission rate, which is updated by the real-time temperature data in AnyLogic software, decreased as temperature increased, so that the infection rate and size of the outbreak decreased.

In the study, we found temperature was an environmental driver of the COVID-19 outbreak in China. Our LOESS showed that the daily incidence was lowest at -10 ℃ and highest at 10 ℃. Our DLNMs indicated that temperature was significantly associated with the daily incidence of COVID-19 with and without time lags. Our M-SEIR model for Wuhan predicted the COVID-19 outbreak would peak J o u r n a l P r e -p r o o f on about March 5, 2020 and would end in late April. Additionally, we found that the transmission rate decreased as temperature increased, and that the increasing temperature contributed to further decreases of the infection rate and size of the outbreak. Therefore, we found that temperature drove the spatial and temporal correlations of the COVID-19 outbreak in China, and should be considered the optimal climatic predictor for the incidence of COVID-19.

Our results indicated a significant association between temperature and 

Consequently, we constructed an M-SEIR model to correct for temperature changes in our simulation of the dynamics of the COVID-19 epidemic in China. Our M-SEIR model predicted that the outbreak would reach its peak on about March 5, 2020, consistent with the actual data released by the CNHC (Wan et al., 2020; Tang et al., 2020; Shi et al., 2020; Pan et al., 2020; Hong et al., 2020) . Our results also predicted that the COVID-19 outbreak in Wuhan would end in late April. In addition, we conducted a sensitivity analysis on the temperature-adjusted transmission rate. These analyses indicate that the transmission rate decreased as temperature increased, so that the infection rate and size of the epidemic declined over time.

Our analysis is subject to some limitations. First, multiple factors, including virus properties, additional climatic factors, socio-economic development, population mobility, population immunity, and urbanization, presumably affected the dynamics of the COVID-19 epidemic in China, but we cannot consider every factor in this study.

Second, we optimized the parameters of M-SEIR model based on previous analysis, and this might have led to bias due to the lack of official data and the adjustment of diagnostic criteria during the outbreak. Third, our study was an ecological analysis that examined a very short period of time, so our conclusions regarding climatic factors as being causative in virus transmission are limited. In particular, we cannot avoid the possible bias caused by other ecological factors that also changed over time.

Fourth, the temperature on the reporting date was not the actual temperature experienced by each case at infection. However, it is not feasible to determine the temperature on actual date when each patient was first infected. Fifth, the current J o u r n a l P r e -p r o o f research on GAM-DLNM analysis was developed in the context of non-infections disease death of incident counts. There may be limitations in applying this method to short-term infectious diseases. But the research of Lowe has made exploration on the infections disease at present (Lowe et al., 2018) . Finally, all confirmed cases for each province in China include "imported" cases and "local" cases. An imported case (i.e., from another province, especially Hubei Province) should ideally be counted for the province of the patient's origin. However, this information was not available.

Temperature was an environmental driver of the COVID-19 outbreak in China. Meta-analysis of the exposure-lag-response association. CI, confidence interval; RR, the relative risk; Weight, the percentage of cases in each province amount to the total cases among 28 provinces; For temperature, lag is distributed over lags 0-5 as described in the text. for different temperature scenarios in Wuhan. As the temperature-corrected transmission index increased, the maximum of the incidence rate curve increased and

Digital epidemiology reveals global childhood disease seasonality and the effects of immunization

Global circulation patterns of seasonal influenza viruses vary with antigenic drift

Differences and similarities between Severe Acute Respiratory Syndrome (SARS)-CoronaVirus (CoV) and SARS-CoV-2. Would a rose by another name smell as sweet?

Effect of Rainfall for the Dynamical Transmission Model of the Dengue Disease in Thailand

Distributed Lag Linear and Non-Linear Models in R: The Package dlnm

Distributed lag non-linear models

Plug-and-play inference for disease dynamics: measles in large and small populations as a case study

An initial investigation of the association between the SARS outbreak and weather: with the view of the environmental temperature and its variation

Prediction of New Coronavirus Infection Based on a Modified SEIR Model

When will the battle against novel coronavirus end in Wuhan: A SEIR modeling analysis

Nowcasting and forecasting the potential domestic and international spread of the 2019-nCoV outbreak originating in Wuhan, China: a modelling study

Open-source analytics tools for studying the COVID-19 coronavirus outbreak