The impact of transmission control measures during the first 50 days of the COVID-19 epidemic in China

Respiratory illness caused by a novel coronavirus (COVID-19) appeared in China during December 2019. Attempting to contain infection, China banned travel to and from Wuhan city on 23 January and implemented a national emergency response. Here we evaluate the spread and control of the epidemic based on a unique synthesis of data including case reports, human movement and public health interventions. The Wuhan shutdown slowed the dispersal of infection to other cities by an estimated 2.91 days (95%CI: 2.54-3.29), delaying epidemic growth elsewhere in China. Other cities that implemented control measures pre-emptively reported 33.3% (11.1-44.4%) fewer cases in the first week of their outbreaks (13.0; 7.1-18.8) compared with cities that started control later (20.6; 14.5-26.8). Among interventions investigated here, the most effective were suspending intra-city public transport, closing entertainment venues and banning public gatherings. The national emergency response delayed the growth and limited the size of the COVID-19 epidemic and, by 19 February (day 50), had averted hundreds of thousands of cases across China.

On 31 December 2019, less than a month before the 2020 Spring Festival holiday, including the 66 Chinese Lunar New Year, a cluster of pneumonia cases caused by an unknown pathogen was 67 reported in Wuhan, a city of 11 million inhabitants and the largest transport hub in Central China. 68 A novel coronavirus (1, 2) was identified as the etiological agent (3, 4) and human-to-human 69 transmission of the viral disease  has been since confirmed (5, 6). Further spatial 70 spread of this disease was of great concern in view of the upcoming Spring Festival ("chunyun") 71 during which there are typically three billion travel movements over the 40-day holiday period, 72 which runs from 15 days before the Spring Festival (Chinese Lunar New Year) to 25 days 73 afterwards (7).

75
As there is currently neither a vaccine nor a specific drug treatment for COVID-19, a range of 76 public health (non-pharmaceutical) interventions has been used to control the outbreak. In an 77 attempt to prevent further dispersal of COVID-19 from its source, all transport was prohibited in 78 and out of Wuhan city from 10:00h on 23 January 2020, followed by the whole of Hubei Province 79 a day later. In terms of the population covered, this appears to be the largest attempted quarantine 80 (movement restriction) event in human history.

82
On 23 January, China also raised its national public health response to the highest state of 83 emergency ─ Level 1 of 4 levels of severity in the Chinese Emergency System, defined as an 84 "extremely serious incident" (8). As part of the national emergency response, and in addition to 85 the Wuhan city travel ban, suspected and confirmed cases have been isolated, public transport by 86 bus and subway rail suspended, schools and entertainment venues have been closed, public 87 gatherings banned, health checks carried out on migrants ("floating population"), travel prohibited 88 in and out of cities, and information widely disseminated. Despite all these measures, the outbreak 89 has continued to spread geographically, within and beyond China, with mounting numbers of 90 cases and deaths.

92
Although the spatial spread of infectious diseases has been intensively studied (9-14), including 93 explicit studies of the role of human movement (15,16), the effectiveness of travel restrictions and 94 social distancing measures in preventing the spread of infection is uncertain. For 95 coronavirus transmission patterns and the impact of interventions are still poorly understood (6, 7). 96 We therefore carried out a quantitative analysis of the impact of travel restrictions and 97 transmission control measures during the first 50 days of the COVID-19 epidemic in China, from 98 31 December 2019 to 19 February 2020 (Fig. 1). This period embraced the 40 days of the Spring 99 Festival holiday, 15 days before the Chinese Lunar New Year on 25 January and 25 days 100 afterwards. The analysis is based on a unique geocoded repository of data on COVID-19 101 epidemiology, human movement, and public health (non-pharmaceutical) interventions.

103
We first investigated the effect of the Wuhan city travel ban, comparing travel in 2020 with that in 104 previous years and exploring the consequences of holiday travel for the dispersal of infection 105 across China. During Spring Festival travel in 2017 and 2018, there was an average outflow of 5.2 106 million people from Wuhan city during the 15 days before the Chinese Lunar New Year. In 2020, 107 this travel was interrupted by the Wuhan city shutdown, but 4.3 million people travelled out of the 108 . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . https: //doi.org/10.1101//doi.org/10. /2020 city between 11 January and the implementation of the ban on 23 January (7) ( Fig. 2A). In 2017 109 and 2018, travel out of the city during the 25 days after the Chinese Lunar New Year averaged 6.7 110 million people each year. In 2020, the travel ban prevented almost all of that movement.

112
The dispersal of COVID-19 from Wuhan was rapid (Fig. 3A) This delay provided extra time to prepare for the arrival of COVID-19 across China but would not 126 have curbed transmission after infection had been exported to new locations from Wuhan. Fig. 1  127 shows the timing and implementation of emergency control measures in 342 cities across China 128 (see also Figs. S2 and S4). School closure, the isolation of suspected and confirmed patients, plus 129 the disclosure of information was implemented in all cities. Public gatherings were banned and 130 entertainment venues closed in 220 cities (64.3%). Intra-city public transport was suspended in 131 136 cities (39.7%) and inter-city travel was prohibited by 219 cities (64.0%). All three measures 132 were applied in 136 cities (Table S2).

134
Cities that implemented a Level 1 response (any combination of control measures; Figs. S2 and S4) 135 pre-emptively, before discovering any COVID-19 cases, reported 33.3% (95%CI: 11.1-44.4%) 136 fewer laboratory-confirmed cases during the first week of an outbreak (13.0,137 n=125) compared with cities that started control later (20.6 cases, n=171;138 difference between groups, U=8197 z=-3.4, P<0.01). Among specific control measures, cities that 139 suspended intra-city public transport and/or closed entertainment venues and banned public 140 gatherings, and did so sooner, reported fewer cases during the first week of their outbreaks (Table  141 2, Table S3). This analysis provided no evidence that the prohibition of travel between cities, 142 which was implemented after and in addition to the Wuhan shutdown on 23 January, reduced the 143 number of cases in other cities across China. These results are robust to the choice of statistical 144 regression model (Supplementary Material, Table S3).

146
The reported daily incidence of confirmed cases peaked in Hubei province (including Wuhan) on 147 4 February (3156 laboratory-confirmed cases, 5.33/100,000 population in Hubei), and in all other 148 provinces on 31 January (875 cases, 0.07/100,000 population; Fig. S1). The low level of peak 149 incidence per capita, the early timing of the peak, and the subsequent decline in daily case reports, 150 suggest that transmission control measures not only delayed the growth of the epidemic, but also 151 greatly limited the number of cases. By fitting an epidemic model to the time series of cases 152 . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . https: //doi.org/10.1101//doi.org/10. /2020 reported in each province (Supplementary Material, Fig. S3), we estimate that the (basic) case 153 reproduction number (R0) was 3.15 prior to the implementation of the emergency response on 23 154 January (Table 3). As control was scaled-up from 23 January onwards (stage 1), the case 155 reproduction number declined to 0.97, 2.01 and 3.05 (estimated as C1R0) in three groups of 156 provinces, depending on the rate of implementation in each group (Tables 3 and S4). Once the 157 implementation of interventions was 95% complete everywhere (stage 2), the case reproduction 158 number had fallen to 0.04 on average (C2R0), far below the replacement rate (<< 1) and consistent 159 with the rapid decline in incidence (Fig. 4A, Fig. S5, Table 3, Table S4).

161
Based on the fit of the model to daily case reports from each province, we investigated the 162 aggregate effect of control measures on the trajectory of the epidemic outside Wuhan city (Fig.  163 4B). Without the Wuhan travel ban or the national emergency response, there would have been 164 744,000 (± 156,000) confirmed COVID-19 cases outside Wuhan by 19 February, day 50 of the 165 epidemic. The Wuhan travel ban alone would have reduced this number to 202,000 (± 10,000) 166 cases, by delaying epidemic growth. The national emergency response alone would have cut the 167 number of cases to 199,000 (± 8500). Therefore neither of these interventions would, on their own, 168 have reversed the rise in incidence by 19 February (Fig. 4B). But together and interactively, these 169 control measures evidently did halt and reverse the rise in incidence, limiting the number of 170 confirmed cases reported to 29,839 (fitted model estimate 28,000 ± 1400 cases), a 96% reduction 171 on the total number of cases expected in the absence of interventions.

173
In summary, this early analysis suggests that transmission control (non-pharmaceutical) measures 174 The number of people who have developed COVID-19 during this epidemic, and therefore the 185 number of people who were protected by control measures, is not known precisely, given that 186 cases were almost certainly under-reported. However, in view of the small fraction of people 187 known to have been infected by 19 February (75,532 cases, 5.41 per 100,000 population), it is 188 unlikely that the spread of infection was halted and epidemic growth reversed because the supply 189 of susceptible people had been exhausted. This implies that a large fraction of the Chinese 190 population remains at risk of COVID-19; relaxing control measures could lead to a resurgence of 191 transmission. Further investigations are needed to verify that proposition, and population surveys 192 of infection are needed to reveal the true number of people who have been exposed to this novel 193 coronavirus.

195
We could not investigate the impact of all elements of the national emergency response because 196 . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi.org/10.1101/2020.01.30.20019844 doi: medRxiv preprint many were introduced simultaneously across China. However, there is firm evidence from the data 197 used in this analysis that suspending intra-city public transport, closing entertainment venues and 198 banning public gatherings, which were introduced at different times in different places, 199 contributed to the overall containment of the epidemic. Other factors are likely to have contributed 200 to control, such as the isolation of suspected and confirmed patients, contact tracing and the 201 closure of schools, and it is not yet clear which parts of the national emergency response were 202 most effective. We did not find that prohibiting travel between cities or provinces reduced the 203 numbers of COVID-19 cases outside Wuhan and Hubei, perhaps because such travel bans were 204 implemented as a response to, rather than in anticipation of, the arrival of COVID-19.

206
This study has drawn inferences, not from a controlled experiment, but from statistical and 207 mathematical analyses of the temporal and spatial variation in case reports, human mobility and 208 transmission control measures. With that caveat, we conclude that these control measures had a 209 major impact on the COVID-19 epidemic, averting hundreds of thousands of cases by 19 Feburary.   author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint  G. F. Gao, From "A" IV to "Z" IKV: attacks from emerging and re-emerging pathogens.   . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint   . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint   (Table S3), the model judged best by the Akaike Information Criterion) was: 463 464 Dependent variable Yj is the arrival time (day) of the first confirmed case in city j, a measure of the 467 spatial spread of COVID-19. The βi are the regression coefficients. α is the intercept. TotalFlowj 468 . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . https: //doi.org/10.1101//doi.org/10. /2020 represents the passenger volume from Wuhan to city j by airplane, train and road during the whole 469 of 2018. Popj is the population of city j. Latj and Lon j represent the latitude and longitude of city j. 470 The binary dummy variable Shutdownj is used to identify whether the arrival time of COVID-19 471 in newly-infected city j is influenced by the Wuhan travel ban. For each city, shutdown was set to 472 0 for arrival before 23 January 2020 and 1 for arrival on or after 23 January 2020. The regression 473 analysis was performed using the R package (R version 3.4.0) MASS (22). All of the candidate 474 models examined (Table S3)  (day 0 of the epidemic). Each city was regarded as implementing an intervention when the official 486 policy was announced publicly (Table S1). Other transmission control measures included 487 delineating control areas, closure of schools, isolation of suspected and confirmed cases, and the 488 disclosure of information. The effects of these interventions could not be investigated because 489 they were reportedly applied in all cities uniformly and without delay.

491
As above, we used regression analysis to investigate the effects of interventions B, S and P. The 492 dependent (Poisson) variable is the total number of confirmed cases that were reported during the 493 first seven days (μ) of an outbreak in any city (i). The analysis was performed using the GLM 494 function in the statistical software R (version 3.6.2) using the model: 495 496 log(i) =  + 1Mi,S + 2Mi,P+ 3Mi,B4Ti,S + 5Ti,P+ 6Ti,B + 7Ai+ 8Di+ log(Qi) + log(Fi)

498
where population size of a city i (Oi) and inflow from Wuhan (Fi) are offset variables, while the 499 distance to Wuhan and the arrival time of the infection are adjustments to control for confounding 500 with other independent variables. The βj's are regression coefficients. Mi,k is a binary variable 501 indicating whether or not control measure k is implemented in city i. Ti,k represents the timing of 502 implementation of control measure k in city i. Di is the distance from city i to Wuhan City. Ai is the 503 arrival time of the epidemic in city i (the date of the first confirmed case).

505
To check and confirm the validity of results obtained with the Poisson regression model, we 506 repeated the analysis with a log-linear model. The first step was to standardize case counts by 507 dividing by the number of people in each city (incidence per capita) and the number of people 508 arriving from Wuhan, giving dependent variable . The log-linear model is then: 509 510 E[log(i)] =  + 1Mi,S + 3Mi,B4Ti,S + 6T.Ci,B + 7Ai.

512
. CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi.org/10. 1101/2020 The subscripts of the coefficients (j) are consistent with a Poisson regression model. To avoid 513 heteroscedasticity, variables describing the distance from Wuhan, and the implementation and 514 timing of P (prohibiting inter-city travel) were removed. Further exploration of the model showed 515 that these variables did not help to explain variation in . Table S3 presents the results of the   516 log-linear regression analysis, which uphold the conclusions reached from the Poisson regression 517 model.

519
Epidemic modelling 520 For each province, we estimated the effect of transmission control measures by fitting an SEIR 521 model (23) to the number of new confirmed cases reported each day from each province using 522 Bayesian Markov Chain Monte Carlo methods (24). The model is: 523 524 where S, E, I, and R are the number of susceptible, exposed (latent), infectious, and removed 530 individuals on day t in province i. This standard SEIR model makes some simplifying assumptions: 531 for example, the human population is homogeneous (e.g. not stratified by age or sex), contacts 532 between infectious and susceptible people are also homogeneous (e.g. not stratified by social 533 group) and infection is fully immunizing (1). However, the model describes the data accurately 534 (Fig. 4A, Fig. S3) and these assumptions are unlikely to affect the principal conclusions of the 535 analysis, which apply only to the first 50 days of the epidemic. The basic reproductive number of 536 the model is R0 =β/γ, where β is the per capita transmission rate per day and 1/δ and 1/γ are, 537 respectively, the mean latent and infectious periods.

539
Variable λ is the estimated number of cases imported from Wuhan City on day t: Iw is the number of reported cases in Wuhan on day t, Pw is the Wuhan population size, and ρw is 544 the proportion of all infected people (including infectious cases) reported in Wuhan. Ti is the 545 number of people leaving Wuhan on day t travelling to province i, derived from data describing 546 . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . https: //doi.org/10.1101//doi.org/10. /2020 mobility 15 days before the Chinese Lunar New Year 2020. The binary variable shutdown is used 547 to identify whether cases were or were not exported from Wuhan on or after 23 January 2020.

549
The effects of control measures at different stages of the outbreak are captured by estimated 550 parameter C (range 0-1), which reduces transmission and R0 proportionally as a multiplicand of β. 551 The timing and implementation of transmission control measures in the 342 cities and 31 552 provinces are shown in Fig. S4. Before 22 January 2020, there were no recorded interventions thus 553 C0=1. From 23 January onwards, provinces gradually scaled up Level 1 emergency responses 554 (stage 1), with effects measured as C1 (Fig. S4). Because the effects of control measures varied 555 among provinces during the scale-up, C1 was grouped into high C1_high, medium C1_medium, and low 556 C1_low. The allocation of provinces to groups was made by proposing several alternative 557 hypotheses and testing each by model fitting (Table 3, Table S4). Stage 2 of control (C2) began 558 when more than 95% of cities in a province had implemented control measures, including the 559 closure of entertainment venues, suspension of intra-city public transport or prohibition of travel 560 by any means to and from other cities (see above). In Hubei Province (except Wuhan city), stage 2 561 included the use of shelter or "Fang Cang" hospitals from early February onwards. Metropolis). Prior estimates of the mean and (Gaussian) variance of R0, δ, and γ were derived from 566 epidemiological surveys (25). There was no evidence to inform a prior for the reporting rate ρ, the 567 proportion of cases that were reported among all latent and infectious individuals in Wuhan. 568 Systematic surveys of infection (e.g. by serological testing) have not yet been reported. In the 569 absence of any guiding data, ρ was given a prior uniform distribution between 0 and 1.

571
After a burn-in of 1 million iterations, we ran the MCMC simulation for a further 10 million 572 iterations, sampled at every 1000th step to avoid auto-correlation. Trace plots and Gelman and 573 Rubin diagnostics were used to judge convergence of the MCMC chains (Fig. S4). Each fitting 574 exercise was repeated three times to test the robustness of results, which converged to the same 575 estimates on each occasion (Fig. S5). We used the fitted SEIR model, with posterior estimates of 576 parameter values, to simulate outbreaks outside Wuhan, with and without the Wuhan travel ban 577 and with and without the national emergency response (Fig. 4B). 578 579 . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity. provinces in relation to distance between provinces. Synchrony is measured by the correlation 589 between the number of cases reported in two provinces on each day, using a spatial 590 non-parametric correlation function (26). 591 592 . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi.org/10. 1101/2020 593 594 595 Figure S2. Percentage of cities that implemented three kinds of transmission control measures 596 before (blue), or on the same day or after (red), the first case was reported. 597 598 599 600 . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint  . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi.org/10. 1101/2020 642 643 Table S2. Summary of interventions and their timing across 342 cities (see Table 2  . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity. is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi.org/10. 1101/2020 652 653 Table S3. Impact of the type and timing of transmission control measures, estimated from a 654 log-linear regression model. This analysis checks and confirms the robustness of results in Table 2  655 of the main text. As described in the main text, the prohibition of inter-city travel, the third 656 intervention that was investigated in this study, did not significantly reduce the number of cases 657 reported during the first week of city outbreaks. 658 659 . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Covariates
is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi.org/10. 1101/2020 663 664 Table S4. Candidate models used to characterize the effect of control measures in different 665 provinces (see Table 3  High, medium and low represent the efficacy of control measures in three groups of provinces. 675 676 . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity. is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi.org/10. 1101/2020