We analyzed epidemiologic data describing the COVID-19 outbreak in South Korea during January 20–March 16. We transcribed daily numbers of reported cases in each municipality from Korea Centers for Disease Control and Prevention (KCDC) press releases (1). We also transcribed partial line lists from press releases by KCDC and municipal governments. All data and code are stored in a publicly available GitHub repository (https://github.com/parksw3/Korea-analysis).

Figure 1

Figure 1. Comparison of daily epidemiologic and traffic data from Daegu (A) and Seoul (B) during the coronavirus disease (COVID-19) outbreak, South Korea. Black bars indicate no. COVID-19 cases; lines represent daily...

We compared epidemiologic dynamics of COVID-19 from 2 major cities: Daegu (2020 population: 2.4 million) and Seoul (2020 population: 9.7 million). During January 20–March 16, KCDC reported 6,083 cases from Daegu and 248 from Seoul. The Daegu epidemic was characterized by a single large peak followed by a decrease (Figure 1, panel A); the Seoul epidemic comprised several small outbreaks (Figure 1, panel B).

We obtained the daily number of persons who boarded the subway or monorail in Daegu and Seoul during 2017–2020. For Daegu, we used data from https://data.go.kr for lines 1–3; for Seoul, we used data from https://data.seoul.go.kr for lines 1–9 (Figure 1). Soon after the first church-related case was reported, traffic volume decreased by »80% in Daegu and »50% in Seoul. To our knowledge, KCDC first recommended social distancing on February 29 (1), and no official guidelines existed regarding public transportation, which suggests that distancing was, at least in part, voluntary.

We reconstructed the time series of a proxy for incidence of infection I_t, representing the number of persons who became infected at time t and reported later, and estimated the instantaneous reproduction number, R_t, defined as the average number of secondary infections caused by an infected person, given conditions at time t (5). We adjusted the daily number of reported cases to account for changes in testing criteria and censoring bias (Appendix) and then sampled infection dates using inferred onset-to-confirmation delay distributions from the partial line list (Appendix) and previous estimated incubation period distribution (Table) to obtain our incidence proxy, I_t. Finally, we estimated R_t on the basis of the renewal equation (5):where w_t is the generation-interval distribution randomly drawn from a prior distribution (Table). We weighted each sample of R_t using a gamma probability distribution with a mean of 2.6 and a SD ± 2 to reflect prior knowledge (S. Abbott, unpub. data, https://doi.org/10.12688/wellcomeopenres.16006.1) and took weighted quantiles to calculate medians and associated 95% credible intervals. We estimated R_t for February 2 (14 days after the first confirmed case) through March 10 (after which the effects of censoring were too strong for reliable estimates) (Appendix). All analyses were performed using R version 3.6.1 (https://www.r-project.org).

Figure 2

Figure 2. Comparison of reconstructed coronavirus disease incidence proxy and instantaneous reproduction number R_t in Daegu (A, C) and Seoul (B, D), South Korea. The instantaneous reproduction number R_t...

We reconstructed incidence proxy (Figure 2, panels A, B) and estimates of R_t (Figure 2, panels C, D) in Daegu and Seoul. In Daegu, incidence peaked shortly after the first case was confirmed (Figure 2, panel A). In Daegu, symptoms had developed in the first case-patient on February 7; this person had visited the church on February 9 and 16, indicating the disease probably was spreading within the church community earlier (1). Likewise, the estimates of R_t gradually decreased and eventually decreased to <1 approximately 1 week after the first case was reported, coinciding with the decrease in the metropolitan traffic volume (Figure 2, panel C). The initial decrease in R_t was likely to have been caused by our resampling method for infection times for each reported case, which oversmooths the incidence curve and the R_t estimates (K. Gostic, unpub. data, https://doi.org/10.1101/2020.06.18.20134858). In Seoul, estimates of R_t decreased slightly but remained at »1 (Figure 2, panel D), which might be explained by less-intense social distancing. Stronger distancing or further intervention would have been necessary to reduce R_t to <1 by March 10.

Although we found clear, positive correlations on a daily scale between normalized traffic and the median estimates of R_t in Daegu (r = 0.93; 95% credible interval 0.86–0.96; Appendix) and Seoul (r = 0.76; 95% credible interval 0.60–0.87; Appendix), these correlations are conflated by time trends and by other measures that could have affected R_t. We did not find clear signatures of lags in the correlation between R_t and traffic volume (Appendix Figure 3). Patterns in R_t were similar in directly adjacent provinces (Gyeongsangbuk-do and Gyeonggi-do), demonstrating the robustness of our analysis (Appendix Figure 4).

The South Korea experience with COVID-19 provides evidence that epidemics can be suppressed with less extreme measures than those taken by China (9) and demonstrates the necessity of prompt identification and isolation of case-patients in preventing spread (4). Our analysis reveals the potential role of social distancing in assisting such efforts. Even though social distancing alone might not prevent spread, it can flatten the epidemic curve (compare Figure 2, panels B, D) and reduce the burden on the healthcare system (10).

Our study is not without limitations. Because of insufficient data, we could not account for differences in delay distributions or changes in testing capacity among cities; line list data were mostly derived from outside Daegu. Nonetheless, the sensitivity analyses support the robustness of our findings (Appendix Figures 5–8). We were unable to distinguish local and imported cases and thus might have overestimated R_t (11). Conducting a separate analysis for Seoul that accounts for imported cases did not affect our qualitative conclusions (Appendix Figure 9). Finally, although the method of resampling infection time can capture qualitative changes in R_t, estimates of R_t can be oversmoothed and should be interpreted with care (K. Gostic, unpub. data, https://doi.org/10.1101/2020.06.18.20134858). Nonetheless, our estimates of R_t are broadly consistent with previous estimates (12).

We used metropolitan traffic to quantify the degree of social distancing. The 80% decrease in traffic volume suggests that distancing measures in Daegu might be comparable to those in Wuhan, China (13). We were unable to directly estimate the effect of social distancing on population contacts or epidemiologic dynamics. Other measures, such as intensive testing and tracing of core transmission groups, are also likely to have affected transmission dynamics.

Our study highlights the importance of considering geographic heterogeneity in estimating epidemic potential. The sharp decrease in Daegu drove the number of reported cases in South Korea. Our analysis revealed that the epidemic remained close to the epidemic threshold in other regions, including Seoul and Gyeonggi-do. Relatively weak distancing might have assisted the recent resurgence of COVID-19 cases in Seoul (E. Shim, G. Chowell, unpub. data, https://doi.org/10.1101/2020.07.21.20158923).