The Effects of Stringent Interventions for Coronavirus Pandemic

2.1 Data collection

We compiled a total of 504 confirmed cases (Health Commission of Wenzhou, 2000) from January 21, 2020 when the confirmed cases were first reported, to February 25, 2020 when no newly reported confirmed cases had been reported for a week, from the official website of the health commission of Wenzhou, and obtained the residential population of Wenzhou from the 2019 statistical yearbook of the Zhejiang Provincial Bureau of Statistics (Statistics Bureau of Zhejiang Province, 2019). The total confirmed cases of Shenzhen during the same period were collected. The cases per 100,000 individuals were calculated by the prevalence of COVID-19 multiplying 100,000 individuals in Wenzhou. We searched the daily number of cumulative confirmed cases in each city of China using the R package “nCov2019” (Wu et al., 2020), and collected the number of cumulative confirmed cases from the top 100 cities for the epidemic of COVID-19 (except Hubei province) on February 25, 2020. The corresponding population density and GDP per capita statistics of these 100 cities were obtained from their latest statistical yearbooks.

2.2 Synthetic Control Method

To identify the treatment effects of the highly stringent interventions implemented on February 1, 2020, we compared Wenzhou with cities without those interventions. The synthetic control method was used to construct a data-driven control-group in the absence of those interventions in Wenzhou (Abadie and Gardeazabal, 2003; Abadie et al., 2010). The observed cases per 100,000 for Wenzhou (active intervention city) at the time t with interventions were denoted by , and cases per 100,000 for cities j at the time t without interventions by . The difference of was the treatment effect of the interventions on the cases per 100,000 for the active intervention city in the post-interventions period. Note that is counterfactual for the active intervention city in the post-interventions period. Reliable estimates of cases per 100,000 for a synthetic intervention city were used to derive , where a synthetic intervention city resembled the active intervention city by possible pre-interventions characteristics. We selected the GDP per-capita (Zhang et al., 2020), population density (Oto-Peralías, 2020), and three-day cases per 100,000 before the implementation of interventions to construct a resembling city of Wenzhou. The effects of unobserved factors varied over time and were controlled by the linear combination of cases per 100,000 of pre-interventions (Abadie et al., 2010). Thus, the treatment effect of interventions in Wenzhou was formulated as (Abadie et al., 2011):

To determine J, the number of cities demographically similar to Wenzhou, we selected the top 100 cities of cumulative confirmed cases on February 25, 2020 in China (except Hubei province) and used hierarchical clustering (Ward Jr, 1963) to group those cities based on population density and GDP per capita. A homogenous group including Wenzhou and other J cities were grouped together. Thus, those J cities could be considered as highly resembling to Wenzhou. The sum of weights for each of J cities was 1 and the values of weights in (2.1) were determined by pre-interventions characteristics. The placebo tests were used to identify the magnitude of the treatment effect (i.e. the causal effect of interventions) (Abadie et al., 2010).

2.3 Regression Discontinuity

For the strict intervention policies of Wenzhou, we vertically compared the difference in the epidemic situation of COVID-19 before and after the implementation of such intervention policies. Regression Discontinuity (RD) (Thistlethwaite and Campbell, 1960; Hahn et al., 2001; Imbens and Lemieux, 2008) was used to evaluate whether the implementation of strict intervention policies of Wenzhou on February 1, 2020 had a significant effect on its own COVID-19 epidemic. The outcome of these stringent interventions was formulated as:

where the logarithmic transformation of the daily number of cumulative confirmed cases as a response variable. The values of t include 1 to 36, representing the number of days starting from January 21, 2020 as a rate variable in Equation (2.2) of RD, for example, the date of February 1 was denoted as 12 of t. h is an indicator variable for the implementation of the strict intervention policies in Wenzhou, given a value of 1 as the implementation of the policies but a value of 0 as no implementation of the policies. This indicates that the values are 0 and 1 before and after February 1, 2020 (including February 1 itself), respectively and the subscript of t indicates the number of days as defined before. c is the number of days corresponding to the cutoff-point whose value is 12, representing the policy implemented on February 1, 2020. The values of k could be denoted as the k^th order of interaction effect between rate variable and policy. We compared the k^th order of interaction effect item in the (2.2) by AIC.

In RD, to determine whether c is an obvious jump point, the plot of both the rate variable and response variable was used to identify their relationship (Calonico et al., 2015a). Different numbers of bins could be used to divide both the left-hand side and right-hand side of the cutoff-point to small intervals. Thus, the scatter plot of both median values of the rate variable and mean values of the response variable could be drawn (Calonico et al., 2015b). The actual rate variables are weighted uniformly in each bin, and the fitted curves of both the left-hand side and right-hand side are used to examine whether the cutoff-point can be considered as a threshold.

2.4 Susceptible-Infected and infectious without isolation-Hospitalized in isolation-Removed (SIHR) model

The dynamic system of SIHR with four classes: Susceptible (S), infected and infectious without isolation (I), hospitalized in isolation (H), removed (R) (Li, 2018; Hsieh et al., 2004) was defined as:

where the transmission rate α is the average rate of being infected given contact over unit time, 1/β is the mean of the incubation period, 1/γ is the mean of the hospitalization period. To consider the effects of interventions, we introduced the time-varying transmission rate, which was defined as (Tan et al., 2020):

where α₀ denotes the maximum transmission rate of COVID-19 during the early outbreak, d represents the time when the interventions start to be effective and the transmission rate starts to decline, m is the duration of a process where the epidemic is to nearly vanish, λ_m is selected as (2 log((1 − ε)/ε))/m and ε is specified to be 0.01. The smaller the values of d and m, the earlier effectiveness and the stronger intensity of interventions were implemented, respectively. We simulated the likely outcomes of delaying stringent interventions from 0 to 5 days by changing the values of d and of mild interventions on the same starting time of Wenzhou policies by changing the value of m to Shenzhen policies.

3.1 Synthetic Control Method Analysis

We selected the top 100 cities outside Hubei province according to the cumulative confirmed cases on February 25, 2020 (Table 1 in the Supplementary Materials). Based on the per capita GDP and population density of each city, these cities (including Wenzhou) were clustered into 4 groups using hierarchical clustering. Among them, a homogenous group consisted of Wenzhou and other 45 cities (Figure 1 in the Supplementary Materials), where Taizhou of Zhejiang province implemented similar policies to Wenzhou and was excluded from our analysis. The remaining 44 cities formed a counterfactual city resembling Wenzhou (Table 2 in the Supplementary Materials).

• Download figure
• Open in new tab

Figure 1: The trends of COVID-19 cases per 100,000 between Wenzhou and Synthetic Wenzhou from January 21 to February 25, 2020.

The black dashed line indicated the time of implementation of the intervention policies in Wenzhou. The grey dashed line indicated 2 days delay of implementation of the intervention policies in Wenzhou.

Before the strict interventions were implemented on February 1, 2020, the trends of actual cases per 100,000 in both Wenzhou and “synthetic Wenzhou” were highly similar, suggesting that this synthetic city can be used to estimate the “counterfactual” results of Wenzhou. After 2 days of the implementation of the policies, the growth rate of actual cases per 100,000 in Wenzhou remarkably slower than that of “synthetic Wenzhou”. The number of cases per 100,000 in “Synthetic Wenzhou” on February 25, 2020 was 10.32, which is 1.69 times the actual Wenzhou (6.08) (Figure 1). In other words, Wenzhou had not implemented the strict interventions on February 1, 2020, the number of cumulative confirmed cases would have reached 954 on February 25, 2020, i.e., the epidemic of COVID-19 in Wenzhou would have expanded to approximately 1.7 times.

A placebo test was performed to determine the significance level of the difference in the trends of COVID-19 cases per 100,000. To this end, we plotted the gap curves between Wenzhou and synthetic Wenzhou by in turn exchanging Wenzhou and one of each 44 cities in the homogenous group of Wenzhou. The gap of COVID-19 cases per 100,000 between Wenzhou and our synthetic Wenzhou was the largest compared to the rest, i.e. the negative effect of the intervention policies on COVID-19 per 100,000 in Wenzhou was the lowest of all. For those 44 cities, the probability of having a gap for Wenzhou under a random permutation of the control measures was 5%, which is conventionally regarded as statistically significant. This suggested that the effect of the implementation of the policies in Wenzhou was significantly different from the implementation of the policies in the remaining 44 cities, indicating that the strict interventions of Wenzhou might have significantly reduced the COVID-19 cases per 100,000 (Figure 2).

• Download figure
• Open in new tab

Figure 2: The permutation test of the treatment effect of the implementation of the policies in Wenzhou and other 44 control regions.

All grey curves represent the placebo tests of COVID-19 cases per 100,000 gaps between a control city (an arbitrary random city of 44 cities) and synthetic control city (a combination of remaining 43 cities and Wenzhou), and the blue curve represents the placebo test of COVID-19 cases per 100,000 gap between Wenzhou and synthetic Wenzhou.

3.2 Regression Discontinuity Analysis

The values of AIC for each regression discontinuity model for different k^th order of interaction effects between rate variable and policy were shown in Table 1.

Looking at Table 1, the model included item Policy*Time³(h_t(t − c)³) had the lowest value of AIC but it had more parameters (df = 7). Also, the slopes of the fitted curves between the left-hand side and the right-hand side of the cutoff-point were different (Figure 3). The points lied on the right-hand side of quadratic curve. Thus, RD could be used to examine the treatment effect of interventions implemented in Wenzhou and the model included the second order of interaction effects between policy and rate variable was reasonable. The coefficient of interventions is –0.350 (p-value: 0.003) and the interaction effect between interventions and time is –0.438 (p-value: < 0.001) indicating that there was a significant treatment effect of highly stringent interventions implemented in Wenzhou on February 1, 2020 (Table 2).

• Download figure
• Open in new tab

Figure 3: The plot of both rate variable and response variable.

The green fitted curves of relationship between median of time within each bin and mean of log(cumulative confirmed cases) within each bin, the black dashed vertical line represents the time of cutoff-point (with value 12)

3.3 Simulation of stringent interventions delay or mild interventions instead

Our simulation projected that the expected cumulative confirmed cases would be 1.84 times of the actual cases for a 2-day delay, 2.45 times for a 3-day delay, 3.26 times for a 4-day delay and 4.30 times for a 5-day delay on February 25, 2020. The corresponding 95% credible interval (CI) for these projected numbers of cases were presented in Table 3. The full simulation results from January 21 to February 25, 2020 are presented in Figure 2 of the Supplementary Materials. According to Table 3, the expected cumulative confirmed cases for 0-day delay (i.e. the predicted cumulative confirmed cases) was very close to the actual cumulative confirmed cases of Wenzhou. The expected cumulative confirmed cases were 925 with corresponding 95% CI (571,1547) for the 2-day delay, 1233 with corresponding 95% CI (702,2233) for the 3-day delay, 1644 with corresponding 95% CI (851,3171) for the 4-day delay, and 2167 with corresponding 95% CI (1034, 4578) for the 5-day delay. The observed cumulative confirmed cases had been stable since February 17, 2020, however, the expected cumulative confirmed continued to increase if the stringent interventions were delayed. Based on Figure 1, if the stringent interventions were delayed by 2 days or more, the epidemic of COVID-19 in Wenzhou could have been more severe than that in synthetic Wenzhou (954 cumulative confirmed cases on February 25, 2020). If the mild interventions as those implemented in Shenzhen were implemented in Wenzhou on February 1, 2020, the expected number of cumulative confirmed cases would have been 2319 with corresponding 95% CI (1145,5189) on February 25, 2020 (Figure 4).

• Download figure
• Open in new tab

Figure 4: The expected number of cumulative confirmed cases for mild interventions implemented on February 1, 2020 in Wenzhou.

The points represent the actual cumulative confirmed cases of Wenzhou.