A new method for the joint estimation of instantaneous reproductive number and serial interval during epidemics

Chenxi Dai; Dongsheng Zhou; Bo Gao; Kaifa Wang

doi:10.1371/journal.pcbi.1011021

Abstract

Although some methods for estimating the instantaneous reproductive number during epidemics have been developed, the existing frameworks usually require information on the distribution of the serial interval and/or additional contact tracing data. However, in the case of outbreaks of emerging infectious diseases with an unknown natural history or undetermined characteristics, the serial interval and/or contact tracing data are often not available, resulting in inaccurate estimates for this quantity. In the present study, a new framework was specifically designed for joint estimates of the instantaneous reproductive number and serial interval. Concretely, a likelihood function for the two quantities was first introduced. Then, the instantaneous reproductive number and the serial interval were modeled parametrically as a function of time using the interpolation method and a known traditional distribution, respectively. Using the Bayesian information criterion and the Markov Chain Monte Carlo method, we ultimately obtained their estimates and distribution. The simulation study revealed that our estimates of the two quantities were consistent with the ground truth. Seven data sets of historical epidemics were considered and further verified the robust performance of our method. Therefore, to some extent, even if we know only the daily incidence, our method can accurately estimate the instantaneous reproductive number and serial interval to provide crucial information for policymakers to design appropriate prevention and control interventions during epidemics.

Author summary

Contagious disease epidemics, such as coronavirus disease 2019 (COVID-19), intermittently threaten global public health. In this situation, policymakers usually hope to timely monitor the severity of the disease transmission and thus evaluate the effectiveness of the prevention and control interventions. Instantaneous reproductive number and serial interval are the two most commonly used monitoring indices. The former refers to the expected number of secondary cases infected by a primary case, and the latter refers to the time interval between a primary case presenting with symptoms and its secondary cases developing symptoms. However, the existing frameworks can hardly estimate these two indicators simultaneously in the absence of additional contact tracing data or other prior information. In the presented study, we developed a likelihood method incorporating interpolation, Bayesian information criterion and Markov Chain Monte Carlo method to jointly estimate the two quantities, which only depends on the daily incidence. The simulation study and several data sets of historical epidemics verified the robust performance of our method. Therefore, even in the case of outbreaks of emerging infectious diseases with an unknown natural history or undetermined characteristics, our method can, to a certain extent, provide crucial information for policymakers to design appropriate measures during epidemics.

Citation: Dai C, Zhou D, Gao B, Wang K (2023) A new method for the joint estimation of instantaneous reproductive number and serial interval during epidemics. PLoS Comput Biol 19(3): e1011021. https://doi.org/10.1371/journal.pcbi.1011021

Editor: Eric H. Y. Lau, The University of Hong Kong, CHINA

Received: June 23, 2022; Accepted: March 9, 2023; Published: March 31, 2023

Copyright: © 2023 Dai et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: All relevant data are within the manuscript and its Supporting Information files.

Funding: Kaifa Wang was supported by the National Natural Science Foundation of China (grant numbers: 12171396, https://www.nsfc.gov.cn/) and by Chongqing Science and Technology Program (grant number: cstc2020jscx-cylhX0003, http://kjj.cq.gov.cn/). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Competing interests: The authors have declared that no competing interests exist.

1. Introduction

Contagious disease epidemics, such as the Spanish flu, smallpox, severe acute respiratory syndrome (SARS), and coronavirus disease 2019 (COVID-19), intermittently threaten global public health [1–4]. Currently, COVID-19 has spread worldwide since its initial outbreak in 2019, leading to an ongoing pandemic [5]. Mathematical models have been important tools to help policymakers design appropriate prevention and control interventions [6–8]. In these models, the serial interval and the reproductive number are regarded as the two most important parameters for understanding disease transmission [9].

The serial interval is defined as the time interval between a primary case presenting with symptoms and its secondary cases developing symptoms [10]. It is a mixture of the incubation period and infectious period, both of which are easy to understand but difficult to measure. Therefore, this quantity can only be estimated through detailed, time-consuming and expensive contact tracing.

The time-varying reproductive number is widely used to assess whether a disease will continue to spread under current control efforts [11], and its value denotes the expected number of secondary cases arising from a primary case infected at time t. In general, if the value is less than one, the outbreak will die out, while if it is greater than one, a sustained outbreak is likely. Thus, control measures must be implemented to decrease the time-varying reproductive number in practice.

Two formal definitions of the time-varying reproductive number have been proposed: case reproductive number and instantaneous reproductive number R_t [12]. By considering all possible transmission trees consistent with the observed epidemic data, the case reproductive number is estimated at each timestep with observed cases [8]. Due to its definition, this parameter can be estimated only in a delayed manner. For instance, if a policymaker wishes to understand the effects of control interventions in real time, this parameter may not be useful because it does not change immediately after interventions are altered [13]. In contrast, the instantaneous reproductive number changes immediately without delay and is therefore a useful quantity for understanding the effects of control strategies in real time [6]. Recognizing this advantage, Cori et al. subsequently developed a method and software (the EpiEstim R package) to estimate the instantaneous reproductive number with two inputs: daily incidence and the distribution of the serial interval [6]. Thompson et al. extended this statistical framework by integrating data on known pairs of primary cases and secondary cases from which the serial interval is directly estimated rather than relying on a priori estimates of the serial interval [13]. Although these methods [6,13] have been widely used, they may be limited in some contexts because of an important limitation, i.e., the serial interval and instantaneous reproductive number cannot be simultaneously estimated based on the daily incidence alone, resulting in the need for additional contact tracing data on the serial interval, which is often difficult and expensive to attain. In fact, in the initial phase of an outbreak, the serial interval is often not available or may be associated with significant uncertainty. This limitation is particularly notable for outbreaks of emerging infectious diseases with an unknown natural history or undetermined characteristics, as well as outbreaks of a known infectious diseases in a new population [14]. Although White et al. introduced a method to estimate both quantities and applied it to data from the 1918 influenza outbreak in a previous study [1], the approach is suitable only for a two-phase epidemic (an outbreak followed by control interventions). Thus, an improved method is required for complicated epidemic curves, such as two-wave epidemics and/or seasonal epidemics.

In this study, we extend the statistical framework [7] for estimating the instantaneous reproductive number and the serial interval directly from daily incidence without additional contact tracing data. The paper is organized as follows. After describing our method, we apply it first to simulated data from two artificial scenarios and then to several historical epidemics, including pandemic influenza [1], smallpox [2], SARS [3], COVID-19 [15,16], and a seasonal epidemic, hand-foot-mouth disease (HFMD) [17].

2. Materials and methods

2.1. Statistical method

Inspired by a previous study [7] (see S1 Text for details), we proposed an improved likelihood-based methodology to jointly estimate the instantaneous reproductive number and the serial interval using only daily incidence. The proposed method is divided into five key steps (Fig 1), and the specific contents of each step are described below. Firstly, we constructed the likelihood function of the instantaneous reproductive number and the serial interval. Secondly, the serial interval was parameterized by known distributions. Thirdly, the instantaneous reproductive number was modeled parametrically as a function of time using the interpolation method; Fourthly, the Bayesian information criterion (BIC) and the trust-region algorithm were employed to maximize the likelihood function and tune the undetermined parameters; Finally, the robust adaptive Metropolis (RAM) algorithm [18] was used to conduct the Markov Chain Monte Carlo (MCMC) procedure and estimate the target distribution of the instantaneous reproductive number as well as the serial interval. Notably, data obtained from Steps 4 and 5 can both be used as the final results. The MCMC method (MATLAB toolbox) was further utilized to enhance the credibility, and obtain the distribution and 95% credible interval (95% CI) of all parameters. Thus, data from Step 4 are usually reported as the final results for simulation study to save computational cost, while data from Step 5 for historical epidemic data sets.

Download:

Fig 1. A precise illustration of the framework for estimating the serial interval and the instantaneous reproductive number using daily incidence time series.

Step 1: Construct the likelihood function of the instantaneous reproductive number and the serial interval. Step 2: Parameterize the serial interval by known distributions. Step 3: Model the instantaneous reproductive number parametrically as a function of time using the interpolation method. Step 4: Employ the BIC and trust-region algorithm to maximize the likelihood function and tune undetermined parameters. Step 5: Use the MCMC method to estimate the target distribution of the instantaneous reproductive number and the serial interval. BIC: Bayesian information criterion; MCMC: Markov Chain Monte Carlo.

https://doi.org/10.1371/journal.pcbi.1011021.g001

2.1.1. Likelihood function of the instantaneous reproductive number and the serial interval.

Let N = {N_t}, t = 0,1,2⋯T represent the number of new cases per day between days 0 and T of the epidemic, which is the unique known data set in the procedure, and let X_ij represent the number of cases that appear on day j that are infected by individuals who develop symptoms on day i, which usually represents hidden, unknown data. An illustration of the model for disease transmission in the population is provided below. in which k denotes the length of the serial interval.

Similar to a previous report [7], we assumed that the total number of cases produced by N_t on day t, , exhibited a Poisson distribution with parameter N_tR_t, where R_t is the reproductive number for cases on day t. Furthermore, was assumed to follow a multinomial distribution with parameters (X_t,., p, k), where p = {p₁, p₂,⋯,p_k} represents the distribution of the serial interval. According to the aforementioned assumptions and several pioneering works [6,7], we constructed a likelihood function (see S1 Text for details). After simplification, it yields the following convenient form: (1) where . Maximization of the likelihood function (1) yields estimates of R_t and p.

2.1.2. Parameterized serial interval distribution by known distributions.

Serial interval instead of generation time was estimated, because the latter is often difficult to acquire in practice and its distribution is difficult to validate. Notably, Weibull, lognormal, gamma, and normal distributions (Table 1) are frequently used to fit serial interval distributions of various diseases [19,20], which are two-parameter distributions that provide rich family with sufficient flexibility to model a large number of infectious disease data sets. The criterion for choosing a distribution is whether it yields the largest value of likelihood function (1) among the four distributions. To reduce the dimensionality of the parameter space and create a concise and reliable model, p were parameterized by these distributions. Therefore, we mimicked p_i as (2) where f(x|a, b) is the probability density function of the serial interval. This formulation means that we discretized the distribution and, since k is finite, truncate it as well. ψ_i was normalized to ensure that the sum is unit so that p can represent a probability distribution. Since k was reported to exert a trivial impact on the results if k is sufficiently large [7], we specified the formula with the constraint that k<T. Furthermore, the normal distribution we used in this study was a truncated form, since our model assumed that only patients on day t would be infected by individuals who developed symptoms before day t. We set a>0, b>0 for Weibull, gamma, and truncated normal distributions, and a∈ℝ, b>0 for lognormal distribution to ensure the positive value of serial interval.

Download:

Table 1. Distributions used to estimate the serial interval.

https://doi.org/10.1371/journal.pcbi.1011021.t001

2.1.3. The instantaneous reproductive number was modeled parametrically as a function of time using the interpolation method.

Similarly, R_t was modeled parametrically as a function of time using an interpolation method with some data points {t_j, r_j}, j = 1,2,⋯,n (Fig 1, Step 3, red points), where n is the number of data points used for the interpolation of R_t. The data points {t_j, r_j} are key variables because they determine the shape of R_t and the value of the likelihood function. Although numerous interpolation methods were available, including linear interpolation, polynomial interpolation, spline interpolation, and nearest-neighbor interpolation [1,21], these methods required {t_j}, j = 1,2,⋯,n in monotonous order, and it was not easy to extrapolate when t<t₁ or t>t_n. Here, we provided an interpolation method without the aforementioned limitations, called the adaptive weighted neighbors method (AWN). Similar to Gaussian process/Kriging, AWN is also a kernel-based method that uses a limited set of sampled data points to estimate the value of a variable over a continuous spatial field (see our previous study for details [22]). Specifically, R_t was estimated using the following formulas: (3) where t_i≥0, r_i≥0, n≥2, i = 1,2,⋯,n. AWN made full use of the information contained in all key variables {t_j, r_j}, j = 1,2,⋯,n and automatically assigned greater weight to nearby interpolating points. The closer t is to t_i, the smaller d_i and the greater w_i, resulting in R_t closer to r_i. The prior distribution of r_j is U(0.5,5), where U represents a uniform distribution. These variables {t_j, r_j}, j = 1,2,⋯,n, parameter c and the aforementioned interpolation method (3) fit any shapes of R_t by increasing n and tuning c, which describe the timing of the intervention and the rapidity with which it affects transmission.

2.1.4. BIC and the trust-region algorithm were employed to maximize the likelihood function and tune undetermined parameters.

When fitting model (1), although the likelihood L(R_t, p|N) was increased by adding parameters, i.e., increasing n, doing so may result in overfitting and lower accuracy. According to a previous study [23], we introduced a penalty term for the number of parameters in the model and used BIC to resolve this problem. Concretely, the BIC was defined as follows: (4) where 2n+3 is the number of undetermined parameters and T is the number of data points of N. Similar to a previous report [24], an iterative method called the trust-region algorithm was used to minimize BIC and tune the parameters by supplying the gradients of the negative log-likelihood −ln L(R_t, p|N) with respect to the parameters (see S1 Text for details). According to the BIC results, we obtained the number of parameters, the number of data points (n), the distribution of the serial interval, and the values for all parameters, p and R_t (Fig 1, Step 4).

2.1.5. The MCMC method was used to generate the distribution of the instantaneous reproductive number and the serial interval.

From Subsection 2.1.4, we obtained the results of parameters as the initial values for the chain, and the parameters were constrained by (see details in S1 Table). The MCMC method based on MATLAB toolbox requires five elements for each parameter, including initial value, minimum value, maximum value, prior mean, and standard deviation (if the parameter is assumed to be normally distributed), but only the initial value must be provided and other elements are optional. S1 Table showed the prior information for each parameter. Then the RAM algorithm [18] was used to conduct the MCMC procedure and estimate the target distribution of p and R_t (Fig 1, Step 5). The algorithm ran for 2×10⁶ iterations after a burn-in of 10⁶ iterations, with the Geweke convergence diagnostic method employed to assess the convergence. Finally, we reliably obtained the distribution and 95% CI of the parameters {a, b, c}, key variables {t_j, r_j}, j = 1,2,⋯,n, p and R_t. Additionally, the codes of the proposed algorithm are available in S1 Codes.

2.2. Simulation study

We designed a simulation study based on two artificial scenarios to assess the ability of our proposed method to quantify the transmissibility of epidemics (see S1 Text for details). In scenario one, we verified the effectiveness of our method by changing different parameters, including the choice of Akaike information criterion (AIC) or BIC, the number of initial cases, the distribution of the serial interval, the mean and variance of the serial interval, the epidemic severity, the subsequent effectiveness of control measures, timeliness and other parameters. Then, we compared our approach with two previous methods (hereafter referred to as White et al method [1] and Cori et al method [6]). In scenario two, the outbreak-control-rebound-control epidemic curve was employed to verify that our method would fit any shape of curve for the instantaneous reproductive number and serial interval.

2.2.1. Scenario one.

Scenario one assumed that an epidemic had a constant R_t (R₁>1) before a certain date (day 40) and then had another constant value (R₂<1) to simulate the effect of control measures, such as school closures. We applied our method to an example in this scenario to illustrate the working principles and verify that it produced sensible results. This example assumed that the initial number of cases (N₀) was 2, the serial interval was lognormal distributed with mean and variance μ = 8, σ² = 9, and R₁ = 2.5, R₂ = 0.9. In addition, we illustrated the reason that the BIC instead of the AIC was chosen in this study.

Furthermore, the influence of different epidemic parameters (serial interval distribution, μ, σ², N₀, R₁, and R₂) on the effectiveness of the methods was simulated by changing the evaluation parameters while keeping the other parameters fixed. We first investigated the effects of the serial interval distribution (Weibull, lognormal, gamma, and normal). Then, the effect of the mean serial interval (μ = 6, μ = 8, and μ = 10) was studied. Next, we tested how the variance of the serial interval (σ² = 4, σ² = 9, and σ² = 16) affected the results. Fourth, the impact of the number of initial cases N₀ (2, 20) was investigated. Fifth, we tested how the epidemic severity (R₁ = 2.5, R₁ = 1.8) affected the results. Last, we quantified how R₂ (0.8, 0.9) affected the inference results to mimic the effectiveness of control measures.

To assess the timeliness of our method, we also clarified the impact of time length of the data on the performance of our method by varying the time length from 20 to 100. Besides, we performed a prospective analysis to predict the transmissibility in the next 20 days.

Note that serial interval has been shown to change over time, even in short periods of time and around the initial stages of an epidemic [25]. We then illustrated the effectiveness of our method by changing the mean serial interval from 8 days at the beginning of the epidemic to 3 days at the end of the epidemic. We also compared the results to the case with a fixed mean serial interval of 8 days.

Finally, we implemented the White et al method [1] and Cori et al method [6], and compared the results of the three methodologies. White et al method implemented a logistic curve that is suitable only for two-stage epidemics, including outbreak and control stages. The likelihood for the method was fitted using a Nelder-Mead maximization procedure, and 100 starting values were used to ensure that the global maximum was reached. This method was summarized in S2 Text. Cori et al. developed a method and software (the EpiEstim R package) for estimating the instantaneous reproductive number, but the method relies on the distribution of serial intervals. Thus, we calculated the results of Cori et al method based on the ground truth of serial interval. We also summarized Cori et al method in S2 Text.

2.2.2. Scenario two.

Scenario two simulated four two-wave epidemics to verify whether our method would fit any shapes of R_t. R_t was constant during the initial outbreak (R₁, T₁), control stage (R₂, T₂), rebound (R₃, T₃) and recontrol stage (R₄, T₄). R₁−R₄ and T₁−T₄ are two kinds of key parameters that determine the epidemic severity and duration of each stage, respectively. We assumed that N₀ = 2 and the serial interval exhibited a lognormal distribution with a mean and variance of μ = 8 and σ² = 9, respectively. Besides, we assessed the timeliness and performed a prospective analysis in this complicated scenario.

2.2.3. Quantitative validation measures.

To evaluate the performance of our method, three indicators, ΔR_t, Δμ, and Δσ, were used in this study.

ΔR_t was employed to evaluate the degree of similarity between the ground truth and the estimate of the instantaneous reproductive number, which was calculated using the following Eq: (5) where R_t and denote the ground truth and estimate, respectively. A lower ΔR_t represents a higher accuracy of .

Δμ and Δσ were used to evaluate the performance of serial interval estimates, (6) (7) where μ and σ are the ground truth mean and standard deviation of the serial interval, respectively, and and are the corresponding estimates. Similarly, lower Δμ and Δσ represent more accurate estimates.

2.3. Application to historical epidemics

To illustrate that our proposed method can provide new insights, we applied it to seven historical epidemics that varied in terms of reproductive number, serial interval, and epidemic scale. We retrieved the epidemic curves, as well as the means and standard deviations of the serial interval from the literatures and systematic reviews (Table 2). Meanwhile, we compared our method with two other methods: White et al method [1] and Cori et al method [6] (see S2 Text for details).

Download:

Table 2. Historical epidemics and their serial interval.

https://doi.org/10.1371/journal.pcbi.1011021.t002

A total of seven historical epidemics were collected. Firstly, data from a troop ship that embarked in the late fall of 1918 were reported. Boonah left Durban and reported the first three definitive cases of influenza on November 29. Researchers who collected the data noted that some initial cases were likely not identified. A total of 427 cases were reported in the 40 days of the epidemic. Secondly, another well-documented influenza outbreak in 1918 was considered. During the pandemic, the United States Public Health Service created special surveys of 18 localities, including Maryland, where Cumberland is one of the communities. In this community, the reported results were derived from house-to-house surveys requesting the date of onset of influenza for all infected individuals. Thirdly, four other data sets, including the smallpox outbreak in Kosovo in 1972, the SARS outbreak in Hong Kong in 2003, COVID-19 in Hunan, COVID-19 in Chongqing, were analyzed. Lastly, to illustrate the effectiveness of our method for complicated epidemic curves, we adopted HFMD data in Wenzhou, mainland China, which has a distinct seasonal pattern with more than two annual peaks in subtropical regions.

2.4. Statistical analysis

The statistical analysis was performed with SPSS 22.0 software (SPSS, Chicago, IL, USA). Since the data for ΔR_t, Δμ, and Δσ were not normally distributed, the Kruskal-Wallis test and its post hoc comparisons, and Wilcoxon test were used to analyze the effects of epidemic parameters (serial interval distribution, μ, σ², N₀, R₁, R₂, fixed and varying serial intervals), and AIC or BIC depending on the number of assigned groups. Friedman’s analysis of variance (ANOVA) with its post hoc test, and paired samples Wilcoxon test were used to analyze the effects of different methods (White et al, Cori et al and our methods), different time length, depending on the number of assigned groups, since their results were generated from the same data set. A two-tailed Student t-test was used for single comparison of standard deviation of serial interval between White et al and our methods). The null hypothesis of aforementioned tests stated that no differences were observed among/between groups. Chi-square test was used to measure the difference of the number of data points for R_t interpolation between AIC and BIC. Pearson correlation analysis was performed among parameters after MCMC analysis. P<0.05 was considered statistically significant.

3. Results

3.1. Simulated data set in scenario one

We applied our method to estimate the serial interval and instantaneous reproductive number throughout scenario one in the simulation study, for which only the daily incidence was available.

An example illustrating the workflow of our method is depicted in Fig 2. The time series of daily incidence was first generated under the condition of scenario one (N₀ = 2, the serial interval obeys a lognormal distribution with μ = 8, σ² = 9, and R₁ = 2.5, R₂ = 0.9). The daily incidence indicated that the epidemic first expanded rapidly during the outbreak and retracted gradually after a control measure on day 40. After our method was applied, the lowest value for BIC was observed when n = 3 and the lognormal distribution was used (Fig 2B), and the corresponding results for p and R_t were also acquired (Fig 2C and 2D). Then, the RAM algorithm was implemented to conduct the MCMC procedure (Fig 2E) and estimate the target distributions of p and R_t (Fig 2F and 2G). The MCMC chains, distributions, and correlations of 9 (2n+3) parameters are shown in S1 Fig. Besides, we chose the BIC since the better performance was observed when compared to AIC, such as lower ΔR_t, Δμ, Δσ, and number of data points for ΔR_t interpolation (see details in S2 Fig).

Download:

Fig 2. An example illustrating the workflow of our method.

The time series of daily incidence (A) was simulated according to the conditions of scenario one. Using the BIC (B), the number of data points and serial interval distribution were determined, and the serial interval (C) and instantaneous reproductive number (D) were obtained. The MCMC method (E) was then used to generate the distribution of the serial interval (F) and instantaneous reproductive number (G). BIC: Bayesian information criterion; MCMC: Markov Chain Monte Carlo.

https://doi.org/10.1371/journal.pcbi.1011021.g002

To better characterize our method’s performance, we tested how six epidemic parameters affected the results (S3 Fig). The simulations were performed based on the assumptions that the number of initial cases was 2, the serial interval exhibited a lognormal distribution with a mean and variance of 8 and 9, respectively and with a constant R before (R₁ = 2.5) and after (R₂ = 0.9) a control measure was implemented on day 40. In addition to the use of ΔR_t, Δμ and Δσ to assess the bias in estimates, the proportion of the 95% CI covering the true value is also important when evaluating the performance. However, we did not consider this indicator, because our method yielded high performance (0.99 [0.96 to 1.00, 95% CI]) and no difference would be detected when investigating the effects of different epidemic parameters. When we varied a parameter to study its effects, 100 epidemics were produced, and all other settings were kept unchanged. Since the distributions of the serial interval and instantaneous reproductive number are not needed when testing the effects of different parameters, we did not execute the MCMC method (Step 5). We found that the serial interval distribution exerted trivial effects on ΔR_t, Δμ and Δσ (S3A Fig). The mean serial interval exerted a marked impact on performance; it increased consistently as ΔR_t, Δμ and Δσ increase (S3B Fig). In addition, the variance in the serial interval was reported to increase with increasing and Δμ, Δσ except for ΔR_t (S3C Fig). The number of initial cases and epidemic severity also produced notable effects: larger N₀ and R₁ tended to yield accurate results, i.e., lower ΔR_t, Δμ and Δσ values (S3D and S3E Fig). However, we did not observe significant effects of R₂ on ΔR_t, Δμ and Δσ (S3F Fig).

To clarify the impact of time length of the data on the performance of our method, we analyzed the same simulated data sets by varying time length from 20 to 100 (Fig 3). Obviously, a longer time length yielded better performance, such as lower ΔR_t, Δμ and Δσ. In particular, the longer time length groups (60, 80, 100) exhibited significantly better performance than the shorter time length groups (20 and 40), including lower ΔR_t, Δμ and Δσ values. No difference was observed among longer time length groups and between shorter length groups, except that a remarkably lower ΔR_t was observed for time length 80, 100 compared to that of time length 60. Besides, we found that if R_t remained unchanged in the next 20 days, our method could accurately predict its trend (the third row in Fig 3A); otherwise, our method would lose its accuracy (the second row in Fig 3A).

Download:

Fig 3. The effects of the time length on the results.

A total of 100 trials were conducted to investigate the performance of our method by varying time length (from 20 to 100). Simulations were performed based on the assumptions that the number of initial cases was 2, the serial interval exhibited a lognormal distribution with a mean and variance of 8 and 9, respectively, and a constant R before (R₁ = 2.5) and after (R₂ = 0.9) a control measure on day 40. The left, middle and right panels are the estimates of instantaneous reproductive number, the estimates of the serial interval and the results of three indices, respectively. The blue lines show the estimates, and the dark lines denote their mean values under 100 simulations. Additionally, the grey lines denote predictions and dashed dark lines represent their mean values from 100 simulations. The red dashed lines represent the ground truth. *: P<0.05, and ***: P<0.001.

https://doi.org/10.1371/journal.pcbi.1011021.g003

We analyzed another simulated data set to show the effectiveness of our method when the serial interval changed over time (S4 Fig). The mean serial interval changed from 8 days at the beginning of the epidemic to 3 days at the end of the epidemic (S4A and S4B Fig). The true serial interval distribution that was averaged over time (red dashed curve) was compared with our estimates in S4D Fig. Surprisingly, our method also accurately estimated R_t and serial interval (S4C and S4D Fig), since no differences were observed between the ΔR_t, Δμ and Δσ for varying and fixed serial intervals (S4E, S4F and S4G Fig).

We analyzed the same simulated data sets to demonstrate our method’s performance relative to that of the White et al method and Cori et al method. We included the White et al method in the performance comparisons because it was designed for the outbreak-control mode with a two-stage function. Cori et al method was employed to demonstrate the robustness of our method even in the absence of serial interval. Since the distributions of the serial interval and instantaneous reproductive number were not needed in the comparison with the White et al method and to reduce computational cost, we did not execute the MCMC method (Step 5). As expected, all methods detected changes in the instantaneous reproductive number, for instance, a decrease in transmissibility following a control measure, and precisely estimate the serial interval distribution (Fig 4C). However, we also observed the inaccurate results of Cori et al method when R_t changed substantially. Moreover, our method performed comparably to White et al method since no differences were observed in ΔR_t, Δμ and Δσ (Fig 4D). We also observed better performance of our method and White et al method over Cori et al method, because of the remarkably lower ΔR_t of the former two methods (Fig 4D).

Download:

Fig 4. Comparison of methods in scenario one.

Daily incidence (A) was obtained based on the assumptions that the number of initial cases was 2, the serial interval exhibited a lognormal distribution with a mean and variance of 8 and 9, respectively, and a constant R before (R₁ = 2.5) and after (R₂ = 0.9) a control measure on day 40. The results of serial interval obtained by our method and White et al method (B). The results of instantaneous reproductive number obtained by all methods (C). The red dashed lines show the ground truth. ΔR_t, Δμ, and Δσ (D). The blue lines show the estimates, and the dark lines denote the mean values for 100 simulations. ns: no significant difference, ***: P<0.001.

https://doi.org/10.1371/journal.pcbi.1011021.g004

3.2. Simulated data set in scenario two

We next analyzed a more complicated epidemic scenario in which an outbreak was first controlled, then rebounded and was finally controlled again. Different choices of parameters, including epidemic severity (R₁, R₂, R₃, and R₄) and duration (T₁, T₂, T₃, and T₄), can lead to distinct epidemic modes (Fig 5, left panels). From each time series of daily incidence, we computed the serial interval and instantaneous reproductive number, and the results were illustrated in the middle and right panels of Fig 5. Our method successfully estimated p and R_t: all dashed lines of the ground truth were within the 95% CI (Fig 5, light blue area). One hundred trials were randomly selected to further verify the performance of our method in all possible epidemics simulated with the given parameters (R₁−R₄, and T₁−T₄). As shown in Fig 6, the serial interval and instantaneous reproductive number can be calculated faithfully because the mean estimates of both curves agreed well with their own ground truth.

Download:

Fig 5. Four examples in which an outbreak was first controlled, then rebounded and was finally controlled again were used for the simulation study in scenario two.

Simulations were performed based on the assumptions that the number of initial cases was 2 and the serial interval exhibited a lognormal distribution with a mean and variance of 8 and 9, respectively. Each epidemic was split into four stages, and R₁−R₄ and T₁−T₄ are key parameters that determine the epidemic severity and end time, respectively. The left, middle and right panels present the daily incidence and estimates of the serial interval and instantaneous reproductive number, respectively. The blue lines show the estimates, and the light blue area denotes their 95% credible intervals. The red dashed lines represent the ground truth.

https://doi.org/10.1371/journal.pcbi.1011021.g005

Download:

Fig 6. A total of 100 trials were conducted to further verify the performance of our method in scenario two.

Simulations were performed based on the assumptions that the number of initial cases was 2 and the serial interval exhibited a lognormal distribution with a mean and variance of 8 and 9, respectively. Each epidemic was split into four stages, and R₁−R₄ and T₁−T₄ are key parameters that determine the epidemic severity and end time, respectively. The left, middle and right panels present the daily incidence and estimates of the serial interval and instantaneous reproductive number, respectively. The blue lines show the estimates, and the dark lines denote their mean values after 100 simulations. The red dashed lines represent the ground truth.

https://doi.org/10.1371/journal.pcbi.1011021.g006

To clarify the impact of time length of the data on the performance of our method in this complicated case, we analyzed the same simulated data sets as Fig 6C by varying the time length from 40 to 200 (S5 Fig). Similar to the results in Fig 3, we found that a longer time length yielded better performance, such as lower ΔR_t, Δμ and Δσ. Besides, we found that if R_t remained unchanged in the next 20 days, our method accurately predicted its trend (the first and fourth rows in S5A Fig); otherwise, our method required additional time to improve the accuracy (the third row in S5A Fig).

3.3. Historical epidemic data sets

The estimated serial interval and R_t were obtained for seven historical epidemics from 1918 to 2020 using our proposed method and the methods developed by White et al. and Cori et al., and the results are presented in Fig 7.

Download:

Fig 7. Application of our method to seven historical epidemics.

The first column shows daily epidemic curves (from top to bottom) for the pandemic influenza in Boonah, 1918; the pandemic influenza in Cumberland, 1918; the smallpox in Kosovo, February–April 1972; the SARS in Hong Kong, February–June 2003; the COVID-19 in Hunan, 2020; the COVID-19 in Chongqing, 2020; and HFMD in Wenzhou, 2010–2011. The second and third columns show estimates of the serial interval p and instantaneous reproductive number R_t, respectively. The red lines and light red area, the blue lines, the green lines and the light green area in the middle and right columns denote the estimates obtained using our method, White et al method and Cori et al method, respectively. The dark dashed lines represent the threshold of one, and the red arrow denotes the onset of strict measures.

https://doi.org/10.1371/journal.pcbi.1011021.g007

3.3.1. Pandemic influenza for the Boonah ship, 1918.

We first considered data from a troop ship that embarked in the late fall of 1918. White et al. previously documented the effectiveness of their method in this simple outbreak-control epidemic and showed that the serial interval was 3.81±1.12 [1]. In Fig 7, a large initial R_t (>10) was apparent for the Boonah ship, which was consistent with the results reported by White et al. [1]. A previous study attributed this result to missing data at the beginning of the epidemic; thus, our method reported that a few initially infected individuals resulted in a large number of cases. Several days later, R_t rapidly decreased to sub-epidemic levels, indicating that many susceptible individuals were first infected and then the transmission decreased. We also observed a consistent trend using the three methods, but White et al method failed to detect the small peak around day 28. Furthermore, consistent with the published results [1], the serial interval calculated using our method was 4.00±1.35.

3.3.2. Pandemic influenza in Cumberland, 1918.

We now considered another influenza outbreak in 1918. Based on the daily incidence (Fig 7, the panel in the first column, second row), we observed a sharp increase on day 30, with a median estimate of 6.32 (95% CI: 4.80–8.67). R_t then fell below 1 after day 44 and this trend persisted until the end of the outbreak. White et al and Cori et al methods produced the same tendency as our method (Fig 7, the panel in the last column, second row). For the serial interval, our estimate (7.14±3.86) was consistent with White et al (8.28±5.00). Thus, the first two cases of pandemic influenza verified the effectiveness of our method for the simple outbreak-control situation using historical data.

3.3.3. Smallpox in Kosovo, 1972.

We then analyzed data from the smallpox outbreak. Here, for a small epidemic with a long mean serial interval, R_t initially had a high value of 10.9 (95% CI: 0–57.6) and this high level was maintained for 40 days. Then, R_t continued to decrease and eventually decreased to the threshold of one on day 48. We observed the same trend for the two other methods, but the turning point (day 40) of White et al method was earlier than that of the two other methods. In addition, our method provided the same results for the serial interval with previously reported values (Table 2), but White et al method underestimated the mean of serial interval (12.06±13.39).

3.3.4. SARS in Hong Kong, 2003.

For the SARS outbreak in Hong Kong in 2003, R_t dramatically decreased before day 15 and then slightly rebounded to 2.14 (95% CI: 1.60–2.82) on day 23. Several days later, R_t decreased below the threshold of one, and the epidemic was finally controlled. A similar trend was observed for Cori et al method, while White et al method produced a slightly lower R_t before day 20. For the serial interval, the reported value was consistent with our estimate (8.4±3.8 vs. 9.70±4.57), but White et al. reported a lower value (3.89±3.25 vs. 8.4±3.8).

3.3.5. COVID-19 in Hunan, 2020.

We next analyzed a recent pandemic, COVID-19 in Hunan. R_t was clearly divided into two phases. In the first phase (from day 1 to 19), R_t was higher than the threshold of one, indicating an outbreak. In the second phase (from day 20 to 45), R_t decreased below the threshold of one, representing effective control of the epidemic. The results of White et al method was consistent with our estimate, while Cori et al method delayed the second phase. The serial interval estimated by our method and White et al method were 3.42±1.48 and 1.98±0.84, respectively, which were slightly lower than the previously reported value of 4.40±3.17 [26].

3.3.6. COVID-19 in Chongqing, 2020.

Another COVID-19 outbreak was then analyzed. On January 1, 2020 (day 0), the first case was imported to Chongqing, and as of February 24, 2020, a total of 576 cases of COVID-19 were confirmed. The initial outbreak led to a high R_t (4.15 with 95% CI 2.25–9.53). On January 24, 2020 (day 24), the local government gradually implemented and strengthened prevention and control measures for the epidemic. As a result, the reproductive number dropped below one on day 28 and continued to decrease until the end of the epidemic. At the same time, the other two methods reported the same trend. In this case, our method (3.79±0.86 vs. 5.2±6.2) and White et al method (3.89±0.97 vs. 5.2±6.2) both obtained lower estimates for the serial interval.

3.3.7. HFMD in Wenzhou, 2010–2011.

HFMD was finally analyzed to illustrate the effectiveness of our method for complicated epidemic curves. We found three peaks in daily incidence (Fig 7), and each peak coincided with a period of R_t>1. The first period occurred from day 1 to 146, with a median estimate of 1.07; the second outbreak occurred from day 403 to 541, with a median of 1.07; and the last outbreak occurred from day 611 to 701, with a median of 1.03. Cori et al method reported the same trend, while White et al method failed to estimate R_t. Our method yielded a similar result for the serial interval (3.86±5.08 vs. 3.7±2.6) as a previous report [27], further illustrating its robust performance, while White et al method reported a lower serial interval (2.39±2.69 vs. 3.7±2.6).

4. Discussion

We developed a new framework to simultaneously estimate the instantaneous reproductive number and serial interval using only the time series of daily incidence. Our approach built on a well-established approach [1,7] and addressed a crucial drawback that many studies have faced. The distinct feature of our framework is that the serial interval distribution and instantaneous reproductive number can be estimated jointly from the latest available data without including additional contact tracing data in the estimation procedure. By incorporating AWN interpolation and BIC, our method precisely estimated transmissibility and described the timing of an intervention and the rapidity with which it affects transmission. In addition, we provided four alternative distributions used for serial interval estimation to further maximize the likelihood function (Eq 1). More importantly, the MCMC method provided a reliable result for determining the distribution of the instantaneous reproductive number and serial interval, which further helped us to understand the variable range of an epidemic. The results for two artificial scenarios and seven historical epidemics suggest that our method has promising advantages for application in estimating transmissibility.

White et al employed a likelihood-based methodology tailored to a two-phase epidemic to estimate the serial interval and the instantaneous reproductive number for each day of an epidemic and then applied this method to data from the 1918 influenza outbreak [1]. In artificial scenario one, this method provided reliable results for estimating these quantities (Fig 4). By extending the statistical framework reported previously [7], the current study provides an alternative and improved method for estimating these quantities for multiphase epidemics. First, our method provided comparable results to those obtained using White et al method for a two-phase outbreak. Second, the reliable estimation results of the instantaneous reproductive number for four-phase epidemics (artificial scenario two, Figs 5 and 6) further extended the application of our method. In contrast to White et al method, which is tailored to a two-phase epidemic, our method automatically divided the epidemic into n stages via AWN interpolation, and n was determined using BIC. Thus, our method can estimate any shape of the curve for instantaneous reproductive number.

The most commonly used approach to estimate instantaneous reproductive number is Cori et al method [6]. The first point that hinders its popularity is that it uses pre-existing estimates of the serial interval distribution as an input. This approach potentially leads to delays between inferring the serial interval and subsequent estimation of transmissibility or indicates that estimates of R_t are based on estimates of the serial interval from earlier outbreaks. However, it may not be accurate when emerging infectious diseases with unknown natural history outbreaks. Secondly, R_t estimated by Cori et al method did not change immediately after interventions were applied (Fig 4C) due to the sliding window used; instead, it changed smoothly in a delayed manner. In contrast, our method reported that R_t changed immediately without delay and produced lower ΔR_t values compare with Cori et al method (Fig 4D). Thus, an estimate of Cori et al method may not be the ideal choice for policymakers who wish to understand the effects of control interventions.

Our method may be promising when serial interval information is missing. First, we observed a continuous curve for the instantaneous reproductive number over time due to AWN interpolation, which provided a clear illustration of current control measures for policymakers (Fig 7). Second, the estimates of the serial interval were much more accurate than those using White et al method (Table 2). Last, the instantaneous reproductive number changed immediately after the intervention. For instance, after the local government implemented strict measures on day 24 (Fig 7, COVID-19 in Chongqing), the instantaneous reproductive number rapidly decreased below the threshold 3 days later.

We next reported the reason why the distribution of serial interval was estimated instead of generation time. Serial interval is the time between onset of symptoms of a case and onset of symptoms of his/her secondary cases, while generation time refers to the time from the infection of a primary case to infection of the cases he/she generates. Thus, generation time differs from serial interval. We applied our method to data consisting of daily counts of symptom onset where the infectivity profile p was approximated from the distribution of the serial interval because the timing of symptom onset is usually known, and such data collected in closed settings can reliably be ascertained after an epidemic. However, time of infection are rarely observed, and the generation time distribution is therefore difficult to measure and validate.

In addition to interventions, other factors may affect the instantaneous reproductive number. Firstly, the results obtained using our method may be sensitive to the time length during the epidemic. This result is clearly presented in Fig 3 and S5 Fig, in a longer time length, particularly a value greater than 40, yielded better performance, including lower ΔR_t, Δμ and Δσ values. Therefore, it should be carefully used especially in the early stage of the epidemic. Secondly, our method may be feasible to predict unchangeable transmissibility. Due to the nature of interpolation, our method performed well when R_t remained unchanged over the next 20 days (Fig 3 and S5). However, when predicting sharp changes in R_t, our method requires more information to improve the accuracy. Thirdly, serial interval that changed over time had a trivial impact on the R_t estimation. Serial interval has been shown to change over time, and this change is driven by enhanced nonpharmaceutical interventions, particularly case isolation [25,35]. Our simulation results showed that even the use of a fixed serial interval did not affect the final result of the R_t estimation (S4 Fig), which largely expands the application of our method. The underlying reason of unbiased estimates of R_t (S4C Fig) may attributed to that our method simultaneously estimates the serial interval and instantaneous reproductive number, which provides the possibility that the estimated serial interval fits the true serial interval distribution that was averaged over time well. Lastly, seasonal variations in the parameters governing disease spread play a crucial role in the transmission of many pathogens. For example, the transmission of HFMD varies due to factors including temperature, relative humidity, and school opening [17]. Our method successfully estimated the instantaneous reproductive number and serial interval based on the daily incidence. We observed that the instantaneous reproductive number in the spring semester was higher than the threshold of one, while its value in the fall semester was lower than the threshold, which depicted seasonal variation.

A previous study showed that an increasing epidemic size led to improvements in the reproductive number estimate [7]. The serial interval distribution, μ, σ², N₀, R₁, and R₂ are key parameters that impact the epidemic size. We thus first tested our method by varying these epidemic parameters. Increasing the number of initial cases, decreasing the mean serial interval, and increasing the epidemic severity (R₁) of a particular disease decreased ΔR_t, Δμ and Δσ (Fig 3B, 3D and 3E) because these parameters may substantially affect the scale of an epidemic. In contrast, the degree of control measures (R₂), variance and mean of the serial interval, and serial interval distribution exerted a noninfluential impact on ΔR_t; the underlying reason was that these factors may not be significantly associated with the epidemic size (Fig 3A, 3C and 3F). This evidence may partially explain the varying performance of our method under different conditions.

However, the present study has some limitations that should be addressed in future studies. Firstly, we assumed that secondary cases were generated in a Poisson distribution according to previous studies [1,6,7,13]. This assumption may not be accurate because it was not verified by real data. In fact, infectious diseases may have different characteristics that would lead primary infectors to generate secondary cases at varying rates, which may produce many possible distributions and increase the difficulty of estimation outcomes. Although this approach may not be perfectly accurate for disease generation, this assumption simplified the estimation process. Secondly, the usefulness was not assessed by determining the effect of right truncation. Due to different delays, the date of infection cases cannot be reported in a timely manner, and thus, reported cases in real or near-real time are systematically right-censored [36]. Thirdly, this study did not address the issue of the imperfect observation of cases. Numerous studies have shown that reported cases particularly in the beginning of an epidemic, are commonly biased [36–38], and only the most severe cases are reported, or diagnostics are weighted differently across subpopulations or age groups. Thus, our method should be carefully applied if the distribution of ascertainment and reporting rates are not distributed homogeneously over time, which may lead to severely biased estimates. Fourthly, the interchangeability between the serial interval and the generation time was not assessed, and R_t calculated from serial interval may suffer from bias. A previous study has showed the relationship between serial interval and generation time, and the latter can be estimated from the former but with inaccurate variance [39]. Further study should be implemented to directly estimate generation time with highly accurate variance, and improve the accuracy of R_t estimation. Fifthly, we also assumed that serial interval was assumed to be positive [6], but it can sometimes take negative values, especially for COVID-19 virus infection. Lastly, from a public health/surveillance perspective, interventions, such as case isolation, can shorten the mean serial interval and change the instantaneous reproductive number, and then may pose challenges to our method. Despite the high performance our method achieved when serial interval changed over time (S4 Fig), serial interval can vary over time during the epidemic due to sampling bias and additional studies should be conducted in the future to construct a model of this issue.

5. Conclusions

In this study, we developed a new framework to jointly estimate the instantaneous reproductive number and serial interval by relying only on the time series of daily incidence. The simulation study revealed that our estimates of the two quantities were consistent with the ground truth. Seven data sets of historical epidemics further verified the robust performance of our method, and our analysis was also in agreement with previously published results. Without the need for additional contact tracing data, our simple method will be helpful for the timely analysis of new outbreaks and the retrospective study of historical epidemics. Further studies are required to verify the robustness of our method, especially surveillance data.

Supporting information

S1 Text. Additional information on model specification.

The file contains four sections: (1) White et al method, (2) simplification of the likelihood, (3) model for generating simulation data, (4) the gradient of the negative log-likelihood −ln L(R_t, p|N).

https://doi.org/10.1371/journal.pcbi.1011021.s001

(DOCX)

S2 Text. Comparative methods.

The file includes White et al method and Cori et al method.

https://doi.org/10.1371/journal.pcbi.1011021.s002

(DOCX)

S1 Table. Prior information of parameters for MCMC method based on MATLAB toolbox.

https://doi.org/10.1371/journal.pcbi.1011021.s003

(DOCX)

S2 Table. Comparison of standard deviation of serial interval between White et al and our method.

https://doi.org/10.1371/journal.pcbi.1011021.s004

(DOCX)

S1 Fig. MCMC results of each parameter.

MCMC chains (A), distributions (B), and correlation matrix (C) of all nine parameters. Results of Geweke convergence diagnostic method were shown in the top of each chain (A), and P>0.05 was diagnosed as convergent chain. The MCMC algorithm ran for 2×10⁶ iterations with a burn-in of 10⁶ iterations. MCMC: Markov Chain Monte Carlo.

https://doi.org/10.1371/journal.pcbi.1011021.s005

(DOCX)

S2 Fig. Comparison between the results of BIC and AIC. ΔR_t.

(A), Δμ (B), Δσ (C), and the number of key variables {t_j, r_j}, j = 1,2,⋯,n used for R_t interpolation (D). Compared to the AIC, BIC required significantly a lower number of key variables for R_t interpolation, meaning that BIC tended to select less complex model. Meanwhile, BIC was a more effective penalized model selection criteria since it produced remarkably lower ΔR_t, Δμ, and Δσ, when compared with AIC. A total of 100 trials were conducted to investigate AIC and BIC results. Simulations were performed based on the assumptions that the number of initial cases was 2, the serial interval exhibited a lognormal distribution with a mean and variance of 8 and 9, respectively, and a constant R before (R₁ = 2.5) and after (R₂ = 0.9) a control measure on day 40. AIC: Akaike information criterion; BIC: Bayesian information criterion. *: P<0.05, and ***: P<0.001.

https://doi.org/10.1371/journal.pcbi.1011021.s006

(DOCX)

S3 Fig. The performance of our method when varying different parameters.

(A) Serial interval distribution. (B) Mean value of the serial interval. (C) Standard deviation of the serial interval. (D) The number of initial cases. (E) The epidemic severity R₁. (F) The effectiveness of control measures R₂. Simulations were performed based on the assumptions that the number of initial cases was 2, the serial interval exhibited a lognormal distribution with a mean and variance of 8 and 9, respectively, and a constant R before (R₁ = 2.5) and after (R₂ = 0.9) a control measure on day 40. When we varied a parameter to study its effects, the others were kept unchanged. *: P<0.05, **: P<0.01, and ***: P<0.001.

https://doi.org/10.1371/journal.pcbi.1011021.s007

(DOCX)

S4 Fig. The effects of changing serial interval on the results.

(A) Distribution of serial interval over time. (B) Daily incidence generated by varying the serial interval. (C) Estimates of instantaneous reproductive number. (D) Estimates of serial interval. The red dashed curve denotes the true serial interval distribution averaged over time. (E) ΔR_t. (F) Δμ is the absolute error between the true mean serial interval averaged over time and our estimates. (G) Δσ is the absolute error between the true standard deviation of the serial interval averaged over time and our estimates. A total of 100 trials were conducted to investigate the performance of our method by varying mean serial interval (from 8 to 3) and fixed mean serial interval (8). Simulations were performed based on the assumptions that the number of initial cases was 2, the serial interval exhibited a lognormal distribution with a variance of 9, and a constant R before (R₁ = 2.5) and after (R₂ = 0.9) a control measure on day 40. *: P<0.05, and ***: P<0.001.

https://doi.org/10.1371/journal.pcbi.1011021.s008

(DOCX)

S5 Fig. The effects of time length on the results.

A total of 100 trials were conducted to investigate the performance of our method by varying the time length (from 40 to 200). Simulations were performed based on the assumptions that the number of initial cases was 2, the serial interval exhibited a lognormal distribution with a mean and variance of 8 and 9, respectively, and it was split into four stages. The left, middle and right panels are the estimates of instantaneous reproductive number, the estimates of the serial interval and the results of the three indices, respectively. The blue lines show the estimates, and the dark lines denote their mean values after 100 simulations. Additionally, the gray lines denote predictions and dashed dark lines represent their mean values after 100 simulations. The red dashed lines represent the ground truth. *: P<0.05, **: P<0.01, and ***: P<0.001.

https://doi.org/10.1371/journal.pcbi.1011021.s009

(DOCX)

S1 Codes. MATLAB main model file.

This MATLAB script (main_an_example.m) provided an example illustrating the workflow of our method.

https://doi.org/10.1371/journal.pcbi.1011021.s010

(ZIP)

References

1. White LF, Pagano M. Transmissibility of the influenza virus in the 1918 pandemic. PLoS One. 2008;3. pmid:18231585
- View Article
- PubMed/NCBI
- Google Scholar
2. Hennein SS. Smallpox and Its Eradication. JAMA J Am Med Assoc. 1988;260: 559.
- View Article
- Google Scholar
3. Cori A, Boëlle PY, Thomas G, Leung GM, Valleron AJ. Temporal variability and social heterogeneity in disease transmission: The case of SARS in Hong Kong. PLoS Comput Biol. 2009;5. pmid:19696879
- View Article
- PubMed/NCBI
- Google Scholar
4. China CDC. The Epidemiological Characteristics of an Outbreak of 2019 Novel Coronavirus Diseases (COVID-19)—China, 2020. China CDC Wkly. 2020;2: 113–122. pmid:34594836
- View Article
- PubMed/NCBI
- Google Scholar
5. Sharma A, Ahmad Farouk I, Lal SK. Covid-19: A review on the novel coronavirus disease evolution, transmission, detection, control and prevention. Viruses. 2021;13. pmid:33572857
- View Article
- PubMed/NCBI
- Google Scholar
6. Cori A, Ferguson NM, Fraser C, Cauchemez S. A new framework and software to estimate time-varying reproduction numbers during epidemics. Am J Epidemiol. 2013;178: 1505–1512. pmid:24043437
- View Article
- PubMed/NCBI
- Google Scholar
7. White LF, Pagano M. A likelihood-based method for real-time estimation of the serial interval and reproductive number of an epidemic. Stat Med. 2008;27: 2999–3016. pmid:18058829
- View Article
- PubMed/NCBI
- Google Scholar
8. Wallinga J, Teunis P. Different Epidemic Curves for Severe Acute Respiratory Syndrome Reveal. Am J Epidemiol. 2004;160: 509–516. pmid:15353409
- View Article
- PubMed/NCBI
- Google Scholar
9. Fine PEM. The Interval between Successive Cases of an Infectious Disease. Am J Epidemiol. 2003;158: 1039–1047. pmid:14630599
- View Article
- PubMed/NCBI
- Google Scholar
10. Svensson Å. A note on generation times in epidemic models. Math Biosci. 2007;208: 300–311. pmid:17174352
- View Article
- PubMed/NCBI
- Google Scholar
11. Nishiura H, Chowell G. The effective reproduction number as a prelude to statistical estimation of time-dependent epidemic trends. Math Stat Estim Approaches Epidemiol. 2009; 103–121.
- View Article
- Google Scholar
12. Fraser C. Estimating individual and household reproduction numbers in an emerging epidemic. PLoS One. 2007;2. pmid:17712406
- View Article
- PubMed/NCBI
- Google Scholar
13. Thompson RN, Stockwin JE, Van Gaalen RD, Polonsky JA, Kamvar ZN. Improved inference of time-varying reproduction numbers during infectious disease outbreaks. Epidemics. 2019;29: 1–11. pmid:31624039
- View Article
- PubMed/NCBI
- Google Scholar
14. Metcalf CJE, Lessler J. Opportunities and challenges in modeling emerging infectious diseases. Science (80-). 2017;357: 149–152. pmid:28706037
- View Article
- PubMed/NCBI
- Google Scholar
15. Wang K, Zhao S, Li H, Song Y, Wang L, Wang MH, et al. Real-time estimation of the reproduction number of the novel coronavirus disease (COVID-19) in China in 2020 based on incidence data. Ann Transl Med. 2020;8: 689–689. pmid:32617309
- View Article
- PubMed/NCBI
- Google Scholar
16. Dai C, Yang J, Wang K. Evaluation of prevention and control interventions and its impact on the epidemic of coronavirus disease 2019 in Chongqing and Guizhou Provinces. Math Biosci Eng. 2020;17: 2781–2791. pmid:32987495
- View Article
- PubMed/NCBI
- Google Scholar
17. Dai C, Wang Z, Wang W, Li Y, Wang K. Epidemics and underlying factors of multiple-peak pattern on hand, foot and mouth disease in Wenzhou, China. Math Biosci Eng. 2019;16: 2168–2188. pmid:31137205
- View Article
- PubMed/NCBI
- Google Scholar
18. Vihola M. Robust adaptive Metropolis algorithm with coerced acceptance rate. Stat Comput. 2012;22: 997–1008.
- View Article
- Google Scholar
19. Vink MA, Bootsma MCJ, Wallinga J. Serial intervals of respiratory infectious diseases: A systematic review and analysis. Am J Epidemiol. 2014;180: 865–875. pmid:25294601
- View Article
- PubMed/NCBI
- Google Scholar
20. Nishiura H, Linton NM, Akhmetzhanov AR. Serial interval of novel coronavirus (COVID-19) infections. Int J Infect Dis. 2020;93: 284–286. pmid:32145466
- View Article
- PubMed/NCBI
- Google Scholar
21. Akima H. A New Method of Interpolation and Smooth Curve Fitting Based on Local Procedures. J ACM. 1970;17: 589–602.
- View Article
- Google Scholar
22. Dai C, Wang K. Adaptive Weighted Neighbors Method for Sensitivity Analysis. Interdiscip Sci Comput Life Sci. 2022. pmid:35426544
- View Article
- PubMed/NCBI
- Google Scholar
23. Schwarz G. Estimating the Dimension of a Model. Ann Stat. 1978;6: 461–464.
- View Article
- Google Scholar
24. Coleman TF, Li Y. An interior trust region approach for nonlinear minimization subject to bounds. SIAM J Optim. 1996;6: 418–445.
- View Article
- Google Scholar
25. Ali ST, Wang L, Lau EHY, Xu XK, Du Z, Wu Y, et al. Serial interval of SARS-CoV-2 was shortened over time by nonpharmaceutical interventions. Science (80-). 2020;369: 1106–1109. pmid:32694200
- View Article
- PubMed/NCBI
- Google Scholar
26. Leung K, Wu JT, Leung GM. Real-time tracking and prediction of COVID-19 infection using digital proxies of population mobility and mixing. Nat Commun. 2021;12: 1–8. pmid:33686075
- View Article
- PubMed/NCBI
- Google Scholar
27. Chang LY, Tsao KC, Hsia SH, Shih SR, Huang CG, Chan WK, et al. Transmission and Clinical Features of Enterovirus 71 Infections in Household Contacts in Taiwan. J Am Med Assoc. 2004;291: 222–227. pmid:14722149
- View Article
- PubMed/NCBI
- Google Scholar
28. Fraser C, Cummings DAT, Burke DS, Ferguson NM, Cauchemez S, Riley S, et al. Strategies for containing an emerging influenza pandemic in Southeast Asia. Nature. 2005;437: 209–214. Available from: http://www.ncbi.nlm.nih.gov/pubmed/16079797. pmid:16079797
- View Article
- PubMed/NCBI
- Google Scholar
29. Riley S, Ferguson NM. Smallpox transmission and control: Spatial dynamics in Great Britain. Proc Natl Acad Sci U S A. 2006;103: 12637–12642. pmid:16894173
- View Article
- PubMed/NCBI
- Google Scholar
30. Chelsky M, Angulo JJ. Analysis of the incidence of variola minor in an outbreak by use of a mathematical model. Am J Trop Med Hyg. 1977;26: 152–158. pmid:842777
- View Article
- PubMed/NCBI
- Google Scholar
31. Lipsitch M, Cohen T, Cooper B, Robins JM, Ma S, James L, et al. Transmission dynamics and control of severe acute respiratory syndrome. Science (80-). 2003;300: 1966–1970. pmid:12766207
- View Article
- PubMed/NCBI
- Google Scholar
32. You C, Deng Y, Hu W, Sun J, Lin Q, Zhou F, et al. Estimation of the Time-Varying Reproduction Number of COVID-19 Outbreak in China. SSRN Electron J. 2020. pmid:32460229
- View Article
- PubMed/NCBI
- Google Scholar
33. Griffin J, Casey M, Collins Á, Hunt K, McEvoy D, Byrne A, et al. Rapid review of available evidence on the serial interval and generation time of COVID-19. BMJ Open. 2020;10. pmid:33234640
- View Article
- PubMed/NCBI
- Google Scholar
34. Koh WM, Bogich T, Siegel K, Jin J, Chong EY, Tan CY, et al. The Epidemiology of Hand, Foot and Mouth Disease in Asia. Pediatr Infect Dis J. 2016;35: e285–e300. pmid:27273688
- View Article
- PubMed/NCBI
- Google Scholar
35. Park SW, Sun K, Champredon D, Li M, Bolker BM, Earn DJD, et al. Forward-looking serial intervals correctly link epidemic growth to reproduction numbers. Proc Natl Acad Sci U S A. 2020;118. pmid:33361331
- View Article
- PubMed/NCBI
- Google Scholar
36. Gostic KM, McGough L, Baskerville EB, Abbott S, Joshi K, Tedijanto C, et al. Practical considerations for measuring the effective reproductive number, Rt. PLoS Comput Biol. 2020;16. pmid:33301457
- View Article
- PubMed/NCBI
- Google Scholar
37. Lipsitch M, Donnelly CA, Fraser C, Blake IM, Cori A, Dorigatti I, et al. Potential biases in estimating absolute and relative case-fatality risks during outbreaks. PLoS Negl Trop Dis. 2015;9: 1–16. pmid:26181387
- View Article
- PubMed/NCBI
- Google Scholar
38. De Salazar PM, Lu F, Hay JA, Gómez-Barroso D, Fernández-Navarro P, Martínez E V., et al. Near real-time surveillance of the SARS-CoV-2 epidemic with incomplete data. PLoS Comput Biol. 2022;18. pmid:35358171
- View Article
- PubMed/NCBI
- Google Scholar
39. Lehtinen S, Ashcroft P, Bonhoeffer S. On the relationship between serial interval, infectiousness profile and generation time: On the relationship between serial interval, infectiousness profile and generation time. J R Soc Interface. 2021;18. pmid:33402022
- View Article
- PubMed/NCBI
- Google Scholar

[ref1] 1. White LF, Pagano M. Transmissibility of the influenza virus in the 1918 pandemic. PLoS One. 2008;3. pmid:18231585
View Article
PubMed/NCBI
Google Scholar

[2] View Article

[3] PubMed/NCBI

[4] Google Scholar

[ref2] 2. Hennein SS. Smallpox and Its Eradication. JAMA J Am Med Assoc. 1988;260: 559.
View Article
Google Scholar

[6] View Article

[7] Google Scholar

[ref3] 3. Cori A, Boëlle PY, Thomas G, Leung GM, Valleron AJ. Temporal variability and social heterogeneity in disease transmission: The case of SARS in Hong Kong. PLoS Comput Biol. 2009;5. pmid:19696879
View Article
PubMed/NCBI
Google Scholar

[9] View Article

[10] PubMed/NCBI

[11] Google Scholar

[ref4] 4. China CDC. The Epidemiological Characteristics of an Outbreak of 2019 Novel Coronavirus Diseases (COVID-19)—China, 2020. China CDC Wkly. 2020;2: 113–122. pmid:34594836
View Article
PubMed/NCBI
Google Scholar

[13] View Article

[14] PubMed/NCBI

[15] Google Scholar

[ref5] 5. Sharma A, Ahmad Farouk I, Lal SK. Covid-19: A review on the novel coronavirus disease evolution, transmission, detection, control and prevention. Viruses. 2021;13. pmid:33572857
View Article
PubMed/NCBI
Google Scholar

[17] View Article

[18] PubMed/NCBI

[19] Google Scholar

[ref6] 6. Cori A, Ferguson NM, Fraser C, Cauchemez S. A new framework and software to estimate time-varying reproduction numbers during epidemics. Am J Epidemiol. 2013;178: 1505–1512. pmid:24043437
View Article
PubMed/NCBI
Google Scholar

[21] View Article

[22] PubMed/NCBI

[23] Google Scholar

[ref7] 7. White LF, Pagano M. A likelihood-based method for real-time estimation of the serial interval and reproductive number of an epidemic. Stat Med. 2008;27: 2999–3016. pmid:18058829
View Article
PubMed/NCBI
Google Scholar

[25] View Article

[26] PubMed/NCBI

[27] Google Scholar

[ref8] 8. Wallinga J, Teunis P. Different Epidemic Curves for Severe Acute Respiratory Syndrome Reveal. Am J Epidemiol. 2004;160: 509–516. pmid:15353409
View Article
PubMed/NCBI
Google Scholar

[29] View Article

[30] PubMed/NCBI

[31] Google Scholar

[ref9] 9. Fine PEM. The Interval between Successive Cases of an Infectious Disease. Am J Epidemiol. 2003;158: 1039–1047. pmid:14630599
View Article
PubMed/NCBI
Google Scholar

[33] View Article

[34] PubMed/NCBI

[35] Google Scholar

[ref10] 10. Svensson Å. A note on generation times in epidemic models. Math Biosci. 2007;208: 300–311. pmid:17174352
View Article
PubMed/NCBI
Google Scholar

[37] View Article

[38] PubMed/NCBI

[39] Google Scholar

[ref11] 11. Nishiura H, Chowell G. The effective reproduction number as a prelude to statistical estimation of time-dependent epidemic trends. Math Stat Estim Approaches Epidemiol. 2009; 103–121.
View Article
Google Scholar

[41] View Article

[42] Google Scholar

[ref12] 12. Fraser C. Estimating individual and household reproduction numbers in an emerging epidemic. PLoS One. 2007;2. pmid:17712406
View Article
PubMed/NCBI
Google Scholar

[44] View Article

[45] PubMed/NCBI

[46] Google Scholar

[ref13] 13. Thompson RN, Stockwin JE, Van Gaalen RD, Polonsky JA, Kamvar ZN. Improved inference of time-varying reproduction numbers during infectious disease outbreaks. Epidemics. 2019;29: 1–11. pmid:31624039
View Article
PubMed/NCBI
Google Scholar

[48] View Article

[49] PubMed/NCBI

[50] Google Scholar

[ref14] 14. Metcalf CJE, Lessler J. Opportunities and challenges in modeling emerging infectious diseases. Science (80-). 2017;357: 149–152. pmid:28706037
View Article
PubMed/NCBI
Google Scholar

[52] View Article

[53] PubMed/NCBI

[54] Google Scholar

[ref15] 15. Wang K, Zhao S, Li H, Song Y, Wang L, Wang MH, et al. Real-time estimation of the reproduction number of the novel coronavirus disease (COVID-19) in China in 2020 based on incidence data. Ann Transl Med. 2020;8: 689–689. pmid:32617309
View Article
PubMed/NCBI
Google Scholar

[56] View Article

[57] PubMed/NCBI

[58] Google Scholar

[ref16] 16. Dai C, Yang J, Wang K. Evaluation of prevention and control interventions and its impact on the epidemic of coronavirus disease 2019 in Chongqing and Guizhou Provinces. Math Biosci Eng. 2020;17: 2781–2791. pmid:32987495
View Article
PubMed/NCBI
Google Scholar

[60] View Article

[61] PubMed/NCBI

[62] Google Scholar

[ref17] 17. Dai C, Wang Z, Wang W, Li Y, Wang K. Epidemics and underlying factors of multiple-peak pattern on hand, foot and mouth disease in Wenzhou, China. Math Biosci Eng. 2019;16: 2168–2188. pmid:31137205
View Article
PubMed/NCBI
Google Scholar

[64] View Article

[65] PubMed/NCBI

[66] Google Scholar

[ref18] 18. Vihola M. Robust adaptive Metropolis algorithm with coerced acceptance rate. Stat Comput. 2012;22: 997–1008.
View Article
Google Scholar

[68] View Article

[69] Google Scholar

[ref19] 19. Vink MA, Bootsma MCJ, Wallinga J. Serial intervals of respiratory infectious diseases: A systematic review and analysis. Am J Epidemiol. 2014;180: 865–875. pmid:25294601
View Article
PubMed/NCBI
Google Scholar

[71] View Article

[72] PubMed/NCBI

[73] Google Scholar

[ref20] 20. Nishiura H, Linton NM, Akhmetzhanov AR. Serial interval of novel coronavirus (COVID-19) infections. Int J Infect Dis. 2020;93: 284–286. pmid:32145466
View Article
PubMed/NCBI
Google Scholar

[75] View Article

[76] PubMed/NCBI

[77] Google Scholar

[ref21] 21. Akima H. A New Method of Interpolation and Smooth Curve Fitting Based on Local Procedures. J ACM. 1970;17: 589–602.
View Article
Google Scholar

[79] View Article

[80] Google Scholar

[ref22] 22. Dai C, Wang K. Adaptive Weighted Neighbors Method for Sensitivity Analysis. Interdiscip Sci Comput Life Sci. 2022. pmid:35426544
View Article
PubMed/NCBI
Google Scholar

[82] View Article

[83] PubMed/NCBI

[84] Google Scholar

[ref23] 23. Schwarz G. Estimating the Dimension of a Model. Ann Stat. 1978;6: 461–464.
View Article
Google Scholar

[86] View Article

[87] Google Scholar

[ref24] 24. Coleman TF, Li Y. An interior trust region approach for nonlinear minimization subject to bounds. SIAM J Optim. 1996;6: 418–445.
View Article
Google Scholar

[89] View Article

[90] Google Scholar

[ref25] 25. Ali ST, Wang L, Lau EHY, Xu XK, Du Z, Wu Y, et al. Serial interval of SARS-CoV-2 was shortened over time by nonpharmaceutical interventions. Science (80-). 2020;369: 1106–1109. pmid:32694200
View Article
PubMed/NCBI
Google Scholar

[92] View Article

[93] PubMed/NCBI

[94] Google Scholar

[ref26] 26. Leung K, Wu JT, Leung GM. Real-time tracking and prediction of COVID-19 infection using digital proxies of population mobility and mixing. Nat Commun. 2021;12: 1–8. pmid:33686075
View Article
PubMed/NCBI
Google Scholar

[96] View Article

[97] PubMed/NCBI

[98] Google Scholar

[ref27] 27. Chang LY, Tsao KC, Hsia SH, Shih SR, Huang CG, Chan WK, et al. Transmission and Clinical Features of Enterovirus 71 Infections in Household Contacts in Taiwan. J Am Med Assoc. 2004;291: 222–227. pmid:14722149
View Article
PubMed/NCBI
Google Scholar

[100] View Article

[101] PubMed/NCBI

[102] Google Scholar

[ref28] 28. Fraser C, Cummings DAT, Burke DS, Ferguson NM, Cauchemez S, Riley S, et al. Strategies for containing an emerging influenza pandemic in Southeast Asia. Nature. 2005;437: 209–214. Available from: http://www.ncbi.nlm.nih.gov/pubmed/16079797. pmid:16079797
View Article
PubMed/NCBI
Google Scholar

[104] View Article

[105] PubMed/NCBI

[106] Google Scholar

[ref29] 29. Riley S, Ferguson NM. Smallpox transmission and control: Spatial dynamics in Great Britain. Proc Natl Acad Sci U S A. 2006;103: 12637–12642. pmid:16894173
View Article
PubMed/NCBI
Google Scholar

[108] View Article

[109] PubMed/NCBI

[110] Google Scholar

[ref30] 30. Chelsky M, Angulo JJ. Analysis of the incidence of variola minor in an outbreak by use of a mathematical model. Am J Trop Med Hyg. 1977;26: 152–158. pmid:842777
View Article
PubMed/NCBI
Google Scholar

[112] View Article

[113] PubMed/NCBI

[114] Google Scholar

[ref31] 31. Lipsitch M, Cohen T, Cooper B, Robins JM, Ma S, James L, et al. Transmission dynamics and control of severe acute respiratory syndrome. Science (80-). 2003;300: 1966–1970. pmid:12766207
View Article
PubMed/NCBI
Google Scholar

[116] View Article

[117] PubMed/NCBI

[118] Google Scholar

[ref32] 32. You C, Deng Y, Hu W, Sun J, Lin Q, Zhou F, et al. Estimation of the Time-Varying Reproduction Number of COVID-19 Outbreak in China. SSRN Electron J. 2020. pmid:32460229
View Article
PubMed/NCBI
Google Scholar

[120] View Article

[121] PubMed/NCBI

[122] Google Scholar

[ref33] 33. Griffin J, Casey M, Collins Á, Hunt K, McEvoy D, Byrne A, et al. Rapid review of available evidence on the serial interval and generation time of COVID-19. BMJ Open. 2020;10. pmid:33234640
View Article
PubMed/NCBI
Google Scholar

[124] View Article

[125] PubMed/NCBI

[126] Google Scholar

[ref34] 34. Koh WM, Bogich T, Siegel K, Jin J, Chong EY, Tan CY, et al. The Epidemiology of Hand, Foot and Mouth Disease in Asia. Pediatr Infect Dis J. 2016;35: e285–e300. pmid:27273688
View Article
PubMed/NCBI
Google Scholar

[128] View Article

[129] PubMed/NCBI

[130] Google Scholar

[ref35] 35. Park SW, Sun K, Champredon D, Li M, Bolker BM, Earn DJD, et al. Forward-looking serial intervals correctly link epidemic growth to reproduction numbers. Proc Natl Acad Sci U S A. 2020;118. pmid:33361331
View Article
PubMed/NCBI
Google Scholar

[132] View Article

[133] PubMed/NCBI

[134] Google Scholar

[ref36] 36. Gostic KM, McGough L, Baskerville EB, Abbott S, Joshi K, Tedijanto C, et al. Practical considerations for measuring the effective reproductive number, Rt. PLoS Comput Biol. 2020;16. pmid:33301457
View Article
PubMed/NCBI
Google Scholar

[136] View Article

[137] PubMed/NCBI

[138] Google Scholar

[ref37] 37. Lipsitch M, Donnelly CA, Fraser C, Blake IM, Cori A, Dorigatti I, et al. Potential biases in estimating absolute and relative case-fatality risks during outbreaks. PLoS Negl Trop Dis. 2015;9: 1–16. pmid:26181387
View Article
PubMed/NCBI
Google Scholar

[140] View Article

[141] PubMed/NCBI

[142] Google Scholar

[ref38] 38. De Salazar PM, Lu F, Hay JA, Gómez-Barroso D, Fernández-Navarro P, Martínez E V., et al. Near real-time surveillance of the SARS-CoV-2 epidemic with incomplete data. PLoS Comput Biol. 2022;18. pmid:35358171
View Article
PubMed/NCBI
Google Scholar

[144] View Article

[145] PubMed/NCBI

[146] Google Scholar

[ref39] 39. Lehtinen S, Ashcroft P, Bonhoeffer S. On the relationship between serial interval, infectiousness profile and generation time: On the relationship between serial interval, infectiousness profile and generation time. J R Soc Interface. 2021;18. pmid:33402022
View Article
PubMed/NCBI
Google Scholar

[148] View Article

[149] PubMed/NCBI

[150] Google Scholar

Figures

Abstract

Author summary

1. Introduction

2. Materials and methods

2.1. Statistical method

2.1.1. Likelihood function of the instantaneous reproductive number and the serial interval.

2.1.2. Parameterized serial interval distribution by known distributions.

2.1.3. The instantaneous reproductive number was modeled parametrically as a function of time using the interpolation method.

2.1.4. BIC and the trust-region algorithm were employed to maximize the likelihood function and tune undetermined parameters.

2.1.5. The MCMC method was used to generate the distribution of the instantaneous reproductive number and the serial interval.

2.2. Simulation study

2.2.1. Scenario one.

2.2.2. Scenario two.

2.2.3. Quantitative validation measures.

2.3. Application to historical epidemics

2.4. Statistical analysis

3. Results

3.1. Simulated data set in scenario one

3.2. Simulated data set in scenario two

3.3. Historical epidemic data sets

3.3.1. Pandemic influenza for the Boonah ship, 1918.

3.3.2. Pandemic influenza in Cumberland, 1918.

3.3.3. Smallpox in Kosovo, 1972.

3.3.4. SARS in Hong Kong, 2003.

3.3.5. COVID-19 in Hunan, 2020.

3.3.6. COVID-19 in Chongqing, 2020.

3.3.7. HFMD in Wenzhou, 2010–2011.

4. Discussion

5. Conclusions

Supporting information

S1 Text. Additional information on model specification.

S2 Text. Comparative methods.

S1 Table. Prior information of parameters for MCMC method based on MATLAB toolbox.

S2 Table. Comparison of standard deviation of serial interval between White et al and our method.

S1 Fig. MCMC results of each parameter.

S2 Fig. Comparison between the results of BIC and AIC. ΔRt.

S3 Fig. The performance of our method when varying different parameters.

S4 Fig. The effects of changing serial interval on the results.

S5 Fig. The effects of time length on the results.

S1 Codes. MATLAB main model file.

References

S2 Fig. Comparison between the results of BIC and AIC. ΔR_t.