ABSTRACT
Objective In 2020, SARS-CoV-2 impacted Georgia, USA. Georgia announced state-wide shelter-in-place on April 2 (implemented the next day) and announced the partial lifting of restrictions on April 27. Time-varying reproduction number estimates might vary depending on methodology.
Methods Daily incidence of confirmed COVID-19 cases by reporting date, March 2-June 14, 2020, in Georgia, Metro Atlanta, and Dougherty County were analyzed. We estimated and compared the COVID-19 time-varying reproduction number, Rt, by R package EpiEstim (with a window of 1 and 2 weeks) and the renewal equation tied to the generalized growth model (GGM).
Results The EpiEstim Rt estimates are more sensitive to fluctuation in incident cases than the GGM estimates. The median EpiEstim Rt estimate (1-week window) in Georgia dropped from between 2 and 4 in mid-March, to <2 in late March and early April. The median Rt dropped <1 for the first time on April 15. Rt fluctuates around 1 from mid-April to June 14. The Rt estimate by GGM dropped below 2 on March 7 and stabilized to the range of 1 to 1.1 by March 31.
Conclusions GGM estimated Rt as an overall trend over a 3-month period, while EpiEstim is sensitive to daily variation in incidence.
INTRODUCTION
In 2020, the pandemic of coronavirus disease 2019 (COVID-19), caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), impacted the state of Georgia (GA) as in other jurisdictions within the United States. Georgia ranked twelve out of fifty states and ten territories of the United States in terms of COVID-19 confirmed cases as of April 18, 2020.1 Metro Atlanta counties have been the hardest hit by the virus with thousands of confirmed cases cumulatively: 6,350 in Fulton, 5,322 in Dekalb and 7,463 in Gwinnett as of June 26, 2020.2 Dougherty County, with Albany as the county seat, was a COVID-19 hotspot in southeast GA and reported a large number of cases (as of June 26, 2020: cumulative number = 1,873, with incidence rate, IR=2,083 per 100,000 individuals).2 In GA, every county government had the power to impose preventive measures to reduce viral transmission as they see fit, before the state imposed a state-wide emergency that overrode autonomy of county governments (Table 1). On March 23, 2020, the GA State Government issued an executive order requesting citizens with underlying conditions and those with a COVID-19 diagnosis to shelter-in-place.3 Certain businesses were to remain closed and no more than 10 individuals could gather in a location without maintaining a distance of at least six feet. The order also called for restaurants to offer only curbside pick-ups or deliveries.3 On April 2, 2020, a state-wide shelter-in-place ordinance was enacted by the Governor allowing only essential services to operate (implemented on April 3).4 The GA State Government announced on April 27, 2020 during a press conference that services, such as beauty salons, barber shops, stores, and restaurants, can reopen if they follow pertinent social distancing measures specified by the state.5 As the COVID-19 epidemic in GA continues, it is important to quantify the epidemiologic characteristics of COVID-19 so that we may formulate policies and implement interventions to minimize transmission and mortality.
To characterize the transmission potential of an epidemic, it is necessary to calculate the reproduction number based on the trajectory of the incidence curve.6 The basic reproduction number, R0, is the average number of secondary cases that one primary case can generate in a completely susceptible population in the absence of behavioral changes or public health interventions.6 The estimated values of the R0 for SARS-CoV-2 vary across geographic locations. An early study of the epidemic in Wuhan reported a R0 of 2.2 (assuming serial interval of 7.5).7 A more recent study of the epidemic in China, adjusted for the changing case definition, estimated R0 of 1.8-2.0 (assuming serial interval of 7.5) or 1.4-1.5 (assuming serial interval of 4.7).8 Our analysis of confirmed COVID-19 cases in Iran estimated the mean R0 as 3.5 or 4.4, depending on the statistical method (assuming serial interval of 4.4).9
In contrast, the time-varying (or time-dependent) reproduction number, Rt, is a time-dependent estimate of the secondary cases that arise from one case at time t, when depletion of the susceptible population, behavioral changes, and measures to control transmission of disease have taken place.10,11 As with R0, if Rt>1, it indicates there is sustainable transmission in the population. When Rt<1, disease transmission cannot be sustained, and it is used as an indication of the effectiveness of infection control measures.6,10
Various statistical methods have been proposed to estimate Rt. Their strengths and weaknesses have been recently assessed by researchers who compared the performance of different methods using synthetic epidemic data,12,13 and observed COVID-19 incidence data.14
An oft-used method is the instantaneous reproduction number method implemented in the R package EpiEstim.10,11 This Bayesian method provides an estimate of the average Rt over a short time window specified by the user (e.g., a 7-day window that ends at time t). It treats the fluctuation in incidence data as signals of an increasing or decreasing reproduction number. This method has been used to estimate COVID-19 Rt in jurisdictions such as mainland China,15 Hong Kong,16 Iran,17 South Korea,18 Italy,19 Nigeria,20 and Switzerland.21
Another method is based on the renewal equation,22,23 which characterizes the incidence curve using the generalized growth model (GGM).24 This method characterizes the mean trajectory of Rt over a longer time period assuming the intermittent fluctuation of incidence data as an underlying noise structure that follows a Poisson or negative binomial distribution. This method has been used to estimate COVID-19 Rt in countries such as, Iran,9 South Korea,25 and Peru.26
This study aimed to use these two statistical methods and compared their Rt estimates for COVID-19 in GA, Dougherty County, and Metro Atlanta counties, analyzing historical data from March 2 through June 14, 2020, as the state incrementally implemented and then relaxed social distancing interventions (Table 1).
METHODS
This study uses historical data from the COVID-19 pandemic, March 2 — June 14, 2020, in the state of GA, Dougherty County, and all Metro Atlanta counties (Supplementary Materials Tables S1-S12). Metro Atlanta is defined by the United States Office of Management and Budget as the “Atlanta-Sandy Springs-Alpharetta, Georgia Metropolitan Statistical Area.”27 The list of Metro Atlanta counties is provided in Table 2.
Data acquisition
We downloaded the cumulative data of confirmed cases from March 2 — June 14, 2020, for the entire state of GA and its counties from the New York Times (NYT) GitHub data repository.28 Every jurisdiction in our study has a start date corresponding to the first reported case for the area, according to NYT (Table 2). The first case in GA was reported on March 2, 2020.28 Our cutoff point for all jurisdictions was June 14, 2020, 7 weeks after businesses in the state reopened on April 27, 2020, following social distancing and prevention measures (Table 1).5
We verified the numbers with official statistical reports from the Georgia Department of Public Health (GA-DPH).2 If any inconsistencies were found, the numbers from the GA-DPH were used as the standard. To estimate Rt, we transformed the cumulative numbers into daily number of cases reported to obtain daily SARS-CoV-2 incidence (Supplementary Materials Table S1-S12). We also searched the local government pages to verify if any control measures were established. Such information is presented in Table 1.
Statistical analyses
The instantaneous reproduction number (EpiEstim package)
The Rt estimated by the R package EpiEstim is also known as an instantaneous reproduction number.11 This measure is defined by Cori et al. as the ratio between It, the number of incident cases at time t, to Λt, the total infectiousness of all the infected individuals at time t. The latter is represented mathematically as , i.e., the sum of infection incidence up to time t-1, weighted by an infectivity function ws, which is a probability distribution describing the average infectiousness profile after infection and is usually represented by the serial interval distribution. Thus, the number of new cases at time t is Poisson-distributed with a mean of . Conditional on the number of new cases in previous time points, I0, …, It-1, and given the reproduction number Rt, the likelihood of the incidence It is as follows: where the total infectiousness of infected individuals at time t,
In other words, . This can be interpreted as the average number of secondary cases that an infectious individual will infect at time t, if conditions remain the same at time t. As this formulation may result in a highly variable Rt estimate over time, the instantaneous reproduction number method as implemented in EpiEstim assumes that the Rt is constant over a specified time window of size τ ending at time t. This will allow the estimate to be less variable and more precise. Given that transmissibility is assumed constant over a time period, from (t-τ+1) to t, and is denoted by a reproduction number, Rt,τ, the likelihood of the incidence during the time period, from I(t-τ+1) to It, conditional on incidence prior to the time period, from I0, to I(t-τ), is as follows:11
Here, the time window used was a period of 7 days (main analysis) and 14 days (sensitivity analysis) respectively.
Using a Bayesian framework with a Gamma-distributed prior for Rt,τ, Cori et al. derived an analytical expression of the posterior distribution of Rt and thus estimated its median, the variance, and the 95% credible interval.11 In this paper, the data was analyzed using EpiEstim version 2.2-3,10,11 implemented in R Version 1.2.1335 Macintosh (R Core Team, Vienna, Austria).
Generalized growth model
The GGM characterizes the COVID-19 daily incidence using the growth rate, r, and a scaling factor, p, of the growth rate as its parameters. The GGM model allows epidemiologists to measure multiple epidemic profiles, including constant incidence, sub-exponential growth, and exponential growth.29 The GGM is represented by dC(t)/dt = C′(t) = rC(t)p, where C′(t) corresponds to the incidence curve over time; C(t) describes the cumulative incidence over time t; r is the growth rate (1/time) and p is the scaling parameter fitting sub-exponential growth.24,29 The scaling parameter of GGM, p, represents exponential growth if p = 1, sub-exponential growth or polynomial growth if 0 < p < 1, and constant growth if p = 0. To estimate the Rt at a specific time, we linked the GGM model fit to the renewal equation that estimates the temporal changes in the effective reproduction number using a discretized probability distribution (ρi) of the serial interval, namely, . In the renewal equation we denote the total incidence at calendar time ti by Ii. Here the denominator represents the total number of cases contributing to the new reported cases at time ti. We simulated 300 curves to estimate uncertainty boundaries.24,29 Because the time series of incident cases displays overdispersion, we assumed a negative binomial error structure underlying the data points (Tables 2).24 The choice of the scaling parameter of the negative binomial distribution, i.e. the ratio of variance / mean, was manually explored (from 1, 2, 5 and then in the increments of 5) to achieve the lowest variance possible with at least 85% of the data points covered by the 95% prediction interval. The serial interval is assumed to follow a gamma distribution with a mean = 4.60 days, and standard deviation = 5.55 days.30
To estimate Rt, we analyzed the incidence data during March 2 — June 14, 2020. Statistical significance was defined a priori at α = 0.05. Our GGM analyses were conducted in MATLAB R2019b (MathWorks, Natick, MA).
Creation of figures comparing EpiEstim and GGM results
The processing of EpiEstim and GGM outputs and the creation of the figures comparing them were conducted using R version 3.6.2 (R Core Team, Vienna, Austria).
Ethics
The Georgia Southern University Institutional Review Board made a non–human subjects determination for this project (H20364), under the G8 exemption category.
RESULTS
Community transmission of SARS-CoV-2 remained ongoing in GA based on incidence data by date of report from March 2 – June 14, 2020. As of June 14, 2020, The median EpiEstim Rt estimates were 1.14 (95% credible interval, CrI, 1.11, 1.17) for a 1-week-window and 1.07 (95% CrI, 1.05, 1.09) for a 2-week-window respectively. In Metro Atlanta, the transmission may have been under control, with Rt estimates being 1.02 (95% CrI, 0.98, 1.06) for a 1-week-window and 0.98 (95% CrI, 0.95, 1.01) for a 2-week-window. Likewise, in Dougherty County, the transmission with Rt estimates being 1.03 (95% CrI, 0.72, 1.41) for a 1-week-window and 0.93 (95% CrI, 0.72, 1.17) for a 2-week-window. The mean GGM Rt estimate was 1.00 (95% confidence interval, CI, 1.00, 1.00) for GA, Metro Atlanta, and Dougherty County (Table 2).
The EpiEstim Rt estimates are more sensitive to fluctuation in incident cases than the GGM estimates. As social distancing measures unfolded and then relaxed in GA during our study period (March 2 – June 14, 2020), the median EpiEstim Rt estimate (1-week window) in Georgia dropped from between 2 and 4 in mid-March, to <2 in late March and early April. The median Rt dropped <1 for the first time on April 15. Rt fluctuated around 1 from mid-April to June 14. The 2-week window EpiEstim Rt estimate largely followed the trend of the 1-week window estimate, even though the 2-week estimate is smoother than the 1-week estimate when the Rt estimate fluctuates around 1 from late April to mid-June. The median Rt estimate by GGM dropped below 2 on March 7 and was stabilized to the range of 1 to 1.1 from March 31 through June 14 (Figure 1).
Regarding Metro Atlanta (Figure 2), EpiEstim Rt estimates (both 1-week and 2-week window) fluctuated above 1.5 in March and gradually decreased to around 1 in April and continued to fluctuate above and below 1. In contrast, the GGM Rt estimate dropped below 1.5 on March 9, and below 1.1 on March 29, and it stabilized between 1 and 1.1 through June 14. The Rt estimates for each of the Metro Atlanta counties are presented in Figures S1 to S29 in the Supplementary Materials.
Regarding Dougherty County (Figure 3), we observed a speedy decline in EpiEstim Rt estimates (both 1-week and 2-week window) from around 2 in late March to a low of 0.32 on April 26. This finding was primarily driven by the early epidemic observed in Dougherty County, which were large clusters of cases infected via two funerals that happened to be superspreading events.31 The median GGM Rt estimate dropped below 1.5 on March 20, below 1.1 on March 28, and it stabilized between 1 and 1.1, through June 14.
DISCUSSION
Community transmission of SARS-CoV-2 remained ongoing in GA as of June 14, 2020 (i.e., 10½ weeks after statewide shelter-in-place order was first imposed on April 2 and 7 weeks after businesses reopened under new social distancing guidance) (Table 1). The 95% CrI of both the 1-week and 2-week window EpiEstim Rt estimates were >1.
On April 27, GA reopened some sectors of the economy, with specific guidelines pertinent to social distancing.5 As the economy reopened and increased unprotected social mixing ensued, an increase in daily number of new confirmed cases was observed in June as SARS-CoV-2 transmission continued unabated.2 Our study documented the decrease in Rt following social distancing interventions in GA and provided further evidence that social distancing measures remained important to keep COVID-19 under control.
Furthermore, many residents in both rural and urban GA are medically vulnerable. A recent analysis by The Surgo Foundation estimated the COVID-19 community vulnerability index (CCVI) for Dougherty County, by combining epidemiologic risk factors for infection and sociodemographic factors, at very high levels (CCVI = 0.87) when compared with counties in Metro Atlanta (Fulton county’s CCVI = 0.42).32 Although the EpiEstim Rt estimates for Dougherty County were <1 from April 14 to April 30, Rt rebounded since late April, and it has been fluctuating with 95% CrI including 1 most of the time from early May to mid-June (Figure 3). This corresponded to the partial reopening of the GA economy. The relaxation of social distancing measures should be implemented with an abundance of caution due to the population’s vulnerability. Another important factor for consideration is access to healthcare and surge capacity in hospitals, especially in rural GA. The healthcare system in Dougherty County was heavily impacted by the surge of COVID-19 patients driven by superspreading events.31
While economic factors are legitimate concerns for decisionmakers, reopening the economy too soon led to the resurgence of COVID-19 cases in GA as observed in June. Further research into the spatiotemporal variation of SARS-CoV-2 transmission potential and its association with economic and medical vulnerability will shed light on the disease and economic burden of COVID-19 in GA.
Our study compared the Rt estimate obtained using two methods: the instantaneous reproduction number method implemented in the R package EpiEstim 10,11 and the renewal equation linked to the GGM.24 One major difference between the two methods is the underlying assumption about the fluctuation in the incident case count number. The EpiEstim estimate is sensitive to fluctuation in daily incident case counts as the instantaneous reproduction number method treats such changes as meaningful signals reflecting genuine increases or decreases in transmission potential. To the contrary, the GGM treats the fluctuation in daily incident case counts as noise and represent such noise with an underlying error structure using either a Poisson or a negative binomial distribution. Each method has its strengths and weaknesses. The instantaneous reproduction number method in the EpiEstim package should be used if the purpose is to identify time-dependent changes in the Rt estimate that reflects the implementation or relaxation of social distancing measures over time. GGM should be used if the purpose is to identify an overall trend of the epidemic trajectory and if the data is noisy, and we have reasons to believe that the fluctuations may reflect artifacts in the surveillance system.
Regarding the time window chosen for EpiEstim, we used a window of 7 days in our main analysis and 14 days in our sensitivity analysis. We did not use a window of <7 days, because a weekend effect was observed in the data, i.e., the daily number of cases reported in the weekend was consistently smaller than those in the weekdays before or after the weekend. The 2-week window EpiEstim estimate lagged behind the 1-week window EpiEstim estimate when the Rt estimate fluctuates around 1. This result can be explained because the 2-week estimate was an average of 14 days and did not reflect the immediate change in transmission potential.
Limitations
Our study is limited by several factors. First, we used the NYT dataset; data was recorded by reported date and not by day of symptom onset. Given the time difference between the transmission events and subsequent symptom onset, testing, and case reporting, the Rt estimates should reflect transmission that had happened several days ago (the sum of the incubation period and the delay to testing and case reporting). Second, our data does not differentiate between imported and community transmission cases. Third, cases may be underreported due to limited testing capacity, or they may be mild or asymptomatic cases. Fourth, our analysis is right-censored by June 14, 2020. Future studies can further extend the analysis as the pandemic progresses. Fifth, in addition to the two methods utilized here, there are other statistical methods that estimate Rt,12,13 e.g., the case reproduction number method as proposed by Walling and Teunis,33 but they are out-of-scope for this paper.
CONCLUSION
In Spring 2020, social distancing measures reduced the transmission of SARS-CoV-2 in GA. However, the epidemic has not yet been suppressed. Rt was fluctuating around 1 for GA, Metro Atlanta and Dougherty County since GA economy reopened in late April 2020. Government agencies should carefully consider the next steps of their COVID-19 plans for their communities considering ongoing transmission across GA.
Data Availability
All data used in the analysis is presented in the Supplementary Materials.
Online Supplementary Materials
Legends of Supplementary Figures: Results for Counties in Metro Atlanta, GA, USA
Acknowledgment
G.C. received support from NSF grant 1414374 as part of the joint NSF-NIH-USDA Ecology and Evolution of Infectious Diseases program. I.C.-H.F. received salary support from the Centers for Disease Control and Prevention (CDC) (19IPA1908208). This article is not part of I.C.-H.F’s CDC-sponsored projects.
Footnotes
Email addresses: km11200{at}georgiasouthern.edu (K. Muniz-Rodriguez); gchowell{at}gsu.edu (G. Chowell); jschwind{at}georgiasouthern.edu (J. S. Schwind); rf08112{at}georgiasouthern.edu (R. Ford); so01935{at}georgiasouthern.edu (S. K. Ofori); co05814{at}georgiasouthern.edu (C. A. Ogwara); md18742{at}georgiasouthern.edu (M. R. Davies); tj10163{at}georgiasouthern.edu (T. Jacobs); westerpants{at}gmail.com (C.-H. Cheung); lcowan{at}georgiasouthern.edu (L. Cowan); ahansen{at}georgiasouthern.edu (A. Hansen); cfung{at}georgiasouthern.edu (I. C.-H. Fung)
Disclaimer: The opinions expressed in this paper does not necessarily represent the official positions of the CDC or the United States Government.