Estimation and Interactive Visualization of the Time-Varying Reproduction Number Rt and the Time-Delay from Infection to Estimation

Rt plays a key role in the development of the COVID-19 pandemic. The methods used for building an interactive website for the visualization of the time-varying reproduction number Rt and a novel way to visualize the time delay from infection to estimation overlayed with the Rt estimate are described and analyzed with regards to the influence of parameters chosen and compared to published estimates for Austria. Visualizing the time delays enables the interactive exploration of suspected changes in transmission and their effect on the Rt estimate.


Serial Interval
Accuracy of the serial interval estimation plays a key role in the accuracy of the estimation of R t,τ .
Representing the uncertainties contained in the current serial interval estimates [9] was realized using the uncertain_si method in estimate_R. This samples the parameters for the serial intervals Gamma distribution Gamma s from a truncated normal distribution. The median of the R t,τ estimate is only affected to a limited extent by this, but the credible interval will be increased around those median values [7], giving a more accurate representation based on our current beliefs about the serial interval.
Parameters used in the implementation are documented in Table 1.
Standard deviations of the mean serial interval and the standard deviation of the standard deviation were obtained based upon the assumption of normality from the 95% confidence interval CI lower , see Equation 1.
Based upon these parameters multiple serial intervals are explored by estimate_R. [8] Explored serial interval distributions are shown in Figure 3. During each estimation 500 discretized Gamma distributions of the serial interval are explored.  Figure 1: Serial interval distributions explored during R t,τ estimation. Quantiles shown in color scaling (blue), with the outer border marking the 0.95 credible interval (0.025 and 0.975 quantiles). Overlayed with samples from the explored discrete serial interval distributions (yellow).

Sliding Time Window τ
Choosing τ is a trade-off between less noise and less delay. Smaller time windows τ lead to a more rapid detection of changes in R t,τ , but also to more statistical noise. [10] Cori, et al. describe a method for choosing the time window size τ based upon the minimum number of cases included in each time window [10]. Based upon a chosen target coefficient of variation for the R t,τ estimate of at least 0.4 a minium of 7 cases in each time window is required. With this in consideration and the usage of a time window of 7 days in Cori, et al. [7] a time window size τ of 7 days for the interactive visualization was chosen. AGES uses a 13 days time window for their estimation, no explanation is given for this choice and this τ was therefore only used for comparisons but can be chosen by the user on the interactive front-end.
2 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint  The effect a change in τ has on the estimation of R t,τ is illustrated in Figure 2. Comparing the first date where the median of τ has fallen below 1 illustrates the trade-off discussed above. For τ = 7 the condition R t,7 < 1 is first met on the 31st of March 2020, for τ = 13 the condition R t,13 < 1 is only met 4 days later on the 4th of April 2020.
A 7-day window has the added advantage that it always contains exactly 2 weekend days.

Crossvalidation with AGES estimates for Austria
The output of the estimation described in this section was compared to the R t,τ estimates obtained by AGES on the 24th of April 2020 [1].
Two different methods were chosen for the comparison, one where the dataset which allows the closest reproduction of the estimates obtained by AGES was used as shown in figure Figure 3a. In the actual interactive visualization a different dataset is used. The comparison against this dataset is shown in Figure 3b. For a further discussion on the datasets see section 4.
3 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (a) Closest reproduction using the data.gv.at dataset [11] and parametric_si .
(b) Dataset used in actual visualization and using uncertain_si in estimate_R. Figure 3: Comparison between own estimates and R t,τ estimates published by AGES. [12] 3 Estimation of the time delay from infection to inclusion in the R t,τ estimation The total time delay from infection to inclusion in the estimate t inf ection,estimation will be split up into three different time delays. The time delay from infection to onset t inf ection,onset , the incubation period. The time delay from onset to reporting t onset,reporting and the time delay from reporting to inclusion in the estimation t reporting,estimation , which is caused by the R t,τ estimation method of assuming a constant R t within the time window τ .

Time delay from infection to reporting t inf ection,reporting
Both t inf ection,onset and t onset,reporting are assumed to be independent random variables and a Gamma distribution was used for the estimation. Estimates of the Gamma distribution parameters for t inf ection,onset were taken from the supplementary appendix of the study by Zhang, et al. [13] based upon case data from Mainland China, excluding Hubei province. The parameters used are α inf ection,onset = 4.23(SD 1.28) and β inf ection,onset = 0.81(SD 0.24) with a mean of 5.2days and a 95% CI of the mean of (95% CI: 1.5 − 11.3) days.
Onset to reporting Gamma distribution parameters were also taken from the estimates from Mainland China, excluding Hubei [13] from the second period (Jan 28 -Feb 17) were used. Those parameters were obtained from 2079 observations and are α onset,reporting = 3.18 and β onset,reporting = 0.59 giving a mean of 5.3 (95% CI: 1.2 − 13.1) days.
In order to obtain the empirical CDF of t inf ection,reporting random samples were drawn from the Gamma distributions of t inf ection,onset and t onset,reporting using the R-function rgamma A total of 10 7 samples were drawn from each distribution and the resulting emperical distribution function F inf ection,reporting was obtained by applying the ecdf function[14] to the list of t i inf ection,reporting .
4 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Time delay from reporting to inclusion in the R t,τ estimation t reporting,estimation
To derive the time-delay from reporting to estimation t reporting,estimation we need to consider the method used for the R t,τ estimation in detail. The mean of R t,τ given the parameters a and b of the Gamma distribution prior is given by Equations 3 and 4.[10] An initial investigation of the time delay was performed by investigating the propagation of a time pulse through Λ t , the key component of the denominator, as defined in defined in Equation 3. This time pulse was modeled as a dirac δ function at t = 0. The estimation process is discretized such that t ∈ N, therefore, the discrete number of cases for the dirac δ is given in Equation 5.

Simulation of the Estimation Driven Time-Delay using a Compartmental Model
The R t,τ response to a step function in the transmission rate β of a SEIR model [15] has been used for deriving the τ dependent time-delay inherent to the R t,τ estimation method.
Parameters of the model itself are not related to COVID-19 and the model has been defined without vital dynamics.
5 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint S, E , I, and R represent the total number of susceptible, exposed, infected, and removed individuals in a total population N , respectively. β denotes the transmission rate, γ is the removale rate, σ is the rate at which individuals transition from the exposed class to the infectious class. The dynamics of the model are defined in equations 6, 7, 8 and 9. [15] For simulating a change in transmission the transmission rate β has been given a time dependency β(t) with a step function, modeled using a logistic function to keep the model continuously differentiable. The model can change between two different transmission rates β 0 , before t swap and β 1 after t swap . The steepness of the curve, k, was set to 20.
Considering the relationship between β(t) and R 0 of the SEIR model, Equation 11[15], a change in β is proportional to a change in R 0,SEIR (t).
Since the dynamics of the R t,τ response to a step function depends on the serial interval distribution, as can be seen in Equation 3, the actual serial interval distribution, as described in subsection 2.1, was used. For this simulation the parametric_si mode was used using the mean of the serial interval mean and the mean of the serial interval standard deviation as shown in Table 1.
In the SEIR model the average time in the infectious compartment γ −1 was set to 3. The average time in the exposed compartment, σ −1 , was set to 4, the transmission rate before the step β 0 was set at 0.6, the center of the logistic function, in Equation 10, t swap was set to 40 and the transmission rate transitions to β 1 around this point, which was set to 0.8.
For each τ a step response R t,τ,sim is obtained from the simulation. This step response is normalized such that it is at 0 before the time step is applied and that the last element R τ,τ,sim = 1.
The resulting time responses from the simulations, R t,τ,sim , are then applied to the empirical distribution function F inf ection,reporting to obtain the empirical distribution function of the total time delay F inf ection,estimation . Each element of this function is calculated using Equation 12, using the definition R 0,τ,sim = 0.
6 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.  Figure 5: Response of the R t,τ estimation to a step function in the transmission rate β(t), and therefore also in R t of a SEIR model.

7
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

Data Sources
Case data for Austria are obtained from a GitHub repository maintained by the Complexity Science Hub Vienna based upon the official data published by the Sozialministerium. This dataset contains the case numbers for the 3 pm CEST closing time.

Data Correction
Some days in the Austrian states contain a negative number of cases on a given day. Those negative numbers are corrected by adding the negative correction to the previous day and adjusting the day containing the negative number to 0, keeping the total sum of cases constant.

Data Quality Issues
Data published by the Austrian Sozialministerium combines data into bins based upon different dates, depending on the availability of the information for each given case.
[16] Currently cases are assigned to a bin by the date in the following order: Symptom onset, laboratory confirmation, reporting. This could potentially have a big impact on the estimation of R t,τ and also on the time-delay involved. It is currently unknown how many of the data entries binned using which date. Estimates of R t,τ published by AGES [1] therefore differ from the reproduction of their method. Comparing the results of the estimation using this dataset and the AGES estimation suggests that there might be a limited difference between those datasets, see Figure 3a.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted September 22, 2020. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted September 22, 2020. . https://doi.org/10.1101/2020.09.19.20197970 doi: medRxiv preprint