Near real-time surveillance of the SARS-CoV-2 epidemic with incomplete data

Pablo M. De Salazar, Fred Lu, James A. Hay, Diana Gomez-Barroso, Pablo Fernández-Navarro, Elena V. Martínez, Jenaro Astray-Mochales, Rocío Amillategui, Ana García-Fulgueiras, Maria D. Chirlaque, Alonso Sánchez-Migallón, Amparo Larrauri, María J. Sierra, Marc Lipsitch, Fernando Simón, Mauricio Santillana, Miguel A. HernánPLOS Computational Biology 2022 · DOI · OpenReview

TL;DR

This paper introduces a novel three-step statistical framework to accurately track infectious disease outbreaks in near real-time, even with incomplete and delayed data. By imputing missing symptom onset dates and using a sophisticated nowcasting method, the approach provides a more reliable estimate of epidemic curves and the time-varying reproduction number (Rt), crucial for timely public health interventions.

What Problem Does This Paper Address?

Monitoring infectious disease outbreaks in real-time is critical for effective public health responses. However, real-world surveillance data often suffers from two major issues: missing information on the exact date of symptom onset (DOS) and retrospective revisions of historical data, including right censoring (where data collection stops before all events have occurred). These data limitations make it challenging to accurately construct epidemic curves – graphs that show the number of infected individuals over time – and to estimate key epidemiological parameters like the reproduction number (Rt) in a timely manner. Traditional surveillance methods often struggle with these data imperfections, leading to delayed or inaccurate insights into an epidemic's trajectory.

Methods

The authors propose a three-step approach to overcome these data challenges:

  1. Imputation of Missing Data: They impute missing dates of symptom onset (DOS) by drawing from a dynamically estimated "backward" reporting delay distribution. This distribution is conditional on the date of report (DOR) and accounts for variations due to the day of the week and short-term trends. This is a more sophisticated approach than simply using an overall average delay.
  2. Nowcasting the Epidemic Curve: Using the imputed DOS data, they employ the NobBS (Nowcasting by Bayesian Smoothing) software package. NobBS uses historical reporting delay data and the intrinsic autocorrelation of the transmission process to estimate the number of cases that have occurred but have not yet been reported. This process adjusts for reporting delays and right censoring, providing a more accurate "nowcast" of the epidemic curve.
  3. Estimation of Time-Varying Reproduction Number (Rt): With the nowcasted epidemic curve, they estimate the time-varying reproduction number (Rt) using two established methods (Wallinga and Teunis, and Cori et al.). Rt represents the average number of secondary infections caused by a single infected individual at a given time.

Datasets

The study utilizes publicly available, anonymized, line-listed data from the Spanish System for Surveillance at the National Center of Epidemiology (RENAVE) for two regions in Spain: Madrid and Murcia. The data covers the early stages of the SARS-CoV-2 outbreak between March and April 2020. The authors also provide code and anonymized data for reproducibility.

Key Insights and Ideas

  • The Power of Imputation and Nowcasting: The study demonstrates that even with significant missingness in symptom onset dates, a robust imputation strategy combined with advanced nowcasting techniques can yield epidemic curves that are far more accurate and timely than those derived from raw, unadjusted data.
  • Earlier Detection of Intervention Impact: The nowcasted curves allowed for the identification of the impact of the country-wide lockdown in Spain approximately a week earlier than when using reported cases by date of report. This highlights the critical role of timely data in assessing the effectiveness of public health interventions.
  • Sensitivity to Assumptions: While the framework is robust, the study emphasizes that its accuracy is dependent on the assumption that missing data is "missing at random" and that reporting delays can be adequately modeled. Significant deviations from these assumptions, or very early stages of an outbreak with limited data, can impact the reliability of the estimates.
  • Rt Estimation Benefits from Nowcasting: Estimating Rt from nowcasted epidemic curves provides a smoother and more accurate picture of transmission dynamics compared to using raw case counts by date of report, which are heavily influenced by reporting delays.

Key Results

  • The proportion of missing symptom onset dates varied significantly by region and epidemic phase, ranging from 11% to 53%.
  • Nowcasted epidemic curves were found to be more accurate and consistent with later-obtained complete data compared to unadjusted curves based on either date of report or date of symptom onset.
  • The nowcast approach identified the impact of the lockdown in Spain approximately 6-9 days earlier than traditional methods.
  • Estimates of Rt derived from nowcasted curves showed an earlier reduction and were less noisy than those from raw case counts by date of report.

Limitations and Future Work

The authors acknowledge several limitations:

  • Assumption of "Missing at Random": The validity of the approach relies on the assumption that missing symptom onset dates are random. If there's a systematic reason for missingness (e.g., severe cases are less likely to have their DOS recorded), it could bias results.
  • Data Availability for Training: The approach underperforms when there's insufficient data to train the nowcasting algorithm, particularly in the very early stages of an outbreak or when reporting patterns are highly variable (e.g., weekend reporting dips).
  • Constant Underreporting: The method assumes a relatively constant degree of underreporting. Significant changes in reporting rates over time can bias Rt estimates.
  • Date of Infection vs. Symptom Onset: The authors suggest that reconstructing the epidemic curve by date of infection, rather than symptom onset, could further improve accuracy, though this would require more complex modeling due to the greater difficulty in estimating infection dates.

Future work could involve developing more sophisticated models to account for reporting dynamics and exploring methods to reconstruct curves based on infection dates. The authors also emphasize the importance of understanding the limitations of any surveillance framework for proper interpretation and decision-making.

Summary updated

Abstract

When responding to infectious disease outbreaks, rapid and accurate estimation of the epidemic trajectory is critical. However, two common data collection problems affect the reliability of the epidemiological data in real time: missing information on the time of first symptoms, and retrospective revision of historical information, including right censoring. Here, we propose an approach to construct epidemic curves in near real time that addresses these two challenges by 1) imputation of dates of symptom onset for reported cases using a dynamically-estimated “backward” reporting delay conditional distribution, and 2) adjustment for right censoring using the NobBS software package to nowcast cases by date of symptom onset. This process allows us to obtain an approximation of the time-varying reproduction number (Rt) in real time. We apply this approach to characterize the early SARS-CoV-2 outbreak in two Spanish regions between March and April 2020. We evaluate how these real-time estimates compare with more complete epidemiological data that became available later. We explore the impact of the different assumptions on the estimates, and compare our estimates with those obtained from commonly used surveillance approaches. Our framework can help improve accuracy, quantify uncertainty, and evaluate frequently unstated assumptions when recovering the epidemic curves from limited data obtained from public health systems in other locations.
BibTeX
@article{salazar2022near,
  author = {Pablo M. De Salazar and Fred Lu and James A. Hay and Diana Gomez-Barroso and Pablo Fernández-Navarro and Elena V. Martínez and Jenaro Astray-Mochales and Rocío Amillategui and Ana García-Fulgueiras and Maria D. Chirlaque and Alonso Sánchez-Migallón and Amparo Larrauri and María J. Sierra and Marc Lipsitch and Fernando Simón and Mauricio Santillana and Miguel A. Hernán},
  title = {{Near real-time surveillance of the SARS-CoV-2 epidemic with incomplete data}},
  journal = {PLOS Computational Biology},
  year = {2022},
  doi = {10.1371/journal.pcbi.1009964},
}
Figure