Further Improving Analysis of Date-Based COVID-19 Surveillance Data

In the December 2021 issue of AJPH, Hennessee et al. reminded readers of an important limitation of COVID-19 surveillance data.¹ As they noted, there are many dates that can be captured in a time series, including dates of SARS-CoV-2 infection, symptom onset, test, and report. These dates provide information useful to public health, from calculating measures of contagion to demand on the health care system. We agree with the authors that, epidemiologically, infection date is most relevant.

However, the authors do not do justice to the complexities inherent in analyzing these data and, worse yet, may inadvertently mislead readers regarding best practices for addressing these complexities. Indeed, the authors state “many estimations methods [for infection date] homogenize substantial heterogeneities,” but they disregard this in their demonstration of recovering infection date in Figure 2: “infection dates were estimated as symptom onset dates minus a median incubation period.”^1(p2129) Although this is straightforward to calculate, it ignores the “substantial heterogeneities,” and leads to invalid conclusions.² For those interested in obtaining infection date, we would suggest deconvolution, which has been applied in infectious disease surveillance for decades,³ and has also been implemented for estimating the reproductive number of SARS-CoV-2 during the pandemic.^2,4

There is another issue with these data that must be dealt with before making any inferences about the outbreak: outcome misclassification.^4,5 Such errors will primarily be underreporting (i.e., suboptimal sensitivity of the surveillance program) through asymptomatic infection or those symptomatic and unable or unwilling to test, but there may also be issues with diagnostic accuracy, including both false positives and false negatives. As such, taking epidemic curves at face value is problematic.⁶

Methods are freely available to rigorously address both the timing of infection and flaws in the capture of cases.⁴ Figure 1 is our application of these methods to the Georgia Department of Public Health time series COVID-19 data used by Hennessee et al. and shows the divergence in the two approaches beyond the stochastic error of the 95% uncertainty interval, which may have implications for pandemic management.⁷ Methodologically rigorous adjustment for biases in the surveillance data smoothed out fluctuations and leads to a more definitive interpretation, reducing the need for qualitative judgements. Although our methods can be improved upon, we urge public health practitioners to adopt the best available methods, because both our collective credibility and, more importantly, success in combating COVID-19 rest on using tools that are state-of-the-art.

FIGURE 1—

Deconvolution of SARS-CoV-2 Infection Date From Symptom Onset Date While Also Accounting for Misclassification Error in the Case Ascertainment of COVID-19 Surveillance Data: Georgia Department of Public Health, December 1, 2020–April 11, 2021

Note. Dashed gray line depicts the double adjustment method described in Goldstein et al.,⁴ with corresponding 95% uncertainty interval. Solid black line depicts the method described in Hennessee et al.¹ of infection date estimated as symptom onset date minus a median incubation period of 5.1 days. Tails of the distribution may be unreliable because of left and right truncation. Analytic codes are available for download from https://doi.org/10.5281/zenodo.5798398.