Abstract
[This corrects the article DOI: 10.3389/fphar.2022.893484.].
Introduction
This corrigendum addresses a technical error in a large-scale evaluation of vaccine safety monitoring designs previously published in Frontiers in Pharmacology. While the primary scientific conclusions regarding the relative performance of different epidemiological designs remain intact, the corrected data reveals that detecting moderate risks is more challenging and time-consuming than initially reported.
The Nature of the Error
The original study utilized "positive controls"—simulated associations where a known signal is injected into the data—to evaluate the sensitivity of various surveillance methods. A software implementation error resulted in the intended effect sizes (Relative Risks of 1.25, 2, and 4) being applied twice. This effectively squared the values; for instance, an intended 4x risk was erroneously treated as a 16x risk.
Because larger effect sizes are significantly easier to detect, this error led to an underestimation of Type 2 errors (false negatives) and an overly optimistic "time to 50% sensitivity" (the duration of data collection required to have a 50% probability of detecting a signal).
Methods and Datasets
The researchers evaluated several standard epidemiological designs using the Optum EHR database:
- Case-control: Comparing cases with adverse events to matched or random controls.
- Cohort methods: Utilizing various comparators such as outpatient visits or historical data, often with Propensity Score (PS) matching or weighting.
- Self-Controlled Designs (SCCS/SCRI): Where individuals serve as their own controls, comparing post-vaccination risk windows to other periods.
The evaluation spanned several vaccines, including H1N1pdm, HPV (Gardasil 9), Seasonal flu (Fluvirin, Fluzone), and Zoster (Shingrix).
Updated Error Rates
The corrected Figure 4 illustrates the trade-off between Type 1 (false positive) and Type 2 (false negative) errors. Empirical calibration—using negative controls to adjust p-values—effectively minimizes Type 1 errors across designs. However, the corrected data shows that Type 2 error rates are higher than previously estimated, particularly for smaller effect sizes.
Time to Detection
The corrected Figure 5 provides a realistic timeline for safety signal detection. For a moderate effect size (e.g., a doubling of risk), many surveillance designs require 6 to 12 months of data collection in a large EHR database to achieve 50% sensitivity.
Conclusion
The "double-multiplication" error serves as a critical reminder of the complexities involved in large-scale computational epidemiology. While the absolute metrics for sensitivity and error rates have been adjusted, the relative rankings of the methods remain consistent. The authors reaffirm that empirical calibration is essential for reducing false positives, though it necessitates a longer surveillance period to reliably detect true safety signals.
Summary updated