TL;DR
This paper introduces the BETS generative model to correct for severe selection biases in early COVID-19 epidemiological data. By accounting for rapid epidemic growth and travel quarantines, the authors found that the virus was doubling every 2 to 2.5 days in Wuhan—much faster than early high-profile estimates—and that roughly 5% of symptomatic cases may exceed the standard 14-day incubation window.
What Problem Does This Paper Address?
In the chaotic early weeks of 2020, researchers scrambled to estimate two "holy grail" numbers: the epidemic doubling time (how fast it spreads) and the incubation period (how long until symptoms appear). Early estimates were all over the map, with doubling times ranging from 2 to 7 days.
The authors argue that these discrepancies weren't just noise; they were the result of systematic selection bias. Early datasets mostly consisted of "exported cases"—people who traveled from Wuhan and were diagnosed elsewhere. These samples were inherently biased by:
- The Travel Quarantine: The abrupt lockdown of Wuhan on January 23, 2020, changed the "selection probability" of cases.
- Exponential Growth: In a rapidly growing epidemic, a person is statistically more likely to have been infected recently than a week ago. Ignoring this "exponential tilt" leads to overestimating the incubation period.
- Right-Truncation: Early studies only looked at cases confirmed by a certain date, naturally missing people with longer incubation periods who hadn't shown symptoms yet.
Methods
The authors proposed a generative model called BETS, named after four key events:
- Beginning of exposure
- End of exposure
- Time of transmission (infection)
- Symptom onset
Unlike previous models that treated infection times as uniformly distributed over an exposure window, BETS explicitly models the exponential growth of the virus within the likelihood function. They derived closed-form formulas to adjust for the January 23rd travel ban and the fact that only symptomatic, "exported" cases were being observed.
To ensure their results weren't just artifacts of a specific distribution (like a Gamma distribution), they also employed a Bayesian nonparametric analysis to model the incubation period's "tail" more flexibly.
Datasets
The study utilized a manually curated dataset of 1,460 confirmed COVID-19 cases across 14 locations (8 cities/provinces in mainland China and 6 international countries/regions). From this, they identified a high-quality subset of 378 Wuhan-exported cases who left the city before the quarantine. This subset was chosen because their travel history provided clear "exposure windows" (the time between arriving in and leaving Wuhan).
The authors have made their dataset and statistical programs publicly available as an R package: bets.covid19.
Key Insights and Ideas
- The "Balanced Out" Coincidence: The authors found that some early influential studies arrived at roughly correct median incubation periods only because two massive biases—right-truncation (which underestimates) and ignoring epidemic growth (which overestimates)—happened to partially cancel each other out.
- The 14-Day Tail: While the median incubation is about 4.5 days, the nonparametric model suggests the "tail" is heavier than a standard Gamma distribution predicts. About 5% of symptomatic patients might not show symptoms until after the standard 14-day quarantine period.
- Gender Differences: The data suggests men might be more likely than women to develop symptoms very quickly (within 2 days of infection), though men also showed a more variable overall incubation distribution.
- Why Doubling Time Matters: By correcting for the travel quarantine, the authors showed that the doubling time was likely 2.1 days, not the 6.4 days reported in some early Lancet papers. This difference is the difference between a manageable outbreak and a global explosion.
Key Results
- Doubling Time: Consistently estimated at 2.0 to 2.5 days across all sub-samples and models.
- Incubation Period:
- Median: ~4.5 days.
- 95th Percentile: ~13.5 to 15 days.
- Selection Bias Impact: The paper demonstrates that ignoring epidemic growth can lead to an overestimation of the incubation period by as much as 9 days in extreme cases.
Limitations and Future Work
- Self-Reporting Bias: The data relies on symptom onset dates reported by patients, which can be subject to memory errors or social pressure (e.g., downplaying symptoms to avoid travel restrictions).
- Under-ascertainment: The model assumes that the likelihood of being "captured" in the dataset isn't related to the length of the incubation period itself, which might not be entirely true if very sick people were prioritized for testing.
- Representativeness: The "exported" population (travelers) might have different demographics or health profiles than the general population of Wuhan, potentially affecting the generalizability of the incubation period findings.
Summary updated