Alice Corbella, Anne M Presanis, Paul J Birrell and Daniela De Angelis's contribution to the Discussion of ‘The Second Discussion Meeting on Statistical aspects of the Covid-19 Pandemic’

We congratulate Storvik and co-authors on this excellent work that combines elaborate models and cutting-edge inference in an impactful setting.

The model used to describe the dynamics that contribute to the generation of COVID-19 data comprises environmental, demographic, and observational randomness, with the final goal of appropriately allocating the stochasticity to each of these three components.

Among several innovations that are presented, we have particularly appreciated the presence of stochastic delays in the severity process and their handling in the Bayesian inference using sequential Monte Carlo (SMC). In our related work (Corbella et al., 2022), we highlight that the severity process plays a substantial role in generating randomness observable in the data, and that the delay between infection and the case becoming detectable via (severe) symptoms should not be ignored. A substantial body of related work to account for delays between infection and symptom onset/observation traces back to stochastic back-calculation methods used to estimate incidence of HIV/AIDS (Brookmeyer et al., 1994) and has advanced since (e.g., Brizzi et al., 2019; Marschner, 1994; Mezzetti & Robertson, 1999; Sweeting et al., 2005).

In some contexts, the use of multiple data streams, such as the community testing data and hospital admissions used by Storvik et al., should be accompanied by an assessment of potential dependency between the datasets. The authors have considered these two streams as independent signals of the same underlying process, which may be plausible for a context where those hospitalised are unlikely to have had time to be tested prior to admission. However, there are other situations (e.g., Corbella et al., 2022) where accounting for possible hidden dependence between data streams is important.

Lastly, we have enjoyed reading the description of the computational methods very much. In our experience, the implementation of the SMC sampler is often non-trivial but enables inference when richer models are assumed, such as the three models adopted for R_t. In our understanding, the authors have exploited results from a previous study on the same data (Engebretsen et al., 2021), to fix several parameters of the transmission model, enabling the fast inference using SMC. Does such use of the data in two stages not risk overstating the authors’ certainty about R_t? Perhaps a two-phase approach, where the static parameters are inferred using data on the first wave, then are assumed known at the beginning of the second wave, to drive the SMC inference, might be more realistic.

Funding

This research was funded in whole or in part by EPSRC grant New Approaches to Bayesian Data Science: Tackling Challenges from the Health Sciences [AC, DDA: EP/R018561/1], and by the UKRI Medical Research Council [DDA, AMP, PJB: Unit Programme number MC/UU/00002/11].

Data availability

This work is entirely theoretical, there is no data underpinning this publication.

References

Author notes