You are on page 1of 1

Correspondence

COVID-19 and the how to usefully incorporate the six The second problem is that,
positive cases from this sample is generically, Bayesian models describe
difficulty of inferring difficult. uncertainty both in the data and in Lancet Infect Dis 2020
epidemiological Case mortality data from China prior beliefs about the studied system. Published Online
provide an upper bound for the IFR, Only when data are informative about May 28, 2020
parameters from and with extra assumptions these the targets of modelling can we be
https://doi.org/10.1016/

clinical data
S1473-3099(20)30437-0
data also supply information on how sure that prior beliefs play a small
the IFR varies with age. Because pre­ role in what the model tells us about
Knowing the infection fatality valence is unknown, the data contain the world. In this case, the data are
ratio (IFR) is crucial for epidemic no information for estimating the especially uninformative: we suspect
management: for immediate absolute IFR magnitude. results are mostly the consequence of
planning, for balancing the life- Because of extensive testing, the what our prior beliefs were.
years saved against those lost to the outbreak on the Diamond Princess, Taken together these problems
consequences of management, and the quarantined cruise liner (used indicate that Verity and colleagues’
for considering the ethics of paying only for validation by Verity and IFRs should be treated very
substantially more to save a life-year colleagues) supplies data on infections cautiously when planning epidemic
from the epidemic than from other and symptomatic cases, with fewer management. While awaiting actual
diseases. Impressively, Robert Verity ascertainment problems. These data measurements, we would base IFRs
and colleagues1 rapidly assembled case appear directly informative about the on the Diamond Princess outbreak
data and used statistical modelling to IFR, although the comorbidity load data, with the Chinese case-fatality
infer the IFR for COVID-19. We have on the Diamond Princess is unlikely data informing the dependence of
attempted an in-depth statistical to fully represent any population of IFR on age. We have included a crude
review of their paper, eschewing serious interest (perhaps having fewer Bayesian model with its IFR estimates
statistical nit-picking, but attempting individuals with very severe but more by age in the appendix. IFR estimates See Online for appendix
to identify the extent to which the with mild comorbidities). for corresponding populations are
(necessarily compromised) data are Second, the modelling assumptions, China 0·43% (95% credible interval
more informative about the IFR than in which we see two primary problems. 0·23–0·65), UK 0·55% (0·30–0·82),
are the modelling assumptions. First, The first problem is that Verity and and India 0·20% (0·11–0·30). The
the data. colleagues correct the Chinese case strong assumptions required, by
Individual-level data for outside data by assuming that ascertainment this approach too, emphasise the
China appear problematic because differences across age groups deter­ need for improved data. We should
different countries have differing mine case rate differences. Outside of replace complex models of inadequate
levels of ascertainment and different Wuhan, the authors replace observed clinical data with simpler models
disease-severity thresholds, even for case data by the cases that would of epidemiological prevalence data
classification as a case. The data’s use have occurred if each age group had from appropriately designed random
in IFR estimation would require model- the same per-capita observed case sampling using antibody or PCR tests.
ascertainment parameters that are rate as the 50–59 years age group. We declare no competing interests.
country specific, about which we have The authors assume complete
*Simon N Wood, Ernst C Wit,
no information. Consequently, these ascertainment for this age group. Matteo Fasiolo, Peter J Green
data provide no useful information on These are very strong modelling simon.wood@bath.edu
the IFR. assumptions that will greatly affect the
Università della Svizzera Italiana, Switzerland (ECW);
Repatriation flight data provide the results, but the published uncertainty and School of Mathematics, University of Bristol,
sole information on the prevalence bounds reflect no uncertainty about Bristol BS8 1UG, UK
in Wuhan, China (excepting the these assumptions. In Wuhan, the 1 Verity R, Okell LC, Dorigatti I, et al. Estimates of
lower bound of confirmed cases). complete ascertainment assumption the severity of coronavirus disease 2019:
a model-based analysis. Lancet Infect Dis 2020;
689 foreign nationals who were is relaxed by introducing a parameter, published online March 30. https://doi.
eligible for repatriation are doubtfully but one for which the data appear org/10.1016/S1473-3099(20)30243-7.
representative of the susceptible uninformative, so the results will be
population of Wuhan. Hence, seeing driven by the assumed uncertainty.

www.thelancet.com/infection Published online May 28, 2020 https://doi.org/10.1016/S1473-3099(20)30437-0 1

You might also like