An indirect method to monitor the fraction of people ever infected with COVID-19: An application to the United States

Abstract

The number of COVID-19 infections is key for accurately monitoring the pandemics. However, due to differential testing policies, asymptomatic individuals and limited large-scale testing availability, it is challenging to detect all cases. Seroprevalence studies aim to address this gap by retrospectively assessing the number of infections, but they can be expensive and time-intensive, limiting their use to specific population subgroups. In this paper, we propose a complementary approach that combines estimated (1) infection fatality rates (IFR) using a Bayesian melding SEIR model with (2) reported case-fatality rates (CFR) in order to indirectly estimate the fraction of people ever infected (from the total population) and detected (from the ever infected). We apply the technique to the U.S. due to their remarkable regional diversity and because they count with almost a quarter of all global confirmed cases and deaths. We obtain that the IFR varies from 1.25% (0.39–2.16%, 90% CI) in Florida, the most aged population, to 0.69% in Utah (0.21–1.30%, 90% CI), the youngest population. By September 8, 2020, we estimate that at least five states have already a fraction of people ever infected between 10% and 20% (New Jersey, New York, Massachussets, Connecticut, and District of Columbia). The state with the highest estimated fraction of people ever infected is New Jersey with 17.3% (10.0, 55.8, 90% CI). Moreover, our results indicate that with a probability of 90 percent the fraction of detected people among the ever infected since the beginning of the epidemic has been less than 50% in 15 out of the 20 states analyzed in this paper. Our approach can be a valuable tool that complements seroprevalence studies and indicates how efficient have testing policies been since the beginning of the outbreak.

SARS-CoV-2 infections in 165 countries over time

“…The above observations imply that comparisons of the epidemic's extent and progression between countries should preferably be done based on infection or death counts, rather than reported cases alone (Flaxman et al, 2020;Sánchez-Romero et al, 2021). For example, as of June 25, 2021 the cumulative per-capita number of cases reported for the Czech Republic (16%) and Slovenia (12%) were much higher than for Paraguay (5.8%), Peru (6.3%) or Brazil (8.6%), while the median predicted cumulative infection fractions for the Czech Republic (52%) and Slovenia (38%) are much lower than for Paraguay (86%), Peru (98%, Fig.…”

“…In contrast to case reports, COVID-19-related death counts are generally regarded as less sensitive to testing effort and strategy (Flaxman et al, 2020;Lau et al, 2021;Lu et al, 2020;Maugeri et al, 2020a), and fortunately most countries have established nationwide continuous reporting mechanisms for COVID-19-related deaths. Hence, in principle, knowing the infection fatality risk (IFR, the probability of death following infection by SARS-CoV-2) should permit a conversion of death counts to infection counts (Bohk-Ewald et al, 2020;Flaxman et al, 2020;Lu et al, 2020;Sánchez-Romero et al, 2021). The IFR of SARS-CoV-2, however, depends strongly on the host's age, and hence the effective IFR of the entire population depends on the population's age struc-ture as well as the disease's age distribution (Dowd et al, 2020).…”

“…In principle, one could also first determine the "effective" (integrated over all ages) IFR for the entire population and combine that effective IFR with total (non-age-stratified) death counts to estimate infection rates. This approach was taken by Sánchez-Romero et al (2021), who first estimated the effective IFR for various states in the US based -among others -on age-specific mortality data and then estimated the cumulative number of SARS-CoV-2 infections across the US as of September 8, 2020. However, such an effective IFR is specific to the population for which it was estimated, and hence applying it to other countries (even if correcting for the local population age structure, which is possible in the framework by Sánchez-Romero et al (2021)) would fail to account for differences (or uncertainty) in the age distribution of infections or deaths.…”

“…However, using a global average age distribution of deaths ignores the fact that the age distribution of infections (and deaths) actually needs to be adjusted for each country's population age structure, even if any given age group were to experience a similar exposure in each country. Further, while the approaches by Bohk-Ewald et al (2020) and Sánchez-Romero et al (2021) can account for the average time lag between infection and death, they cannot account for its actual probability distribution and considerable spread around the mean (Linton et al, 2020), which further complicates the estimation of time-resolved infections from deaths. Lastly, all of the above studies cover only an early portion of the pandemic (Bohk-Ewald et al, 2020;Flaxman et al, 2020) or only focus on a single time point (Sánchez-Romero et al, 2021), and focus on a small number of countries (1-11).…”

Underdetection in a stochastic SIR model for the analysis of the COVID-19 Italian epidemic

“… 2020 ) comparing the case fatality rate (CFR) to an estimate of the key parameter, infection fatality rate (IFR) (Sánchez-Romero et al. 2021 ); from exported cases and air travel volume (Tuite et al. 2020 ); from epidemiological and population data using deterministic models (Krantz and Rao 2020 ); by post-stratification sampling techniques (Bassi et al.…”