Time Series Analysis of Covid-19 Excess Mortality Rates

Date: 
Monday, May 4, 2020 - 4:15pm
Location: 
Virtual
Affiliation: 
UC Berkeley
Public Lecture: 
Yes

 

Professor Uroš Seljak speaks about how machine learning methods originally applied for astrophysical purposes give insight into the mortality of COVID-19.  Vir Bulchandani introduces a branching process model for COVID-19 suggesting that the efficacy of app-based contact tracing may be controlled by a novel phase transition. 

Abstracts

Professor Uros Seljak
How deadly is COVID-19? A time-series analysis of mortality data

How many people have died of COVID-19? The answer is surprisingly difficult to obtain, because so many fatalities occur outside hospitals without being tested. We propose a counterfactual analysis, with data from the previous five years as control, which we evaluate using a Gaussian Process. When applied to 2020 mortality data reported from towns in Italy, it reveals a large excess of deaths in March 2020, and far above official COVID-19 mortality,  increasing with age above 70 years, suggesting there is a large population of predominantly old people missing from the official fatality statistics. We estimate that the number of COVID-19 deaths in Italy is 52,000 +/- 2000 as of April 18 2020, a factor of 2 higher than the official number. How deadly is COVID-19 once infected? This is difficult to answer because we do not know the infection rate, but we can bound it from below using total mortality and upper bounds on infection rates. We determine infection fatality rate (IFR) lower bound of 0.8% for Lombardia and infection rate of 23% for Lombardia, a factor of 35 above the number of positive tests. Such analyses can help predict the corresponding numbers here in USA: we observe that the COVID-19 mortality tracks closely the overall age-specific yearly mortality rate of the underlying population. With this we predict 0.5% lower bound on IFR for NYC and California, in a good agreement with the existing data. 

 

Vir Bulchandani 
Digital Herd Immunity and COVID-19

According to the traditional notion of “herd immunity”, a population can be immune to epidemics even if not all of its individual members are immune to the disease. We introduce an analogous notion of “digital herd immunity”, which is similarly an emergent characteristic of the population. This immunity arises because contact-tracing protocols based on smartphone capabilities can lead to highly efficient quarantining of infected population members and thus the extinguishing of nascent epidemics. As usage decreases there is a novel "contact-tracing phase transition” to an epidemic phase. We present and study a simple branching-process model for COVID-19 and show that digital immunity is possible regardless of the proportion of non-symptomatic transmission. Applications to the US and India are briefly discussed.