Accurate knowledge of population exposure at the outset of a pandemic has critical ramifications for preparedness plans for future epidemic waves. In this thesis, I developed a mechanistically informed statistical model to integrate multiple epidemiological datasets in different settings and in different population and to estimate key epidemiological parameters as well as population exposure using Bayesian inference.
First, I present a dynamic model to link together three key metrics for evaluating the progress of COVID-19 epidemic in England: seroprevalence, PR-PCR test positivity and death. While estimating the IgG antibody seroreversion rate and region-specific infection fatality ratios, I find that epidemic progression resulted in an increasing gap between measured serology prevalence levels and cumulative population exposure to the virus. Ultimately, this may mean that twice as many, or more, people have been exposed to the virus relative to the number of people who are seropositive by the end of 2020.
Moreover, I demonstrate that the model could reconstruct the first, unobserved, epidemic wave of COVID-19 in England from March 2020 to June 2020 as long as two or three serological measurements are given as inputs, with the second wave during the winter of 2020 validated by the estimates from the ONS Coronavirus Infection Survey. Comparing with the inferred exposure, I find that the UK official COVID-9 online dashboard reported COVID-19 cases only accounted for less than ten percent by the end of October 2020. I then generalise the model to account for the undocumented COVID-19-related mortality and sparse measurements of seroprevalence. I apply this in the context of Afghanistan COVID- 19 epidemic and find the population exposure in nine regions of Afghanistan were all higher than the seroprevalence survey suggested by July 2020.
Finally, I assess the impact of shielding among pregnant patients by comparing their exposure with the estimated exposure of the general population. To approach this, I develop a dynamic model to link RT-PCR and antibody testing results from patients who gave birth and then apply Bayesian inference to estimate transmission parameters and exposure among pregnant patients. I find that after considering the duration of each pregnancy pre-COVID onset and after, the impact of shielding on reducing the level of exposure among pregnant patients during early 2020 who gave birth in this New York City hospital were approximately 50%.
The thesis is accessible from Oxford University library https://ora.ox.ac.uk/objects/uuid:562dda97-ca81-40e2-96ec-0dc5cdb51d19.