10 04
Αίθουσα Α31
Ώρα 14:15
|
A Bayesian approach to the analysis of infectious disease data using continuous-time stochastic models.
Πέτρος Μπαρμπουνάκης
Abstract: The aim of this work is the development of stochastic epidemic models focused
on disease outbreaks in humans, as well as livestock. Statistical methodology is
developed aimed at informing public health policies and their communication
as implemented by the governing organizations, specifically at a time of crisis
like the Covid-19 pandemic.
The first part is concerned with the results of a simulation-based evaluation of several policies for vaccine roll-out. Particular focus is placed upon
on the effects of delaying the second dose of two-dose vaccines. In the presence of limited vaccine supply, the specific policy choice was a pressing issue for
several countries worldwide, and the adopted course of action affected the extension or easing of non-pharmaceutical interventions (NPIs). We used a suitably
generalised, age-structured, stochastic SEIR (Susceptible, Exposed, Infectious,
Removed) epidemic model that accommodates quantitative descriptions of the
major effects resulting from distinct vaccination strategies. The different rates
of social contacts among distinct age-groups (as well as other model parameters)
are informed by a recent survey conducted in Greece, but the conclusions are
widely applicable. The results are summarised and evaluated in terms of the
total number of deaths and infections as well as life years lost.
A number of NPIs had been implemented in order to reduce transmission,
thus leading to multiple phases of transmission. The disease reproduction number Rt, a way of quantifying transmissibility, has been a key part in assessing
the impact of such interventions. In the second part of this work we discuss
the distinct types of transmission models used and how they are linked. We
consider a hierarchical stochastic epidemic model with piece-wise constant Rt,
appropriate for modelling the distinct phases of the epidemic and quantifying
the true disease magnitude. The location and scale of Rt changes are inferred
directly from data while the number of transmissibility phases is allowed to vary.
We determine the model complexity via appropriate Poisson point process and
Dirichlet process-type modelling components. The models are evaluated using
synthetic data sets and the methods are applied to freely available data from the
United Kingdom and Greece as well as California and New York states. We estimate the true infected cases and the corresponding Rt, among other quantities,
and independently validate the proposed approach using a large seroprevalence
study.
The final part is concerned with a class of models where the Ornstein-
Uhlenbeck (OU) process is embedded within Poisson-type point processes. We utilise a general OU model with Student’s t-distribution marginals and a Cox-
Ingersoll-Ross model for the latent infection rate of the spatio-temporal model. We also propose a class of Bayesian Neural Nets using horseshoe priors for the
weights. Real data from Foot and Mouth and Sheep-pox outbreaks in livestock
within the Evros region of Greece are studied. The predictive ability of each model is being assessed using proper scoring rules within the prequential analysis framework. Our investigation concludes that the Student-t OU and the CIR models improve upon the previously introduced models with Gaussian OU for
the latent rate of the Poisson-type point process.
Βιβλιογραφία
- --
- --
|