A common issue that arises with the modelling of count data (Poisson, but also binomial when trial size is greater than one) is the need to account for over-dispersion.
You will focus on the case where over-dispersion has been introduced by an excess of zero counts. Typically, this is modelled by introducing a latent variable z to indicate if an observed count is a true zero or not. This means you are working with two models simultaneously, one for the true counts, and one for the extra counts.
In this project, you will develop Variational Bayes approaches for fitting these models and compare to MCMC based approaches. If there is time available, you would explore whether some combination of these methods could deliver the speed improvement of a Variational Bayes approach with the accuracy of MCMC methods.
You can either zero inflated Poisson, or both zero-inflated Poisson and Binomial models.
Supervisors
Primary Supervisor: John Holmes
Key qualifications and skills
Strong coding skills in R/Rcpp would be very useful.
Strong mathematical/statistical skills are needed. You will need to derive conditional posteriors and develop algorithms as part of the project.
Does the project come with funding
No - Student must be self-funded
Final date for receiving applications
Ongoing
How to apply
Apply by email to primary supervisor
Keywords
(approximate) Bayesian methods, Statistics, Generalised Linear Models
