Non-parametric Bayesian updates by kernel densities
One of the big attractions for people adopting Bayesian methods is the promise of "updating" their parameter estimates and predictions as more data arrive. Yesterday's posterior becomes today's prior. In practice, this is not always simple, requiring at the very least a complete set of sufficient statistics, random (conditional on the model) samples from an unchanging population, and no changes of probability distribution for the priors. Sometimes, one would like to update without imposing an a priori distribution on yesterday's posterior and without estimating lots of statistics. This project evaluates a kernel approach, which is easily incorporated in Stan by an additional target+= statement, or in BUGS/JAGS via the "ones trick" with uniform proposal densities. We compare this with parametric updates, and explore the potential to reduce computation by using kernels weighted by counts of posterior draws inside hypercubes of parameter space.
You can access more detail and all the code (in R+Stan) here.
Rehabilitation after spinal fusion surgery
A collaboration with University College Hospital and St George's, University of London to design and run a clinical trial of an enhanced physiotherapy rehab programme. Rather than just running some simple endpoint hypothesis test, we are helping to design a Bayesian structural equation model that will take all the aspects of recovery (physical, psychological and social) into account together.
Simulation methods in agent-based models for sustainability transitions
We are contributing statistical advice to this project of the Global Climate Forum. The goal is reliable but complex modelling of how social networks and economies evolve as environmentally sustainable technologies are adopted, for example switching from hydrocarbon to electric cars.
Bayesian meta-analysis of exercise for osteoarthritis
This project re-analyses the Cochrane review of exercise interventions for osteoarthritis, adopting a Bayesian approach to allow all stats from studies to be included. The aim is to investigate the feasibility of such a fully Bayesian evidence synthesis approach, in accounting for the real-life shortcomings in the studies and their reporting, in achieving consensus on model and priors, and in communicating it to a clinical / researcher audience. We are working with Professor Mike Hurley and Dr Rachel Hallett at St George's, University of London.
International Workshop on Computational Economics and Econometrics
This event is organised by the National Research Council of Italy's Research Institute on Sustainable Economic Growth (IRCrES) each summer, and attracts researchers, both applied and methodological, from across Europe. BayesCamp contribute to organisation, invite one of the invited speakers each year, and – because communication outside our own fields is important – provide a prize for the talk with the clearest exposition.
More research coming soon, including meta-analysis code, algorithms to deal with huge numbers of group-level parameters efficiently, and alternatives to random forests or xgboost, when you have a small dataset.
You can contact Robert at email@example.com to discuss bespoke training for your team, or one-to-one coaching.
BayesCamp Ltd is incorporated in England and Wales, number 10666858.
VAT registration: 317349595
The term 'BayesCamp' and the Gaussian tent logo are registered trademarks.