- Type
- Research
- Focus
- Bayesian inference · Survival analysis · R
- Context
- MSc research project · Western University
My contribution
Independently derived and implemented a mean-field variational Bayes algorithm for right-censored Weibull AFT models, using piecewise linear and quadratic approximations for closed-form coordinate-ascent updates.
Method and implementation
I developed the method as my MSc research project at Western University and independently implemented it in R. The model handles right-censored survival outcomes, while the approximations make coordinate-ascent updates available in closed form.
Simulation study
I benchmarked the method against maximum-likelihood estimation and Stan MCMC across six simulation scenarios, with 500 retained datasets per scenario.
The study recorded an average 453× fitting-time speed-up on successful fits, while producing similar point estimates. This is a success-only fitting-time result: it does not describe total experiment runtime or guarantee the same speed-up on other datasets.
METABRIC application
The application used 1,815 complete METABRIC breast-cancer records, with 42.6% censoring and 11 covariate terms. The method reproduced MCMC posterior summaries and yielded a reasonable Cox–Snell model-fit diagnostic.