Developing methods for model selection in causal health analyses.

点击翻译按钮获取中文摘要

This PhD project aims to conceptualise and implement an automated Bayesian system that can make decisions about statistical modelling in the context of causal health - these decisions include model form selection, variable selection, parameter fitting techniques and model evaluation methods / success metrics. Present developments are looking into causal inference, specifically the role of confounders and mediators in variable selection. We investigate different variable selection rubrics and confounder selection strategies to assess the suitability of different methods at forming causally valid statistical models. We seek to, from here, consider attempts at classifying variables (confounders, mediators or other types / properties) using Bayesian weighting where such categorisations are both subject to real-world interpretation and not typically identifiable from within the data. The broader project requires making use of larger more complicated data sources, plus priors for key variables. Lastly, we consider taking into account missing-ness (categorising types of missing data such as MCAR, MAR and MNAR), measurement error, variable categorisation and the underlying causal diagram (DAGs). The novelty of the research arises from the consideration of effect-interactions between variables and use of novel Bayesian techniques. The intended impact of the research would be on statistical models and their evaluation within medical papers, such as randomized control trials for new pharmaceuticals, wherein one typically employs a unit-block treatment model as a means of measuring treatment effects of different medicines on patients. This would improve the strength and statistical validity of papers in the medical literature, in turn improving decision making by medical practitioners in the real-world, thus improving patient outcomes. This project falls within the EPSRC mathematical sciences research area with a goal towards transforming health and healthcare. This project is supervised jointly by three professors: Kate Tilling (University of Bristol, Bristol Medical School), Jonathan Sterne (University of Bristol, Bristol Medical School) and Rhian Daniels (University of Cardiff, Department of Statistics). This PhD is funded jointly from three sources: 25% by NHS BRISTOL, 50% by EPSCRC COMPASS CDT, 25% by Faculty Funded PGR Studentship CDT funding. The project does not have industry partners as such, but Emma has worked alongside Southmead Hospital Bristol as part of her research.