Applications of Bayesian latent variable models to challenges in health and social care data

This portfolio seeks to demonstrate that: novel Bayesian models are useful for challenging data in health and social care, are computationally feasible with current software, can be agreed with subject expert colleagues, and findings can be communicated effectively. Latent variables are unknowns which take different values for observations or groups.

An opinion paper highlighted concerns about replicability of statistical findings, linking these to challenges in health and social care data. Bayesian models were identified as a preferred way of working. A tutorial paper considered the problem of converting a common but poorly understood statistic into a more accessible al ternative in a semi-Bayesian way. A scoping review of Bayesian meta-analyses searched two major biomedical databases, and examined papers for reporting quality and methodological rigour.

Four papers use item-response theory models for educational attainment data in health professions, and a quality of life scale. Four papers use latent variable models to investigate: diagnostic tests, rehabilitation after stroke, attrition from longitudinal studies, and performance of child protection services. Models were agreed with non-statistical colleagues and fitted to data using a variety of software. A methodological paper tested new methods to analyse large, rapidly arriving datasets, using non-parametric priors. Two papers present and test a new interface to Bayesian modelling software.

The applied papers demonstrated new substantive insights into their topics. Bayesian analyses can account for many common deficiencies in data. Models can be agreed with non-statistical colleagues and informed by existing research, theory and expert opinion. Non-parametric updating of Bayesian models in big data settings is possible using a novel method, and a Stata command was developed to broaden accessibility of Bayesian software to researchers.

In conclusion, Bayesian models with latent variables made several substantive insights possible despite challenges in the data. There are opportunities for further methodological developments.